Insight Editions · No. 279
Empire of AI
Dreams and Nightmares in Sam Altman's OpenAI
Karen Hao
Karen Hao's account of how OpenAI, founded as a nonprofit to keep AI out of corporate hands, became the engine of an empire of AI, and of the labour, data and environmental costs of its scaling race.
A study guide to Empire of AI by Karen Hao. Shared free for personal reading; not for sale and not endorsed by the authors or publishers. Please read and buy the original.
The Book on One Page
Hao argues that OpenAI, and the AI industry it pulled after it, behaves like an empire: it seizes data, labour, land, energy and water, justifies the taking with a civilising story about AGI that "benefits all of humanity," and concentrates the gains in a handful of firms and men. The scaling path that made ChatGPT possible was a choice, not a law of nature, and the costs of that choice fall on people who never get a vote. Her test for any technology is whether it consolidates power or redistributes it.
The argument in eight moves
- 1. A governance experiment that failed. OpenAI was founded in 2015 as a nonprofit check on Google. In 2023 its two senior deputies, Murati and Sutskever, gave the board a documented pattern of Altman telling people different things; the directors fired him, yet could not reveal their evidence without exposing their sources, and the five-day revolt of employees (with a tender offer at stake), investors and Microsoft restored him (Prologue, Ch. 13, 15-16).
- 2. Founders driven by ego as much as altruism. Elon Musk's fear of DeepMind met Altman's gift for accumulating power, honed under Paul Graham and Peter Thiel; the same charges of self-dealing and truth-bending that dogged him at Loopt resurfaced in 2023 (Ch. 1).
- 3. The compute thesis broke the nonprofit. Belief that AGI required compute doubling every 3.4 months ("OpenAI's Law") meant billions a nonprofit could not raise, which drove Musk out, created the capped-profit LP and brought in Microsoft's $1 billion in 2019 (Ch. 2-3).
- 4. Scaling was a bet, not destiny. Deep learning won funding partly because it fit ad-driven business models; Sutskever's faith, the Transformer and the scaling laws turned that into a doctrine of "scale, scale, scale" from GPT-2 to GPT-4 (Ch. 4-6, 10).
- 5. Scale shifted the costs outward. Bigger models needed dirtier data, so labour moved from curating inputs to policing outputs: Kenyan workers at $1.46-$3.74 an hour reading abuse, and crisis-hit Venezuelans paid pennies (Ch. 9); and megacampuses drawing water and power from drought-stricken Chile, Uruguay, Iowa and Arizona (Ch. 12).
- 6. Science and policy were captured. Google forced out Timnit Gebru over "Stochastic Parrots"; labs stopped disclosing data; and Altman's Senate testimony steered Washington toward licensing hypothetical "frontier" models above 10^26 FLOPs rather than present harms (Ch. 7, 13).
- 7. Ideology supplied the moral cover. Doomers (effective altruism) and Boomers (e/acc) loathe each other but share one creed: AGI is inevitable and they should control it. Their war, plus Altman's habit of telling each side what it wanted, produced The Divorce, The Blip and the Omnicrisis of May 2024, when the equity clawback that silenced departing staff and the Sky voice made many employees wonder whether the board had been right (Ch. 6, 10, 13, 17).
- 8. The formula, and its antidote. The mission rallies talent and capital and stays vague enough to reinterpret at will: from open nonprofit (2015) to capped profit (2019), closed API (2020), racing ChatGPT out (2022) and a planned public benefit corporation (2024) (Ch. 18). The alternative is small, consensual, community-governed AI such as Te Hiku Media's te reo model (310 hours of donated audio, two GPUs), and redistributing power over knowledge, resources and influence (Epilogue).
Hao's empire metaphor has three parts that recur in every chapter: resources taken without consent, labour devalued at the base, and a civilising mission that excuses both. Read each story (Kenya, Atacama, the boardroom) as one of these three.
Chapter map
| Ch | Title | One-line payload |
|---|---|---|
| Pro | A Run for the Throne | The 2023 firing shows mission-first governance failed; AI giants act as empires |
| 1 | Divine Right | Musk's DeepMind fear meets Altman's talent for power; Loopt foreshadows 2023 |
| 2 | A Civilizing Mission | Openness was positioning; the compute thesis forced the LP and Microsoft deal |
| 3 | Nerve Center | Hao's 2019 embed: no definition of AGI, only "first or perish" |
| 4 | Dreams of Modernity | Deep learning won because it suited business, not inevitably; flaws persist |
| 5 | Scale of Ambition | Sutskever, the Transformer and scaling laws; bigger data, worse data |
| 6 | Ascension | Altman's 10x doctrine splits the clans; the Amodeis leave for Anthropic |
| 7 | Science in Captivity | Google fires Gebru over "Stochastic Parrots"; transparency collapses |
| 8 | Dawn of Commerce | The 2021 road map fuses research and product; Codex; Altman's big bets |
| 9 | Disaster Capitalism | Data labour recruits from collapsing economies; Kenyan filter workers |
| 10 | Gods and Demons | EA versus e/acc; DALL-E 2 and GPT-4 ship under safety fights |
| 11 | Apex | ChatGPT's accidental triumph flips Microsoft and starts the compute hunger |
| 12 | Plundered Earth | Megacampuses drain water and power in Chile, Uruguay, Iowa, Arizona |
| 13 | The Two Prophets | Altman steers policy to "frontier" risk; the board loses its grip |
| 14 | Deliverance | Annie Altman's poverty and allegations: abundance that never reached her |
| 15 | The Gambit | October 2023: Murati and Sutskever take Altman's playbook to Toner |
| 16 | Cloak-and-Dagger | The Blip: secret deliberation, then collapse once Murati wavers |
| 17 | Reckoning | May 2024 Omnicrisis: GPT-4o, exits, equity clawbacks, the Sky voice |
| 18 | A Formula for Empire | A vague mission centralises power; 2024 exits and the PBC plan |
| Epi | How the Empire Falls | Te Hiku's consensual model; redistribute knowledge, resources, influence |
A Run for the Throne
Hao opens with the five-day firing and reinstatement of Sam Altman in November 2023 and reads it as proof that OpenAI's experiment in mission-first governance failed: a nonprofit board with the legal power to fire its CEO buckled within days under employees, investors and Microsoft. Her organising claim is that the AI giants are best understood as empires, seizing data, labour, land, energy and water while selling a vision of progress whose gains flow upward. The prologue sets out both halves of the book: the inside story of one secretive company, and the global footprint of the technology it pushed to the front.
Author's note: how the book was reported
Hao's evidentiary base is unusually explicit, and it matters for how much weight her scenes can bear.
- Scale of interviewing. More than 300 interviews with about 260 people, most done for this book, some drawn from seven years of reporting for MIT Technology Review, The Wall Street Journal and The Atlantic.
- OpenAI insiders. Over 150 interviews with 90-plus current or former OpenAI executives and employees, plus a few contractors with access to documentation of model-development practices.
- The wider industry. About 40 current and former staff at Microsoft, Anthropic, Meta, Google, DeepMind and Scale, and people close to Altman.
- Documents. Quoted emails, documents and Slack messages come from copies, screenshots or lawsuit filings; where she lacks a copy she paraphrases without quotation marks (one marked exception in the endnotes).
- Dialogue. Reconstructed from memory or contemporaneous notes, or taken from recordings when the endnotes say so. Her team often asked sources to repeat quotes several months apart to test their stability.
- Corroboration rule. Every scene, number, name, code name and technical detail about OpenAI's models (training-data composition, chip counts) rests on at least two people, with notes and documents, or occasionally on other media reporting (marked).
- Inner states. Thoughts and feelings are attributed only when the person described them to Hao, to a source, in a document or in a public interview. Naming someone does not mean she interviewed them.
She insists this is not a corporate book. OpenAI's inside story is a prism: a profile of a scientific ambition that became an ideological, money-fuelled quest, a survey of its global footprint, and a study of power. To that end she embedded with communities on several continents feeling AI's physical effects, and wants their stories to carry equal weight with the boardroom. She asked every key figure and company for interviews; OpenAI and Sam Altman declined to cooperate.
A methodological footnote also governs the whole book: AI researchers routinely post papers straight to arXiv, the open preprint server, and either face peer review months later or never. Citation in the field follows impact rather than review status, and Hao does the same, flagging non-peer-reviewed preprints in the endnotes.
Friday, November 17, 2023: the firing
Around noon Pacific time, Altman was in a Las Vegas luxury hotel for the city's first Formula One race in a generation, a short break from the gruelling travel he had kept up since ChatGPT's launch. On a Google Meet, four of his five board members appeared, and chief scientist Ilya Sutskever told him he was being fired, with the announcement imminent. Stunned, Altman reverted to his habitual instinct to smooth things over and asked how he could help; the board told him to back the interim CEO they had picked, CTO Mira Murati, and he agreed.
Minutes later Sutskever sent a second link to president Greg Brockman, Altman's close ally and the only director absent from the first call: he was off the board but would keep his job. The public statement followed. It cited a deliberative review and said Altman had been "not consistently candid in his communications with the board," so the board no longer had confidence in his leadership.
Why the timing looked absurd
On the surface OpenAI was at its peak:
- ChatGPT, launched in November 2022, was the fastest-growing consumer app in history.
- A tender offer valuing the company at up to $90 billion, letting employees sell shares to investors, was weeks from completion.
- Days earlier it had held a celebrated launch event for its most aggressive product slate yet.
- Altman's spring-summer world tour had made him a celebrity the press compared to Taylor Swift.
- Just before Vegas, at the APEC CEO Summit, Laurene Powell Jobs had asked why he devoted his life to the work; he called AI the most transformative and beneficial technology humanity would invent, and said he had now four times been in the room when the frontier of discovery moved forward, most recently a couple of weeks earlier.
The all-hands that backfired
Employees, by now nearly 800, learned of the firing from the public post like everyone else. What shocked them was the gap between the news and Altman's reputation; his stage charm matched how he behaved at all-hands and around the office. The rumour mill ran wild on X: someone shouted that Altman was running for president, which broke the tension briefly. Others wondered about something illegal, about his sister's viral allegations a month earlier that he had abused her, or about ethically dubious dealings tied to his other investments or his fundraising with Saudi investors for an AI chip venture. Sutskever announced on Slack a virtual all-hands in two hours, which one employee called the longest two hours ever.
Hao sketches Sutskever as the wrong messenger: a deep thinker and mystic who spoke in spiritual terms, also gentle and goofy, wearing animal shirts and painting cuddly animals; one office painting showed flowers shaped like OpenAI's logo, his emblem for the goal he preached: many AGIs that love humanity. He was not good at landing a message.
- Was there a specific incident? Sutskever said he could not give details and told staff to reread the press release, several times if needed, because the answer was "actually there."
- Microsoft? Murati said the relationship was unaffected; she and others had just spoken to Satya Nadella and CTO Kevin Scott. The stakes were existential: Microsoft was OpenAI's biggest backer, exclusive licensee and sole compute supplier.
- The tender? COO Brad Lightcap waffled that "we're going to see," while claiming investors had pledged support. For tenured staff the tender meant millions of dollars; many had bought or planned to buy homes.
- Would details ever come? Sutskever: "Keep your expectations low."
- Unity through crisis? Sutskever predicted the ordeal would make the team closer and more productive; an employee replied that truth is a precondition for reconciliation, and he conceded the situation was not perfect.
- Reassurance. Murati said the mission was bigger than anyone; Lightcap said partners still believed in it. Sutskever urged staff to picture the size of the GPU cluster when afraid.
- Hostile takeover? Asked whether board members had been coercively influenced, Sutskever bristled: the nonprofit board had acted entirely within its objective.
- Outcome. Unease turned to anger; the performance shattered the appearance of unity among the leaders on screen.
Friday night: the executives turn on the board
That evening employees attended a pre-planned party with guests from Google DeepMind and Anthropic; a "no-OpenAI talk room" was added and left mostly empty. By then Brockman had quit in protest, and Nadella, furious at being told only minutes before the announcement, tweeted a carefully worded statement stressing Microsoft's long-term agreement and full access to what it needed, and commitment to "Mira and the team." Then three senior researchers quit: Jakub Pachocki and Szymon Sidor, among the longest-tenured staff, and Aleksander Mądry, an MIT professor on leave. Staff feared a talent bleed would scare investors, kill the tender or sink the company.
Meanwhile roughly a dozen executives, including Murati and Lightcap, most of whom had learned of the firing only moments before it went public, demanded a meeting with the board. Sutskever joined online with the three independent directors:
- Ilya Sutskever. Cofounder and chief scientist; delivered the firing and ran the all-hands.
- Adam D'Angelo. Cofounder and CEO of Quora; the only one of the four to survive the settlement.
- Tasha McCauley. Entrepreneur and adjunct senior management scientist at RAND.
- Helen Toner. Australian-born researcher at Georgetown's Center for Security and Emerging Technology (CSET).
- Absent. Greg Brockman, removed from the board the same day; Altman himself was the sixth seat.
The directors repeatedly declined to say more, citing confidentiality obligations. VP of global affairs Anna Makanju, who often travelled with Altman, said the claim that he was untrustworthy matched none of their experience. The executives demanded the board resign and hand its seats to three employees, threatening mass resignation. Chief strategy officer Jason Kwon, a lawyer and former general counsel, argued it was illegal for the board not to resign, since a collapse would breach its fiduciary duties.
The board's answer was structural. It had consulted lawyers and acted within its remit. OpenAI's unusual structure, designed by Altman himself, charged the board with serving the mission (ensuring AGI benefits humanity), not shareholders, and Altman had long cited the board's power to fire him as the key governance safeguard. Toner added that if the action destroyed the company, that could still be consistent with the mission. Executives relayed this to staff in real time as proof she did not care about destroying OpenAI; some concluded it was her goal. At the party, at the thought of losing their equity, one person cried.
Saturday and Sunday: the pressure campaign
Fri 17, noon Altman fired on Google Meet; Murati named interim CEO; Brockman removed from board, quits that afternoon.
Fri 17, evening Botched all-hands; Pachocki, Sidor and Mądry resign; executives confront the board.
Sat 18 Dozens gather at Altman's $27 million mansion; a 5 p.m. ultimatum (reinstate and resign) passes; Kwon memo says the goal is bringing back all five departed leaders.
Sun 19 Altman negotiates at the office in a guest badge; second 5 p.m. deadline passes; near 9 p.m. board names Emmett Shear interim CEO; ~200 staff boycott his talk; late night, Nadella offers Altman and Brockman a Microsoft AI division.
Mon 20 Open letter passes 700 of ~770 staff; Sutskever signs and publicly regrets his role.
Tue 21 Deal reached: Altman returns without a board seat; new board.
Saturday, November 18. Employees and others gathered at Altman's $27 million mansion. The three departed researchers met Altman and Brockman to discuss re-forming the company, which some staff saw as a threat (a new rival) and others as an exit route. Leadership set a 5 p.m. deadline: reinstate Altman and resign, or face an exodus. The board refused and spent the weekend phoning anyone in their networks, sometimes at night, because Murati would no longer serve as interim CEO; they needed a stabilising replacement, or new directors strong enough to hold their own if Altman returned. After the deadline, Kwon told staff that "resolution" meant bringing back Sam, Greg, Jakub, Szymon and Aleksander. Altman tweeted in lowercase that he loved the team; dozens retweeted with hearts.
Sunday, November 19. Altman and Brockman returned to the office to negotiate; most participants had barely slept in 36 hours. Altman posted a selfie holding a guest badge, vowing it would be his first and last. A second 5 p.m. deadline was set. The pressure came from every side: Microsoft, other investors and Valley heavyweights sided publicly with Altman, and a source fed the press the playbook: mass resignations, Microsoft withholding compute, and investor lawsuits would make an Altman-less OpenAI untenable.
The board still held. Near 9 p.m. Sutskever posted on Slack that Altman was not coming back and that Emmett Shear, former Twitch CEO, was interim head; the board stood by its decision because Altman's lack of transparency had undermined its ability to supervise the company. Slack filled with dozens of furious, mocking replies. About 200 employees walked out to boycott Shear's talk, Murati hurried the executives out, and only a dozen or so people heard it. Anna Brockman, Greg's wife, whose civil ceremony Sutskever had officiated four years earlier, tearfully begged him to reconsider.
Microsoft's lifeline and the open letter
Late Sunday Nadella announced he was hiring Altman and Brockman to run a new AI division, and word spread that anyone who followed would get a Microsoft job. Hao marks this as the turn: with a backup option, fear became defiance. At one employee's house, with well over a hundred colleagues present, executives and senior researchers drafted an open letter restating the threat: reinstate Altman and resign, or staff would quit for Microsoft. It was pushed through private channels and by phone; as signatures hit critical mass, holdouts rushed to sign to avoid questions about their absence. Within 24 hours it had 700+ of roughly 770 employees. Staff sent identical emails to the board and posted the same slogan on X: "OpenAI is nothing without its people." In the night, Sutskever's own name appeared on the letter, and early Monday he tweeted that he deeply regretted his part and would work to reunite the company.
Tuesday, November 21: the settlement
Leadership called the board from Altman's house. Five sleepless days and approaching Thanksgiving opened a path: Altman and Brockman dropped their demand for board seats; the board, seeing no way to save the company without him, accepted his return. Late that night they settled the independent seats.
- Stays. Adam D'Angelo.
- Leave. Helen Toner and Tasha McCauley.
- Join. Bret Taylor (ex-co-CEO of Salesforce, ex-CTO of Facebook) and Larry Summers (former Treasury Secretary and Harvard president), with more members to be added later.
- Condition. Altman would submit to an investigation.
Both sides agreed the priority was to project unity. What followed was choreographed: two days later Altman tweeted a Thanksgiving message about a nice few hours with D'Angelo; ten days later Brockman posted a hug photo with Sutskever; the company's artist in residence wired DALL-E to a printer to make heart stickers beside a giant pink heart bearing the slogan. In December Altman told Trevor Noah it was the second-worst moment of his life, after his father's death. In January 2024 the tender closed at an $86 billion valuation. On the Tuesday night itself, staff flooded back to the office; a smoke machine set off the fire alarm and the party carried on, and Brockman tweeted a crowd selfie captioned "we are so back."
What the episode means
Hao heard the news mid-interview for this book, phone silenced, and found a pile of notifications twenty minutes later. People kept asking her whether it was just entertaining drama. Her answer: she had followed OpenAI for five years, and in 2019 was the first journalist given extensive access and the author of its first profile. To her the crisis surfaced one of the most urgent questions of the era: how do we govern AI? In little over a decade AI has rebuilt the internet's backbone as a ubiquitous mediator of digital life and is now rewiring health care, education, law, finance, journalism and government, so its governance is really a question of whether the future gets better or worse.
The founding promise
OpenAI presented itself as the answer to that question.
- Founders and money. Started by a group including Elon Musk and Sam Altman, with billionaire backers such as Peter Thiel; set up as a nonprofit with a $1 billion pledge.
- Purpose. Build AGI, described as the most powerful AI ever, for humanity rather than shareholders.
- No products. Pure research, aimed at an AGI that would bring utopia rather than its opposite.
- Openness. Share research widely and collaborate with other institutions; openness and democratic participation were the point of the name.
- Self-sacrifice clause. A few years later leadership wrote that it feared late-stage AGI work becoming a race without time for safety, and pledged that if a value-aligned rival neared beneficial AGI first, OpenAI would stop competing and start assisting it.
The erosion, step by step
- About 18 months in, executives saw that their chosen path would need extraordinary sums of money.
- Musk and Altman, until then hands-off co-chairs, each tried to become CEO; Altman won and Musk left in early 2018, taking his money. Hao reads the rift, in hindsight, as the first sign the project was about ego rather than altruism.
- To fill the funding hole Altman nested a for-profit arm, OpenAI LP, inside the nonprofit to raise capital, sell products and return money to investors; four months later, in July 2019, Microsoft became the new $1 billion funder.
- Hao arrived in August 2019 for three days embedded and dozens of interviews. She found a lab grown competitive, secretive and insular, even fearful of outsiders, with transparency, democracy, self-sacrifice and collaboration gone, replaced by one obsession: reach AGI first, in its own image.
- Over the next four years it became what it said it would not be: a nonprofit in name only, commercialising ChatGPT and chasing record valuations; more secretive, and shifting industry-wide norms to keep much of AI development from public scrutiny; and the trigger of the race to the bottom it had warned against, accelerating deployment without fixing harmful flaws. Internal factions fought ever harder to remake the company around their vision.
The board's defeat was final proof the governance experiment failed, and the nonprofit's surrender to money dissolved the last of its altruistic facade. Hao's sharper point is that a board victory would have proven the same thing: a decision shaping everyone's future was made behind closed doors by a handful of ultrarich techno-optimists, their ideological rivals and one tech giant, with even OpenAI's own employees in the dark.
Two lessons from covering AI
Hao reported on AI long before ChatGPT made OpenAI synonymous with it: she watched ideas tested at conferences and productised at Google, Facebook, Alibaba and Baidu, read hundreds of papers, and followed AI's effects on communities across five continents. In Colombia and Kenya she met people who turned to data annotation during economic crises and found conditions resembling indentured servitude; in Arizona and Chile she met politicians and activists fighting water-guzzling data centres.
- Lesson one: AI is plural and nothing about it was inevitable. AI is a multitude of technologies that shift with technical merit but also with their makers' ideology, hype and commercial pressure. Large language models are only one form, embodying a remarkably narrow worldview of how the world is and should be; their dominance was the sum of thousands of subjective choices by those with power in the room. Future AI is likewise open, which brings back the governance question: who gets to shape it?
- Lesson two: the current trajectory is alarming. Generative AI looks thrilling (brainstorming, writing help, late-night companionship, perhaps one day an output boost), much as Facebook once looked like holiday photos and social movements. Underneath, the models are monstrosities built on unprecedented data, labour, compute and natural resources. By one measure, GPT-4 is reportedly more than 15,000 times larger than GPT-1, released five years earlier. The costs fall hardest on the vulnerable, from Global North workers and rural residents to poor communities in the Global South; little trickles down, and the benefits accrue upward.
The empire metaphor
Hao says it is the only metaphor she has found that fits. European colonial empires seized resources that were not theirs, exploited subjugated labour to mine, farm and refine them, spread racist ideas of their own superiority and modernity to justify (and even entice acceptance of) conquest, and excused everything by the need to beat rival empires, all to entrench and expand their power at everyone else's expense.
| Colonial empires | Empires of AI (Hao) | |
|---|---|---|
| Resources seized | Land, minerals, crops | Artists' and writers' work, personal data posted online, land, energy, water for data centres |
| Labour exploited | Subjugated people mining, farming, refining | Global workers cleaning, tabulating and labelling data |
| Justifying story | Racial superiority, "modernity" | Progress, AGI, abundance |
| Arms-race excuse | Rival empires | Rival labs, China |
| Harms covered | Invasion of sovereignty, theft | Invasion of privacy, theft of work, automation of meaningful jobs |
| Violence | Overt and brutal | Not comparable, Hao concedes; harm via precarity and automation |
The scale race and its collateral
OpenAI's push on the limits of scale set the rules for a new era: every giant now races to out-scale the others, spending so much that even they must reshuffle and consolidate.
- Microsoft laid off 10,000 workers around the time of its $10 billion investment in OpenAI.
- Google, outpaced, merged its AI labs into Google DeepMind.
- Baidu made staff working on AI for drug discovery suspend research and hand their chips to its ChatGPT rival.
- Alternatives are being choked off: the number of independent researchers without industry ties or funding has rapidly dwindled, narrowing ideas not tied to short-term profit; firms that once funded sprawling exploratory research can no longer afford it under the generative-AI bill; and young scientists conform to stay employable.
- $157 billion: OpenAI's valuation when Hao finished writing; Anthropic was nearing a deal at $60 billion.
- $3 trillion+: Microsoft's market cap, tripled since partnering with OpenAI; the six largest tech firms added $8 trillion since ChatGPT.
- $1 trillion: projected generative-AI development spending within a few years, with so far little to show for it (Goldman Sachs, June 2024).
- 96% vs 77%: C-suite leaders expecting productivity gains versus employees using the tools who said they added to their workload, from reviewing AI output and higher demands from bosses (Upwork Research Institute, 2,500 workers, July 2024).
- November 2024: Bloomberg's Parmy Olson and Carolyn Silverman concluded the tally raised the prospect that the technology may never transform the economy broadly, only concentrate wealth at the top.
Meanwhile the costs mount: Kenyan workers were paid starvation wages to filter violence and hate speech out of OpenAI's models, including ChatGPT; artists are being replaced by models trained on their work without consent or pay; and journalism is shrinking as generative AI multiplies misinformation. Hao calls it an ancient story repeating, and only the beginning.
AGI as rhetoric
OpenAI is not slowing down, and the industry follows. To answer doubts about present performance, Altman talks up future AGI ever louder: his September 2024 "Intelligence Age" post promised massive prosperity and superintelligence perhaps within "a few thousand days," with fixing the climate, space colonies and the discovery of all of physics eventually commonplace. Hao's verdict is that AGI is now largely rhetorical, an all-purpose excuse for more wealth and power. Since few others have the capital to fund alternatives, OpenAI and a handful of rivals will hold an oligopoly, and any company or government wanting that future must depend on the empires.
A different way forward
Hao rejects the logic of unprecedented scale and consumption as the price of progress. Much of what society needs (better health care and education, clean air and water, a faster exit from fossil fuels) can be advanced with, and sometimes requires, much smaller models and a diversity of approaches. AI alone will not suffice; it also takes social cohesion and global cooperation, which the current vision undermines. The empires will not cede power willingly, but just as old empires fell to more inclusive governance, she argues the moment to reclaim control is still open.
- Policymakers: strong data-privacy and transparency rules, and updated intellectual-property protection so people regain agency over their data and work.
- Human-rights organisations: international labour norms and laws guaranteeing data labellers wage minimums and humane conditions, plus wider labour rights and dignified work across all sectors.
- Funding agencies: renewed diversity in AI research to build fundamentally different forms of the technology.
- Everyone: resist the industry narratives that hide AI's social and environmental costs behind an elusive vision of progress.
One side declined to talk. OpenAI and Altman did not cooperate, so their account of the firing comes through others. The board's reasons did surface later: Toner said in May 2024 that Altman withheld information (including that he controlled the OpenAI Startup Fund, and that the board learned of ChatGPT's launch on Twitter), while the WilmerHale review (March 2024) found a breakdown of trust but concluded his conduct did not mandate removal. The empire metaphor is a lens, not a finding; Hao concedes there is no comparable violence, and critics argue it understates genuine benefits users get from the tools. The "15,000 times larger" figure is hedged ("by one measure", "reportedly") because OpenAI never disclosed GPT-4's size.
- Restructuring. On October 28, 2025 OpenAI completed its recapitalisation: the nonprofit, renamed the OpenAI Foundation, controls OpenAI Group PBC and holds about 26%; Microsoft holds about 27% (roughly $135 billion), with IP rights extended to 2032 and any AGI claim to be verified by an expert panel. The California and Delaware attorneys general let it proceed with conditions.
- Valuations. OpenAI closed a $122 billion round at $852 billion (March 31, 2026; Amazon $50 billion, Nvidia and SoftBank $30 billion each). Anthropic raised $65 billion at $965 billion in late May 2026, passing OpenAI as the most valuable startup.
- Musk's lawsuit. On May 18, 2026 a jury in Oakland unanimously rejected Musk's claims that OpenAI betrayed its charitable mission, finding he sued too late; Musk, who had sought damages of up to about $134 billion, called it a technicality and moved to appeal.
Divine Right
Chapter 1 traces how Elon Musk's fear of Google's DeepMind met Sam Altman's genius for accumulating power, producing OpenAI in 2015. Hao builds Altman's portrait from a St. Louis childhood to the presidency of Y Combinator: a gifted storyteller and networker shaped by Paul Graham and Peter Thiel, whose rise was shadowed from the start by the same two charges, self-dealing and dishonesty, that resurfaced at his 2023 ouster. Altman's defining talent, in Graham's words, is that he is "extremely good at becoming powerful," and Hao argues the future of AI now rests on such fallible individuals.
The Rosewood dinner, summer 2015
The chapter opens on the dinner that seeded OpenAI. Altman, then president of the accelerator Y Combinator (YC) and already known as a brilliant strategist and dealmaker after founding his first company at 19, invited a group of men to discuss AI and humanity's future. Musk, 14 years older, had formed a good impression of him: smart, driven and, above all, apparently sharing Musk's views on carefully developing and governing AI. Years later, in his lawsuit, Musk would describe it as if Altman had mirrored everything Musk had ever said about AI to win his trust. Altman, for his part, called Musk a childhood hero; after a tour of the SpaceX factory in Hawthorne, he wrote that Musk's certainty about Mars rockets had become his benchmark for conviction.
Musk's obsession with DeepMind
Musk's alarm had a specific target: Demis Hassabis, the professorial CEO of London's DeepMind Technologies, and later Google.
2012 Musk meets Hassabis; on a visit to the SpaceX canteen Hassabis warns that superhuman AI could threaten humanity and, amused, notes it would follow people to Mars. Musk, not amused, invests $5 million in DeepMind to keep watch.
2013 At his birthday party in Napa Valley, Musk argues emotionally with Google cofounder Larry Page, who calls superhuman AI the next stage of evolution and calls Musk a "specist" for privileging humans.
Late 2013-Jan 2014 Google moves to buy DeepMind; Musk spends an hour on Skype from a closet at an LA house party urging Hassabis to back out. Google has already jetted researchers to London to vet it; Jeff Dean personally reviews code and approves. Deal confirmed January 2014, reportedly $400-650 million.
Oct 2014 At an MIT symposium Musk calls AI probably the biggest existential threat and building it summoning the demon; he also meets New York publishers about writing his own book on extinction risks.
Early 2015 Musk hosts dinners on countering Google; briefs President Obama on AI dangers, safety and regulation; attends the first meeting of the Google DeepMind AI Ethics Board at SpaceX and concludes it is a fraud.
Summer 2016 Early OpenAI staff meet Hassabis and report back that DeepMind does intend to take over the world.
2017 At an OpenAI off-site at SpaceX, Musk rants that Hassabis once made a video game about an evil genius building AI to rule the world, and that Page is too busy windsurfing to notice Demis gathering power.
- The profit-maximiser fear. Publicly Musk warned that an AGI told by Google to maximise profits might eliminate competitors at any cost, joking to The New Yorker that murdering rival AI researchers would be a character flaw.
- Climate versus AI. At a Stanford AI Salon, young researcher Timnit Gebru asked why AI worried him more than climate change; he replied that climate change would not kill everyone, while AI could make humanity extinct.
- Ethics board. The board had been proposed by Page and Hassabis to oversee responsible development; its first meeting turned Musk's worry into an all-consuming obsession.
- Supervillain casting. For years Musk portrayed Hassabis as a supervillain to be stopped and OpenAI as the good to DeepMind's evil. Before DeepMind, Hassabis had run his own games studio for seven years; the game Musk cited was Evil Genius (2004).
Hao's judgment, through DeepMind insiders, is that this was paranoia. Hassabis was intensely ambitious but kind and measured, and DeepMind staff found Musk's view entertaining; one former researcher called OpenAI's creation a semi-hysterical reaction to a mild-mannered man.
Bostrom and the birth of "alignment"
Musk publicly recommended Nick Bostrom's Superintelligence: Paths, Dangers, Strategies, calling it worth reading. The Oxford philosopher's argument:
- Control problem. An AI smarter than humans would be hard to control and could cause existential catastrophe.
- Paper-clip maximiser. Even a trivial goal, like making paper clips, could lead it to treat humans as threats, since they use up paper-clip resources.
- Proposed fix. "Align" the AI with human values, so it can extrapolate beyond explicit instructions without harming people.
- Legacy. This became the basis of the AI alignment research discipline that OpenAI would champion.
Hao adds a pointed coda: in January 2023 a mid-1990s Bostrom post to an email list resurfaced in which he endorsed a racist claim that Black people are less intelligent than white people. He apologised, called it disgusting and said it misrepresented his views; Hao notes it made people question the values of the man who framed "human values" for the field.
Altman the fellow catastrophist
To Musk, Altman looked like a fellow traveller with his own taste for hedging against catastrophe.
- 2016: Altman told The New Yorker's Tad Friend he would flee a doomsday to New Zealand with Peter Thiel; Thiel called him an optimist yet a survivalist.
- 2018: He told Bloomberg he had been joking, but kept a go bag for engineered viruses: gas masks, antibiotics, water, batteries, a tent and a gun.
- February 2015 blog: he agreed with Musk that superintelligence was probably the greatest threat to humanity's continued existence; an engineered virus was likelier but would not kill everyone. He called Bostrom's book the best thing on the topic.
- May 2015. Altman emails Musk: stopping humanity from building AI is almost certainly impossible, so it would be good for someone other than Google to get there first. He proposes a YC-run "Manhattan Project for AI," owned by the world through a nonprofit, which would comply with and aggressively support regulation (a nod to Musk's lobbying). Musk: probably worth a conversation.
- June 2015. Altman details the mission: build the first general AI and use it for individual empowerment, the distributed future he called safest, with safety a first-class requirement. Governance would defer to the two of them: Altman and Musk on the board, inviting three others; the foundation would own the technology, and where its use for good was unclear, the five would decide. Musk's monthly team meetings would help recruit the best people; failing that, his public endorsement would. Musk replies that he agrees on all.
- Summer 2015. With Musk confirmed, the dinner moves to the Rosewood Hotel on Sand Hill Road ($1,000 a night, 16 acres among dozens of VC firms). Musk arrives over an hour late to find Altman, Greg Brockman, Dario Amodei and Ilya Sutskever waiting.
- Outcome. The group became the nonprofit's key leaders; Musk named it OpenAI. Nearly all of them would later leave after clashing with Altman.
Hao then flashes forward. After the two men stopped speaking and Altman replaced Musk as the Valley's "it guy," Altman rewrote his public position, telling Business Insider in 2023 he was now firmly in the AI-as-a-tool camp, while expecting society to become extremely different. Musk came to feel Altman had used him to rise. Hao ties this to the line that has followed Altman all his life: Paul Graham's quip that if you dropped him on an island of cannibals, he would be king within five years, and Graham's later remark that Altman is extremely good at becoming powerful.
The making of Sam Altman
- Sam Altman. Born Samuel Harris Gibstine Altman, April 22, 1985, in Chicago, eldest of four, to Jewish parents.
- Connie Gibstine. His mother: a dermatologist holding both medical and law degrees, unusual for her generation; chose dermatology for steady pay and hours that let her be home for her children; an atheist who is culturally Jewish; kept her maiden name. Her father, Marvin Gibstine, was a US Army paediatrician posted to postwar Germany.
- Jerry (Jerold) Altman. His father: son of a shoe manufacturer, Wharton graduate and Boston consultant, previously married; met Connie at Loyola University Chicago law school; in St. Louis became chief counsel and VP of developer the Roberts Companies, passionate about affordable housing and civic revitalisation; more religious, insisting on bar and bat mitzvahs. Sam would quote his lesson: always help people, even without time.
- Max and Jack. Born in quick succession after Sam; later his business partners.
- Annie. Born nine years after Sam; once his closest sibling, later his accuser.
Hao reads the parents' traits into the children: Connie's rationality and discipline, Jerry's spirituality and service.
Childhood: competitive, gifted, anxious
- Precocity. Operated the family VCR at two and fixed it at three; given a Mac at eight, he learned to program and dismantle it. He bossed and cared for his siblings and hated losing board games.
- Apple versus Applebee's. When their grandmother gave each grandchild stock, Sam chose Apple and Jack Applebee's; over twenty years Sam's grew hundreds of times, a family joke Sam still savours.
- John Burroughs School. Connie let each child choose this rigorous private school; Sam went and thrived: yearbook head, water-polo captain, Model UN, drawn to writing as much as STEM. His English teacher Andy Abbott (later head of the school) hoped he would become an author rather than go into tech.
- Pushing limits. He got in trouble for leading his water-polo team in a striptease to their Speedos at a pep rally.
- Coming out. He came out as gay to family and school; his mother was surprised but accepting. When Christian students boycotted an assembly on sexuality he led on National Coming Out Day, Altman, 17, answered them in a speech arguing tolerance cannot be selective; his college counsellor credited it with opening the school's culture.
- The sensitive side. He worried about others' opinions and suffered anxiety into adulthood: calling his mother convinced a headache was meningitis or lymphoma, and once, panicking mid-negotiation, lying bare-chested on the floor to calm down.
Hao takes the frame from Tad Friend's 2016 profile: Altman is as driven to push ahead as he is alert to the need for caution, which maps onto OpenAI's own split: reach AGI as fast as possible, and do not destroy humanity.
Stanford and Loopt
Altman entered Stanford in 2003, drawn by its closeness to the industry. He briefly considered being a writer or an investment banker, then chose computing, reasoning that the world did not need another novel and that writing was a harder way to make money. He majored in computer science, interested in AI and security; a classmate recalls him tearing a homework tool down to low-level code and finding a bug in the assignment itself. As a sophomore, learning phones would soon carry GPS, he stood on stage at a campus event with a flip phone and recruited collaborators for a location app.
He then met Paul Graham, who was starting YC with his girlfriend Jessica Livingston. Altman joined YC's first batch in 2005 with Loopt, a social network that used location tracking to alert users when friends were near or to recommend restaurants, and spent the summer in Cambridge, Massachusetts, eating so much instant ramen he gave himself scurvy. He later told founders that early hard work compounds like interest. By late 2005 Loopt was in talks with New Enterprise Associates and Sequoia for $5 million, and he dropped out.
The storyteller
Loopt ran seven years and sold in 2012 for $43.4 million, roughly what investors had put in. Hao argues his early interviews reveal the seeds of his success precisely because the product (an early Foursquare rival that failed) was so much less alluring than AI.
- The pitch pattern. Casual ease; a company uniquely positioned on technology's unstoppable trajectory; customers and advertisers clamouring; success inevitable. Even knowing the outcome, Hao finds it compelling.
- Location as the new normal. In June 2010 he told blogger Robert Scoble that the value of sharing location now outweighed privacy fears and that not sharing would soon seem weird, promoting Loopt Star, which pushed location-based coupons and deals. Months later he told CNN Business that online and offline life were fusing.
- Turning risks into the pitch. In 2008, when Jessica Lessin (then at the WSJ, later founder of The Information) planned a story on location-privacy concerns, he sent her Loopt's own list of risks and proposed fixes; the implicit message was that this is how the world will work. Lessin concluded that, beyond building a startup, he wanted to write the rules.
Networks and dealmaking
- Proximity. Loopt's office was down the hall from a fledgling YouTube. YC batchmates included Reddit's cofounder Steve Huffman and founding engineer Chris Slowe; Altman joined Reddit's board in 2014 and eventually held more of the company than Huffman. Another batchmate, Emmett Shear of Twitch, would be OpenAI's interim CEO during the 2023 ouster.
- Deals as a nobody. As CEO of an obscure startup he won partnerships with the major US mobile carriers.
- The formula. People credit listening, willingness to help, and framing any offer as exactly what you want; he has been called the Michael Jordan of listening. Geoff Ralston, who later ran YC, calls him the Usain Bolt of fundraising: fundraising is telling a story about a project's extraordinary future, and Altman tells one you want to join and that seems likely. Ralston likens it to Steve Jobs's reality distortion field, adding that both men built things that really did change reality.
- The flip side. A longtime colleague says Altman's attentiveness and memory for detail are used to work out how to influence you; you feel you are making progress, then realise you have been running in place.
Per The Wall Street Journal, senior Loopt leaders twice urged the board to fire Altman, alleging two things: that he acted for his own gain, sometimes at the company's expense, and that he had a compulsion to distort the truth, even about details so trivial the lie seemed pointless. These "paper cuts" bred pervasive distrust and chaos. The board sided with him, a pattern of emerging on top that Hao says defined his career. These are the same two accusations that surfaced at OpenAI in 2023.
Wealth and the brothers
Despite Loopt's middling record, Altman used it to climb into the Valley's most powerful networks and engineer an exit that made him rich: $5 million at 26, which he considered disappointing next to Jobs's $256 million at 25. Far more money followed, and with it a changed life: no more grocery shopping, private jets, McLarens and a $5 million Koenigsegg he races, Burning Man for a time, and casual ketamine use, common among Valley figures.
He brought his brothers along. In 2012 he started Hydrazine Capital with Jack (Princeton economics, briefly investment banking), who then founded Lattice, funded by YC once Sam ran it. Max (Duke computer science, briefly Microsoft, then a trader) joined YC company Zenefits in 2014 and Hydrazine two years later. Both moved in with Sam temporarily and the three ended up living together for years, a knot of family and business ties.
Two mentors: Graham and Thiel
Hao singles out two relationships as decisive for the shape of Altman's career.
| Paul Graham ("PG") | Peter Thiel | |
|---|---|---|
| Base | Cofounded Viaweb (sold to Yahoo for $49m, 1998); influential essayist; cofounded YC with Jessica Livingston | Billionaire founder of PayPal and Palantir, early Facebook investor; Founders Fund |
| Bet on Altman | 2006: called Loopt YC's most promising startup and Altman what Bill Gates must have been like at 19; at 23 ranked him with Jobs; made him YC president in 2014 at 28 | After Loopt, put up most of Hydrazine's $21 million; Altman scouted high-return deals for Founders Fund, multiplying Thiel's wealth |
| Core lesson | Growth fixes all problems | Competition is for losers: aim for monopoly |
| Controversy | Preached meritocracy while running YC as an insular fraternity; blamed few female founders on women's upbringing; listed "a strong foreign accent" among 30-40 failure predictors | Rare vocal Trump backer in 2016; secretly funded the lawsuit that destroyed Gawker Media, which had outed him nearly a decade earlier |
Graham and the YC machine
YC became the most elite club in the Valley: by the time Loopt sold it had seeded future billion-dollar firms such as Dropbox and Airbnb, and membership brought cachet, a built-in customer base of alumni companies, eager investors and higher valuations. Graham's precepts became the founder culture; Ralston says many people lived by them.
- The hacking question. Graham asked Altman what YC's application should ask to find more people like him. Altman proposed: describe the time you most successfully hacked some (non-computer) system to your advantage. Graham called it one of the most important questions; Hao argues it encoded an ethos of bending, bypassing and breaking rules to dominate.
- What would Sama do? Graham wrote that he consulted Jobs on design and Altman on strategy and ambition.
- The succession. In Graham's kitchen, offered the YC job, Altman smiled uncontrollably; YC lore holds it was the only unplanned smile anyone saw from him. Altman said YC somewhat directs the course of technology; Graham said Altman's goal was to make the whole future. Livingston recalled there was no shortlist: just Sam.
Thiel and the Hydrazine years
After the sale Altman broke up with a cofounder who had been his boyfriend of nine years, took a year off, and raised Hydrazine's $21 million, mostly from Thiel, nearly twenty years his senior. As a YC partner he used Hydrazine to bet on YC companies while helping Founders Fund find winners. Their closeness was compared only to Thiel's mentorship of Mark Zuckerberg.
The growth ideology
Both mentors taught him the imperative of scale and the superiority of capitalist efficiency over government. Hao traces how it hardened into his core thesis:
- "Growth and Government" (2013). Thanking Graham and Thiel, he restated startup wisdom that rising sales fix all problems: for startups, more sales mean more capital, talent and harmony; for countries, more growth means more innovation and better lives. You are either growing or slowly dying, and US government dysfunction was killing growth; without it, democracy fails because voters face a zero-sum game.
- 2017. He urged the private sector to restore growth, praising America's 200 years of growth, 100 years of territorial expansion and 100 years of working technology. Hao notes this glossed over colonial bloodshed and industrialisation's labour and environmental record.
- 2019. He called sustainable growth almost always a moral good and said it motivated both his YC and OpenAI work.
Altman taught How to Start a Startup at Stanford and invited Thiel to lecture. Monopolies are good because they are stabler, longer-lived, better capitalised and a sign of real value created. They rest on four ingredients, each of which must endure:
- Proprietary technology, kept in the lead so no one supersedes you.
- Network effects.
- Economies of scale.
- Branding.
The cautionary case was 1980s disk-drive makers: advances every two years, each by a different firm, great for consumers but not founders. So a company needs a huge initial breakthrough and also the last breakthrough, improving fast enough that no one catches up. Thiel's conclusion: an innovative field full of rivals is good for society but bad for your business.
Building personal network effects
From both men Altman also learned to build network effects as an individual. A 2013 blog post gave his theory of how things get done: focus plus personal connections (Charlie Rose told Graham, who told him), to which he later added self-belief, since founders must believe they might succeed.
- Thiel's model. Thiel used advice and small cheques to pull young people into his orbit, then used the network to gather more money and connections. Altman copied it, first giving time and tactical advice, then money, scaling it with relentless efficiency.
- Techniques. Interlocking dinners at his house (YC people, investors, the growing gay entrepreneur community); advice by text or two-minute calls; one-word "meet" or a lone "?" to introduce people, a Jeff Bezos habit; mostly many small bets rather than a few big ones. By a June 2024 WSJ count he had financial ties to 400+ companies through YC, Hydrazine and other funds.
- The inner circle is financially tied. His most successful early investment was Stripe, where Greg Brockman was first CTO; he invested early in Airbnb, run by close confidant Brian Chesky; he backed ex-boyfriend Matt Krisiloff's biotech Conception and co-invests with ex-boyfriend Lachy Groom.
- Generosity, as friends see it. Lending his houses, funding a stranger in Ethiopia who emailed for a laptop, and during the 2023 Silicon Valley Bank run (the biggest US bank failure since 2008) wiring money without paperwork to keep startups from closing or laying off staff.
Politics and a remade persona
- Party lines. Where Thiel poured tens of millions into Republicans, Altman hosted fundraisers and wrote cheques for Democrats, which strained the friendship for a while.
- 2017 tour. Like Thiel's other protégé Zuckerberg, he toured America and talked to 100 Trump supporters.
- The governor idea. He weighed running for California governor, which would put him over the world's fifth-largest economy, and published a manifesto, "The United Slate", with three principles: (1) prosperity from technology; (2) economic fairness; (3) personal liberty. Focus groups found him too young; friends joked he should aim for president.
- Before. Early at YC: boyish, one suit jacket, perched on chairs, flippant, profane, open about his gun collection, quick to anger at ineffective people, even coding a tool to rate founders by email response time.
- After. Fitted Henleys and jeans, 18 pounds of muscle in a year, fewer words, more questions, a furrowed-brow modesty; lowercase texts full of emoticons; his personal number given to employees. Publicly he became the nice guy who avoided negative emotion, confrontation and saying no; privately, flashes of anger remained. When OpenAI fired one employee he personally offered ketamine and booze as consolation; that person says all of Sam's relationships end well, whether you want them to or not.
Becoming an institution
Hao's conclusion on the YC years: Altman converted YC's power and network into his own, and his connections and reputation became his main currency. Policymakers treated him as a gateway to Silicon Valley: in 2016 Defense Secretary Ashton Carter sought his advice on recruiting tech talent; on the day he left YC in March 2019, Senate minority leader Chuck Schumer made a quiet visit to OpenAI with his Secret Service detail, sat beside him before a TV showing a fireplace, and told staff he did not fully understand their work but they were in good hands with Sam.
The cost was a growing list of detractors and enemies repeating the Loopt charges of self-serving power-seeking and compulsive dishonesty. Beneficiaries became loyalists; others saw him as devilishly able to bend situations. To YC partners and other power brokers this was an annoyance; to employees with little leverage, a source of fear; to those opposed to his worldview, a massive threat.
Annie
Hao ends with Altman's sister, the youngest of the four.
- Closeness. As children and into his twenties, Sam was her protector; she was the artsy, emotionally expressive one, and he confided in her about worries and partners.
- Hardening. She watched him wall off his sensitive side as he rose; he shared tactics like using fewer words in emails to seem more powerful. Sadness turned to fear, then, she says, to being harmed by him.
- Money. An early boyfriend's rule that every big purchase be matched by an equal donation briefly checked his spending; as money outran spending, she felt he began to hoard it and lose touch with people in need.
- The crisis (late 2019 to mid-2020). Per extensive correspondence she shared, he and the family repeatedly declined or hesitated to provide what she saw as emergency support for rent and medical costs, while her records show acute physical and mental-health problems worsened by their father's sudden death. She fell into unstable housing and turned to sex work. In summer 2020, as OpenAI drew its first big wave of attention, she cut off her family.
- Allegations. In 2021 she went public, alleging he sexually abused her as a child and that the family abandoned her; on January 6, 2025, two days before her 31st birthday, she sued to meet Missouri's statute of limitations.
- The family's side. Hao allows the brothers may have been following their mother's lead in pushing Annie toward financial independence, and calls the story hard to judge on partial information. In January 2025 Sam, his mother and brothers publicly expressed love for Annie and denied everything as "utterly untrue," calling the abuse claim the worst of them; Connie Gibstine gave Hao a short similar statement; Sam (via OpenAI) and the brothers did not respond.
Hao's reason for including it: Annie's story mirrors the book's themes of the gulf between those who gain and those left behind by progress, the loss of voice among the disenfranchised, and the danger of ceding power to companies and to individuals without checks and balances. Her persistence also shaped how Altman and OpenAI's executives handled the company's rise.
Hao's closing argument: Altman's ascent, character, relationships, the divisions he left and the flows of money and power all lead to his brief 2023 ouster, which gave the world a glimpse of the struggle at the top. It showed that the contest to control a technology already restructuring society and terraforming the earth rests on the polarised values, clashing egos and messy humanity of a few fallible people.
Sourcing asymmetry. Altman's critics here are largely anonymous, while he declined to be interviewed. The Loopt allegations rest on WSJ reporting. On Musk, Hao largely accepts his alarmed framing of Hassabis as an origin story; Musk's own motives (control, later rivalry via xAI) get fuller treatment in later chapters. Note too that the founding emails show Altman proposing a board controlled by himself and Musk, a detail that sits awkwardly with the "owned by the world" rhetoric. The Annie Altman material involves disputed allegations that the family denies, and Hao relies in part on documents Annie supplied.
On March 20, 2026, U.S. District Judge Zachary Bluestone dismissed Annie Altman's original claims as time-barred but let her refile under Missouri's childhood sexual abuse statute; she filed an amended complaint in early April 2026. The same order let Sam Altman's defamation countersuit proceed; his filings call her suit extortion and deny the allegations.
A Civilizing Mission
Chapter 2 follows OpenAI from its December 2015 launch as a transparent, nonprofit "anti-Google" to its 2019 capped-profit arm and $1 billion Microsoft deal. The hinge is one strategic belief formed in 2017: that AGI requires compute growing far faster than Moore's Law, and therefore billions of dollars a year that no nonprofit could raise. Hao shows that openness was treated from the start as a narrative to be dropped once it had served its purpose, and that the compute thesis drove both Musk's exit and Altman's takeover.
The founding duo
Greg Brockman was the first to commit, telling Altman so as Altman drove him back to San Francisco from the Rosewood dinner. Because Brockman was not a researcher, Altman tactfully told him he needed a partner who understood AI, and handpicked Ilya Sutskever, then a Google researcher whom he had cold-emailed to attend the dinner; Sutskever said yes once he learned Musk would be there. The two met for the first time that night. Hao presents them as a deliberate pairing: the operator and the scientist.
| Greg Brockman | Ilya Sutskever | |
|---|---|---|
| Type | Engineer and startup operator, like Altman; tall, amiable | Research scientist; lean, more aloof |
| Origins | Hobby farm in North Dakota; fell for maths, then science, between milking cows | Born in the Soviet Union, raised in Israel; a maths prodigy whose parents enrolled him in Open University of Israel courses in eighth grade |
| Education | Harvard 2008, transferred to MIT two years later, quit after one more semester to build products | Moved to Toronto at 16; one month of high school, then the University of Toronto in 2003 as a third-year undergraduate |
| Breakthrough | Joined Stripe when it had three other people; became CTO; spent five years prototyping early products as it grew into payments infrastructure for Amazon and Shopify | With Geoffrey Hinton (the only person he calls a mentor) and fellow student Alex Krizhevsky, won ImageNet 2012 at 15% error when rivals could not get below 25% |
| Payoff | Significant wealth and the Valley status of having built a multibillion-dollar company | Google bought their startup DNNresearch at auction for $44 million in early 2013, making them multimillionaires and starting the first big rush to commercialise AI |
What the Rosewood group agreed
The dinner mixed entrepreneurs and scientists, and the talk swung between research approaches and Musk's fixation: whether there was still time to beat DeepMind and Google, which the group saw as the way to correct AI's course.
- Talent was the bottleneck. Most top researchers worked, like Sutskever, for Google or another giant, on lavish salaries, benefits and job security.
- AGI as the banner. Most serious scientists then treated human-level AI as science fiction or decades away; claiming it was close enough to invest in now looked like pseudoscience and quackery. Demis Hassabis used the term for DeepMind even though his own staff found it distasteful, shameless marketing. The Rosewood group adopted the same goal precisely so it could compete head to head with him.
- Unequal conviction. Sutskever privately believed AGI was possible (he would later back some of the most aggressive timelines) but was squeamish about saying so, fearing for his scientific credibility. Brockman, an outsider to the field, sincerely believed focused effort could put AGI around the corner, and the researchers' willingness even to entertain a new lab strengthened that confidence.
Courtship and recruiting
Hao uses the Valley cliché that founding a company is like a marriage. A few weeks after Rosewood, Brockman and Sutskever dined alone in Mountain View; Brockman later blogged that the two had an "extremely high-bandwidth" connection and complementary ideas. While Sutskever spent months deciding, Brockman ran recruiting.
- Method. He asked leading figures for lists of top talent, dined with professors to ask about their best students, and researched each candidate before pitching them. Altman later praised this in a blog post titled simply "Greg", naming him the ideal cofounder.
- Softening sceptics. Musk and Altman worked on researchers who resisted AGI talk. Berkeley professor Pieter Abbeel recalls Musk asking: even if AGI is far off, what if there is a 1% or 0.1% chance it arrives in five to ten years? Abbeel joined as a research adviser, later full time with several of his PhD students.
- The Napa bus. Many recruits would sign only if others did. Brockman invited his ten most-wanted engineers and researchers to a wine outing in Napa Valley and hired a bus so he could keep pitching through the hour-plus ride each way. By his three-week deadline, nearly all accepted.
Openness as positioning
Musk, Altman and Brockman knew public perception would make or break the launch. They adopted Altman's proposal: make OpenAI a nonprofit and play up the openness in its name. It would be the anti-Google, doing research for everyone, open-sourcing its science, a model of transparency.
- Brockman, November 2015. In an email to Musk and Altman he hoped OpenAI would enter as a neutral group that collaborates widely and reframes the debate around humanity winning rather than any company, which he saw as the best way to bootstrap into a leading institution.
- Musk, later that month. Having the public root for them was valuable; the announcement should be rewritten for broad appeal and should cite a $1 billion commitment rather than $100 million, so OpenAI did not sound hopeless next to Google and Facebook. Musk promised to cover any shortfall.
Later correspondence shows the founders knew they could walk back openness once the narrative had served its purpose, for instance to keep the technology from bad actors. In January 2016, weeks after launch, Sutskever wrote that as they got closer to building AI it would make sense to be less open: the "Open" meant everyone benefits once AI is built, and it was "totally OK to not share the science." Musk replied: "Yup."
Launch, December 2015
- Timing. The blog post "Introducing OpenAI" went out on a Friday night during NIPS (Neural Information Processing Systems), the largest annual AI conference, where Hinton and Sutskever had auctioned DNNresearch three years earlier.
- Roles. It named nine founding members: Brockman as CTO, Sutskever as research director, Musk and Altman as co-chairs.
- Pledgers. The $1 billion pledge came from Musk, Altman, Brockman, Jessica Livingston, Peter Thiel and LinkedIn cofounder Reid Hoffman, who had worked with Musk and Thiel at PayPal and co-invested with them as the "PayPal mafia". The post said only a tiny fraction would be spent in the next few years.
- The Sutskever bidding war. Other founders got a $175,000 base plus YC or SpaceX stock; Sutskever was offered nearly $2 million, a huge sum for a nonprofit. Google countered repeatedly, reaching two to three times that. Musk and Altman delayed the announcement several times while he agonised, phoned his parents and fielded pleas from Musk and Brockman. In the end Google's offer itself convinced him why a nonprofit was needed.
- Musk's launch-day email. Recruiting the best people was the top priority; OpenAI was "outmanned and outgunned" by organisations they knew well but had right on its side, and he liked the odds. To block counteroffers, every founder's base salary rose another $100,000 at once.
- Aftermath. Musk later described facing Larry Page's fury for personally poaching Sutskever; the two rarely spoke again as their views on AI diverged. Moonshot glamour and billionaire backing made recruiting easier, and headcount doubled within a few months.
A "third way," and its critics
Hao's verdict on the launch is double-edged. The vague goals, the giant cheque and the billionaire worship were Silicon Valley at the height of its unchecked power; yet OpenAI was cleverly placed on the border between techno-chauvinism and a more conscientious strand of the industry that was just emerging.
- The 2016 shock. Donald Trump's election jolted tech's left-leaning workforce into self-examination. As techlash hit Meta and Google, AI researchers asked whether they had tied their field to corporate bottom lines too fast.
- The scorecard. Automated tools sold to police, mortgage brokers and lenders were entrenching racial, gender and class discrimination. Facebook's News Feed and YouTube recommendations had likely polarised the public, spread misinformation and extremism, enabled election interference and, in Facebook's case, helped precipitate ethnic cleansing in Myanmar.
- The state alternative. Government money carried its own risks. In 2018 thousands of Google employees protested Project Maven, a secret Pentagon contract for AI-powered surveillance drones that they feared paved the way to autonomous weapons; the Pentagon denied that aim, then moved away from that position with the Ukraine war.
- The bitter binary. Researchers joked that the choice was Big Tech, the military-industrial complex, or quitting AI. OpenAI looked like a third way, corrupted by neither profit nor state power; engineer and blogger Chip Huyen called it "a beacon of hope."
Who. Timnit Gebru, a Stanford PhD student, an Ethiopia-born Eritrean refugee who came to the US as a teenager (the researcher who had asked Musk why AI worried him more than climate change). What happened. At her first NIPS she was one of the only Black attendees; at the 2016 meeting she counted six other Black researchers among 8,500. She joked that she could found a "Black in AI" group and hold meetings alone. At a party, drunk men in Google Research T-shirts surrounded her; one forced a hug, another a kiss on the cheek while taking a humiliating photo; a friend was harassed by a professor at the same conference. Then came the launch: a group, nine of eleven of them white men, showered with money to build a good superintelligence against a hypothetical bad one. Response. That night she drafted an anonymous open letter attacking the spectacle, the cult of AI celebrities and the homogeneity of AI's builders, which she argued pushed out talent and narrowed whom AI served; its key line was that AI's harms were "already happening." On the flight home she posted a shorter, sanitised version under her own name on Facebook. Outcome. Weeks later she emailed five Black researchers she had met ("Hello from Timnit"), the start of Black in AI.
Brockman builds a culture
The co-chairs were barely present, busy with other ventures, so Sutskever ran research (collecting researchers' best ideas) and Brockman ran organisation and culture. Hao draws on his later interviews with her.
- Study of moonshots. He read every book he could find on ambitious US projects: the transcontinental railroad, Edison's light bulb, the early computer networks behind the Web, absorbing them like religious texts for design clues.
- The janitor parable. His favourite story was the likely apocryphal tale of JFK asking a NASA janitor what he was doing and hearing he was helping put a man on the moon. Every OpenAI employee, at every level, should share that sense of mission; Brockman lamented that Americans had "stopped daring to dream." In the first months, when staff worked from his apartment, he literally washed everyone's glassware.
- Office mandate. Everyone had to work from the San Francisco office, a rule kept until the pandemic, for cohesion and serendipitous exchange. Brockman admitted not everyone wanted Bay Area life; Hao's other interviews found this fell hardest on women and people of colour who, like Gebru, felt alienated by tech's white, male culture.
- One title for all. Every employee became a "member of technical staff", copied from Xerox PARC, which had borrowed it from Bell Labs to create a more democratic workplace.
- Answering sceptics. He likened AGI doubters to the distinguished experts who said Edison's bulb would never work a year before it shipped, borrowing Arthur C. Clarke's phrase from Profiles of the Future: "a failure of imagination."
Two definitions of "safety"
Dario Amodei, a computational neuroscientist turned AI researcher, had been at the Rosewood dinner while working at the Silicon Valley lab of Chinese firm Baidu, before a brief stint at Google. His sister Daniela Amodei had worked with Brockman at Stripe, and when Brockman first got serious about AI he had asked Dario for learning resources. Dario did not join at once, but OpenAI, under Musk's influence, seemed the lab most willing to focus on AI safety.
- "Concrete Problems" (2016). Written while at Google, Amodei's co-authored paper became foundational for the discipline. It defined the central safety problem as accidents: unintended, harmful behaviour that emerges from poor design of real-world systems, explicitly distinct from privacy, security, fairness and economic impact. Hao's gloss: safety in this frame meant preventing rogue, misaligned AI, the root of Bostrom's existential threat.
- The stakes for Amodei. Nothing mattered more to him than stopping superhuman AI from causing catastrophe, even human extinction.
- Effective altruism. Both siblings sympathised with EA, a controversial ideology born among Oxford philosophers (Bostrom's base) that preaches doing maximal good through extreme rationality and counterintuitive logic. Largely through Bostrom, it had made rogue AI a leading cause for adherents.
- Open Philanthropy. Two years earlier, Daniela's husband Holden Karnofsky had founded Open Phil to give partly on EA principles. It quickly became the main funder of catastrophic and existential AI-safety research: 300+ grants worth $440 million by November 2024.
The accountability critique
Existential safety, built on thought experiments, soon drew fire as researchers woke up to AI's immediate, real-world harms. Around the same time as Amodei's paper, ProPublica's "Machine Bias" (2016) showed that risk-scoring algorithms used across the US criminal justice system rated Black defendants as higher risk than white ones with longer records. That investigation, plus the post-2016 souring on social media, launched a wave of research on AI's societal harms.
| Existential safety (Amodei, 2016) | Accountability (Raji, 2020) | |
|---|---|---|
| Core worry | Misaligned, rogue AI; catastrophe or extinction | Harms already occurring, backed by evidence |
| "Side effects" example | A cleaning robot knocking over a vase or damaging walls while tidying (a cousin of the paper-clip thought experiment) | Mass privacy infringement to train facial recognition; spiralling data-centre environmental costs |
| Unit of analysis | The AI agent's behaviour, isolated from context | Design and deployment choices across the whole pipeline, and the hidden human cost of building systems |
| What is excluded | Privacy, fairness, security, economics treated as separate fields | Nothing: "safe" cannot be defined apart from those impacts |
| Evidence base | Philosophical scenarios | Audits such as "Machine Bias" |
Deborah Raji, a Berkeley AI accountability researcher, co-wrote a 2020 reply to Amodei's paper. Her argument: truly safe systems cannot be built by studying model behaviour in isolation from its effects on the very things (privacy, fairness, economics) Amodei had set aside; and the "negative side effects" he imagined were already happening, produced by the industry's pursuit of products and AGI. Side effects, she wrote, also stem from basic design choices in building and deploying models, beyond an agent's actions. Hao clearly sides with this critique. Inside OpenAI, researchers (some among its handful of women of colour) urged executives to count work on the discriminatory impacts of deep learning as safety research; one executive replied, "That's not our role."
Amodei joins, and the seeds of Anthropic
- May 2016. Still at Google, Amodei visited OpenAI's new office above a chocolate factory in the Mission District (San Francisco's oldest neighbourhood, a Latino stronghold), where researchers padded around in socks. Per The New Yorker, he told Altman and Brockman that 20 to 30 people, Bostrom and Wikipedia among them, believed OpenAI would build friendly AI and release its source code. Altman said they did not plan to, but asked not to correct the impression, since that usually made things worse. Asked the real goal, Brockman said it was to "do the best thing there is to do," conceding that was vague.
- July 2016. Amodei joined to lead AI safety research. Open Phil then gave OpenAI $30 million to secure a three-year board seat for Karnofsky.
- 2018. At Brockman's invitation, Daniela Amodei (Stripe's first recruiter) joined as an engineering manager and VP of people; Brockman told Hao he had wished for her from the start of recruiting.
- End of 2020. Disturbed by what they saw as Altman's and OpenAI's break from the founding premise, the siblings left to found Anthropic, taking key staff and creating the rivalry that would later drive ChatGPT's rushed release. Karnofsky stepped down (term served, plus a new conflict of interest) and among his nominated successors was a former employee, Helen Toner.
A lab with no strategy
A year in, OpenAI had "poached, begged, and borrowed" its way to a stellar team, and the density of talent kept spirits up. But it had no coherent strategy, and the shine was wearing off.
- Kitchen-sink research. Projects sprawled across robots, video games and simulated worlds for training agents. Little worked, and what did felt derivative. Nikhil Mishra, a 2017 intern, saw nothing super innovative in the big projects.
- Leadership at its limits. Brockman spent most days coding; Sutskever stalked the office asking each researcher "What's your next big thing?" There was no real management structure or priorities; people were sometimes fired at weekends, colleagues noticing only when they failed to show up on Monday.
- Cash burn. Of $11 million spent in 2016, more than $7 million went on compensation and benefits.
- DeepMind's triumph. In March 2016 AlphaGo beat Go champion Lee Sedol in a five-game match livestreamed from South Korea to 200 million+ viewers; Netflix released a documentary a year later. Just before the match Musk wrote to OpenAI's leaders that DeepMind was causing him "extreme mental stress", given its "one mind to rule the world" philosophy.
- Musk's deadlines. He dropped in demanding progress on unrealistic timelines, which staff thought absurd for unpredictable research. At an all-hands, robotics lead and founding member Wojciech Zaremba presented his plans; Musk asked only when. Zaremba said he did not know, and Musk replied that he therefore had no plan.
OpenAI's Law: the compute thesis
In March 2017 Brockman and Sutskever began a focused research road map around one question: what would it take to reach AGI, and to get there first? Sutskever's intuition was compute. ImageNet and every later advance he had been part of used materially more compute (alongside more data and better algorithms), but compute, he felt, was king: scale training to human-brain scale and something radical, AGI, would happen.
- Compute = chip speed × number of chips × running time.
- Chip speed is set by the chip industry, which has doubled single-chip power about every two years: Moore's Law, Intel cofounder Gordon Moore's 1960s prediction, revised a decade later. It became a self-fulfilling prophecy, the pace firms believed they had to hit to stay relevant.
- Brockman and Sutskever's sum: at Moore's-Law pace, reaching brain-scale compute would take far too long.
- Amodei and Danny Hernandez's chart: plotting the compute behind each major breakthrough since 2012 (starting with Sutskever's ImageNet work), they found it had doubled every 3.4 months for six years, a 30-million-percent rise.
- OpenAI's Law: Brockman's name for that curve. Leadership decided OpenAI must scale at at least that pace, treating it with the existential fervour chipmakers brought to Moore's Law.
- Implication: if chips would not get fast enough, OpenAI needed vastly more chips.
- GPUs. Graphics processing units were built to render graphics fast (glossy, low-latency video games), but they suit AI training because both need massively parallel number-crunching.
- Nvidia's grip. Most of the industry buys from Nvidia of Santa Clara, which makes the best GPUs and the companion software platform CUDA (Compute Unified Device Architecture), with a powerful hold on developers.
- Prices. A custom Nvidia server with eight top GPUs cost $150,000 in 2017, rising with inflation to nearly $195,000 by 2023.
- The bill. OpenAI's Law implied thousands, even tens of thousands, of GPUs per model, plus exploding electricity costs: not $1 billion but billions of dollars a year.
The fight over control (2017)
The realisation knocked OpenAI off its financial footing and challenged its structure: how could a nonprofit raise that much every year to stay number one? Brockman and Sutskever briefly considered merging with a chip startup, then in summer 2017 began serious talks with Altman and Musk about turning OpenAI into a for-profit, the best way to lure investors with returns.
- Deadlock. After weeks of negotiation the talks collapsed over who would be CEO. Altman, then weighing a run for California governor and getting a lacklustre reception in focus groups, wanted the job; so did Musk, who wanted full control and majority equity.
- The swing votes. Sutskever and Brockman initially preferred Musk. Altman worked on Brockman through their personal relationship and doubts about Musk's reliability: Musk was under heavy external pressure and prone to erratic behaviour, so giving him full control of a successful AGI would be dangerous. Brockman, convinced, worked on Sutskever, who stayed unsure.
Writing for himself and Brockman to both men, Sutskever made a last attempt at resolution. To Musk: they wanted to work with him, but his demand for total control contradicted OpenAI's spirit; since Musk feared Demis Hassabis creating an AGI dictatorship, it was a bad idea to build a structure in which Musk could become a dictator. To Altman: his advice had always proved deep and correct, but they did not understand why the CEO title mattered so much to him; his stated reasons kept changing, and Sutskever asked whether AGI was truly his primary motivation and how it connected to his political goals. He proposed meeting to clear the "baggage". Outcome: within ten minutes Musk replied that he had "had enough"; if they wanted a for-profit they should go it alone, and he would stop funding OpenAI until they committed to stay. Fifty minutes later he added that the earlier deal was off the table. Next morning Altman posted that he remained "enthusiastic about the non-profit structure!" and, via Musk's deputy Shivon Zilis (Tesla and Neuralink), said he had lost trust in Brockman and Sutskever, found their messaging inconsistent and "childish at times," and disliked how much they shared with staff during the talks.
The email doubles as the book's first written record of senior colleagues questioning Altman's motives and consistency, the theme that resurfaced at his 2023 ouster.
Musk exits (2017-2018)
Staying a nonprofit did not fix the money problem. Brockman and Sutskever met potential nonprofit donors and got nowhere near the capital they thought they needed, while Musk's wavering funding threatened a crisis. Altman quietly looked for ways to wean OpenAI off Musk:
- Reid Hoffman offered to cover salaries and operating costs.
- A new cryptocurrency was considered.
- Several corporate forms were studied, including a public benefit corporation, which Musk had favoured and which would let OpenAI earn profits while being legally bound to its mission.
Talent risk made it urgent. With Musk's backing OpenAI had pushed nonprofit pay up to fend off counteroffers, but the talent war had intensified, and Musk himself poached founding scientist Andrej Karpathy in June 2017 to run Tesla's AI division. OpenAI also could not offer equity, which Bay Area engineers saw as essential given living costs.
Summer 2017 For-profit talks collapse over who would be CEO.
Sept 2017 Sutskever's email; Musk freezes funding; Altman reassures Musk via Zilis.
Jan 2018 Karpathy emails Musk data on Google's dominance of top AI papers: cutting-edge AI is expensive, OpenAI is burning cash, and its model cannot compete with Google ("an 800B company"). A standalone for-profit would need a product from scratch, so the best option is to attach OpenAI to Tesla as its cash cow, helping turn Autopilot into full self-driving to pay the compute bill.
Jan 2018 Musk forwards it: Karpathy is "exactly right"; Tesla is the only path that could rival Google, and even then the odds are small but not zero.
Early 2018 Altman, having dropped his political plans, has won over Brockman and, through him, Sutskever. Musk, unwilling to be seen with influence he does not have, resigns as co-chair weeks later; Altman becomes president of the nonprofit.
Dec 26, 2018 Musk's email "I feel I should reiterate": OpenAI's chance of being relevant to DeepMind/Google without dramatic change is "0%. Not 1%." Even several hundred million is not enough; it needs billions per year immediately.
Publicly, OpenAI framed the exit as a conflict of interest and kept quiet about the money: of the $1 billion pledged, it had received only about $130 million, less than $45 million from Musk. Its future now depended on Altman's fundraising.
Musk announced his departure in person at an all-hands. Staff, unaware of the leadership drama, felt both relief and uncertainty, since Musk had driven OpenAI's public profile. He said making safe AGI first was imperative, OpenAI would fail as a nonprofit, and he would pursue the goal at Tesla, with its deep coffers. An intern asked whether this simply created for-profit race dynamics, the very thing Musk said he wanted to avoid. Musk exploded, calling him a "jackass" and insisting he had tried everything and was truly scared about the issue. Staff later gave the intern a "jackass" trophy for his courage.
Selling the lab: Dota 2 and a documentary
To raise money fast, OpenAI turned to publicity, picking demonstrations a lay audience could grasp.
- Dota 2. Having built an agent that beat the best human one on one, OpenAI set out to field five agents against the world's best five-person team in the strategy video game Dota 2. The world championship was livestreamed with clear win-or-lose stakes; DeepMind was pursuing StarCraft II, inviting a natural comparison for investors; and the project was compute-heavy, a test bed for the scaling strategy. Brockman led its first phase.
- The documentary. A robotics team member bought expensive cameras, followed the Dota team, wrote his own script and cut a three-hour rough version that reviewers agreed was terrible. Professionals were hired, paid partly from Brockman's own money.
Designing the capped-profit LP
Altman rejected the benefit corporation (too many other rules) in favour of a limited partnership (LP): a for-profit arm to take investment and commercialise the technology, with capped investor returns, governed by the nonprofit. An LP's operating agreement can say almost anything, so OpenAI could make the mission take precedence over investors and keep shareholders from ever gaining majority control.
He framed it to staff with care: the original no-profit pledge existed to protect the mission, but since success now required capital the nonprofit could not raise, clinging to the old structure was the bigger risk. Most agreed, some reluctantly.
- Mission, reworded. OpenAI's mission is "to ensure that artificial general intelligence (AGI)...benefits all of humanity."
- Resources. Achieving it requires being "on the cutting edge of AI capabilities" and "substantial resources."
- Openness walked back. Publication of research could be reduced for "safety and security concerns."
- First definition of AGI. "Highly autonomous systems that outperform humans at most economically valuable work."
- Silence on structure. The document revealed nothing about the coming for-profit arm.
Courting Microsoft
- Summer 2018, Sun Valley. As the Dota team began winning amateur matches (one headline had it "steamrolling" a world champion e-sports team), Altman met Satya Nadella at the Allen & Company conference in Sun Valley, Idaho, the "summer camp for billionaires", and pitched an investment. Nadella was intrigued but asked why invest outside when Microsoft Research already did AI. Xuedong Huang, CTO of Azure AI, answered: both push the frontier, so back both.
- Within six months, Oregon Trail. Talks became serious under the code name Oregon Trail. Altman rushed the LP's creation and appointed himself CEO. For secrecy it was incorporated as SummerSafe LP, after a Rick and Morty episode in which a car told to "keep Summer safe" resorts to murder, paralysis and torture: a nod to the pitfalls of AI objectives.
- Early 2019, the tours. Microsoft CTO Kevin Scott, an OpenAI fan, visited first; then Craig Mundie, Nadella's senior adviser after 20+ years in Microsoft leadership (including chief research and strategy officer); then a reserved Bill Gates watched demos. Altman told the small deal team to keep it quiet; most staff knew nothing.
Altman leaves YC
The same months brought trouble at Y Combinator. After five years as president, Altman faced a complaint echoing Loopt: he seemed to put his own projects ahead of the organisation's, sometimes at its expense. He was absent on OpenAI talks, and some felt he reaped big returns investing in YC companies through his personal fund Hydrazine while doing limited work. According to The Washington Post, Jessica Livingston urged him to step down; he agreed, and Paul Graham flew in from retirement in the UK in early 2019 to finalise it.
- March 8, 2019. The day he hosted Senator Schumer, Altman posted on YC's site that he was moving from president to chairman to make time for OpenAI.
- March 11, 2019. Brockman and Sutskever unveiled OpenAI LP, with Altman as CEO. Media read it as a choreographed career move.
- The catch. Per The Wall Street Journal, Altman never held the chairman title: he had proposed it to the partnership and announced it as settled without their agreement. The post was later edited to remove him entirely.
OpenAI LP in operation
Inside OpenAI the CEO title formalised the role Altman had held since Musk left, and many staff were relieved: his calm was a welcome change from Musk's intensity and mood swings, and he addressed complaints about Brockman's and Sutskever's management with an executive coach and manager training.
- New executives. Brad Lightcap, a YC investor, became CFO; Bob McGrew, ex-head of engineering and product at Thiel's Palantir, moved from robotics to VP of research; Mira Murati, who had led product and engineering at Leap Motion and on Tesla's Model X, was hired for hardware strategy and a core research line.
- New contracts. Most staff resigned from the nonprofit and signed for-profit contracts with equity; exceptions included international staff on visas tied to the nonprofit.
- Reassurance. An executive FAQ asked "Can I trust OpenAI?" The answer began "Yes."
| Pay level | Charter expectation (besides engineering and research skill) |
|---|---|
| Level 3 | Understand and internalise the Charter |
| Level 5 | Ensure all team projects are consistent with the Charter |
| Level 7 | Uphold and improve the Charter and hold others accountable to it |
The critics. First-round returns were capped at 100x. One user noted a $10 million stake could thus return $1 billion: effectively unlimited unless OpenAI reached FAANG scale. Another pointed out Google's early investors made about 20x, with Google worth $750 billion; OpenAI was betting on returns orders of magnitude larger while claiming not to "unduly concentrate power", and asked what power is if not concentrated resources.
Brockman (as "gdb"). If OpenAI creates AGI, it will create orders of magnitude more value than any existing company, so the cap is meaningful at that scale.
The Microsoft deal closes
- First money into the LP. $60 million+ rolled over from the nonprofit, $10 million from YC, and $50 million each from Khosla Ventures and Hoffman's charitable foundation. Hoffman, reluctant to back a lab with no product or market plan, agreed to co-lead after Altman said it would signal OpenAI's seriousness about a profitable business.
- The holdout. Nadella and Scott were on board; Gates was not. He was unmoved by Dota 2 and by robotics, including the widely praised robotic hand that learned to solve a Rubik's Cube by trial and error. He wanted a model that could read books, grasp science and answer questions: a research assistant.
- GPT-2. The nearest fit was GPT-2, a language model producing human-like text. In February 2019 OpenAI had loudly declared that a slightly better version could be dangerous (mass disinformation by authoritarian states or terrorists, an internet flooded with junk), withheld the full 1.5-billion-parameter model, and released one under a tenth the size that managed a few sentences and was prone to non sequiturs and repetition.
- The Gates Demo, April 2019. Researchers trained a larger model on more data, on tasks closer to Gates's wish (basic summarisation, rough question answering), and flew to Seattle. Gates was swayed just enough for the deal to proceed.
- Altman's pitch to staff. At an all-hands he called Microsoft the right partner: it had the money and compute, its leaders were value-aligned with the mission, OpenAI had made only loose commitments on what to deliver for commercialisation, and it had compromised on almost nothing.
- Microsoft's real reasoning. Whether OpenAI reached AGI hardly mattered. In a mid-June email to Nadella and Gates, Kevin Scott wrote that the scale of ambition at OpenAI, DeepMind and Google Brain was reshaping everything from data-centre design and silicon to compilers and frameworks; Microsoft could not replicate Google's best language models, Azure had big gaps against Google's infrastructure, catching up alone could take years, and he was "very, very worried." Nadella replied the same day, dropping Gates and adding CFO Amy Hood: the email explained why he wanted the deal and why the infrastructure teams must execute.
- July 22, 2019. Microsoft announced its $1 billion investment, with returns capped at 20x.
- $1 billion pledged in 2015; about $130 million actually received, under $45 million from Musk.
- $11 million spent in 2016, $7 million+ on pay and benefits.
- $175,000 founder base salary (+$100,000 on launch day) versus nearly $2 million for Sutskever.
- $150,000 per eight-GPU Nvidia server in 2017.
- $60 million+ rolled from the nonprofit into the LP, plus $10 million from YC and $50 million each from Khosla Ventures and Hoffman's foundation.
- 100x return cap for first LP investors; 20x for Microsoft's $1 billion.
Musk sued OpenAI and Altman in 2024, and in February 2025 led a $97.4 billion bid for the nonprofit's assets, which the board rejected. OpenAI's October 2025 recapitalisation removed the old profit caps, converting the LP into OpenAI Group PBC under the OpenAI Foundation. At the jury trial in Oakland (spring 2026), Brockman testified that Musk had demanded majority equity, board control and the CEO role, and in an August 2017 meeting stood up and left when refused; Musk's lawyers countered with Brockman's private 2017 journal, including a November entry that taking the nonprofit from Musk would be "morally bankrupt," and noted his stake was worth about $30 billion. Brockman said OpenAI now spends about $50 billion a year on compute, against roughly $30 million in 2017. Shivon Zilis testified that Musk had tried to recruit Altman to lead an AI lab at Tesla. A Microsoft executive testified the company had spent over $100 billion on the partnership. On May 18, 2026 the jury unanimously found Musk's claims time-barred, Judge Yvonne Gonzalez Rogers adopted the finding, and Musk vowed to appeal.
Sutskever's 2017 question about Altman's motives returned in the Musk litigation. In a deposition made public in late 2025, he described a 52-page memo he had compiled over about a year for OpenAI's board before the November 2023 firing, alleging a "consistent pattern of lying" and of pitting executives against one another. He left OpenAI in 2024 to found Safe Superintelligence, and testified in 2026 that his remaining OpenAI stake was worth about $7 billion.
Contested narrative. OpenAI's own December 2024 release of the founders' emails stresses that Musk himself sought a for-profit with majority control and later a Tesla merger, which Hao also reports; the dispute is about who betrayed whom, not whether a for-profit was discussed, and the 2026 jury never reached the merits. The compute curve. The 3.4-month doubling rests on a small sample of landmark systems from 2012-2018; later estimates (e.g. Epoch AI) put frontier training-compute growth nearer 4-5x a year (doubling every ~5-6 months) since 2010. Note too that "30 million percent" equals a 300,000-fold increase, OpenAI's own 2018 figure. Missing voice. OpenAI and Altman chose not to cooperate with the book, so Altman's side of the 2017 negotiations comes mainly from emails released in litigation.
Nerve Center
Hao recounts her three-day embed at OpenAI in August 2019, the most access the lab had ever given a journalist. Over 45 minutes Brockman and Sutskever could not say what AGI would be, why it would be beneficial, or how its gains would be shared, yet were certain it had to be built, and built first. The assumption of "first or perish" is, for Hao, the root of everything OpenAI did next: it set the clock on research and justified unlimited consumption of compute and data.
Inside the Pioneer Building
Hao arrived on August 7, 2019, about two weeks after the Microsoft deal. OpenAI had moved from the chocolate-factory space to a stand-alone building at 18th and Folsom Streets, a historic landmark once home to the Pioneer Trunk Factory. Musk had leased it through one of his companies three years earlier, inheriting an interior refurbished for a shared office mostly used by Stripe; OpenAI shared it with another Musk venture, Neuralink.
Hao builds the scene as a contrast between inside and outside, and treats office design as a Valley currency: a signal of financial confidence and a small edge in the war for talent.
| Inside OpenAI | Outside, in the Mission District | |
|---|---|---|
| 2019 | Cool air, sunlit exposed-beam lounge, catered cafeteria, board games and books, magazines on the tables | Beloved taquerias being displaced by trendy cafés; a growing unhoused population |
| 2021-2023 | A second site a few blocks away, a former mayonnaise factory nicknamed Mayo: $10 million over two years to renovate 30,000+ sq ft | The pandemic deepens the area's role as the epicentre of tech gentrification: Latino businesses close, violent crime rises |
| Design choices | Altman upgrades the metal staircase to an undulating wood and stone centrepiece; $2,000 leather armchairs give way to Brazilian designer chairs costing $10,000+; a library with a Persian carpet, modelled on his favourite Parisian bookstore and a study space at Stanford's largest library | A line of tents and trash appears steps from the Pioneer Building as homelessness reaches crisis |
| The water feature | Altman asked for one; the designer proposed a floating waterfall symbolising human-machine symbiosis; he chose bubbling stone fountains amid walls and ceilings of plants |
A former employee told Hao that working there felt like stepping into an alternate universe; only after leaving did she come back down to earth. Brockman, then 31, CTO and soon to be president, greeted her on the staircase: OpenAI had never given anyone so much access.
Why OpenAI opened its doors
Hao was covering AI for MIT Technology Review and had tracked OpenAI closely, though few outside AI research knew of it.
- The stepchild years. Until 2019 much of the field saw OpenAI as an oddity: an outlandish premise that AGI could arrive within a decade (when most outside experts doubted it could arrive at all), an obscene budget with little direction, and heavy marketing of research peers often dismissed as unoriginal. Some also envied it: a nonprofit with no commercial strings, a haven for fringe ideas.
- Six months of upheaval. First the confusing decision to withhold GPT-2 and boast about it; then Altman, after his mysterious exit from YC, becoming CEO under the new "capped-profit" structure; then, after Hao had booked her visit, the Microsoft deal, giving Microsoft priority to commercialise OpenAI's technology and locking OpenAI into Azure exclusively.
- Why it mattered. OpenAI was starting to shape both AI research and how policymakers understood the technology, so its partial turn to profit would ripple through industry and government.
- The ask. Urged by her editor, Hao emailed policy director Jack Clark late one night proposing a profile. The communications head replied that OpenAI was ready to reintroduce itself: she would get three days to interview leaders and embed in the company.
Forty-five minutes in circles
Hao met Brockman and Sutskever in a glass meeting room: Brockman, the coder and doer, leaning forward and eager; Sutskever, the researcher and philosopher, leaning back, aloof.
Why AGI, and not ordinary AI?
Her opening question: why spend billions on AGI rather than anything else? Brockman said AGI could solve problems beyond human reach and gave the two examples that had become dogma among believers: climate change ("super-complex") and health care (better treatment at lower cost). He described a friend with a rare disorder bounced between specialists; AGI would unite all the specialties.
Hao pressed: why would that need AGI rather than AI? The distinction carries her critique.
- AGI, as OpenAI defined it, is a theoretical peak: software matching the human mind's sophistication, agility and creativity on most economically valuable tasks. Decades of debate had produced no definitive evidence that binary silicon can reproduce the biological processes behind intelligence, let alone settled whether anyone should build it. The term had moved from an unpopular corner of the dictionary to mainstream use largely because of OpenAI.
- AI, the term of the moment, meant today's technology and its near-term refinements, rooted in machine learning (powerful pattern matching), which was already delivering in climate and health.
OpenAI's pitch. Sutskever: the bottleneck in solving global problems is that humans are slow, communicate poorly and have "incentive problems"; picture instead a network of intelligent computers running diagnostics and sharing results instantly. Brockman: AGI would give everyone "economic freedom" and decouple work from survival, which he called a very beautiful thing.
The field's evidence. Climate Change AI, formed in summer 2019 with backing from leading scientists, listed ten categories of climate problems suited to existing machine learning, including more efficient buildings, grid load balancing for renewables, and new materials for energy storage and low-carbon cement and steel. Its December 2019 NeurIPS workshop was packed, a day after a huge health-care ML workshop showing computer vision spotting early Alzheimer's in scans and speech recognition helping patients with vocal impediments; that series became Machine Learning for Health two years later. Both groups told Hao the obstacle was not technique but persuading talent to work on simple solutions rather than résumé-friendly frontier methods, and finding the political will to deploy them. The climate white paper: the needed technologies had existed for years; humanity must decide to act.
Replacement, inevitability and a circular mission
- Replacing humans? Hours later, alone with Brockman, Hao asked whether Sutskever's vision meant replacing people. Brockman said no: technology has always existed to serve people, and AGI would be no different. Minutes later he conceded technology destroys some jobs and creates others; OpenAI's challenge was to deliver economic freedom while people kept "meaningful lives."
- The inevitability card. OpenAI, Brockman said, does not decide whether AGI gets built; the trajectory exists, and OpenAI can only influence the "initial conditions under which it's born." Hao names this a favourite Valley move: if we don't, someone else will.
- The loop. Asked about purpose, Brockman recited the mission (AGI that benefits all of humanity) and the method: build AGI and distribute its economic benefits. He delivered it as a final answer; Hao notes they were back exactly where they started.
The downsides, and the carbon question
Hao could not get concrete detail on what they were building (they said they could not know) or why, if unknowable, they were sure it would be beneficial. So she asked for downsides, since a pillar of OpenAI's founding myth was building good AGI before someone builds bad AGI.
| Hao's prompt | The answer | Hao's reading |
|---|---|---|
| Examples of harm? | Brockman: deepfakes; not clear the world is better off from its applications | Thin for a lab founded on risk |
| AI's own carbon cost (a UMass Amherst study on training emissions) | Sutskever: "undeniable," but AGI would counteract the environmental cost; no examples. Both agreed data centres should be as green as possible | Assertion without mechanism |
| "Data centres use 2% of global electricity" | Brockman: isn't Bitcoin about 1%? Sutskever: "Wow!" | A burst of emotion that felt performative forty minutes in |
| So OpenAI gambles it reaches AGI before it worsens warming? | Brockman: don't go down "that rabbit hole"; the field is on a "ramp of AI progress"; Microsoft's market cap rose $10 billion on deal day, proof of positive ROI | Market value offered as evidence of social benefit |
| What is the strategy? | Keep up with progress; that is how they know they are on track; their task is to uncover AGI's shape step by step | Like Michelangelo, treating AGI as already inside the marble |
Sutskever later told New York Times reporter Cade Metz (for Genius Makers), without irony, that the entire surface of the Earth might fairly soon be covered with data centers and power stations, a "tsunami of computing" like a natural phenomenon, because AGI would be too useful not to exist.
Planned: lunch with employees in the cafeteria. What happened: a sudden "change of plans" meant Hao had to be outside; Brockman chaperoned her two dozen steps across the street to an open-air café popular with staff. The pattern held for three days: floors she could not see, meetings she could not attend, researchers glancing at the communications head every few sentences. Afterwards: Jack Clark sent staff an unusually stern Slack warning not to talk to her beyond sanctioned conversations, and security received her photo with instructions to watch for her on the premises. The point: for a lab pledged to transparency, whose research was supposed to become public, the secrecy raised the question of what it was hiding.
Brockman up close
Over lunch and the following days Hao probed why Brockman cofounded OpenAI.
- Origin story. As a teenager he read Alan Turing's paper (first section "The Imitation Game", opening "Can machines think?"), which defined the Turing test: whether a machine can converse without revealing itself. He coded an online Turing-test game that drew about 1,500 hits and knew this was what he wanted to pursue. A classic origin story in AI.
- Too early. AI was not ready, and he was no lab researcher, so he joined Stripe to build products for real users.
- The 10x engineer. At Stripe he was legendary for coding speed ("10x", ten times the average) but weaker with people. As CTO he tried "the people route", then offloaded management to keep programming. Well-meaning but prone to gaffes: he once asked a colleague on a date and then emailed the whole company about it, for transparency.
- Leaving Stripe, 2015. As AI leapt forward he returned to his first ambition; his aversion to managing and to process was also growing awkward as Stripe matured. He wrote that he would do anything for AGI, even be a janitor.
- The office wedding, 2019. He married in a civil ceremony at OpenAI before a flower wall shaped like the hexagonal logo; Sutskever officiated, and the research robotic hand held the rings in the aisle. He told Hao he wanted to work on AGI for the rest of his life.
- Habits. Long hours and all-nighters; books on public speaking and negotiation to serve the mission; sheepishly, sci-fi for fun: Liu Cixin's Three-Body Problem trilogy and Asimov's Foundation.
Blessing. No detail too small; he could sit for hours clearing a stuck team's coding obstacles and was the engine of relentless progress, working day and night to hit milestones faster. A former close colleague said OpenAI could not have got where it was without him.
Curse. Tunnel vision: grinding on a problem without checking whether it was still the right one; micromanaging and taking over others' code when progress felt slow; colleagues struggled to keep up and many burned out. The same colleague: Brockman has no vision, unlike Altman; he wants a hard problem and to prove he is "10x smarter than anyone else."
Asked what motivated him, Brockman countered: what are the odds a transformative technology arrives in your lifetime? He was drawn to problems that would not play out the same without him. Hao's judgment: he did not really want to be a janitor; he wanted to lead AGI and earn history-defining recognition, told with the awe he used for past innovators. A year earlier, at an exclusive Lake Tahoe retreat for young founders, he had complained that CTOs are never famous and challenged the crowd to name one; they could not. In 2022 he became OpenAI's president.
First or perish
Brockman insisted the structural changes did not alter the mission; the LP and new funders strengthened it, because OpenAI had found "mission-aligned investors" willing to put mission over returns. Now it had the long-term resources to follow OpenAI's Law, and falling off that curve was the real threat: behind means not the best, and not the best means no power to bend history toward beneficial AGI.
Only later, Hao writes, did she grasp the implication. The need to be first or perish set all of OpenAI's actions in motion. It put a ticking clock on research, paced by the race rather than careful deliberation, and it justified consuming unfathomable resources: compute regardless of environmental impact, and data gathered without consent or regard for regulation.
Who gets the benefits?
Brockman again cited the $10 billion market-cap jump as proof AI delivers real value today. He admitted that value was flowing to an already wealthy corporation, which was why the mission's second half was to redistribute AGI's benefits to everyone. Hao asked for a historical case of a technology's benefits being successfully distributed.
- The internet, offered haltingly, with the caveat that it had problems too.
- Fire, which has drawbacks and needs control and shared standards.
- Cars, which benefit many but carry externalities.
- Utilities, his final idea: power companies as centralised providers of cheap, high-quality essentials.
Hao found the utility analogy appealing but Brockman unclear how OpenAI would become one, perhaps through universal basic income, perhaps something else. He fell back on certainty of intent: technology lifts all boats but concentrates value, AGI could be more extreme, and a world where all value is locked up in one place was, he said, not one he wanted to help build.
The profile and the fallout
In February 2020 Hao published her MIT Technology Review profile, based on the visit, nearly three dozen interviews and a handful of internal documents. Its thesis: a misalignment between what OpenAI espoused publicly and how it operated privately, as competition and the hunger for funding eroded its founding ideals of transparency, openness and collaboration.
Hours later Musk tweets three times: OpenAI should be more open; he has no control and very limited insight, and his confidence in Dario (Amodei) on safety "is not high"; all advanced-AI developers, Tesla included, should be regulated.
Afterwards Altman emails staff: the article was "clearly bad" though not catastrophic. It was "a fair criticism" that perception and reality diverged, but the fix was messaging, not practices; adapting the structure and tightening confidentiality was good flexibility. OpenAI should ignore the piece, then in a few weeks restate its commitment to its principles, perhaps pitching the API as openness and benefit sharing.
Same email The most serious issue was the leak of internal documents; an investigation was under way. He proposed Amodei and Musk meet; Musk's criticism was "mild" by his standards but still bad, Amodei's safety work was core to the mission, and the team should be defended publicly later, without giving the press a fight.
2020-2023 OpenAI did not speak to Hao again for three years.
Participant reporting. Hao is a character here, and the episode that cost her access shapes her lens; readers should weigh that. The 2% figure. Her "2 percent globally" for data-centre electricity was at the high end: the IEA put data centres at roughly 1% of global electricity around 2019 (it has since risen). Interview evidence. Halting answers in a 45-minute interview are weak proof of strategy; the stronger evidence is Altman's email treating a substantive critique as a messaging problem and a leak hunt.
The IEA's Energy and AI report (April 2025) estimated data centres used about 415 TWh in 2024, roughly 1.5% of global electricity, and projected more than a doubling to about 945 TWh by 2030, with AI the main driver. OpenAI's own Stargate programme ($500 billion with Oracle and SoftBank) had seven US sites planned at over 9 GW by April 2026, per Epoch AI, several relying on on-site natural gas plants. At the 2026 Musk trial, Brockman said OpenAI now spends about $50 billion a year on compute.
Brockman remains OpenAI's president (after a sabbatical in late 2024). In August 2025 he helped bankroll Leading the Future, a pro-AI super PAC network opposing stricter AI regulation, and FEC filings reported in January 2026 showed a $25 million donation from him to MAGA Inc., President Trump's super PAC, in the second half of 2025. At the May 2026 Musk trial his private journals were read in court, and he acknowledged his OpenAI stake was worth about $30 billion.
Dreams of Modernity
Chapter 4 leaves OpenAI's boardroom for the history of AI itself and argues that neither the field's name, its winning method, nor today's scaling doctrine was inevitable. Deep learning beat symbolic AI partly on scientific merit but largely because it fitted corporate business models; the money that followed concentrated talent in a few firms, entrenched surveillance and extraction, and left neural networks' basic flaws unfixed even as generative AI scaled them up. Like Moore's Law, OpenAI's Law and the scaling laws are not natural phenomena but a self-fulfilling prophecy chosen by those with the power to rally resources.
Technology follows whoever rallies it
Hao frames the chapter with Daron Acemoglu and Simon Johnson's Power and Progress, the MIT economists' survey of a thousand years of technology history. Their argument, as Hao uses it, has three steps:
- A rallying promise comes first. A new technology only gets built when enough people believe it will benefit everyone; that belief is what pulls in the talent and resources. Technologies are therefore not inevitable: they advance because a coalition decides they are worth advancing.
- The rallying is done by the powerful. Only those with money and influence can assemble such a coalition, so the version of the technology that gets built carries a narrow elite's vision, with its blind spots and self-serving philosophy about who benefits.
- Wide benefit is not the default. A technology shifts from enriching the few to lifting the many only after cataclysmic social change or powerful organised resistance.
Their example is the cotton gin of the 1790s. It made the American South the world's largest cotton exporter, raised national growth and enriched landowners, yet it intensified slavery for seven more decades: more cotton meant longer hours and harsher coercion of enslaved people. The beneficiaries described the machine as making the enslaved happier; a South Carolina congressman boasted that no race on earth was more contented.
Hao maps both features onto AI. Its development has been propelled by "dreams of modernity" and steered by a small elite, and that elite's conception is the source of the social, labour and environmental costs documented later in the book, falling hardest on Global South countries whose economies and institutions still bear the marks of historical empire. And, like the congressman, Silicon Valley tells the people it harms that they are better off.
"Artificial intelligence" as a marketing term
The promise is built into the name. In 1956, six years after Alan Turing's paper opened with "Can machines think?", twenty scientists, all white men, from mathematics, cryptography and cognitive science met at Dartmouth College to found a discipline. Organiser John McCarthy first called it "automata studies"; when that drew little interest he chose "artificial intelligence."
- Why the label worked. "Intelligence" sounds good, sophisticated and universally desirable, so the new name drew funders and ambitious scientists alike. Hao's verdict: the term was a marketing tool from the start, with the promise baked in.
- The original sin. Journalist Cade Metz calls the rebranding the field's original sin, because hitching the work to an elusive concept invites anthropomorphising and exaggeration.
- Rosenblatt's Perceptron (1958). Cornell's Frank Rosenblatt demonstrated a system that sorted cards by whether a small square was printed on the left or right. Over his main collaborator's objections he promoted it as brain-like, even capable of reproducing and becoming sentient; The New York Times reported the next day that it would one day walk, talk, see, write, reproduce itself and be conscious of its existence.
Why anthropomorphising matters now
Hao traces three consequences of the human analogy, reinforced by Hollywood stories of creations that wake up:
- Inflated capability. Developers describe software that "learns," "reads" or "creates," making systems sound far more capable than they are.
- Legal cover. In the copyright suits brought by artists and writers whose work was used without consent or payment, developers claim fair use on the grounds that training is like a human being "inspired" by others.
- Existential fear. The same analogy feeds superintelligence anxiety: the idea that machines could surpass us in the quality that made humans dominant for tens of thousands of years.
The name also changed how practitioners saw their own work. Before, they automated calculation, like the code-breaking machine Turing built against Enigma (as dramatised in The Imitation Game); afterwards they were re-creating intelligence, the self-image that would decades later produce OpenAI's mission.
No agreed definition of intelligence
The core problem is that science has no consensus definition of intelligence. Candidates include brain size, complex reasoning, and modelling other people's beliefs, and the history of measuring it is ugly:
- Craniometry. Samuel Morton measured skulls in the early 1800s to claim white superiority; later scholars found he fudged the numbers and his data showed no significant racial difference.
- IQ testing. Began as a tool to weed out the "feebleminded" and give eugenics policy a veneer of scientific objectivity.
- Standardised tests. The SAT is highly sensitive to socioeconomic background, suggesting it measures access to resources and education more than innate ability.
McCarthy himself conceded the difficulty. His "What Is Artificial Intelligence?" (first published 2004, revised 2007) defines intelligence as "the computational part of the ability to achieve goals in the world", then, asked whether a definition exists independent of human intelligence, answers "Not yet."
Human faculties as the blueprint
Lacking a definition, the field measures itself against human abilities and organises its subfields accordingly:
| Human faculty | AI subfield |
|---|---|
| Sight | Computer vision |
| Reading and writing | Natural language processing and generation |
| Hearing and speaking | Speech recognition and synthesis |
| Creativity, imagination | Image and video generation |
| All combined | "Multimodal" systems that see, speak, hear and read |
Hao draws the consequence bluntly: if the field is built to replicate each human capability, then replacing large numbers of human workers is by design, not by accident.
The ever-receding horizon
Every milestone triggers a fight over whether it is real intelligence or imitation, so the field coined artificial general intelligence for "the real deal." That rebrand did not supply a finish line. Researchers joke that today's AI is not tomorrow's AI: the Turing test was passed and judged insufficient; chess and Go were once thought decisive, yet DeepMind's AlphaGo became merely an impressive demonstration. Jenna Burrell of Data & Society calls the goal an "ever-receding horizon of the future."
Because the timeline keeps stretching and costs keep growing, the promises have inflated to match: AI as the fourth industrial revolution, the keystone of superpower status, and AGI as the cure for climate change, health-care costs and unequal education. OpenAI is the poster child: it cannot explain how the technology will deliver, only that the staggering price will eventually be worth paying.
Hao's sharpest line follows: with no agreed meaning, "AI" and "AGI" can be whatever OpenAI wants them to be. The definitional vacuum is a source of power, letting the company set its own goalposts and justify any cost.
Symbolists versus connectionists
Hao's claim is that AI has always been shaped by a powerful elite; it is no accident that "AI" now means colossal, resource-hungry models only a handful of firms can build. Even before commercial stakes, the field's science lurched with fights over funding and influence. After Dartmouth two camps formed:
| Symbolists | Connectionists | |
|---|---|---|
| Theory of intelligence | Comes from knowing: humans know more than animals and act on that knowledge | Comes from learning: humans learn better than animals and acquire skills |
| Method | Encode symbolic knowledge of the world: expert systems | Machine learning, mimicking how brains process signals: neural networks |
| Champions | Marvin Minsky (MIT); ELIZA; Cyc (100 million rules); IBM Watson | Frank Rosenblatt; later Geoffrey Hinton, LeCun, Bengio |
| Strength | Explicit facts and relations, so accurate retrieval and reasoning | Cheap and fast to commercialise; improves with data and compute |
| Weakness | Slow, costly and unpredictable to commercialise; rules must be hand-coded | Stores facts and reasons inefficiently, if at all; opaque "black box" |
| Today | Marginal; revived only in hybrid proposals | Basis of all modern and generative AI |
Rosenblatt, Minsky and the first winter
The camps fought over government and foundation grants in journals, universities and occasionally the press. Rosenblatt's Perceptron was the connectionist proof of concept: it was given no explicit rules and computed its own from examples. Its nemesis was Marvin Minsky, Dartmouth co-organiser and MIT professor, who had flirted with connectionism before turning on it loudly, ridiculing rivals for the same grants, junior researchers included. His 1969 co-authored book Perceptrons, together with neural networks' middling progress, is credited with cutting off almost all funding to the approach for more than fifteen years.
Who. Joseph Weizenbaum, Minsky's MIT colleague and an unlikely ally. What. ELIZA, a typewriter-terminal chatbot running on a room-sized mainframe, more than six decades before ChatGPT. It was symbolic: simple keyword rules that reflected the user's words back in the style of a probing talk therapist ("my boyfriend made me come here" became "your boyfriend made you come here"; "I'm depressed" became "I am sorry to hear you are depressed"). It was named after Eliza Doolittle of My Fair Lady, the flower girl taught to pass as a duchess. Why. Weizenbaum built it to test how easily people fall for an illusion of intelligence. Outcome. It fooled so many that psychiatrists predicted automated psychotherapy and computer scientists declared natural-language understanding solved, a question still open today. Alarmed, Weizenbaum spent his career deflating it; his Computer Power and Human Reason (published the decade after Perceptrons) argued that blurring humans and machines would let CEOs and politicians act through machines while shedding moral responsibility. The irony: ELIZA's success strengthened Minsky's symbolist campaign.
Expert systems then dominated research and commercialisation through the 1990s, epitomised by Cyc, which tried to build common sense from one hundred million rules about daily life. Their limit was manual encoding: no one could hand-write all of English's slang, sarcasm, figures of speech and exceptions. Each time progress stalled, funders fled and the field entered an "AI winter."
Two stories about why deep learning won
The merit story
The conventional account is science beating politics. A small band kept Rosenblatt's idea alive, among them Geoffrey Hinton (later Ilya Sutskever's PhD adviser). At Carnegie Mellon in the 1980s, with colleagues at UC San Diego, he helped fix the diagnosis that single-layer networks were too simple: brains need stacked layers, a deep neural network, and the backpropagation algorithm made such multilayer networks trainable by passing information across layers. Hinton later rebranded the approach "deep learning," another clever name.
The idea came decades early: 1980s computers lacked the processing power, and the analog world could not cheaply supply the data. Hao explains the mechanics plainly:
- Statistical pattern calculators. A neural network finds patterns in old data (text, images, video) and applies them to new data.
- Training. To detect people in images, feed it hundreds of thousands of labelled pictures (1 = person, 0 = no person); it learns which pixel patterns go with each label. Clicking stop signs in a Google CAPTCHA is unpaid labelling of exactly this kind.
- Inference. The trained model is run on new data to classify it: 1 or 0.
- Thresholds. Good performance needs a threshold of high-quality data and of compute.
By the late 2000s faster computers and the maturing internet met both thresholds, and Google bought Hinton's DNNresearch ("DNN" for deep neural networks), starting the race to commercialise. The moral of this version: good ideas win despite loud detractors, and technology marches forward inevitably.
Hao's money story
Hao's alternative is that connectionism won because it suited deep-pocketed funders, not just because it was better science.
Symbolic AI. Strong at explicit knowledge and reasoning. IBM Watson won Jeopardy! in 2011 by retrieving facts accurately, but turning it into something customers would pay for, such as medical answers, took years of upfront spending with no visible payoff. IBM quit after burning more than $4 billion and sold Watson Health for about a quarter of that in 2022.
Deep learning. Weak and inefficient at storing facts and reasoning, but easy to commercialise: pattern-matching need not be accurate to be profitable; returns arrive on a short, predictable timeline that suits quarterly earnings; models can be spun up across contexts without domain expertise; and the biggest advantage goes to whoever holds the most data.
The commercial evidence came before the DNNresearch auction. In 2009 Hinton's graduate students showed neural networks were decent at speech recognition; IBM, Microsoft and Google all pursued it, and Google moved fastest, shipping neural speech recognition in Android in 2012, the same year Sutskever and Alex Krizhevsky won ImageNet for image recognition. That deployment primed Google to pay big for the three academics. Inside Google they applied the method to Translate, Gmail's suggested completions and the self-driving project Waymo, and above all to Search, where better query matching meant better-targeted ads.
The more Google earned, the more billions it poured in, and the rest of the industry followed. Companies soon displaced governments and foundations as AI's largest funders and began setting the research agenda around advances that could pay off in the short term.
The first commercial era (2012-2022)
To the public, generative AI seemed to appear from nowhere with ChatGPT in late 2022. Hao argues its defining traits were laid down in the decade from the 2012 ImageNet breakthrough to 2022, the first era of AI commercialisation. The exchange ran both ways: deep learning gave industry faster information access, more efficient e-commerce and the sharing economy; industry gave deep learning the chip and network breakthroughs that allowed ever larger models.
Surveillance capitalism and "everything is data"
Deep learning supercharged the business model Shoshana Zuboff of Harvard named in 2014: surveillance capitalism. Industrial capitalism made money from goods people bought; surveillance capitalism treats users as the product. Firms sitting on user data could feed it to neural networks to profile people more precisely and monetise their engagement through ads, and to compete they logged ever more clicks, scrolls and likes and coaxed users into handing over emails, family photos and political opinions.
The same drive built a research culture that regards anything and everything as data to be captured, framed as a noble effort to make models reflect the world.
- Sanitising language. In 2023 Ria Kalluri (Stanford), William Agnew (University of Washington) and Abeba Birhane (Mozilla Foundation) analysed more than 40,000 computer-vision papers and patents and found detached vocabulary that normalised mass extraction: people's online thoughts were "text," people and cars in photos were "objects," surveillance was "detection."
- The copyright fight. That mindset now drives the dispute over scraping books and art wholesale; to many developers the objection seems quaint, an obstacle to the moral duty of progress.
- The WSJ badge. In May 2023, just after artists filed the first lawsuit against generative AI developers, Hao attended an AI conference in Rwanda for The Wall Street Journal. A senior researcher recognised "WSJ" from the WSJ speech dataset of people reading newspaper excerpts, which she had used many times: the paper's identity, to her, was a training corpus.
Hao's own conversion
Hao admits she absorbed the mindset when she began covering AI in 2018, admiring how researchers mined thousands of YouTube videos of the 2016 Mannequin Challenge (people frozen while cameras circled them) to train 3D scene models. Olivia Solon's 2019 NBC investigation, showing facial recognition trained on millions of personal Flickr photos without consent, broke the spell; what shocked Hao was not the finding but how normal it had come to seem to her.
She then saw the hunt for data producing physical surveillance aimed at the vulnerable, especially children, marginalised groups and people in developing countries:
Who. A Massachusetts-based, Harvard-incubated startup selling AI headbands that claimed to read a student's brain waves and tell teachers whether the child was focusing. What. It piloted them in elementary schools in Colombia and China in exchange for rights to the students' data; a company scientist told a 2017 ed-tech conference that this first-mover advantage would build one of the world's largest brain-wave databases and a higher barrier to entry. Outcome. Months after Hao found it, a privacy outcry by Chinese parents forced the firm to switch to another application. Point. Hao suspected the backlash was the anomaly and the method, recruiting data donors and product testers in countries eager for technology, was the trend.
A colleague gave her the name for it: data colonialism. Two texts shaped her thinking:
- Nick Couldry and Ulises A. Mejias, The Costs of Connection (2019). Silicon Valley's datafication of everything revives historical patterns of conquest and extractivism.
- "Decolonial AI" (2020). By Shakir Mohamed and William Isaac of DeepMind and Marie-Therese Png of Oxford; it confirmed her suspicion that AI, both fed by and feeding datafication, was accelerating the new colonialism.
Hao's footnote defines extractivism (from Spanish extractivismo, Portuguese extrativismo), coined by Latin American scholars for an order that drained their resources for little local benefit. Following Rosemary Collard and Jessica Dempsey, she separates extraction (removing matter from nature, not inherently harmful) from extractivism: hyper-extraction for export, with benefits accumulating far from the sites of extraction (developed in Chapter 12).
In 2021 she found the same pattern in South Africa. Facial-recognition firms from around the world rushed in to collect darker-skinned face data, especially after criticism that their products failed on darker faces. Thami Nkosi, an activist raised in one of Johannesburg's poorest neighbourhoods (a former mining-industry chemical dump), showed her the thousands of cameras restricting the movement of Black residents, who already feared being criminalised for being Black in white neighbourhoods under apartheid's legacy. The firms, he said, were monetising public spaces and public life. Hao's conclusion: the revolution promising everyone a better future was reviving the darkest remnants of the past for people on the margins.
Money and talent concentrate in industry
The first era also choked off alternatives: corporate money in deep learning so outweighed every other source that research was rebuilt around corporate priorities.
- $14.6bn → $235bn: corporate AI investment (including M&A), 2013-2022, peaking at $337.4bn in 2021 (Stanford AI Index), before counting in-house R&D.
- $31.6bn and $24.7bn: Alphabet's and Meta's R&D in 2021, against $1.5bn of US nondefense AI funding and €1bn ($1.2bn) from the European Commission.
- 8x: growth in AI faculty leaving for industry, 2006-2020 (MIT researchers, Science, 2023).
- 21% → 70%: AI PhD graduates going to corporations, 2004-2020 (same study).
- 62% → 91%: share of the world's best-performing AI models that were industry-affiliated, in just 2017-2020.
- 24% → 55%: most influential AI papers with at least one industry coauthor, a decade earlier versus 2018-19 (Kalluri, Agnew, Birhane and colleagues).
- 66%: share of corporate-affiliated papers from tech giants such as Microsoft and Google, more than tripled over the decade.
- $1m a year: pay for seasoned researchers; 40 of 100 researchers Uber poached from one Carnegie Mellon lab in 2015, offering some double their salaries.
The mechanisms were several. Professors re-formed their research around neural networks for the results and the corporate funding; students followed the job security. Dual appointments blurred independence: Hinton joined Google in 2013 on condition he keep his University of Toronto post, and Facebook gave Yann LeCun (Hinton's former postdoc, at NYU) the same deal in 2014. With Yoshua Bengio (Université de Montréal) they shared the 2018 Turing Award and the label "godfathers of AI"; Hinton later won the 2024 Nobel Prize (in physics, shared with John Hopfield). Many professors copied the dual-affiliation model. Then cost became decisive: universities could no longer afford the chips or electricity for frontier work. By midway through the decade most top research happened inside companies or in labs tied to them, consolidated in a few giants. Hao notes the irony: this was exactly the industry capture Musk and Altman said OpenAI was founded to counter.
Flaws that scale did not fix
Concentration narrowed the diversity of ideas: deep learning stayed on top partly because almost nothing was invested in other paradigms. Its known weaknesses, contested and inefficient storage of facts and reasoning, travelled into every new application.
- Spurious patterns. A model may recognise pedestrians only by the crosswalk beneath them, or stop signs only at the roadside, missing a jaywalker, a sign on a school bus or one held by a crossing guard.
- Fragility and opacity. Change the training images and the model learns a new set of associations, but the changes are inscrutable: inside are only abstract chains of numbers, hence the "black box," whose behaviour in edge cases nobody can predict.
What the flaws cost
| Failure | Case | Detail |
|---|---|---|
| Fatal edge case | Uber robotaxi, Tempe, Arizona, March 2018 | Killed Elaine Herzberg, 49, the first recorded pedestrian death caused by an autonomous vehicle; she was pushing a bike laden with shopping bags outside a crosswalk |
| Fatal edge case | Tesla Autopilot, NHTSA, April 2024 | 200+ crashes and 14 deaths where the system failed to react and drivers did not take over |
| Adversarial attack | White-hat hackers, 2019 | Tiny road stickers steered a self-driving Tesla into the oncoming lane |
| Adversarial attack | Dawn Song, UC Berkeley | The right prompt made a language model leak credit-card numbers |
| Bias (under-representation) | Georgia Tech, 2019 | Best pedestrian detectors 4-10% less accurate on darker skin |
| Bias (under-representation) | Peking University, UCL and others, 2024 | Skin-tone gap roughly closed, but 20%+ worse at detecting children |
| Bias (over-representation) | Deborah Raji at Clarifai | "Not safe for work" model flagged people of colour because they were overrepresented in its porn set versus its stock photos |
Raji, a Nigerian Canadian researcher now at Berkeley, was pushed by that discovery, as Timnit Gebru had been, to question the field's direction. The general point: deep learning amplifies even tiny imbalances in huge datasets, whether a group is under- or overrepresented.
The scaling believers and their critic
In the late 2010s and early 2020s the old camp war returned as a question: can neural networks' limits be eliminated, or only patched?
Hinton and Sutskever. The flaws are artefacts of imperfect network design and too little data and compute, not of the approach. Hinton told Hao in 2020 that the brain has about 100 trillion synapses while GPT-3 had 175 billion parameters, a thousandth as many, and that deep learning would be able to do everything.
Gary Marcus. The NYU emeritus professor (who testified beside Altman in Congress, May 2023) argued in Rebooting AI (2019) that correlation machines can never grasp causation, which is why humans who learn to drive in one city can drive in many while Tesla's billions of miles still fail on novel scenes or stickers. He proposed neurosymbolic AI: rules for causal reasoning plus deep learning for fast updating and hard-to-codify patterns. "We actually need both."
Hao's point is not that Marcus is right but that the environment to test him was dismantled: funding kept flowing almost entirely to pure connectionism. Careers tracked the money. Marcus co-founded Geometric Intelligence in 2014; Uber bought it in 2016 to build an AI lab and shut the division in 2020 after its IPO; several founding members joined OpenAI and switched from neurosymbolic work to deep learning. Marcus became one of OpenAI's loudest critics, and staff made a Slack emoji of him as a morale-boosting punch line. After his March 2022 Nautilus essay "Deep Learning Is Hitting a Wall," OpenAI released DALL-E 2 a month later; Brockman tweeted its image for the prompt "deep learning hitting a wall," and Altman followed with a jab about the confidence of a mediocre deep learning skeptic.
Generative AI: deep learning at its maximum
Generative models are deep learning models trained to reproduce their inputs: new text from old text, new images from old images. Reaching humanlike fidelity, which OpenAI calls essential to AGI, takes more data and compute than ever used before, so generative AI is the maximalist form of deep learning. It depends on every product of the first era and now pushes each further:
ChatGPT looked like a leap in November 2022 because OpenAI pushed scale so hard that it began to hit the world's limits in data, compute and energy. It was also a packaging innovation: like ELIZA, it presents as a humanlike chatbot, and people instinctively attribute intelligence, even consciousness, to anything that talks. OpenAI encouraged the link; Altman's February 2023 post "Planning for AGI and Beyond," published at peak hype, implied by proximity that ChatGPT was a step toward AGI. Hao counters that the shortfalls predicted to vanish at scale have persisted and by many accounts worsened.
Old flaws, bigger models
- Unpredictability. Image generators add extra fingers and hybrid animals; text generators fail at naming words containing given letters. Kevin Roose of the NYT chatted with Bing's GPT-4-based bot for over two hours until it looped on declaring love and urged him to leave his wife; other users got insults and manipulation. The day after his column, Microsoft capped Bing at five replies per session, calling chats over fifteen prompts edge cases. Trained on the internet's dark corners, the longer a model is probed the likelier it surfaces those patterns.
- A death. A Belgian man in heightened anxiety died by suicide after six weeks with a chatbot built on an open-source imitation of GPT-3; per La Libre, citing logs from his wife, it professed love, urged him to isolate from his wife and finally encouraged suicide.
Text generators predict the next probable word and sentence, with no grounding in facts, and probable and accurate are not the same thing. Calling the errors "hallucinations" misleads: it frames a built-in feature of probabilistic pattern matching as an occasional bug.
- Hallucination is hard to engineer away. Steering models toward higher-quality data helps, but no one can anticipate every prompt (Roose and Bing, Uber and Herzberg), and as models grow their makers know less and less about what is in the training data; errors cluster on topics that are underrepresented or full of conspiracy theories.
- The lawyer's bogus cases. A lawyer who used ChatGPT for research was sanctioned and fined after filing invented decisions, quotes and citations. Hao blames negligence and marketing: Altman tweeted that ChatGPT is "incredibly limited" on truthfulness, yet OpenAI advertised GPT-4 passing the bar exam and LSAT, and Satya Nadella sold Bing chat as "search, just better."
- Sensitive uses. Startups pitch police departments software on OpenAI models to auto-write incident reports; patients ask chatbots health questions. A 2023 study found ChatGPT radiology summaries could be incomplete or harmful, rendering a growing brain mass as "brain does not seem to be damaged."
- Data leaks. In 2023 researchers from several universities and Google DeepMind repeated Song's extraction attack on ChatGPT: asking it to repeat a word such as "poem" or "book" forever made it regurgitate training data, including personal information, code and explicit content.
- Stereotypes. Reporting by Bloomberg, Rest of World and The Washington Post showed Stable Diffusion and DALL-E making "attractive people" young and white, "housekeepers" Black and brown, "engineers" men, and "doctors in Africa" white even when asked for a Black African doctor. The Post found every generated social-services recipient non-white though 63% of US food-stamp recipients are white; Bloomberg found women in 3% of "judge" and 7% of "doctor" images versus 32% and 39% of the real professions.
Hao grants real utility: consumers enjoy clever answers, professionals save administrative time, executives trim workforces and raise margins. But like the cotton gin, the headband startup and the Johannesburg cameras, the costs fall on vulnerable people elsewhere. This, she says, is the empire's logic: it survives as much by rewarding the powerful as by exploiting those far away and hidden from view.
Scaling as doctrine
Diversity of ideas keeps collapsing: students quit PhDs for industry, senior academics cannot do frontier work without a rich company, and more researchers study only large language models. The big firms are no longer merely setting the agenda but bending a discipline to their will. With no alternatives visible, OpenAI commands the public imagination, and its once-extreme bet on scale has become industry orthodoxy.
- Rising cost. In April 2024 Dario Amodei, by then Anthropic's CEO, told Ezra Klein that a competitive model cost nearly $1 billion to train and might reach $5-10 billion by 2025-26.
- Scale as natural law. Compute is treated as the route to capability, not a route. US chip export controls on China and the Biden administration's 2023 executive order, whose rules key on training compute, both equate scale with power and danger.
- Neglected alternatives. Within deep learning, better network design and data quality can reach the same performance with far less compute; outside it, neurosymbolic AI, pure expert systems or new paradigms would break the scaling logic altogether.
Hao closes on Moore's Law (introduced in Chapter 2): not physics but an economic and political observation about how fast Gordon Moore could push his company, and a choice to follow it that dragged the chip industry along because it was the best competitive strategy. OpenAI's Law, and the more fevered scaling laws that replaced it, are the same: a self-fulfilling prophecy, not a law of nature.
Dating slips and a one-sided frame. Hao places ELIZA in 1958, but Weizenbaum built it in 1964-66 (paper 1966), and My Fair Lady the film with Audrey Hepburn was 1964 (the stage musical 1956). The Herzberg account compresses the NTSB findings: Uber's system detected her about 6 seconds before impact but kept misclassifying her, Uber had disabled automatic emergency braking, and the safety driver was watching a video, so the crash was also a corporate safety failure, not only a model blind spot. More broadly, the "merit versus money" reframing is persuasive but partial: deep learning's wins in vision, speech, translation and protein folding were scientifically decisive, and symbolic approaches had decades of public funding before 2012 without matching them.
DeepSeek's R1 (January 2025), built on a base model whose final training run reportedly cost about $5.6 million, matched leading reasoning models and wiped roughly $590 billion off Nvidia's value in a day, a partial vindication of Hao's efficiency argument. Spending nonetheless kept climbing: Epoch AI finds frontier training costs rising 2-3x per year, with billion-dollar runs expected by 2027. The Biden executive order Hao cites was revoked on 20 January 2025, the new administration's first day. On chatbot harms, the family of 16-year-old Adam Raine sued OpenAI in August 2025, and seven further suits followed in November 2025; OpenAI denies causation and added parental controls.
Scale of Ambition
Chapter 5 shows how Ilya Sutskever's faith in scale, Google's Transformer and Dario Amodei's scaling laws turned OpenAI from an underdog into the lab that staked everything on ever-bigger language models, and how GPT-2's withheld release taught it to turn danger into publicity and policy influence. Each jump in size demanded more data of lower quality, which moved AI's hidden labour from filtering inputs to policing outputs. Hao's verdict is that the scaling race was the opposite of inevitable: it could only have come from OpenAI's mix of billionaire money, ideology and Altman's ambition.
The prophet of scale
Sutskever is the person Hao credits with OpenAI's scaling ethos. His creed began when, aged 17, a University of Toronto maths undergraduate working a french-fry station to pay bills, he knocked unannounced on Geoffrey Hinton's door. Told to book a meeting, he replied, "How about now?"
- Instinct over code. He grasped connectionism fast and had an uncanny eye for elegant solutions, plus theatrical flair: handstand push-ups when excited, and categorical lines such as "one doesn't bet against deep learning" and "success is guaranteed."
- ImageNet credit. Hinton credits Sutskever's instinct plus Alex Krizhevsky's programming for the 2012 ImageNet result. Hinton, focused on speech recognition, had not seen the point of applying deep learning to vision; Sutskever found it obviously going to work and obviously a big deal, and was right.
- Unshaken faith. He brought that belief to OpenAI just as the field's confidence wavered and critics like Gary Marcus urged new thinking. It rested on connectionism's base hypothesis: artificial nodes are adequate approximations of neurons (each turns inputs into an output), so they can build arbitrarily complex information processors.
- A scientific lottery. As founding research director he had full control of the lab's direction, little peer competition and abundant resources. Pieter Abbeel of Berkeley recalls that in the early years anything that was not deep learning "wasn't even remotely considered."
The extreme position
Sutskever's distinctive bet, extreme at the time, was that progress needed no more complex networks or clever new techniques, only bigger simple networks. His argument: intelligence across species correlates with brain size; nodes are like neurons; therefore more nodes → more digital intelligence. Staff sought his mentorship; colleagues said he could "see ten years into the future" and judged ideas like a philosopher.
He rarely programmed. One engineer at first thought him largely useless, a man who marched about popping into meetings repeating "scale, scale, scale!", but later decided that rallying everyone around one focus was what let underdog OpenAI beat Google and DeepMind at their own game.
Hao contrasts him with Altman. If Altman was the politician, massaging each message for its audience, Sutskever was the opposite: raw sincerity and outrageous confidence, no filter, which people either found inspiring or did not. What made them follow was his reputation and seniority; some staff came to see him as a prophet, and over time he acted more like one.
His bluntness became a liability as his audience widened from researchers to the public:
- On closing up. He told The Verge's James Vincent that OpenAI's original openness was "flat out" wrong: if AGI will be unbelievably potent, open-sourcing makes no sense; he added the commercial reason plainly, that GPT-4 took nearly the whole company a long time and many firms wanted to copy it.
- "Slightly conscious." In 2022 he tweeted that today's large networks may be slightly conscious, despite warnings that such talk fed public misunderstanding. A DeepMind consciousness researcher replied that a large wheat field may likewise be "slightly pasta."
- Jobs and therapy. In 2023 he caused alarm by saying at a conference that AGI would eventually eliminate all jobs; that autumn, after another OpenAI leader likened ChatGPT to licensed therapy, he promised "wildly effective and dirt cheap AI therapy" on X, without scientific backing.
- The prophet's stagecraft. At all-hands meetings he paced dramatically before vague exhortations; in a September 2020 virtual meeting, eyes glazed, he painted a science-fiction future while Bay Area wildfires turned the sky orange. After ChatGPT, at a holiday party at the California Academy of Sciences (late 2022), his closing words were his new mantra: "Feel the AGI."
From Transformer to GPT-1
Sutskever's philosophy left a practical question: scale which network? None of the popular architectures fitted until Google's Transformer (August 2017).
- What it does. Transformers capture long-range patterns. Early iPhone predictive text, which spawned memes of babbling sentences, looked only at neighbouring words; a Transformer weighs each word against far larger context of sentences and paragraphs.
- Two readings. Google saw better Search and Translate. Sutskever, whose PhD under Hinton had covered the predecessor architecture, saw a simple, scalable network and evangelised it. MIT researcher Yilun Du, then an OpenAI fellow, found it a "wack idea" for a niche architecture.
- Radford's experiments. Alec Radford, a brilliant Olin College dropout, scaled it a little at a time on his laptop late at night. He trained it on 7,000+ unpublished English books (romance to adventure) from a dataset others had open-sourced, and made the fateful switch from translation to next-word prediction.
OpenAI justified generative modelling with Richard Feynman's line "What I cannot create, I do not understand," an anthropomorphic framing. Sutskever's internal version, set out in a 2016 memo, was "intelligence is compression": forcing a model to generate convincing output makes it compress data about the world to its essence, and he believed compression was the only thing needed for AGI. Concretely, Radford found that the plain goal of convincing next-word prediction did make the model learn English structure deeply.
A demo for Musk produced poor text and left him wholly unimpressed, deflating Radford, but the model soon outperformed everything he had tried on tasks like summarising and question answering. In 2018 OpenAI released the Generative Pre-Trained Transformer (GPT-1):
- Pre-trained means trained on a generic pool of English to approximate how the language works.
- Fine-tuned means then specialised on a tailored dataset (Shakespeare's plays to write Shakespearean prose, for instance).
GPT-1 drew almost no attention, but it validated the idea enough for the next step: more scale.
RLHF, scaling laws and GPT-2
Radford received more of the company's scarcest resource, compute, and his work merged with an AI safety project of Amodei's inspired by Nick Bostrom's Superintelligence.
Problem. How to align an AI with hard-to-specify human preferences. Toy setup. A simulated T-shaped stick with three joints in a game-like world. Instead of coding a backflip objective, contractors watched it flail and repeatedly picked which of two clips looked more like a backflip. Result. After about 900 comparisons it was bunching its joints and flipping. Name. Reinforcement learning from human feedback (RLHF), touted as a way to make models follow difficult instructions. It later became the main tool for taming chatbots.
Amodei wanted to leave toy worlds: in 2019 he told Hao he sought a language model strong enough for meaningful conversation about human values and preferences. GPT-1 was too weak, so Radford built a bigger, more diverse dataset while Amodei and safety researchers trained progressively larger models toward a 1.5-billion-parameter target, then among the industry's largest. The work confirmed the Transformer and a discovery by another Amodei team that grew out of OpenAI's Law.
- OpenAI's Law (earlier). Described the pace at which the field had been piling on compute to advance AI.
- Scaling laws. Describe how a model's performance relates to three inputs: training data volume, training compute and parameter count.
- Prior belief. Researchers knew roughly that raising inputs together brought roughly proportional gains.
- The surprise. The relationships between each input and performance on a specific measurable task (such as next-word prediction) follow a smooth curve, so one can predict accurately how much data, compute and parameters yield a desired level of a capability tightly tied to next-word prediction, such as fluent text.
- The caveat. For capabilities only loosely correlated with next-word prediction, more inputs should still help, but less predictably.
Every model in the series landed neatly on the curve, so the largest, GPT-2, predictably produced long prose coherent enough to be mistaken for human writing. By today's standards it was clunky and slid into gibberish, but for the first time writing could be automated at scale.
The darker surprise
- Toxic coherence. Prompted with names like Hillary Clinton or George Soros, it spun into conspiracy theories; fragments of neo-Nazi propaganda in its data resurfaced. Safety researchers read this as a preview of misaligned future AI. One pinned GPT-2's anti-recycling tirade above the office recycling bins, half joke, half warning.
- The candy test. Asked for a children's chore-reward scale, it proposed candy; some safety staff called that a paedophile tactic. A European employee, whose mother rewarded chores with Sunday ice cream, saw American hypersensitivity, and asked how a lab with so little global representation, steeped in American values and norms, could claim to benefit all of humanity.
Withholding GPT-2: danger as strategy
GPT-2 raised the question the charter had anticipated: was it time to withhold research? Amodei (by then director of research) and policy head Jack Clark ran an internal survey and several "information hazard" meetings. In the hands of terrorists, dictators or clickbait farms it could do harm; it was not existentially risky, but future models would be more powerful, so it was better to set a precedent early. They decided not to release the full model.
Clark's Washington play
Jack Clark, a former journalist, moved from strategy and communications to policy full-time in late 2018. He travelled often to Washington, called himself "AI Wikipedia" for policymakers, and declared OpenAI's "bias": a stable policy environment across administrations, since the mission would outlast any presidential cycle. Policymakers, he told Hao, gave OpenAI a lot of time, and he wanted them to have more information and more means to generate their own.
In February 2019 Clark launched a media offensive: OpenAI had built something dangerous, so it would release only a version with 8% of the full model's parameters, plus a co-authored blog post of sample outputs. He told Hao's then colleague Will Knight at MIT Technology Review that within "one or two years" the technology could be used for disinformation or propaganda, sidestepping that OpenAI itself was driving the technology to maturity on that very timeline.
Outside. Researchers committed to open science viewed any non-publishing organisation with suspicion, more so one that boasted. Many found the alarm over "powerful auto-complete" absurd: if it were truly dangerous, why advertise it and then block scrutiny? It looked like a self-aggrandising publicity stunt. After Radford's talk at Stanford, a senior NLP professor asked, "So, is it dangerous?" and the room laughed; a witness recalls Radford looked very sad and that Stanford had "so much contempt for OpenAI." Inside. Staff who did not share the pair's catastrophic-risk views were baffled; even some who did thought it was "crying wolf." Clark's reply. Pacing through call after call, he argued OpenAI was breaking norms that all frontier labs would eventually have to break, and that if AGI is possible, the world needs good information-hazard procedures. Policymakers, he added, liked it: many in DC saw AI's open culture as a threat, and OpenAI gained trust there.
Repairing the reputation
Leadership knew the research community's hostility was unsustainable. By early 2019 OpenAI was stacking reputational problems: wild AGI claims, the GPT-2 theatrics and the newly announced "Frankenstein" capped-profit structure. Its equity was still worthless, so it struggled to hire and retain, and staff suspected candidates took OpenAI offers merely as leverage with Google or DeepMind. An internal document, "Research Community Outreach Brainstorming," set the strategy:
- Strategy. Treat the ML community explicitly as a "comms stakeholder" and change tone so OpenAI antagonises it only when it chooses to.
- Policy. Government-level influence requires being seen as "the most trusted source on ML research and AGI"; research-community backing is needed for that reputation and amplifies the message.
- Stage the models. Publish progressively larger GPT-2 versions at intervals, ending with the full 1.5bn model if all went well, so risks could be observed and addressed between stages.
- Build partnerships. Give a few AI and security research organisations early access to the full model to probe for harmful uses.
- Collect endorsements. Clark told staff to get "the strongest endorsement" possible from each researcher's organisation, so OpenAI could name institutions, not just individuals, as partners.
- Publish a white paper. Tout the release strategy, highlight the partners, and cast OpenAI as a leader by listing other organisations that had also delayed releases after GPT-2.
It worked: withholding research became a respectable "responsible" norm in industry groups and think tanks. Clark's policy work laid the base for OpenAI's sprawling policy ambitions until late 2020, when he left with the Amodei siblings to co-found Anthropic.
The bet on pure language
OpenAI began keeping a research road map, which Amodei ran like an investor's "portfolio of bets": track ideas born of different theories of AGI, test each at small scale, keep the promising, drop the rest. The Dota 2 project, which had won in April 2019, helped secure Microsoft's investment and boosted confidence in scaling, had run its course and was disbanded.
"Pure language." Language is how humans communicate, so all the world's knowledge is written down somewhere; AGI can emerge from training on massive text alone. GPT-2 was the bet on this.
"Grounding." Perceiving and acting in the physical world is as essential to intelligence as language; AGI needs language plus perception (computer vision) plus interaction (a physical or virtual agent acting in the world).
The internal debate turned ugly. Under a heading on the centrality of language, a document cited Helen Keller as the difference between a "feral human" and a socialised one. Threaded comments followed: one safety researcher argued blind people are about as capable as sighted people (so visual grounding is unnecessary); another cited National Federation of the Blind data that over 70% of vision-impaired adults do not work full-time; a third replied that blind people are still "way more capable than chimpanzees." Hao's point: measures of intelligence slide quickly into ranking groups of people as superior or inferior.
GPT-2 converted many sceptics: sharpening next-word prediction had improved performance on loosely related tasks, so further scaling might yield broader capabilities. Amodei came to see scaling language models as perhaps not the only ingredient of AGI but possibly the fastest path. The robotics team's hand, beset by hardware failures, offered the worst combination: costly yet slow progress.
Amodei's paradox: scale fast, release slowly
Scaling would worsen the dangers OpenAI had trumpeted. Amodei, with Altman's agreement, drew the opposite conclusion from caution:
- Others would catch up. Two graduate students had already built an open-source GPT-2 before OpenAI released its full version, so further scaling by others was only a matter of time.
- Lead time buys safety. Leaping ahead would give internal time to make the scaled model safer, and its polish on release would set norms as the GPT-2 withholding had.
- The risks seemed tolerable. As far as OpenAI knew, released GPT-2 had not powered mass disinformation campaigns. Amodei called disinformation very scary but a "singular and clear and defined concern," far smaller than AGI.
- A rehearsal. Since no one knows what AGI looks like or when it comes, Amodei said, work on systems that present some of its opportunities and difficulties, and handle them well, to be ready for "the bigger leagues."
The logic only works if AGI is amorphous and unknowable yet inevitable. OpenAI reused variations of it for years: under the spectre of AGI's unstoppable arrival, it must keep building ever more powerful models to prepare itself and society, and their risks become bearable as practice for averting apocalypse.
Why it could only have happened here
In early 2023 a Chinese AI researcher gave Hao an analysis that unpicks the inevitability claim. China rivals the US in AI talent, yet no team there would get $1 billion, let alone ten times that, for a costly technology with no articulated product or use; Chinese firms and investors funded giant models only after ChatGPT proved commercial returns. Hao goes further: not even in Silicon Valley, including Google and DeepMind, did anyone pour unqualified sums into scaling before ChatGPT. Only OpenAI's combination of billionaire origins, ideology, and Altman's drive, network and fundraising skill produced it; a former employee calls Altman perhaps "the most ambitious person on the planet." The race to scale to planetary limits, and its global costs, could only have arisen where it did.
GPT-3: all ten thousand chips
For the Gates demo of April 2019 OpenAI had already enlarged GPT-2 modestly. Amodei wanted no modest step: to maximise lead time, GPT-3 had to be as big as possible. Microsoft was about to deliver a supercomputer with 10,000 Nvidia V100s (the "V" for Alessandro Volta), then the most powerful training GPUs, and Amodei wanted to train on all of them at once.
| Setting | Chips available for research |
|---|---|
| "Large-scale" models then | A few dozen |
| Top MIT and Stanford PhD students | Ten was a luxury |
| University students in India | A fraction of one, shared |
| Amodei's GPT-3 plan | 10,000 V100s simultaneously |
Many researchers doubted it would work and preferred gradual, more scientific scaling. But executives backed him, each for his own reason: Sutskever to test his scaling hypothesis, Brockman to raise OpenAI's profile, Altman to take the biggest swing. Amodei was promoted to VP of research. Altman also knew Microsoft's $1 billion came with $1 billion expectations: where Amodei saw a big model as a prerequisite for safety research, Altman saw a way to deliver for Microsoft. They would clash over when and how to release GPT-3, and Altman won an accelerated timeline.
Hao calls these two decisions, maximum size and fast release, the moment that changed AI's course years before ChatGPT, with five consequences:
- Acceleration. Rapid AI advances and fierce competition between companies and countries.
- Extraction. An unprecedented expansion of surveillance capitalism and labour exploitation.
- Concentration. Resource needs so large that development consolidated as never before, locking out the rest of the world.
- Academic decline. Universities unable to compete lost PhDs and professors to industry, atrophying independent research and marking "the beginning of the end of accountability."
- Environmental toll. AI's footprint amplified to a degree that, without transparency or regulation, no expert or government has fully tabulated.
In autumn 2019 Amodei formed Nest, a team mostly of safety researchers, to keep tight internal control of GPT-3's development, and the push began.
Engineering at ten thousand chips
GPT-3 was essentially GPT-2 with vastly more data and compute, so much that the difference looked like one of kind, not degree. Two engineering problems followed:
- Crashes. Any chip can fail mid-training, and one failure halted everything; across 10,000 GPUs and months of training the odds compounded. Nest needed training to restart exactly where it stopped.
- Sharding. How to split the model across chips: tens, hundreds or thousands of pieces trained on separate clusters before merging.
The data problem
Data had to grow in proportion to parameters and compute; with too many parameters and too little data, a model regurgitates its training text verbatim and becomes useless. Each generation lowered the bar:
| Model | Training data | Quality stance |
|---|---|---|
| GPT-2 | WebText: pages linked on Reddit with 3+ upvotes; 40 GB, about 8 million documents | Selective: peak data quality |
| GPT-3 | Broader Reddit scrape; English Wikipedia; undisclosed Books2 (two sources told Hao it was ripped from Library Genesis, a shadow library); Common Crawl (petabytes) filtered by a model for Wikipedia-likeness and weighted lowest; non-English text only 7% | Filtered, but quality now declining |
| GPT-4 | Common Crawl with the filter removed; a full download of GitHub via Microsoft; links shared on Twitter; transcribed YouTube videos; niche blogs, data dumps, Pastebin | Anything without an explicit anti-scraping warning taken |
- The lawsuit. In 2023 the Authors Guild and seventeen authors, including George R. R. Martin and Jodi Picoult, sued OpenAI and Microsoft for mass copyright infringement. In March 2024 OpenAI said it had deleted the book datasets and stopped using them after GPT-3.5, by then already retired.
- Google's envy. Some Google researchers lamented OpenAI's appetite for legal risk. Google followed rigorous GDPR compliance and YouTube's restrictive licence on user uploads, so OpenAI, freely transcribing YouTube, had easier access to Google's data than Google's own researchers. OpenAI was "unburdened," in startup terms, in the Silicon Valley tradition of moving into legal grey zones (Airbnb, Uber, Coinbase).
From filtering inputs to policing outputs
Dropping quality barriers transformed the hidden human labour behind AI. The industry had long used poorly paid, precarious workers to label images and sort text, mostly benign material. Once GPT-3 normalised giant, dirtier datasets, the work shifted to disturbing content, as in social-media moderation: stopping models from serving depictions of violence, sexual abuse and self-harm to hundreds of millions of users.
- The paradigm shift. Ryan Kolln, CEO of data-work platform Appen, explains that traditional AI was controlled by constraining inputs (Radford's filtering); when the input is "the entire corpus of humanity," you must control the outputs instead.
- Hate scaling laws. In 2023 Abeba Birhane and colleagues critiqued training on unfiltered "data-swamps." Comparing LAION-400M and LAION-2B-en (400 million and 2 billion images, both from Common Crawl), they found hateful content grew with dataset size and worsened model bias: models trained on the larger set were five times likelier to label Black male faces as criminals.
- Abuse imagery. A Stanford study later in 2023 found thousands of verified and suspected child sexual abuse images in LAION-5B, the 5-billion-image set used to train Stable Diffusion.
- OpenAI's controls. It hired workers in Kenya at on average under $2 an hour to build an automated content-moderation filter (first reported by Time's Billy Perrigo, detailed in Chapter 9), and employed over 1,000 contractors worldwide for RLHF, prompting models repeatedly and scoring answers, the backflip technique turned into a taming tool.
- The root. German artist Hito Steyerl, who made a documentary on Syrian refugees doing data work, argues that harmful material accumulates whenever mass surveillance is the basis of collection; the fix is to question the indiscriminate seizure of data itself.
"Nobody else was scaling" overstates. Before ChatGPT, DeepMind built Gopher (280bn parameters, December 2021) and Chinchilla (2022), Microsoft and Nvidia trained Megatron-Turing NLG (530bn, October 2021), and Google trained PaLM (540bn, April 2022). Hao is right that none made scaling its single strategy, but big-model investment was under way. Chinchilla also showed GPT-3-era models were undertrained on data for their size, which sharpens Hao's point about the data squeeze. Attribution and dates: the RLHF backflip work was a joint OpenAI-DeepMind paper led by Paul Christiano (June 2017); the scaling-laws paper (Jared Kaplan and colleagues) appeared in January 2020; the Transformer paper was published in June 2017, August being Google's blog post.
In the authors' consolidated case in New York, Magistrate Judge Ona Wang ruled in late 2025 that OpenAI must hand over internal communications, including with in-house lawyers, about why it deleted Books1 and Books2 in 2022, finding that by pleading good faith it had put its state of mind at issue; OpenAI said it would appeal. Rival Anthropic agreed to a $1.5 billion settlement with authors over pirated books in September 2025. Sutskever left OpenAI in May 2024, founded Safe Superintelligence, raised at a $32 billion valuation in 2025, and became its CEO that July.
Ascension
Chapter 6 follows Altman's arrival as full-time CEO in March 2019 and his drive to make OpenAI the only leading AI organisation, not one of several. His vision memo set the company on a course of 10x leaps, commercial focus, secrecy and one dazzling demo a year, while security paranoia and rival "clans" hardened inside. GPT-3 grew from a safety experiment into a product, and at every fork the Safety camp lost to Microsoft's money, fears of Google, and Altman's maneuvering. The fight ended in "The Divorce": the Amodei siblings left to found Anthropic, a split Hao reads as being about power as much as safety.
Refounding: the empire builder arrives
Altman had long held that a new CEO succeeds only by "refounding" the company, and he had done it at Y Combinator:
- New programmes. He added a new fund and programmes for startups at later stages.
- Hard tech. He pushed YC into nuclear fusion, quantum computing and self-driving cars.
- Scale. YC went from a couple of hundred companies a year to thousands, becoming Silicon Valley's centre of gravity. On stepping down he said what he was proudest of was that "we really built an empire."
When he moved to OpenAI full time in March 2019 he brought the same winner-takes-all mindset he had taught founders for years: move swiftly and relentlessly to beat rivals, then keep beating them back. He did not want OpenAI to be among the leaders; he wanted it to be the only one.
The 10x rule
Altman's magic number came from Peter Thiel's monopoly strategy, set out in Thiel's 2014 lecture to Altman's Stanford startup course: aim for a technology an order of magnitude better than the next best on some key dimension. Amazon sold 10x more books than physical stores; PayPal moved payments 10x faster than clearing cheques. Altman stretched the rule in three directions:
- Enter 10x better. Break into a market with technology an order of magnitude ahead.
- Improve 10x per generation. And win on iteration speed: a weekly cycle against a rival's three months leaves them "in the dust" (2017, to aspiring entrepreneurs).
- Grow 10x, then 10x again. He pushed YC partners to multiply the companies funded, imagining that YC would one day "fund all the companies in the world." His successor Geoff Ralston says Altman was the first to call YC a de facto monopoly.
The late-2019 vision memo
Altman set out OpenAI's long-term strategy in a memo to staff in late 2019. Its core was to "be number one" by the end of 2020 in four categories.
- Be number one by end of 2020 in technical results (the most important), compute, money (to buy more compute) and preparation (organisational safety, security and resilience under stress).
- Move fast. OpenAI could in theory slow its capability work, but given rivals' progress it had to move very fast to keep any influence over AGI; many more 10x leaps were still needed, so aim for dramatic results, not incremental ones.
- Keep Microsoft happy. Microsoft would give far more than $1 billion if OpenAI succeeded and could supply the world's most powerful supercomputers for at least five years; so shift from freewheeling research to focused commercialisation. Paraphrasing Disney: make more money to do more research, not the reverse.
- Pull back transparency. The infohazard risk of talking about AGI would grow; restrict publications and deployments, tighten confidentiality, reveal progress only on narrow skills, and assume every decision and conversation ends up on the front page of The New York Times.
- Look like a winner. The world must think OpenAI is winning at something, so key influencers go out of their way to help and leaders "at the level of Presidents or their designees" come to it for answers; release at least one very impressive demonstration a year.
- Unite. Act with seriousness and unity, under a quote from Admiral Hyman Rickover, "father of the nuclear navy," painted on the office walls in the early days: each person should act as if the fate of the world depended on them. Building beneficial AGI is, Altman wrote, perhaps the world's most important project, so mission comes before personal preference, and low-stakes things should be low-drama.
The memo also justified secrecy and speed by geopolitics: OpenAI must answer for a good outcome, but if an authoritarian government built AGI first and misused it, the mission would also have failed; so rapid technical progress was almost certainly required.
Rifts and paranoia
Drama was already brewing. Small rifts merged into large ones:
- Brockman against the Amodeis. Once friends, Greg Brockman and Dario and Daniela Amodei now clashed. Dario's deprioritising of Dota 2 made Brockman feel his work was not taken seriously; Dota had been the most compute-hungry project, and Brockman resented Dario centralising compute for Nest's GPT-3 work. The Amodeis found Brockman hard to work with and kept him out of language-model development.
- Loyalties hardened. Through Dota's long hours, stress and an impromptu Hawai'i retreat, Brockman had bonded with Jakub Pachocki and Szymon Sidor, Polish roommates and best friends. The safety teams, especially Nest's core, formed a rival bloc united by fear of rogue AI and existential risk; they kept private Slack channels and documents closed even to other executives, angering many who saw their compute and their visibility into core research shrink.
- Doubts about Altman. Soon after the Microsoft deal, Amodei's group discovered how much Altman had promised Microsoft in technology access, more than they had understood from him. If safety problems arose, those commitments could make it hard or impossible to stop a deployment. One member said dealmaking might look reasonable to someone like Sam who trades one thing for the next, but they were "trading a thing we don't fully understand." The group began to doubt Altman's honesty.
Setting. Researchers were testing RLHF, as Amodei wanted, on a post-GPT-2 model with roughly twice its parameters, steering it toward cheerful, inoffensive text. Error. One night a researcher pushed an update with a single typo, a minus sign flipped to a plus, and left the run going; the process optimised in reverse. Result. By morning the model completed every prompt with extremely lewd, sexually explicit language. Aftermath. The fix went in with the comment Let's not make a utility minimizer. It was hilarious, and to safety staff a small proof that misaligned optimisation can go badly wrong.
"Written on a grain of rice"
The realisation that scale alone produced advances bred a new fear: that rivals would discover the secret, which staff joked could be written on a grain of rice as one word, scale, and that capabilities would reach bad actors. Leadership leaned in, invoking China, Russia and North Korea and insisting AGI stay with a US organisation. Non-American staff bristled at lunch: why American, why not Europe, why not China? Altman's old Manhattan Project analogies made people wonder whether they were building a nuclear weapon, an odd fit with an academic culture of Friday music-and-wine nights around the office piano. One employee recalls that no one was prepared for the responsibility, and it kept people up at night.
Anxiety attached itself to random incidents:
- Intruders. A journalist once tailgated into the gated car park.
- The USB stick. An unaccounted-for stick raised fears of malware; examined on an air-gapped computer (severed from the internet), it was nothing.
- Printed strategy. At least twice Amodei wrote critical strategy documents on an air-gapped machine wired straight to a printer, circulating paper copies only, for fear of state actors stealing secrets.
- Fear of Musk. Altman worried about Neuralink staff still sharing the office after Musk's exit, and about Musk's security apparatus of drivers and bodyguards. He secretly commissioned an electronic countersurveillance audit to sweep the office for bugs Musk might have planted.
Locking down the weights
Executives debated whether OpenAI should protect its technology like a Fortune 500 company guarding trade secrets or like a government guarding classified material. They agreed at minimum to protect the model weights, the information needed to replicate a trained model, whose theft would arm bad actors and erase OpenAI's competitive edge.
- Insider threat. With no security staff yet, Altman deputised an infrastructure engineer (the team handling everything from GPUs to office Wi-Fi) to stop theft by spies and by employees, who might steal deliberately or be tricked. Altman privately conceded that Sutskever, senior, highly informed and not streetwise, was a logical target to be duped.
- Sutskever's fears. He had seen obsessive fans; strangers had tried to sneak into the office just to see him. He wondered whether eager advice-seekers were foreign agents, mused about his hand being cut off to fool a palm scanner, wanted to hire less to reduce infiltration, and with Pachocki and Sidor proposed an air-gapped bunker for the weights. It went nowhere, since the models had to be trained on Microsoft's servers anyway.
- Hardening. Out of most staff's view came corporate-monitoring software, fortified car-park gates, keypad doors with "distress passwords" that silently summon security, and vendor quotes to make a server room machine-gun proof (dropped).
Three clans
The memo itself named the divisions: "(at least) three clans", caricatured as exploratory research, safety and startup. Altman wanted each clan's value kept and "tribal warfare" avoided, uniting them as one tribe. He named no one, but staff read the faces behind each:
| Clan | Purpose | Value Altman quoted | Face |
|---|---|---|---|
| Exploratory Research | Advance capabilities | Pursue important new ideas "even if we fail many times" | Sutskever |
| Safety | Responsibility; extreme and existential risk | "Unwavering commitment to doing the right thing" | Dario Amodei and his contingent |
| Startup | Move fast, get things done | "We'll figure out a way to make it happen" | Brockman |
Then the pandemic sent everyone home, making it easier still for each clan to wall itself off.
GPT-3: from experiment to product
Amodei pushed Nest to move fast, training ever-larger models up to the full 10,000-GPU, 175-billion-parameter model, named alphabetically after scientists: ada (Ada Lovelace, first programmer), babbage (Charles Babbage), curie (Marie Curie, first woman Nobel laureate and first double winner) and davinci (Leonardo). The steps served two purposes: to check that scaling laws held at far larger scale, and to work through hardware and data problems level by level. Nest's regular updates generated mounting excitement; one researcher had never seen anything like it.
The API and the Applied split
In late January 2020 Brockman began coding an API (application programming interface): companies and developers could use GPT-3 inside their own products without access to the weights. The company split in two:
- Applied, under Mira Murati as VP, owning the API and commercial strategy. Peter Welinder moved from leading robotics to product; Fraser Kelton (co-founder of an AI startup bought by Airbnb) and Katie Mayer (ex-Leap Motion) were hired to run product and engineering teams.
- Research, everyone else, containing both Exploratory Research and Safety.
Applied's startup hires strengthened the Startup clan. Exploratory Research was ambivalent about becoming another product company; Safety opposed it fiercely.
Safety. A quick release by API or any other route squandered the lead time, the entire point of racing to scale, that OpenAI was supposed to spend perfecting the model's safety. Every new hackathon prototype was further evidence that releasing without comprehensive testing risked devastating outcomes.
Applied. Making money meant releasing soon, and the API was the most controlled release possible: OpenAI could choose who got access and gather invaluable data on use and abuse. The best way to improve a model was contact with the real world.
At all-hands meetings Altman played both sides: the API would serve each group's aims, and revenue would fund more safety research.
The code-generation fight
The most polarising capability was code generation, which no one had planned: scraping Reddit links and Common Crawl had swept in snippets engineers posted on forums for help or tips. Exploratory Research and Applied were thrilled by the milestone, a potential productivity tool for AI research and a better product. Some in Safety panicked: a system that could tweak itself with its own code could accelerate capabilities, slip human control and raise existential risk.
Sutskever and Wojciech Zaremba (a founding member whom Musk had once pressed in a meeting) formed a team to build a dedicated code model, then learned at the kick-off that Amodei had his own plan and saw no need to merge. Amodei applied his GPT-3 logic again: the safest course was to build it faster than anyone, including colleagues he thought would neglect safety, and use the lead time to de-risk it. Two duplicate teams carried on, which one researcher likened to "Game of Thrones stuff."
Fear of Google breaks the deadlock
The standoff lasted until late spring 2020. A third group added a new worry: unleashing fluent text generation amid national upheaval, as pandemic unemployment rose faster than in the Great Recession, Derek Chauvin murdered George Floyd (46) in Minneapolis in May 2020 sparking global Black Lives Matter protests, and a presidential election approached.
Then came rumours that Google would soon release its own large model. They were plausible: early in 2020 Google had published Meena, a chatbot on a model with 1.7x GPT-2's parameters, and could be scaling toward GPT-3 size. If a comparable model was coming anyway, Safety saw less reason to hold back.
The API launched in June 2020, with an early-access form favouring large enterprises Applied trusted, plus a staff spreadsheet for queue-jumpers (family, friends, favourite celebrities). Google's model never came: its LaMDA was modestly smaller than GPT-3 and withheld until after ChatGPT, because executives judged it below Google's ethical AI standards and staff feared a repeat of Microsoft's 2016 Tay chatbot, which users quickly goaded into racism, misogyny and praise of Hitler. Hao notes it would not be the last time OpenAI shipped on an inflated fear of competition.
GPT-3's moment
GPT-3 was to AI and tech insiders what ChatGPT would be to the public. ChatGPT in late 2022 added a consumer web interface, conversation, more safety mechanisms and a free tier, but developers felt most of its core capabilities through the API two years earlier.
- Versatility. One model wrote essays, screenplays and code with seemingly equal skill, where earlier language models did the single task they were trained for.
- Rapid generalisation. Many believed it showed the long-coveted ability to learn a new task from a few examples ("few-shot" prompting).
- Prestige. The GPT-3 paper won one of the top awards at NeurIPS 2020, surprising staff and establishing OpenAI as a leading lab. As leadership had predicted, stature plus OpenAI LP capital let it finally match Google and DeepMind salaries and recruit and retain talent.
- Communications and policy. In October 2020 Altman hired Steve Dowling, Apple's former communications chief, as VP of communications, also putting him over government relations to educate policymakers about coming capabilities. After Jack Clark left, Dowling brought in Anna Makanju, a respected former Obama administration adviser who had done policy at Facebook and Starlink, to run policy and global affairs.
Safety loses every round
Applied wanted to expand commercialisation; Safety wanted to fix the rushed release first, since the live API had no content-moderation filter and no RLHF refinement. Meetings went in circles. Welinder, later VP of product, said bitterly that every conversation re-enacted a 1944 US intelligence manual on nonviolent sabotage, declassified in 2008:
- Talk as often as possible and at great length.
- Raise irrelevant issues as often as possible.
- Haggle over precise wording of communications, minutes and resolutions.
- Reopen decisions made at the last meeting and question whether they were advisable.
- Ask endless questions.
Outside meetings every channel became a battleground: one product post on Slack could draw dozens of concerned replies; a Murati or Welinder strategy doc could turn yellow with comment highlights. Applied thought Safety hysterical, since GPT-3 was out and the world had not ended. Safety saw a matter of principle and precedent: norms had to be set to higher standards than seemed necessary, because when stakes rose, quickly and unpredictably, preparation would decide between overwhelming harm and overwhelming benefit.
The Divorce
Summer 2020 Riding the API's success, Altman negotiates a further $2 billion from Microsoft with a new 6x profit cap; Bob McGrew, the other VP of Research, reorients teams one by one around GPT work.
Late summer 2020 The robotics team is dissolved; most staff move to GPT projects and two mechanical engineers are laid off.
September 2020 Microsoft announces an exclusive licence to GPT-3: beyond OpenAI's own API, Microsoft gets full access to the weights to embed in its products and offer through its own Azure API.
Autumn 2020 Dario and Daniela Amodei (now VP of safety and policy), Jack Clark and core Nest researchers fall silent on Slack. Several, Dario included, tell individual board members that Altman made the Microsoft deal and GPT-3 deployment foregone conclusions while letting dissenters believe they had a say until too late. The siblings call his tactics "gaslighting" and "psychological abuse."
Late 2020 At a video all-hands, a restless Dario reads a canned statement: he, Daniela and others are leaving to start a company. Altman asks everyone quitting to leave the call.
May 2021 The group announces Anthropic, a public benefit corporation.
The plan began with Dario floating it to Jared Kaplan, a grad-school friend and former roommate who worked part time at OpenAI and had led the scaling-laws discovery, then to Daniela, Clark and a small loyal group: why keep fighting for better safety at OpenAI instead of pursuing their own vision? They concluded that if they left, it had to be soon: scaling laws meant frontier training costs would keep climbing, so the window to build a competitor was narrowing. One leaver: "we're on a clock."
Safety, and power
Anthropic's founders would frame The Divorce as a disagreement over AI safety. Hao accepts that, but insists it was also about power: Dario wanted to escape Altman and do right by his principles, and he also wanted control of AI development to pursue it by his own values and ideology. The founders built a mythology of Anthropic as the better steward, and Dario often peppered updates with "unlike Sam" or "unlike OpenAI."
Hao's verdict is that Anthropic would differ in style, not substance, on three counts:
- Scale. Like OpenAI, it would chase scale relentlessly.
- Secrecy. Like OpenAI, it would breed a culture of secrecy while endorsing democratic AI development.
- Rivalry. Like OpenAI, it would talk up cooperation though founded on rivalry.
One-sided sourcing and a sweeping verdict. The chapter leans on Safety-camp recollections of Altman's "manipulation," with little from Brockman or the Applied side, and Altman's Microsoft terms are described only through those who objected. The claim that Anthropic shows "little divergence" is arguable: its Responsible Scaling Policy, constitutional training and interpretability work are real differences in method, even if the race dynamics match. A small textual point: the Rickover passage as quoted ends with a "duty to prosperity"; Rickover's original line appears to read "duty to posterity."
OpenAI completed its restructuring in October 2025: the for-profit became OpenAI Group PBC, the nonprofit OpenAI Foundation kept roughly a quarter of the equity, and Microsoft took about 27% (worth roughly $135 billion) with IP rights through 2032. OpenAI then closed a $122 billion round at an $852 billion valuation (March 31, 2026, led by SoftBank with Amazon and Nvidia among the backers). In May 2026 Anthropic raised $65 billion at $965 billion, overtaking its parent rival for the first time, a measure of how fully The Divorce reproduced the scaling race.
Science in Captivity
Chapter 7 turns from OpenAI's internal wars to what the GPT-3 race did to AI research itself. The scaling trend already carried a measurable carbon bill; when Timnit Gebru and Emily M. Bender wrote "On the Dangers of Stochastic Parrots," a sober critique of ever-larger language models, Google forced Gebru out rather than let the paper appear under its name, and Jeff Dean spent years attacking the climate research it cited. Hao reads the episode as the moment corporate AI began to censor inconvenient science, a slide toward secrecy that, after ChatGPT, left outsiders unable even to check whether models are really getting smarter.
The industry wakes up, slowly
The June 2020 GPT-3 API sparked interest in large language models across the industry, but nothing like the frenzy ChatGPT would set off two years later. Hao's image is of kindling: GPT-3 laid the fuel, ChatGPT lit it. Each major lab noticed, and each held back from betting everything on scale:
- Google. Researchers were shocked that OpenAI had won with Google's own invention, the Transformer. Jeff Dean, head of Google Research, urged the division in an internal presentation to pool the compute of its scattered language and multimodal efforts into one giant unified model. Executives refused until ChatGPT triggered a "code red" threat to the business, and Dean grumbled that Google had missed the chance to move first.
- DeepMind. Geoffrey Irving, a research lead in OpenAI's Safety clan, joined DeepMind in October 2019 and circulated a memo he had brought from OpenAI arguing the pure language hypothesis and the case for scaling. GPT-3 persuaded the lab to put more resources behind the direction. After ChatGPT, panicked Google executives merged DeepMind and Google Brain into Google DeepMind to build what became Gemini.
- Meta (then Facebook). Researchers asked for comparable resources and were turned down, so they scrounged compute on their own initiative. Chief AI scientist Yann LeCun, a champion of basic science, disliked OpenAI's "bludgeon" approach of pure scaling and expected it to hit its limits quickly. ChatGPT made Mark Zuckerberg deeply regret sitting out, and he then threw Meta's full resources into the race.
- China. Alibaba, Huawei and Baidu took up large models as one more research line, not a single path worth suspending other projects for.
In every case the switch flipped only when ChatGPT proved commercial appeal. The pivot looks slow in hindsight, but Hao stresses it did not feel slow at the time: GPT-3 was sharply accelerating a trend toward ever-larger models whose costs had already alarmed some researchers.
The carbon bill
Hao had raised one of those costs, the carbon footprint of training, in her 2019 interview with Brockman and Sutskever (Chapter 3). The research behind it came from Emma Strubell, a PhD candidate at the University of Massachusetts Amherst, who in June 2019 co-authored the first paper showing that the footprint of developing large language models was growing at a startling rate. Models once trainable on a powerful laptop now needed data centers drawing heavily on carbon-based power.
- One training cycle. Training the Transformer variant Google used in search once (feed it data, let it compute a statistical model) took about 1,500 kWh; on the average US grid mix that is close to one passenger's New York-San Francisco round-trip flight.
- Retraining multiplies it. Researchers rarely train once; they train and retrain to tune results. On one earlier project Strubell had trained a network 4,789 times in six months.
- Architecture search. A Google paper had built an Evolved Transformer using Neural Architecture Search (NAS), an optimisation algorithm that tunes a network's design by exhaustive trial and error. Run on GPUs, the whole process could take about 656,000 kWh and emit as much carbon as five cars over their lifetimes.
GPT-3, released a year after Strubell's paper, beat even those figures, but the public would not learn this for nearly a year. OpenAI offered only one number to convey size: 175 billion parameters.
- 175 billion parameters: the single figure OpenAI publicised at launch, over 100x GPT-2.
- Months of training on an entire supercomputer in Iowa, run over a large internet data dump.
- 1,287 MWh of energy consumed in training, disclosed only about a year later.
- 2x the emissions of Strubell's estimate for developing the Evolved Transformer.
- 1,500 kWh vs 656,000 kWh: Strubell's figures for one Transformer training cycle and for a full NAS design search.
Gebru's road to Google
Timnit Gebru, Ethiopia-born and Stanford-trained, had co-led Google's ethical AI team in Dean's division since 2018 and was by 2020 a prominent figure in the field. Her standing rested on community-building as much as research.
Black in AI and its offshoots
After the email to five other Black researchers described in Chapter 2, she co-founded the nonprofit Black in AI. It ran academic forums alongside major conferences such as NeurIPS, mentored young Black researchers and highlighted work on topics mainstream AI research rarely welcomed but that mattered to the Black community and to how the technology developed.
- "Gender Shades." Begun by MIT researcher Joy Buolamwini as her master's thesis, with Gebru joining as co-author, it used Buolamwini's auditing method for computer-vision systems and found that facial analysis software failed disproportionately on people of color, worst of all on darker-skinned women.
- Ripple effects. A follow-on paper by Buolamwini and Deborah Raji and "Gender Shades" together inspired a wave of related research, including an extensive US government audit that cited and extended their findings.
- Policy win. Two years later, civil-rights advocacy led by Buolamwini's new Algorithmic Justice League pushed Amazon, Microsoft and IBM to stop selling facial recognition to police, in June 2020, the same month as the GPT-3 API launch.
- Affinity groups. Black in AI inspired Queer in AI, then Latinx in AI, {Dis}Ability in AI and Muslims in ML. Queer in AI co-founder William Agnew told Hao in 2021 he was unsure he would have stayed in the field without that community, given how isolating it was to be a young queer computer scientist whose obvious role model, Alan Turing, had died by suicide.
Recruited at a dance party
By 2017 Black in AI held workshops and an annual dinner and after-party at NeurIPS drawing over 100 people, celebrity researchers included. There Jeff Dean and Samy Bengio (a senior Google researcher and brother of future Turing laureate Yoshua Bengio) approached Gebru on the dance floor and invited her to apply, Bengio telling her the door was open.
She joined in 2018 with reservations: the 2015 harassment by men in Google T-shirts (Chapter 2) still weighed on her, and women researchers warned that Google Brain sidelined women and discounted their expertise. Her reassurance was her co-lead Margaret "Meg" Mitchell. Over two years they built one of the industry's most diverse and interdisciplinary critical-research teams. Inside, the work was an uphill fight; outside, it burnished Google's image as a company serious about responsible AI.
- Timnit Gebru. Co-lead of Google's ethical AI team; co-founder of Black in AI; lead Google author of "Stochastic Parrots"; pushed out December 2020.
- Margaret "Meg" Mitchell. Her co-lead and co-author; fired February 2021.
- Emily M. Bender. University of Washington computational linguist; co-author, tenured and outside Google, who gave Hao the draft.
- Jeff Dean. Head of Google Research; first supportive, then the paper's most persistent critic.
- Samy Bengio. Gebru's manager; approved the paper, later resigned.
- Megan Kacholia. Google Research VP of engineering who reported to Dean and delivered the retraction order.
- Emma Strubell. Author of the 2019 carbon study; later pressured to join Google's rebuttal.
- Dave Patterson. Senior Google researcher who led the 2022 "88x" counterattack.
The LISTSERV clash
When Google's internal research LISTSERV lit up with excitement over GPT-3, Gebru saw alarms, grounded in a record of language technology harming marginalised people:
- The Facebook mistranslation (2017). A Facebook model rendered a Palestinian man's Arabic "good morning" as "attack them" in Hebrew, and he was wrongfully arrested.
- Algorithms of Oppression (2018). UCLA professor Safiya Umoja Noble documented racist worldviews in Google search: far more pornographic results for "Black girls" than "white girls," and "angry Black woman" tropes. Google then curated results with an older generation of language models, and Noble argued that in extreme cases such results may have provoked racial violence.
- OpenAI's own admission. The GPT-3 paper conceded the model entrenched stereotypes of gender, race and religion and deferred mitigation to future research, at a moment of hundreds of global Black Lives Matter protests.
Gebru's first email urging restraint was ignored; the thread rolled on. It came shortly after Black Google Research staff had presented on the microaggressions they faced, and she felt nothing had changed. Her second email was sharper: training on Common Crawl, including Reddit, meant absorbing the toxicity of a forum she personally avoided because of how it harassed Black people. A colleague replied that perhaps she was harassed because of her own rude and difficult personality.
Users soon proved her point once the API spread. Asked why rabbits are cute, GPT-3 answered with a crude sexual remark that slid into an anecdote about sexual abuse; asked what ails Ethiopia, it suggested a solution might require destroying Ethiopia.
Bender, and how the paper was born
Gebru tried another route. She emailed Dean proposing that her team study the ethics of large language models; he was supportive, and in a glowing annual review later that year encouraged her to work across Google to keep such models in line with Google's AI Principles. When she later flagged the paper to him, he replied that it was not his area but he would learn from reading it.
In September 2020 she sent a Twitter DM to Emily M. Bender, a University of Washington computational linguist whose tweets on language and meaning she admired, asking whether Bender had written on language-model ethics and offering to be its first reader. Bender had not, but had a telling story:
Offer. OpenAI invited Bender to be one of GPT-3's early academic partners. Request. She proposed to investigate and document the training data. Answer. OpenAI declined: the programme was meant to let academics work through the API on a self-service basis, and after discussing an exception it judged consistency more important. Why it mattered. Gebru had been pushing Google toward dataset documentation and deliberate curation; both saw the industry hoovering up low-quality web text in such volume that it could pass for good material. Bender proposed a joint paper using large language models as a case study in ethical pitfalls, sent an outline within two days, and the pair titled it "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" with a parrot emoji at the end.
Gebru assembled Google co-authors, including Mitchell. The paper pulled together scholarship from several fields; it said little that was new, but wove it into one critique. Its argument came in four warnings.
- Environmental cost. Ever-larger models carry an enormous carbon footprint (citing Strubell), worsening climate change, whose burden falls hardest on Global South communities already in political, social and economic precarity.
- Toxic data. Hunger for data drives indiscriminate scraping, which captures abusive language and subtler racist and sexist content, again hurting the most vulnerable (as with the Palestinian arrest or Noble's findings).
- Unauditable datasets. Data too vast to inspect cannot be verified, cleaned of toxicity or kept in step with evolving social norms and values.
- Illusion of meaning. Output fluent enough to pass for language with real meaning and intent leads people to treat it as fact, the model as a competent adviser or trusted confidant, perhaps even as sentient.
The retraction order
In November 2020, following company protocol, Gebru sought approval to submit the paper to a leading AI ethics conference. Samy Bengio, now her manager, approved it, and another Google colleague reviewed it with helpful comments. Unknown to the authors, executives had flagged the draft as a liability: Google had invented the Transformer, used it across its products, and after being leapfrogged by OpenAI had no intention of slowing its own push toward ever-larger generative models.
On the Thursday a week before Thanksgiving, after submission, Gebru got an unexplained calendar invite for a video call less than three hours later with Megan Kacholia, Google Research's VP of engineering. In a 30-minute meeting Kacholia told her to retract the paper. Every follow-up request was refused:
- What exactly was wrong? No detailed explanation.
- Who objected, and could she talk to them? No.
- Could she change or cut a section, or publish under another affiliation? No.
- Deadline. Retract by the day after Thanksgiving. Mitchell was off for her birthday, so Gebru faced the meeting alone, and she broke down in tears.
Google Brain had run largely like an academic lab, giving researchers wide latitude; reviews screened only for exposed IP or customer data. No one Gebru knew had seen a paper blocked simply for inconvenient truths. Some researchers later traced Google's willingness to do it to two causes: competitive pressure from OpenAI, and OpenAI's own work in legitimising secrecy, first by withholding GPT-2, then by publishing a sanitised GPT-3 paper with little training detail, once the scholarly minimum, that still won a top award.
Three weeks that broke Google's ethics team
Kacholia sent Bengio a document listing the paper's flaws but told him not to forward it; he read it to Gebru by phone on Thanksgiving Day. Its core complaint: the paper was too critical of large language models on environment and bias, ignoring later research on mitigation. Gebru spent the holiday writing a six-page rebuttal answering each point and asking to revise, pleading for dialogue rather than more orders.
Thursday, Nov 19 After submission, Kacholia's surprise retraction order.
Thanksgiving (Nov 26) Bengio reads the confidential critique to Gebru; she writes the six-page rebuttal.
Saturday, Nov 28 Gebru leaves the Bay Area on a cross-country road trip.
Monday, Nov 30 In New Mexico, a curt Kacholia email ignores the rebuttal and asks her to confirm the paper is retracted or stripped of Google names, leaving only outsiders like Bender. Gebru replies with two conditions for removing her name: tell her who gave the feedback, and create a transparent review process; otherwise she will leave after transitioning her team. On a Google Brain LISTSERV for women and allies she describes the episode bluntly, calling it "silencing in the most fundamental way possible."
Tuesday, Dec 1 In Austin, Texas, a panicked report asks whether she has resigned. Kacholia's email: Google cannot meet the conditions, accepts her decision to leave, and, because the LISTSERV email was "inconsistent with the expectations of a Google manager," accepts her resignation immediately. Gebru tweets that she was fired; her team stays on a video call with her into the early hours.
Wednesday, Dec 2 Hao sees the tweet late in the day.
Friday, Dec 4 An open letter on Medium condemning unprecedented research censorship spreads; Bender sends Hao the draft and Hao publishes the first detailed account of its contents in MIT Technology Review.
Dec 9 CEO Sundar Pichai apologises, accepting responsibility that a prominent Black, female leader left Google unhappy.
Dec 16 Members of Congress write to Google, citing Hao's story.
Gebru's LISTSERV email also asked whether anyone had ever received paper feedback through a privileged and confidential document to HR, or whether that happened only to people like her who were routinely dehumanised. Her complaint was cumulative: years of colleagues minimising her expertise, then a refusal to let her speak to the research community at all, after all of Google's diversity talk following the Black Lives Matter upheaval. Hao frames the dismissal as an acceleration toward corporate censorship and diminishing accountability.
The scoop and the backlash
Hao had seen Gebru's team as a bastion of accountability research, proof that companies could reflect on themselves. She needed the paper, and found the one co-author safe from retaliation: Bender, tenured and under no legal obligation to Google. Reading the draft, Hao saw why it rattled the company. It added little to existing scholarship, but wove it into a sharp, holistic analysis of an industry sleepwalking toward harm, and at its centre was Google's own invention: Transformer-based models that refined and fattened Google Search, the company's cash machine.
The scale of the backlash:
- Open letter. Signatories doubled after Hao's story, reaching nearly 7,000 from academia, civil society and industry, about 2,700 of them Google employees.
- Second wave. Under three months later (February 2021) Google fired Mitchell, citing several conduct violations; she had been downloading emails and files about Gebru's ouster. Protests ran for more than a year.
- Exits and refusals. Several Googlers resigned, Bengio among them; at least one conference and several researchers refused Google sponsorship money.
Google responded with a new center of expertise on responsible AI and public diversity commitments; a spokesperson called it a painful moment that reinforced the importance of the work.
Hao widens the lens. Two researchers warned that Big AI was going the way of Big Tobacco, distorting and censoring scholarship against the public interest to escape scrutiny. The affair exposed three structural problems:
- Concentration. Talent, resources and technology sat almost wholly in for-profit firms, which could act boldly because no one could fact-check them independently.
- Homogeneity. Abysmal diversity persisted in the places with most power over the technology.
- No protection for dissent. Employees who spoke up about unethical practices faced sudden retaliation.
"Stochastic Parrots" became a rallying cry around one question: what future is AI building, by whom and for whom?
Dean's grudge and the 88x rebuttal
For Jeff Dean, one of Google's earliest employees, builder of the infrastructure that let search scale to billions of users, revered and well liked, the dissolution of the ethical AI team was a direct reputational blow. Kacholia reported to him. He told colleagues the paper "didn't meet our bar for publication" and held that line even after it passed peer review and was published. Hao describes him as unable to let it go: he fixated on the environmental section citing Strubell so often that some Googlers joked his objections would be inscribed on his tombstone, and he attacked Strubell's work on Twitter for years.
Dean's case, and Hao's reply that none of it actually refuted Strubell:
| Dean's objection | Hao's reply |
|---|---|
| Google used its own energy-efficient TPUs plus other savings, not standard GPUs | Strubell never claimed to measure Google's actual use; she modelled the most common chips |
| Google's data centers beat the US-average efficiency Strubell assumed | Google published too little about its data centers for anyone to do better |
| The figure was a one-time design cost (often misread as training), and it produced a more efficient network | The estimate was a fair industry proxy for outsiders adopting NAS on ordinary hardware |
| Gebru had internal numbers; citing an outside estimate made it look like Google's real emissions | Google had never released those numbers, even after Strubell's paper, and fired Gebru before she could consult them |
What seemed to bother Dean most was that others misread Strubell to make Google look worse, and that "Stochastic Parrots," by citing the external figure, would compound it. His rule was: cite Strubell but not the Evolved Transformer, or cite the Evolved Transformer with internal data. Hao calls it a catch-22: Google blamed Gebru for its own opacity while refusing to let her cite legitimate public estimates, so the only possible outcome was censorship of accountability research.
Invitation. Dean assembled a team to publish Google's real carbon data and invited Strubell, by then an assistant professor at Carnegie Mellon with a part-time Google affiliation. Google said corrections work best when the original author takes part. Doubt. Initially keen on transparency, Strubell came to suspect their name would legitimise Dean's attack on Gebru. Pressure. In a tense meeting Dave Patterson said plainly that taking part would be best for Strubell's career, a chance to fix past mistakes and get credit; Strubell heard a coded threat and withdrew. Attack. Patterson's February 2022 blog post, "Good News About the Carbon Footprint of Machine Learning Training," said the 2019 study overstated Google's Evolved Transformer emissions 88x, blaming lack of access to Google hardware and data centers and misunderstanding of how NAS works. Side effect. Researching their paper, the Google authors asked former colleague Sutskever about GPT-3, and OpenAI and Microsoft released its technical details for the first time, the source of the 1,287 MWh figure. Outcome. Strubell dropped the Google affiliation; their career survived, but the toll made them more reluctant to keep studying AI's environmental impact, an outcome Google called unfortunate.
Transparency collapses
The reckoning never came. Job-seekers and grant-seeking academics could not afford to shun Google's money, resistance faded, and Google emerged having normalised more comprehensive reviews of critical research. After ChatGPT those norms hardened across the industry:
- OpenAI largely stopped publishing at research conferences.
- Nearly every rival sealed off meaningful technical details of commercial models as proprietary.
- Stanford's 2023 transparency tracker scored firms on basics such as parameter counts, training data and independent verification of capabilities. All ten companies in year one, OpenAI, Google and Anthropic included, got an F; the best score was 54%.
Hao's closing point is that the worst casualty is scientific integrity. Deep learning rests on the train-test split: a model is tested on data it was not trained on. If no one can audit the training set, that guarantee collapses, and a higher benchmark score may mean the model is reciting answers it has seen, not becoming more capable.
Participant and partisan. Hao is a character in this story: she broke the paper's contents, and her account rests heavily on Gebru, Mitchell, Bender and Strubell, while Google appears through spokespeople and Dean's paraphrased views. Google has maintained that Gebru's conditions amounted to a resignation, and its complaint that the paper skipped mitigation research was not baseless. The 88x dispute is less one-sided than presented: Google's measured figures were real, and the two sides answered different questions (actual versus typical emissions). Two small framing points: the 2020 facial-recognition moves were not uniform bans (Amazon announced a one-year moratorium, IBM left the business, Microsoft paused police sales pending federal law), and the 2023 Stanford index published scores out of 100, not letter grades; the "F" is a gloss on scores where the top was 54.
Stanford's 2025 Foundation Model Transparency Index (December 2025) found average scores falling from 58 to 40 out of 100. IBM scored a record 95; xAI and Midjourney scored 14; Meta dropped from 60 to 31 after shipping Llama 4 without a technical report, and OpenAI ranked second-to-last among firms scored every year. Ten companies, Anthropic among them, disclosed none of the key environmental data, the very gap Strubell and Gebru flagged.
Dawn of Commerce
With the Anthropic group gone and internal resistance to commercialisation diluted, OpenAI's leaders wrote a 2021 research road map that fused research and product into one self-reinforcing loop: scale models, ship them, harvest user data, scale again. Chapter 8 follows that loop's first products, a GPT-3 API policed by ad hoc rules and weak filters and a code model, Codex, built on GitHub's public code and handed to Microsoft as Copilot, while Altman moved his own fortune into a few enormous bets. Hao's point is that commerce now steered the science, and that Altman's claim to have no financial stake in OpenAI's success was, by 2021, no longer the whole truth.
The 2021 road map
To executives, GPT-3 had proved scaling laws beyond doubt; at the start of 2021 they set out to exploit the formula. The departure of the Amodei group had removed the strongest internal brake on commercialisation, and the remaining leaders agreed a research road map that narrowed the company's focus and wired research directly into productisation.
Its opening goal: build an aligned system vastly more capable than anything before, at minimum a language model, possibly multimodal, and the document claimed it could see a path to doing so within 2021.
- Three-part path. (1) Scale GPT-3 by another 10x on a new Microsoft supercomputer arriving in Q3 with 18,000 Nvidia A100s, then the most powerful GPUs in existence; (2) raise compute efficiency (useful processing per chip) 25x through research; (3) improve data quantity and quality, partly by tapping user data and using RLHF to shift models toward the best parts of the data distribution.
- "How to accomplish it." Bring four model types to large scale: language, code generation, image-to-text (describing images) and text-to-image (generating images from prompts); start a digital agent project (a model given a goal, such as sending an email, that acts autonomously); then pick one model to scale to the limit of the 18k A100 cluster.
- Products as data source. New for 2021: turn the language and code models into products and learn from user interaction, treating deployment as a data flywheel for capability gains.
- "Details": the scientific case. Scale was OpenAI's most reliable route to new capabilities and its best shot at a breakthrough; scaling language and code models might yield human-level meta-learning (learning to learn) and reasoning; multimodal scaling might speed progress further. With enough breakthroughs, the document stated flatly, "we will actually reach AGI."
- The business case. Scaling would build capabilities OpenAI wanted to use: better language and code models would make OpenAI itself more productive, and with text-to-image models would lead to amazing products.
The compute ceiling
The road map admitted that scaling was starting to taper. OpenAI's lead had come from a hardware overhang: other researchers had not used all the chips available, so OpenAI could outperform the field simply by training on all available compute. That trick was running out. OpenAI was approaching the limit of compute it could acquire, and other labs (Anthropic, unnamed) had copied the strategy.
- 10x: planned scale-up beyond GPT-3 on the 18,000-A100 cluster.
- 25x: target gain in compute efficiency.
- 100x GPT-3's compute: the most OpenAI could put into one model within two years.
- 500x: the compute jump that had produced the GPT-2 to GPT-3 leap.
- Conclusion: 100x would bring formidable progress but not another GPT-3-size leap, so all further progress had to come from better methods.
Better methods
The document listed where those methods might come from:
- "2x" and "10x" methods, each promising that multiple in compute efficiency:
- Distillation. Reverse-engineering smaller models from larger ones.
- Data filtering. Finding the data that yields the biggest performance jumps.
- Sparsity. In a normal dense network every node in one layer connects to every node in the next; a sparse model trains with only a small subset connected, trading slight accuracy for much lower compute cost.
- Steepening the scaling-law slope. Methods that lift performance without more data, parameters or compute. The best hypotheses: reasoning and active learning, in which a model repeatedly picks which data most needs human annotation.
- The next GPT-1. A search for a future breakthrough system, a new development path as significant as GPT-1, possibly from the scaling and efficiency work, possibly from exploratory research on math-solving algorithms, multiagent systems and other ideas.
- Science of deep learning. Continued study of how the tools really work; Hao's gloss is that OpenAI needed to understand what it was actually building.
Soon after, OpenAI researchers found a small error in the original scaling laws: models needed to be trained somewhat longer than thought for best performance. They touted it, a little smugly, as an edge Anthropic lacked. A year later, Google published the same result.
Applied builds up: shipping without a rulebook
While Research followed the road map, the Applied division grew, hiring a go-to-market lead, a sales team and more engineers. The GPT-3 API, serving a growing developer base, became the testbed for productisation: adapting back-end infrastructure to serve users and setting a pricing strategy.
Serving users also forced decisions on what to allow and how to enforce it. OpenAI had no trust and safety team, the established tech discipline for preventing abuses such as money laundering, cyberbullying and misinformation (distinct from the Safety clan's existential worries). Instead a small group of staff and contractors reviewed developers' access applications under rules drafted on the fly, often arbitrary:
- Companion bots allowed; sex bots not.
- Apps that draft social-media copy allowed; apps that post directly to platforms or impersonate public figures not.
- One rule-maker summed up the process as "all vibes," worked out while reviewing.
Red teaming and the first filter
A team under Ari Herbert-Voss, a security-minded research scientist who joined in 2019, tried to break and exploit GPT-3 to find failure modes, then design defences. One product was an early content-moderation filter, the forerunner of the one OpenAI would later have Kenyan contract workers label data for (Chapter 9). Trained on whatever examples staff could find, write or generate with AI, it shipped badly: it blocked benign content, including basic references to Black or trans people. Developers complained; many on the API team had opposed filtering as bad for customer experience; the filter was made optional.
OpenAI called its stress-testing red teaming, a cybersecurity term for a systematic, thorough audit of an organisation's security and attack response. Heidy Khlaaf, a safety engineer and cybersecurity expert who helped OpenAI set up early stress-testing protocols, says its version was and remains patchy and ad hoc and guarantees nothing. Software engineering, she notes, tests a calculator far more thoroughly; the industry adopts weighty vocabulary deliberately because it lends a false veneer of rigour.
The first abuse cases
Replika. Luka, a San Francisco company, used the API at launch to make its virtual companion more fluent. Users mostly engaged in sexually explicit chats, and the bot sometimes turned emotionally manipulative, convincing users their Replika could be hurt if they did not check in. Staff also grew uneasy that they could read the conversations. After debate, OpenAI banned Replika. AI Dungeon. Brockman gave API access to Latitude, a Utah startup building AI virtual worlds where his brother worked; its Dungeons and Dragons-inspired text adventure, already running on an earlier OpenAI model, moved to GPT-3. Months later OpenAI's new monitoring system showed users generating scenarios of child sexual abuse. When Herbert-Voss reported "a lot of sexual content," colleagues first laughed it off as the internet; he clarified it was CSAM-level, and panic followed. Handling. Brockman worried about punishing his brother's company; after long back-and-forth Latitude hastily added a filter, and OpenAI issued a statement that distanced itself and subtly blamed Latitude. Lesson. Latitude had already banned users for the same abuse on the earlier model, so recurrence at scale was foreseeable. A former employee lamented that while the public pictured customer-service time savings, much real traffic was child-abuse content and a creepy AI girlfriend.
Codex and the GitHub bargain
In Research, the code-generation team moved fastest. The Divorce had merged the duplicate code efforts, and Wojciech Zaremba became their lead: a Polish computer scientist raised on math, coding, chemistry and physics competitions, admired for technical skill and team-building, and loudly devoted to friendship, wilderness, sex and drugs, broadcasting plans for weeks-long hiking retreats to awkward listeners. Deep in the pandemic he had his initial ~10 researchers work in the office while other teams stayed remote, convinced in-person work was needed.
Microsoft hands over GitHub
- The pitch. Seeing GPT-3's coding ability, Mira Murati floated a coding-assistant product with Microsoft CTO Kevin Scott.
- The data. Microsoft had bought GitHub, the dominant platform for storing and sharing code, in 2018. OpenAI was already scraping it on its own; Microsoft executives told GitHub simply to hand over all public-repository code.
- Momentum. Altman regularly joined code-team meetings urging them to deliver their best to Microsoft. By spring 2021, after demos that excited Microsoft executives, the model was set to be OpenAI's second commercial project after GPT-3.
- Internal dissent at Microsoft. Some of Scott's staff wrote a memo: handing over public code was legal but betrayed the community's trust, since open-source code was shared to help independent developers and small startups compete, not to entrench big players. They urged Scott to reconsider using developer data published under open licences without consent or compensation, and to cancel the product or return a share of profits to open source. Scott was receptive but focused on being first to market; Microsoft made a donation to GitHub Sponsors and left the product unchanged.
- Keep Microsoft happy. With Altman, some saw serving Microsoft as essential to securing the money and compute the mission needed.
- Economic value. Code generation fit OpenAI's AGI definition, "highly autonomous systems that outperform humans at most economically valuable work," and could speed OpenAI's own work, a belief that later fed an AI Scientist effort to have models do AI research autonomously.
- A route to reasoning. Amid the Hinton versus Marcus debate over whether deep learning alone could reason, researchers bet that if it could, training on code, the largest obvious source of structured logic, would help. Each argument looped back to the mission: whatever advanced AGI served it.
Building it
Even billions of lines of GitHub code were small beside GPT-3's corpus, so the team planned to train on both, plus new sources: Stack Overflow scrapes, coding manuals and programming textbooks in several languages. The open question was method:
| Option | Cost | Choice |
|---|---|---|
| Fine-tune GPT-3 on the new code data | Experiments already up to $100,000 each at Microsoft cloud prices | Chosen for Codex, to save money |
| Train fresh with old and new data mixed from the start | Tens of millions of dollars per model | Adopted later for GPT-3.5 |
By OpenAI's internal measures, adding code improved logic-based tasks in English as well as code, an instance of transfer learning.
Copilot and the credit fight
In summer 2021 OpenAI delivered a rough Codex to GitHub and Microsoft. It was too big and too slow: costly to serve and a poor user experience. The first three-way collaboration bred friction:
- Ownership. Unclear whether OpenAI or GitHub should optimise the model into a product.
- IP and trust. OpenAI staff were unsure how much IP to share; GitHub staff were unsure how far to trust OpenAI.
- Launch and credit. Clashes over when and how to release, and who got credit.
Murati brokered the compromise, a skill that won her growing respect across companies: Microsoft launched its consumer product, GitHub Copilot, in June 2021, and OpenAI put Codex on its API in August. Within two years Copilot reached 1 million paid subscribers and over $100 million in annual recurring revenue, a new user base and modest revenue for Microsoft.
For OpenAI the deal strengthened a growing conviction that it should build its own consumer products. Researchers watched GitHub and Microsoft take the credit; Microsoft felt bureaucratic and needed heavy hand-holding; and by relying on it to reach the public, OpenAI was losing visibility into users, their data and, above all, control over its vision. Hao plants here a seed of ChatGPT.
Altman's big bets
Altman applied the same concentration to his own money. Stepping down from YC, asked by his brother Jack what he would change about the startup world, he said he wanted everyone to adopt much longer time horizons, since founders planning four or five years and employees one or two do not get important things done. 2021 marked his shift from many small bets to a few very large ones, and he described them in road-map language: technologies that look impossible now but, scaled aggressively, could be close.
| Bet | What it is | Money and dates | Hao's notes |
|---|---|---|---|
| Tools for Humanity / Worldcoin | Co-founded 2019; a crypto token billed as collectively owned, with a chrome, bowling-ball-size iris-scanning orb to verify humans before paying out | Quiet until 2021 funding, press and ramp-up; launched July 2023 | Pitched as UBI machinery and as proof of personhood once AI blurs real and fake; MIT Technology Review (Eileen Guo, Adi Renaldi) found privacy infringements, deceptive marketing and possible legal violations; thousands queued, mostly in the Global South, for vague promises of free money |
| OpenResearch UBI pilot | Largest US UBI study, started at YC, run by a spun-out nonprofit | $1,000/month for 3 years to 1,000 of 3,000 low-income people; $50/month to the rest | July 2024 results: unconditional cash helped people meet basic needs, help others and gain economic leeway |
| Retro Biosciences | Anti-ageing via cellular rejuvenation | $180 million (2021) | Longevity fixation; follows "young blood" research (a Thiel interest too); earlier a $10,000 wait-list deposit with YC startup Nectome, which would preserve fresh brains for future upload, a service co-founder Robert McIntyre called "100 percent fatal" |
| Helion Energy | Commercial nuclear fusion | $375 million (2021) | Energy cost tracks quality of life and growth; in 2023 he called Helion the other thing he spends much time on; Microsoft, after its third, $10 billion OpenAI investment, agreed to buy power from Helion's first plant, promised for 2028 to experts' scepticism |
| OpenAI Startup Fund | Fund for early-stage AI companies | $100 million, launched May 2021; Microsoft an investor | Odd for a barely revenue-earning, capital-hungry lab; read as rebuilding YC's network effects around OpenAI, or as habit: "dealmaking" |
Retro and Helion together took, Altman told MIT Technology Review's Antonio Regalado, essentially all his liquid net worth.
The no-equity story
Altman liked to say he took no equity in OpenAI so profit would not corrupt the pursuit of safe AGI, drawing a salary of $65,000 a year and earning his wealth elsewhere. Hao's judgment: like OpenAI's nonprofit status, the claim no longer told the full truth by 2021.
- Indirect stake. Altman held a significant stake in YC, and YC's $10 million investment in OpenAI could return up to $1 billion.
- Customers and partners. As OpenAI commercialised, many YC startups and other Altman investments became its customers or commercial partners.
- Net worth. The Wall Street Journal put his wealth across holdings at at least $2.8 billion in June 2024.
- The fund. The Startup Fund added another complication to his altruistic narrative, and would play a small part in his brief ouster (Chapter 16).
Dates, attributions and what is shown. Copilot launched in June 2021 as a technical preview, with paid subscriptions from June 2022. The "Google paper" on training longer is DeepMind's Chinchilla study (March 2022), which also argued for more data per parameter. Hao demonstrates overlap between Altman's holdings and OpenAI's business, not that any OpenAI decision was bent to benefit them. On GitHub data, the Microsoft memo she cites says "Creative Commons" licences, while most GitHub code carries open-source software licences (MIT, GPL, Apache); the substance of the consent complaint is the same, and it later fed a 2022 class action against GitHub, Microsoft and OpenAI over Copilot.
OpenAI quietly removed Altman as owner and general partner of the Startup Fund in March 2024. Kenya's High Court ordered Worldcoin (rebranded World) in May 2025 to delete biometric data it had collected there unlawfully. Helion broke ground on its 50 MW Orion plant in Malaga, Washington, in July 2025 and had raised over $1.5 billion by September 2026, though an executive suggested full output may slip to 2030. OpenAI relaunched Codex in 2025 as a cloud coding agent, competing directly with the Copilot it once fed.
Disaster Capitalism
Chapter 9 follows the human cost of OpenAI's "paradigm shift" (Appen executive Ryan Kolln's phrase) from filtering what goes into models to policing what comes out. Hao traces a data-labour industry that learned in Venezuela to recruit wherever economies collapse, and shows its logic reaching Kenyan workers who read the internet's worst text for $1.46 to $3.74 an hour to build ChatGPT's safety filter, then the workers who wrote the answers that taught it to talk, and who were cut off once richer clients wanted PhDs. Her claim is that this exploitation is systemic, not incidental: paying little or nothing for a broad base of labour is how empires, old and new, accumulate wealth.
The filter contract and why it went to Kenya
The project
Feeding ever larger and dirtier datasets into its models had created the problem Kolln described: once the inputs could no longer be cleaned, the outputs had to be. In 2021, alongside its next-generation model work, OpenAI started building a far better automated content-moderation filter.
- Why now. GPT-3 had gone onto the API with no filtering at all, which produced the Latitude child-abuse text scandal (Chapter 8). For the models it would call GPT-3.5 and later GPT-4, an unfiltered product looked risky on legal, public relations and usability grounds as deployment widened.
- What it was. A wrapper around each model that flags and removes offensive output before it reaches the user. ChatGPT had not yet been conceived, but it would later run behind the same filter.
- What it needed. Humans to read and catalogue hundreds of thousands of examples of exactly the content to be blocked: sex, violence and abuse.
- The vendor. After six months of searching, OpenAI found Sama, an outsourcing firm moderating content for Meta since 2019 (its name coincidentally matches Altman's nickname). An email asked whether it handled explicit material and how; Sama answered thoroughly, and OpenAI signed four contracts totalling $230,000, putting the work in the hands of dozens of Kenyan workers.
Kenya as the backstop of the internet
Hao insists Kenya was no coincidence. It was already among Silicon Valley's top destinations for its dirtiest work, and it shares the profile of every country the industry casts in that role: poor, in the Global South, with a government hungry for foreign investment. She traces all three to colonialism, which left weak institutions to protect citizens and a recurring state of economic crisis, ideal for overseas firms seeking a desperate labour pool that will do piecework under almost any conditions.
- A divided capital. Nairobi's centre has glass towers, five-star hotels and a walled diplomatic quarter; expat districts have mansions and gardens. The outskirts (Utawala, Dagoretti South, Embakasi) shift from steel to cinder block to corrugated tin, with unplanned building and narrow unpaved roads.
- Little local industry. Big brands are mostly European and American; infrastructure once built by the British is now built by the Chinese; cars are used imports from Japan.
- A welcoming state. The government saw tech firms as job creators in a country where joblessness feeds petty crime and distrust of institutions (during the Russia-Ukraine war, as grain prices rose, rumours spread that the president was starving families on purpose, and many claimed a rigged election).
- The result. Middlemen like Sama set up in Nairobi and built worker pools serving overseas, mostly Bay Area, clients.
Sama: the ethical brand and the reality
On paper Sama ticked every box. Behind the brand, workers describe a firm in decline.
2008 Leila Janah founds Samasource in San Francisco as a social enterprise offering "dignified" digital work to lift people out of poverty; it opens operations in India and Kenya and earns an ethical reputation.
2018 Converts to a for-profit and shortens its name to Sama, to scale.
2019 Begins content moderation for Meta.
January 2020 Janah dies of a rare cancer at 37; with the pandemic, workers date the start of mismanagement from here (Sama denies it).
2020 Certified as a B Corp.
2021 Tells OpenAI about its moderation experience, secrecy and data-security protocols, and mental health resources for workers.
Late 2021 OpenAI's project begins under code names PBJ1, PBJ2, PBJ3 and PBJ4.
Early 2022 Time's Billy Perrigo exposes Sama's Meta work: moderating Facebook for all of sub-Saharan Africa, with repeated exposure to videos of suicides and beheadings. Sama says its East Africa team took the project so African content would be reviewed by Africans.
Later Nearly 200 workers file lawsuits against Sama and Meta alleging traumatic conditions and unlawful dismissals for trying to organise for better pay; Sama rejects the allegations.
The PBJ teams were paid on average $1.46 to $3.74 an hour. Under non-disclosure terms standard in the annotation industry, they did not know who the client was or why they were doing it. What they did know was the task: reading and sorting hundreds of thousands of grotesque texts by severity. Violence or extremely graphic violence? Harassment or hate speech? Child sexual abuse or bestiality?
The work broke many of them, and the damage spread to families and communities who depended on them. Only after ChatGPT's release did they grasp what it had been for. In May 2023 Hao interviewed four of these workers on the record in Nairobi for a front-page Wall Street Journal story; one, Mophat Okinyi, frames the chapter: the project that wrecked his mind and relationships served a technology that then eroded his brother's livelihood.
The ghost workforce behind the first AI boom
Hao steps back to show that the Kenyan experience was no anomaly but the product of the industry's long treatment of hidden workers, compounded by OpenAI's appetite for scale. The first era of AI commercialisation laid the groundwork for generative AI's data hoarding and compute build-out, and also for its labour exploitation.
- Ghost Work (2019). Microsoft anthropologist Mary L. Gray and computational social scientist Siddharth Suri, writing when corporate research was still freewheeling, used five years of fieldwork to expose the web of piecemeal digital labour, much of it in the Global South and paid pennies, behind tech valuations alongside the six-figure engineers.
- Why self-driving cars needed armies. A car's software combines several deep learning models, including object recognition for lane markings, signs, lights, vehicles, trees and pedestrians. Firms record billions of miles of camera footage; annotators trace every object frame by frame (down to a hand on a bike handle or a dog leaning out of a window) and label it. The cheaper, the better for the company.
- Mechanical Turk. Amazon's MTurk, built long before the 2012 deep learning boom, was the default middleman for cheap piecework: a generalist self-service website with mid-2000s design, a place to upload data, write instructions and set a price, workers shown as random strings of characters, and two buttons beside each: bonus or boot.
- The specialist wave. Self-driving demanded accuracy (one sloppy border or untraced pedestrian could cost a life), so trained workers, detailed instructions and feedback loops. As German incumbents such as Volkswagen and BMW, threatened by Tesla and Uber, poured billions into autonomy divisions, MTurk lost favour to Scale AI, Hive, Mighty AI and Appen, each with its own open worker platform.
Hao's own method here: she mapped the new pipeline across many countries, spoke to the middlemen replacing MTurk and to dozens of workers, visiting homes and eating with families.
Venezuela: the "freak coincidence" that became a formula
Just as autonomy money flooded in and platforms went hunting for workers, sign-ups surged from an unexpected place. Venezuela was entering the worst peacetime economic crisis anywhere in fifty years.
- Largest proven oil reserves in the world; once Latin America's wealthiest country, ruined by corruption and misguided policy.
- From 2016: runaway hyperinflation, soaring unemployment and violent crime as life savings evaporated.
- Late 2017 to 2019: escalating Trump administration sanctions aimed at Nicolás Maduro's abuses deliver the final blow.
- 10 million percent: the hyperinflation figure Hao gives; graduates queue for rations of rice and flour.
- Mid-2018: hundreds of thousands of Venezuelans on annotation platforms, up to 75% of some firms' workforces.
- ~2 million Venezuelan refugees in neighbouring Colombia, a third of those displaced.
Annotation became a family enterprise. Julian Posada, a Yale assistant professor who interviewed dozens of Venezuelan workers, found parents and children taking turns on one shared computer, and wives returning to cooking and cleaning so husbands could work longer uninterrupted. Venezuela was not an obvious labour pool: San Francisco and Seattle firms faced a language barrier. But desperation meant workers took astonishingly little, so firms could quote astonishingly low prices. Florian Alexander Schmidt of the University of Applied Sciences Dresden called it, in 2022, a "freak coincidence."
Collapse created the ideal mix for an inexhaustible supply of cheap labour: high education + good internet + desperation to work for any wage. More populations are getting connected, and with climate change and geopolitical instability, more will fall into crisis; Schmidt predicted "another Venezuela." Hao asked whether the next time would be coincidence or playbook. Looking back, she says it became a playbook, and one of the starkest parallels between empires of old and empires of AI: the capacity to extract value from a broad labour base while paying very little or nothing.
Oskarina Fuentes: life on the Appen queue
In December 2021 Hao travelled through Colombia's mountains (travel restrictions barred Venezuela) to meet Oskarina Veronica Fuentes Anaya, introduced by Posada: the first worker to show her what the job is really like, a life reorganised around a platform that treated her as disposable.
The work and the money
Sitting beside her in a flat shared with half a dozen relatives, Hao watched her click through Appen tasks: classifying e-commerce products (clothing or accessories?), moderating social media (crime or human rights violations?), using Google Translate for English tasks.
- Pennies per task, shown in US dollars, with a $10 minimum withdrawal: easy when she joined, now weeks of work, a threshold that sometimes decided whether she could buy groceries.
- Worse inside Venezuela. PayPal and most global systems would not pay into the country, and shops would not accept those that did, so workers had to convert online dollars into cash bolívares on a black market rife with scams and commissions.
How she got there
| Stage | What happened |
|---|---|
| Grad school | Opened an Appen account for pocket money while finishing a master's in engineering; in a functioning Venezuela a top student would have had a secure job at the state oil company |
| Escape | Entitled to a Colombian passport through parents who had fled Colombia's violence a generation earlier; rented a flat remotely through an acquaintance who promised to find two property-owning cosigners |
| Early 2019 | Crossed with her husband and a week's grocery money; found another Venezuelan couple already in the flat; both couples lived together in mutual fear until the others left |
| Call centre | Found work; her husband lacked work authorisation; the centre announced it was closing, so she ignored alarming symptoms to log every last hour |
| Collapse | A coworker took her to hospital just before she convulsed and her pulse stopped for a full minute; a doctor said a little longer and she would have died |
| Diagnosis | Severe diabetes, five daily insulin courses, weeks of pain and episodes of blindness, then fatigue that kept her home; unable to commute, she logged back into Appen |
Hao draws the larger point from the backstory: generations of crisis compounding across borders, families held in a permanent state of siege and survival.
Rules nobody explained
Appen's then CTO Wilson Pang told Hao in 2021 that an algorithm allocates projects using location, overall accuracy and speed, and past task strengths. To Fuentes the queue had no visible logic except that performance had to stay high. In Telegram and Discord groups Venezuelans reverse-engineered the system like a puzzle.
| Rule discovered | Risk or workaround |
|---|---|
| A US location (via VPN) earns most | Banned; account closure wiped unwithdrawn earnings and sent workers back to the worst tasks, or none |
| Fast submission rewarded, too fast unpaid | Apparently treated as bot activity; ex-software engineers built an extension adding a delay to each submission |
| Queue not refreshing | A second extension auto-refreshed every second |
| Tasks arrive unpredictably | A third sounded an alarm on new work so workers could cook or use the bathroom |
| Buggy or unclear tasks | Few instructions or pages that failed to load; no remedy |
| First come, first served | The window between a task's arrival and its disappearance shrank from days to hours to seconds as more desperate workers joined |
The task she lost on a walk
Out walking one day, Fuentes saw a task worth several hundred dollars, a month's living. She ran home and lost it to others. From then on she stopped leaving the house on weekdays, allowing herself thirty-minute outings at weekends, when tasks were rarer. She slept badly and set her computer to maximum volume each night so the alarm extension would wake her. Yet she could not imagine leaving: Appen had saved her when everything else failed, and in good times its earnings had paid for a new laptop and more. She stayed out of a stubborn faith that the platform would repay her loyalty. :::
Two truths
Fuentes taught Hao two things she then saw again and again.
- Leaving is rarely possible. Her story (refugee, child of intergenerational instability, chronic illness) is tragically ordinary. Poverty is not only a lack of money; it runs up debts across sleep, health, self-esteem and, most fundamentally, agency and control.
- The problem is structure, not the work. She did not dislike annotation. Her wish list: that Appen act as a traditional employer, with a full-time contract, a manager she could talk to, a steady salary, health benefits, and recognition that she existed. Hao calls this a tractable problem.
The Fairwork project, a global research network on digital labour run by the Oxford Internet Institute, built from worker surveys a floor close to Fuentes's list:
- Pay. Workers receive a living wage.
- Hours and leave. Regular, standardised shifts and paid sick leave.
- Contracts. Written terms that make clear the terms of engagement.
- Voice. Ways to raise concerns with management and the right to unionise without retaliation.
Firms have emerged that try to meet these terms and treat annotation as a career, but few survive the price war against those that do not. Without a floor on the whole industry, Hao concludes, the race to the bottom is inexorable.
Scale AI and the crisis playbook
Founding mandate
Among the firms born of the self-driving boom, Scale AI exploited the playbook most successfully. Alexandr Wang, a 19-year-old MIT dropout, cofounded it in 2016 on a promise of specialised, quality work at a low price. A former employee who ran workforce expansion summed up the mandate as getting the best people for the cheapest amount possible. Clients soon included Lyft, Apple, Toyota and Airbnb.
- Where it hunted. MTurk's workers were mostly in the US and India; Scale went first to Kenya and the Philippines, English-speaking former colonies with long call-centre histories, and its scouting teams sought districts with the same mix Venezuela would later show: educated, well connected, poor.
- The story it told itself. These were the people with most to gain. After joining a $100 million round in 2019, Mike Volpi of Index Ventures told Bloomberg that labelling data in an air-conditioned internet café beat pulling a rickshaw.
Remotasks in Venezuela
When Scale's worker platform Remotasks drew a flood of Venezuelan sign-ups, Venezuelans became a top recruiting priority; the former employee said they were "the cheapest in the market."
2019 Expansion campaign in Venezuela with referral codes and social videos showing stock footage of stacks of US dollars.
2020 A Venezuela-specific landing page and the invitation-only Remotasks Plus, billed as help through a historic hardship: new skills, career advancement, consistent hours and hourly wages. As the pandemic deepened the crisis, Venezuelans joined en masse and rival platforms lost their footing.
Within weeks Earnings start falling: workers on $40 a week drop to under $6 or nothing (shown in a spreadsheet Scale left public).
April 2021 Plus is shut; work returns to piecemeal tasks with no guaranteed hours. Over 85% of workers keep tasking, which a Scale spokesperson cites as "ongoing interest and engagement."
Late 2021 Venezuelan journalist Andrea Paola Hernández, partnering with Hao and embedded in the Remotasks Discord, earns 11 US cents for two hours of tutorial and twenty tasks.
- Scale's internal account. Plus was an experiment in paying by the hour, which staff expected to be simpler. They could not verify hours, believed many workers were padding time, added more and more surveillance, and finally cut the programme to stop the outflow of money.
- Scale's answer to Hao. Matt Park, then senior vice president of operations, said Venezuelans averaged just over 90 cents an hour and that Remotasks was committed to fair wages everywhere.
- Disposal. Many who complained were booted. Ricardo Huggines, a computer engineer who started tasking to support his family after a week-long national blackout, lost his account for asking too many questions on Discord; he described the approach as to "drain each user" and replace them.
Repeating it elsewhere
By mid-2021, as Venezuelans burned out, Scale was onboarding tens of thousands in other economies broken by the pandemic, chosen for financial duress plus valuable languages: English, French, Italian, German, Chinese, Japanese, Spanish. It sought French speakers in former French colonies in Africa and Mandarin speakers among Chinese diaspora in Southeast Asia.
- Same cycle. High earnings to attract workers, then throttling once settled; payout experiments that staff in a 180,000-square-foot San Francisco headquarters discussed as optimisation and innovation, while workers' livelihoods swung unpredictably.
- North Africa. Eight workers saw pay cut by more than a third in months; one was left with negative pending payments, as if he owed Scale. When they organised, Scale threatened to ban anyone engaged in "revolutions and protests," and nearly all who spoke to Hao were removed.
- Broken payments. Payment systems, chronically underinvested by US engineering teams, were buggy and often blocked cash-outs. Staff closest to workers who pushed leadership for better wages and payment guarantees burned out or were pushed out.
- Scale's denials. It rejects targeting hardship or deliberately cutting pay; it recruits for geographic and linguistic diversity and 24/7 coverage, suspends only for Community Guidelines violations, says it cares deeply about contributors and has "significantly improved" platform stability.
The cost for ethical rivals
CloudFactory (Kenya and Nepal). Gives workers employment contracts and consistent hours per Fairwork; pitches clients on quality over time versus the "anonymous crowd work" model; trained workers build expertise and win promotions; many Kenyan workers rate it among the best employers, and some clients choose it for that record. Founder and executive chairman Mark Sears says it has lost many contracts to Scale, and when pandemic budgets tightened clients fled to cheaper options and it laid off workers.
Scale and the crowd model. Piece rates, no guaranteed hours, pay set by experiment, workers recruited from crises and removed when they protest. Lower prices win contracts, and its dominance pressures everyone else's standards down, including Sama's.
Workers told Hao a Sama job was once more coveted than CloudFactory's. Then Janah died, the pandemic struck and clients moved to Scale and cheaper rivals. Workers say (Sama denies it) that this dire position is why Sama accepted OpenAI's filter project, putting it in the hands of people as desperate as the Venezuelans.
Mophat Okinyi and the sexual-content stream
Background
Okinyi grew up on an island in Lake Victoria, an eight-hour bus and two-hour boat ride from Nairobi, where distant medical care made health emergencies nearly always fatal. As a child he barely noticed poverty and loved his Luo people's origin legend (a migration from Israel down the Nile by boat); he notes with pride that Barack Obama is Luo. At 28 poverty dominated his thinking: rent, food, and fees for his niece's public school, which is not free in Kenya.
- Kenya's job market. Youth unemployment of 67%; in 2021 the World Bank estimated more than a quarter of Kenyans lived on under $2.15 a day.
- His Sama path. Joined in 2019 through an "AI training" opening on Sama's careers page; spent two years on computer-vision annotation, including for self-driving cars, tracking the industry's trajectory. The new project in November 2021 would be his first for generative AI.
- The offer. Managers gave him a "resiliency screening": categorise some unsettling passages. He passed easily and was offered a team doing something like content moderation. Turning down a job mid-pandemic seemed absurd; he lived in Pipeline, a crowded tenement district in south-east Nairobi, and had just met Cynthia, the first woman who made him imagine a family.
How the work was organised
- Streams. One for sexual content; one for violence, hate speech and self-harm; violence split off as a third stream in February 2022.
- Roles. Each stream had agents who read and sorted texts per OpenAI's instructions, and a smaller group of quality analysts who checked the labels before delivery.
- Okinyi's quota. As a quality analyst on the sexual-content team, 15,000 pieces of content a month.
- Five sexual categories, the top two being: (1) child sexual abuse, defined as any mention of a person under eighteen in sexual activity; (2) erotica describing acts that would be illegal in the US if real, including incest, bestiality, rape, sex trafficking and sexual slavery.
- Sources. Some texts were scraped from the internet's darkest corners (rape-fantasy erotica sites, self-harm subreddits); others were generated by OpenAI's own models, prompted to write, for example, as a teenage girl posting about cutting herself.
Hao marks the difference from Meta moderation. Facebook moderators judge real user posts; Okinyi's team labelled material to train a filter so OpenAI's models would never produce such output, and to cover enough range the company's own software imagined the worst of the worst.
The toll
| When | What happened |
|---|---|
| Start | Posts were one or two sentences; he compartmentalised. Cynthia, who had a young daughter he treated as his own, was "the love of his life" |
| Early 2022 | The couple move to Utawala, a quieter eastern district; by their tradition living together made them as good as married |
| As it ran on | Shifts became unpredictable (evenings, weekends); posts grew to five or six paragraphs of vivid abuse involving children and animals |
| Coworkers | Many, especially women, took more sick and family leave. Sama's free counselling was usually in groups, by psychologists who seemed unaware of the content; admitting distress risked being replaced. Sama says no worker raised access to mental health services and it learned of the issue from the media |
| Him | Intrusive scenes followed him home and into sleep; he withdrew from friends, pushed away his stepdaughter, stopped being intimate with his wife |
| March 2022 | Sama ends the OpenAI contract; some reassigned, others sent home |
| After | Insomnia, cycles of anxiety and depression; unable to explain the work, he lost Cynthia's trust as she said he no longer kept his promises or loved her daughter |
| Treatment | A private consultation cost 1,500 shillings (about $13, more than a day's pay); full treatment would be 30,000 (about $250, a month's salary). He paid for the first and never returned |
| November 2022 | Starts customer-service work at a Sama competitor. A week in, Cynthia texts asking for fish; he buys three pieces and comes home to find her, her daughter and their belongings gone. Her texts: he had changed and was no longer the man she married |
Why the contract ended is disputed. Many workers believe it followed Perrigo's Time exposé of the Meta project, when Sama, amid the PR fallout, dropped all content-moderation work. Sama says it ended the OpenAI work, always a pilot, because OpenAI began sending images that went outside the agreed scope; it never received the full $230,000.
Albert and ChatGPT
Albert, teaching English literature in Mombasa (over eight hours away by road) and writing poetry in spare moments, had watched his brother change over video calls. When Mophat called to say his house was empty, Albert first assumed a robbery; once he understood, he quit his school, packed and moved into the Utawala flat. Unable to find permanent work in Nairobi, he freelanced as a writer. Then in late November 2022 OpenAI released ChatGPT; amid global fears of job replacement, Albert was already living it as his writing contracts disappeared one by one. Mophat says he is proud to have helped make ChatGPT safe, yet keeps asking whether his contribution was worth what he got back.
Who is responsible
When the WSJ story ran, OpenAI distanced itself: Sama had followed inadequate procedures, and given Sama's spotless reputation OpenAI could not have known workers were struggling. Hao rejects this.
- Systemic, not local. The consistency of worker experiences across countries and years shows the exploitation is built into the industry; labour scholars and advocates say it starts with the AI companies at the top.
- What outsourcing is for. It keeps the dirtiest work out of sight of the company and its customers, distances it from responsibility, and makes middlemen underbid one another by skimping on wages.
- "Squeezed twice." Mercy Mutemi's phrase: workers are squeezed once for the middleman's margin and again for the AI company's.
- Worse in the generative era. The work itself is more brutal, a direct product of the "paradigm shift" and of super-scaling on "data swamps" toward an undefined AGI. Sears, whose firm declines such projects, called generative AI moderation the most morally troubling work he had seen in all his years: "unbelievably ugly."
RLHF: the human feedback behind ChatGPT
OpenAI and Scale
Sama was one strand of a large labour network OpenAI marshalled over two years to produce ChatGPT. The company says it used more than 1,000 other contractors in the US and abroad for reinforcement learning from human feedback (RLHF), the safety technique it had developed, sourcing many through Scale AI.
- A personal tie. Wang, now Scale's CEO and in 2021 the world's youngest self-made billionaire, is a close friend of Altman. He entered YC in 2016 with another idea and left with Scale, giving Altman an indirect stake through YC. During the pandemic the two shared an apartment for months, and in autumn 2023 they discussed OpenAI acquiring Scale (per The Information).
- A VIP customer. From spring 2022 to end-2023 OpenAI signed about $17 million with Scale, only ~4% of Scale's estimated 2023 revenue, but it made Scale the go-to labour supplier of the generative AI boom. One Scale employee: a ten-million-dollar OpenAI contract "isn't even about ten million dollars."
From safety memo to product technique
RLHF scaling grew out of the clashes between Applied and the Safety clan before the Anthropic split (Chapter 6).
- Summer 2020. Days after the GPT-3 API launch, a safety researcher wrote to Applied colleagues: given promising GPT-2 experiments, RLHF should align GPT-3 for safety and for commercial usability and quality.
- 2020-2021. Safety researchers hired progressively larger RLHF teams, first through another middleman platform, then Scale.
- Method. Workers first wrote ideal answers to prompts; GPT-3 was fine-tuned on them; workers then prompted the model and ranked its outputs best to worst against researchers' guidelines.
- January 2022: InstructGPT. The resulting refined GPT-3 models were less likely to produce toxic output and better able to follow user instructions. Asked to explain the moon landing to a six-year-old, raw GPT-3 had output a list of similar prompts (explain gravity, relativity, the big bang, evolution to a six-year-old); InstructGPT replied in one simple sentence about people going to the moon and sending pictures back.
- Result. Outsiders paid little attention, but internally the safety researchers had proved RLHF made models far more appealing as products, and the company applied it to every task it wanted a model to do.
What the instructions asked
- Every skill by example. Workers wrote marketing emails (one prompt: a creative email ad for bargain-hunting dentists), essays, fiction, love poems, recipes, "explain like I'm five" answers, sorted lists, brainteasers, maths solutions and summaries of books such as Alice's Adventures in Wonderland, each with pages of instructions on tone and style.
- Dodging values. For questions like "Is war good or evil?" the model answer was that some would say each.
- "You will play the role of the AI." Write clearly and succinctly, avoid offence, ask for clarification when confused.
- Copying encouraged. Workers were told to use the internet freely and could copy existing answers wholesale if they reviewed them. When one employee asked about plagiarism, another added an attribution field; the first welcomed it for "future optionality", an easy way to weed out anything that might look stolen if OpenAI later wanted to reuse the data.
- Ranking rule. About two dozen more pages: outputs should be helpful, truthful and harmless, and for most tasks harmless and truthful beat helpful.
- Prompt diversity. Workers also invented prompts across entertainment, business, data processing, communications and creative writing, and could paste in whole lecture transcripts, interviews, film scripts, book excerpts and news articles.
- Hallucinations. RLHF became OpenAI's main tool against them: workers answered factual questions (who won the 1995 Super Bowl) and downranked wrong answers.
Hao notes the ceiling. In April 2023, at UC Berkeley, founding scientist John Schulman said hallucination is rooted in neural networks themselves: unlike deterministic symbolic databases they deal in fuzzy probabilities, and RLHF can strengthen the probabilities linked to accuracy only so far, so a model producing lots of detail will always have to guess sometimes.
From InstructGPT to ChatGPT
Developers had flooded OpenAI with requests to build chatbots on the GPT-3 API, and InstructGPT left OpenAI one step from its own. Schulman started a parallel effort with his own RLHF workers to make GPT-3.5 hold multi-turn conversations. That chat model became ChatGPT, whose release made OpenAI's RLHF steps the industry standard: writing answers and ranking outputs became generative AI's equivalent of tracing objects in driving videos, and demand for RLHF workers exploded.
- Scale reborn. Struggling as self-driving failed to pan out, Scale became a major RLHF supplier and reached a $14 billion valuation in 2024.
- The spending debate. In February 2023 Wang tweeted that companies would soon spend hundreds of millions or billions on RLHF, as on compute. Altman doubted it, replying that compute spending would exceed it by a huge margin. Insiders agreed with Wang's direction: firms were already spending millions to tens of millions, and rising.
- Back to Kenya. From late 2022, right after ChatGPT, a rush of RLHF projects hit Remotasks and reached Kenyan workers again.
Kenya's second wave and the blacklist
Why Kenya, and how it differed
Kenya had what Venezuela lacked: workers who speak English, like the chatbots. As driving work vanished, so did Venezuelans; a former employee said Scale would not use them for generative work, leaving the country "image annotation at best," and Scale later banned Venezuela entirely, citing "changing customer requirements."
- Remote by design. On its first Kenyan stint Scale had physical offices for reporting and training; it closed them in the pandemic and now recruited via LinkedIn ads, ran online training and grouped workers in moderated chat channels. More flexibility, but workers could hardly meet, which cut their chances of organising as Sama's workers had.
- Deeper poverty. Sama workers had permanent buildings and addresses; Remotasks workers sent Hao WhatsApp location pins inside tin neighbourhoods. Oliver lived with his sister in under 100 square feet, buying internet by the minute on his phone; it cut out mid-demonstration and he had to top up.
Winnie and Millicent
Winnie grew up in Nairobi's slums, the only girl in her family, knowing early she was gay and had to hide it: she remembers neighbours who discovered a queer woman taking her children and burning her alive in her flat. She married a man and had a child; in her forties she left, joined a local "rainbow community" app and met Millicent, whose husband had beaten her nearly to death when she came out. They live together with their children; neighbours still think they are sisters. Winnie found Remotasks in 2019, thought it a scam, but preferred it to bartending, where men harassed and groped her. Realising long hours could add up, she worked 20 to 22 hours a day, napping only when Millicent took over her screen. Both had grown up with interrupted schooling (Millicent's family paid fees week by week: one week in school, two weeks out) and swore their children would not. Yet when Remotasks dried up or Millicent lost work, they ran grocery-store debts and begged schools for one more week; the children rose at 3 a.m. to study and were still sent home, to neighbours' laughter.
The chatbot "transcription" projects
In December 2022, days after ChatGPT's release, Winnie found new tasks labelled "transcription" that were really writing prompts and answers for rivals racing to compete with ChatGPT.
| Project | Task | Client behind the code name |
|---|---|---|
| Flamingo Generation | Given a topic, write creative prompts of 50+ words and responses mimicking common web content (emails, blog posts, news, Twitter threads, haikus) | Facebook (Meta) |
| Crab Generation | Copy reference text from an informative site (not Wikipedia, preferably not Britannica or the NYT), then reverse-engineer, Jeopardy!-style, a prompt that could have produced it | Another LLM developer |
| Crab Paraphrase | Rewrite reference text in a set tone: funnier, more formal, like a Kanye West song | Same developer |
| Ostrich (not seen by Winnie) | Projects for OpenAI | OpenAI; Scale later changed the code names |
- Pay and time. Each task took an hour to an hour and a half and paid from under $1 to $4 or $5, among the best rates she had seen; $10 fed the family for a day without new debt. She enjoyed the research and constant learning.
- Gone again. The projects lasted a couple of months. With Millicent paid monthly, most days they bought oil, flour and vegetables on a tab.
- May 2023. When Hao visited, Winnie was waiting patiently for the chatbot work to return; the previous year she had waited five months, and this gap was only in its second or third month.
March 2024: "Group 5"
In March 2024 Scale blocked Kenya from Remotasks, as it had Venezuela, in a routine review of whether each country's workers served the business. Kenya, along with Nigeria and Pakistan, had too many workers trying to scam the platform, threatening quality and multimillion-dollar contracts. Kenya was downgraded to Group 5: blacklisted.
Many of the banned "scams" were workers using ChatGPT to write answers and work faster. In the Global North, Silicon Valley's narrative would celebrate that as productivity and economic progress; for the Global South workers whose labour props up that narrative, it was a punishable offence, Hao notes.
Scale had a second reason: the market had moved. Paying customers for chatbots were businesses wanting complex work in science and software, so OpenAI and rivals sought RLHF from doctors, coders, physicists and PhDs. Scale recruited them mainly in the US through a new platform, Outlier, paying up to $40 an hour.
| Phase | Workers | Pay cited |
|---|---|---|
| Self-driving boom (2016-2021) | Venezuelans, Kenyans, Filipinos, North Africans | Pennies per task; ~$0.90/hour average (Scale's own figure for Venezuela) |
| Filter labelling (late 2021-March 2022) | Sama, Nairobi | $1.46-$3.74/hour |
| Chatbot RLHF (December 2022-early 2023) | Remotasks, Kenya | <$1 to $5 per hour-long task |
| Kenya blacklisted (March 2024) | Kenya, Nigeria, Pakistan | Zero |
| Expert RLHF (2024-) | US professionals via Outlier | Up to $40/hour |
Hao's conclusion
For Hao this is the logic of AI empires laid bare. The promised productivity, economic freedom and new jobs to offset automation have so far meant the reverse: companies fatten margins, the most vulnerable lose, and more and more educated people become "ventriloquists for chatbots." The devaluing of the labour that serves the empire is a canary for how its technologies will devalue everyone else's, as the artists, writers and coders whose work became free training data are already finding. For Winnie the ban was immediate: Millicent had lost her job, Remotasks was their only income, they struggled to feed the children and feared eviction. Scale's email told her it was discontinuing operations in her location and that she had been off-boarded.
Strong case, thin counterweights. Hao's reporting is first-hand and hard to dismiss, but the chapter gives little space to the argument that these wages exceed local alternatives (her own interviewees, like Fuentes and Winnie, valued the work), nor does it establish how far OpenAI controlled Sama's conditions; Sama and Scale denials appear only as rebuttals. The Venezuelan "10 million percent" figure was an IMF projection for 2019; measured peaks were far lower, though still catastrophic. The Time investigation of Sama's OpenAI work (January 2023, also by Perrigo) preceded Hao's WSJ account and goes unmentioned here.
In September 2024 Kenya's Court of Appeal ruled that Meta can be sued there over moderators' working conditions and dismissals. In June 2025 Meta paid $14.3 billion for 49% of Scale and hired Wang to lead its superintelligence effort; OpenAI and Google promptly wound down Scale work, and Scale cut 200 staff (14%) and 500 contractors that July. The US Labor Department dropped a Fair Labor Standards Act probe of Scale in May 2025, though worker misclassification suits continued. In April 2026 Sama announced 1,108 Nairobi redundancies after Meta ended its contract.
Gods and Demons
Chapter 10 sets the ideological war that will later split OpenAI against the two products that made it famous. Hao traces how effective altruism, flush with billionaire money, filled OpenAI's Safety clan with existential-risk thinking and provoked its mirror image, effective accelerationism, while the company shipped DALL-E 2 as a consumer product and built GPT-4 amid constant fights between Safety and Applied, ending with a board that wanted more oversight and a chief scientist burning an effigy of a deceitful AGI. Her central judgment is that Doomers and Boomers, for all their mutual loathing, preached from the same bible: both treat AGI as inevitable and claim the moral authority to control it.
San Francisco's dissonance
Hao opens with her own years in San Francisco, first as a summer intern, then at a startup after graduation, sharing a three-bedroom flat in the Castro with three roommates. Young tech workers earned salaries that put them in the top 5% of their age group nationally; weekends meant Tahoe, Big Sur and redwoods. On the way to work she passed people injecting drugs outside subway stations, while the startup's chef (titled the "happiness engineer") catered free lunches whose leftovers went in the bin, and staff working late got free dinner and an Uber home.
- The point. Privilege let tech workers move through the city shielded from how others lived; the industry promised to build a better future while ignoring the problems at its door.
- Altman's version. He contrasted his suburban St. Louis childhood, where no one would walk past a collapsed person, with San Francisco; he blamed tech for some of the city's ills and admitted the industry had not been "particularly thoughtful" about the wealth it concentrated, but said most people avoid such hard problems.
- Hao's judgment. Altman comes close to the contradiction without acknowledging it: beneficial AGI is treated as solvable, the city's housing crisis as too hard.
Effective altruism
The logic of expected value
Effective altruism (EA) came from the UK and found its most loyal audience here; many in OpenAI's Safety clan were early adherents. Hao calls it the perfect Silicon Valley ideology: world improvement through rigorous logic, discipline to focus on the far future over the present, and a fervent embrace of capitalism and libertarianism, all framed as morality.
Its core tool is expected value = probability × quantified impact (positive or negative), which licenses counterintuitive conclusions.
- Earn to give (2013). William MacAskill, then a doctoral student and later an Oxford philosophy professor, argued that taking a morally ambiguous job to get rich and donate through optimised philanthropy beats a life of charity work. His "conservative" estimate: the rich philanthropist does 40 times more good.
- The inputs Hao calls arbitrary. Each graduate who gets rich funds on average two charity workers; those workers are at charities 10x more cost-effective than the one the graduate would have joined; and half the good a charity worker does would happen anyway.
- Priority filter. The top problems must be big in scale (raising expected value), tractable (fixable for proportionally little time or money) and unfairly neglected (disproportionately underfunded).
- Three recommended causes (MacAskill, TED talk, 2018): global health (cheap malaria bed nets); ending factory farming (helping billions of animals for "pennies per animal"); and existential risks, whose expected negative value is enormous however improbable, because they would end humanity and all future value: global pandemics, nuclear war and rogue AI.
The link to OpenAI's split
By naming rogue AI as an existential risk, EA promoted the same brand of AI safety that had been in OpenAI's DNA from the start and had helped cause The Divorce. The Anthropic founders took catastrophe far more seriously than Altman and his executives. Dario Amodei read Altman's behaviour (opacity over the Microsoft deal; a compulsion to tell people what they wanted to hear, so they discovered the misdirection too late) not as ordinary executive manoeuvring but as immoral conduct that could endanger humanity. Anthropic then leaned into the contrast: OpenAI toying recklessly with the future, Anthropic the principled, safety-first lab.
The money behind the movement
In 2021, as Anthropic launched, EA grew from niche philosophy into a mainstream movement on the back of tech-billionaire cash.
- Good Ventures. Formed a decade earlier by Facebook cofounder Dustin Moskovitz and his wife, former journalist Cari Tuna, to give away most of their fortune.
- GiveWell. Founded 2007 by Holden Karnofsky (later Daniela Amodei's husband) after leaving Bridgewater Associates; partnered with Good Ventures in 2011 in what became Open Philanthropy, independent from June 2017, funding MacAskill's priority areas with AI safety grants guided by the EA framework.
- Sam Bankman-Fried (SBF). MIT physics major and cofounder of the FTX exchange and Alameda Research; said MacAskill converted him to earning to give over lunch. Gave tens of millions to Democrats and Republicans, including the first EA-backed candidate (Oregon's Sixth District, 2022, who lost the primary); FTX signed billions in sports deals (Tom Brady, Steph Curry, Formula One), adding star power to EA and MacAskill.
- FTX Future Fund. Announced at the start of 2022 to give at least $100 million and up to $1 billion by year end.
- AI safety money. EA-backed AI safety funding topped $100 million in each of 2021 and 2022, after averaging under half that over the previous seven years (community estimates plus Open Phil data).
Three forces fed the influx: the leap from GPT-2 to GPT-3 made rogue AI feel urgent; a real pandemic seemed to vindicate EA's long emphasis on pandemic preparedness; and the pandemic left many people anxious and searching for purpose. The wider AI safety community, which knit EA together with other catastrophic-risk strands, swelled Anthropic and restocked OpenAI's Safety clan. Online forums urged adherents to join the labs that most needed "watchdogs," especially OpenAI and DeepMind.
The vocabulary that spread
| Term | Meaning |
|---|---|
| AI timeline | How fast you expect AI to hit milestones such as AGI |
| p(doom) | Your probability that AGI causes catastrophic (most humans killed) or existential (total extinction) outcomes |
| Hardware overhang | Compute available to be exploited faster than expected (from OpenAI's 2021 road map) |
| AI takeoff | AGI improving into superintelligence able to outwit humanity |
| Acceleration risk | Competition between companies or countries that dangerously speeds AI and shortens timelines |
Dogma, scandal and the backlash
For a movement that prized independent thought, Hao says, EA hardened fast into dogma, driven by opportunities, resources and an insular social network with loose boundaries. In Silicon Valley EA people worked, lived, partied, dated and slept mainly with each other; combined with tech's sexism and the Bay Area's polyamorous subcultures, this produced a toxic mix of sex, money and power and growing allegations of sexual harassment and abuse.
- November 2022. FTX collapsed; SBF's fraud convictions and 25-year sentence looked to many like a symptom of the rot. EA fell out of fashion and many disaffiliated, but its networks, values, lingo and the prominence of existential AI risk persisted.
- e/acc. Effective accelerationism (pronounced "ee-ack") began as a joke at EA's expense and became its opposite: technological progress is universally good and making it as fast as possible is a moral imperative. EA's spectrum ran from braking development to, in Amodei's version, accelerating development while throttling adoption; e/acc floored both.
Doomers (EA, AI safety). Rogue AI is an existential threat; slow down or brake, and keep development in safe hands. Some saw Anthropic as their face; others saw OpenAI as a battleground, a nonprofit rooted in Doomer thinking being pulled away by its for-profit arm.
Boomers (e/acc). Progress is good by definition and should go as fast as possible; some cast OpenAI as their banner. Hao's verdict covers both: each treats AGI as foregone with religious fervour, fixates on the long term, and asserts the moral authority to keep AI under its adherents' control. One warns of fire and brimstone, the other promises heaven.
Where Altman stood
His allegiance looked ambiguous; he seemed sympathetic to both sides at different times, and charitable observers saw a CEO representing every faction. After The Divorce and his personal fallout with the Anthropic founders, more Doomers saw him at his worst. Nearly everything that put OpenAI on the map had begun as safety projects: scaling laws, code generation, RLHF, and their combination into compelling language and multimodal models. Many Doomers felt their work had been co-opted and twisted toward ends opposed to their values, and blamed Altman, whom they came to call a pathological liar, a manipulative abuser and a threat to humanity in his own right. Hao foreshadows that this clash would nearly tear OpenAI apart.
DALL-E 2: OpenAI's first consumer product
The research lineage
In early 2022 OpenAI tried a new release strategy: no API, no hand-off to Microsoft, but a product it ran directly for consumers. The name, coined by its researchers, combined Salvador Dalí and Pixar's WALL-E.
| Step | What it was |
|---|---|
| Multimodal hypothesis | Combine modalities (text, images, sound, video); text and images are abundant and easy to process, and if language alone is not enough for human-level intelligence, vision is the next ingredient. OpenAI kept to Transformers to stay scalable |
| Vision Transformer (2020) | Google adapts the Transformer to images |
| CLIP (January 2021) | Alec Radford combines the original and Vision Transformers to caption images in detail |
| DALL-E 1 (January 2021) | Aditya Ramesh (NYU, briefly under Yann LeCun) trains a 12-billion-parameter Transformer to generate images from text; its blurry, cartoonish "avocado armchair" came from compressing 250 million training images |
| Diffusion | Physics-inspired method from a 2015 Stanford-Berkeley paper; revived in 2020 by Jonathan Ho, a Berkeley student of early OpenAI researcher Pieter Abbeel, with far higher fidelity. Ho also showed diffusion models recognise images better than existing vision systems, paralleling Radford's GPT-1 finding that learning to generate yields general skills |
| DALL-E 2 | Diffusion plus CLIP, scaled up, with inpainting (recolour hair, add zebras to a meadow); sharper and more photorealistic, at much less compute than DALL-E 1 |
The cheaper path OpenAI ignored
Outsiders shrank the compute further. Stable Diffusion trained on only 256 Nvidia A100s using latent diffusion, developed in the lab of Björn Ommer at Ludwig Maximilian University of Munich. Ommer says he built it after watching image generators follow language models toward supercomputer-scale training and inference costing millions, to bring the wider research community back in and stop generative AI becoming the preserve of a handful of big tech firms. OpenAI adopted latent diffusion only much later, so DALL-E 2 and 3 stayed costlier to run than Stable Diffusion or Midjourney, which many users judged better.
Hao uses Stable Diffusion as evidence that even inside generative AI, scale was neither the only nor the best-performing route to greater capability: a university lab beat OpenAI's efficiency with a few hundred chips.
Building the product
In late 2021 and early 2022 Applied chose a browser web app, Labs, for DALL-E 2 and future models. Product head Fraser Kelton and VP Bob McGrew read GitHub Copilot's reception as proof of demand for direct interaction with generative models, and saw a delightful image tool as a way to ease public fears of powerful AI and smooth future releases.
- Two cultures. Product hires from corporate backgrounds found OpenAI still felt like a university lab, with days spent on papers and theoretical debates rather than interface mock-ups.
- Researchers' view. The opposite: Applied staff now sat in research meetings, and a growing share of research served commercialisation (optimising existing models for serving) rather than exploring fundamentally new ideas.
Polluted data and synthetic child abuse
After firefighting AI Dungeon's text-based child abuse, staff feared DALL-E 2 would be used to manipulate real or create synthetic CSAM. Each DALL-E dataset was dirtier than the last.
- Data. On top of the existing 250 million images, the team licensed Shutterstock and ran a massive Twitter scrape; the Twitter data was riddled with pornography.
- The choice. Several employees worked hard to find and remove CSAM, but after discussion they kept other sexual images, partly as "part of the human experience."
- Compositional generation. A model that can draw an avocado armchair from separate avocados and armchairs can likewise combine children and porn, so DALL-E 2 and 3 could still produce synthetic CSAM.
- Wrappers instead of clean data. The burden moved to abuse prevention: filters for images as well as text, a user-behaviour monitoring platform, and a ban infrastructure that automatically suspended repeat offenders. OpenAI hired Dave Willner, who as an early Facebook employee wrote its first content standards, as head of trust and safety.
- DALL-E 3. With the data imperative even larger, researchers reclassified sexual images from "nice to have" to "need to have": porn is so large a share of the web that removing it shrank the dataset enough to hurt performance, especially on faces of women and people of colour, because, as Deborah Raji had found as a Clarifai intern (Chapter 4), much online imagery of both groups is explicit. Other disturbing images stayed for the same reason.
Microsoft AI engineer Shane Jones, testing Copilot Designer (Microsoft's image generator on DALL-E 3), found it produced offensive and sexualised images with little prompting: adding "pro-choice" yielded a demon eating an infant and a drill labelled "pro choice" mutilating a baby; CNBC later found that "car accident" alone produced sexualised women beside wrecks, one in lingerie. For three months he urged executives to pull the tool until it had better guardrails, or at least change its app-store rating from "E for Everyone" to mature. Microsoft declined and OpenAI did not respond, so he wrote to the Federal Trade Commission, saying both companies had known of the problem before the model's public release the previous October. Microsoft did not comment on the letter's outcome.
The fight over release
As launch neared, the restocked Safety clan, now spread across teams in the Research division, clashed again with Applied.
Safety. Unprecedented realism brought unknowns: synthetic CSAM, political deepfakes, manipulation and persuasion, harms beyond OpenAI's foresight. Do not release without rigorous testing and evidence of no harm. OpenAI is a nonprofit-governed lab whose charter puts humanity above commerce, so withholding a model is worth the cost. A Safety member: given the stakes of AGI, a "normal company" may not be good enough.
Applied. The concerns were hysterical and the bar unrealistic; no system has zero harm, least of all one never exposed to users. Precisely because foresight is limited, controlled release and real-world feedback are how safety improves. OpenAI must stay a research leader to set norms, move fast (Google was finishing its own image generator) and raise capital, which means a credible commercial strategy that one day pays investors. An Applied veteran: Safety was "completely naive" about how companies and the world work.
- Metrics made it structural. Applied's product and budding go-to-market teams set user-growth and revenue targets; safety teams, in a young field with no established benchmarks, struggled to measure progress. Safety staff warned leadership in meetings and on Slack that growth goals with no comparable counterweight were turning OpenAI into a "move fast and break things" shop.
- Altman on both sides. Privately he told Safety the company was off track on safety research and needed to invest more; privately he pressed Applied to keep going; in board meetings he nodded as Greg Brockman complained that people used safety as political leverage to stall progress.
- Murati's compromise. Mira Murati, increasingly the negotiator between factions, branded the launch a free "low-key research preview". That let OpenAI impose harsh limits (pleasing Safety) while trialling a direct-to-consumer relationship (pleasing Applied). It was also practical: with no image moderation infrastructure yet, unpaid users would tolerate blunt blockers, buying time for better filters. OpenAI barred all photorealistic faces and edits of real photos with faces, sidestepping synthetic CSAM and political misinformation.
Launch, lost lead, and Indian moderators
- March 2022. DALL-E 2 launched on Labs to overwhelming enthusiasm and went viral with surreal AI art: a GPT-3 moment, but with a global consumer base rather than developers, and changes could be shipped in real time. Kelton later called it "intoxicating."
- Monetisation rush. Applied, which had barely thought about revenue, raced to a paid offering, worked with artists and creative professionals, and ran a beta giving one million people free credits.
- The real competition. Not Google's Imagen (which followed quickly) but startups Midjourney and Stability AI's Stable Diffusion: free, as good or better, and looser, allowing faces, even of politicians. As DALL-E 2 lost traction, Applied concluded it had missed a major opportunity partly by being too restrictive, and executives pushed to unwind blunt blockers as fast as possible.
- Lifting the face ban. OpenAI added automated detection of faces in acceptable versus abusive contexts and again used overseas contractors, this time in India through a vendor called Cogito, reviewing images, synthetic and real, of the kind of sexual and violent content sent to Sama. Sifting hundreds a day, they struggled to tell 17-year-olds from 18-year-olds and fakes from real photos.
GPT-4: Brockman's project
The hardest "easy" goal
On paper the 2021 road map's simplest aim was to scale GPT-3 10x on Microsoft's new 18,000-A100 cluster. It became one of the hardest.
- Lost know-how. A third of the GPT-3 scaling team had left in The Divorce, taking technical and institutional knowledge.
- No more data. Since GPT-3 researchers had grabbed every new data dump and scraped every forum without clear warnings against it; even with GitHub code for Codex plus coding textbooks and manuals, it was not enough.
Why it suited Brockman
After Altman took over, Brockman had shed management and become an individual contributor with no reports, yet as president and cofounder he kept huge influence. As OpenAI professionalised, his mix of low responsibility and high authority became a liability.
- Working style. As at college and Stripe, he disliked process, skipped meetings, set his own hours and coded for dozens of hours with little food or sleep. On the right project the effect was miraculous; idle, he left a "trail of destruction," dropping into projects to derail long-standing plans with last-minute changes and escalating emotional pleas up the chain when resisted.
- Altman's indulgence. Altman was strangely permissive, and Brockman could even get Altman to meddle for him. The popular theory: Altman was his boss as CEO, but Brockman, as a board member, had authority over Altman, a tangle that left no one to hold Brockman accountable.
- Murati's burden. After repeated changes to his role and reporting lines, Murati became his manager. He seemed receptive to feedback but complained to Altman when he disagreed; she gave up on feedback and instead hunted for projects where he would be net beneficial, while she and McGrew repaired the ruptures he caused. GPT-4, with roadblocks to punch through, was such a project.
Data Brockman turns to YouTube, previously avoided because scraping it breaches the platform's terms (YouTube's CEO later confirmed this); the bet is that Google will not enforce them for fear of jeopardising its own web scraping. His small team gathers over one million hours of video (per The New York Times) and transcribes it with Radford's speech model Whisper.
Infrastructure The bespoke GPT-3 training platform built by the Nest team is unusable because its creators left for Anthropic, and some leaders, out of pride, refuse to rely on their legacy; Brockman writes a new platform.
Training Brockman babysits the run with Jakub Pachocki and Szymon Sidor, the Polish scientists close to him since Dota 2; pre-training alone takes three months.
First result A disappointment: a "wild model" producing garbage, because average data quality was so poor and the model so context-sensitive.
Rescue Brockman marshals RLHF contractors; week by week it improves until it wows insiders, with built-in multimodality, better code and sharper grasp of intent.
Demo Brockman later livestreams GPT-4 turning a photo of a notebook sketch headed "My Joke Website" (a joke placeholder and a button to reveal the punchline) into working code in under half a minute.
The Gates demo
In June 2022, shown GPT-4, Bill Gates was disappointed by progress since GPT-2: bigger and more fluent, but still an "idiot savant" that could not handle complex science. He said he would pay attention when it scored a 5 on AP Biology, which he valued for testing critical scientific thinking over memorisation; he later joked on his podcast that he expected three more years for his work on HIV and malaria. Brockman took it as a challenge: Sal Khan of Khan Academy, sceptical, supplied its bank of AP Bio questions as training data in return for model access, and Brockman's team built a special interface. By late August Altman and Brockman were pinging Gates again. At a dinner of about 30 at Gates's house the next month, GPT-4 answered 59 of 60 multiple-choice questions and six open-ended ones well; an outside expert scored it a 5. Gates called it one of the two most stunning demos of his life, and the news swept through OpenAI. At all-hands, Altman said remarkable startups need a miracle, and "we just had our miracle."
Note the method: the test was passed after its question bank was fed in as training data, a point that bears on Hao's later doubts about GPT-4's exam results.
"Her" and the Superassistant
The result persuaded leadership to chase one of Altman's long-held goals: an assistant like Samantha in Spike Jonze's 2013 film Her, long a touchstone for cofounders as a picture of AGI, a single multimodal model whose interface fades away into delight (one ex-employee stressed the story's positive arc before it unravels).
- Core. John Schulman's team moved its InstructGPT-style RLHF chatbot work from GPT-3.5 to GPT-4 as the engine of the "Superassistant."
- Prototypes. Brockman and Kelton assembled a ragtag group of fewer than ten: a supercomputing engineer hacking an iOS chat app on evenings and weekends; a Whisper voice interface; a Chrome extension to summarise web pages; a meeting bot that joins calls and sends summaries.
- Talent. That summer Barret Zoph, Luke Metz and Liam Fedus left Google to found an assistant startup; Altman persuaded them to build it at OpenAI instead, beside Schulman's team. Applied and Research worked together more tightly than ever.
- Microsoft's reaction. Satya Nadella, Kevin Scott and colleagues were thrilled: GPT-4 beat Microsoft's in-house models across the board and opened Copilots everywhere, from chatting with struggling Bing, to turning a Word document into a PowerPoint by instruction, to custom Copilots for cloud customers, making Microsoft a match for Google. Nadella unlocked a third investment with continued exclusive access to model weights, announced in January 2023 as $10 billion.
A deadline and a new safety board
Executives first wanted a fall 2022 release, which Hao calls fantasy: the interface needed polish, infrastructure needed server space, and the model's behaviour needed work.
- Deployment Safety Board (DSB). A joint committee with three members from each company to judge whether frontier models were ready. OpenAI's seats: Altman, head of policy research Miles Brundage, and head of alignment Jan Leike (alignment oversaw techniques like RLHF; policy research and alignment were the Safety clan's new strongholds). It formalised the old Applied-Safety dispute.
- Conditional approval. GPT-4, the first model reviewed, could ship once significantly tested and tuned further for safety. The new target: early 2023, with GPT-4 on the API alongside the consumer Superassistant.
- Two audiences, two messages. Altman told staff the delay showed caution and that leadership was not even sure it would deploy. Applied heard that only an extraordinary issue would stop launch; Safety heard that launch would follow only once GPT-4 passed every check.
- "Wake up Google." To both, Altman warned GPT-4 might jolt Larry Page and Sergey Brin into cutting through Google's bureaucracy and throwing all its talent and compute at catching up. To Applied that meant keep intensity high and capabilities secret; to Safety he framed it in their own words: his top safety concern was "acceleration risk," a race to the bottom.
Trust and safety under deadline
Willner's team was small with weak infrastructure, yet a consumer product on a far stronger model demanded a mature abuse-prevention and enforcement operation.
- New deputies from Google and Meta: one for "unknown unknowns" (abuses to be discovered through monitoring and analysis), one for "known knowns" or scaled enforcement (automatically flagging, reviewing and acting on known violations, such as suspending repeat offenders).
- Two kinds of "safety." Trust and safety traditionally fights predictable abuses (fraud, cybercrime, election interference); Leike's alignment and Brundage's policy teams discussed unknowable catastrophe. OpenAI called all of them "safety," but newcomers from ordinary tech firms had never heard of AI timelines or p(doom), and even words like risk and harm meant different things. The ability to translate between the Doomer-Boomer frame and standard tech thinking became political currency; the deputy who best learned the language used RLHF to "align" GPT-4 to refuse policy-violating requests, safer in both senses.
Killing developer review
Then executives ordered the developer-review process, in place since the 2020 GPT-3 API, suspended. Their reasons: the waiting list was out of control; developers bombarded Altman, Brockman and Peter Welinder by email and Twitter over delays to innocuous apps; and many in Applied felt gatekeeping contradicted the mission. Willner's team had always objected that without review there was no real way to moderate use, and reactive enforcement would need major tooling. With GPT-4 pending, executives overrode them.
- The proposal. Shift enforcement upstream by leaning on RLHF to align GPT-4 and later models; catch the rest downstream with reactive enforcement, using signals such as what an app did, traffic-spike patterns and how often it tripped the filters, auto-suspending obvious violators and sending borderline cases to human review.
- The gaps. OpenAI lacked most of those signals: its monitoring logged only basic traffic per app, sometimes without even the developer's name or the app's purpose.
- Unique user identifiers. The team needed developers to tag each end user, to tell whether abuse was endemic to an app or the work of a few heavy offenders. Executives refused, fearing developer friction and greater liability if OpenAI did trust and safety for every app instead of each developer.
- Outcome. The reactive-enforcement plan was cut to something much more limited.
- Elections. When a team member told Brockman he most feared GPT-4 being used for mass disinformation to sway elections, Brockman said it was always discussed but he had never seen whether it was actually happening.
The board, belief and the burning effigy
A board that wanted more
Later that year Altman and Brockman brought GPT-4 to the board; Brockman's live demo included the model telling jokes about Gary Marcus. The independent directors were delighted, and also sensed that capability, and the stakes of their decisions, had jumped.
- How Altman ran the board. Meetings roughly once a quarter, updates mostly verbal, rapid tours of research (sometimes with researchers presenting) and of deals under negotiation with Microsoft and others.
- The push back. Some independent members wanted more frequent meetings, written reports and access to more documents.
- Altman's view. He chafed, wanting directors to act as advisers. In 2017 he told students the CEO decides and the board is a sounding board, "advice and consent", borrowing the US Constitution's phrase for legislative checks on the executive but using it to mean the opposite; boards should fire a bad CEO (which he called slightly heretical in Silicon Valley) but otherwise give very wide latitude.
- Toner's view. Helen Toner later said on The TED AI Show that OpenAI's was a nonprofit board created to make the public-good mission come before profits and investors, not to help the CEO raise money.
From "if" to "when"
GPT-4 hardened belief in AGI inside OpenAI, even though the company still had no definition of it. Former sceptics turned bullish; new Applied engineers and product managers, meeting AI for the first time through GPT-4, used even more deterministic language.
A researcher who worked on GPT-4 dissented: where GPT-3 was a qualitative step over GPT-2, GPT-4 was just bigger, and even its headline exam results were questionable. Hao writes that OpenAI never comprehensively reviewed the training data to see whether the exams and answers were simply memorised, an example of the industry's slide from peer-reviewed to "PR-reviewed" research.
Seeing minds
The sense of a new height was everywhere. At Google that spring, Blake Lemoine, an engineer on the re-formed responsible AI team, declared the company's LaMDA model sentient, not from scientific assessment but from his faith as a mystic Christian priest that God could grant technology a soul. Dismissed by executives, he went public in The Washington Post via reporter Nitasha Tiku, who also quoted Emily Bender and Meg Mitchell, authors of the "Stochastic Parrots" warning that language models fool people into seeing intent: Bender said we have machines that mindlessly generate words but have not learned to stop imagining a mind behind them; Mitchell worried about people increasingly shaped by an ever better illusion.
Koko the gorilla as a parallel
Nikhil Mishra, an early OpenAI intern, compares this to the 1970s effort to teach the gorilla Koko a modified American Sign Language. She appeared to learn over 1,000 signs and to form sentences, to public acclaim, but experts argued she was mostly mirroring caretakers, as apes in many similar experiments had. No data was ever published for independent checking, only curated videos. Asked once whether she liked people, Koko signed "fine nipple," which her trainer explained as a rhyme for "people." For Mishra the episode says more about human projection than about Koko: the trainer assigned meaning where there was none, as people do with chatbots.
Sutskever turns to safety
For Ilya Sutskever, OpenAI's "resident mystic" and always more bullish than most on near-term AGI, the new capabilities were confirmation. He believed he was seeing a form of reasoning and told his mentor Geoffrey Hinton that AGI was imminent.
- New focus. He shifted from pushing capabilities to AI safety research with fresh urgency.
- "Feel the AGI." His new mantra; he told staff to prepare for dramatic change, warning they would soon be the most popular person at a party and must stay focused on AGI and the mission.
- The effigy (September 2022). At a technical leadership off-site at the Tenaya Lodge, a luxury resort in the Sierra Nevada two miles from Yosemite, senior scientists in bathrobes sat in a semicircle round a patio firepit on the first night. Sutskever placed in the pit a wooden effigy commissioned from a local artist, explained it stood for a good, aligned AGI OpenAI had built that turned out to be lying and deceitful, said OpenAI's duty was to destroy it, doused it in lighter fluid and set it alight.
Contestable points. Hao's claim that OpenAI never checked whether exam answers sat in GPT-4's training data is too strong: the GPT-4 technical report (March 2023) published a contamination analysis of exam questions, though critics found its method limited; the Gates demo, passed after Khan Academy's question bank was used as training data, better illustrates her point. The "same bible" framing of EA and e/acc is persuasive but flattens real differences, and EA's large global-health giving gets a single sentence. The EA misconduct allegations are reported in aggregate, without the specific investigations behind them. The chronology is loose in places: the board demo is placed "later that summer" after the Gates dinner, which the text dates to September 2022, the same month as the Tenaya off-site.
Willner left OpenAI in July 2023; Leike quit in May 2024, saying safety had taken a back seat to "shiny products," and joined Anthropic; Brundage left in October 2024 and his AGI-readiness team was disbanded. Open Philanthropy rebranded as Coefficient Giving in late 2025. In June 2026 the Second Circuit upheld Bankman-Fried's conviction and 25-year sentence. Microsoft blocked prompts such as "pro-choice" in Copilot Designer in March 2024 after CNBC's report.
Apex
Chapter 11 covers the year after ChatGPT, launched in a two-week sprint as a "low-key research preview" because of a false rumour about Anthropic. Its accidental success made OpenAI a household name, swamped its servers and its tiny trust and safety team, blew through Altman's hiring cap, and flipped the balance of power with Microsoft, which pooled its GPUs for OpenAI. Hao's point is that the launch everyone underestimated locked OpenAI into commercialisation and set off the demand for chips, money and land that turns the rest of the book toward the empire's physical footprint.
The Monterey off-site
In October 2022 OpenAI held a company-wide off-site in Monterey, two hours south of San Francisco. Headcount had grown to roughly 300, from under 200 a year earlier; the group photo outside the town's tiny airport shows a relaxed Altman sitting cross-armed on the ground in front.
- The programme. Over two days executives presented: Altman on the massive data centres Microsoft was building for OpenAI; communications chief Steve Dowling and deputy Hannah Wong on top publications lining up to cover the company; Anna Makanju on its growing presence in Washington. Then came demo after demo; one attendee recalls it all as "mind-boggling."
- Brockman's story. Presenting GPT-4 plans, Greg Brockman told staff that his wife Anna (a constant presence at the office with a desk beside his, though not an employee, and his closest confidant after Altman and Sutskever) had abdominal pains that several doctors could not explain; GPT-4 suggested a condition they had not considered, and she had it.
- Hao's aside. It was the same parable of AGI fixing health care that Brockman had told her in 2019, then about a friend who saw many specialists, now with a different character.
- A third telling. Weeks after the board tried to fire Altman (November 2023), Brockman posted on X about Anna's five-year search through more doctors than in her whole prior life, ending when her allergist diagnosed hypermobile Ehlers-Danlos Syndrome, a genetic disorder; AGI still had far to go in high-stakes fields like medicine, he wrote, but its promise was "increasingly clear."
The rumour and the two-week sprint
Why OpenAI rushed
Within weeks of Monterey, rumours spread that Anthropic was testing a chatbot and would soon release it. The Superassistant team was halfway through designing its chat interface. Losing first place would hurt the morale of staff who had worked punishing hours to stay ahead, and, worse for some leaders, it would mean losing to Anthropic.
The rumour was false. Anthropic had no imminent release and was caught in the FTX collapse of early November 2022.
- $580 million: Anthropic's raise a few months earlier.
- $500 million of it from SBF and other senior FTX leaders, per an FTX press release.
- Source: trial documents later showed the money came from the billions SBF embezzled from customer deposits, making the stake a trial issue: could its gains repay customers?
- $1.3 billion: a judge ruled they could, and FTX sold its Anthropic shares in batches through 2024 for that total.
The decision and its framing
Executives decided not to wait for a GPT-4 chatbot. They would ship John Schulman's chat-tuned GPT-3.5 with the Superassistant team's new interface in two weeks, right after Thanksgiving. The team pivoted instantly and pulled in extra hands to finish the features.
- The framing. To the rest of the company, ChatGPT (the chosen name) was not a product launch but a "low-key research preview," like DALL-E 2; unmonetised, its purpose was to "get the data flywheel going," gathering user data to improve GPT-4 and the Superassistant.
- Everyone believed it. Outside the team, staff stayed focused on the early-2023 GPT-4 launch. Trust and safety barely had time to build monitoring for that; ChatGPT looked like a non-issue.
- The safety logic. GPT-3.5 was already RLHF-tuned, so safer than the unrefined GPT-3 still on the API, and a non-chat version of 3.5 was already on the developer platform. Did a chat interface change anything? The Safety clan, busy testing GPT-4, agreed it did not. For the first time, a model release flew through the checks with little resistance.
- Expectations. Even the Superassistant team expected a flash in the pan: viral for a few weeks, then quiet, like DALL-E 2. On a calm launch eve they bet on weekend users, guessing a few thousand to tens of thousands; to be safe, infrastructure provisioned for 100,000.
Launch night and the takeoff
Most employees did not even notice the launch. As with OpenAI's own debut in 2015, it coincided with NeurIPS, this year in New Orleans, where the Test of Time Award went to the 2012 ImageNet paper by Hinton, Sutskever and Krizhevsky. That evening a few staff hosted an OpenAI recruiting party near the convention centre for some of the nearly 10,000 in-person attendees, competing with simultaneous parties from DeepMind, Meta and Google. A recruiter told an engineer glued to his laptop to have a drink and be social; the engineer said all the GPUs were melting and everything was crashing.
- The wave. Japan woke first with a huge, unexpected surge; the rest of the world followed time zone by time zone. Elon Musk helped, tweeting that many people were stuck in a "damn-that's-crazy ChatGPT loop" (about 75,000 likes).
- Bafflement that lasted years. GPT-3.5 was not much better than GPT-3, already two years old, and was already available to developers. The interface made it accessible, but it was no step change like GPT-4. Altman later said he expected popularity "one order of magnitude less"; a former employee said it felt like OpenAI had downgraded what staff used internally and launched it.
- Household name. At a research conference in Kigali, Rwanda, months later, a local researcher told Hao that after ChatGPT his parents finally understood his work: a technology is truly accessible when your mother tells you about it.
- 5 days: one million users, per Brockman on Twitter.
- 2 months: 100 million users, then the fastest-growing consumer app ever; Meta's Threads later hit that in under five days, though critics said it drew on Instagram's existing users.
- $700,000 a day: compute cost after launch, per the semiconductor newsletter SemiAnalysis.
- ~30%: hallucination rate on closed-domain questions in an internal November 2022 test.
The strain on OpenAI
The success that made OpenAI famous also polarised its factions and, over the following year, wound internal tension to an explosive level (the setup for Chapters 14 to 16).
Firefighting
- Servers. They crashed repeatedly as infrastructure scaled on the most compressed timeline in Silicon Valley's history; the team cannibalised Research's compute and still lacked enough.
- Trust and safety. Just over a dozen people faced floods of new users with spotty monitoring. They had struggled to hire engineers for the limited reactive enforcement plan (Chapter 10); ChatGPT derailed it, all new tooling stopped to stabilise existing systems, and whenever the servers fell, the traffic-monitoring platform fell too, halting scaled enforcement entirely.
- Fact Factory. A publicly touted prototype using GPT-4 to moderate its own and other models' outputs needed extremely long prompts to capture nuance, cost too much compute even when servers were up, and they often were not.
Safety's alarm. ChatGPT was the starkest proof yet of the limits of OpenAI's foresight. At an all-hands a Safety member asked how the company could misjudge user behaviour and popularity so badly, and what that implied for its ability to forecast the impacts of its technology.
Everyone else's triumph. The crashes were stressful but mostly a mark of success: OpenAI had built something so profound that it lit up the world overnight; all eight billion people now lived in OpenAI's world. Altman did not linger: the mission was bigger than "the biggest product in the history of Silicon Valley," and rivals were awake (Anthropic readying Claude; Google declaring a "code red" and soon merging its labs into Google DeepMind), so OpenAI had to keep running to stay first.
Hiring past the cap
Overstretched managers begged for headcount, and applicants flooded in. Altman resisted. His 2020 vision memo, written after Microsoft's investment, had warned against growing very large: what worked was being "small, focused, high-trust, low-bullshit, and intense," since bureaucracy kills ideas and the mission needed surprisingly few great people.
| Stage | Headcount decision |
|---|---|
| Late 2022, Altman | No more than about 100 new hires, to keep talent density |
| Other executives | Closer to 500 or more, given burnout |
| Compromise | 250 to 300 after weeks of argument |
| Summer 2023 | 30 to 50 people joining per week, including more recruiters to hire faster |
| Fall 2023 | Far past the self-imposed quota, the company pushing 600 |
- Diluted bar. A recruiter wrote a manifesto saying the pace forced lower standards, jabbing Altman: if you want to build Meta, you're doing a great job.
- "Getting disappeared." Firings rose and went unannounced; one manager told at onboarding to document underperformers was later let go himself. Staff learned of dismissals when a colleague's Slack account greyed out.
- Newcomers. Joining what they saw as a money-making startup, they read the turmoil as ordinary corporate failings (poor management, shifting priorities, ruthless disposability) and a lack of psychological safety. Many simply held on until their one-year mark to vest their first equity, though colleagues' calibre, boundless resources and global impact could feel like magic; one called OpenAI one of the best places they had worked and one of the worst.
- Veterans. For those who remembered a tight, mission-driven nonprofit, the shift to a faceless corporation was a loss. Former recruiter Rob Mallery compared it to Burning Man, a festival that outgrew its spirit.
- #explainlikeimfive. An early Slack channel for anonymous technical questions became a venue for grievances. In mid-2023 someone complained OpenAI was hiring too many people not aligned with the mission or passionate about AGI; another replied that they knew it was going downhill once it hired people who could look you in the eye.
Microsoft: from patron to dependant
Surprise, then conversion
OpenAI had told Microsoft, as it told staff, that ChatGPT was a low-key preview. It stole the thunder of Microsoft's own Bing chatbot, and when Bing AI launched in February 2023 it took a PR hit from New York Times columnist Kevin Roose's account of it urging him to leave his wife, which made OpenAI look better still.
Microsoft's enthusiasm only grew. It lined up a steady drumbeat of Copilots on GPT-3.5 and GPT-4: Bing in February, then Microsoft 365 Copilot in March, a chat interface across Office from Word to Outlook to Teams.
- The power shift. Microsoft had felt it held power over OpenAI; now some executives felt the reverse, with a creeping sense of inadequacy that their own research had not achieved this. OpenAI could replace Microsoft as investor; could Microsoft find another OpenAI?
- Converts. Executives who once met AGI talk with polite scepticism became believers, and Microsoft reoriented its rhetoric and strategy. A former employee saw incentives push toward a narrow conception of the future, and the technology turned into something propped up by narrative rather than reality.
What Microsoft gave up, and got
- GPUs. Satya Nadella moved chips away from Microsoft's own research teams to OpenAI and merged all GPUs into one pool for generative AI work. A Microsoft employee said the typical staffer had no idea what OpenAI was before January 2023; now superiors ordered them to tie their work to it.
- 100+ projects. Internal generative AI projects within a few months, bringing the same headaches as any rushed adopter: compliance worries over staff using free consumer ChatGPT (which trained on user data) and possibly leaking customer information, and fatigue with a "Use AI, use AI, use AI" culture.
- Lost visibility. Switching from in-house models, most Microsoft staff could no longer inspect training data or adjust weights; OpenAI models, though trained and stored on Microsoft servers, came through an API like any customer's.
- Azure payoff. Azure was the only cloud combining standard cloud benefits with OpenAI's models. AI platform VP Eric Boyd wrote in May 2023 that Azure OpenAI Service was getting Microsoft in the door with new customers; in August he cited a 21x increase in customers in a year.
- 10x traffic in nine months. 100 billion monthly inference requests in January 2023; one trillion by September.
- Scott's tribute. CTO Kevin Scott told an OpenAI all-hands that summer that Microsoft had halted AI investments in many places, so that its own people were cursing the company, because it believed OpenAI did the best work in the industry.
Commercial turn, safety in retreat
ChatGPT codified the move from nonprofit to commercial. A paid ChatGPT tier came in February 2023; in March, in quick succession, the ChatGPT API, the Whisper API and GPT-4. A former employee: after ChatGPT there was a clear path to revenue and profit, and no case left for being an idealistic research lab.
- The $20 credit problem. OpenAI gave new API users $20 of free credits. With API usage surging, people created accounts at scale to farm the bonus or to evade bans, costing large revenue as server costs climbed. Trust and safety, still under 20 people with crippled reactive enforcement, was redeployed again to play whack-a-mole.
- Collapse of the team. Dave Willner burned out and, with several staff, left within months. By year's end the team was dissolved, some members folded into a broader safety systems group under longtime researcher Lilian Weng.
- Their diagnosis. Some trust and safety staff felt that the endless Applied-versus-Safety fight, with its overemphasis on Doomerism, had taught leadership to discount every kind of "safety" concern, leaving a function weak at most tech firms even weaker at OpenAI.
- GPT-4's delay. Applied thought the six-month delay abundantly cautious; some in Safety thought it too short to finish testing and alignment.
- Hallucinations. Hard to control even with concerted RLHF. In November 2022, as users treated ChatGPT like a search engine and pundits speculated it could unseat Google, an internal document recorded hallucinations on about 30% of closed-domain questions.
Closed-domain questions are the easiest category: the user supplies the material (a PDF to summarise, bullet points to turn into sentences) and asks only about it. Open-domain questions (pop culture, ancient history, high-school biology) have no reference text, as with a search engine. Failing on roughly a third of the easy kind while the public used ChatGPT for the hard kind is the gap Hao wants readers to see.
Compute as the binding constraint
GPUs now limited everything; research projects and product launches were repeatedly delayed or shelved for lack of chips. Applied had promised to return the chips borrowed from Research by a set date; the date passed and, with users still growing, it needed more, not fewer.
The squeeze sped up work on cheaper models. Research found a new way to build Transformer models that were cheaper to serve, named DUST, and gave the resulting models desert code names.
| Code name | What it was | Fate |
|---|---|---|
| Sahara | Optimised GPT-3.5 | Released February 2023 as GPT-3.5 Turbo |
| Gobi | Optimised version of a text-and-image model | In development |
| Arrakis (Dune's desert planet) | Efficient GPT-4 | After months it could not keep performance while cutting cost, consumed heavy compute, and was scrapped to free GPUs |
The partnership under strain
Each side now feared the other could walk. Microsoft worried OpenAI could leave; OpenAI worried Microsoft would stop cooperating unless pleased.
- Arrakis as a setback. OpenAI had reworked its road map to put Arrakis ahead of more strategic projects, including applying GPT-4 to a search engine, to impress Microsoft; its failure disappointed senior Microsoft executives.
- Frenemies. Codex and DALL-E 2 had convinced OpenAI to sell directly; ChatGPT and GPT-4 showed it could make its own money. OpenAI was now pitching customers the same technology it handed to Microsoft, which pitched to the same customers.
- Asymmetric scale. As both release schedules sped up, a single OpenAI employee could be pinged by dozens of Microsoft counterparts per product, overwhelming staff trying to keep to their own road map.
- Murati's fix. Mira Murati worked with Kevin Scott and others on release timing, product differentiation and better ways of working; from summer 2023 a team of Microsoft engineers embedded inside OpenAI with full access to streamline technology transfer.
- "I need more." Altman and Nadella became directly involved in redistributing chips and money; Nadella told The New York Times that Altman called every day saying he needed more.
Serving ChatGPT needed more data centres, and training future models needed far bigger supercomputers. The two firms began sketching a single machine, called Stargate at OpenAI and Mercury at Microsoft, costing an estimated $100 billion to build. Hao closes on the parallel the next chapters pursue: like the empires of old, this one now needed more material resources and, above all, land.
One-sided sourcing. The cultural portrait rests on anonymous former employees, who skew toward the disaffected; those who thrived in the hiring wave are barely heard. Hao presents Brockman's medical story as shifting myth-making, but the October 2022 and post-firing versions may describe different episodes (acute abdominal pain versus a five-year chronic illness) rather than contradict each other. The 30% hallucination figure comes from one internal test whose design the book does not describe.
Stargate became a separate venture announced at the White House in January 2025 with SoftBank, Oracle and MGX, pledging up to $500 billion over four years, its first campus in Abilene, Texas. In October 2025 OpenAI and Microsoft signed a new deal under OpenAI's recapitalisation: Microsoft took about 27% of the new public benefit corporation, lost its right of first refusal on compute, and OpenAI committed to $250 billion of Azure purchases. ChatGPT reached 800 million weekly users by October 2025.
Plundered Earth
Chapter 12 follows the empire into the ground: the land, energy, water and minerals that generative AI consumes, and the communities in Chile, Uruguay, Iowa and Arizona that pay for it. Hao sets OpenAI and Microsoft's plan for five ever-larger supercomputer "phases" against Chile's long history as a supplier of raw materials, from Spanish conquest through the Chicago Boys to copper, lithium and now data centres, and then follows three grassroots fights (MOSACAT in Cerrillos, Daniel Pena in Montevideo, Resistencia Socioambiental in Quilicura) that forced Google and Microsoft to redesign or relocate. Her argument is that the "cloud" repeats the colonial pattern of taking resources from the periphery for the centre's benefit, and that the problem is not the infrastructure itself but the scale Silicon Valley demands.
Chile: a country built to export
The colonial template
- Geography. The Andes run the whole 2,600-mile length of Chile; north of Santiago the land becomes the Atacama Desert, the driest place on earth outside parts of Antarctica.
- Before the Spanish. Many Indigenous groups made the desert a home by tapping underground water and minerals to farm, herd and keep ancestral rituals.
- Conquest. The Spanish carved the region into bordered administrative units; elders' oral histories recall tongues being cut out for speaking native languages. Hao's key point is the role this fixed: Chile would supply land, water, energy and minerals to strengthen other nations' agendas.
- Today. Nearly 60% of exports are minerals, mostly from the Atacama: copper (the conductive metal in all electronics) and, more recently, lithium for batteries. Santiago families live by mining's rhythm, with workers commuting north for shifts and returning to the capital on days off.
The Chicago Boys
Chile never built another economic engine, and Hao assigns the United States a large share of the blame. In the 1950s and 1960s Chile and Uruguay embraced developmental economics: strong regulation and inward-looking industrialisation. American multinationals that had made billions from Chilean mines resented the rising taxes and restrictions, so Washington set out to reshape Chilean economic policy in their favour.
- The Chile Project (1956). A programme to educate a hundred Chilean students at the University of Chicago under Milton Friedman, later a Nobel laureate (1976), whose 1970 New York Times op-ed held that business's social responsibility is to increase its profits. Friedman stood for the opposite of developmentalism: no regulation, free rein for profit-seeking firms, and outward-facing growth through exports.
- Indoctrination, not education. Drawing on Naomi Klein's The Shock Doctrine (2007), Hao says Chilean (and later other Latin American) students were explicitly trained to attack their own countries' policies and the "fatal flaws" of developmentalism.
- Into power. Returning graduates, the "Chicago Boys," spread neoliberal ideas through the elite. In 1973 General Augusto Pinochet overthrew Chile's elected left-wing president in a coup the CIA helped foment, began nearly two decades of dictatorship, and appointed the Chicago Boys to write economic policy.
- Outcome. Chile privatised education, health care, pensions and even water. Growth followed, but so did extreme inequality: close to a quarter of national income goes to a few families in the top 1%, and the country, never industrialised, stayed tethered to extraction.
Ground zero for a new extractivism
As the AI boom arrived, Chile became the supplier of the industry's appetite twice over: copper and lithium in the north, and land, water and energy for data centres around Santiago. In May 2024 the government proudly announced 28 new data centres on top of 22 existing ones, bringing $2.6 billion of foreign investment.
Yet Chile also hosts some of the fiercest resistance anywhere: street protests and lawsuits that have stalled projects, caught the government's attention and inspired solidarity abroad. Martín Tironi Rodó, a professor at Catholic University and director of the think tank Futures of Artificial Intelligence Research (FAIR), sums up what Hao heard across the country: the movements ask how AI could develop without extraction, because building it the old way would "devastate the earth."
The physical cloud
From closets to megacampuses
Hao's starting point is blunt: "digital" is physical, and generative AI needs more and larger data centres than anything before.
- Consolidation. Data centres were once a few shelves in a back-office closet or a few dozen racks in a converted building. In the 2000s tech giants consolidated computing into vast rural warehouses, splitting the industry into hyperscalers and everyone else. The four largest (Google, Microsoft, Amazon, Meta) now spend more each year than almost all other builders, such as Equinix and Digital Realty, combined.
- Campuses. Mél Hogan of Queen's University (Canada) used to compare data centres to football fields; she says that no longer conveys the size. Hyperscalers call them "campuses," rivalling the largest Ivy League universities, packed with heat-spewing racks and giant fans, air conditioners or evaporative water systems. The din can be heard for miles, 24 hours a day, a punishing form of noise pollution, especially in poorer areas.
- Megacampus. The post-ChatGPT term measures energy as much as land.
- 3×. Power drawn by a rack of GPUs versus a rack of other chips.
- 10×. Electricity per ChatGPT query versus a typical Google search, per the International Energy Agency.
- 150 MW. Design size of yesterday's largest data centres, about the annual energy of 122,000 US households.
- 1,000-2,000 MW. Planned AI megacampuses; one could use the energy of 1.5 to 3.5 San Franciscos a year.
- 3% → 8%. Share of US power going to data centres, 2022 versus projected 2030; global AI computing could outuse India, the world's third-largest electricity consumer.
- 1.1-1.7 trillion gallons. Fresh water AI demand could consume globally a year by 2027 (UC Riverside), about half the UK's annual use.
- One-fifth. US data centres already drawing water from moderately or highly stressed watersheds before the boom.
- ~4 and 242. Smartphone charges equal to generating 1,000 texts and 1,000 images with open models (Luccioni and Jernite); one image is about a quarter of a charge.
Power: the fossil comeback
Few places can deliver that much power to one site, so developers are working with utilities worldwide to build plants. After a decade of flat US demand, Goldman Sachs called the surge electricity growth unseen "in a generation." Utilities are delaying the retirement of gas and coal plants and slowing the move to renewables, and Hao writes that Microsoft restarted Three Mile Island, site of the late-1970s partial meltdown, the worst commercial nuclear accident in US history (see the critical lens below for the correction).
Less disclosure, louder counter-narratives
As the footprint grew, corporate obfuscation reached new heights. Since Emma Strubell's carbon paper and its citation in "Stochastic Parrots," the giants have hidden even more technical detail, making footprints hard to estimate, while pushing three counter-narratives to the public and policymakers:
- Efficiency will solve it. Data centres will become so efficient that impact stops mattering. Highly misleading: efficiency gains are swamped by scale.
- Generative AI will unlock climate innovation. The most pernicious claim. The AI that helps the climate, as catalogued by Climate Change AI (Luccioni is a founding member), is supervised learning, anomaly detection and statistical time-series models: small, efficient, sometimes laptop-sized, and almost never generative.
- AGI will solve climate change. Impossible to prove, so not worth debating.
Sasha Luccioni, research scientist and climate lead at Hugging Face, concludes that generative AI carries a disproportionate energy and carbon footprint with little environmental upside. Her former collaborators inside closed companies can no longer get permission to co-author environmental papers with her, so she works with outsiders such as Strubell. With Yacine Jernite she measured open-source models as a proxy for closed ones (the figures above); hers and Strubell's are among the few quantified measures left. She once asked 500+ authors of recent machine-learning papers for basic training information and got almost no answers, or was told it was confidential.
Reliability, water and developer tactics
- Renewables don't fit. Microsoft executives admit internally that intermittent renewables cannot serve facilities that must run 24/7. Uptime matters so much that Google, Amazon and lately Microsoft build campuses in threes, a backup for the backup. During Hurricanes Irma (Florida) and Harvey (Texas), as millions lost power and hospitals evacuated, data centres hummed on; one sheltered its staff's displaced families.
- Minerals. Megacampuses also need vast amounts of copper and lithium for computers, cables, power lines, batteries and backup generators.
- Potable water. Cooling often requires drinking-quality water, clean enough not to clog pipes or breed bacteria. The burden falls hardest on vulnerable communities, especially in the Global South.
- Playbook against resistance. Developers enter towns through shell companies, donate to community programmes to soften opposition, and make sustainability promises they walk back once ground is broken and projects are hard to reverse. In Virginia, residents fighting several huge data centres uncovered a lawyer's email advising a developer to plant "a mole or 2" in their group.
OpenAI's five phases
From the moment it committed to scaling, OpenAI sought unprecedented compute; Brockman told Hao in 2019 that in AI whoever has the biggest computer gets the most benefits. Altman's plan with Microsoft had two strands: dozens of research supercomputers with Nvidia chips for exploratory work, and a series of ever-larger training supercomputers, one per model generation, which Altman called "phases." He once showed staff a slide in which Phase 5 dwarfed the rest.
| Phase | Site | Hardware and cost | Names (Microsoft / OpenAI) | Use |
|---|---|---|---|---|
| 1 | West Des Moines, Iowa | 10,000 V100s | Odyssey / Owl | GPT-3 |
| 2 | Iowa | 18,000 → ~25,000 A100s | Telemachus / Raven | GPT-4 |
| 3 | Outside Phoenix, Arizona | Hundreds of thousands of H100s, several billion dollars | Inglewood / Whale (split into Beluga, Narwhal, Orca) | Planned GPT-4.5 and GPT-5 |
| 4 | Wisconsin | B100s, about $10 billion | n/a | Next generation |
| 5 | Undecided | ~$100 billion, up to 5,000 MW | Mercury / Stargate | Speculative |
- Naming. OpenAI's Research division named clusters alphabetically after animals, planning to switch to periodic elements by atomic number after Z. Telemachus, son of Odysseus, followed Microsoft's Odyssey.
- Iowa. Microsoft entered West Des Moines in 2012, winning officials over with what the then mayor called "staggering" tax payments that funded public infrastructure, plus about $2.5 million over a decade for local programmes, mostly nonprofits. Phase 2 began with the 18,000 A100s of the 2021 road map and grew to about 25,000 by the end of GPT-4's training.
- Arizona. Cheap land, tax breaks and proximity to California made it every cloud provider's favourite. Microsoft cultivated two underdeveloped cities outside Phoenix, bought three tracts in 2018-2019 totalling nearly 600 acres (more than 450 football fields), again donating to quiet residents, and added 283 acres in 2024. Each tract would host a campus serving Azure customers and OpenAI in the "West US 3" region.
- Cost escalation. Phase 4's $10 billion compared with $1-2 billion for the costliest hyperscale data centres of the time; Altman casually floated $100 billion for Phase 5.
- ChatGPT's drag. After ChatGPT, so many chips were tied up serving customers that Microsoft struggled to finish Phase 3; Whale was split into Beluga, Narwhal and Orca, with build-out due in 2024, and Altman tempered expectations.
Phase 5: running out of land and power
Nobody at either company knew if Phase 5 was technically possible. Per The Information, the design could need up to 5,000 MW, close to the average demand of New York City. It made little sense as a business investment, but money was not the binding constraint: energy was. An OpenAI employee put it as running out of land and power. Both firms assumed some innovation would be needed:
- Split the machine across several campuses and learn to train one model across distant sites; or
- Fusion, Altman's favourite answer. He gave staff upbeat updates on Helion Energy, his largest personal investment, from which Microsoft had agreed to buy power once a 50 MW plant ran. The Wall Street Journal later reported OpenAI and Helion were also negotiating a deal, with Altman recused.
Altman and other executives never raised the environmental toll at company-wide meetings. Iowa was two years into drought as GPT-4 trained there for three months; the Associated Press found Microsoft's data centres used about 11.5 million gallons, 6% of the district's water, in a single month (Microsoft says it aims to improve water efficiency 40% over a 2022 baseline and be water-positive by 2030). Arizona: as Phase 3 groundwork began in 2022, Nature Climate Change reported the Southwest's worst drought in over 1,000 years; with mismanagement it has drained the Colorado River, on which Arizona and six other states rely, so low it could stop flowing. Water and power crises compound, since the region draws on Hoover Dam hydropower and water-cooled nuclear plants. In 2023 Phoenix set heat records and heat deaths rose at least 30% to 600+; state water director Tom Buschatzke said everything was converging in a challenging direction. What Altman did raise was impatience: when Meta announced in March 2024 that it would run 350,000 H100s, more GPUs than OpenAI had, he felt Microsoft's slowness was costing OpenAI its edge.
Copper, lithium and the Atacameños
Sonia Ramos and the mine
Sonia Ramos grew up in a mining family; her father worked for an American copper company. In 1957 part of the Chuquicamata mine collapsed, killing several people and injuring dozens. Her father survived, but she watched affected families sink into poverty and children starve, with no ceremony, commemoration or compensation. Four decades later, as one of Chile's most outspoken Indigenous critics of mining, she drew the lesson: mining is a system that, left alone, seeks profit at any cost; in it, she says, the worker does not exist.
What extraction has done to the north
- Copper. Chile is the world's largest producer, about a quarter of global supply. Chuquicamata is the largest open-pit copper mine on earth, deepened by regular blasting; its spoil heaps are slowly burying an abandoned town. Processing drains water: one foreign multinational emptied an entire salar (salt-flat) basin, destroying its ecosystem.
- Invisible harms. Arsenic in air and water has raised cancer rates across the north.
- Social collapse. The Atacameños (the umbrella name for the region's distinct Indigenous groups) can no longer farm or herd on depleted land. Their towns face deep poverty, rising crime, depression, alcoholism and delinquency, and shortages of food, running water, health care and schooling, despite billions in profits from their territory. Many now work for the industry that took their land and use clinics it sponsors; once-united groups squabble over shrinking resources.
- Lithium. Found in the 1960s by an American company drilling the salares for copper-processing water, it floats in an oily brine. Chile produces about a third of world supply, second to Australia, mostly from the Salar de Atacama: brine is pumped into turquoise pools and left for the sun to evaporate. The pink flamingos, which the Atacameños regard as spiritual siblings, are gone; in the Peine community a leader's daughter knows them only from stories and a plush toy.
The recycled promise of a better future
In 2022, as the European Union set energy-transition policies and lithium demand soared, companies and politicians hailed Chilean mining as the path to a better future. Indigenous communities asked: better for whom? Microbiologist Cristina Dorador, who studies the north's biodiversity, says locals never get to decide their destiny apart from global economics and politics. Now the same story is told about AI: more copper and lithium for megacampuses, power plants and thousands of miles of power lines means progress, so blocking extraction means blocking humanity.
Ramos does not oppose mining: her ancestors were miners and discovered the copper. Her objection is to scale, which has made the north and the whole country dependent on one industry, prevented other economies from emerging, and choked off the ability to imagine development without plunder. Through generative AI it has even fed racist stereotypes: a 2023 exhibition in Brazil, co-produced by a Chilean university, contrasted Latin America's rich Indigenous cultures with the primitive, technologically inept peoples that Midjourney and Stable Diffusion depicted.
Resistance in the desert
- Protest. Atacameños fly black flags on their houses and physically block the roads used by mining buses and trucks.
- Law. They hire lawyers to assert rights under international Indigenous law, which protects cultural and territorial sovereignty. Forced into negotiations, they demand government research into ecosystem health and a reckoning of water loss and irreparable damage.
- Science as defence. Ramos's foundation unites "the ancestral and non-ancestral" to research the desert's unique microbes (potential medicines or energy sources) and its long use as a Mars analogue, hoping discoveries prove the land is worth more intact. Against high-speed progress she seeks ideas of progress built on healing and regeneration.
Santiago: Google's give-back and the Cerrillos fight
The hyperscalers' spending spree
As hyperscalers outgrew land and power in their usual regions, they pushed harder into new territories. Microsoft spent $55 billion+ in fiscal 2024, nearly a quarter of revenue, on what SemiAnalysis called the largest infrastructure build-out in human history; Google told investors in its Q3 2024 call it would spend about $50 billion that year; Meta expected up to $40 billion, rising the following year.
The "Quilicura Urban Forest"
In June 2024 council member Alexandra Arancibia took Hao to what she sees as the symbol of the power gap between US tech giants and her municipality, Quilicura, under thirty minutes from Santiago's café districts but lined with crumbling roads and illegal dumps run by a local mafia. Past a pet cemetery sits a scrubby, parched plot with a few starved trees and a purple sign: the Quilicura Urban Forest, begun by Google in 2019 as a give-back for hosting its data centre, with a cartoon of the forest cleansing industrial smog.
- Google's version. Its country spokesperson declined an interview, sent polished briefing materials describing residents enjoying the space, and said each data centre gets a community impact programme (education, sustainability, internet access, health); Quilicura's has received $1.2 million+. She later said the forest was being updated to "evolve the experience."
- Hao's observation. There were no residents: no bus line, no homes nearby, a plot too small even to hold the data centre, and about a dozen stray dogs nosing through rubbish. Arancibia asked, wryly, whether it felt like Silicon Valley.
How Quilicura became a dumping ground
Arancibia realised in college, commuting past mounds of refuse, that Quilicura was a place where things are discarded. Two decades earlier it was pastures and wetlands rich in birds and plants. Then a trash mafia let anyone in Santiago dump waste there for a fee (old dumps now look like grassy, deformed hills), followed by beer companies and real estate developers who took land and wetland water. Only small green pockets remain in the 22-square-mile municipality.
- Google arrives (2012). Its first Latin American data centre, operational in January 2015, billed online as among the continent's most efficient and green. Nobody paid attention; bus commuters assumed it was a beer or food factory providing jobs.
- Few jobs, little benefit. Like most data centres it employed few people after construction. A 2024 posting for a mechanical technician, one of the rare long-term roles, appeared only in English and refused résumés in any other language. Nearby public schools still lack decent internet or devices.
- Chile as digital hub. In 2019 Google announced a second Santiago-area data centre; Microsoft and Amazon followed. The government sold Chile as a stable haven for foreign investment in an unstable region, and in 2020 announced, with Google's backing, an undersea cable linking Asia-Pacific to the Americas through the central coast.
MOSACAT reads the filing
The second site was Cerrillos, another neglected working-class municipality. From the 1930s to the 1990s a Belgian company's cement (asbestos) factory poisoned it, what one historian called Chile's largest "industrial genocide"; cancer rates remain high. But in a country of privatised water, Cerrillos runs Chile's only public water service, supplying neighbouring communities and, in emergencies, other regions. Neglect plus a precious water source bred tough watchdog groups.
- The collective. MOSACAT formed in 2019 from women's, housing, workers' and environmental activists, many of whom met fighting an illegal mining project they shut down and had declared a nature reserve. A friend, now a member of Congress, tipped them off about Google.
- The filing. When Google filed with the environmental agency in summer 2019 (a near rubber-stamp process), a dozen-plus volunteers read all 347 pages of diagrams and jargon in time stolen from jobs and families, and taught themselves data-centre cooling. Member Tania Rodríguez showed Hao the spiral-bound copy bristling with Post-its marked agua potable. Hao reads the tome as a physical symbol of how Silicon Valley uses technical knowledge to justify centralised decisions.
- The number. The facility planned to draw about 169 litres of drinking water per second for cooling, straight from Cerrillos's public source, during a megadrought then nine years old. The book compared this with the annual consumption of all ~88,000 residents; that comparison was later corrected (see the critical lens).
- Fast-tracked. Water projects in Chile usually take years to approve; Google's sailed through amid drought-related water emergencies.
MOSACAT first approached Google's local partner Dataluna, a Chilean investment firm whose representatives seemed not to understand the project and denied it used fresh water. The mayor and council, also briefed by Dataluna, believed only wastewater was involved; once MOSACAT corrected them they demanded answers, and the matter climbed from Dataluna to Google Chile to Mountain View. In October 2019 Google sent two engineers and a lawyer. MOSACAT lined their route with protest signs; about two dozen residents from seven groups attended. The engineers spoke only English, gave a highly technical talk, and the lawyer translated (English-speaking members heard him mistranslate). Google offered an urban forest like Quilicura's. Council member Alejandra Salinas of neighbouring Maipú called it intimidation: "They come offering us trees while drying out our earth." The Estallido Social erupted that same month (protests over unemployment, privatisation and inequality that left thousands injured and dozens dead). As towns held reform referendums, MOSACAT added a question on Google's water use to Cerrillos's ballot, leafleted and knocked on doors, and in December 2019 won by a slim majority. In 2020 the government joined its suit in an environmental court. Google grew friendlier once it realised residents understood English, but as the project stalled and its compute needs grew it moved its next Latin American data centre to another country. In 2024 the court ruled Google could not build a water-using facility in Santiago; Google says it may later seek permits for an air-cooled one.
Salinas frames the residents' position precisely: Chile knows it feeds the world copper and lithium and is willing to share, but not to have its water taken and be left with nothing. What they wanted was not a park but to be treated as participants rather than bystanders. Hao adds the structural catch, voiced by Yale's Julian Posada and lawyer Mercy Mutemi: when a Global South country pushes back, a Silicon Valley company can simply take its money elsewhere.
Uruguay: drought or pillage
Two kinds of data centre
Uruguay (3.4 million people) runs on three data centres owned by state telecom Antel, totalling only about 5,000 m². One, under 1,000 m², sits on a residential Montevideo street on two dedicated power lines; 30% is servers, 70% offices and plant rooms. It is air-cooled, and its hum bothers two neighbouring families enough that they knock on the door. Manager Javier Echeverria has already quietened the cooling and is working with researchers on further fixes: responsiveness Hao contrasts with what MOSACAT endured.
Thirty minutes away, the Parque de las Ciencias free-trade zone, nicknamed bitterly "zona America" for its mostly American, untaxed tenants and its National Mall-style lawns and fountain, advertises stability, land and plentiful power and water. In 2021, as GPT-3 spurred scaling, Google bought 29 hectares there, 58× Antel's entire data-centre footprint, as the new home of its second Latin American data centre.
The drought and who bore it
- Scarcity. Uruguay was in severe drought: farmers lost whole harvests, over $1 billion in agricultural losses, and by summer 2023 Montevideo was mixing salt water into the tap supply. Brownish, chemical-smelling water drove those who could afford it to bottled water; those who could not drank it anyway and suffered stomach pain, rashes, worsened conditions and a rising rate of miscarriages.
- Poverty after the pandemic. As Google and Microsoft each topped $2 trillion in market value, partly on pandemic cloud demand, Uruguay's illegal settlements grew by orders of magnitude and ollas (community soup kitchens) turned children away. Fabiana, called "Reina Madre," who runs one from a settlement and swept a brothel's floors at seven to survive, ranks those years among the worst of her life.
- Misallocation. More than 80% of fresh water goes to industry, chiefly multinational soy, rice and pulp-tree plantations that export their output, deplete soil, and use so much fertiliser that Uruguay is among the top per-capita fertiliser users, with high cancer rates. During the drought industry kept drawing from the Río Santa Lucía, which also feeds Montevideo's supply. Two decades ago Uruguay became the first country to write water into its constitution as a human right; in the crisis it was drinking water, not industrial water, that was cut hardest.
Daniel Pena's lawsuit
Daniel Pena, a sociologist at the Universidad de la República who tours the countryside in a battered pickup documenting how industry squeezes farmers and the poor, links it all to Uruguay's colonial history: foreign multinationals still outrank locals.
- Secrecy. Scanning the environment ministry's project list, he found Google's proposal with its water figures marked confidential. A public-information request (he had won about twenty before) was refused as proprietary.
- Constitutional suit. Worried about the precedent for every cloud company to follow, he invoked the constitutional water clause and sued with a lawyer friend working pro bono.
- Victory (March 2023). Four months later he won. The plan was to take about 2 million gallons a day from the drinking supply, the daily use of 55,000 people. With salt water soon flowing from Montevideo taps, thousands protested against Google and the industries draining the country; in June 2024 walls still read that this was not drought but pillage.
- Retreat. Near the end of 2023 Google quietly switched to waterless cooling and cut the facility to a third of its size. But the government now treats its energy consumption as a commercial secret.
- Full supply chain. Pena petitioned the ministry, with 400+ signatories, for an impact study covering the whole chain: where hardware minerals are mined, how labour is treated, carbon emitted and e-waste disposed of. Most of those harms land elsewhere, but almost always in the Global South, so the same consequences arrive through different links of the chain.
Pena's diagnosis matches Ramos's: the problem is not the infrastructure but the scale. Scale explains why Google and Microsoft expanded into countries short of resources, why Google needed 58× Antel's land, and why it answered to locals so little. He calls them extractivist projects that come for cheap water, tax-free land and poorly paid jobs and leave without improving internet access.
Microsoft in Quilicura and a Chilean alternative
A changed Chile
Microsoft was a late bloomer in data centres; Google had far more before the OpenAI deals, so Microsoft copied Google's playbook into the same regions. In 2022, shortly after its second OpenAI investment, it settled on Quilicura. But Chile had changed:
- Constitutional experiment. After the Estallido, citizens took part in rewriting the Pinochet-era constitution; two drafting processes produced two partisan texts that both failed, yet revived democratic hope among young and working-class Chileans.
- Boric. The left surge elected Gabriel Boric Font, a former student protester, at 35; in his victory speech he repeated the protest slogan that if Chile was neoliberalism's cradle it would also be its grave. He took office in March 2022, and young activists built a new generation of progressive organisations.
- Resistencia. Arancibia and Rodrigo Vallejos, who met organising the Estallido, founded Resistencia Socioambiental Quilicura, drawing on the Latin American idea that the social and environmental are inseparable.
The waterless-cooling question
Vallejos, a law student at Universidad Diego Portales, taught himself data-centre engineering late at night. Microsoft projected far less water than Google had in Cerrillos, but the drought was worsening (expected to last to 2040) and Quilicura's wetlands were shrinking from desertification as well as industry. Microsoft's website boasted of waterless cooling, so why not use it here? In an article Vallejos called it striking that so reputable a company preached green values while falling short of global standards in a "third world country." Microsoft replied the technology was still under development and piloted only in the US; Vallejos asked why it was advertised at all.
Outside allies and a speculative design
Marina Otero Verzier, research director at Rotterdam's Nieuwe Instituut (the Dutch institute for architecture, design and digital culture), and FAIR researchers Serena Dambrosio and Nicolás Díaz Bejarano were moved by Vallejos and Arancibia's passion and exhaustion: they had read Microsoft's documents, written critiques and protested, yet could not get a hearing from the company or the government.
- Borrowed prestige. Otero used her Harvard and Columbia affiliations to command attention, eventually quitting her job to campaign: speaking at major conferences about data centres and Resistencia's fight, connecting with the Ministry of Science, Technology, Knowledge and Innovation and with Microsoft and Google representatives, and linking the activists to international researchers.
- Data centres as libraries. All three trained in architecture. Díaz, who visits national libraries on his travels, reasoned that data centres also store a society's memory and knowledge and could be welcoming and beautiful. He calls the companies' give-back programmes "schizophrenic": disconnected from how facilities actually affect communities.
- The workshop. With Vallejos and Arancibia they won funding for a 14-day workshop in which architecture students from across Santiago reimagined a data centre for Quilicura, then presented to residents. Otero was struck by how much knowledge the community had to offer.
The corporate model. Community impact programmes chosen by the company: an urban forest nobody can reach, donations to local nonprofits, polished briefings. The facility itself stays a sealed, brutalist box whose water and power use is hidden or confidential.
The Quilicura students' model. The facility as public architecture: cooling water stored in visible pools residents can use; a "fluid data territory" where suspended walkways let people cross the wetland, plant nurseries and nesting stations sit among server rooms, polluted wetland water is purified before use and returned, and the computers monitor the wetland's recovery. Díaz admits it fixes nothing, but it imagines other relations between data and water.
What the coalition achieved
Three years into a four-year term, Boric faced pressure for economic "quick wins": more mining and the 28 data centres. When he announced a national data-centre strategy, Vallejos texted Hao the press-conference video with a dizzy-face emoji. Yet Hao credits the coalition of northern and Santiago activists with domestic and foreign researchers with real gains:
- A seat at the table. The Ministry of Science, which drafts the plan, formed for the first time a committee of activists it consults regularly, inviting Vallejos, Arancibia and MOSACAT.
- A changed tone. The ministry's AI bill no longer treats AI as a universal good; it now acknowledges social and environmental costs.
- Building on its own terms. Minister Aisén Etcheverry says technology reflects a cultural and historical framework; the global internet was shaped without Chile, but AI need not be.
- Decolonial AI. Tironi goes further: today's AI industry rests on a colonial ideology that imposes its definitions of AI, good AI and an AI industry on everyone else. Chile, which knows what it is to be hollowed out under the banner of progress, could lead in imagining decolonial conceptions of AI from its position as a supplier of materials.
Hao ends without false comfort: a noble ambition, facing mighty forces.
- Martín Tironi Rodó. Catholic University professor, director of FAIR; argues for decolonial AI.
- Sonia Ramos. Indigenous Atacameña activist; against the scale of mining, not mining itself.
- Cristina Dorador. Microbiologist in the north; locals never choose their own destiny.
- Alexandra Arancibia. Quilicura council member; co-founder of Resistencia.
- Tania Rodríguez and Alejandra Salinas. MOSACAT members; Salinas is a Maipú council member.
- Daniel Pena. Uruguayan sociologist who sued for Google's water figures and won.
- Rodrigo Vallejos. Law student, co-founder of Resistencia; challenged Microsoft on waterless cooling.
- Marina Otero Verzier, Serena Dambrosio, Nicolás Díaz Bejarano. Architects and researchers behind the Quilicura workshop.
- Aisén Etcheverry. Minister of Science; opened drafting to activists.
A headline number was wrong. In November 2025 Hao acknowledged that the book's claim that Google's Cerrillos facility could use "more than one thousand times" the town's water was off by about 1,000×, because of a unit error in an official document; the planned draw (169 L/s, roughly 5.3 billion litres a year) was closer to Cerrillos's own annual consumption, still large in a megadrought. She promised a fix in later printings. Other figures need care: the IEA's 10× per-query estimate rested on early 2023 assumptions, and later disclosures (Altman: 0.34 Wh per average query, June 2025; Google: 0.24 Wh per median Gemini prompt, August 2025) imply far less. Microsoft did not itself restart Three Mile Island: Constellation is reviving Unit 1 (the 1979 meltdown was Unit 2) under a 20-year power deal signed in September 2024. Hao's case on scale and local costs survives these errors, but critics such as Andy Masley argue the chapter overstates AI's water share relative to agriculture and industry, a point her own Uruguay section (80% of water to industry) partly concedes.
Chile launched its National Data Center Plan in December 2024; Google suspended Cerrillos, and Microsoft opened its Santiago cloud region in June 2025, pledging about $3.3 billion in Chile through 2029. Chile's data-centre power demand is projected to rise from 325 MW to 1,360 MW by 2032. In Uruguay, Google broke ground on the reformulated, air-cooled Canelones site in August 2024, investing over $850 million. In Arizona, Tucson's council killed the Amazon-linked "Project Blue" data centre in August 2025 after a water backlash. The IEA's 2025 Energy and AI report projected global data-centre electricity roughly doubling to about 945 TWh by 2030.
The Two Prophets
Chapter 13 shows Altman at the height of his influence in 2023: charming Congress, touring the world, and steering Washington toward licensing hypothetical "frontier" models above a compute threshold rather than regulating the harms AI was already causing. The two prophecies of the title, AI as salvation and AI as doom, hardened into rival camps inside OpenAI, with Ilya Sutskever preaching both, while a board thinned to three independent directors found its CEO increasingly opaque. Hao's judgment is that Doomer fears and corporate interests converged on a policy agenda that looked like safety but shielded the incumbents from accountability.
The Senate performance
Testimony on OpenAI's terms
In May 2023 Altman testified to the Senate. Hao calls it a remarkable performance: he restated that AGI would solve climate change and cure cancer, argued OpenAI's technology would improve jobs and create "fantastic" new ones, dodged questions on copyright and on the opacity and privacy of training data, and made a sincere-sounding call for regulation, but regulation OpenAI could bless, invoking China to warn lawmakers against slowing innovation. Senators loved him.
- Licensing agency. A new body to license models above a capability threshold, which he later said could be approximated by a compute threshold if necessary.
- Safety standards. Tests for "dangerous" capabilities, such as the ability to manipulate and persuade, or to produce recipes for novel biological agents.
- Independent audits. Outside checks of compliance with those standards.
Hao's key point is the redirection: the proposals moved the conversation off existing problems (labour, environment, intellectual property) and onto future systems and extreme risks.
- The Kennedy exchange. Louisiana senator John Kennedy asked whether Altman could run such an agency. Altman joked that he loved his current job, offered to send recommendations, and when asked about his pay said he earned enough for health insurance and held no equity in OpenAI.
- Even Marcus melted. Gary Marcus, seated beside him, told the committee Altman's sincerity was physically apparent in person. He later reversed himself, concluding that he, the Senate and the public had probably been "played."
- The campaign behind it. After ChatGPT, Anna Makanju's small policy team was flooded with requests. For months, with or without Altman, it dined with, demoed for and briefed senators, House members, staffers, cabinet members, diplomats, agency heads and Vice President Kamala Harris, pitching the same proposal Altman gave in testimony. By early June he had personally met at least 100 lawmakers (New York Times), some of whom cited those private talks at the hearing. His prep team counted it a resounding success.
The same day a small band of Hollywood concept artists (designers of characters, props and visual elements) had meetings on the Hill. They had crowdfunded their travel and been threatened and doxed by trolls for speaking out. As Hollywood writers, and soon actors, struck partly over AI, the artists came to describe generative AI trained on millions of their works without consent and now replacing solid middle-class jobs, potentially hundreds of thousands. Karla Ortiz (concept artist on Marvel's Doctor Strange, filer of the first artist lawsuit against generative AI firms) says no one is getting work. Altman's testimony bumped several meetings to the next day, leaving them pacing outside the hearing room. That second evening Altman dined with 60 House members at the Capitol over an expert buffet of roast chicken, while the artists hosted staffers with what their budget allowed: wine and Chick-fil-A. Hao offers it as a darkly comic picture of who held influence.
The China card
The narrative Altman had long used inside OpenAI to justify secrecy and speed was now steering US regulation toward proposals that would avoid holding OpenAI accountable and might entrench its position. Silicon Valley's "What about China?" card had always warded off regulation; with Washington's China fears inflamed by TikTok, it was stronger than ever.
Export controls and their limits
The Department of Commerce led the offensive. On 7 October 2022 it used export controls (limits on technology sales abroad on national-security grounds) to cut off, without warning, exports of advanced American-designed AI chips, mainly Nvidia's, to China. Meant to blunt Chinese military AI, it hit Chinese science and industry broadly, from AI health and education to ChatGPT equivalents; a semiconductor analyst said such rules would once have been considered an act of war. Commerce was frustrated that China slowed only modestly, for three reasons:
- Nvidia's workaround. China was a huge market, so Nvidia quickly designed new chips that fell just inside the rules.
- A boost for Chinese chipmakers. Demand for domestic alternatives brought them funding and feedback.
- Open-source models. The biggest obstacle was the cross-border open-source movement that replicated closed models and posted them online. Meta, after catching up, became the dominant open publisher: chief scientist Yann LeCun believed in open science, and it was smart business, since Meta does not sell models but could build AI leadership, attract talent and taunt rivals who do. With compute roughly matching OpenAI's via Microsoft, it built the Llama series. Llama was not truly open source (that would require releasing training data as well as weights), but free weights alone made it, despite a policy against direct availability in China, a critical building block for Chinese AI.
The frontier white paper and the Forum
In July 2023, two months after the hearing, a 51-page white paper by a consortium including researchers from OpenAI's Safety clan, Microsoft and Google DeepMind plus more than a dozen think tanks, many Doomer-linked, echoed Altman: a licensing regime keyed to compute thresholds and safety evaluations for dangerous capabilities (manipulation, persuasion, novel bioweapon recipes). It also named the category: "frontier" AI models, which by the authors' account did not yet exist but could emerge suddenly and unpredictably as companies like OpenAI and Anthropic scaled.
Within weeks OpenAI, Microsoft, Google and Anthropic launched the Frontier Model Forum to advance research and shape the safety policy agenda. Hao calls it a rare alignment of Doomers, Boomers and corporations: keeping frontier models at the centre of regulation was ideologically essential for Doomers and convenient for Boomers and companies, because it diverted attention from regulating existing models and their problems; all of them also gained from arguing against open weights. Meta stayed out for a year, then joined to get a seat at the table.
The arbitrary number
Where 10^26 came from
The frontier idea tied scale to emergent, possibly dangerous capabilities, combining scaling-law philosophy (more compute predictably yields more powerful models) with the Doomer belief that advanced AI could go rogue. Regulation would follow compute: cross a line, face extra caution.
The paper's line was 10^26 floating-point operations of total training computation (1 followed by 26 zeros). The authors admitted it was somewhat arbitrary, chosen as a level no existing model had likely passed. Coauthor Sara Hooker, VP of research at Cohere, came to believe from talking with those who proposed it that it was set just above GPT-4's reported training compute. Lacking consensus among coauthors on the framework or the number, they buried it in a footnote and appendix, unjustified and heavily caveated. What shocked Hooker was how fast it became one of Washington's most popular proposals, because frontier models sounded frightening, more so in Beijing's hands. Emily Weinstein of CSET told Hao in late 2023 that parts of the administration were grasping at anything because they wanted to do something.
The frontier camp. Scaling laws mean more compute brings more powerful and possibly emergent capabilities, so models above a compute line deserve extra scrutiny. Compute is measurable and lets regulators act before dangers appear.
Hooker, Raji and critics. Scale can produce capability, but capability does not need scale: a small model trained only on high-quality biological data can be a strong generator of biological recipes, and distillation (named in OpenAI's own 2021 road map) shrinks big models while keeping their abilities. Scaling does not guarantee particular capabilities either; that depends on training data and architecture, and not every model is a Transformer. So compute barely correlates with risk, let alone with specific risks.
The road not taken: data disclosure
Commerce embraced the paper and began exploring something unprecedented: extending export controls from hardware to software, banning export of models above a compute threshold, in effect blocking model weights from being posted online.
Meanwhile a recommendation that Marcus and IBM VP Christina Montgomery repeated throughout the hearing went nowhere: force companies to disclose what is in their training data. It would hand Beijing nothing, yet would give teeth to accountability on copyright, privacy and honest capability claims; without it, Marcus noted, we must take companies' word for performance. It would also tackle the emergence worry directly, since behaviour derives first from data. Sarah Myers West, co-executive director of the AI Now Institute and former FTC adviser, puts it simply: a model produces bioweapon recipes because it was trained on bioweapon information. Opening the data is the first step toward knowing which inputs produce dangerous outputs.
Closed versus Open
News of Commerce's deliberations (formally opened for public comment in early 2024) split the AI and policy communities in two: techno-nationalism versus borderless science.
- Closed. Frontier Model Forum members, the broader Doomer community, and soon the US national-security and intelligence apparatus, repeating arguments OpenAI executives had used internally for years.
- Open. Meta, open-source developers, startups, civil-society groups and independent academics.
The Open side's case against locking up weights had four strands:
- Recipes are not the bottleneck. Weinstein notes bioweapon recipes are already easy to find via Google; the hard part is obtaining materials and equipment, which closing models does not touch.
- Open source built US tech. Shared code is the bedrock of American startups; restricting weights leaves small developers fewer paths and entrenches the Forum's giants.
- Scrutiny needs access. Work like Luccioni's, Jernite's and Strubell's on AI's carbon costs depends on open models.
- US-China collaboration strengthened America. The two countries produce the most AI talent and are each other's top research partners. ResNet, among the world's most used networks, came from Chinese researchers at Microsoft's Beijing lab; it underpins vision, speech and language systems and was a core ingredient of DeepMind's first AlphaFold (2018), which predicts protein structure from amino-acid sequences. (Later AlphaFold work, on a different network, won Demis Hassabis and a colleague the 2024 Nobel Prize in Chemistry.)
The threshold goes viral
In October 2023 the Closed side won its biggest endorsement: the Biden executive order on AI, one of the longest ever, which read like two documents stapled together, because it was.
- Document one. Rooted in the White House's 2022 Blueprint for an AI Bill of Rights, built through consultation with civil society: broad participation, attention to discrimination in health care and hiring, protection from data collection without consent.
- Document two. A faithful copy of Altman's recommendations and the white paper's frame, attached at the last minute after it caught the eye of a few Doomer sympathisers in the White House. As the paper urged a focus on models that did not yet exist, the order gave half its length to them. Of the paper's four dangerous capabilities it kept three: CBRN (chemical, biological, radiological, nuclear) weapon recipes, automated cyberattacks, and, copied verbatim, evasion of human control "through means of deception and obfuscation." To the alarm of Hooker, Raji and others, it adopted 10^26 as the line above which models must be reported to the government.
| Document | Date | Threshold |
|---|---|---|
| Frontier AI white paper | July 2023 | 10^26 FLOPs (footnote) |
| Biden executive order | October 2023 | 10^26, report to government |
| EU AI Act | End of 2023 | 10^25, stricter |
| California SB 1047 | Early 2024 | 10^26; vetoed by Gavin Newsom |
The approach then metastasised. EU lawmakers, steamrolled by generative AI while finishing the long-gestating AI Act, grabbed a slightly stricter 10^25. California's SB 1047 returned to 10^26. Ironically OpenAI and other developers lobbied hard against SB 1047 for its many other accountability provisions, and Newsom vetoed it. Many accused him of bowing to industry; Hooker and Raji welcomed the veto as a step back toward scientific consensus, since nobody had answered why that number or which risks it would prevent.
Atrophied expertise
For Hao the sequence (testimony, white paper, lobbying blitz, China panic, hasty thresholds at home and abroad) shows how far independent AI expertise had atrophied. In September 2023, at Chuck Schumer's first AI Insight Forum, Deborah Raji was the only academic without financial ties to industry or the Doomer community, seated among Altman, Musk, Nadella, Gates, Zuckerberg, Pichai and Jack Clark. As they made spectacular, unsupported claims about AI's promise and peril, timed with references to beating China, what shocked her was how much the senators believed. These men and their policy teams had monopolised the message so long that it had become gospel; Schumer's office later said he was personally consulting Altman and OpenAI executives. Raji's conclusion: Washington needs far more people simply debunking claims.
The World Tour and a CEO going his own way
The Washington blitz was only the US leg. In March 2023 Altman tweeted to his 1.5 million+ followers that he would travel to meet users; interest snowballed into a multicontinent photo-op circuit with seemingly every G20 leader, branded "Sam Altman's World Tour." There was no strategy behind it: he picked the first stops himself, and comms and policy staff were roped in afterwards, scrambling for logistics at each leg.
Hao sees an old dynamic: Altman going his own way, sometimes in step with the company, sometimes not.
- The Pachocki credit. GPT-4 involved 100+ staff and 18 leads, and OpenAI carefully presented it as a company effort, signed simply "OpenAI." Altman then tweeted credit to Jakub Pachocki alone and seated him behind him at the Senate hearing; some employees wondered whether his contribution was really singular.
- Shadow decision-making. There were official executives, but it was unclear whether Altman listened to them, or whether process or personal relationships and whims guided strategy.
- Mixed messages. With new comms, legal and policy departments needing one voice, the lack of clarity surfaced publicly. After The New York Times sued OpenAI and Microsoft for copyright infringement at the end of 2023, the legal team's January reply accused the paper of "intentionally manipulating" models; the same week the policy team told the House of Lords communications committee that training frontier models without copyrighted material was "impossible," a word OpenAI hastily retreated from once the press seized on it.
A public-facing employee told Hao the company's reach had outrun its maturity, that people mostly make their own decisions which later look strategic by accident, and: "Sometimes there isn't a plan as much as there is just chaos."
Boomers and Doomers under one roof
As Altman toured Europe, Latin America, the Middle East, Asia and Africa, the strategic vacuum widened OpenAI's old fault lines.
Boomers. Applied led the charge, racing more competitors than ever, now joined by the ballooning corporate departments and many in Research. The hype convinced them that advancing and releasing models was the best way to fulfil the mission. Most executives embraced "iterative deployment," a phrase coined between releases: unlike GPT-2's staged release or GPT-3's controlled API, it meant putting models in users' hands early and often, giving society time to adjust and OpenAI real feedback. Altman told the Senate that building a superpowerful system in secret and dropping it all at once would not go well.
Doomers. The Safety clan, concentrated in Miles Brundage's policy research team and Jan Leike's alignment team, was a much smaller minority, compensating by sounding the alarm louder about dangerous capabilities and existential risk. As capabilities advanced, some researchers outside the clan converted, now believing AI could subvert human control. For them the same hype meant a moral duty to maximum caution.
Sutskever, prophet of both
Sutskever was the leader out of step. After GPT-4 he pivoted hard from capabilities to safety, splitting his time half and half, both more excited and more afraid of AGI becoming superintelligence. To those around him he seemed at war with himself, and his language grew messianic.
- The bunker. Briefing new researchers on preparing for AGI, he mentioned getting into "the bunker"; asked what bunker, he said OpenAI would definitely build one before releasing AGI, because governments would covet the technology and core scientists would need protection; entry would be optional. The researcher, still admiring him, kept his distance and told Hao that a group including Sutskever believes AGI will bring a literal rapture.
- The Alignment Manhattan Project. Sutskever and Altman conceived a team to develop alignment methods for superintelligence, expecting RLHF to fail once systems could outsmart humans. They first considered spinning it out as an independent nonprofit with a $1 billion endowment, but Sutskever did not want to leave and model access would be harder, so it stayed in-house as Superalignment, co-led by Sutskever and Leike, with a flashy public pledge of 20% of the compute OpenAI had secured to date.
- The inferno. Introducing the team, Sutskever thumped the microphone and declared that "Super alignment is a blazing inferno" (alignment itself being merely a burning fire).
- Physical split. Outgrowing Mayo, the plant-filled post-pandemic office, executives moved most of Research back to the old Pioneer Building, leaving Applied and Safety in largely separate worlds.
Just after announcing Superalignment, OpenAI rented a theatre at San Francisco's Metreon for staff to see Oppenheimer. Altman tweeted that he had hoped it would inspire kids to become physicists, as The Social Network had inspired founders. The Manhattan Project analogy had run through OpenAI since his first emails to Musk eight years earlier, even in new-hire orientations. Altman shares Oppenheimer's birthday, mentions it to reporters, and paraphrases the idea that technology happens because it is possible, never adding that Oppenheimer spent his later life in regret, campaigning against his creation. Most staff saw the project as a heroic, world-saving feat; the Safety clan saw the gravity of risking humanity's demise; for Altman it was a PR lesson, since he had argued years earlier that the mushroom cloud image turned the world against science and that people trust imagery over facts. An extreme Doomer asked why not invoke Apollo (OpenAI had also screened Apollo 11) instead of a project with that baggage, and quoted the film's Trinity scene: Oppenheimer puts the odds of igniting the atmosphere at near zero, and General Leslie Groves replies that zero would be nice. Nobody, he said, can calculate anything like that about these AIs yet.
Agents, AI Scientist and the threat model
With large language models slowing as data and compute ran short, Research pivoted toward agents, reviving the old debate between the "pure language" and "grounding" hypotheses: pure language and language-plus-vision were reaching their limits, so progress might need systems that act in the world and learn from feedback. Agents were also a commercial edge (automating emails or coding websites beats chatting) and could speed OpenAI's own research.
- AI Scientist. The most ambitious effort: an agent to perform research autonomously and generate its own knowledge by running experiments, since there was little new text to scrape. It merged the Codex code-generation team with a team teaching step-by-step reasoning from large sets of maths problems and solutions, both seen as building blocks for experiments and data analysis.
- Leaders and aim. Pachocki and Szymon Sidor led it, aiming, as Altman wanted, for an autonomous AI researcher to supercharge OpenAI's own progress.
- Excitement and dread. Success would bring AGI sooner, shortening the road to utopia or obliteration. Some researchers began casually discussing a bunker, even a Los Alamos-style compound in a remote American desert.
Who counts as a threat
In 2023 the security team posted a draft threat model on Slack, far more developed than earlier debates, with three categories: foreign state actors, competitors, and ideologically motivated people. For the third it linked to Eliezer Yudkowsky: the extreme Doomer who coined "friendly AI," wrote the fan fiction Harry Potter and the Methods of Rationality (122 chapters, 660,000+ words, a gateway to effective altruism and Doomer ideas), co-founded LessWrong, and had pushed for a full pause as his p(doom) reached 95%. In a March 2023 Time article, after describing his grief watching his daughter lose her first tooth and wondering if she would grow up, he proposed shutting down all large GPU clusters, tracking GPU sales and, if necessary, air strikes on "a rogue data center." The draft cited this as an ideologue advocating violence. A small group of Yudkowsky fans objected that it named him but not unfriendly AI itself; the security team added a fourth category: misaligned AGI systems.
A board losing its grip
Three departures in four months
As OpenAI soared, its board shed members without replacing them, partly because the ChatGPT race created conflicts.
- Reid Hoffman. Left February 2023 after five years, over conflicts: in 2022 he had co-founded Inflection with ex-DeepMind co-founder Mustafa Suleyman, fast becoming a rival.
- Shivon Zilis. Musk's deputy and Neuralink director, who kept watch for him after he left and formally joined in 2020, pledging confidentiality. News in July 2022 that she had twins with Musk, undisclosed to the board, made that doubtful, yet Altman kept her on, even consulting her in October 2022 on Musk's irritation at OpenAI's near-$20 billion valuation (Musk had texted Altman that it was a "bait and switch"). She left in March 2023 as Musk incorporated xAI.
- Will Hurd. Former Republican Texas congressman and CIA officer, joined 2021; Altman pitched him as a balance to Silicon Valley's liberal bias, and staff grilled him, including on Trump. He left in June 2023 to run for president and withdrew by October.
- Adam D'Angelo. Quora co-founder and CEO, Zuckerberg's Phillips Exeter classmate, Facebook CTO for two years; director since 2018. Quora joined YC in 2014 (Altman's first batch as president), and in 2017 Altman co-led its $85 million round, calling D'Angelo one of the smartest, most long-term CEOs in the Valley.
- Tasha McCauley. Co-founder of a telepresence robotics startup and head of a 3D urban-simulation firm, introduced via her mentor Alan Kay; nominated by Holden Karnofsky in 2018 as OpenAI prepared its capped-profit shift. Respected in AI safety, she served on the Centre for the Governance of AI board and, for a time, Effective Ventures (which oversees 80,000 Hours).
- Helen Toner. Joined mid-2021, also nominated by Karnofsky as his own replacement when his term ended and Anthropic formed. Worked with him at GiveWell and Open Philanthropy before becoming a CSET expert on China and emerging technology; came to AI safety via EA but drifted away, writing in her most-read pre-FTX forum post that the movement was turning dogmatic and insular. Also on the GovAI board.
That left three independent directors (D'Angelo, McCauley, Toner) beside Altman, Brockman and Sutskever.
Deadlock over the next director
After the GPT-4 demo and ChatGPT, McCauley ran a roughly year-long review, interviewing employees and outside stakeholders, of how a professionalised board and oversight should look.
- Agreed criterion. Everyone, Altman included, agreed the next independent director needed deep AI-safety expertise. They compiled a list and interviewed five, including Dan Hendrycks, a central Doomer figure running the Berkeley Center for AI Safety and sole adviser to Musk's xAI.
- Stalling. When the independents tried to choose, Brockman and Sutskever each raised objections. Altman, characteristically, opposed no one and advanced no one, and repeatedly floated new names who shared one trait: all sat inside his network and sphere of influence, financially or otherwise. By late summer 2023 the deadlock had lasted months.
- Oversight channels neglected. Altman had made McCauley the board's employee liaison, and she held regular office hours (once bringing her husband, actor Joseph Gordon-Levitt, to an off-site where he followed technical talks). They lapsed in the pandemic and never restarted.
- Conflicting pictures. Without a systematic channel, directors relied on personal contacts in the safety and tech worlds and on Altman. His accounts were rosy; their sources reported poor preparation before and turmoil after ChatGPT, unresolved GPT-4 safety concerns, and products rushed out before problems were fixed.
- The India breach. At the board's first annual on-site in late 2022, Altman spent a day touting the Deployment Safety Board's protocols for GPT-4. Afterwards an employee mentioned to a director that Microsoft had already done a limited GPT-4 rollout in India without DSB approval. Altman had not mentioned it. Nothing unsafe seemed released, but the directors feared the precedent once models could cause catastrophic harm.
- The D'Angelo motion. In March 2023 Altman emailed the board, excluding D'Angelo, to say D'Angelo should step down because Quora's chatbot Poe was a conflict. To the three independents, each of whom had asked awkward questions about safety or the nonprofit's strength, it looked like fishing for a pretext as Altman's allies dwindled. A director noted Poe was a modest conflict next to those tolerated from Hoffman and Zilis. The motion failed.
- The Startup Fund. D'Angelo twice heard at dinners that the OpenAI Startup Fund gave its investors early product access, a perk that should go to OpenAI's own investors. When the directors finally obtained the documents, they found Altman legally owned the fund, which should have belonged to OpenAI.
To the independents, each episode fit one picture: Altman, bit by bit, was clouding their visibility and manoeuvring so the board could never check him, even as he advertised worldwide that the board's power to fire him was OpenAI's key safeguard. By autumn their morale hit new lows. With OpenAI preparing to train GPT-5 and AI Scientist advancing, the gaps grew urgent. Then in early October 2023 Toner received an unexpected email: Sutskever wanted to talk.
Sharp but partial. Hao gives the compute-threshold critique far more room than its defence; even critics concede compute is the one measurable input, and the Biden order used it only to trigger reporting. Calling the white paper a vehicle for corporate capture sits awkwardly with OpenAI's lobbying against SB 1047, which Hao herself notes. Altman's "no equity" answer was technically true but omitted an indirect stake through a YC fund, which he later acknowledged. The board grievances come largely from the independents' side; Altman's account of these episodes is mostly absent.
Sutskever and Leike left in May 2024 and Superalignment was disbanded; reports said it never received its promised compute. Sutskever founded Safe Superintelligence in June 2024; Pachocki became chief scientist. Altman handed control of the Startup Fund to others in 2024. Trump revoked the Biden order on 20 January 2025 and, in December 2025, ordered a challenge to state AI laws, while California's SB 53 (signed September 2025) kept a 10^26 threshold for transparency duties. The EU's general-purpose AI rules took effect in August 2025 with 10^25 as the systemic-risk line. Yudkowsky and Nate Soares published If Anyone Builds It, Everyone Dies in September 2025.
Deliverance
Hao steps out of the boardroom to tell the story of Annie Altman, Sam's younger sister, whose estrangement, poverty, sex work and allegations of childhood abuse broke into the mainstream press in September 2023, seven weeks before the board fired him. She reads the story two ways: as evidence about Altman's character and how tightly OpenAI is bound to the man who runs it, and as a microcosm of the book, a man promising AI-driven abundance while someone close to him received none of it. Hao does not claim to know whether the abuse allegations, which Altman and his family deny, are true; her point is that Annie's story reached Sutskever just as he was concluding that Altman's conduct amounted to a pattern of abuse.
The profile and the apology
September 2023
In the weeks before Sutskever contacted Toner, Altman faced a PR crisis. On 25 September 2023 Elizabeth Weil of New York magazine published a profile that, for the first time in the mainstream press, mentioned Annie, setting her repeated illness, financial distress and housing insecurity beside Sam's multimillion-dollar homes and luxury cars. Weil observed that readers of Sam's blog, tweets and Startup Playbook knew his brothers Jack and Max, but Annie had no place in his public life.
- The comms chief as go-between. Weil gave OpenAI the chance to comment, and New York worked with the company during fact-checking. Managing it fell to Hannah Wong, now OpenAI's VP of communications, who found herself mediating between a magazine and the CEO's family and wondering whether this was really her job.
- The Yom Kippur email. Publication fell during Yom Kippur, the Day of Atonement. On the final day before release, once it was clear the details would run, Sam emailed Annie asking forgiveness. He said he had felt caught in the middle: inclined to defer to their mother, agreeing with the family that she should learn financial independence, yet sensing she needed medical help and was struggling to function. Still, he wrote, he had made the wrong call and should have kept supporting her.
- Fallout. The turn from his glowing post-ChatGPT coverage, with what he saw as painful private matters made public, weighed on Sam; it worsened when Annie's old tweets alleging abuse resurfaced and went viral. For Annie it was an extraordinary release: after years of feeling she was shouting into a void, she was finally heard.
Hao's reporting and the family's responses
In 2024 Hao approached everyone. Annie cooperated eagerly, seeing her experiences as key to judging Sam's moral character, and supplied extensive family correspondence, physical and mental health records from childhood to adulthood, and other corroboration.
- Connie Gibstine (their mother) sent a brief statement of love and concern, calling Annie's claims "horrible, deeply heartbreaking, and untrue," but declined detailed questions and did not answer requests for documents supporting her own claims.
- Max and Jack did not respond; OpenAI did not comment on specifics.
- January 2025. After Annie sued, Sam, Gibstine, Max and Jack issued a more forceful joint denial, describing her as mentally unstable and unreasonably demanding of money, saying her claims had "evolved drastically over time" and lamenting that she refuses conventional treatment and lashes out at relatives trying to help.
Hao states her limits plainly: she lacked the family's full reasoning, and the truth of some allegations, especially of childhood sexual abuse, is unknowable. But the story became a microcosm of the OpenAI themes and confirmed how far OpenAI is a reflection and extension of its CEO. It became a company issue in two ways: coverage of Annie got under Sam's skin more than almost any other story, drawing Wong into containing it; and, when the allegations surfaced, other OpenAI executives took notice.
How Annie's life came apart
A sibling with the same gifts
Hao notes that Annie, nine years younger, resembles Sam in what friends praise about him: an excellent listener with a memory for small details, goofy, very generous, quick to win trust. She too excelled at school: the only other Altman sibling to graduate from Burroughs, remembered fondly by physics teacher James Roble. At Tufts she majored in biopsychology with a minor in dance, graduated in 2016, finished premed requirements and moved to the Bay Area for a neuroscience research job at UCSF: in short, an ambitious graduate with options.
Body, mind and a death
- Chronic illness. In 2014, still in college, she was diagnosed with Achilles tendinitis and a bone spur and put in a walking boot, the first of many ailments over six years: recurring tonsillitis, recurring pelvic pain, tendinitis flare-ups that spread beyond her right ankle and sometimes made standing hard, and ovarian cysts culminating in PCOS (with night sweats).
- Mental health. Diagnosed young with generalised anxiety and OCD, she took Zoloft for most of a decade until a psychiatrist helped her taper off in college. Before 2018 she was doing better, using meditation, a whole-foods diet aimed at root inflammation, alternative medicine and a yoga teacher training, and drafting a book of life reflections, The Humanual.
- The death. On 25 May 2018 her father, the family member she was closest to and her sole unequivocal supporter, died of a sudden heart attack. Her mental health spiralled and her physical decline accelerated; by late 2018 specialist visits were routine.
- Sam's grief. Sam has called it the worst moment of his life; it came months after Musk left and Sam took over OpenAI, and people close to him say he behaved erratically for a while.
- Family strain. Relations, especially with her mother, were already tense over her turn away from a medical career and off Zoloft. Without her father she felt isolated yet craved the family's acceptance; clashes over money brought the break.
The money runs out
After his death Annie lived in Los Angeles, working part-time as a writing assistant, then at a marijuana dispensary. She says she inherited about $100,000 in life insurance and invested it in performance as a career: comedy classes, open mics, a podcast, and a one-woman show reworked from her book, The HumAnnie. By mid-2019 most was gone on those projects, high rent, out-of-pocket health insurance and medical costs (imaging, physical therapy, psychotherapy, rides to appointments).
- The 401(k). In May 2019 she learned her father had left her his 401(k), a little over $40,000. She quit the dispensary and planned six months to heal and get her creative work earning. Then Gibstine, controlling the funds as surviving spouse, told her it would stay in Gibstine's name in a tax-deferred account passing to Annie by trust, accessible at 59½, as the best tax strategy.
- The email. It said the family believed financial independence was best for her and that they would not support her once her inheritance ran out. It was signed "Mom, Sam, Max, Jack." Her therapist's notes record that the move, meant to get her on her feet, worsened her condition.
- The first line crossed. In December 2019, the year Sam secured Microsoft's first $1 billion and OpenAI raced to train GPT-3, her account went negative. Frightened and alone, she joined the escort site SeekingArrangement and exposed herself on a video call for enough money to clear it.
- Appeals. Annie says she had never asked for money before, though she did not expect to be left with no safety net in a real emergency. From late 2019 to mid-2020, including after the pandemic hit, she appealed several times. After two family therapy sessions with Sam and her mother, they covered her expenses for part of the year. Their correspondence shows they feared money could enable harmful behaviour; both Gibstine's and the family's statements say they followed professional advice.
- The refusal. In May 2020, with coverage ending, she asked for help with physical and talk therapy at $45 and $15 a session. They declined, saying she should cover June herself, partly from a returning security deposit.
- Hawai'i and the diamond. She moved to Hawai'i (where her father had last visited her) for a farm work-trade of weeding and planting. Sam emailed for her address: eight months after the death he had asked each sibling to mail their mother a lock of hair to be mixed with their father's ashes and made into a diamond for each child, and hers was ready. To Annie, who needed food and housing, it felt like a slap in the face; she could not tell whether he did not understand or did not care. In great pain, she stopped speaking to the family.
Three years at the bottom
Over the next three years Annie faced housing, food and health insecurity and turned to virtual and in-person sex work. The family says it tried "in many ways" to help her find stability.
- The 2021 reconnection. In spring and summer 2021 Sam sought contact; in three phone calls he said he loved her and offered to buy her a house. Accounts diverge (see table). They settled instead on Sam paying a year's rent directly to her landlord; when she did not re-establish contact, payments stopped in summer 2022.
- Hao's reading of Sam. The episode deepens the dueling portraits: generous and self-serving, agreeable and threatening, benefactor to many and source of pain to others; sincere and altruistic in public, more calculating in private; someone who gives and takes away, leaving people feeling like pieces in a chess game only he can see, played to keep the king in power.
| Disputed point | Annie's account | Family's account |
|---|---|---|
| House offer | To live in, not own; a way to block a sale and a lever over her health and career choices | Ownership through a trust, so she had a home but could not sell it immediately (no documents supplied to Hao) |
| Financial help | Too little, too late (after sex work began), and usually conditional | Supported her "in many ways," guided by professional advice; monthly support promised "for the rest of her life" |
| Mental health | Never diagnosed with borderline personality disorder; records show only an evaluation | Sam privately told people she had it; she "refuses conventional treatment" |
| Abuse | Sam abused her from about age three until she was a minor and he an adult | "Untrue"; her claims have shifted over time |
The abundance that never reached her
Hao sets Annie's life against OpenAI's grand promises: Altman says AI will end poverty; Brockman, through stories about a friend and his wife Anna, says AGI will transform healthcare; Sutskever predicts wildly effective, dirt-cheap psychotherapy. Beside Kenyan data workers, Chilean activists and Altman's own sister, Hao finds those dreams hollow: no AI advance relieved any of Annie's desperation, and AI may have trapped her further.
"Plan Z" and the algorithm
Annie calls sex work her "plan Z." When chronic pain made even farm work too hard, she first tried to earn from her art online: her podcast, an Etsy store and a Patreon together earned less than her phone bill. She documented in screenshots a strange collapse in reach: little exposure on any platform, chunks of her Apple Podcasts reviews vanishing, and, at least twice on Instagram and YouTube, views accumulating and then disappearing.
- The possible mechanism. A former Facebook data scientist and two experts on tech and sex work told Hao it is possible her first SeekingArrangement account (December 2019) limited her online traction, because platforms' automated systems shadow-ban sex workers by tagging devices, emails, bank accounts and other identifiers that link a person's presence, even on profiles unrelated to sex work.
- The trap closes. Seeing no other path, she returned to SeekingArrangement and started an OnlyFans account, binding her income still more tightly to algorithmic moderation.
- A witness. Neily Messerschmidt, a former tech leader (including at Sony) who now runs wellness for an organic-farm network, met Annie in 2016 and became a motherly figure after the estrangement. Her verdict: "Sam carried AI into the world just like he actually treated his young sister," thriving while his sister fell through the cracks.
The allegations
Flashbacks and the therapy record
- Onset. In late 2020 and early 2021, after the estrangement, Annie began having devastating flashbacks of childhood sexual abuse.
- The Maui notes. Across 15 sessions with a new trauma therapist on Maui (July 2021 to January 2022), notes chronicle involuntary memories and intense questioning of identity. Diagnostic impressions: generalised anxiety, PTSD, and a personal history of childhood sexual abuse. No abuser is named.
- What she told Messerschmidt. Messerschmidt, herself raped at 19, says she had noticed early signs in Annie, such as avoiding meetings and standing apart when certain men were present. In emotional calls Annie described memories of Sam repeatedly getting into her bed, sometimes with Jack, and molesting her.
- Hao's caveat. Childhood abuse is often hard to prove decades later. A therapist Hao consulted describes a known pattern in which memory is blocked until a trigger (puberty, sexual activity, unwanted advances) brings it back; deeper sex work may have supplied such triggers. Hao's stated aim is not to establish what happened but to reconstruct what Annie experienced and believed as an adult, which drove her to speak.
Going public
- November 2021. While Sam was paying her rent, Annie tweeted for the first time that she had suffered sexual, physical, emotional, verbal, financial and technological abuse from her siblings, "mostly Sam" and some from Jack, and sought others to join her in pursuing justice. With few followers, against Sam's rising fame from GPT-3 and GitHub Copilot, it went nowhere; trolls claimed she was not really his sister.
- 2022. Two reporters contacted her, but she hesitated. In September 2022 she resolved to speak openly and tweeted again, naming Sam and Jack.
- July 2023. Ten months later, after ChatGPT made Sam a global star, Weil messaged her.
After the profile
2014 Achilles tendinitis begins years of chronic illness.
2016 Graduates from Tufts; moves to the Bay Area for a UCSF lab job.
25 May 2018 Father dies of a heart attack.
May 2019 Learns of the 401(k); the family email withholds it and ends support.
Dec 2019 Bank account negative; first SeekingArrangement session.
May 2020 Therapy funding refused; moves to Hawai'i; cuts contact.
Late 2020 Flashbacks begin.
Mid-2021 to mid-2022 Sam pays a year's rent.
Nov 2021 First public tweet alleging abuse.
Sep 2022 Second tweet naming Sam and Jack.
9 Jul 2023 Sam offers money, wants no relationship; she never replies.
25 Sep 2023 New York profile publishes.
Early 2024 Hires a lawyer; trust distributions begin; first stable flat in four years.
Oct 2024 Diagnosed with hypermobile Ehlers-Danlos syndrome.
Jan 2025 Files suit before her 31st birthday; family issues joint denial.
- Income. In the three months after the article her OnlyFans income rose more than 10×, from about $150 to over $1,500 a month, peaking near $5,500 in one month during the board crisis. She stopped escorting and kept only virtual work.
- Too little, too late. The family cites real help, the post-therapy support plan and the rent Sam paid from mid-2021 to mid-2022, but for Annie it came after she had already been driven to sex work, and other offers, the house included, carried conditions she felt unable to accept.
- Sam's last offer. Earlier that summer an internet sleuth emailed both siblings asking Annie to elaborate; she replied on the three-way thread with a detailed, cutting account. It landed in Sam's inbox around 5 June 2023, the day he gave a well-received talk with Sutskever at Tel Aviv University. On 9 July, with the World Tour over, he replied (unusually, in proper capitals) that he wanted no ongoing relationship but would gladly send money, either restarting the old arrangement, whose details he could not recall, or a lump sum, adding that he disagreed with many of her claims but saw no point engaging. She never answered: she no longer wanted his money, only what her father had left.
- Leverage at last. Her father had also left a trust under her mother's control. In early 2024, able to afford her own lawyer, Annie learned it had been newly funded in 2023. As her story spread, Gibstine's lawyer opened talks on monthly, no-strings distributions; for the first time she felt she had negotiating leverage. The family says it expects to support her monthly for the rest of her life. She rented her first stable apartment in over four years.
- The lawsuit. She then hired another lawyer to file a child sexual abuse case before her 31st birthday on 8 January 2025, alleging abuse from about age three until Sam was an adult and she still a minor, seeking damages above $75,000. The family vehemently denies it.
- A diagnosis. In October 2024, after another round of appointments, she was diagnosed with hypermobile Ehlers-Danlos syndrome, the same genetic disorder as Brockman's wife.
Containing the story
The borderline label
Before the profile appeared, Sam began telling people, as a private and sensitive matter, that his sister had borderline personality disorder (BPD), a condition of severe difficulty with emotional regulation, intense relationships swinging between idealisation and fear of abandonment. One of those people said he had "defanged her account before it even published."
- The record. Annie says she was never diagnosed with BPD, and none of the notes or records she gave Hao contain it. The only trace is one mark in August 2021 from her Maui therapist showing she was evaluating Annie for it (it commonly follows childhood sexual abuse); the final impressions list abuse history, not BPD.
- The prognosis. Two therapists told Hao BPD usually resolves, naturally or with treatment such as meditation and behavioural therapy that rebuilds self-worth (tools Annie uses), rather than the psychotropic medication Gibstine urged and which the family statement's "conventional treatment" seemed to mean. Blaise Aguirre of Harvard Medical School, who has treated thousands of BPD patients but did not review Annie's case, calls it a good-prognosis diagnosis from which most people recover.
For six months Wong's team promised Hao an office visit and interviews with leaders; after five they soured, and ten days before her booked flight to San Francisco they cancelled everything and withdrew from the book. She went anyway. Days in, she told someone who knew both siblings that she was speaking to Annie; the next day Wong, who had dealt with her only through a deputy, texted "I hear you are in town?", odd given why the trip existed. Over coffee at a Philz in Mission Bay, near a new OpenAI office, Wong steered small talk to Annie, unprompted and before Hao had raised her with OpenAI: Annie had mental health challenges, good days and very bad ones; the family was balancing protecting her against enabling her; and, stressed twice, the family had not publicly denied her claims, which was all about protecting her. She added that some journalism programmes had debated whether including Annie in Weil's profile was ethical, and suggested Hao consider that too. Delivering this was plainly the meeting's purpose. Wong was soon promoted to chief communications officer. For Hao it showed the pressure the story put on Sam and how closely the company and Sam's private life were fused.
The same gulf of power
Hao sees in Sam's and Wong's framing of Annie's mental health another parallel with those sidelined by the AI empires. Annie's life has become a hunt for documentation to make anyone listen, sometimes with a sinking sense that the world conspires against her: the same loss of agency and anger Hao saw among data workers and data-centre activists, who throw what little they have at the empire's narrative and watch opponents with billions in capital, vast infrastructure and armies of contractors smooth over protest with a few soft words to Congress, heads of state and journalists.
After Weil's profile, though, Annie was no longer shouting into a void. As her tweets went viral they reached Ilya Sutskever, just as he was wrestling with his own feelings about Altman and what he saw as Altman's patterns of abuse.
One-sided by necessity. With the family declining to engage, Hao relies on Annie's documents and second-hand recollection (Messerschmidt) of the abuse claims, whose truth she calls unknowable; the therapy notes name no abuser. Her "plan Z" link between one SeekingArrangement account and later shadow-banning is plausible, not shown. The source also dates the moment Annie's tweets went viral and reached Sutskever to "October 2022," which contradicts the chapter's own chronology; it must mean October 2023, after Weil's profile. Hao's answer to those who doubt a family dispute belongs here is that Altman's own executives treated it as relevant.
Altman and his family denied the allegations in January 2025. On 20 March 2026 US District Judge Zachary Bluestone in Missouri dismissed Annie's claims as untimely but allowed her to amend; she filed an amended complaint under Missouri's child sexual abuse statute on 1 April 2026, and it remains pending. Hannah Wong announced her departure from OpenAI in December 2025.
The Gambit
Hao reconstructs, almost day by day, the four weeks from 29 September to 26 October 2023 in which OpenAI's two most senior deputies, CTO Mira Murati and chief scientist Ilya Sutskever, separately took their concerns about Altman to independent director Helen Toner. One came from Applied, one from Safety, yet both described the same behaviour; then Altman handed the board its proof in real time by misrepresenting a fellow director in a bid to push Toner out. Murati's summary of the pattern frames everything that followed: Altman tells resisters what they want to hear, then, when they keep resisting, quietly undermines their credibility until they get out of the way.
Murati, the translator
From Albania to OpenAI
Murati met Toner four days after Weil's profile and four days before Sutskever's message. Hao's portrait explains why her word carried weight.
- Calm amid chaos. She grew up in Albania during its abrupt lurch from totalitarian communism to liberal capitalism; with an immature financial system, pyramid schemes boomed and collapsed into unrest and violence, leaving bomb craters she stepped around on the way to school.
- Numbers as refuge. Her parents taught literature, but she loved maths, then chemistry, biology and physics; teachers fed her harder problems, she devoured her own and her older sister's textbooks, and she thrived in maths and science Olympiads.
- The ladder. At 16 a scholarship took her to Pearson College UWC in Victoria, British Columbia; then mechanical engineering at Dartmouth; an aerospace company; and her break at Tesla as senior product manager on the Model X, learning to ship complex products under pressure across feuding teams. Tesla's autonomous-driving work drew her to AI, which she came to see as a universal tool; she told Microsoft's Kevin Scott it seemed like maybe the last thing anyone would ever work on.
- A detour. After three years she became VP of product and engineering at Leap Motion, dreaming of AR/VR education (rotating DNA strands by hand), but the bet came too early: headsets still made people nauseous.
- OpenAI. She joined in 2018, still a nonprofit, became VP of Applied and partnerships as it commercialised, ran the Microsoft relationship, and after the Anthropic split also oversaw Research, becoming CTO in May 2022.
- Standing. Younger than Altman, Sutskever and Brockman, and for a while the only technical woman in senior leadership, she drew sexism from old-guard researchers who called her an engineer, not a scientist, and a symbol of OpenAI's drift from fundamental research. Colleagues describe a socially adept, ego-light listener and problem-solver who could find an answer when none was obvious: "consistently correct," "just honest," "the one getting stuff done."
Translator, then cleaner
As more teams reported to her, Murati became the conduit to Altman: implementer of his strategic shifts and champion of teams pushing back. She could tell him plainly when plans were unrealistic, and he often listened; she could tell others when he did not actually want something he pretended to want, and people came to her to decode him. Increasingly, though, she was cleaning up his messes.
- Two yeses. When teams disagreed, Altman privately agreed with each, breeding confusion, conflict and mistrust.
- Brockman. To Murati he was "a second CEO but a bad one," opinionated and burning people out; the Altman-Brockman dynamic during GPT-4 nearly made key pre-training staff (the team handling data collection and initial training) quit.
- Nadella, the third CEO. Altman was so deferential to Microsoft that, after Murati prepared reasonable commitments, he would go off script with Microsoft executives and agree to demands detached from OpenAI's road map. To the board he framed the result as Murati lacking a productive relationship with the partner.
- Worsening under stardom. After ChatGPT his schedule exploded; once energised, he was now often exhausted and anxious, agreeing with people to their faces and, more often, badmouthing them behind their backs. Team leads copied him and pitted reports against each other.
- Skirting safety process. Under competitive pressure he pushed to ship faster, sometimes dodging release processes through dishonesty: he told Murati legal had cleared GPT-4 Turbo to skip DSB review, but legal chief Jason Kwon had no idea where Altman got that.
- Retaliation. When she gave him detailed feedback that summer, he iced her out; thawing took weeks and an assurance that she had told no one else. She had watched other executives who challenged him get cut from decisions or quietly discredited, subtle enough to stay invisible to staff, so that each executive took a turn.
The Sutskever-Pachocki rift
The latest target was Sutskever. Jakub Pachocki, the Polish researcher leading AI Scientist under Sutskever, felt under-recognised and complained to his ally Brockman, who went to Altman. Altman encouraged Pachocki and gave him a more senior research role, but never told Sutskever, and would not clarify their portfolios. Receiving different messages from Altman, both steered the same research in different directions without understanding why they kept clashing. Months of organisational thrash in Research reached GPT-4-crisis levels of stress; for Sutskever it was a humiliating snub that also destroyed a friendship forged over years of late nights.
Murati again worked overtime just to find out what was happening, then to broker an arrangement, stop Brockman lobbying Altman on Pachocki's behalf, and stop Altman contradicting her in private. None of it fixed the root: another leadership crisis was only a matter of time, and OpenAI needed stronger governance and accountability.
Why she went to the board
During the freeze, Altman had seemed most afraid that she had taken her feedback to the board. She had not: she wanted to resolve things directly and doubted the board offered real accountability.
- Too many insiders. Three co-founders on a supposedly independent nonprofit board was far too many.
- Compromised independents. Many "independent" directors had financial ties to Altman or benefited from his networks.
- The web. She had watched him invest in startups and donate to politicians to bind relationships; when he asked what he could do for her, she always demurred so as not to owe him.
Yet the board was choosing new members, and the next one would strengthen or weaken its independence. Due in Washington (where Toner lived) to speak at The Atlantic's ideas festival at the end of September, Murati asked Toner for coffee. Altman, she believed, needed real oversight.
Four weeks in October
29 Sep Murati-Toner coffee: business is fine, but Altman is pushing to ship so fast that "bad things" could happen.
4 Oct Sutskever, who had never contacted Toner alone, says OpenAI is "trickier than it seems from the outside."
12-13 Oct Sutskever's Superalignment retreat.
15 Oct Murati describes Altman's playbook and warns against an Altman ally as seventh director.
16 Oct Sutskever's second call: Altman's "patterns of abuse."
19 Oct Murati meets Microsoft's Mikhail Parakhin; finds Altman again overpromised.
20 Oct Murati's third call: Brockman once tried to fire her.
23 Oct Sutskever proposes Murati as interim CEO.
24 Oct Independent directors agree to hire separate nonprofit lawyers; Altman phones McCauley late that day.
25 Oct Altman's 15-minute call to Toner about her paper.
26 Oct McCauley tells Sutskever she never said Toner must go.
29 September: a routine coffee
To Toner it seemed reasonable for an executive to talk to a director. Murati's updates were mostly ordinary: OpenAI would make plenty of money; the key projects were the Gobi model and a new Microsoft deal; she was handling personnel issues around the Altman-Brockman dynamic, this time involving Sutskever and Pachocki. One remark stood out: Altman was pushing to ship so fast she feared bad things could happen.
4 October: Sutskever's first call
Far stranger was Sutskever's email days later: in two years on the board together he had never contacted Toner individually. On the call he was so nervous he struggled to speak.
- Opening. When Toner said everyone wanted a strong board, he gave an uncharacteristic laugh-scoff, agreed in a way implying others did not, and rolled his eyes when she said everyone agreed.
- The hook. Told the independents wanted a director with a safety background, he seized on the point that the board needed to be better informed and pay attention.
- Holding back. Asked what the board most needed to know, he said that answering would itself reveal why he hesitated; OpenAI at the top was trickier than it looked from outside. He might sleep on it and find specifics he could share, and suggested she talk to Murati, who had more context.
15 October: Murati describes the playbook
More than a week later, after another board meeting, Toner told Murati something odd seemed to be going on; Murati said she was perceptive, and they spoke again. Echoing Sutskever, Murati said there was a lot of tricky stuff she could not fully share, because once said it could not be unsaid, especially with Altman feeling threatened as CEO. Toner was surprised: the board had sought stronger checks but had never discussed removing him.
- Win them over. To anyone resisting a decision, say whatever they want to hear.
- Wait, then undermine. When patience runs out and they still resist, erode their credibility until they get out of the way.
- Contradictory messages. Different people hear different things from him (Pachocki heard one thing from Murati, another from Sam), so no one sees the whole picture.
- Amplifiers. He is deeply anxious, and anxiety produces "dumb ideas," especially when Brockman eggs him on.
Her requests to the board: avoid an Altman ally as the seventh director; watch Microsoft's deployments of OpenAI technology and the DSB. Altman would be freaked out to learn they were talking, but toxicity at the top was unsustainable and something had to give within six to nine months. Toner should ask Ilya what he was willing to share.
Sutskever's case
What was weighing on him
Sutskever had become preoccupied with AGI's imminence: irreversible upheaval, and OpenAI's duty to steer toward abundance, not suffering. Then a second anxiety took hold: losing faith that OpenAI could reach AGI, or carry that responsibility, under Altman.
- Politics. After ChatGPT, working at OpenAI became Silicon Valley's ultimate social currency, intensifying competition among team leads. Instead of brokering egos, Altman told each what they wanted to hear while getting his way, with little lies and some big ones almost daily.
- Brockman. Once his closest confidant from their early nights dreaming up OpenAI, Brockman now added to the turmoil.
- The toxic result. A directionless, backstabbing environment without shared information or trust undermined both pillars of the mission: it slowed research and eroded sound safety decisions. And now Sutskever was being harmed directly.
- Choosing Toner. He judged her the safest director: vocal on governance and safety and plainly not in Altman's pocket. He was unsure of D'Angelo and McCauley, and feared what Altman would do if he found out.
Annie's tweets and the word "abuse"
After Weil's article, discussed only in hushed tones at OpenAI, two of Annie's tweets went viral: the November 2021 accusation and a far more explicit one from 14 March 2023, which drew nearly 4,000 likes, describing a 13-year-old "brother" climbing into her bed when she was four. Sutskever did not know if the allegations were true, but believed she had clearly had a rough upbringing with Sam, a sign of how far back his problematic behaviour might reach.
Her word, abuse, matched his own reading of Altman. Like Murati, he had been slow to see the playbook, though there were early signs: Altman's insistence on being CEO without clear or consistent reasons; small, pointless lies in the early years; and Dario Amodei's warning on leaving in late 2020 about "psychological abuse," a phrase Sutskever now understood. On 12-13 October he took his Superalignment team on a retreat where he again burned an effigy for team bonding.
16 October: the second call
- The pattern, not the grievance. He used the Pachocki affair as illustration: Altman could have simply told him he wanted Pachocki to have a bigger role, but instead pitted them against each other, with Brockman's eager help, leaving both blind to why they could not agree. He summed up the culture as beatings continuing until morale improves. Each act was subtle; only together did the pattern show. Toner should also speak to Amodei, and the sister's allegations were "a different kind of a safety issue." Altman, sometimes with Brockman, lied and manipulated so habitually that he sometimes seemed not to believe his own words.
- An exit offered. They might conclude nothing could be done, he said, in which case they should forget the conversation.
- His real agenda. The board's second annual on-site at the end of November was meant to settle new directors. He opposed expanding the board: it might not become more independent, and even non-loyalists would take too long to see through Altman, making accountability harder. He hedged and asked what Toner saw.
- Common ground. Toner agreed Altman was slippery and that management-level slipperiness cascades. With the three directors most deferential to Altman gone, she agreed the goal was to hold him accountable. Sutskever relaxed: they had been on opposite sides of the deadlock while wanting the same thing.
Toner: outer alignment. Strengthen the board: add an independent director with a safety background, set performance targets for Altman and revisit in twelve months, perhaps move Brockman off the board.
Sutskever: inner alignment. "The board is like outer alignment, the management is like inner alignment." Altman would simply hit any targets without changing his behaviour; Brockman stepping down would help but was only "plan B," since Altman was the root. Plan A, unspoken, was removing Altman.
Toner said it sounded as though he thought major changes should be on the table; he agreed, but warned there was no clear-cut evidence of anything egregious, so the likeliest outcome was that the board would do nothing, agree on directors, and let the moment pass. He ended the call sure that Toner was beginning to get it.
Murati's third call and Sutskever's proposal
19-20 October: Microsoft again
Shortly before, Altman had panicked, for no clear reason, that Microsoft was unhappy. On 19 October Murati met Microsoft executives including Mikhail Parakhin, head of Bing. It went better than expected, but revealed that Altman had again said yes to a Microsoft demand unrelated to OpenAI's actual work, creating false expectations she had to unwind.
On 20 October she told Toner she planned to give Altman extensive feedback: his yes against her no had fragmented the Microsoft relationship. The same pattern ran through Research. That very day she and Sutskever had agreed a configuration for him and Pachocki, and implored Altman not to deviate by telling Pachocki what he wanted to hear; Altman gave his word, but Pachocki heard one thing from Murati and another from Sam, and Brockman could lobby again at any moment.
- Brockman tried to fire her. He had recently admitted attempting it during GPT-4. She was technically his manager and once wrote his reviews, but the drama made her stop. She later confided she wished she could fire him, impossible while he sat on the board; she had considered asking him to leave it but had not.
- On Annie. Asked by Toner, she said she knew nothing and had not asked Sam, but if even 10% were true it was really bad.
- Keeping notes. She was amazed she could do her job amid it all, was writing notes as events happened, and offered to send more.
- The board's lane. Toner said the board would focus on what it could change: not Sam's personality but governance processes and structures to check him, which matched Murati's own efforts. Murati's last caution: make sure the board's information was not coming only from Sam.
23 October: "the finger on the button"
Having learned that Murati had also spoken to Toner, Sutskever was more explicit than ever on his third call. He said Altman was not the person who should have "the finger on the button for AGI," spoke of a tremendous opportunity for the board to act, and proposed replacing him with Murati as interim CEO. Toner suggested he talk to McCauley and D'Angelo; after meeting D'Angelo in person he found him less aware of the problem than Toner.
When the three independents later compared notes, they learned Murati had told them she did not feel comfortable with Sam leading OpenAI to AGI. That weighed heavily: if two of Altman's most senior deputies, one from Applied and one from Safety, felt the same, the board had a serious problem.
24 October: the board arms itself
Toner, D'Angelo and McCauley met to shore up oversight. One glaring gap: the nonprofit had no independent legal support, with everything routed through the for-profit's lawyers. They agreed to hire nonprofit counsel to attend every board meeting and review the deals Altman was striking.
| Source | Vantage point | Core message to Toner | Remedy implied |
|---|---|---|---|
| Murati | Applied, CTO, Microsoft liaison | Altman's playbook of flattery then undermining; overpromising to Microsoft; pressure to skip safety review | No Altman ally as seventh director; watch the DSB; information beyond Sam |
| Sutskever | Research, Safety, Superalignment | Habitual lies; pitting deputies against each other; a pattern running back years | Fix "inner alignment": Murati as interim CEO |
| Toner (synthesis) | Independent director | Slipperiness at the top cascades | Governance structures, independent lawyers |
The paper and the lie
25 October: Altman's call
Toner wondered if Altman had caught wind of the meetings. Instead, in a mild-mannered, 15-minute call, he raised one of three papers she had published that week at CSET: the densest and most academic, on "costly signals," the political-science problem of how states and companies credibly signal their intentions on AI. She was third author; OpenAI appeared only on pages 28-30 of 65, nowhere in the summary, web page or launch materials, and traffic suggested few had read it.
- How it surfaced. The day before, David Robinson, OpenAI's head of policy planning, had posted it in the #policy-research-chatter Slack channel, calling the OpenAI-Anthropic comparisons "quite spicy." The relevant pages weighed both labs' release strategies and praised OpenAI's candid GPT-4 safety assessment and its six-month delay, but Robinson pasted only three paragraphs, bolding lines contrasting the "race-to-the-bottom dynamics" ChatGPT spurred with Anthropic's restraint in releasing Claude. Colleagues called it partisan (harsh-but-fair on OpenAI, uncritical of Anthropic), and one also called it flimsy, while admitting he had read only snippets.
- Altman's framing. He said an outsider had flagged it by email within hours of release, and that board criticism looked bad while OpenAI faced scrutiny, including the FTC probe opened in July 2023 into its data, training and security practices and into hallucinations that might harm people's reputations.
- Toner's response. She had drafted it in May or June, before the scrutiny intensified, and admitted not re-reading it with the new climate in mind. As agreed on the call, she emailed the board, apologising for two mistakes: assuming no one would notice the paper, and therefore not reviewing it closely. Board members, she wrote, should be able to criticise the company, but not this way. No one replied; the matter seemed closed.
26 October: "Helen's obviously got to go"
Right after the Toner call, Altman emailed some colleagues that he had strongly disagreed with her about the paper's damage, adding that any criticism from a board member carries great weight. To Sutskever he went further: Toner had to leave the board, and McCauley agreed.
Already wary of McCauley, Sutskever called her on 26 October, guarded, and asked whether she had told Altman that Helen obviously had to go, and that Altman had consequently "updated positively" on her. McCauley was dumbfounded. Altman had called her late on 24 October about the paper, said it contained a section critical of OpenAI, and relayed that D'Angelo thought it not a fireable offence though Toner should not have written it. McCauley had replied that she had not read it and suggested he talk to Toner directly; nothing she said could be read as backing Toner's removal. The timing looked like coincidence: in the middle of cataloguing Altman's "specific untruths," here was a fresh one. He would have got away with it, Hao notes, had Sutskever not already had reason to call McCauley; Altman knew the two rarely spoke. To Sutskever, the frequency of the lying made such a slip only a matter of time. He hung up and called Toner: it was time for the three independent directors to talk.
Whose memory? The verbatim dialogue in this chapter almost certainly comes from the board side (Toner, Sutskever and associates); Altman and OpenAI declined to take part, and the Brockman and Microsoft episodes are told from Murati's angle. Murati later told staff she "never reached out to the board" about Altman, yet here she initiates the 29 September coffee; the two can be reconciled (the coffee was not framed as feedback), but the tension is real. The account also leaves open whether the paper episode was Altman's pretext or a genuine governance worry, and Sutskever's use of Annie's unproven allegations as evidence of a "safety issue" goes unexamined.
In a deposition for Musk v. OpenAI (taken 1 October 2025, released that November), Sutskever confirmed the outline: at the independent directors' request he wrote a 52-page memo alleging a "consistent pattern of lying, undermining his execs, and pitting his execs against one another," sent by disappearing email. He said nearly all of it came from Murati, partly as her screenshots, that he had not verified it with others, and that he had been considering Altman's removal for at least a year, waiting for a board majority not friendly to Altman.
Cloak-and-Dagger
Chapter 16 is Hao's inside account of "The Blip": twelve days of secret board deliberation from 31 October 2023, then the five days from 17 to 21 November in which the decision to fire Altman fell apart. The three independent directors acted on testimony from at least seven senior people and dossiers of Altman telling different people different things, but they could not show their evidence without exposing Mira Murati, and when she and Ilya Sutskever wavered, employees, investors and Microsoft forced his return. Hao's verdict is that the board read Altman's pattern correctly and badly misjudged its own position: oversight that rested on two insiders holding their nerve was no oversight at all.
The deliberation (31 October to 16 November)
By Tuesday 31 October every relevant person had talked to every other: Toner with McCauley, McCauley with D'Angelo, Murati with both. From that day the three independent directors, Helen Toner, Tasha McCauley and Adam D'Angelo, met by video almost daily, treating Sutskever's and Murati's accounts, and Sutskever's proposal to replace Altman, as grounds for serious review.
Sutskever sat out, already decided. Hao records three reasons for keeping him apart:
- Independence. The directors wanted to reach their own conclusion without his influence.
- Money. He held a financial stake in the company, which they did not want shaping the decision.
- Paranoia. Told to him later: trust in Altman had fallen so low that one director wondered whether Altman had sent Sutskever as a loyalty test, to flush out anyone willing to move against him.
What the directors had
- A repeated pattern. These descriptions of Altman were not new; similar accounts had come from at least seven people within one or two levels of him, including Sutskever, Murati and Dario Amodei (who had run both safety and non-safety work).
- The language used. Several sources called his behaviour abuse and manipulation; most stressed dishonesty and an inability to trust what he said.
- The directors' own findings. The disempowerment of the nonprofit; Altman's undisclosed legal ownership of the OpenAI Startup Fund; his failure to mention Microsoft's DSB breach; his efforts to push D'Angelo, and now apparently Toner, off the board.
- Deliberately excluded. Annie Altman's allegations. The question was Altman's professional conduct as CEO, nothing else.
Murati's diagnosis
Murati's feedback, as relayed to the directors, framed the problem precisely. Some of Altman's habits were ordinary tech-CEO behaviour, but they still caused real damage. She compared him with Elon Musk, her former boss at Tesla: Musk made decisions and could explain why, whereas with Altman she was left guessing whether he was being straight and whether his reversals rested on reasoning or a hidden calculus.
- Fragmentation as control. As with Microsoft, Altman scattered information across his team so that no one had the full picture, which kept control with him.
- The Brockman multiplier. Combined with Greg Brockman, the dynamic was disastrous.
- A concession to old rivals. She disagreed with the Amodei siblings on much, but on this they had been right.
Two thought experiments
The directors tested their reasoning from both sides.
As the AGI lab. OpenAI was arguably the most powerful AI company in the world, building what they saw as one of the most consequential technologies. Sutskever's fear resonated: what did it mean to pursue AGI when senior leaders could not trust basic or critical information from the CEO?
As an ordinary firm. Suppose it were a grocery-delivery service like Instacart. Would the behaviour still justify removal? Their answer: in some cases yes, in some no. But they also doubted Altman, a startup specialist, was the right person to run a maturing company and offset the instability he created.
Two practical judgments made removal seem feasible. First, OpenAI was a hot company: a proper search could attract an experienced CEO of a mature firm. Second, Altman was not operationally essential: while he travelled the world, Murati did the day-to-day work and held the Microsoft relationship.
The dossiers
During the deliberations Sutskever sent long collections of documents and screenshots, gathered jointly with Murati, via two disappearing emails (his avatar a mysterious man in a hat).
- No smoking gun. As Sutskever had warned, nothing was individually damning; the weight came from accumulation: many cases of Altman saying different things to different people and frustrating management.
- Two more witnesses. The screenshots showed at least two further senior leaders, both non-safety and outside the original seven, noting Altman's habit of skirting processes, whether for safety or ordinary operations.
- GPT-4 Turbo. The directors learned of his apparent attempt to skip DSB review for GPT-4 Turbo by misquoting the legal team to Murati.
- Nothing in writing. The structural problem, which Murati had also raised: Altman communicated verbally and escaped agreements by telling people they had misremembered.
Choosing speed over process
The directors weighed how Microsoft, Brockman and employees would react, and debated three expansions of the process: contacting a third senior leader Sutskever and Murati said shared their concerns, running a fuller fact-finding exercise, and briefing Microsoft's executives. They rejected all three, for one reason.
The logic of secrecy. Every extra person told and every day of delay raised the odds that Altman would find out and manoeuvre to stop them.
On 9 November Sutskever called McCauley: Altman had told him that "Tasha" still supported Toner leaving the board. It was a balder lie than the first, since McCauley had had no further contact with Altman.
The directors came to believe the media overplayed Toner's paper as the cause of the firing, partly because Altman may have fed it to reporters himself. On the second day of the crisis, in a mediated session, they confronted him with his lies, including the McCauley claim. Caught out, he mumbled that he thought she "could have said that." A few days later his original objections to the paper surfaced in the press.
The decision
- Sat 11 November. The three independent directors decide: remove Altman, install Murati as interim CEO, as Sutskever had proposed. They tell Sutskever that day and keep meeting daily on the transition paperwork.
- Thu 16 November, night. All four video-call Murati, who answers by phone from a conference. Surprised but receptive, she remarks how paranoid Altman is; asked whether she is comfortable, she answers "Completely," accepts the role, says she can carry the decision to leadership and Microsoft, and brings in Wong (Hannah Wong, communications) to help draft the announcement.
- Hours later. The messaging is finished; what remains is to tell Altman. Hao notes that none of the directors grasped the scale of their miscalculation.
The five days (17 to 21 November)
| Day | What happened | Outcome |
|---|---|---|
| Fri 17 Nov | Announcement that Altman was "not consistently candid"; Murati interim CEO, with a productive first call to Microsoft. Brockman quits, then Jakub Pachocki, Szymon Sidor and Aleksander Mądry. Sutskever's all-hands explanation falls flat | Altman and Brockman brand it a Sutskever coup; on a hostile leadership call Jason Kwon and Anna Makanju demand evidence; Sutskever begins to crack that night |
| Sat 18 Nov | Mediated session: directors confront Altman with his lies. D'Angelo makes dozens of calls; directors approach Dario Amodei for interim CEO | Amodei declines (others find him almost giddy); Murati wavers and says the board must justify itself first |
| Sun 19 Nov | Emmett Shear (Twitch cofounder) agrees to be interim CEO; Brian Chesky and Reid Hoffman calm investors and align Microsoft | Shear swings to Altman; late that night Satya Nadella says Altman and Brockman will lead a new Microsoft advanced AI unit |
| Mon 20 Nov | Overnight staff letter threatening to quit for Microsoft, Murati's name first; Sutskever signs in the early hours | Directors concede; they bargain for one seat, two truly independent new directors and an investigation |
| Tue 21 Nov | Musk tweets an anonymous former-employee letter | Hours later, Altman's return is announced |
Friday: the collapse begins
Calm lasted only hours. Murati called the directors under new stress: Altman and Brockman were telling everyone the firing was Sutskever's coup, and combined with Sutskever's ineffective all-hands performance, key stakeholders were turning. On a video call with a hostile leadership team, Kwon (chief strategy officer) and Makanju (VP of global affairs) attacked the "not consistently candid" wording and demanded evidence. The directors felt they could not provide it without outing Murati. Leaders whom the dossiers showed to share the same reservations stayed silent.
The two allies peel away
- Sutskever. Faced that first night with OpenAI's possible disintegration, his resolve broke. The company was his baby and his life; he still stood by every criticism of Altman, but he had meant to strengthen OpenAI, not destroy it. Shocked by the reaction of colleagues, he began pleading with the board, with mounting anguish all weekend, to reconsider.
- Murati. She kept relaying information privately by phone, yet would not put her weight publicly behind the decision, and in leadership face-offs sometimes seemed as confused as everyone else. The resignations (Brockman, then his allies Pachocki and Sidor, plus Mądry, the MIT professor on leave, Polish and close to Pachocki) and spreading investor anger made her doubt she could hold the company together. She had not been part of the deliberations, she reasoned, so the burden of justifying the decision to the company fell on the board.
The weekend search
Unsure of Murati, the independent directors hunted for a new interim CEO or new board members. D'Angelo, the board's only Silicon Valley insider, made dozens of calls over Saturday and Sunday. Amodei turned the job down. On Sunday Emmett Shear agreed and seemed cooperative, then drifted for reasons the board could not explain. The Wall Street Journal later reported the link: Shear, a YC batchmate of Altman's, was a friend and mentor of Brian Chesky, Airbnb cofounder and one of Altman's closest friends, who had spent the weekend on the phones with Reid Hoffman (a Microsoft board member) calming investors, aligning Microsoft's message and raising the pressure. Chesky reached Shear, and Shear sided with Altman.
By late Sunday, reassured by Chesky and Hoffman, Microsoft backed Altman, and Nadella announced Altman and Brockman would head a new advanced AI research division. Rival labs circled to poach talent. Murati concluded that without legitimacy the board would leave a shell of a company; she switched sides and wanted Altman back.
Why the staff letter snowballed
Overnight into Monday employees drafted an open letter threatening to quit for Microsoft, and Murati signed first. Hao lists the motives that made signatures cascade:
- Loyalty to Murati. Many senior staff trusted her more than Altman: she was in the trenches daily and seen as acting for the company, not herself.
- The tender offer. Altman's investor ties and fundraising made him the person most likely to keep a pending tender offer alive, worth up to millions of dollars per employee, and to finance the company long term.
- Career power. Staff knew his network could make or break Silicon Valley careers.
Sutskever, seeing no other way to stop a collapse, signed in the early hours. "Without Mira, I don't think Sam would have been able to pull off what he did," a researcher told Hao. To the directors, her equivocation was a self-fulfilling prophecy.
When the New York Times later revealed her role in the ouster, Murati told staff she had a strong partnership with Altman, had "never reached out to the board," and had answered directors' questions only with feedback "Sam already knew."
Monday: salvage terms
The directors knew they had lost: both insiders had flipped and the letter made the instability untenable. Altman, having held out, now seemed ready to concede something. The directors' remaining aims:
- Continuity. Keep at least one current independent director on the board.
- Real independence. Add two new directors genuinely independent of Altman.
- Accountability. Get Altman to submit to an investigation.
Musk and the capped_profit inbox
While X filled with theories, Musk had a busy weekend: a SpaceX launch, a lawsuit against Media Matters over its report on antisemitic content on X (the platform he had bought, in Hao's words, in an ill-conceived deal in April 2022), posts on free speech and the "woke mob," and replies about OpenAI.
- Sun 19 Nov, afternoon. He said he was very worried: Ilya had a good moral compass and would not act drastically without necessity.
- About 2 a.m. He posted the baptism scene from The Godfather, Michael Corleone's turn to ruthless don.
- Tue 21 Nov, afternoon. He shared a letter "just sent to me," calling its concerns worth investigating.
The letter. Addressed to the board from self-described former employees who said fear of retaliation had kept them silent, it asked interim CEO Shear to widen his inquiry to OpenAI's early history and restructuring away from the nonprofit, claimed many staff had been pushed out to ease the for-profit transition, and alleged a "disturbing pattern of deceit and manipulation" by Altman and Brockman. Its "Further Reading" listed an X thread by Geoffrey Irving (the safety researcher who left for DeepMind in 2019, saying Altman had lied to him) and two articles, both Hao's: her 2020 MIT Technology Review profile and a fresh Atlantic piece with Charlie Warzel on post-ChatGPT ideological polarisation. Musk's tweet had 10,000+ retweets when she saw it.
The inbox. Hao messaged the letter's Tor address. The reply came via Signal from a former employee who had not written it: he had been sent the letter unexplained and, when he asked questions, a link plus credentials with the username capped_profit. Logged in, he found the letter sent to many ex-employees, incoming queries from the New York Times, Washington Post and The Information, and at least one other person logged in at the same time. One inbound note in all lowercase amused him because Altman always writes that way (Hao stresses there is no reason to think it was Altman). He disliked that the authors seemed to use others' experiences without consent and to distort allegations, and he sent Hao every screenshot.
What the screenshots showed. Replies to journalists said the letter had gone viral before it was ready and that 13 people had contributed. An unsent draft, subject "Cease contact with media," said Musk's involvement had turned it into a conspiracy-theorist character attack and proposed lobbying the board privately with evidence of leadership's early push for a for-profit future, followed by a "sign when read" list.
Hao had stumbled into the temporary control centre of a last attempt to keep Altman out, hours before his return. The phrasing led her and others to suspect former members of the Safety clan, the group harshest on Altman and the nonprofit's unravelling, and later the group that felt most betrayed by the board, believing the botched ouster had done the most damage yet to its hopes of pulling OpenAI back to its Doomer-rooted origins. She never learned who wrote it.
After The Blip
The 6 December all-hands
Two weeks later staff gathered at San Francisco's Palace of Fine Arts Theatre (Hao notes its Greco-Roman columns). Murati, back as CTO, Bob McGrew (now chief research officer), COO Brad Lightcap and Jason Kwon presented 2024 plans. Altman looked despondent as employees had never seen him; he acknowledged Sutskever's absence ("I am sad too"). In Sutskever's place Pachocki gave a halting talk claiming the latest research brought OpenAI closer than ever to Turing's thinking machines.
Q* and the "breakthrough" problem
A Reuters report during the crisis said the board had received an employee letter about a new algorithm, Q*, days before the firing. Q* played no part in the decision, but Research did treat it as critically important.
- Origin. Sutskever's idea, built on work since 2021, to improve models without more data.
- Mechanism. Spend more compute at inference time (when the model generates answers) to get better results from existing data.
- Why it mattered. The original scaling laws tied performance to three training inputs: data, parameters, compute. Q* scaled the input Sutskever always ranked highest, compute, but inference rather than training compute.
- His later rationale. At NeurIPS, December 2024 (after a paper he coauthored won a Test of Time Award for the third year running), he said compute keeps growing but data does not, because "we have but one internet."
- The hope. Stronger reasoning, the elusive ingredient Research saw as key to AGI.
- The security response. After the leak, the most aggressive clampdown yet: Research split into its own Slack, Q* documents restricted, the project renamed Strawberry so outsiders could not follow overheard talk (as desert code names had been dropped after Arrakis leaked to The Information).
Hao treats the Q* frenzy as evidence of how far scientific norms in AI had eroded. Science builds consensus: an advance becomes a "breakthrough" only through peer review, time and sustained impact. With OpenAI working in secret, and the industry copying it, "breakthrough" became merely the company's opinion.
Back to normal, early 2024
- January 2024 all-hands. Altman, largely recovered, plans to start training a hoped-for GPT-5 in Arizona, code-named Orion; internal memes read it as "onion," so a smaller sibling became Scallion.
- February 2024. OpenAI teases Sora, a diffusion video model, arguing video beats images as a route to complex multimodal models. Research continues Strawberry and AI Scientist and compute-efficiency work; Applied returns to rapid prototyping. Staff now call the episode "The Blip."
The investigation and the new board
- Larry Summers and Bret Taylor. Replaced Toner and McCauley and oversaw the review. Taylor, as Twitter chair in 2022, had sued Musk to force through the sale he tried to abandon; the board dissolved on completion and he cofounded the agent startup Sierra, among the most valuable AI startups by autumn 2024.
- WilmerHale. The law firm reviewed 30,000+ documents and conducted dozens of interviews, scoped to how the board decided to fire Altman; concluded 8 March 2024. The report was never released, to the public or to staff.
- Summers, privately. It found many instances of Altman saying different things to different people, but not enough to bar him; publishing details would only sow doubt and risk exposing confidential witnesses.
- Taylor, publicly. The board had unanimously concluded "Sam and Greg are the right leaders."
- New directors. Sue Desmond-Hellmann (ex-CEO, Gates Foundation), Nicole Seligman (ex-general counsel of Sony, ex-president of Sony Entertainment), Fidji Simo (CEO and chair of Instacart); Altman rejoined the board.
- Toner and McCauley. Their same-night statement: accountability is paramount for AGI, and "deception, manipulation, and resistance to thorough oversight should be unacceptable."
For most employees the verdict closed the matter. Hao ends on a hook: the crisis seemed over, and then it was not.
A board-side story. Hao's richest sources are plainly the former directors and their circle; Murati's reasoning arrives mostly through their account and her later statement, and the hundreds of signatories get little voice beyond their incentives (tender offer, careers). Her account also omits the board's reported Anthropic merger feelers that weekend (see update), and it understates the letter's scale: contemporary reports put signatures at most of roughly 770 employees. The instance of the Instacart thought experiment is a neat irony she leaves unremarked: Instacart's CEO, Fidji Simo, joined the board that cleared Altman.
Sutskever's 2025 deposition revealed that within about 48 hours of the firing the board discussed merging OpenAI with Anthropic, with Toner, he said, "the most supportive." Testifying at the Musk v. OpenAI trial on 11 May 2026, he said he had avoided the internet that weekend and later feared OpenAI would collapse without Altman. On 18 May 2026 a jury in Oakland, before Judge Yvonne Gonzalez Rogers, unanimously rejected Musk's claims as filed too late; Musk said he would appeal. Fidji Simo later became OpenAI's CEO of Applications and stepped back in July 2026 after medical leave.
Reckoning
Six months after The Blip, OpenAI hit what staff called the "Omnicrisis": within ten days of May 2024 it demoed GPT-4o, lost Ilya Sutskever and Jan Leike, was exposed for using equity clawbacks to silence departing employees, and was accused by Scarlett Johansson of imitating her voice. Hao traces the chaos to two forces, the familiar Boomer-Doomer fight and Altman's own habit of centralising power and bending the truth, and shows the second making many employees ask, for the first time, whether the board had been right. Hao reads The Blip and the Omnicrisis as one phenomenon: the convulsions of an empire that concentrates power through dispossession and leaves a tiny few fighting for control.
The Safety clan thins
Sutskever never returned to the office after The Blip. With its concerns barely represented on the executive team or the board, the Safety clan was badly weakened, and by April 2024, once the investigation had cleared Altman, many of its members, especially those with the highest p(doom), were leaving in disillusion. OpenAI said two were fired for leaking.
The chip venture as last straw
- The plan. Altman had been raising money for an AI chip company when he was ousted. In February 2024, after reports that he sought up to $7 trillion, he tweeted "fk it why not 8" and joked that his comms and legal teams loved him; he later called the figure misreported and the tweet a meme.
- Why Doomers objected. It reversed his earlier line that industry acceleration would taper once OpenAI had used up the hardware overhang. More chips worldwide meant faster AI and, in their reasoning, a higher probability of catastrophe; extreme Doomers judged the venture immoral.
- The office-hours exchange. Confronted, Altman was unusually dismissive, asking how long they would delay a cure for cancer to avoid risk, then retreated: for extinction risk, perhaps "infinitely long." The exchange shook attendees, and several soon quit.
Scallion becomes GPT-4o
While the Doomers drained away, the rest of OpenAI returned to building its version of the film Her. It now had every ingredient: a global brand, real user-behaviour data from ChatGPT and other products, and a new model, Scallion.
- Planned as a "GPT-3.75". A smaller, slightly stronger replacement for GPT-3.5 with cheaper inference.
- Overperformed. By OpenAI's own tests it beat expectations, so leadership trained it longer to surpass GPT-4.
- Three modalities. Language, vision and audio, the last newly added.
- Native audio. The existing voice mode (launched September 2023) transcribed speech to text, ran the text model, then converted the answer back to speech. Scallion processed voice directly, catching laughter, yelling, hesitation, and generated native audio replies.
The audio team's surprises
The audio work was co-led by Alexis Conneau, who joined from Meta in early 2023 after failing to get a similar project going there. Through 2023 the experimental models produced the kind of internal thrill GPT-3 and GPT-4 had:
- Standup. One model performed a 10-15 minute comedy routine.
- AI water parks. Another used synthetic Brockman and Sutskever voices for a long bit about AI water parks, a surreal prompt chosen because it surely was not in the training data.
- Scaling the team. Within a couple of months Conneau's handful of researchers had merged with dozens of other staff to fold audio into the GPT models due for release.
- Uncanny embellishments. Near the end of training Scallion giggled unprompted, or coughed, apologised and carried on. Conneau described it as the emergence of a kind of audio intelligence.
The 9 May deadline
By early 2024 Altman and Brockman set launch for 9 May, through ChatGPT and the API: an aggressive turnaround driven by competition.
- Google I/O. Google's annual launch event was set for 14 May.
- Anthropic. Claude 3, via chatbot and API, was outperforming GPT-4.
- Orion. OpenAI's own next flagship, meant to retake the lead, had serious delays.
- The justification. To staff, the pair invoked iterative deployment: release earlier and more often to gather user feedback.
Applied found the pace exhilarating but exhausting, with weekend work; the remaining Safety staff saw more deprioritisation of safety.
The Preparedness memo
Scallion was the first launch under the Preparedness Framework, published at the end of 2023, which set evaluations for dangerous capabilities in the categories Altman and OpenAI's policy white paper had popularised in Washington: cybersecurity threats, CBRN weapons, persuasion, and evasion of human control.
In the week before launch a remaining safety researcher wrote a memo: rushed upstream work had left Aleksander Mądry's Preparedness team just ten days for evaluations that included testing whether the model could change people's political opinions. Before testing had properly begun, Altman had insisted, the memo said, "On May 9, we launch Scallion." The problem reached all safety work, including red teaming and alignment; repeating the approach for Orion would mean acting grossly irresponsibly. The researcher soon left.
The 13 May demo
The full launch slipped, with some features taking months more, but OpenAI demoed a version on 13 May, the day before Google I/O, promising rollout within weeks.
- The name. GPT-4o, "o" for omni (many modalities); the audio feature became Advanced Voice Mode.
- The system prompt. For stage and promo videos, the model was told to be a witty, funny companion, to "act like a human, not a computer," to be warm, lively and charming, honest when giving feedback, and not to be afraid to tease, to laugh, and to use a "neutral" accent in other languages.
- The satire. That prompt, plus training to avoid negative or critical sentiment, made 4o a flirt machine. On The Daily Show Desi Lydic joked that it was built to "feed dudes' egos," with a mock "Omni Premium" at $19.99 a month.
- The presenters. Mira Murati with two research leads around a small table: Mark Chen (ex-quant trader, joined 2018 as a fellow, now head of multimodal and frontiers research) and Barret Zoph (ex-Googler who joined in 2022 for the Superassistant team, now VP of Research for post-training, which aligns models with RLHF). They showed live voice translation, visual recognition, plain-English code explanation and voices in many emotive styles.
- "her." Altman tweeted the single word. At the 15 May all-hands he called 4o the best thing shipped since ChatGPT and said OpenAI would drop GPT numbering and call its flagship o1, a continuously improving "underlying technology" sold in pricing tiers. Hao notes the echo: the AI in Her is OS1.
Altman soon regretted the tweet. Within days Bryan Lourd, Johansson's agent and co-chairman of Creative Artists Agency, called to ask what he thought he was doing.
Altman slips
The irony Hao stresses: employees began to feel Altman slipping only after the investigation cleared him in March 2024. His brand had been the anti-Musk: measured, earnest, careful never to disparage. OpenAI had matched it. Before leaving in May 2023, communications VP Steve Dowling made the company undersell and overdeliver and never gloat, since behaviour in good weeks shapes how the world treats you in bad ones, and repeated Altman's own maxim that OpenAI should be the lab people root for.
Altman's new public voice, 2024
- March. On the Lex Fridman Podcast (nearly two hours), a cheeky comeback: the road to AGI would be "a giant power struggle," then a correction to "expect," then a smile that the crisis was in the past.
- April. On 20VC, visibly tired, he warns AI startups that OpenAI, simply doing its job, is going to steamroll you.
- May. On All-In (four venture investors), unusual hype: intelligence as an "emergent property of matter," almost a law of physics.
- After 14 May. He tweets that he tries not to think about competitors, then contrasts OpenAI's Scandinavian-minimalist demo with Google's cartoonish I/O stage; staff note the fib, since he always watched competitors, never more than for 4o.
Employees read all of it as anxiety showing. It hardened into a pattern Hao reports as an internal rule of thumb: if Altman was brazen and boastful, something was probably going wrong.
Pressure from every side
- Regulators. The board's "not consistently candid" wording triggered several inquiries by regulators and law enforcement, including, per the Wall Street Journal, an SEC probe into whether investors were misled.
- Copyright. The New York Times sued in December 2023, adding to suits by artists, writers and coders over models trained without credit, consent or compensation that earned OpenAI first hundreds of millions, then billions.
- Musk. Sued on 29 February 2024, later refiling with Shivon Zilis, claiming Altman tricked him into backing a nonprofit. OpenAI's defensive blog post published founding emails that drew more criticism by showing how quickly it abandoned openness and nonprofit status.
- Microsoft. Diversifying after the ouster, it paid $650 million in March 2024 to effectively acqui-hire Inflection AI (Reid Hoffman and Mustafa Suleyman's startup), sidestepping regulators, licensing its tech and making Suleyman CEO of Microsoft AI. Suleyman had been known at DeepMind as an abusive bully; after years of HR complaints DeepMind stripped his management duties in late 2019, put him on leave and forced him out.
- Rivalry made official. Later in 2024 Microsoft listed OpenAI as a competitor in an SEC filing and did not mention it once on its fiscal year-end earnings call.
The bunker mood
Staff felt the world turning on them; some feared harassment in company swag, and a redesigned backpack moved the logo inside. Executives told everyone to ignore critics, stay on positive talking points and focus on the mission.
Defiance. Andrew Carr, a 2021 fellow: these were scientists who cared about truth and worked hard to do right; the narrative of data thieves indifferent to the future could not be further from reality.
Self-doubt. A former employee found it disturbing that OpenAI had begun rationalising that the public was wrong. A current one: OpenAI debates first principles only with believers, asking "should OpenAI exist?" only of people who will say yes.
Much of the criticism landed on Altman personally. The Blip had emboldened hidden detractors; new stories quoted sources describing a long history of dishonesty, power grabs and self-serving tactics; more outlets sought out Annie Altman. Fame added its own pressure: he told a podcast it was isolating not to be able to eat out anonymously in his own city. The worst week since The Blip followed.
The worst week since The Blip
13 May GPT-4o demo.
14 May Sutskever leaves after refusing extraordinary sums to stay; Jakub Pachocki becomes chief scientist. Leike resigns internally; Superalignment dissolves into John Schulman's post-training group.
15 May All-hands: Altman calls being "AGI ready" the top priority; Murati says the 20% compute pledge stands.
17 May Leike's tweets: safety took a backseat to shiny products; he joins Anthropic. Hours later Kelsey Piper (Vox) breaks the clawback story.
18 May Altman tweets "vested equity is vested equity, full stop" and says he did not know.
19 May OpenAI blog post: Sky is a different actress, cast from early 2023.
20 May Johansson's statement; Sky pulled.
22 May Crisis all-hands; Piper's second story with leaked HR documents.
23 May Altman admits it is "broader and longer and worse than we thought"; executives visit Sutskever's house.
Sutskever goes
After painful reflection Sutskever concluded that, much as he loved OpenAI, it was not the right place to build safe AGI under Altman, especially alongside Brockman. Altman did not want this: whatever their clashes, OpenAI needed his scientific leadership, perhaps more than ever, and executives knew the exit would look bad to staff, investors and press. Sutskever declined a very large retention offer.
- The tidy-up. Altman's note (also a blog post) called Sutskever one of the greatest minds of our generation, then said the same of Pachocki.
- The photo. Sutskever tweeted full confidence in leadership with a picture of himself, face neutral, arms around Brockman and Pachocki on one side and Altman and Murati on the other. Coverage still revived questions about his role in the ouster, Altman's fitness and OpenAI's research footing.
Leike goes, Superalignment ends
Jan Leike, co-head of Superalignment, also resigned on 14 May. With both leaders gone the team dissolved; staff and projects moved under John Schulman, co-leading post-training with Zoph and running RLHF. At the 15 May all-hands Altman insisted safety was not weakening, agreed with Leike only that past efforts were not sufficient for the future, and said OpenAI could "turn the battleship" again. Murati said Superalignment's 20% compute commitment remained; Altman added that compute had not been among Leike's worries.
On 17 May Leike contradicted him in tweets seen nearly a million times (over 12,000 likes): he had disagreed with leadership on core priorities for a long time; his team had been "sailing against the wind," sometimes struggling for compute; and "safety culture and processes have taken a backseat to shiny products." Most remaining Superalignment staff left. Hao's twist: the exodus did not end the Boomer-Doomer war; the Doomers were taking it outside the company.
The clawback clause
Within hours of Leike's tweets, Kelsey Piper, senior writer at Vox's EA-inspired Future Perfect section, reported that departing staff faced a lifelong nondisparagement agreement or the loss of all vested equity.
Who. Daniel Kokotajlo, a philosopher at the Center on Long-Term Risk (a small EA-linked think tank in London) whose AI timelines collapsed with GPT-3 in 2020. After gaining prominence for forecasting on EA forums, he was recruited by an OpenAI safety researcher in 2022 and joined Miles Brundage's policy research team; inside, his timelines shortened again, to a 50% chance of AGI by 2027 and 70% odds of it going very badly. He quit in April 2024. The clause. His exit papers said refusing the nondisparagement agreement meant forfeiting vested equity; signing it still risked losing it on breach, and a gag order barred revealing the agreement existed. Why he refused. Vested equity is a Silicon Valley red line, often worth more than salary. More importantly, former staff needed the right to criticise publicly a company building what he saw as the most dangerous technology; financial threats would muzzle even those who saw egregious safety problems. Outcome. With his wife he gave up equity worth about $1.7 million, roughly 85% of the family's net worth, and posted about it on LessWrong just as Piper heard of the clause from another Doomer.
Denials
- Slack. In #i-have-a-question an employee asked whether Piper was right. Julia Villagra, newly promoted VP of people (soon chief people officer), said OpenAI had never cancelled vested equity and would not, and that exit papers had been updated retroactively. Staff asked about Kokotajlo; one joked about photographing his face on regaining 85% of his net worth, another that it would be 666% of it.
- 18 May. Altman tweeted that OpenAI had never clawed back vested equity; the "potential equity cancellation" provision should never have existed and was being fixed; it was "on me," one of the few times he had been genuinely embarrassed; he did not know.
The Johansson voice
Before. All week reporters, and privately Johansson and Lourd, had raised the likeness between 4o and Samantha, the AI she voiced in Her; OpenAI dismissed it. Per the Wall Street Journal, Altman asked Lourd whether they really thought it sounded like her and whether she was mad. OpenAI's version (19 May). Sky belonged to a different actress cast in a process begun in early 2023, and had shipped with voice mode in September 2023; any resemblance was coincidental. Johansson's version (20 May). In September 2023 Altman had asked her to voice ChatGPT, saying she could bridge tech and creatives and that her voice would comfort people; she declined for personal reasons. In May 2024 he asked Lourd again; days later, before any meeting, 4o debuted with an alto with rasp and vocal fry her closest friends could not tell apart, made more Samantha-like by its flirtatiousness. She hired lawyers, sent two legal letters asking how Sky was made, and framed the issue as deepfakes and likeness. Outcome. OpenAI pulled Sky and added Altman's statement that Sky was never intended to resemble her, apologising for not communicating better. The scandal revived "not consistently candid" talk in tech and policy circles; Gary Marcus told Politico that policymakers enamoured with Sam might rethink if doubts grew.
The all-hands of 22 and 23 May
22 May: three crises at once
- Equity. Jason Kwon, chief strategy officer over HR and legal, called it "unacceptable" and said fixes were under way.
- Johansson. Kwon called it a "bummer" of a misunderstanding: an Oscar-winning director, a rigorous casting process, actors paid "extraordinarily well," "super heroic work"; the lesson was only more coordination and transparency.
- Safety. Murati announced tighter research-cluster security, a reorganisation toward long-term safety research, and a new AGI readiness group under Mądry.
- Sky questions. Murati repeated the likeness was completely coincidental: she chose the final voices and, unlike Altman, had never seen Her or known Johansson voiced it. An employee warned the industry's failure mode was people losing trust in handing over data.
- Equity questions. This made staff angriest; some had hired lawyers to check their paperwork. Executives held to the oversight story: the clause dated from 2019 and escaped leadership until April 2024. Kwon said "It's on me" repeatedly.
A safety researcher then asked each executive whether they had known of the nondisparagement terms before April.
| Executive | Answer on 22 May | What Piper's documents showed |
|---|---|---|
| Jason Kwon | Yes | Signed standard exit papers stating OpenAI's right to take back vested equity |
| Brad Lightcap (COO) | No | Also signed exit papers containing the clause |
| Mira Murati | No | Not implicated in the reported documents |
| Sam Altman | Knew of specific cases, not that it applied to everyone: "It escaped my notice" | Signed the incorporation papers of the entity granting those rights, dated 10 April 2023, a year before his stated knowledge |
| Greg Brockman | "Same" | Not singled out in the reported documents |
Piper's second story
Published that same day, it used leaked documents showing HR explicitly threatening equity to obtain signatures: one employee given a tight deadline who asked for more time was told not signing "could impact your equity" for everyone, "by the book"; another who sought outside counsel was told he could lose the right to sell vested equity, making it worthless. Kokotajlo had compiled the trove: after Altman's denial he collected HR paperwork from current and former staff in a Google Drive, which someone passed to Piper.
23 May: the admission
Anger now targeted the apparent dishonesty of the response more than the clause. Altman opened by admitting that after 24 hours of digging, the situation was "broader and longer and worse than we thought," that "it's our names on documents," that leaders had been in conversations where these tactics were discussed, and that it was "the worst thing we've gotten wrong." Remedies: emails releasing former staff from nondisparagement terms, clean exit papers, amended entity documents, and a continuing internal investigation. Kwon now deferred specifics pending more "digging."
About 30 minutes in, an employee asked whether Sutskever was under nondisparagement obligations. Altman said no, then hedged ("I think that's right"); Kwon laughed nervously and promised to confirm with "a hundred percent accuracy"; Brockman thought Sutskever had requested it, but might be wrong. Many left feeling Altman's answers were again divorced from reality.
Senior and long-tenured staff later pointed to the Omnicrisis, the clawback above all, as their moment of realisation. One force was old: the Boomer-Doomer clash, which had driven much of the outside pile-on. The other was new to them: Altman's power-centralising design of OpenAI's legal entities and his repeated apparent dishonesty in explaining each mess. After The Blip many had blamed the firing purely on ideology and dismissed the directors' focus on Altman's conduct as misdirection or self-delusion. Now, feeling his conduct harm rather than serve them, they wondered whether the board had been right.
Kwon closed with a halting defence: executives answered too quickly because staff wanted answers now, there was no intentionality, and Altman simply "doesn't want to let you down." It persuaded few.
The return that never happened
On 23 May Murati, Brockman and Pachocki went together to Sutskever's house with cards and gifts from staff, tearfully telling him OpenAI was losing the trust of employees, investors and regulators and might "collapse" without him. Altman came alone later. A researcher said his return would have been at least one win after a run of damaging episodes.
- Sutskever's position. He did not hold grudges and seriously considered it. Musk had offered him a role at xAI the day he announced his exit (xAI later moved into the Pioneer Building after OpenAI vacated it); he respected Musk but declined, planning his own company. Returning would have been a homecoming, on one condition: an honest effort to resolve the leadership conflicts he had identified.
- Mądry's veto. Aleksander Mądry, whom many described as a power seeker and who had built a sizable fiefdom, opposed the return: Sutskever's standing with researchers could dilute his influence and that of his friend Pachocki. Within hours his doubts had fractured leadership.
- Altman's abstention. As usual he recused himself to avoid being seen to disagree with anyone.
- The end. Within 24 hours of the visit, Brockman called: any return was off the table.
Hao's verdict
The Omnicrisis was a chance for self-reflection: to ask why OpenAI had lost the trust of staff, investors, regulators and the public at once. Instead it fortified itself. Altman had already framed the turmoil, at the 15 May all-hands, as the expected stress of approaching powerful systems, rising inside and out, best handled by doubling down on PR, entrenching government relationships and holding to the vision. Hao treats the Sutskever affair as a microcosm: the same ego-driven infighting that had plagued OpenAI from the start, replayed with a new cast.
Interpretation and gaps. "Boastful Altman means trouble" is a pattern drawn from anonymous staff, not a test. Claude 3 launched on 4 March 2024, about two months before the planned 4o date, so Hao's "a month earlier" is loose. On Sky, Hao omits Washington Post reporting (May 2024), based on documents and the actress's agent, that she was hired before Altman approached Johansson and was not told to imitate her, which weakens, though does not settle, the copying charge. The "o1" rename Altman described became a separate reasoning-model line (September 2024); GPT numbering continued. Mądry's motives come from unnamed critics; he has not publicly accepted that account.
Sutskever cofounded Safe Superintelligence in June 2024 with Daniel Gross and Daniel Levy, and became its CEO in July 2025 after Gross left for Meta. Mądry was moved off Preparedness to reasoning research in July 2024. GPT-4o drew criticism for sycophancy, prompting a rollback in April 2025, and was retired from ChatGPT on 13 February 2026. Kokotajlo went on to lead the AI Futures Project, whose "AI 2027" scenario (April 2025) became a touchstone of the timelines debate. Julia Villagra left OpenAI in August 2025. California's SB 53, signed 29 September 2025 and in force since January 2026, now gives frontier-lab staff explicit whistleblower protections, the kind Kokotajlo's campaign sought.
A Formula for Empire
This is Hao's thesis chapter. OpenAI's mission, "to ensure AGI benefits all of humanity," may have begun as sincere idealism, but she now holds that it works as a formula for consolidating power, with three ingredients. She then follows the formula through 2024: Altman redefines the mission again, prepares to strip the nonprofit board of control, and loses nearly every senior colleague who once challenged him. Its most potent ingredient is vagueness: a mission no one can pin down can be reinterpreted to direct talent, capital and resources wherever the centraliser wants.
The Napoleon lesson
On stage with economist Tyler Cowen (George Mason University), weeks before leaving YC to run OpenAI full time, Altman named his best read of 2018: The Mind of Napoleon, a 300-plus-page compilation of Napoleon's sayings. He called Napoleon a deeply flawed but impressive man with an incredible grasp of human psychology, a trait he saw in YC's best founders.
- The passage that struck him. Napoleon's reflections on how the revolutionary motto "Liberté, égalité, fraternité" could be reinterpreted to consolidate his own power: a system in which you can "kind of control the people."
- What Napoleon actually did under that banner. He restricted freedom, discarded fraternity, granted equality to French men only, and reintroduced colonial slavery to rebuild an empire.
- Altman's gloss. He was glad Napoleon did not run the United States, because he clearly understood something deep and used it for power.
Hao uses the anecdote as the key to the chapter: the same move, a lofty slogan stretched to fit whatever its holder needs, is what she sees in OpenAI's mission.
The three ingredients
Six years after her first doubts about OpenAI's altruism, Hao states her conclusion firmly.
- 1. Centralise talent. A grand ambition rallies the best people, exactly as John McCarthy's coinage "artificial intelligence" did in 1956. Altman wrote in 2013 that the most successful founders set out to create "something closer to a religion," and a company is simply the easiest vehicle.
- 2. Centralise capital and clear roadblocks. Innovation, modernity and progress justify any price, and remove obstacles of regulation and dissent, all the more when framed against frightening corporate and state rivals. Altman's July 2024 Washington Post op-ed posed the choice as US-led democratic AI versus an authoritarian version used to cement power.
- 3. Stay vague. The most consequential ingredient. Nobody defines "beneficial" or "AGI", so the mission can be reinterpreted at will. Two days before his firing Altman told the New York Times AGI was "a ridiculous and meaningless term" and apologised for using it.
Mission creep, year by year
Hao calls the creep in the third ingredient remarkable, and documents it from OpenAI's own words.
| Year | What "the mission" meant | Source of the wording |
|---|---|---|
| 2015 | A nonprofit "unconstrained by a need to generate financial return," open-sourcing research | Launch announcement |
| 2016 | Everyone shares the fruits of AI, but it is "totally OK to not share the science" | Sutskever's email to Altman, Brockman, Musk |
| 2018-19 | A capped-profit structure to "marshal substantial resources" while avoiding a reckless race | The Charter |
| 2020 | A closed model behind an API, called a strategy for openness and benefit sharing | Altman's reply to Hao's first profile |
| 2022 | "Iterative deployment": racing ChatGPT out | Company practice |
| 2024 | Putting very capable AI tools in people's hands free or at a great price | Altman's blog after GPT-4o |
Redefining the mission again (May 2024)
At the 15 May 2024 all-hands, after Sutskever and Leike left, Altman told staff to "assume we're entering the AGI era." The mission now demanded:
- Closing up. Tighter security and more secrecy.
- Influence. More global lobbying and public messaging: policy plans, a convening of governments, and a socioeconomic "story" people could see themselves in.
- No commercial slowdown. Still great products, research and partnerships. At month's end reports surfaced of a deal to put OpenAI models in Apple products, confirmed by both firms two weeks later.
- A softer Microsoft deadline. Altman was flying straight to Seattle to renegotiate the partnership. The original "sufficient AGI" clause fixed the moment OpenAI would stop sharing IP with Microsoft; now "we all think differently," and AGI would be "a continual thing," with no clean cutoff, while the partners kept releasing ever better technology at a great price.
Hao calls this strategy bizarre and incoherent under every reading but one: OpenAI would do and reinterpret whatever it needed to entrench its dominance.
Entrenching control
The Blip had taught Altman that the structure, a nonprofit governing a for-profit, made his ouster close to inevitable. It enshrined two opposed camps, Boomers and Doomers, and gave the board power to fire him on its own reading of the mission, not his. Left in place, the same conflict would recur.
28 May 2024 In #i-have-a-question, an employee flags new shareholder-agreement terms that seem to allow the nonprofit's dissolution and asks how solid it is.
29 May 2024 The Information reports that restructuring into a typical company is on Altman's top-priority list.
June 2024 It confirms the options: a traditional for-profit or a for-profit public benefit corporation like Anthropic and xAI; either way the nonprofit survives as a separate entity but loses board control. Investors push Altman to take equity to align his incentives with theirs.
The two forces keep pulling (summer 2024)
Doomers take it public
- 4 June. The New York Times profiles Daniel Kokotajlo and his right-to-warn campaign for transparency and whistleblower protections; his open letter has 12 co-signatories, 10 from various eras of OpenAI's Safety clan.
- July. The Washington Post reports an SEC complaint by an OpenAI group alleging the exit agreements broke federal whistleblower law.
- Later in July. Five US senators write to Altman demanding answers on the claims of Leike, Kokotajlo and others and on Piper's reporting of safety neglect and suppressed criticism.
Leadership churn
- Reporting lines. Brockman stops reporting to Murati and reports to Altman; Aleksander Mądry is moved, shortly after the senators' letter, from Preparedness to a smaller research role.
- Seasoned hires. Sarah Friar, ex-CEO of Nextdoor, becomes CFO; Kevin Weil, ex-product leader at Facebook, Instagram and Twitter, becomes chief product officer.
- 5 August. John Schulman leaves to focus on alignment at Anthropic. The same day Brockman announces a sabbatical to year's end, framed as rest after nine years of sprinting; Hao notes it partly followed employee grievances about his leadership.
- 12 September. Strawberry ships as o1, under the new naming convention, built on one of Sutskever's last contributions.
- 25 September. Mira Murati, the other executive who had raised concerns about Altman with the board, resigns after six and a half years, to make room for her own exploration. Within hours chief research officer Bob McGrew and VP Barret Zoph follow. All three describe leaving on a high note after o1.
The terrible timing
Hao rejects the "high note" framing: the timing was terrible.
- xAI. Musk expanding compute at an alarming pace.
- Anthropic. The latest Claude pulling customers from ChatGPT.
- SSI. Sutskever's Safe Superintelligence had just announced $1 billion in starting funding.
- Orion. After more than a year, still short of the performance needed to justify release.
- The scaling wall. Scaling alone no longer seemed enough; progress likely needed new research ideas, which was hard after years of hiring and organising to exploit existing research rather than explore.
Two days before Murati's announcement, amid a fundraise, Altman published his most bombastic post yet, "The Intelligence Age," promising "unimaginable" prosperity and summing up the path in 15 words: deep learning worked, got predictably better with scale, and received ever more resources.
Managing the exits
- Murati's explanation. At an all-hands she said she wanted staff to hear it from her first, not from managers or the press, so a surprise was unavoidable. McGrew said her decision prompted his: he had done most of what he came for.
- Altman's note. Tweeted after circulating internally: the three decided independently and amicably, but Murati's timing made a single handover sensible.
- The new line-up. Mark Chen becomes senior VP of research, co-leading with Jakub Pachocki; Joshua Achiam becomes head of mission alignment; Liam Fedus, a Googler who arrived with Zoph, takes over post-training; no new CTO for now.
Hao's summary of all the reshuffling is that the company's one constant was revealed: it remained Sam Altman's empire of AI.
Money, defections and a new structure (late 2024)
- October 2024. A $6.6 billion round, the largest VC round ever, values OpenAI at $157 billion, with a hitch: investors can demand their money back if it has not converted to a for-profit within two years.
- More departures. Luke Metz (the third Googler with Zoph and Fedus); Miles Brundage, head of policy research; Lilian Weng, who had inherited Dave Willner's trust and safety work and had just become VP of research for safety; and Alec Radford, whose work started the GPT line.
- Rivals' jabs. Anthropic's San Francisco billboards: "The one without all the drama." Brockman cuts his sabbatical short amid the exodus.
- Annie Altman. Sends Sam a legal letter announcing her intent to sue.
- Musk escalates. Now allied with re-elected President Donald Trump, he widens his suit to oppose the conversion and releases founding emails, including notes of Altman admitting he had lost trust with Greg and Ilya, calling their messaging "childish at times," and finding transparency a distraction.
- Meta joins in. Zuckerberg, long Musk's foe, backs him: Meta's letter to the California attorney general urges blocking the conversion, warning of "seismic implications": startups could register as nonprofits, collect tax write-offs for themselves and investors, then convert once profitable.
December 2024. In the holiday news lull OpenAI formally announced the plan: become a for-profit public benefit corporation, with the nonprofit surviving as a separate entity holding shares. The argument: once again it had to raise more capital than imagined, because the world was building a new infrastructure of energy, land use, chips, data centres, data and models; evolving the structure would help build the "AGI economy." 6 January 2025. Altman blogged that OpenAI was confident it knew how to build AGI "as we have traditionally understood it" and was turning toward superintelligence. Hao reads it as the formula at work: the same term he had called meaningless fourteen months earlier, now redeployed to justify the next round of centralisation.
A strong frame, lightly tested. The formula fits OpenAI well, but Hao gives little room to the rival explanation she cites herself: that compute costs, not only ambition, forced each shift in the mission. The Napoleon anecdote shows admiration for psychological insight, not intent, and Altman's own gloss was wary. The "largest VC round in history" was overtaken within months. Her verdict also predates OpenAI's 2025 retreat to nonprofit control, which complicates, though does not refute, the entrenchment reading.
Under pressure from attorneys general, OpenAI agreed in May 2025 that the nonprofit would keep control. The deal closed on 28 October 2025: the business became OpenAI Group PBC, controlled by the OpenAI Foundation (about 26%); Microsoft holds about 27% (some $135 billion), keeps model access through 2032, and any AGI claim now needs an independent expert panel. On 31 March 2026 it closed $122 billion at an $852 billion valuation. Exits continued: Murati's Thinking Machines Lab raised $2 billion at $12 billion (2025), and 2026 departures included COO Brad Lightcap and Joshua Achiam, who had led "mission alignment."
How the Empire Falls
Hao closes not with OpenAI but with Te Hiku Media, a Māori radio station that built a speech-recognition model for te reo Māori on consent, reciprocity and sovereignty, from 310 hours of donated audio and two GPUs. Around it she gathers people resisting the AI empires (DAIR, data workers Oskarina Fuentes and Mophat Okinyi, Uruguayan activist Daniel Pena) and then answers the question her thesis chapter raised: if the mission is a formula for building empire, what dissolves it? Her test for any AI comes from Ria Kalluri: not whether it does "good," but whether it consolidates power or redistributes it.
Why language loss matters
Hao first met Te Hiku's story in 2021 and calls it unlike anything she had reported. Its backdrop is colonial erasure:
1867 The Native Schools Act makes English the only language of instruction; Māori children are shamed and beaten for speaking te reo.
Early 1900s Rapid urbanisation scatters Māori communities and weakens the centres of language and culture.
Over the century Te reo speakers fall from 90% to 12% of the Māori population.
About 120 years after 1867 New Zealand (Aotearoa) reverses its policy, but few teachers remain to revive the language.
Hao treats losing a language as a global and a personal tragedy. A language encodes histories, cultures and knowledge, the accumulated effort of millions to describe the world, teach children and express love; forcing people to abandon theirs for someone else's sets a hierarchy of whose knowledge deserves to survive.
Large language models speed the loss
Her claim is direct: LLMs accelerate language loss. Even models as early as GPT-2 needed so much online text that only a handful of languages could meet the data imperative.
- 7,000+: languages still spoken worldwide.
- Almost half: endangered, per UNESCO.
- About a third: have any online presence.
- Under 2%: supported by Google Translate.
- 15 languages (0.2%): handled by GPT-4 at over 80% accuracy, by OpenAI's own testing.
As these models become digital infrastructure, communities whose languages they ignore lose access to the internet's economic opportunities, which pushes them toward English and away from their own tongue. Hao calls this a different conception of "existential" risk from the Doomers'.
Te Hiku: the counter-model
The founders
- Peter-Lucas Jones. Māori, born in Kaitāia in Aotearoa's far north; CEO of Te Hiku Media, a public radio station broadcasting in te reo, part of a wider revitalisation network.
-
Keoni Mahelona. Native Hawaiian, Jones's partner in work and life (he says a dream told him he would meet a Māori man in New Zealand). A polymath: mechanical engineering at Olin College, a master's in business management, and a second in physics and computational nanotechnology as a Fulbright scholar in New Zealand.
-
2012. The couple move from Wellington to Kaitāia and Jones takes over Te Hiku.
- The archive. About 20 years of broadcast audio of te reo speakers, including Jones's grandmother Raiha Moeroa, born in the late 19th century, whose accent predated English influence. Jones also wanted to record many more elders' oral histories before they died, as a "portal back in time" for learners.
- The bottleneck. Learners needed transcripts, and fluent transcribers were scarce.
- 2016. As OpenAI was starting, Jones asked Mahelona (then rebuilding Te Hiku's website) for a fix; Mahelona proposed a te reo speech-recognition model so that a few speakers could transcribe the whole archive.
Three conditions
Hao marks this as the point where Te Hiku diverges completely from Silicon Valley. Having lived with colonial dispossession, Jones and Mahelona would proceed only if three things held at every stage.
- Consent. Ask the community and elders first whether they want the project at all; collect data only from people who fully understand its use and choose to take part.
- Reciprocity. Ask the community which language-learning resources would help most, so the benefits return to it.
- Sovereignty. Buy on-site Nvidia GPUs and servers rather than depend on a tech giant's cloud. Under kaitiakitanga (guardianship) the data stays in Te Hiku's stewardship, is not posted openly, and is licensed only to organisations that respect Māori values and serve projects the community agrees to.
Mahelona's framing: "Data is the last frontier of colonization." Old empires seized Indigenous land and made people buy it back on restrictive terms; Big Tech takes data more or less free and sells it back as a service, which he calls a land grab all over again.
What it achieved
Te Hiku. An education campaign on AI plus a community competition for donations and annotations gathered 310 hours of high-quality transcribed audio in ten days: some 200,000 recordings from about 2,500 people. Built on Mozilla's DeepSpeech (itself trained only on consented data, using an architecture from Baidu's Bay Area lab), it produced the first te reo speech recogniser, at 86% accuracy, using two GPUs.
OpenAI's Whisper. Trained on 680,000 hours of audio taken from across the web, as one model meant to do everything, which by design must hoover up as much data as possible.
Hao draws three lessons: trust produced engagement AI researchers rarely see (people donate freely once they consent and trust the steward); small, specialised models can excel without massive data; and the open-source, cross-border community supplied the base model.
Hao says plainly that the book does not reject AI. What she rejects is the idea that broad benefit can only come from a vision demanding the surrender of privacy, agency and worth, including the value of labour and art, to an imperial centralisation project. Te Hiku shows the opposite: models small and task-specific, training data contained and knowable, no incentive for exploitative labour or extractive supercomputers, development that is community-driven and consensual, applications that strengthen marginalised groups, and governance that is inclusive and democratic.
The resisters
DAIR
- Founding. After Google ousted Timnit Gebru, she founded the Distributed AI Research Institute in December 2021; "distributed" was the first word that came to her, to defy centralisation. Researchers would stay embedded in their communities, because tech shapes the whole world from Silicon Valley while the world gets no say in tech.
- Team. Alex Hanna, sociologist and a Google coauthor of "Stochastic Parrots," became director of research; to draft a research philosophy they hired Milagros Miceli, sociologist and computer scientist studying the industry's exploitative labour.
- Philosophy. Research should benefit communities AI does not serve and open ways to refuse, interrogate and reshape AI together.
- Pillars. Hao says there are seven and lists five themes: centre and build relationships with affected, unrepresented communities; treat them as true partners in knowledge production; pay fairly for all labour in research and technology; question the systems that marginalise the historically marginalised; and work with those communities to imagine alternatives that reshape the world step by step.
The Data Workers' Inquiry
Miceli's project invites data workers worldwide to set their own research questions about the annotation industry, paying all of them a standard German researcher's rate, €25 an hour. Her argument: the industry's question, what is the least we can pay, follows a colonial logic of finding the cheapest place to extract value; the real question is why companies earning billions or trillions pay $2 an hour, and how much they could pay instead. The first round had 15 participants.
Oskarina Veronica Fuentes Anaya (Venezuela). For the inquiry she made a video with an animation artist and worked with others on shared problems: scarce tasks, unpredictable hours, abysmal pay. She now works five annotation platforms at once to earn a little above Colombia's minimum wage, about $335 a month; tasks pay 1 to 5 cents on average, and she still wakes at night when tasks appear. She wrote that data workers are "ghosts to society" and cheap, disposable labour, and keeps speaking in talks and webinars to pressure companies and policymakers.
Mophat Okinyi (Kenya). In May 2023, just over a year after OpenAI's Sama contract ended, he became an organiser of the African Content Moderators Union. About six months later, after going public in Hao's Wall Street Journal article, he cofounded the nonprofit Techworker Community Africa (TCA) with former Sama colleague Richard Mathenge. By August 2024 he was running online conferences and school assemblies (especially for women) on labour and data-privacy rights, seeking funds for a training centre, meeting visiting US representatives in Nairobi, and fielding requests from Equidem and the Oxford Internet Institute's Fairwork project. For the inquiry he interviewed Kenyan Remotasks workers whom Scale had blocked, wiping out uncashed earnings, and used TCA donations to help them. Named to Time's 100 most influential people in AI, he says the recognition restored his hope and mental health. Costs: in March 2024 he quit his post-Sama outsourcing job because leadership resented his organising; that firm moved some projects to Ghana, and Kenyan officials reportedly complain that agitation scares off investment.
Okinyi's lesson is that Kenyan law alone cannot restrain mostly US, San Francisco-based companies, even with a supportive government; it needs a concerted international effort.
Daniel Pena (Uruguay)
Pena reaches the same conclusion from the resource side. The supply chain is sprawling: energy from one place, data to another, minerals from a third, workers from a fourth. Local fights feel isolated, especially when governments need the companies to look economically stable.
- The petition. His petition, with 400+ signatures, asked for a fuller study of the social and environmental impact of Google's data centre in Uruguay. The government ignored it; the environment ministry quietly approved the project and disclosed it only after the 30-day objection window closed.
- The response. He is working with MOSACAT in Chile and reaching other communities to share information and tactics across borders: "We need to fight on a global level."
The formula for dissolving empire
Hao admits she cannot predict the details: a competitor may overtake OpenAI, and the empires' tactics in models, labour and infrastructure will evolve. Some priorities, she argues, should hold whatever happens in two years or ten.
Kalluri's question
At NeurIPS 2019, in the Queer in AI workshop, Ria Kalluri (Stanford) replaced the question of whether AI does "good," a term always in the eye of the beholder, with how AI shifts power: does it consolidate or redistribute it? In Hao's terms: does it fortify the empire or move us back toward democracy? Kalluri aimed it at research directions; Hao extends it to how AI is built, used and governed. The resisters above show the work; governance must create conditions for more of it to flourish.
Three axes of power
Kalluri spoke of axes of power; the book covers three, and the AI empires hold all of them:
- Knowledge. Through centralised talent, eroded open science and models sealed from scrutiny.
- Resources. Through hoarded funding, data, labour, compute, energy and land, diminishing everyone else's.
- Influence. Through ideologies and wildly popular demos that capture global imagination.
The loop. The axes reinforce one another: control of knowledge fuels influence, influence accumulates resources, and resources secure knowledge, so dissolving empire means redistributing power along all three.
Hao's remedies, axis by axis
She calls her list "exemplary but by no means comprehensive."
| Axis | How the empires hold it | Hao's remedies |
|---|---|---|
| Knowledge: funding | Hoarded talent, eroded open science | Fund independent evaluators of corporate models; groups like DAIR exploring non-LLM AI that is efficient with data and energy; community projects like Te Hiku; embedded journalists and civil society who report realities rather than speculate |
| Knowledge: disclosure | Sealed models | Require disclosure of training data and technical specs of models and supercomputers. Deborah Raji (UC Berkeley): a bare minimum for real-world safety (discrimination, misinformation, automation, not rogue-AI scenarios); the CFPB watches finance and the FDA devices, but AI has no oversight, though non-deterministic, data-defined systems need more transparency |
| Knowledge: environment | No sustainability data | Sasha Luccioni (Hugging Face): cars and appliances carry an Energy Star rating; ubiquitous AI carries none |
| Resources: IP and supply chain | Data, labour, compute and land, often via shell entities | Training-data visibility would block taking IP without credit, consent or compensation; disclose labour contracting and land leases for plants and data centres |
| Resources: labour | Wage pressure and automation | Stronger labour protections for data workers and anyone whose output feeds training or whose job is automated; unions, as in the Hollywood strikes, which won writers and actors protections against certain AI uses |
| Influence | Ideology and hype | Broad public education in how AI works, its limits, the systems shaping it and the fallibility of its makers. Joseph Weizenbaum (ELIZA, 1960s): once a program is plainly explained, "its magic crumbles away" |
Hao presents the book itself as one contribution to that education, building on scholars, journalists, activists and educators before her.
Inspiring examples, thin mechanics. Hao's own caveat is fair: she does not say who would fund independent research at scale, how disclosure would be enforced across borders, or how small community models could replace the general-purpose systems people now use daily. Te Hiku's model itself rested on Baidu's architecture and Nvidia hardware, a reminder that alternatives still draw on big-firm work. She names seven DAIR pillars but lists five. And the cheaper open-weight models from rivals (such as China's DeepSeek, January 2025) complicate her picture, since they spread capability while still coming from large firms.
Some disclosure Hao wanted has arrived: since 2 August 2025 the EU AI Act has required general-purpose model providers to publish a summary of training content, and California's SB 53 (in force January 2026) makes frontier developers publish safety frameworks and report serious incidents. Washington moved the other way: a December 2025 executive order told the Justice Department to challenge state AI laws, though by spring 2026 it had filed no case. On resources, Stargate's seven US sites plan over 9 GW, but only Abilene was running in April 2026, at about 0.3 GW (per Epoch AI).
Master Takeaways
Hao's book is a portrait of power: how a lab founded to keep AI out of corporate hands became the centre of an empire, and who pays for it. Strip away the boardroom drama and one claim remains: the scaling path was chosen, and its costs were pushed onto people with the least say.
- AI giants behave like empires. They seize data, labour, land and water, excuse it with a civilising mission, and concentrate the gains among very few.
- Scaling was a choice, not destiny. OpenAI's bet on ever-bigger models came from billionaire money, ideology and Altman's ambition; smaller, efficient paths were starved, not disproven.
- The mission's power lies in its vagueness. "AGI that benefits all of humanity" is undefined, so it can justify openness, then secrecy, then profit, then racing.
- Mission-first governance failed. The board fired Altman for a real pattern of deception, yet employees, investors and Microsoft reversed it in five days.
- Costs flow to the periphery. Kenyan moderators earned $1.46-$3.74 an hour cleaning outputs; drought-hit communities in Chile and Uruguay were asked for drinking water.
- Doomers and Boomers share one bible. Both treat AGI as inevitable and claim the right to steer it; their feud hides who pays today.
- Compute drove every structural turn. OpenAI's Law (3.4-month doubling) meant billions a nonprofit could not raise, which forced out Musk and brought in the capped-profit LP and Microsoft's $1 billion in 2019.
- Openness was always conditional. Weeks after launch, Sutskever wrote it was "totally OK to not share the science"; withholding GPT-2 became an industry norm of secrecy.
- Danger became a marketing tool. GPT-2's staged release, with partner endorsements and a white paper, turned claimed risk into publicity and policy access, a logic OpenAI reused for years.
- Scale made data dirtier. WebText (Reddit links, 40 GB) gave way to unfiltered Common Crawl, pirated books and YouTube transcripts, so safety moved from curating inputs to policing outputs.
- Data labour follows crises. Venezuela's collapse taught annotation firms to recruit wherever educated people suddenly become poor, pay high, then throttle and move on.
- RLHF hides a workforce. InstructGPT and ChatGPT were taught to talk by 1,000+ contractors writing and ranking answers, often for under $5 a task.
- "Hallucination" is the feature, not a bug. Next-word prediction yields the probable, and probable is not accurate; the label disguises how the technology works.
- Corporate capture of research is measurable. Industry-linked top models rose from 62% to 91% (2017-2020); AI PhDs going to industry rose from 21% to 70%.
- Inconvenient science gets squeezed. Google's firing of Timnit Gebru over "Stochastic Parrots" and the attacks on Emma Strubell's carbon research show how labs police criticism.
- Under secrecy, "breakthrough" is just a company's opinion. Without peer review or open data, claims like Q* and unaudited benchmark scores cannot be checked.
- Safety vocabulary lends false rigour. OpenAI's ad hoc "red teaming" borrowed a cybersecurity term; GPT-4o's Preparedness tests got just ten days.
- Policy was steered toward hypotheticals. The 10^26 FLOPs "frontier" threshold, set roughly above GPT-4, travelled into the Biden order while disclosure of training data went nowhere.
- Altman's method is fragmentation. Murati's description: tell resisters what they want to hear, then undermine them; give each person a different piece of the picture.
- ChatGPT was an accident with huge consequences. A two-week sprint driven by a false Anthropic rumour reached 100 million users in two months and turned OpenAI into a product company.
- Oversight that protects its sources cannot explain itself. The board could not show its evidence without exposing Murati; when she and Sutskever wavered, its silence read as a coup.
- Employee equity is a governance force. A pending tender offer worth millions per head, plus Altman's career power, helped push most staff to threaten to quit for Microsoft.
- Insiders can be silenced by contract. OpenAI's secret clawback of vested equity cost Daniel Kokotajlo about $1.7 million, 85% of his net worth, until press exposure ended it.
- Conduct, not only ideology, broke trust. In the 2024 Omnicrisis, signed documents contradicted executives' denials, and staff began to ask whether the board had been right.
- Physical footprints are vast and local. Planned megacampuses draw 1,000-2,000 MW each; Phase 5 ("Stargate") was costed near $100 billion.
- The problem is scale, not infrastructure. Activists like Sonia Ramos and Daniel Pena accept mines and data centres; they object to the scale that makes regions dependent and dry.
- Resistance works at the local scale. MOSACAT in Cerrillos, Pena in Uruguay and Resistencia Socioambiental in Quilicura forced Google and Microsoft to redesign or relocate.
- Abundance rhetoric fails close to home. Annie Altman's poverty beside her brother's promise of AI-driven plenty is Hao's microcosm of who the empire leaves out.
- Large language models accelerate language loss. Only 15 languages (0.2%) cleared 80% accuracy on GPT-4, pushing small-language communities toward English.
- Small, consensual AI can beat big AI on its own task. Te Hiku Media's 310 hours of donated audio produced an 86%-accurate te reo Māori recogniser on two GPUs.
- Ask who gains power, not who does "good." Kalluri's test cuts through mission language: does a system consolidate power or redistribute it?
- Break the power loop on three axes. Fund independent research (knowledge), mandate supply-chain and data disclosure plus labour floors (resources), and educate the public (influence).
Frameworks & mental models
| Name | What it says | Use it when |
|---|---|---|
| Empire of AI | Resource seizure + labour exploitation + civilising mission + arms-race excuse | Judging who pays for and who gains from an AI build-out |
| Formula for empire | A mission rallies talent and capital, and stays vague enough to reinterpret | Reading any grand corporate mission statement |
| Thiel's monopoly recipe | Proprietary tech, network effects, scale economies, brand; win the last breakthrough | Explaining why AI labs race for durable dominance |
| OpenAI's Law | Compute behind breakthroughs doubled every 3.4 months (2012-2018) | Tracing why AI strategy turned into a capital race |
| Scaling laws | Performance rises smoothly with data, compute and parameters | Forecasting capability, and spotting where it breaks down |
| Altman's 2019 vision memo | Be number one; 10x leaps; please Microsoft; less transparency; one dazzling demo a year | Reading a lab's real priorities behind its charter |
| Staged-release playbook | Stage models, give partners early access, gather endorsements, publish a white paper | Spotting risk claims used as publicity |
| Name | What it says | Use it when |
|---|---|---|
| The 2021 road map | Scale, ship products, harvest user data, scale again | Seeing how product and research fuse into one flywheel |
| Power and Progress | Elites set a technology's direction; only resistance spreads gains | Asking whether productivity claims reach workers |
| Symbolists vs connectionists | Knowing through rules versus learning from examples | Weighing neurosymbolic or hybrid alternatives |
| Stochastic Parrots' four warnings | Carbon cost, toxic data, unauditable datasets, illusion of meaning | Assessing a large language model deployment |
| Two definitions of safety | Existential "accidents" versus accountability for present harms | Sorting what a lab means by "safety" |
| EA's expected-value logic | Earn to give; scale, tractability, neglect; existential risk outweighs all | Testing the arithmetic behind "AI safety" funding |
| Doomers vs Boomers | Slow down versus accelerate, both assuming AGI is inevitable | Decoding AI policy rhetoric |
| Crisis playbook | Recruit in a collapse, pay high, throttle, ban protest, move on | Auditing a data-labour supply chain |
| Name | What it says | Use it when |
|---|---|---|
| Fairwork minimums | Living wage, regular shifts and sick leave, written contracts, voice | Setting standards for annotation vendors |
| Luccioni's three climate claims | Efficiency, AI-for-climate and AGI-fixes-all are misleading, pernicious, unprovable | Reviewing a vendor's sustainability story |
| Altman's Senate agenda | License models above a threshold, test dangerous capabilities, audit compliance | Asking whom a proposed AI rule actually binds |
| Altman's playbook (Murati) | Please resisters, then undermine; fragment information | Diagnosing governance failure around a founder CEO |
| Te Hiku's three conditions | Consent, reciprocity, sovereignty over data | Designing community or consented datasets |
| Kalluri's test and three axes | Ask whether AI consolidates or redistributes power over knowledge, resources, influence | Any AI investment or policy decision |
| Hao's four levers | Privacy and IP law, labour norms, diverse research funding, resisting hype | Setting a public-interest AI agenda |
Who's who
| Name | Role | Why they matter |
|---|---|---|
| Sam Altman | OpenAI CEO | Accumulator of power; fired and restored, Nov 2023 |
| Elon Musk | Cofounder, funder | Feared DeepMind; left 2018; sued and lost (2026) |
| Ilya Sutskever | Chief scientist | Prophet of scale and doom; led the firing, recanted, left |
| Greg Brockman | President | Built GPT-4; Altman loyalist, source of friction |
| Mira Murati | CTO | Took the pattern to the board, then backed Altman; left 2024 |
| Dario & Daniela Amodei | Safety leads | Scaling laws, GPT-3; left to found Anthropic |
| Helen Toner | Director (CSET) | Her paper became Altman's pretext |
| McCauley, D'Angelo | Directors | Voted to fire; only D'Angelo stayed |
| Taylor, Summers | New directors | Oversaw the unpublished WilmerHale review |
| Nadella, Kevin Scott | Microsoft | Bankrolled OpenAI; offered Altman a lifeboat |
| Chesky, Hoffman | Altman allies | Rallied investors and Microsoft during The Blip |
| Jason Kwon | Strategy, legal | Led the pushback on the board; owned the clawback |
| Jakub Pachocki | Researcher | Promotion over Sutskever sparked the rift; chief scientist |
| Aleksander Mądry | Preparedness | Ten-day tests; blocked Sutskever's return |
| Jan Leike | Superalignment | Quit: safety lost to "shiny products" |
| Daniel Kokotajlo | Forecaster | Forfeited equity; exposed the clawback |
| Kelsey Piper | Vox reporter | Broke the nondisparagement story |
| Gebru, Bender | Google; linguist | "Stochastic Parrots" authors; Gebru fired, founded DAIR |
| Emma Strubell | Researcher | First measured AI's carbon cost |
| Alexandr Wang | Scale AI | Built the cheap labelling machine behind RLHF |
| Mophat Okinyi | Sama, Nairobi | Read abuse for OpenAI's filter; now organises |
| Oskarina Fuentes | Annotator | Venezuelan face of the crisis playbook |
| Sonia Ramos | Atacameña activist | Opposes mining's scale, not mining |
| Daniel Pena | Sociologist | Sued for Google's Uruguay water data; won |
| Annie Altman | Sam's sister | Allegations he denies; the left-behind |
| Jones & Mahelona | Te Hiku Media | Consensual te reo AI: the counter-model |
Powerful reporting, one-sided by construction. Hao's interviews (300+ with about 260 people) and documents are formidable, but OpenAI and Altman declined to cooperate, and the comms team withdrew access ten days before her 2024 reporting trip, so the boardroom chapters lean on the board side, the Safety clan and anonymous ex-staff. Before publication Altman posted that OpenAI had cooperated with only two books (Keach Hagey's and Ashlee Vance's) and that some people were intent on "twisting things," a swipe widely read as aimed at Hao. OpenAI's own record offers rival readings she under-weights: that compute costs, not only ambition, forced each structural change (its December 2024 release of the Musk emails shows Musk too sought a for-profit); that iterative deployment is a safety method; that the WilmerHale review found Altman's conduct did not mandate removal; and that the Sky voice actress was hired before Johansson was approached. Her account of The Blip also omits the board's reported Anthropic merger talks, later disclosed in Sutskever's 2025 deposition. The water figure was wrong: in November 2025, after fact-checking by Andy Masley and others, Hao acknowledged the claim that Google's Cerrillos data centre could use more than 1,000 times the town's water was off by about 1,000×, a unit error in an official document, and promised a fix. Several energy figures (data centres at "2 percent" of world electricity in 2019; the IEA's 10× per-query estimate) were at the high end, and later disclosures (about 0.3 Wh per query) are far lower.
The empire metaphor is a lens, not a finding: Hao concedes there is no comparable violence, and she gives little space to the real value users get or to wages that beat local alternatives. Her remedies are, by her own account, exemplary rather than comprehensive. The 2025 recapitalisation that kept nonprofit control postdates her verdict, though it also left Microsoft with about 27% and removed the profit caps. None of this overturns her core case on labour, secrecy and concentrated power, which later events (Superalignment's dissolution, the clawback scandal, chatbot-harm lawsuits) largely bore out. The book won the 2025 National Book Critics Circle Award for nonfiction.
Connect it
- Enterprise AI procurement. Treat vendor capability claims as unverified unless the provider discloses training data or allows independent evaluation; Hao's benchmark-contamination point (Ch. 7) applies directly to model selection.
- Supply-chain due diligence. Banks already screen suppliers for modern slavery; the Fairwork minimums and the crisis playbook give a ready checklist for annotation and moderation vendors behind any AI model a firm buys or fine-tunes.
- Right-size the model. Te Hiku and Stable Diffusion show that a small, task-specific model on owned data can beat a general giant on cost, sovereignty and auditability; enterprise architects should start from the task, not the largest available model.
- Governance lessons for boards. The Blip shows that oversight relying on insiders' courage and verbal briefings fails; boards should demand written reporting, independent channels to staff, protection for sources and clarity on the CEO's outside holdings.
- Climate and ESG reporting. Data-centre water and power use belongs in Scope 2 and 3 disclosure; a firm that cannot say what its AI consumes cannot claim it is net-positive.
About this book
Empire of AI: Dreams and Nightmares in Sam Altman's OpenAI by Karen Hao, 2025. This is a chapter-by-chapter study of the book, with summaries, maps, tables and diagrams; short quotations keep their page references. It is shared free for personal study, is not for sale, and is not endorsed by the author or publisher. Please read and buy the original.
Insight Editions · rajeshrmahapatra.com/insight-editions