Insight Editions · No. 39
Project Maven
The Pentagon's AI Project and the Future of War
Katrina Manson
Katrina Manson traces the Pentagon’s Project Maven from its origins in battlefield surveillance to AI-enabled targeting systems and debates over accountability in military operations.
A study guide to Project Maven by Katrina Manson. Shared free for personal reading; not for sale and not endorsed by the authors or publishers. Please read and buy the original.
Overview
| Core claim | Scope | Key figures | Best for |
|---|---|---|---|
| Project Maven turned Pentagon computer vision from a pathfinder into a kill-chain accelerator: find, fix, finish, then feed the next war. The story is procurement, personality, and moral injury as much as models. | Prologue, Introduction, 26 chapters in four arcs (Find / Fix / Finish / Feedback), Epilogue. History / non-fiction by Katrina Manson (2026). | Drew Cukor; Jack Shanahan; Brian Ward; Alex Karp / Palantir; Joseph O'Callaghan; Trey Whitworth; Chris Donahue; Erik Kurilla; Kirsten Cukor. | Readers of defense AI, procurement, Ukraine targeting, Palantir/DoD history; EA and banking risk leads who need a concrete case of AI entering a high-stakes workflow. |
Every COVER chapter keeps her order of scenes, people, numbers and decisions; adds diagrams; marks contested claims with a closer look; and, only where a concrete public update exists, a brief October 2026 note.
The Book in Brief
- Old wars invent the hunger. Iraq and Afghanistan screen overload, fratricide trauma, and SOCOM impatience create the demand for machines that watch when humans cannot.
- Maven time vs bureaucracy. Cukor's pathfinder culture (label, model, field, ignore "meh") collides with acquisition law, FOIA, Google walkout, and service indifference.
- Google exits; the market absorbs. Moral outrage in Mountain View becomes Microsoft, AWS, Clarifai, Xnor, Scale, and eventually Palantir's interface bet.
- Interface is the product. Hot servers and potato-chip models do not matter until detections live inside the operator's picture (Gaia/Gotham → Maven Smart System).
- Data hell and fiefdoms. Volunteer doodles, The Reboot, and the JAIC data war show pathfinder status turning into a reason to hoard the path.
- Ukraine is the exam. Snow breaks the models; reteach; Task Force Dragon; tens of thousands of points; Taiwan is the intended sequel.
- Custodianship is the epilogue. Whitworth drinks the Kool-Aid; UN bans stall; Replicator/Goalkeeper/Whiplash chase 2027; Kirsten asks who gets haunted.
Figure 0. Four arcs and four cast clusters: pathfinder, vendors, operators, moral ledger.
Chapter Map
| Ch | Title | In one line |
|---|---|---|
| P | Prologue | Opening frame for algorithmic warfare reporting |
| I | Introduction | Stakes, method, cast |
| 1 | Old Wars | Screen overload and the hunger for help |
| 2 | Tilting at Windmills | Early bets and bureaucratic dragons |
| 3 | We Do What We Want | Maven time culture |
| 4 | They Call It Algorithmic Warfare | Naming the thing |
| 5 | The First Mavenites | Founding team |
| 6 | Relaxed About Fury | Team psychology and early fielding |
| 7 | The Colonel and the Math Whiz | Cukor meets the technical engine |
| 8 | Somalia | First theater proof and failure modes |
| 9 | Moral Outrage | Google walkout and the exit |
| 10 | The Algorithms Have No Clue | Ward, Afghanistan, donkey train |
| 11 | Harbinger of Doom | Submarine acoustics and AUKUS |
| 12 | Arms Race | NSCAI, FOIA, allies, China rhetoric |
| 13 | Daddy Karp | Palantir boardroom, Shield AI |
| 14 | Palantir, Palantir, Palantir | Interface Darwin race, Dam Neck |
| 15 | Palantir Splits the Team | MSS vs smart sensor; July 4 email |
| 16 | A Striking Operation | Baghdadi van; Barakat; doubt |
| 17 | Data Hell | Labels, Reboot, JAIC data war |
| 18 | We'll Find It and We'll Strike It | 18th Airborne; Scarlet Dragon; Abbey Gate |
| 19 | Nobody Knows Targeting Better Than Trey | Whitworth blow-up; NGA landing |
| 20 | Kill Chain | IG complaint; Taliban self-joke; JPM |
| 21 | Ukraine Fights Back | Bowling alley; reteach; Dragon |
| 22 | Tens of Thousands of Targets | HIMARS; Taiwan brief; encryptors |
| 23 | We've Drunk the Kool-Aid | Program of record; CENTCOM; weapons debate |
| 24 | Machines Shouldn't Kill People | UNGA; Kabba; Goalkeeper whisper |
| 25 | Trump's Robots | Replicator; rogue GARC; Whiplash |
| 26 | The Winchester House | Kirsten; PTSD; Gaza; Hegseth ROE |
| E | Epilogue | Mafia diaspora; custodians question |
Prologue
Marine colonel Drew Cukor spends an evening rewinding a 2017 documentary about AlphaGo versus Lee Sedol, the eighteen-time Go world champion. The machine's unorthodox Move 37 floors the human. Cukor's lesson is immediate: a human alone would never be enough; a human plus a machine would always be better, with hard consequences for war.
Netflix, notepad, AlphaGo
Cukor watches like an analyst, not a fan. At home with Netflix, notepad and pen beside him, he rewinds the same passages as many as eight times. The film covers DeepMind's AlphaGo against Lee Sedol of South Korea. Go had long been the holy grail of AI challenges: far more complex than chess, with more possible board configurations than atoms in the universe.
Prologue arc: from documentary to battlefield mandate.
Sedol's panic is the point of the evening. The champion faces a move first dismissed as dumb and formulaic. DeepMind's backers later call it "a moment of exquisite algorithmic ingenuity." It hands AI the winning edge in game two. Sedol loses the match four games to one. Cukor watches a human at the peak of his craft discover the humiliating limits of his own mind.
The lesson for war
The takeaway is human-plus-machine, not machine alone. To Cukor, AlphaGo proves that a human alone will never be enough and that a human paired with a machine will always be better. That has profound implications for his field: war. He already believes he knows the future. He needs to bring AI to the battlefield.
Cukor's contrast after AlphaGo.
He is not the only significant viewer. Cukor hears that China's president Xi Jinping watched the same documentary. The prologue plants the race frame that will run through the book: whoever industrialises AI for combat first will reshape power.
Who watched and why it mattered.
From evening viewing to Project Maven motive.
Drew Cukor: US Marine colonel (later JPMorgan AI lead); watches AlphaGo and decides war needs human-plus-machine AI.
Lee Sedol: eighteen-time Go world champion; loses 4-1 to AlphaGo.
AlphaGo / DeepMind: Google subsidiary program whose unorthodox play becomes Cukor's proof point.
Xi Jinping: Chinese president; reportedly watched the same film, signalling a parallel strategic audience.
Introduction
In mid-2024, after nearly a year of pursuit, Katrina Manson finally sits opposite retired Marine colonel Drew Cukor in the emptied café of J.P. Morgan's New York tower, where he now leads AI transformation for Jamie Dimon. Cukor intones "War is terrible" three times, then opens into passions that define the book: the 1942 Dieppe Raid, the US military's failure to match firepower to intelligence, and Project Maven, the Pentagon effort he led to put AI at the heart of how America makes war. The introduction frames Maven as more than drone-footage sorting: from the start it aimed to target people and objects with AI, remake bureaucracy, tether commercial AI to America rather than China, and force minimally tested systems into hot wars so fighters would learn to trust them.
The interview that almost does not happen
Cukor interviews the interviewer. He offers Manson water and takes nothing. He barely lets her write. His hair is tight-cropped, his demeanor stiff, his opinions uncompromising, until he softens into history and Maven. Over months she learns that almost everyone around the project, whether they fete or hate him, treats Cukor as a leading historical figure in a war that has not yet happened.
How Manson finally gets Cukor to talk.
The problem with war, he says, has always been the humans. They are "materially corrupt, inefficient, and they get tired." When they die it affects the campaign. He believes humans do better with machines, and that AI can pierce the fog of war. More immediately he wants to fix bureaucracy that let down fighting forces overseas, bring intelligence into combat operations, prefer software over hardware, and test emerging technology in real wars. A former Joint Special Operations Command (JSOC) member is caustic: "We were really good at killing people, but it didn't get us very far." Cukor wants fewer mistakes that kill civilians, allies and US troops, and a reconception of victory: not destroying the enemy, but defeating it.
What Maven really was
Launched in 2017, Project Maven is sold as computer vision to sort thousands of hours of drone footage across Asia and the Middle East. Manson learns Cukor and his backers always intended more: AI-aided targeting of people and objects. He also wants defense spending to tether a nascent commercial AI market to America rather than let it chase customers in China. In a draft 2017 paper she reviews, he writes that it is the government's role to help the venture capital community monetize its investments.
Introduction's central claim about Maven's dual face.
The colonel's advocacy "monetizes" startups and academic AI researchers. He recruits Big Tech better known for shopping and office software: Amazon Web Services and Microsoft now deliver algorithmic warfare for Maven. Palantir Technologies, celebrated as an S&P 500 force and often called overvalued by short sellers, was on the way out of the Defense Department when it won its first Maven contract; Manson argues it arguably owes its rebirth to Cukor. Scale AI, in which Meta later takes a 49 percent stake, could say something similar. Nvidia was there from the start. Anthropic and OpenAI plunge into defense work as they seek commercial homes for generative AI. Even Google, whose workers protested "the business of war" after discovering Maven involvement in 2018, later embraces national security work. Notably, Apple has said no.
Who Maven pulled into defense AI.
Alex Karp, Palantir's billionaire CEO, later calls Cukor the "founding father of AI targeting."
Twenty years, five years of resistance, Mavenite culture
Cukor thought remaking the military would take twenty years. For the first five he met controversy and resistance as he pushed Congress to fund new firepower to keep America at the apex of global power. He infuriated doubters inside his own team and civilians who fear Terminator-style extinction. Many "Mavenites" saw themselves as renegades inside DoD, carrying tech-startup insouciance into the button-down Pentagon, while also reflecting the overconfidence and anguish of a superpower repeatedly hitting its own limits. The team dynamic was a rollercoaster as Cukor bulldozed toward his dream at personal and professional cost.
Cukor's twenty-year clock as the intro sketches it.
One of his most controversial choices was pushing minimally tested systems into hot wars. He always argued that putting AI on the battlefield before it was ready was the only way to improve it and build trust in a new generation of fighters. He was exacting in every demand. "To Cukor" meant tremendous hours, pursuit, invention and intensity. "Getting Cukor-ed" meant having to pursue the same at his instruction. One person calls him a mastermind who bent bureaucracy and people to his will: admired and feared. "It's a mental thing." Cukor says there is nothing unusual about him. Everyone else Manson speaks to says otherwise: you cannot tell this story without Cukor.
The moral question and the competing theories
The rise of AI warfare poses the biggest moral and practical question: who or what gets to decide to take a human life, and who bears that cost?
Poles of the AI-war debate in the introduction.
- Cheerleaders argue AI and automation will save lives, bring precision that limits civilian and friendly-fire casualties, deter conflict with China, or help win World War III at a pace faster than humans can understand.
- Detractors think AI has already led to civilian deaths, will spread uncontrolled destruction, and could hasten the end of the world.
- Skeptics of the hype think the claims are grandiose and the truth more prosaic: rickety infrastructure, adoption gaps and trust problems.
- Pragmatists argue an incremental mix of humans and machines will forge that trust.
Manson's method is specifics over theory. She wants the people making AI warfare real and the US military members using it. What is inside the black box?
Ten years on: what Maven became
Ten years after Cukor started, AI decision-making systems developed under Maven, and some of the Pentagon's eight hundred other AI projects, are used on the battlefield. Maven Smart System (MSS) develops targets with AI help, is deployed in every branch of the US military and worldwide, incorporates more than 150 data feeds and the work of more than fifty companies. NATO started using a version in spring 2025; by October 2025 ten NATO members were lining up for their own militaries.
Reported target-throughput and platform scale.
Pace of war accelerated. An official at the National Geospatial-Intelligence Agency tells Manson that with computer vision the US went from being able to hit under a hundred targets a day to a thousand. With large language models integrated into Maven, that rose fivefold to five thousand targets a day.
Reach widened. Maven algorithms deploy in submarines and space operations; in subsea sonar systems belonging to America and two closest intelligence allies (UK and Australia) for nuclear deterrence; on autonomous drone boats. AI targeting lives in at least two highly secretive systems, one aerial and one aquatic, that could surveil, select and kill targets entirely on their own, intended for the defense of Taiwan. The US will have to define use cases, guardrails and doctrine carefully if it wants to stick to the Geneva Conventions and avoid shooting civilians and allied forces.
- 2017: Project Maven launches.
- ~800 other Pentagon AI projects alongside Maven.
- MSS: 150+ data feeds, 50+ companies, every US military branch.
- Target throughput: under 100/day → 1,000/day (computer vision) → 5,000/day (with LLMs).
- NATO: version in use spring 2025; 10 members lining up by Oct 2025.
Why Manson wrote it
Three personal strands plus a beat. She became the Financial Times US foreign policy and defense correspondent in 2017, the year Maven started, watching America humbled in Afghanistan and Iraq reach for AI as a shortcut to sustained dominance. The first Trump administration's 2018 national defense strategy predicted commercial technology "will change society and, ultimately, the character of war." After she became a Bloomberg correspondent on emerging tech and national security, chatbots and AI agents accelerated the shift. Under the second Trump administration, DoD reemerges as the "Department of War" devoted to AI and autonomy, under a secretary who wants easier weapons acquisition and freer "rules of engagement."
The reporter's three experiences plus the beat.
1. Jan Bloch, 1899. Since 1999 she has carried the Polish banker's book retitled Is War Now Impossible?, which asked whether industrial-scale lethal weapons would make war obsolete and instead foreshadowed stalemate and horror. Would AI make great war impossible, or prove Bloch wrong again? 2. Africa reporting. A dozen years covering business, investment and politics across African countries taught her the distance between policy and reality; she wants the same test for AI warfare. 3. Afghanistan flight, 2009. British soldiers beside her on a military plane showed phone videos of friends just killed, wanted to leave the military but felt trapped by contracts and unable to survive civilian life. Could AI alleviate the burden of war on those sent to fight?
Maven sits at colliding trends: American insecurity about its place in the world; AI forced into life and war; fraught civil-military relations; Big Tech dominance; China's military and tech ambitions; ubiquitous sensors and commercial software enabling all-encompassing surveillance.
Closing frame
The next ten years are still open. Russia's invasion of Ukraine upturned military expectations. Pentagon strikes against boats in the Caribbean gray the rules of war and underline war at a remove. US commanders say China rehearses a Taiwan takeover. Rival superpowers arm for conflict. Campaigners argue for AI redlines. A new generation of VC-backed Silicon Valley leaders chases defense contracts with braggadocio about Western superiority and AI-enabled killing. National security strategists worry no country can win a war without AI. The UN aim to ban lethal autonomous weapons that select their own targets by 2026 is a lost hope. And yet AI remains a narrow, faulty tool whose limits the US military is still discovering.
Closing strategic frame of the introduction.
Maven as a procurement and platform pattern.
1. Old Wars
Chapter 1 drops the reader into the first months after 9/11: Sergeant Dave Spirk and Captain Drew Cukor ride a blacked-out helicopter toward Kandahar while a door gunner's .50-cal tears into the night. At Cukor's command a bulky computer rides in place of a lance corporal. Loaded with Excel, Word, Google Earth and PowerPoint, it is supposed to find targets and brief commanders. It is nearly useless. The chapter's charge is that America brought unmatched munitions to Afghanistan and still fought blind, because intelligence software designed for office work is where battlefield data goes to die.
Helicopter to Kandahar
Task Force 58 inserts the first conventional forces. Spirk and Cukor begin weeks earlier on a warship in the North Arabian Sea under Brigadier General Jim Mattis, whose call sign "CHAOS" once meant "Colonel Has Another Outstanding Suggestion." The goal is chaos in the south around Kandahar so the Taliban, hammered in the north, cannot regroup. Each man flies about four hundred miles inland over Pakistan's Balochistan to a Helmand airstrip: twice the usual Marine distance from the sea. Mattis is about to send the amphibious wing into a landlocked mountain country, and doctrine be damned.
Spirk and Cukor's path into Afghanistan.
FOB Rhino is a bustard-hunting camp renamed. Army Rangers take a spartan desert strip with no water supply and one dirt runway. Air Force pilots improvise nighttime dirt landings with night-vision goggles. Early days bring wounded US and Afghan personnel, Taliban probes, and one night a lone camel shot for fear it carries explosives. Leaving Rhino, Spirk and Cukor ride the first wave of helicopters to seize Kandahar Airport. Packed under infantry gear, they listen to the 50-cal. The helicopter is a person short: Cukor's computer took the lance corporal's seat. "I carried the damn thing," Spirk remembers.
Microsoft is where data goes to die
The computer is Cukor's "brain" and Spirk's burden. It is meant to help find and assess targets, detect threats, plan missions and brief the commander. Tools include Excel, Word, Google Earth, PowerPoint and disliked in-house military software. "It didn't do anything, it was no good to us," Spirk says. At Rhino, Cukor sets up computer and map as ground intelligence officer for the 26th Marine Expeditionary Unit (Special Operations Capable). Spirk is the intelligence analyst ("which means one could read and write"). Cukor had made Spirk a pet project two years earlier, never treating the enlisted man as lesser, and later urged him to leave the Marines for faster impact as a defense civilian. Spirk would become DoD's first chief data officer.
What rode in place of a Marine.
America had years of intelligence and tools that could not join it. Logs of Taliban sites, caves, terrain, weapons tipoffs; signals from phones or rogue desert lights; phones, computers and paperwork from raids; detainee interrogations that would later scar America's self-image; overlays of attacks. Cukor's stack could not fuse it. Al-Qaeda targets lived in long Excel tables. PowerPoint lines and circles tried to draw people connections. Word held analysis. Google Earth held geography. Briefings were copy-paste slide decks. Nothing translated to war.
The wound Chapter 1 diagnoses.
An Army artillery officer later sums up target lists on quarterly-report software: "We've killed more people on Office than you'd ever imagine." Cukor wanted information more than anything, but what he had was not battle-ready. His brain held only so many connections. Approximating a web for one high-value target still left several thousand al-Qaeda suspects. Missed patterns meant death traps for US forces. "It's just criminal," he tells Manson nearly twenty-five years later.
Early blood and a GPS that resets to self
Eight Americans die in the first weeks after operations begin 7 October 2001. Before withdrawal twenty years later at least 2,465 US dead and 20,149 wounded. Suicides among post-9/11 war veterans: 30,177. Mattis called the insertion "Operation Swift Freedom"; seniors renamed it. The war would be long. Enduring.
Post-9/11 tallies as Manson reports them (suicides exceed combat deaths).
5 December: the Precision Lightweight GPS Receiver kills its own. Special operators and Afghan forces try to install Hamid Karzai near Kandahar. Taliban fire from a cave across a bridge. A US combat air controller changes the battery on his handheld GPS receiver; powering back on resets strike coordinates to his own position. A B-52 drops a 2,000-pound bomb on the US location, killing three Americans and several Afghans, wounding dozens more. Karzai, among the wounded, is evacuated to Camp Rhino. Cukor does not notice: head down, "intense but unflappable," building the next package with primitive tools.
The 5 December friendly-fire sequence.
Kandahar Airport before dawn. Enemy probes the north. Spirk lugs the computer in as conventional forces pour in. The US-built 1960s arches already show war: blown windows, cratered airfield soon filled by Navy Seabees, rose garden turned ablutions block. Troops name tents "Jihad Motel" and draw "Santa Taliban" cartoons through the holidays. On 16 December three Marines are hurt by a land mine inside the commandeered airport. One uses his belt as a femoral tourniquet, later loses his left leg below the knee, and receives the war's first Purple Heart.
High-tech munitions, low-tech information
The Taliban know the ground; America knows how to collect, not fuse. Every service has its own analytic software and target-list processes, often built by rival contractors for hundreds of millions or billions. Cukor is blunt: "They were crap. Just horrible." Too complex for combat, or blue screens of death, or refuse to boot. Internal word processor "Enable" is dubbed "Disable" after crashing past three pages. Once Cukor loses everything on his hard drive and wants to go home: "I just wanted to cry."
Asymmetry of means versus meaning.
Troops go back to paper. Cukor and Spirk fix a paper map on the wall with layers of acetate by hand: more efficient than the computer for threats. Pilots tape physical maps over their knees. Operators cross-check landmarks as clockfaces or phone keypads, exchanging imaginary times under fire. Bases put cutout "pattern wheels" on walls, pizza-sliced by day or month, to mark attack timing. Sometimes the only tools are printouts and highlighter pens. Still they drag the useless server from Rhino to Bagram to Kabul for one reason: they are told to. Cukor finds one reliable use: "The box made the room warm." Blown-out windows cool fast. "I always took that box because it was good as a heater."
Low-tech workarounds that beat the stack.
- 7 Oct 2001: official US operations begin.
- 8 Americans dead in first weeks.
- 2,465 US dead / 20,149 wounded by withdrawal ~20 years later.
- 30,177 suicides among post-9/11 war veterans (exceeds combat deaths).
- 5 Dec: 2,000-lb bomb on own GPS position; 3 Americans + Afghans killed; Karzai wounded.
- 16 Dec: first serious Marine injuries at Kandahar Airport; first Purple Heart of the war.
Before the towers: Cukor's formation
Until 9/11 he thought he had a normal Marine career. Then: "All of a sudden we're in these wars." He grew up in 1980s Los Angeles, skinny and physical, among kids who thought Reagan-era nuclear war might incinerate them. The military was education; becoming "a soldier of the sea" felt like Boy Scouts. By 1993 the Berlin Wall was down, history had "ended," and nuclear war slid toward cartoon threat. First assignment: Twentynine Palms in the Mojave. Fun, hard, absorbing. "For me, it was perfect."
Cukor before the towers.
Marine Corps culture fits him and does not. Smaller and insular versus Army or Navy; expeditionary; flexible; aggressive speed; first in, then hand ground to the Army. He asks for intelligence and is one of only two who make the grade. Marine intel officers rarely rise high: since 2000, only eight Marine intel colonels make major general and four lieutenant general. At the training base he runs up a $30,000 phone bill in 1993 on dial-up for unclassified news, burning the unit's intelligence budget for the year. Naming exercise areas, he turns to wife Kirsten's literary favorites: Austen, Byron.
Kirsten, Don Quixote, the thesis
They meet winter 1993 through her housemate. Approaching the living room she hears him appraise her books and wonder what she sees in Joyce's Ulysses. Born without respect for authority, she enters ready to argue. He is on one knee at the shelves, looks up, smiles. "That's her," he tells himself. When she learns he works Marine intelligence she starts "But isn't that . . ." and he finishes "an oxymoron," delighting her. Contradictions pile up: a Marine with a low opinion of the Marines; likes Phantom of the Opera and Les Mis; wired for war yet children trust him. Favorite book: Don Quixote (1605), the tragicomic hero on a doomed quest to right wrongs in a world that does not match his ideal. "He's such a Don Quixote-ish kind of person." She gets paralyzed in theory; he is all action, "willing enough to be wrong that he'll actually try to solve a problem." Her bedroom poster of The Philadelphia Story lists director George Cukor; she jokes privately about sharing the surname. They marry within four months.
Personal origins that name the quest.
Three years and one child later, at the Naval Postgraduate School in Monterey, he writes a master's thesis on "urgent reform" of the Marine Corps intelligence enterprise. He is obsessed with the failure of Marine ground intelligence for twenty-first-century threats. He has a solution that relies on a way of seeing that does not yet exist. The chapter closes on the quest that Chapter 2 will name: tilting at windmills.
Close of Old Wars into the thesis arc.
Drew Cukor: Marine intel captain; hauls a useless Office stack into Afghanistan; later Maven lead.
Dave Spirk: enlisted intel analyst; carries the computer; later DoD's first chief data officer.
Jim Mattis: Brig. Gen. leading Task Force 58; call sign CHAOS.
Hamid Karzai: wounded in the 5 Dec friendly-fire bomb; future Afghan leader.
Kirsten Cukor: writer-thinker wife; literary base names; reads him as Don Quixote.
Collection-without-fusion loop.
2. Tilting at Windmills
Captain Cukor's 1997 Naval Postgraduate School thesis brands the Marine Corps "centralized, hierarchical and slow" and barely changed since World War II. Gulf War victory, he argues, came from firepower while intelligence failed: Americans fought blind against Iraqi surprise night assaults. He wants wholesale change, "white dots" fused into an accurate battlespace picture, and industry-academia partnership. The chapter traces that vision through Desert Storm friendly fire, the ignored Van Riper plan, an early Taiwan missile warning, Pentagon budget pathology, Iraq mediocrity, Haditha, Project Legacy, and the November 2010 death of First Lieutenant Robert Michael Kelly, son of three-star John Kelly, which leaves Lieutenant Colonel Cukor "unbelievably pissed off."
The 1997 thesis
Gulf War intel failed even in victory. Against the Iraqi Army, Americans were attacked on the ground with no idea where the enemy was. Victory came by bludgeoning with overwhelming firepower, not competence that would survive a tougher foe. "Military victory would need to come through military competence, not from sheer superiority of men and material as it had in previous conflicts," wrote the thirty-year-old captain.
1997 thesis contrast.
Marines are built for wars of maneuver, hitting weak points with ferocious speed from multiple directions so the enemy gives up, avoiding attrition but raising friendly-fire risk. In Desert Storm, Cukor says they fought outdated "movement to contact": keep moving until you slam into the enemy, then fight. It risks mass casualties Americans would no longer stomach. Deployed like a giant army, they sometimes fired on their own. More than half of Marines killed in the Gulf War (14 of 24) died by friendly fire, mostly air-dropped munitions; another 92 were wounded.
From maneuver ideal to Gulf War practice.
Information flowed the wrong way. New intelligence went to headquarters, not tactical commanders. Troops had small-scale maps with little detail; sometimes they rifled truck tubes for the right curled map. Congress had directed fixes in 1992; the Van Riper Plan (Lt. Gen. Paul Van Riper) still unimplemented by 1997. Even if successful, Cukor judged it "incremental, peripheral," mostly training and promotion access.
White dots and Sun Tzu
Sun Tzu's question: Where is the enemy? Cukor watched the Corps fail to answer it. He wanted something "vastly different": turn data into knowledge and action; know precise locations of allied and enemy forces on a meaningful map; build "a collaborative partnership" with industry and academia combining human and machine. In 1997 he wrote: "White dots fused with an understanding of the enemy and validated by other collection assets can contribute to a highly accurate battlespace picture." Making those white dots real became the drumbeat of his life.
The drumbeat that becomes Maven.
Urgency: Taiwan. He warned surprise would be central to any Chinese takeover: missiles easy to conceal, hard to track, "perfect weapons of surprise." Taiwanese forces could be overwhelmed; "Within 24 hours, Taiwan could be in PLA hands." "Caught by surprise, America would not have time to put adequate forces in theater." He denies prophecy: "These were things that were just common analysis... it just took forever to pivot the machine." 1990s US-China tensions spiked over Taiwan; April 2001 midair collision over the South China Sea and Hainan spy-plane dismantling. Things would have worsened, he says; 9/11 got in the way.
Cukor's 1997 Taiwan warning.
Budgets, colors, sensor-to-shooter
Pivoting the machine meant the Pentagon. Three budgets always run: spending, in Congress, next. Enormous yet never enough. Annual five-year resubmissions; Congress allocates in five "colors": procurement, research, operation, personnel, construction, each with a use-by date. Services equip their own forces; some budget classified; military intelligence on a different stream; an entire university for defense acquisitions; jobs and contracts as constituency prizes. The Pentagon had never passed an audit.
Budget pathology feeding the omniscience dream.
IT fell between stools. Little thought to buying bytes versus bullets. Software update pace outran the budget cycle; annual licenses rankled. Cukor wanted budgets reoriented for software that blends intelligence into operations. Emerging holy grail: a single digital grid combining intelligence, communications and operations; ubiquitous sensors; one comprehensive real-time screen; aspirational single pane of glass; one day a brain to fight the war. Omniscience, omnipresence, omnipotence. In 1997 he complained: "The human element . . . slows down processing time." Sensors could connect directly to weapons: sensor-to-shooter warfare reducing the human element to "minimal."
9/11 birth day, then Iraq
Start of 2001: Captain Cukor at Camp Lejeune as intel officer of the 26th MEU. Kirsten pregnant again; he due to float the Mediterranean two weeks after birth. On 11 September 2001 their third child is born and his mission changes instantly.
Iraq as continuous mediocrity.
Anbar, Haditha, Project Legacy
April 2005, Anbar, Major Cukor as fusion officer for II MEF. 456 Marines already dead in Iraq. Unit keeps missing essential, sometimes fatal, information; dozens of daily skirmishes; 189 Marines die during his tour. That November: Haditha. Marines allegedly shoot dead twenty-four unarmed Iraqi civilians after a roadside bomb kills a Marine. A three-year-old girl photographed lifeless with "12" in red pen on her cheek beside her dead five-year-old sister. Cukor is on watch; sees reports. He says Marines did not bury it and studied what happened. Manson notes: this is disputed. "Then it's 2006 and 7 and 8 and 9, and it's just a mess."
Anbar tour into Legacy.
Someone forgot intelligence. US trained Iraqi security in logistics, artillery, infantry, aviation, special forces. "Someone forgot to involve intelligence." He builds Project Legacy on civilian policing methods in Iraq and Afghanistan. Matt McKnight, young Marine officer in Anbar: "Drew is one of the most innovative thinkers the Marine Corps has had in twenty years." Legacy later draws public controversy over methods and funding, investigations into contractor behavior and overcharging. By then Cukor is at the Pentagon. "I probably lost ten years of my life in stress and pain."
Sangin and Robert Kelly
9 November 2010. Marines took Sangin in Helmand two months earlier; southern provinces nicknamed "Marineistan." British predecessors took heavy losses: a third of Britain's war dead. Manson visited Helmand the year before: tracing a FOB perimeter without an explosion counted as success. More than a thousand US combatants already dead; two or three Marines dying weekly. First Lieutenant Robert Michael Kelly, twenty-nine, platoon commander, son of three-star John Kelly, killed by a makeshift landmine on foot patrol. John Kelly, highest-ranking American officer to lose a child in Afghanistan, asks prayers for the rest of the platoon. By tour end Robert is one of twenty-five dead from the battalion; more than a hundred limbs amputated.
The death that ends the chapter.
Cukor knew John Kelly from Marine Reserves command and Iraq Legacy work. "I have all of this anger in me." Tired of not connecting dots; of units relearning the enemy every six months; of no one remembering who lived where or who took which road; of unnecessary death. Lieutenant Colonel Drew Cukor was unbelievably pissed off.
- Gulf War Marine deaths by friendly fire: 14 of 24; 92 wounded.
- Iraq: ~4,600 US troops killed; ~186,000 Iraqi civilians (as reported).
- First two months Iraq invasion: 56 Marines killed.
- By Apr 2005: 456 Marines dead in Iraq; 189 during Cukor's Anbar tour.
- Haditha allegation: 24 unarmed civilians.
- Sangin battalion: 25 dead, 100+ amputations.
Arc from thesis to Palantir hunger.
Paul Van Riper: Lt. Gen.; namesake of unimplemented intel reform plan.
Robert Michael Kelly / John Kelly: son killed in Sangin Nov 2010; father three-star Marine.
Matt McKnight: Anbar contemporary; praises Cukor's innovation.
Kirsten Cukor: third child born 11 Sep 2001 as mission flips.
3. We Do What We Want
After Robert Kelly's IED death, Lieutenant Colonel Cukor, newly analysis and futures chief at Marine Corps Headquarters Intelligence, decides the answer to cheap bombs that caused two-thirds of US casualties is not only MRAP armor but information. His chosen platform is Palantir. Chapter 3 follows Joe Larson's introduction of Palantir to Cukor, the company's CIA/In-Q-Tel roots, aggressive field sales, the i2 espionage lawsuit, Cukor's ghostwritten JUON for Michael Flynn, the $10 million CTTSO grab, servers flown to Camp Leatherneck, go-live 21 March 2011, Army fury, and Cukor's giddy Marine credo: "This is the Marine Corps; we do what we want."
IEDs, MRAPs, and Leatherneck
America spent about $8 trillion on the global war on terror and still lost to IEDs for two-thirds of American casualties. Thousands of heavy MRAPs arrived after painful Pentagon and Congress fights, but they insulated Marines from locals, kept them on roads, were never enough, and did not save anyone on foot. Cukor wanted information. Plan stated plainly: "My plan is nothing short of deploying Palantir to the entire Camp Leatherneck."
Why information beats armor alone.
Larson, the right arm
Faith dates to spring 2009. Junior Marine intel officer Joe Larson, finishing a Stanford law degree on reserve duty, had met Cukor at Camp Fallujah in 2008 and found him "brilliant." Cukor later calls Larson "my right arm." Larson emails that analysts spent hours doing "Ctrl-F" in Windows folders for network analysis. He steers Cukor toward Stanford engineers at Palantir for Afghanistan. Three years later Larson and wife lawyer Ann Marie Rosas (later general counsel at Anduril) both work for Palantir: an insurgent twist on the revolving door.
Larson as bridge.
Cukor's January 2008 paper with co-authors tests 183 public al-Qaeda texts: human reading versus network text analysis. Software misses what humans catch and vice versa, but machines need far less time and fewer people. Conclusion: "an integrated human-centered automation support approach." Co-author Kathleen Carley (Carnegie Mellon), already working with CIA, builds tools to find network centers and predict reconstitution after arrest or killing. For some this foreshadows death lists and domestic surveillance; for others, safety.
Human-centered automation conclusion.
Palantir finds its conduit
Jason Payne (Palantir) lunches with Larson; internal memo: "LTC Cukor is trying to shake up Marine intel... conduit to get us to the right folks." Karp in a 2009 TV interview: "We have no salespeople," yet outreach is aggressive to operational users with skin in the game, not procurement bureaucrats.
Origin story. Founded 2003 after Peter Thiel wonders if PayPal fraud tools apply to jumbled government data; named for Tolkien seeing-stones; Valley vibe (bean bags, free meals) for spy clients. Early investors balk until tip: see the CIA. 2005: In-Q-Tel invests $2 million. Susan Gordon (later principal deputy DNI) calls Palantir one of In-Q-Tel's two best bets (other: company that became Google Earth). Engineer Akash "Aki" Jain, one of first five, spends years on product and facility clearances. Selling point: live with users, shape bespoke platforms, mold captive audiences. Jain's metaphor: "two thousand Lego pieces in six bags with no manual: How do we create the Death Star?" then corrects to "Falcon 9."
Company roots Cukor buys into.
Karp's pitch: get ahead of the terrorist; see patterns they do not realize they give off. "Compare us to what you have." Marines have almost nothing. Larson is "blown away," depressed at teaching Analyst's Notebook (i2, later IBM): "He_hates_ANB." Users cite confirmation bias and weak metadata. Palantir tags data and every link, with audit trails of which analyst linked what, from which report, across agencies; can detect circular reporting. McKnight calls it large-scale indexing tying disparate databases in one secure environment.
Misspellings, black sites, DCGS-A
Cukor visits Palo Alto next year: "I knew it was special." Rules can link Dru Cooker to Drew Cukor. Analysts stitch phone transcripts, interrogations, captured docs across Mosul, Kabul, Riyadh. Controversy: systems fed information from black sites and torture-tainted confessions; conflating Mohamed/Mohammad/Muhammad; "flying while Muslim"; fear that war tools migrate to domestic FBI/homeland use before Snowden 2013.
Power and hazards of the platform.
Army detractors later say DCGS-A does the hard collection while Palantir is a pretty display app; whole iPhone versus one app (2013 Politico). Cukor: DCGS-A horrible, sometimes would not open, could not hand data to incoming teams.
i2 lawsuit (2010): alleges Shyam Sankar (joined March 2006) illegally accessed i2 via fake company Xoom, parents' Florida registration, copied icons and export techniques. Palantir claims i2 had "unclean hands." Settle 2011, reportedly Palantir pays i2 $10 million. Cukor, who read the suit: early sales practices "absolutely pernicious"; Pentagon unused to such startup aggression. Still desperate for the tech. Obstacle: not accredited, not a program of record.
Ghost JUON and the $10 million grab
Stewart's obsolete-request parable (told to Peter Dixon): as a young officer Stewart requested a project; years later as one-star it lands on his desk obsolete; he cancels it. Cukor will not wait. Obama's 33,000-troop surge is on; heavy casualties. More Pentagon employees work acquisition than the entire Marine Corps. Marine intel is a "basket case."
Lt. Gen. Michael Flynn, top US military intel officer in Afghanistan, early 2010 excoriates intel as "marginally relevant," secret over useful: "The real intelligence hero is Sherlock Holmes, not James Bond." July: Flynn files a JUON (Joint Urgent Operational Needs). Cukor: "most powerful document a commander can launch"; leapfrogs hierarchy. Specs describe Palantir so closely Army officials later say "clearly ghost-written by a Palantir engineer." Manson learns it was ghostwritten by Cukor and team. Dixon (Project Navigator on future threats like drone IEDs) drafts requirements with Cukor after visiting Google and Palantir. "We jammed that thing through."
JUON as crowbar.
Army rejects: already spending $550 million/year on DCGS-A ($1.6 billion network). Cukor goes Marine-only; half-wonders about firing or jail. Dixon: fueled by saving lives; none more relentless than Cukor.
Money hunt. FY2010 DoD budget $691 billion, every dollar pre-allocated. Cukor hears of a plus-up partly via Palantir lobbying to CTTSO (Combating Terrorism Technical Support Office): $36 million extra on $82 million baseline. Army turns it down; Cukor grabs it. Final budget week: $10 million reprogrammed "in view of the urgent needs assessment." Sole-source expedited funding; Army Research Lab statement of work becomes a solicitation only Palantir can meet. Three legs of the stool: "Money, contract, statements of work."
Three legs: money, contract, SOW.
Leatherneck go-live
Stewart would have said no if he knew; Kelly's son's death motivates action. Two Palantir engineers with clearances haul $20 million servers on a C-17 to Camp Leatherneck end of February 2011. Palantir sends a $100,000 generator. Still stuck until Maj. Gen. John Toolan (2nd Marine Division), who had already taken a Palantir meeting, orders use. 21 March 2011: live. Within an hour 15 users at Delaram; eight weeks: 250 users across thirty subgroups. Aircraft Wing plans routes and LZ threats; biometrics builds forensic IED networks; even weather cell correlates moon phases with bomb-maker habits. Army soldiers fly in to see; Brits ask what the hell Marines are using.
Fielding sequence.
Stewart takes heat. An Army colonel storms Cukor's empty office: fall back in line with program of record. Dixon shocked: DCGS-A not due for three years; saving lives versus defending a defense company. Army hatred of Palantir grows (cost, arrogance, data trap). Stewart "wasn't super mad"; liked Marines out front. Cukor, giddy: "This is the Marine Corps; we do what we want." "The politics around this was fundamentally the Marines had the balls."
Pulse, Mattis letter, bigger hunger
Operational data pulses in. Raids reveal one place hit twenty times. "OMG we look like idiots." August 2011: Palantir worker shoves a letter into Jim Mattis's hand at Marines' Memorial Club in San Francisco. Later: explosive 2013 Congress session; 2016 Palantir lawsuit forcing Army sales access. Cukor: "Palantir gets a black eye but wins and wins ugly... my name isn't on any of this stuff." Still small wins. Analytic problem unsolved: Where is the enemy? What is going on? What was the point? Mantra: "We always see a past that was predictable but not predicted." Draft geostrategic analysis: "A nation not prepared to wage existential war because it is rare faces catastrophe." 2014 essay: modernization of intel analysis "will largely determine the success of US forces." New tool coming. "It's now the age of AI. And Manson is asking where the fuck is the department's AI?"
From small win to AI question.
- War on terror cost cited: ~$8 trillion; IEDs: 2/3 of US casualties.
- In-Q-Tel into Palantir (2005): $2 million.
- DoD budget 2010: $691 billion; CTTSO plus-up $36M on $82M; Cukor reprograms $10M.
- DCGS-A: $550M/year, $1.6B network claim.
- Servers $20M; generator $100k; go-live 21 Mar 2011; 15 users in 1 hour → 250 in 8 weeks.
- i2 settlement (reported): $10 million.
Further beats from the record
3 WE DO WHAT WE WANT “My plan is nothing short of deploying Palantir to the entire Camp Leatherneck.”AMERICA BROUGHT $8 TRILLION and dominant firepower to the two-decade fight it called the “global war on terror.” But the US military toiled, largely unsuccessfully, to find an answer to the cheap improvised explosive devices (IEDs) that killed and maimed its troops. IEDs accounted for two-thirds of all American casualties. The Marines used thousands of heavy and expensive mine-resistant ambush protected (MRAP) vehicles to protect their troops, the result of a painful, delayed, and ongoing fight at the Pentagon and in Congress over production, design, and money. They insulated the Marines from the local population and kept them on roads. What Marines really needed, Lieutenant Colonel Cukor figured, was information. When First Lieutenant Robert Michael Kelly was killed by an IED in November 2010, Cukor was a handful of months into his new job at the Pentagon. He had a useful title, analysis and futures chief at the Intelligence Department of Marine Corps Headquarters, a good boss, and the task he wanted: to fix intelligence. He thought he had the perfect solution: a data management platform made by a company named Palantir Technologies Inc., a Silicon Valley startup founded a few years earlier that was already working extensively with the US intelligence community. Cukor wanted to field the software to the Marine Corps base in Helmand to help pinpoint the next roadside bomb: “My plan is nothing short of deploying Palantir to the entire Camp Leatherneck,” he told Manson. Cukor’s faith in Palantir went back to spring 2009, when a junior Marine Corps officer named Joe Larson wrote to the Silicon Valley startup to introduce the company to Cukor’s work. Captain Larson, an intelligence officer completing his law degree at Stanford during reserve duty, interrupted his studies to deploy on active duty to Iraq in 2008. He was bowled over by Cukor when he met him at the Camp Fallujah intelligence center: he thought him “brilliant.” Cukor would later come to regard Larson as “the right arm.” And now Larson was bowled over by Palantir too. His job had been to track down high-value al-Qaeda targets. But he, like Cukor, despaired of “the failure” of defense intelligence systems to do modern, complex data analytic work. “My analysts spent many hours in Iraq doing a ‘Ctrl-F’ search in a windows folder to find relevant documents for building their network analysis, and that just isn’t the way to do business,” Larson emailed Cukor. He had previously discussed Palantir with him at a Marine intelligence conference. Marines “may need to look” at this new company of Stanford computer engineers to tackle the Afghanistan intelligence problem, he continued. And so began Larson’s key role in bringing Palantir to Cukor and Cukor to Palantir. Larson and his wife, lawyer Ann Marie Rosas (who is now general counsel at Anduril Industries), would three years later both go on to work for Palantir. The upstart startup would develop its own insurgent take on the long-standing “revolving door” that defines America’s defense-industrial complex. Larson also sent Palantir a paper penned by Cukor. In it, Cukor explored the best way to create and analyze a terrorist network database for connections and clues. Raking over America’s failure to thwart the 9/11 bombers’ plot, the 9/11 Commission Report had highlighted the need to do exactly this. Inside disconnected US systems had been clue-fragments of multiple links among a few men training in American and European flight schools since 2000 and a Saudi construction millionaire who long dreamed of felling American skyscrapers the same way Lebanon’s towers once burned. In his January 2008 paper, Cukor and his co-authors tested out how to digest and learn from 183 public texts about al-Qaeda. One approach had human intelligence analysts simply read them. The other tried an emerging branch of software named network text analysis, which sought to extract meaning by finding linked words among masses of text. Cukor’s study didn’t come down on the side of the machine alone, the software missed things humans picked up. But humans also missed things the machine picked up. And the machine took far less time and fewer people. The authors argued for both: “an integrated human-centered automation support approach.” The network text analysis software they assessed was developed under Cukor’s co-author Kathleen Carley, a computational social scientist at Carnegie Mellon University who was already working with the CIA to automate intelligence analysis. In combination with other software she developed, her network text analysis program would hunt individuals at the center of any web and predict the impact that arresting, or killing, that person might have. She modeled how soon the network might reconstitute itself without them, and through whom. For some, amalgamating and analyzing data en masse like this presaged a terrifying future. Automation machines could propose unwitting individuals for death in war and help build an overpowering domestic surveillance state. For others it would be a smart way to keep America safe.
Bypass loop that foreshadows Maven.
Joe Larson / Ann Marie Rosas: Marines to Palantir (she later Anduril GC).
Jason Payne, Shyam Sankar, Aki Jain, Alex Karp, Peter Thiel: Palantir side.
Michael Flynn: JUON requester.
Peter Dixon: Cukor captain; drafts requirements.
Vince Stewart / John Toolan: Marine generals; heat and rescue.
Kathleen Carley: CMU network-text co-author.
4. They Call It Algorithmic Warfare
Mid-2016, back at the Pentagon under three-star Jack Shanahan, Drew Cukor crashes the deputy's Breakfast Club and pitches Bob Work on AI for drowning drone feeds. Work wants autonomy for China deterrence; Cukor wants a real program, not a study. By 26 April 2017 the memo creates the Algorithmic Warfare Cross-Functional Team. Officially intelligence. Unofficially, as Grant, Work, and Cukor later admit: Find, Fix, Finish targeting compression.
Breakfast Club and China's carrier killers
Cukor rejoins the Office of the Under Secretary of Defense for Intelligence in mid-2016 and still finds intel treated like a flower arrangement. Combat people ask for a report the way they ask for weather, then tell intel to sit down and shut up. He wants into the Breakfast Club, a ~20-person group meeting every two weeks since late 2014 in the conference room next to the deputy secretary. No breakfast or coffee is served. The name echoes a Cold War Pentagon working group that planned how to beat the Soviet Navy. This reincarnation, forty years later, hunts how to beat China.
Club that green-lights Maven.
Greg Grant, who convenes it, calls the club an attempt to bring back "the magic of getting smart people around the table" on specific operational challenges: almost the deputy's brain trust. Cukor self-invites. Boss Lt. Gen. Jack Shanahan (more than 2,800 hours in reconnaissance aircraft, bombers, and fighters; now a defense-intelligence director) green-lights the seat. Meetings scout tech, vet use cases, and surface threats aimed at US weak points: hypersonics, space weapons, the stealthy J-20, and anti-ship ballistic missiles nicknamed carrier killers. Ten active US carriers, each about $13 billion, have projected power into the South China Sea and Taiwan Strait for decades; they now look like sitting ducks.
Location matters. Ordinary planning meetings languish in anonymous offices; Breakfast Club sits just down the hall from the secretary of defense, feeds quarterly discussions among top defense and intelligence leaders, and lets Cukor "chin-stroke" at scale. Deputy Defense Secretary Robert Work, in the role since 2014, often pokes his head in or joins the table. The secretary is the face abroad; the deputy runs the building and future plans. Work constantly pushes the group toward artificial intelligence and autonomy.
Work's leapfrog doctrine
Work's core claim: "The US has never tried to match any adversary tank for tank, airplane for airplane, ship for ship. It's always looked for military technical superiority." He fears China closed the technological gap while America fought Iraq and Afghanistan. China spent fifteen years learning US weak points. Deterrence rests on global "military dominance." Size alone favors China and Russia; America needs something else.
Leapfrog doctrine timeline.
Work is a former Marine Corps artillery officer who grew up in the Cold War and keeps his eyes on World War III. He retires as colonel in 2001, never fights the terrorism wars, then directs future-warfare studies at a Washington think tank that publishes papers including "The Promise of Directed-Energy Weapons" and "Rethinking Armageddon." His model of US advantage is leapfrogging:
- Nukes first. America invents them and alone detonates them in war. Advanced nukes answer Soviet mass mobilization (a million men on railroad security alone). Technology substitutes for numbers; a bomb at arm's length is cheaper than a large conventional force. US nuclear warheads peak at 31,255 in 1967.
- Russia catches up. Tactical warheads and pugnacious use theories follow. By 1978 Russia produces more warheads than the US and shows no intention of slowing.
- 1980s leapfrog. Long-range guided missiles dropped by stealth planes, GPS satellites run by the US Air Force, overwhelming force on target.
He wants the US to keep aggressive first-mover status for each new era: a "never-ending" stream of next-generation weapons. Critics hear warmongering; Work says commanders who refuse put their people at risk. He likens slow US planning, development, and fielding to serving badly defrosted vegetables trucked across the country: sometimes too old, does not taste good. He wants new ways to make it harder for Beijing to catch up. His answer is autonomy. AI is means to that end.
Autonomy at rest and on the move
Autonomy, for Work, means two things. First: better, more relevant decisions faster via AI and human-machine collaboration (autonomous mission control and targeting, situational awareness, course-of-action development, cognitive electronic warfare, predictive maintenance, logistics). He calls this autonomy at rest. Second: humans off weapons platforms altogether: swimming, flying, driving unmanned systems. Autonomy on the move. Public remarks in 2015: "Manson is telling you right now, ten years from now if the first person through a breach isn't a fricking robot, shame on us. We can do this."
Work's two autonomies.
Under Work the department wrings its hands over autonomy arriving everywhere except the Pentagon. He thinks Air Force, Army, and Navy drag their feet the way they once resisted GPS, which underpinned guided munitions, until Deputy Defense Secretary William Perry overruled them. Work wants to play Perry. A June 2016 Defense Department study (same month Cukor returns), with experts from Amazon, Google, and IBM among authors, warns the US military is far behind the commercial sector: IBM has Watson, Google pursues self-driving cars, and despite funding AI research for sixty years the Pentagon has little. Recent leaps call for "immediate action" on military autonomy: fewer troops in harm's way, higher quality and speed of decisions, entirely new missions. "Advances in AI are making it possible to cede to machines many tasks long regarded as impossible for machines to perform."
Services do not budge. Money is one reason: huge 2016 budget, dearth of cash for new starts. Many balk at handing life-and-death decisions to machines; the report itself calls autonomy in command and control "one of the most contentious applications." Trust is harder than technology. Many top commanders have never heard of machine learning. Practical ingredients missing: cloud, compute, machine-readable data. Almost nothing in DoD lives in the cloud. Even though the internet is a Pentagon creation and DoD spends $38 billion a year on IT, cloud meaning for war arrives late. Every weapon, satellite, and computer produces data, but it sits bundled on old slide decks and hard drives as good as discarded. No one brings raw data together to share, sort, or understand.
Why services resist autonomy.
The advisory group's final report argues adoption will come not by diktat but by experiments, pilot programs, and prototypes that can take hold. A 2017 draft document Manson reads says prior niche projects had "little chance of meaningful transition to the warfighter." Work wants confidence and case studies so fighters themselves change their minds. And he wants AI for combat: "I do not want it to just be about intelligence. I want to have some type of direct warfighting applications."
Cukor's pitch: a thing, not a study
Grant hits it off with Cukor from arrival: "super interesting," "simpatico." One day Cukor shows up with an idea. He does not want a two-year study, a prototype, or a test. He wants "a thing." Something real. In 2014 he wrote "Operate to Know," focused on "actively" hunting the enemy. Now he announces a program that will demonstrate AI's importance and why the services should partner with leading-edge AI companies that do not usually work in defense. Grant's reaction: Bob is going to be so excited.
Pitch in Work's office, Grant alongside: use AI to help analyze drone data. America's drone wars took off on the belief remote combat could save troops. From 54 battlefield drone sorties in 2001, numbers reach close to 8,000 in 2010. By 2016 the US spends $3 billion flying drones over Afghanistan, Iraq, North Africa, East Africa. More than thirty countries have developed armed drone programs. US drones watch lives and take them: mostly al-Qaeda, Taliban, and ISIS; sometimes dozens of civilians at weddings, hospitals, markets, roads, farms. Sometimes America apologizes; sometimes it covers up; sometimes it does not even check.
Problem Cukor sells.
Drones produce pictures and video the US cannot process. ISR data recorded in 2016 is so overwhelming it equals about eighty full years of video. Senior acquisition official Frank Kendall is appalled during a Baghdad visit that no one is even looking at it. Drone numbers keep growing: a war factory without a workforce. Cukor's pitch: AI can help with processing, exploiting, and disseminating (PED) drone feeds and take load off humans tackling only a fraction. In Mosul, five airmen stare at intelligence feeds twenty-four hours a day. Laborious process wastes brainpower, morale, and time. "Why can't we have algorithms do this work?"
People in Cukor's orbit later say he could sell salt to slugs. Work is "super psyched." He dispatches Cukor to dummy up a demo.
Valley tours, Clarifai, Spirk, IDenTV
Cukor visits Silicon Valley and emerging tech for best AI talent, not Boeing, Booz Allen, Lockheed Martin, or Raytheon primes. Much thinking behind tomorrow's AI wars is inspired by autonomous car companies: Tesla, Waymo, Uber. At Uber's Pittsburgh plant a dozen cameras process everything around a car and help it (mostly) avoid cars, people, obstacles. AI plasters billboards. He reads a mountain of AI papers. Startup Clarifai offers a public portal: run images through AI on the open internet; it coughs up contextual words. The team runs a picture of Bob Work. Portal replies: "Military." "Leader."
Now the Pentagon needs algorithms of its own. Cukor turns to Dave Spirk, the Marine who once helped carry his computer around Afghan airfields. Spirk has left the Marines, left a senior intel job at Special Operations Command, and swerved into investing in Tampa coffee roaster Buddy Brew (named after the owners' dog). Part-time he still advises the Intelligence Systems Support Office (Air Force). Cukor shows up in Tampa: "Hey Dave, I need you to go full time."
March 2017: Spirk goes to San Francisco, pays IDenTV (founded four years earlier by Mohammad and Amro Shihadah) $120,000 to build a computer-vision model on unclassified drone video, and ships it back. AI puts a white dot on screen every time a specific object shows up. Makeshift demo beats an immobile PowerPoint list and elicits wows. Work's reaction is immediate: go do this. Now they need money.
Demo path to green light.
Money, Maven name, April memo
Grant casts around for budget and finds an existing Congress-funded project. Will Roper, a senior defense expert, runs a program initially so secret he could not tell his wife. He has been reading public Microsoft machine-learning research and struck a project using computer-vision providers at defense contractors to pick objects on still satellite imagery. "The Pentagon had tried automatic target recognition programs many times, and none of them had worked very well," Roper tells Manson. With masses of data, more compute, and research advances, maybe AI could understand more about the battlefield than human operators and act as an expert, a maven, to guide the human. At Work's request Roper allocates his hard-won Congress-approved money and the name of his effort to Cukor's new project.
Cukor, with light glee: "I ended up with all of Will Roper's money, which he was always mad about." (Roper has always given loud public support to Project Maven.) Cukor briefs Congress and rummages an obscure fuel-price reserve pot. "Bam; I got money." Eventually Project Maven becomes DoD's largest AI investment to date, starting with $40.8 million for the first year (document Manson reviewed), with much more to come.
26 April 2017: Work signs the memo outlining scope. General Shanahan is director. Cukor gets the most consequential job of his career: chief of Project Maven. Memo text: apply AI, big data, and deep learning to drone feed footage focused on the fight against ISIS; consolidate the department's algorithm-based technology initiatives related to defense intelligence; eventually address unspecified "other defense intelligence mission areas." Aim: "actionable intelligence" and insight at speed.
Money and charter.
Catchy name may be poached from Roper, but the dry bureaucratic title thrills Cukor most: The Algorithmic Warfare Cross-Functional Team. Not for ease of saying (some soon translate AWCFT as "aw shit"), but for what it implies: AI is going to war, and the whole Department of Defense will send it there.
Skynet fear, targeting denial, targeting confession
Popular conception of AI at war is Skynet: US military and a defense contractor create it; an Air Force three-star activates it; it becomes self-aware, takes nuclear controls, and an hour later kills three billion. "A war between man and machines." Some US military officials are as scared as moviegoers; others email Terminator clips around in approval.
The Maven memo is careful to delineate scope. Joe Larson, the Marine reservist and lawyer Cukor calls back, tells Manson Cukor would have "winced" if anyone described Maven as a targeting program. It simply was not that, Larson insists: intelligence systems only, no connection to any munition on a weapons platform, strictly research and development.
But the official name appeals to Cukor and Work precisely because it makes ambitions clear: algorithmic warfare. Asked if targeting and offensive weapons strikes were intended, Grant answers: "Yah, of course. It's not like we're doing it for kicks. The goal of the intel is to take out high-value targets." The ambition to do "targeting" appears in a 2017 draft Maven memo Manson sees. Years later Cukor makes no bones: he always had operational targeting in mind. He always wanted intelligence into the business of operations; Maven was the way. Efforts went to the essence of targeting: AI would help select and prioritize targets and match the appropriate response. Ambition encapsulated in one idea on a complex recipe: a white dot on a screen preloaded with an accurate coordinate. Shanahan makes broader intent public: "The Department of Defense should never buy another weapons system for the rest of its natural life without artificial intelligence baked into it." Another early Mavenite describes their vision: "We're gonna fucking kill people all the time."
Packaging vs intent.
Work has a vision too: "I really do believe it will lead to a revolution at war," combined with autonomous drones on land, sea, and air. Looking back, he tells Manson sorrowfully that if he could rewrite the memo to make clear the focus on AI-enabled targeting and weaponry, he would. "I have taken a lot of heat." Narrow focus on the intelligence side, rather than explicit combat-operations AI, may have backfired. "We were trying to make sure that we started on the right side of Congress. But it's totally fair to say that it probably caused confusion inside the department." Manson's judgment: from the outset Maven was always about precisely what would most scare detractors: a revolutionary targeting application with AI to compress the kill chain from seeing a target to hitting it: Find, Fix, Finish.
Knife fight and the bureaucratic ninja
Internal friction follows. "Anything in DoD is a competition for money," Grant says, even more so for a deputy-pushed program on new territory. "People either want control over it or they want to kill it." Design: with Work's authority, three-star Shanahan as director gives Cukor top cover; Colonel Cukor as chief goes make it happen. Grant's verdict: "If it wasn't for Drew Cukor, Maven would never have existed."
Work thinks Cukor uniquely combines three worlds: he has fought wars, is technically savvy, and knows contracting and acquisition "down cold." "He was a bureaucratic ninja like no other." Standing order: whenever you hit a problem, come into my office. Cukor remembers Work saying: "You will fire anybody that resists you." Cukor to Manson: "So now the game begins."
Algorithmic revolution briefings
Ambition runs big. In a 2017 presentation to defense officials and potential commercial partners (slide notes Manson reviewed), Cukor says AI will transform war, outthink enemy commands, and run trillions of simulations about what an enemy might do next. One arresting slide runs a clip from the 2012 Bond film Skyfall: Daniel Craig's high-speed Istanbul chase; a widely available algorithm identifies objects in the fast-moving video, pinging boxes labeled "person," "umbrella," "horse," "chair." If the commercial sector can already do this, the Pentagon needs to catch up.
2017 briefing arc.
DoD needs to function more like a software company than a weapons factory against a "tsunami" of data; making the switch "may determine how successful we are in the next war." Regular updates to DoD software code could matter as much as training and morale. "I believe Google updates their core software several times a minute . . . we're lucky to update meaningful processes once a year." "Algorithms at the end of the day may be the key variable in how we win." He wants data processed "instantaneously" to bring target, weapon, and operator together and calls for an "algorithmic revolution." As with Work, AI is a route to automation. He warns: AI will make mistakes; false detections and constant correction; "We are still figuring out the golden utilization of AI."
He also goads the workforce whose support he needs. Recasting the Pentagon as a software company "is a massive and painful undertaking" filled with cultural, organizational, fiscal, and policy "pain." Software culture requires new people. Pace "not suited for everyone." Some will lose their jobs. "No doubt there will be some who are currently part of our major weapon systems who will need to move to other more stable tasks such as fifteen-year-long lead physical projects . . . trucks . . . ships . . . airplanes." However delicate the memo wording, Cukor seizes the controversy: "I know this is now going to be a knife fight."
The effort to put AI at the heart of how America makes war has begun. But Project Maven has almost no people. Besides a barebones crew, the Pentagon is not going to staff his project. Chapter 5 must invent the first Mavenites.
5. The First Mavenites
With almost nobody on payroll, Cukor builds a scrappy cult of Marine reservists who soon call themselves Mavenites. Deputy Garry "Pink" Floyd arrives June 2017; Colin Carroll joins without knowing what AI is; Sy Poggemeyer, Joe Larson, Mike Rhoads, and Dave Spirk take data, contracts, diplomacy, and integration. Eric Schmidt bets on "five pissed-off Marines." The chapter closes on Cukor's Mormon missionary youth in Panama as the origin of persuasion under chaos.
Scrappy cult inside a stodgy building
Colonel Drew Cukor sets out to build a team that will regard itself as a scrappy, subversive cult inside a stodgy, aging Pentagon. Before long they call themselves Mavenites. Besides Cukor, Maven has only one other officer on the Pentagon payroll: Air Force Lieutenant Colonel Garry Floyd, call sign Pink (from his last name). Pink developed lasting AI interest at NSA Alaska; Lt. Gen. Jack Shanahan brings him over as Cukor's deputy in June 2017.
Team that names itself.
Delivering the future of war requires more than one paid officer. "We had to stand this thing up out of thin air," Cukor says. Reservists are unlikely to know AI, but they are cheaper than contractors, Cukor knows how to call them up, and he wants people he can set on a task and trust to figure it out.
He has little respect for the Pentagon. He jokes the onsite pharmacy makes most of its money from statins and cracked-teeth remedies. He brings younger, more aggressive officers better suited for frontline combat than cubicles in the world's largest office building. He has followed some careers for years. He rates them hardworking thinkers who will not stop until they succeed, and he gives them license to break norms. The risky choice will make his Maven legacy and taint his career.
Schmidt's five pissed-off Marines
Eric Schmidt, then executive chairman of Alphabet, who urged DoD toward autonomy, is deeply skeptical of the Pentagon's ability to turn around. Per a New York Times report, he told four-star SOCOM commander Tony Thomas: "You absolutely suck at machine learning." Thomas wanted to throw him out of the vehicle. Yet Schmidt puts faith in Maven: "These guys are going to pull this off. Because you've got five pissed-off Marines who aren't just going to do it the same way."
Why Schmidt backs Maven.
Commander's intent is the Marine central tenet: assign tasks without specifying how. On this Cukor is no iconoclast: he gives the team almost no direction. The Corps endorses fury as method. An official service document: "Marines may use anger to accomplish missions with energy and focus. Properly channeled anger can motivate us to work toward goals and it alerts us that something is wrong and we need to respond."
Later reservists are amazed at freedom: casual attire, first names. One recruit has never addressed a colonel, much less given a view, without asking permission to speak freely. A Marine colonel usually owns at least two thousand troops. "It's a big deal just to be standing in front of this guy."
Cukor insists he needs "big personalities," not least so they tell him when he is wrong: "We train them to be like that, we reward them for being like that." Some agree he hired big personalities and disagree he wanted to hear he was wrong. Others say he always onboarded negative feedback to improve Maven.
Carroll: "I don't know what the fuck AI is"
Among the first: Captain Colin Carroll, Marine intel, age thirty-two. He has worked under Cukor for years: deception planner befuddling China about US Pacific intentions; most recently Iraq focused on ISIS. Cukor: "Come do this for six months. We need a deception plan." Carroll: "I don't know what the fuck AI is." Then: "Sure."
Joe Larson encourages him: it will be awesome. Carroll soon realizes he signed up to deceive China about something that does not exist. Maven barely has one piece of paper. Nothing to protect. Carroll, three Afghanistan deployments, figures Cukor called people "who in his past life he had abused and who could put up with it for twenty-four-hour days for seven days a week."
Separated by about fifteen years, some see Carroll as a "mini-Cukor." Another long acquaintance: "super smart and driven," wants the right things for the country, then adds the Big Lebowski line: "You're not wrong, you're just an asshole." Carroll announces on podcasts he is "just a general asshole," "competitive as shit" with a twin who went Navy while he picked Marines. Basic-officer instructor: would not be surprised if he ended on the front page for Medal of Honor or court-martial. "Ma'am, I'll take that as a compliment."
Later arc (foreshadow): fired twice from the Pentagon (Biden August 2021; second Trump administration ~2025). He suggests abrasive accusation first time, false leak accusation second. AI experience and shock at Pentagon tech turn him into an advocate of autonomous weapons. Lands at Anduril between firings; in 2025 starts a company for a $15 million autonomous plane.
First he spends more than two years on Maven, later calling it a "one in one hundred" team. He loves Maven, work, and Cukor. Posts to old teammates about waking up to wage an uphill battle against the "frozen middle" in the Pentagon basement among bureaucrats who work six-hour days and exist to say no, and still enjoying the camaraderie. Asked about the sociopath remark, he considers the definition and says he does not think he is one: he cares too much. "I don't know why I am the way I am."
Room 1B855 and the job split
First day: Carroll walks into Room 1 Bravo 855, a secure, windowless enclosure on the Pentagon ground floor. He has a Naval Academy aerospace engineering degree and has read a single fifty-page Boston Consulting Group AI guide: to make working AI you need data, contracts, and integration. That is almost everything he knows. He looks questioningly at the men; they stare back, unclear what is happening.
Day-one org chart.
Among them: Sy Poggemeyer, recently promoted Marine captain, age twenty-seven. Cukor first met him in 2010 as a new Marine; rated brilliant, dedicated, hardworking, promising. Poggemeyer remembers amphibious assault ships at age four on stepfather Navy captain's stateroom on USS Peleliu; two decades later serves on the same ship under "Peace through Power." Studied in St. Petersburg 2011. In 2014, as second lieutenant, he, Cukor, and two others published an award-winning paper on reforming intelligence. Others later think Cukor looked on Carroll and Poggemeyer almost as sons; the two never approximate brothers. Their dynamic will track Maven's ups and downs and eventually come apart.
Carroll assigns jobs. Poggemeyer, five years junior, is discounted from a leading role in Carroll's mind. Mike Rhoads, Marine artillery officer shot in front of Carroll in Afghanistan in 2012, is recruited. Larson should do contracts (Stanford lawyer); in a title-light office he becomes Cukor's unofficial chief of staff instead. Spirk, still in Tampa, is the diplomat getting SOCOM on board. Left: data and integration. Rhoads does data, Poggemeyer soon in support. Carroll takes integration (engineering: roll AI out to troops in working systems). Within days of visiting drone bases he is convinced AI will not work. He throws himself in anyway.
Pace as warfare; 7 percent overhead
To Cukor there is no difference between battlefield pace and Pentagon pace. "Why would I be different when Manson is in Kandahar than when Manson is here in the Pentagon? My argument was always the pursuit of technology is warfare." He wants a skinny operation: staffing and admin under 10 percent of budget (usual 20-30 percent). He says he ends at 7 percent. Soon a microcosm of Marine reservists, Pentagon civilians, and companies led by ECS Federal as prime contractor work Maven.
Operating tempo and myth.
Shanahan, whose key task is top cover, describes Cukor as a hard-charging, bullheaded disrupter who does not take no and breaks a little glass. "Everything we were doing was different and new enough that somebody, somewhere was guaranteed to be upset at him at some point, including people on his own team, because he was brusque. It was such a violent shakeup to the status quo."
Pace is too much for some. Cukor works eighteen-hour days, rarely appears to eat, expects the same. Infamous for double- or triple-booking. A 2011 contractor email: "Apparently he never sleeps." Wife Kirsten: he does not get grumpy; he suddenly stops. "Boom and he's down. It's like, oh, his battery just wore out." Office food mainstay: pub mix. Liquid: caffeine-free Diet Mountain Dew. No booze. Not even coffee.
Quirks. Rain might mean a good day for photosynthesis. "Should be fun" soon means nothing of the sort. Long stand-up morning meetings. Voicemail unchecked since 2005; send a text. Late-evening and weekend calls. Carroll stops setting an alarm because Cukor calls every morning at 6:30 a.m. Even today, if Spirk's phone rings before 7 a.m. or after 8 p.m., his children say: "Say hi to Drew Cukor, daddy." Returnees from six-month Afghanistan deployments get little thanks on day one back. Never personally insulting, rarely raises voice, withering toward failure. Some detect he gets names wrong when they mess up: failure rewarded with anonymity. He changes mind so often some ache that toil is destined for the shredder. Some feel humiliated.
Shanahan: some who worked for him in Iraq said they would never work for "that fucker" again, then relent: "Dammit he's a great boss Manson is coming back." One person finds him "Elon Musk-esque": you did not really want to work for him, except the work was so important to national security or saving lives or finding bad guys that even hate still saw brilliance. Another wonders about Stockholm syndrome: "I felt it was the most important thing in the world to work on." Larson: brilliant, capable, tech-oriented, unafraid of risk, innovative. Spirk credits him for his career. Brian Ward (ECS contractor, summer 2018): "a psychopath," sometimes "in the best way." A former Mavenite: annoying then; now "I miss it to death. Because it kind of made a lot of us what we are now."
Origin: LA, Mormon baptism, Panama
Cukor arrives at the Marine Corps in 1992 without a father, siblings, or much money. At Quantico he meets other young men from fatherless families: "Usually single moms, usually lower economic class, and usually lower opportunities. Suddenly we're all the same. We're all modestly aggressive and we know we're working within a system and we want to follow rules because we appreciate them. Manson was not exceptional."
Origin of the persuader.
He grows up in a small apartment in poor Los Angeles public housing with a young single mother in Reagan's 1980s. She works as bookkeeper then schoolteacher; he remembers they were not particularly close. Father's side: Hungarian Jews who landed in midtown New York; a small branch headed to Southern California. "He wasn't much in my life. He kind of walked away."
Around age four, a friend of his mother takes him to church Sundays. Mother "was just happy to see me gone probably on Sunday so she could rest." It is the Mormon church. At eight, youngest baptism age for the LDS Church, he is baptized: future without tea, coffee, alcohol, or tobacco (prohibitions he observes). At fourteen his mother converts too. Eager to leave home at eighteen. "It's not like it's the most loving family... Just a normal single mom trying to be helpful, and probably a very arrogant son who didn't want so much help." Soft, steady voice: "So you launch, you just go." Wife Kirsten later says he is private; he may not have said this aloud much.
Without funds, best path is ROTC: university education for officer service. "The military in the 1980s was really the only way to go to college, honestly, for a lot of us." In 1987 he interrupts USC, losing military scholarship, for a two-year Mormon mission. Church sends him to Panama: spreading an intrinsically American faith (Book of Mormon says Jesus visited ancient America) under Manuel Noriega, in Washington's crosshairs over Communism and drugs.
White shirt, paired with a Panamanian missionary who speaks no English. "You truly are on your own. You get used to that independence." US troops largely confined to base; Cukor's remit is to go all over. "Manson had no clue. We loved it." Mission leaders pull him out as tensions rise before the 1989 invasion; he starts over in Costa Rica. Sudden change, techniques to inspire. "I certainly was no St. Paul, but I did my part." Moderately successful conversions. "It was just a gentle message." A nineteen-year-old Cukor has discovered his powers of persuasion.
6. Relaxed About Fury
Maven's basement office becomes controlled chaos: HBO Silicon Valley cosplay, a water bottle labeled Federal Budget, Putin toilet paper, and Pretzelgate when Pink explodes at Carroll's "Chair Force" slur. Cukor is relaxed about fury. Illegal interview questions ("Are you married?") sit beside the celebrated answer "I want to reduce the non-American population." Parallel track: Spirk hunts SOCOM video, Work clears sanitizing Secret FMV for uncleared commercial AI talent, and Cukor sets out to invent America's first AI defense contractors.
Silicon Valley cosplay in a basement
The most some new Maven recruits know about the Silicon Valley ethos they are supposed to imitate is the HBO comedy Silicon Valley: young developers trying to change the world while not getting run over by dueling billionaires. Some Mavenites watch hoping to see themselves; some shield eyes from the show's intent to skewer people who think all tech is good. Satire is saturated with financial crises, firings, scheming, and AI. Mavenites soon gorge on all four.
Office as theater.
The group unites as insolent contrarians, proud of a windowless Pentagon basement. A small water bottle arrives labeled "Federal Budget": anyone from Maven walking past can drink; the project is guzzling cash. Other paraphernalia: toilet paper with Vladimir Putin's face. Office dynamic turns heads. "People are loud," one worker recalls. "There's some yelling going on, you don't know if they're mad at somebody or they're just on the phone speaking louder 'cause everybody else near them is also loud."
Sometimes Colin Carroll yells at Joe Larson at the whiteboard: "Joe, you're like fucking dumb. This is the dumbest shit." New ECS starters freeze; old hands: "Ignore them; they always yell at each other like this. They still love each other."
Pretzelgate
It is not always love. Air Force Lieutenant Colonel Garry "Pink" Floyd, only other active-duty officer besides Cukor, models the grown-up. Carroll seems too abrasive; Pink counsels slow down. Carroll seems bent on insulting other Pentagon officials; Pink notes those officials will outlive any single reservist. Better to work constructively than steamroll if AI adoption is to take off.
Pretzelgate sequence.
Pink's veneer can disintegrate. He overhears Carroll complaining about the "fucking Chair Force" (phone autocorrects to it). Pink storms over: "You have no idea what you're talking about! Manson is tired of cleaning up your messes! We are never going to get anywhere with you bullying these guys!" He slams a snack container; it explodes. Food flies to locker tops. Pretzelgate. Carroll stands: "Hey, first off, I hate the Air Force, but I hate the Marines more and Manson is a Marine." Larson steps in. Then Major Carroll issues Colonel Pink a command: "Go fucking clean all this up."
Cukor is unmoved by office explosions and uninterested in complaints, yet constantly offers personal counsel to irate individuals at his door. He knows how hard he works them. In this fracas he counsels Pink and Carroll: two strong personalities who must work together. Outbursts are not productive; pace is frustrating.
Besides, Cukor is generally relaxed about fury. "There'd be these flareups, right? And Manson had be like, go back to work, man. Like, this is how we do tech. We have flareups, we work through it, we apologize, and we move on." He sort of likes it: "Did I have a culture where people could tell Colonel to go pound sand? Yes. That allowed for that flexibility." To him it is military culture: office corridors no different from combat. "These are wonderful minds. Let's let them do their thing."
That does not work for everybody. Some cannot stand pace or dynamic, especially as OG Mavenites from Afghanistan and Iraq leave and a looser, less aggressive team replaces them. Cukor eventually goes through six rounds of people: "A copy of a copy of a copy." Others see tension as necessary cost of an innovation office, part of the glamor that appealed.
Recruitment, cartoons, Lululemon sacrilege
Word spreads through reservist networks. Transfers and friends-of-friends. Interviews can be perfunctory. Carroll, who recruits many in the first two years and says he and Larson "built an awesome team," sometimes runs interviews for as little as fifteen minutes: "You've got a clearance, you're a reservist, when can you start?"
Culture clash.
Jaim Coddington, Marine reservist recruited by Poggemeyer partway through 2018, is startled then amused by how little training new starters get on complex, little-understood technology. He sketches cartoons: "Go do AI!!!" to a bewildered newcomer; Cukor squirrelled away "working like a madman" while Mavenites line up at 7:13 p.m. Friday to clear slides ("Well Manson is sorry they can't have their early weekend; it's only 7"). Cartoons pin on the noticeboard for a couple of years.
Cukor always wears uniform; reservists do not. Team adopts startup fashion: gym pants, white-soled sneakers. One recruit feels embarrassed in smart Pentagon attire and ditches it the moment hired. Newer recruits look up to Carroll and Poggemeyer's fashion and verve; others find it untenable. The Pentagon is not Palo Alto. In a building of rules, uniforms, and hierarchy, Lululemon, Allbirds, and three days of stubble erase the divide between officer and tech worker and also range from unorthodox to sacrilegious. Marines, smallest service, derided as "crayon-eaters," use outward discipline to distinguish themselves.
May 2018 Marine Corps Uniform Board regs carry stunning specificity: one allowable visible tattoo from second knuckle of index to first knuckle of pinky; lower-leg tattoo must fit the Marine's own hand with fingers extended and joined, thumb flush; creases, polish, follicles all ruled. Commanders (like Cukor) must disseminate and enforce; "matter of tradition and pride, and of good order and discipline." Rules meant values. Values meant respect. Unclear if Maven was showing any.
Twelfth day of Maven; Kelly Martin; illegal questions
By end of 2018 a Maven Christmas carol lands in Cukor's honor: "On the twelfth day of Maven, Col Cukor gave to us / A Friday evening phonecon / Deployment with no notice / Disgrace for drinking coffee / A red eye to the Valley... Demands for 100 million / Shame for Joe's Detritus / TERRRRRRRRRY MITCHELLLLLL... Shit for leaving early / 4 marine reservists / A crude algorithm / 8 impromptu meetings / 103 slides to brief." (Joe Larson's mess reputation; an unaligned "Terry Mitchell" Manson could not track down.)
One recruit calls the vibe "controlled chaos" and looks up to Kelly Martin, Air Force lieutenant colonel who replaces Pink midway through 2019. Besides Cukor she is the only other active-duty officer; Brooklyn father taught her to curse. Mother of three at home; in office she becomes "mother of the misfit boys," kind, determined, clever, tactful "heat shield" against Cukor's wrath.
An unapproved June 2019 Pentagon-steps photo (Pink, Larson, Cukor, Carroll, Shanahan, Christ, Poggemeyer, Kovarna, Ward behind) circulates with Death Wheeler's email: "We'll see whose book it ends up in first."
Recruitment interviews by junior officers would not pass employment lawyers. "Are you married? Do you have kids?" Once a pregnant woman shows up and the representative wonders if he even has to interview. Compassion laces the illegal questions: whoever gets the job will work hours wildly inhospitable to family life and travel at no notice to war zones. Recruits can learn on day two they are going to Afghanistan. Some start married and leave unmarried. One divorcé blames Maven like the Apollo program's divorce spike. No one wants that attrition on their conscience.
Hiring red flags.
Third question sometimes: "Why do you want to work here?" Celebrated answer, per two people familiar: "I want to reduce the non-American population." Applicant gets the job.
Three drone classes and Spirk the data dealer
Maven plans AI onto three drone types, each needing separate algorithms on different sensors.
Three sensor stacks.
- Tactical (Aerosonde, Boeing ScanEagle): SOF-flown, 1,000-20,000 feet, close to troops.
- MALE attack (General Atomics MQ-1 Predator): 28,000-50,000 feet (source: 25,000-50,000), piloted from US control rooms (California, Nevada, Virginia). Weapons; FMV with timestamps and metadata; measure distances, heights, speeds; playback rates. Narrow field of view: looking through a soda straw.
- Bigger hunter-killers (MQ-9 Reaper): Hellfire, laser-guided bombs, JDAM GPS bombs; synthetic aperture radar through clouds, smoke, dust, dark. By early 2016, Air Force fleet of more than two hundred MQ-9s has flown nearly a million flight hours. WAMI wide-area motion imagery via Gorgon Stare (hundreds of cameras). Infrared for heat. Each system needs its own algorithm. Getting training data is more than half the battle.
Dave Spirk owns securing data. April 2017, ride in Larson's Honda Accord to Navy/Marine small-drone office PMA-263 at Patuxent River, Maryland. Closet of dusty hard drives from Marine Afghanistan 2010, years before the platforms they need. Driving home: "That data we saw was no good. We gotta find a way to get data."
Data-dealer path.
Idea: call a special-operations friend just back from Djibouti (Horn of Africa drone ops). "Hey, what do you guys do with all of that video that you were taking over Somalia?" Answer: with the Navy SEALs. Spirk: "Can I pillage and plunder some of your video?" Then-Rear Admiral Tim Szymanski, Naval Special Warfare: "Have as much of it as you like."
9 May: huge dump from Group 10 (stood up 2011 for drones and specialized surveillance for SEAL teams). Spirk feels like a data dealer. First-year Maven data largely from that Djibouti-friend outreach. SOCOM sees itself as vanguard; receptive to Maven gospel. Commanding generals take MIT AI courses; field AI into the battlefield becomes thinkable. In 2018 Spirk becomes first director of AI and chief data officer for SOCOM; two years later for whole DoD. General Richard Clarke (SOCOM spring 2019) later wants SOCOM first AI-enabled command.
Sanitize Secret for commercial AI
Group 10 footage brings problems: hours of arid nothing; deadly raids with classified imagery illegal to share with uncleared (sometimes non-US) AI experts. Plan: chop into three- or five-second useful clips; labelers identify objects; portion trains algorithms; separate portion tests.
Breakthrough: Work signs a decree that Maven alone may input full-motion video classified Secret, sanitize it, and share declassified imagery with commercial companies whose (possibly foreign) computer scientists carry no clearances. Teams scrub frames with strikes, gory casualties, or operational details (location, date, event) that could help an enemy or cause a stir. Spirk: "We didn't want anybody traumatized." Even bite-size, dozens of frames per second need intense per-frame labeling of people and vehicles.
Sanitize loop for uncleared talent.
Big primes (Booz Allen, Lockheed, Northrop) dominate. Cukor wants faster, smarter, less bureaucratic firms. He wants a new defense-industrial base. Drew Cukor wanted to create America's first AI defense contractors.
7. The Colonel and the Math Whiz
Cukor hunts America's first AI defense contractors and finds Matt Zeiler, Clarifai's one-man ImageNet sweep. Sunday Amtrak meeting, SUNet remote training, ninety-day algorithm horse races, and a disputed African river parable sell Zeiler's hipster staff on "save lives." Parallel: Joe Redmon (YOLO) is horrified at Hawaii and later quits computer vision. Clarifai stays; a couple quit; years later the site brags JSOC. Zeiler becomes an AI defense manufacturer whether he wanted the label or not.
One-man ImageNet sweep
Colonel Cukor knows America's future AI defense contractors will not be Lockheed-sized, at least not yet. One of the first companies he woos started with head count precisely one: a New York math whiz hungry for revenue who knew nothing about military work.
Matt Zeiler is an AI starlet. Company starts November 2013. Within a month it delivers the top five most accurate image-recognition models in the world, beating Google and others. He has not made his first hire; he does it himself.
Zeiler path to Clarifai.
Grew up outside Winnipeg, son of a doctor and a nurse; meant for medicine like his brother. "They all wanted me to go to medical school." A year of premed, then snag: physics and calculus feel more interesting than biology and chemistry. Switches to engineering science at University of Toronto, torn between computing and nanotechnology. Dorm-mate shows a flickering flame generated entirely by AI; the student studies under Geoffrey Hinton, who uncovered neural-network wiles. Zeiler writes an undergrad thesis under Hinton quantifying lovestruck pigeons' head-bobbing. Toronto is then one of only four schools deep into neural nets and image recognition (others: NYU, University of Montreal, Stanford).
Hooked, he goes to NYU under AI godfathers Yann LeCun and Rob Fergus. In 2012, while still there, an ImageNet breakthrough suggests AI detections with error rates far lower than one in four. He works on finding house numbers in street-view images; pinched for money, interns at Google Brain (only started 2011). One day Jeff Dean, Google Brain co-founder, asks him to present research to four hundred people; Zeiler realizes his approach works better than anything Google has. He considers Microsoft, Apple, Facebook (meets Mark Zuckerberg in one interview), declines what Google tells him is its largest-ever graduate offer, and starts Clarifai in November 2013. Three weeks later the brand-new company sweeps ImageNet with the top five most accurate models.
Visualizing the network; apartment GPUs
Zeiler's victory leads to a groundbreaking 2014 paper that does something more important: edges closer to visualizing how convolutional neural networks actually work, confirming for the first time what greats like LeCun and Hinton said for a decade or more: deep learning mimics how the human brain learns.
CNN visualization claim.
He explains gazing at the contest-winning machine under Perspex in his office (dusty plastic and wires): first layer of eyes filters orientations (vertical line, horizontal, angles, small circles). Next layer composes Ts, rods, cones. Third or fourth: eyeball feature, finger. Layers above: face or person. "My paper showed that these neural networks actually learn in that exact same way."
Neural networks rely on multiple layers of differently weighted artificial neurons, functioning like synapses, sluice gates that decide what information passes. Convolutional nets apply this specifically to images.
Zeiler uses his own money for the first server; first angel investment funds two more. Three machines packed with $60,000 of GPUs (specialist chips processing pictures and video at speed, storing code and memory, many calculations at once) overheat a small apartment in Manhattan's financial district. Each machine on a different circuit because of power draw: living room, kitchen, bedroom. GPUs draw so much heat he opens windows even in winter to avoid a sauna.
Startup compute reality.
First customers: stock media, travel, food. Biggest client: a wedding blog wanting tags so people find content. He builds a wedding-specific model: "cake," "veil," "bride," "groom," "suit." More images through models, more pay. Training mirrors brains: feed information, compare prediction to ground truth. Cake really a cake? If yes, invoice. If no, comb every mistake via backpropagation, update filter parameters so next three-tier stack predicts "cake!" better. "You go through millions of examples iterating like that, improving the parameters over and over and over."
War is bad; tech should prevent wars
No personal military experience; fascination honed watching WWII movies with grandfather. One friend goes Reserves; Zeiler amazed at days awake in mud. "It just sounded brutal." Broad view undented: "Like, obviously war is bad in general." Way he sees it: if technology will be used at all, it should prevent future wars.
Sunday Amtrak; SUNet; ninety-day horse race
Cukor proposes a meeting, swaps cammies for service charlies, boards Amtrak with Joe Larson. First Sunday customer meeting Zeiler has held. Still in his twenties, amazed to see the pair with military backpacks on New York streets, but thinks they "had their shit together." An hour into the demo, Cukor's commanding interjection: "We definitely need you to be part of this program."
Deal friction.
Catch: Clarifai and peers must remake approach. Swap training data from wedding paraphernalia for images of warfare. Details sparse. Cukor will not let any government data leave government systems. Clarifai must remote in via VPN to clunky, slow, secure SUNet. As clouds go, risible, but it counts as unclassified-but-protected workspace on government data centers where AI startups can train. Zeiler appalled; used to his own speedy high-capacity environment. Cukor knows the offer is hard and less performant: "It's government grade, which means it's not very good." He will not hand "exquisite" government data on a hard drive and risk hack or embarrassing leak. As he will say at a July conference: you do not buy AI like you buy ammunition.
Unorthodox Pentagon solution, novel commercial path: government owns underlying data and pays a recurring annual license to use algorithms; Clarifai retains IP but not access to its own model. "We'll pay you to use your tech but you cannot take it out; it's the government's," stiff pitch. "I didn't want any of our tech to land in China." Zeiler's concession with grumpy caveats: "Just know it'll be slower and dumber." Cukor barely hears. Government does not pay much, but amazing money for a startup. Demand: "Every ninety days you gotta give me a new algorithm."
Many Matts; Redmon's horror at Hawaii
Cukor is appealing in an age when most people who build sophisticated technology are "not very favorable" toward America's drone wars. At the July 2017 conference he blasts a public appeal to sell AI to the Pentagon, then flies to Honolulu with defense officials to scout talent at the top annual computer-vision research gathering. Each AI company he tries to enlist "had a Matt." Worth the effort: "These are guys that are literally building some of the greatest tech on the planet at the time."
One of those Matts: Joe Redmon, prior-year plaudits for work that quickly recognizes images, award-winning milestone You Only Look Once (YOLO). Thrilled CV can detect objects within twenty milliseconds, down from twenty seconds, using free open code and source models. AI can track threatened gorillas, find cancer, explore oceans. He works at Xnor, which secretly signs Maven (Redmon not on that team).
Redmon refusal arc.
February 2018 he describes horror at being approached at the 2017 Hawaii conference by a US military official. "His technology worked at levels no one else had been able to replicate," Cukor says, remembering a beeline for the gentle, thoughtful scientist preparing to present. Redmon sitting on the floor; Cukor joins. "He was so generous with his time. I gave him my spiel; we want to be more precise and accurate."
Redmon remembers slightly differently: a military official (he said army general; Cukor thinks maybe him) discussed video of his algorithms tracking and finding vehicles on battlefield drone footage. "Manson was kind of horrified," he later tells a TEDx audience in India. US drone strikes have killed thousands across the Middle East; hit weddings, schools; torn apart communities. "I couldn't imagine that my work could be a part of something like this." He perceives an urge to be stronger than anyone else regardless of human cost. Reminds him of Einstein on science's "disastrous by-product": mechanization and depersonalization leading to "abominable deterioration of ethical standards." Fear technology becomes a tool of the elite to consolidate power, not equalize it.
He later says he gave up computer vision entirely because of military applications and privacy concerns. "I loved the work, but the military applications and privacy concerns eventually became impossible to ignore." Views strengthen. June 2025 post: "The US is the largest state sponsor of terrorism in the world. Don't work on tech for the military."
Pilot AI chill; horse racing; dinner parable
Cukor has more success convincing others. For some, military work appeals because it is new and unfamiliar; for others it feels more like diving into war movies than real life. Pilot AI becomes a team favorite: unlike traditional vendors they are always "so chill" about changes not itemized on the contract, says ECS project manager Kateryna (Katya) Volkovska. "They were these crazy gamers from California.
Each startup gets a new task, graded ninety days later; miss the cut, dropped. "Manson is a Marine," Cukor says as if that explains it. "If we need a helicopter we have three on standby," channeling "two is one and one is none." "It's a mentality of no-fail. You need three Humvees because two will break. Always three backups." He likes to call it horse racing. Zeiler likes to call it unrealistic. Goalposts for judging algorithms can change halfway; teams: "What the hell?" Evaluation metrics make no sense to experienced AI practitioners.
Horse-race vendor loop.
But first Cukor needs his yes. As they discuss price by phone, Zeiler hems and haws. Negotiating tactic? Uncomfortable moral questions? Instead of brides with bouquets they will look for jihadis with Kalashnikovs. At least some of his brainy hipster New York workforce will swap cakes for bombs, and that may not go down well.
Clarifai likes to say its mission is to accelerate humanity's progress with continually improving AI; Zeiler's workforce is under the firm impression they refuse work for the military or for pornography "because they didn't improve life." "It's always, like, a divisive thing, like working with the military or not. I knew that was gonna be, like, friction."
Cukor senses the delay. He returns to New York one evening for dinner. The colonel always conducts himself exactly as Zeiler imagines a military officer: "regimens like to the tee." But Cukor also has, when he wants, the charm of a good storyteller.
Cukor tells Manson he does not remember telling Zeiler such a story. Manson strains to take the foundational parable seriously: that the military needed AI to stop needlessly shooting pastoral animals on which nomadic people rely. That was hardly Maven's driving idea. Bob Work wanted AI and autonomy to outpace China's rising military prowess, not save cattle in Africa. Plus a surveillance drone without AI might have been just as helpful. She waits to see if Zeiler has more.
Narrative gap.
But the military also undertakes offensive missions, she offers. Head in the sand? Zeiler says targeting decisions are tough but AI could do a more reliable job. "It doesn't get tired, it doesn't get hungry, it doesn't get stressed out," arguing targeting relies on lots of approvals and other steps, and full automation would be a long way off. "Ultimately it can make better decisions." He does not discuss targeting back at the office. Cukor's yarn has given Zeiler just the (nonlethal) ammunition he needs.
Zeiler rationalizes it is good remaining workers feel the same way and are aligned to company values. Those values have, via an unlikely war story set in the African brush, just taken a huge swerve.
In years to come, more employees leave as more details emerge, but Clarifai doubles down. Eight years after the dinner with Cukor, Zeiler's company creates more than a million algorithms, including facial recognition and sentiment analysis, and for a time proudly displays on its website that it worked for Joint Special Operations Command, home to some of America's most lethal and secret units. Whether he wants to acknowledge it or not, Zeiler has just become an AI defense manufacturer.
8. Somalia
Cukor: twenty years to rebuild Army after Vietnam; fifteen for night vision; AI would take twenty unless operators touched it now. "Field to learn." Christmas deadline. Carroll: Somalia friendly free-for-all. Fort A.P. Hill test; Baledogle SCIF with cell phones; network down forty-five minutes; school-bus clouds; fat yellow pens. Order: find someone who knows FMV and has tattoos. Justin Guzzardo, rubber ducky, sunflower seeds, eight suicides, AI in a closet, then a person in the bushes no human saw. "Just get it in their hands."
Twenty years; Boyd Rickover Doolittle; Christmas
Ch8 close.
Ch8 scale
Ch8 cycle.
Ch8 compare.
Ch8 ladder.
Somalia free-for-all; IA vs AI; school-bus clouds
Carroll figured Somalia good first landing: training data from there so detections likelier; operational risk lower than Afghanistan; friends in SOF, Naval Academy contact company commander MARSOC. "I chose a friendly site in an area where it was more of a free-for-all." SOCOM more willing on data; elite JSOC already "schwacking bad guys on a daily basis"; filtered to conventional. Internal nonpublic review: "JSOC wanted AI on everything... the first AI-driven organization"; Maven "would go on to put AI in every single formation they had." Somalia civil war low-tech: Manson 2013 beach with armed protection, camel milk; white arches pockmarked; UN-backed African troops pushed jihadis; al-Shabaab regained; twin bomb October 2017 500+ dead capital. Still: Christmas. Call into TIPS: different department said can't just plug algorithm into network; started talking "IA." Carroll: "It's called AI. I fucking know about it." Correspondent meant information assurance, Pentagon checklist secure software against physical/digital vulnerabilities. Carroll googling frantically: "Fuck me... Like what is all this shit?" Ended call; speed-read IA. Couldn't simply plug without approval; right way paperwork maybe up to a year for Authority to Operate. Mavenites not fans of paperwork; "insurgency mentality"; effective because didn't properly understand rules: "We weren't bogged down in our head." Cukor little interest sticking to policy; check: break law or just policy? Just policy, go ahead. Days later Carroll dangerous-enough: put algorithm on unclassified part before data entered secret network; fewer hoops; only AI inference onto classified. Hack algorithms? No one; still encrypted. Questioned: Marine charm, "Just shut the fuck up." Late September: test Fort A.P. Hill south of Fredericksburg; ScanEagle through TIPS; models against footage; just about worked. Baledogle Military Airfield fifty miles NW Mogadishu: looser; Somali nationals sometimes rolled into SCIF on cell phones during ops planning. "You could never do that in Afghanistan, Iraq. So we chose places like that where it would be less disruptive." (Year later al-Shabaab massive assault; hour-long shoot-out; 5,000-pound truck bomb largest on continent then; RPG/MG/rifles; sniper mortar drone killed all.) Hot December 2017 morning plug-in: something awry with IP. "We fucked it up somehow and we just went to lunch." Return jolt: accidentally shut down network; spamming JSOC with rogue packets; overloaded. "We brought the whole fucking thing down for forty-five minutes." Dangerous; Afghanistan/Iraq might never recover. Finally up: met Christmas 2017 deadline, first CV algorithm combat zone overseas within eight months of start. No one impressed. Boxes everywhere overwhelming; clunky bright yellow line so thick obstructed hands, bag or Kalashnikov mattered; looked like four-year-old fat pen. Wrong calls: every twenty minutes MQ-9 pan for weather check; alerts school bus in sky; mislabeling clouds. SOF decided easier and safer turn AI off. Maven took hint: narrow fainter lines; six updates in five days December; later little dots instead of boxes; cloud debate, industry want every pixel accurate; Carroll: never display box certain size/angle skyline, "Just never show it to a human." Vendors: cheating; Carroll: adjustment for user accuracy. Pentagon "ruthlessly candid feedback" slowed to drip; SOF seemed stopped using. If Congress heard, bad for funding and improvement. Cukor needed help: someone who stared FMV all day eyeball-to-pixel horrors; gentle consistent persuasion; sit alongside SOF get them turn AI on. Order: "Go find someone who knows FMV and has a bunch of tattoos."
Guzzardo; sunflower seeds; eight suicides; bushes
Justin Guzzardo came with both. First tattoo eighteen after dare, rubber ducky head above swimsuit line. Year after adoptive father (former Marine sniper married mom when Justin two) dissuaded music scholarship Texas: "'Don't do music.'" Born military base; go bag ready; no personal interest; dad advised Navy/Air Force for technical skill less struggle. Air Force training as secret drone wars took off; years Beale AFB California glued to screens calling details as bombs rained. Advice: stay busy, sensory, remember ground life or death. Cocktail fourteen-hour shifts: bag sunflower seeds per shift; Red Bull; others gum or rubber band wrist. Mostly kept asking questions; deeper into picture to stay awake; four or five hours sleep. Started "eyes on", couldn't take eyes off even to sneeze; call out time place building person clothes direction almost all day two years. Teal glow muddy fertilizer drying rooftops, possible IED ingredient. Graduated FMV screener: type furiously without looking what eyes-on said; Predators, Reapers, U-2s, Gorgon Stare; databases and slideshows; people counted sometimes cemeteries. Conventional then SOF, where everyone wanted; villages IEDs terrorist leaders; no room error; more kinetic. Single year: two hundred SOF assaults. First cheered high-fived bad guys; America's work. Changed: life is life; targets people too; less rejoice; relentless death. Five thousand hours ≈ two hundred days/nights emptiness, daily existence, or violence. Kandahar, Erbil, Beale; tiny black specks attacking bases or disappearing. Hush-hush: couldn't talk home; couldn't ask help, removed; whisper weakness fired; sock down emotions. Doesn't remember first teammate suicide; knows count: eight. Breakdowns; one gibberish on table start of shift. Tattoos everywhere: full right sleeve, chest, half left, back rib, pelvis; mementos friends lost; little unmarked. Less innocent; less starry-eyed. 2014 head honchos realized drone function crisis; policy changed late; psychiatrist and therapy dog; Guzzardo packing for nursing; tug back to war. Told Manson only east Africa contractor 2018; three others: Baledogle months after first algorithm. First of long ECS Federal "field service representative" line, integrators, salespersons, ambassadors. Bosses wondered lone wolf; niche connect with screeners. Cukor wanted language of FMV to talk forces into swapping exhausted watchers for indefatigable AI. Tattoos suggested what mattered launching robot wars was human rapport. C-130 from Djibouti heat: steel against passing out before Marines. Arrived dark: algorithm not only off but disconnected; closet. Week of intro; hand-holders home; plywood shack hot scrub; pulverizing AC. No rank: "Manson would have told a contractor to go pound sand." SOF aloof; unsure anyone eat with him. "No one owes me anything." Volunteer free labor; back to screener. ScanEagle small tactical; two analysts can't move eyes; now Guzzardo's too. After shifts convinced turn AI back on. Success still low, more than half wrong. Emerging tasks: count marketplace people; follow own convoys; track people, check targets, friendlies, civilians. Screeners wanted help; wasn't getting people/vehicles as needed. Parameters: not mythical magic; depends how used and circumstance. Raid scatter all directions; soda-straw FOV two pairs eyes struggle; all armed firefight extremely dangerous, scatter, return fire, wait grenade. One day algorithm spotted person hiding in bushes no one else saw by eye. "Holy crap," Guzzardo thought. "This could actually work."
Ch8 flow.
Further beats from the record
Ch8 hub.
9. Moral Outrage
Cukor wants Google's DeepMind "A" team; gets Cloud "B" team via Aileen Black, Mattis visit, and a potato-versus-landmine pitch. Secret $15M (path to $250M/year) deal breaks 6 March 2018. Whittaker, Fitzgerald, Nolan organize; petition; June Google refuses renewal and writes AI principles against weapons. Backups (Microsoft, AWS) step up. Years later principles softens; Cukor texts: "What a difference eight years makes." Google's models still go to Afghanistan via Xnor.
DeepMind A-team fantasy
Inspired by AlphaGo's March 2016 victory over the best human go player, Cukor always considered Google's DeepMind the global "A" team for AI. He wanted that team playing for the Pentagon. Ahead of the 2004 public listing, co-founders put into print the unofficial motto: "Don't be evil."
Ch9 ladder.
A-team blocked.
DeepMind co-founded by Demis Hassabis, North London child chess prodigy who bought a computer with winnings aged eight. Teen gamer dreaming AI could solve disease. Early backing from Peter Thiel and Elon Musk; ravenous for compute. Google snaps up the London startup for more than $500 million in 2014. Breakthroughs come quick, including AlphaGo over Lee Sedol. Hassabis absorbed by intellectual joy toward AGI; Cukor sees unexpectedly creative, risky gameplay as transformative potential of algorithmic warfare.
Unlikely target: Hassabis signed a 2015 open letter against AI in warfare with nearly five thousand researchers. "Most AI researchers have no interest in building AI weapons... autonomous weapons will become the Kalashnikovs of tomorrow." Kalashnikov (1947) became the ubiquitous weapon killing millions; fear is not a new nuke beyond use but a ubiquitous killing machine. Fears prefigure millions of deadly drones in Ukraine, a fraction able to kill autonomously. Letter signed by the Pope, Dalai Lama, Musk, Hassabis.
Stop Killer Robots; DIUx; Hassabis silence
Signatories inspired partly by Stop Killer Robots (2012): preemptive ban on fully autonomous weapons after founders read USAF documents planning their arrival. Mary Wareham (HRW coordinator): could not stop armed drones, but could stop the point of losing human control over force. Emotionless robots would make wars more likely and killing easier; civilians bear the brunt. Pakistan first country to call for a ban after hundreds of unacknowledged CIA drone strikes. UN takes it up; 2016 Iceland's Alþingi first parliamentary resolution endorsing a ban.
Ch9 flow.
More immediately consequential: thousands of AI-skilled scientists refuse warfare work. Military spending built Silicon Valley for decades; now Pentagon scrambles after drone wars and 2013 NSA revelations. 2015: Defense Innovation Unit Experimental (DIUx) under Secretary Ashton Carter. Wareham sees jeans-and-sneakers outreach as reframing an unconscionable arms race as a talent quest.
Hassabis sticks to unarmed guns. Cukor's outreaches land flat; Hassabis does not even respond. "Those guys were high on their horse. They're very anti-military." View, per Cukor: AGI would destroy humanity; "They wanted to figure out AI and then trap it in a box." (Google declined comment.)
Google Brain no; Cloud yes
Next-best Google Brain also no. Co-founded by Andrew Ng (GPU deep learning); 2016 Chinese-English translation breakthrough; Jeff Dean led; hired Zeiler's former professor Hinton in 2013.
Commercial path: Google Cloud, new division eager for business, developing computer vision, only two significant contracts in 2015, hunting clout and customers. Cukor thinks Cloud sets the commercial standard versus AWS and Azure. Cloud-based AI targeting.
How Cloud B-team signs.
C-suite fan: Eric Schmidt (DARPA money, Burning Man, "adult supervision" of founders; initially thought "Don't be evil" "the stupidest rule ever"). Mid-July 2017 conference: Cukor praises Schmidt calling Google an AI company. August: Defense Secretary Jim Mattis visits Google HQ (Marcus Aurelius in rucksack; advisers waking him to cloud). Pentagon will get better at integrating AI, he tells reporters.
Aileen Black's airport pitch; Google signs
Late Friday, policy/legal/contracting gauntlet still running. Cloud CEO Diane Greene calls Aileen Black (public-sector cloud) racing through SFO to her East Coast flight. Black loves the military: father two-star; West Point and Air Force Academy in family; she is one of few who did not serve. Maven is chance for duty. Enchanted by Cukor: weekend cell calls; "Manson had want to be in a foxhole with him any day." He portrays Maven as Google giving back. Also tells her about Lt. Gen. John Kelly's son who died. She is convinced Maven saves all innocent lives, not just American: identify at night someone carrying potatoes versus shovel and landmine. (No WAMI algorithm identifies so few pixels at that distance; maybe backtracking movement. Guardrail pitch to Greene: human involved; nothing to do with autonomous weapons.)
Ch9 hub.
Terminator does not exist yet. Google signs. Cloud has only four sales reps; huge coup; Google has no classified cloud at all. AWS person: "all sad face"; AWS already clouds CIA/IC, chased Cukor for weeks, sponsored a booth just to run into him. Cukor on snagging Cloud: "It was lovely."
Fei-Fei Li; $15M secret; Mountain View kickoff
Not for long. Furor will tear Google's workforce and set back Pentagon-Valley relations. Cloud chief AI scientist Fei-Fei Li (Stanford; China to NJ at sixteen; built ImageNet-scale sets of 14 million images / 22,000 categories; Cukor dismisses as "Flowers and nature... nothing one could use") joins Cloud 2017. Black wants public pride announcement; five-month race; Google beat AWS, Microsoft, IBM; "really exciting great win"; will leak anyway. Li warns weaponized AI is "red meat"; "Avoid at ALL COSTS any mention or implication of AI." Years later she tells Manson comments misinterpreted: Maven was "table stakes cloud," she was anti-hype. Black unwavering in docs: "this is a AI specific award." Award $15 million over 18 months; expected growth $250 million/year; also "priceless" sponsorship of secure classified cloud.
Ch9 compare.
Black loses. Signed September; not public; Advanced Solutions Lab quietly works. 13 October weekly update: deploy Cloud AI models on government platform for field tests. Next week fifty-strong Maven team visits Mountain View two-day kickoff; Google Cloud AI "the core." Focus: detection, classification, tracking in city-wide WAMI; ultimate aims scene recognition, event alerts, "very important" UI. End October: Cloud engineering at Beale AFB with drone pilots. Mid-December: 251 combat-theater images; by February 2018, 5,000 labeled in-theater WAMI. Cukor: Maven ultimately shipped "tens of thousands" of expensive annotated data; Google "of course" made CV models that subsequently ran in combat in Afghanistan. Models already detecting vehicles human labelers missed. Team assesses GPU cooling for WAMI/Reaper and how to airlift tractor-trailer kit to Afghanistan.
Four months into secret work, CEO Sundar Pichai tells a forum: "Countries need to demilitarize AI." Cukor already warns Shanahan: problems at Google; people do not like what they are hearing.
Fitzgerald, Whittaker, Nolan: three become three thousand
William Fitzgerald, Irish, County Waterford, head of communications Google Cloud AI, nine years at Google (sales; policy Pakistan/Hong Kong; typed Schmidt tweets; also protested Iraq war, read Chomsky, helped Snowden 2013, cautioned against Schmidt-Trump 2017). Watched Shanahan court Google; saw Cloud eagerness. Copied on Maven email; tells others not to take it; ignored; takes furtive action. "I didn't think it was a good thing for society or for Google."
Ch9 cycle.
Internal revolt chain.
Tries legal, policy, ethics; late 2017 reaches Meredith Whittaker (AI ethics). Stash of documents: Google developing algorithms for America's "illegal drone wars" on Project Maven. What do we do?
Whittaker: internal dissenter, academic; hates sci-fi AI debate when present bias is already crisis (Google Photos dogs/horses; dark-skinned people tagged gorillas). 2016 White House meeting on AI embedding injustice. With Kate Crawford founded AI Now Institute at NYU; 2017 report calls for strict limits on AI in weapons. Did not know it was in her workplace. Maven: "a very, very scary recipe." Eisenhower's MIC; Snowden fury. "Manson had a sense of how fucked this was, and of the real dangers of yoking an unprecedented surveillance powerhouse like Google at a global scale to the incentives of the world's most lethal military." Vanity about dignity: will not stay and green-light that.
Onto Signal with a third person. Campaign begins; three become three thousand.
Laura Nolan, Ireland-based engineer (self-taught mid-1990s), separately discovers Google planning an air-gapped system to run code and models for the Pentagon in its own data centers. Leadership says save lives; she does not see how a CV target-ID tool does only that. Warmongers always claim safety; deaths continue. "People always say it will save lives. And they don't really specify how." Dreams; gastric reflux; stress. Fear: after another 9/11, huge target lists plus political will equals thousands of strikes. Colleagues look at her "like Manson was a crazy person." Clarity: cannot work where work contributes to innocents blown up by missiles.
Fitzgerald: nothing short of canceling Maven suffices; ethics board is "professional dodge." Coordinates affinity groups, civil society; decides to leak. 6 March 2018: news breaks; "firestorm" inside company. Nolan can breathe. Fitzgerald already signed off for nerve condition (later attributes to stress); resigns same month.
Petition, June exit, principles; UN baton of shame
Pentagon refuses to confirm; Google: "pilot," "save lives," "non-offensive uses only." Fitzgerald: do not talk to us like we are stupid; this will kill people. Wareham writes leaders: project could contribute to fully autonomous weapons; object ID blurs into target ID for lethal force. (Manson: she was right; see Ch. 25.)
Ch9 scale
Petition draft: "We believe that Google should not be in the business of war." NYT makes public in April. Whittaker: quotidian until grand narrative. Sees on-ramp to Big Tech MIC; "There are many more Mavens." Defense leaders rail about her; many predictions prove right. More resignations; June Google will not renew Maven. New AI principles: will not design or deploy AI in "weapons or other technologies whose principal purpose or implementation is to cause or directly facilitate injury to people."
Moral cascade 2018.
Significant Pentagon setback. Some see China's hand in protests; generals fume Google helps China but not America; cannot control workforce. Campaign spreads: engineers, activists, Stanford students. Wareham appeals to UN SG António Guterres; landmine-ban analogy; ban as "baton of shame." September GA: weaponization of AI "morally repugnant." November Lisbon unscripted: "should be banned by international law." Furor extinguishes DeepMind battlefield dream. Shanahan: "Holy shit, what else do we not know about what's happening out in tech companies?"
Clarifai rumble; backups; eight-year snag
Dissent at Clarifai too. Zeiler June 2018 blog: active Maven participant; goal save soldiers and civilians. Never read into classified use; Maven AI getting into targeting. "Google didn't know because it was classified," one Mavenite. Employee letter: will Clarifai participate in projects that might lead to large-scale warfare, mass privacy invasions, or genocide?
Ch9 close.
Cukor sunny: "just a lot of noise." "We already had multiple backups." Orchestrates Google apology tour. Pichai quietly visits Pentagon; feels bad. Chairman Joseph Dunford accuses Google of working with China but refusing Pentagon. Pichai "took the blows."
Door closes; others open. Microsoft shows up loaded with algorithms; AWS finally gets the cloud work. Cukor: great relationship with Microsoft "turned the tide." Remembers Satya Nadella: "Whatever you need." Andy Jassy same. AWS: "We're not going to quit on him." Flow of contributors "just massive."
Absorption pattern.
Slowly C-suite objections fall. February 2025: Cukor sends news link; Google updated AI Principles; gone is refusal to develop AI for weapons. No commentary needed. DeepMind fuses into Google DeepMind 2023. Hassabis still says autonomous weaponry a very bad idea (2024 documentary) but February 2025 blog: AI that "supports national security"; "duty bound"; Western democratic values under threat. Months later Cukor texts a conference photo: "Google AI for ISR Operations," map of Iran, "Target Area." "What a difference eight years makes. Google is now all in supporting DoD." On his twenty-year horizon, he can stomach an eight-year snag that resolves his way.
Maven time; Google algorithms still go to war
During dissent, Maven keeps "Maven time" (no letup): overnight red-eyes, straight to work. Somalia just the start: over 2018, AI furtively into more than sixty sites in multiple countries including Jordan and UAE. Bad networks, reluctant users. Beyond drones: security cameras; blimp balloons over bases; phones and digital files from jihadi raids. Models detect rooftops, walls, entry/exit, roads.
Models still go to war.
Google's war algorithms live on, Goliath to David. Leadership simply gives Maven-developed work to the government. Xnor (one of first four algorithm companies) maintains Google's algorithms after contract ends 2019 (March 2019 docs; multiple sources). Despite moral fury, algorithms created by Google went to war in Afghanistan after all.
Further beats from the record
9 MORAL OUTRAGE “Manson had a sense of how fucked this was.”INSPIRED BY ALPHAGO’S MACHINE victory in March 2016 over the best human go player in the world, Cukor had always considered Google’s DeepMind as the global “A” team for AI. Ahead of the company’s 2004 public listing, the duo put into print the company’s unofficial motto: “Don’t be evil.” DeepMind was co-founded by Demis Hassabis, a child chess prodigy from North London who bought a computer with his winnings aged eight. As a teen gamer, Hassabis dreamed that AI could solve disease and fix the world’s health problems. In DeepMind’s early days, the company won backing from Peter Thiel and Elon Musk, but was soon ravenous for more computing power to develop its algorithms. Google resolved those problems when it snapped up the London startup for more than $500 million in 2014, making a big bet on AI as the technology that would one day rule the world. Breakthroughs at the company started to come quick, including AlphaGo’s victory over Lee Sedol in 2016. Hassabis seemed absorbed by the sheer intellectual joy of combining computer games with his pursuit of artificial general intelligence. The AI’s unexpectedly creative and risky gameplay revealed to him the transformative potential of algorithmic warfare. DeepMind was an unlikely target: Hassabis had signed a 2015 open letter against the perils of bringing AI into warfare, along with nearly five thousand AI and robotics researchers. “Most AI researchers have no interest in building AI weapons,” the letter read. “If any major military power pushes ahead with AI weapon development, a global arms race is virtually inevitable, and the endpoint of this technological trajectory is obvious: autonomous weapons will become the Kalashnikovs of tomorrow.” The reliable automatic assault rifle, created by young Soviet sergeant Mikhail Kalashnikov in 1947, became the world’s most ubiquitous weapon, killing millions of people across the world in wars large and small. The contention behind the letter, signed by the Pope, the Dalai Lama, and Elon Musk, as well as Hassabis and thousands of AI researchers, was that AI would not create a new weapon so powerful as to be beyond the reach of use, like a new and terrifying variant of nuclear deterrence. Instead, it would create a ubiquitous killing machine. Their fears prefigured the arrival of millions of deadly drones in the Ukraine war, a fraction of which can kill autonomously, with more underway. The letter’s signatories were inspired in part by the Stop Killer Robots campaign. When it formed in 2012, several groups came together seeking a preemptive ban on the development, production, and use of fully autonomous weapons after one founder of the group read some US Air Force documents and realized the US military was already planning for their arrival. “We couldn’t stop armed drones, but we could kind of stop the point at which you lose human control over the use of force,” Mary Wareham, an arms expert at Human Rights Watch who became coordinator of the campaign, told Manson. Sending “emotionless robots” into combat would make wars more likely and make it easier to kill, her group argued. Pakistan, which had suffered hundreds of unacknowledged CIA drone strikes that killed thousands of its citizens, became the first country to call for a ban. The UN took it up as an issue for discussion, and, in 2016, Iceland’s Alþingi passed the world’s first parliamentary resolution endorsing the call to ban fully autonomous weapons. Perhaps more immediately consequential, thousands of computer scientists with AI skills refused to work on warfare. Military spending had supported Silicon Valley for decades. Now, the Pentagon scrambled to make defense work appealing to a new generation of tech workers put off by the drone wars and the 2013 revelations of the National Security Agency’s mass spying programs. Worried by the trend, the Pentagon set up an outreach office in Silicon Valley in 2015 called the Defense Innovation Unit Experimental (DIUx) to bring in tech startups. Ashton Carter, the theoretical physicist turned US defense secretary, drove the initiative. But one man’s hero is another woman’s foe. Wareham didn’t buy this appeal to techies by defense officials in jeans and sneakers. She saw it as an attempt to reframe an unconscionable AI arms race as a more palatable AI talent quest. Hassabis also stuck to his unarmed guns: Cukor’s outreaches to Hassabis landed flat. Hassabis didn’t even respond to Cukor’s entreaties, Cukor told Manson. “Those guys were high on their horse,” Cukor told Manson. “They’re very anti-military.” There was just no way DeepMind was going to work on defense. They had the view, Cukor said, that artificial general intelligence, AI that could match or surpass that of humans, would destroy humanity. “They wanted to figure out AI and then trap it in a box.” (A Google representative declined to comment.) Cukor was striking out with what he saw as the next best option too: he couldn’t get Google Brain on board either. Co-founded by Stanford professor Andrew Ng, a computer scientist who argued for the use of AI chips known as GPUs in deep learning, Google Brain captured Cukor’s attention with its 2016 breakthrough on machine translation, when it scored a huge sudden improvement in translating Chinese into English. Jeff Dean, the engineer who led Google Brain, previously worked with Clarifai’s Matt Zeiler and in 2013 hired Zeiler’s former professor, AI pioneer Geoffrey Hinton. Cukor wondered if he would have more luck convincing a more commercial part of Google to open discussions with Project Maven. Google Cloud, a separate new division of the company eager for business, was developing its own computer vision algorithms. The new cloud division had only two significant contracts in 2015 and was in search of two central commodities: clout and customers.
10. The Algorithms Have No Clue
Brian Ward: Impossible bonhomie; Venn-diagram "What / The / Fuck" cap; two-hundred-pound people pleaser who wants to suffer. April 2011 Sangin fratricide: whispered "shooting W," Hellfire hits Marine and Navy medic. Carroll: "An algorithm would never have taken that shot." Afghanistan fall 2018: trees as people; "Don't say strike"; potato-chip AI; Philippines 70%→30%; Ranger newbie; Gorgon Stare "motorcycle" and "sidecar"; donkey train secondary explosions; toy donkey parting gift.
Ch10 cycle.
Ward the garbage man; Sangin whisper; unicorn hire
Brian Ward spoke with impossible bonhomie. Everyone fighting a war, working for country, or delivering a drink was A Great American. Instant friends; baseball cap with Venn diagram "What"/"The"/"Fuck"; throaty laugh; surliest waitstaff fist-bumping within moments. Drink beer, cider, sour, or double whisky; one meal a day; protein shake; stacked two hundred pounds, sixty of it working out from pole-vaulting Indianapolis teen. Thirty-five; people assumed decade older. Everyone who knew him would do anything for him: warm, wild, never-ending passion. "Violence was the only answer", not malice; nothing improves except through crisis; also his métier. Fifteen years in America's drone wars; three hundred songs in home studio as marriage fell apart 2016; straight back to Afghanistan with JSOC for Trump's "Mother of All Bombs" (largest nonnuclear US ordinance; Nangarhar caves). Rap lyric: "We are not heroes or villains, but when it comes to this work, we are more than willing." Bent over backward; friends moving called him; "I knew I wasn't their first call." Two-hundred-pound people pleaser; Ma'am, service, patriotism, duty, generosity, acceptance of violence. No one in Pentagon understood him either; left jobs regularly refusing bureaucracy. "He obstructed my ability to suffer", phrase he said often: want to suffer; life meant for suffering; won't put up with anyone obstructing ability to suffer. Friend: barely anyone had a clue what he meant. Pentagon clocked off 4 p.m. for bus; Ward driven by something else: "the warfighter" and his mama. Fourth of five; single parent; sent paycheck since Air Force 2009; drove Maryland to California overnight helping her move. "You're too nice," she'd say. Maven burned as all-consuming passion; flew bases; installed servers; ran jungle; didn't sleep; invisible "No" always just ahead to swat. "Drop me off somewhere terrible at night alone and Manson is happy." Got Maven onto nineteen different networks (two people). "Loud and obnoxious American"; longest tenure; farthest flights; bureaucratic blows; undaunted love; under radar. Called himself "garbage man": dirty work, unrewarded, how he liked it. Siege mentality after Google; Ward just wanted it to work. Afghanistan, Iraq, Syria, Jordan, Djibouti, Philippines as contractor. Maven would save lives, he'd say.
Ch10 scale
Ch10 ladder.
Enlisted Air Force 2009 age nineteen; geospatial intel analyst; digitally immersed in combat zones: "You see all these different parts of the world via a freaking screen." Patterns of life; kids; markets; phone switches; legal targeting packages. Early April 2011, Terre Haute Air National Guard; fifty-plus combat missions, 480 hours; live Predator over Sangin Valley Helmand; land mine and battles daily. Wednesday morning: Marines taking fire; scattered. Coming up midnight Indiana; Ward watched feed twenty minutes after fighting started; three men prone muzzle flashes pointed west, shooting away from other Marines. Chat: "3 friendlies in FOV... pers are shooting W." Senior officer second opinion: maybe facing opposite directions covering ground. Elsewhere others thought shooting east toward hunkered Marines. Ward couldn't positively ID friendlies; retracted ("disregard. not friendlies. unable to discern"). Ground team requested Hellfire prep. After strike request geospatial analysts barred from main chat unless women/children or LOAC about to break. Fourth person arrived two minutes later; Ward whispered on different platform: "pers are shooting W and the convoy is to the E" to California mission center. Assessment not passed to ground commander, radio controller, or drone pilot. Twenty-one-year-old didn't interrupt further; functional minimum. Seven minutes later Hellfire from same MQ-1B; seventeen seconds later hit own: Marine and Navy medic. First known drone-strike fratricide. Capt. Colin Carroll four or five kilometers away on foot another part of Sangin; heard boom; radio; paced to evacuation; chaos, blood, insoluble fury. Fourteen years later: "An algorithm would never have taken that shot", Taliban don't fight like that, shoot and run; AI trained on patterns and allied tactics would ID three Americans fighting defensively prone. Cavalier despair: "We kill the wrong people all the time... The machine can't be worse than a human." Jerry Smith, father of Marine Sgt. Jeremy Smith killed: would hug whoever at joystick; no ill feelings; daddies proud. Investigation: "lack of overall common situational awareness." Ward instantly lost respect for protocol; resolved never hold tongue; everything else unless life or death became noise. Next year deployed Afghanistan; twelve-hour shifts; nine things listening ten watching; sensory overload still did job; ISR tactical controller, drone strike guy; SEALs, 75th Rangers, CIA. "Tons of missions you'd hear in the news," friend. "I don't have a personal life"; marriage barely two years; "Manson is between divorces." Attracted to chaos. One long DC bar chat early evening; hands to face; weight of pain. "Sometimes Manson would come back from deployment..." Expected tired/hurt; instead slowed: "And Manson was just. So. Disgusted." Disgusted people home went about business complaining; couldn't stomach fripperies; only life-or-death; forty-eight hours without sleep; deaths; mindlessness back home. All that death so Americans could live free complaining.
Ch10 close.
Cukor unsure hiring Ward for global ops at Pentagon: knew deployment; not technical enough for AI systems. Contractor: "Well, sir, you're looking for a unicorn." Ward got job; plane back to Afghanistan within a week; turned out unicorn. Never slowed; no supervision needed; Cukor: fantastic; "I needed Brians." Green light from Gen. Tony Thomas SOCOM; still many rings to kiss; Afghanistan harder than Somalia; stakes higher; units skeptical to hostile; bases austere. By Ward's fall 2018 arrival Carroll and others already into eight sites starting Bagram summer. Soviets built Bagram 1950s; after 2001 fortress of US culture, Burger King, Pizza Hut, pools, movies, chai, war. Some joked Skynet flown in for jobs. Algorithm: trees as people; rocks as buildings; buildings as cars; distracting boxes; pointless; dangerous; suspicious unknown crew tampering equipment; legality questions; nobody wanted to touch. Models from Somalia; none trained on Afghanistan data. Operators minimized on screen running unseen. Former screener contractor recruited to convince colleagues; not even a believer; soon concluded wouldn't work. Ward's zeal; unlikely missionaries hunting vignettes. Honey not vinegar. Easier once models on MQ-9 Reaper feeds not just smaller tactical drones. Screeners scanning civilians before deadly strike; analyst spotted farmer with sheep in forty seconds; aborted; far too close. Replay with Maven: AI detected man "as soon as he walked onto the view, within a second, and tracked him the entire time." Email to Pentagon: useful in strike scenario. Cukor sharp: Don't say strike again, especially by email. Mid-2018 Google roiling; tech outrage over AI targeting. Exactly where it might save life, Mavenites thought. First two years fight for budget; Cukor didn't want AI blamed; wary scapegoat; single mistake could kill program; one bad call or civilian deaths even if not Maven might blame Maven. Later told Manson vision always about strikes; shorten targeting cycle. "I won't lie. That was always in my head, it was in all of our heads." Locate allied and enemy on map, complicated thing perfected hundred years. "Could we just use the AI to identify everything?" Nowhere near good enough. "The AI is just a bag of potato chips," he'd tell Ward; pissed; counted on algorithms and data getting better; build system AI could live in; operator feedback; wait. Microsoft forwent vehicle subclasses (tank/truck); soon only two classes: people and vehicles; scores up, flickering down; frustrated users with ethics codes and LOAC who cared deeply about child vs man not sedan vs truck. Ward: success 70% Afghanistan to as low as 30% Philippines. Dawned: Afghanistan white cars and black Hiluxes; Philippines carts with umbrellas; dusty yellow backgrounds vs thick green jungle edge lines.
Ranger newbie; Gorgon Stare motorcycle; donkey train; toy donkey
Not every user hated Maven. October 2018: soldier 75th Ranger Regiment to JSOC unit eastern Afghanistan. Newcomers rarely trusted; patches pecking order; arrived limping: one scroll, no tab (Ranger selection without Ranger School). "Some people aren't very nice." Deviated: experimented with Maven; pored over WAMI; AI detections; backtracking routes (Cukor's term) over hours to highlight missed patterns. Focus: ISIS-Khorasan after 2015; wanted new caliphate; stiff fight with Taliban after mutual war declaration; Mattis in Kabul September 2018 with Manson: "pretty hard fighting." Rangers quipped "Taliban Air Force", so much help targeting ISIS some felt Taliban calling in US strikes. Taliban negotiating government role; secret deal ("A superpower just lost a war," regional official). ISIS-K: Save the Children Jalalabad January (six dead, twenty-seven injured); late August rocket presidential palace; Manson landed two days after suicide bombing twenty-one dead ninety injured; US service members killed eastern Afghanistan same week. "My team was hunting them"; three or four strikes a day. Never really understood ISIS network. "One of my greatest regrets is we didn't get it better", same group Abbey Gate August 2021 (thirteen US, 170 Afghan civilians). Hundreds ISIS-K suspected southern Nangarhar mountains; hard tell mountain vs Kabul mission control; often word-of-mouth tips. One case: paid informant (wrong before) said ISIS-K meeting Kunar building; looser ROE for deal pressure; mission ahead; IR three people leaving; thermal outlines walking street; no other corroboration; Hellfire; next day local news three civilians dead. "Extremely sad... Everyone was pretty messed up." New Ranger first to give Maven time; Gorgon Stare on Reapers (Google algorithms). "When it worked it was beautiful; you would see an entire city move"; greatest value tracing routes of specific dots for overlaps and meetings. "We can just fly and watch you all day." Maven flew AI systems for two Gorgon Stare Reapers from Dover on C-5 Galaxy. Heavy compute at outposts: Kandahar WAMI powerhouse "the motorcycle"; smaller FOB Fenty Jalalabad "the sidecar." $15 million GPUs, backup generators, chiller in box size of two tractor trailers. Almost no one wanted to experiment; fire hose of data; AI initially worse. "Maven was never our bread-and-butter... pretty crude." Eyes spotted vehicles; AI saved time tracking. "Definitely had bugs, but they were refining it." Explosives for suicide vests high-priority ISIS Kabul and against Hazara; moving homemade explosives required infrastructure. Newbie experimented filters; began think Maven might ID complex network moving explosives. Aggregating AI track detections: low-tech donkey train through Hindu Kush on winding paths too rugged for wheeled carts. "Gorgon Stare was very capable with the donkeys." Mapped routes and assembly points; confidence one particular train carried not pine nuts but explosives. Pitched afternoon targeting meeting; commander first dismissed; no one believed random donkey train from so high; brief compelling; exacting questions; came around. Donkeys heavy-laden pouches; guys walking with them. Deliberately striking unarmed civilians a war crime; lives and careers; moral authority left. "Highly permissive environment" southern Nangarhar; focused on destroying ISIS not deal; "mission... focused on their elimination"; presumption anyone struck there was ISIS; "so routine." Could go very wrong: next year strike killed thirty civilians injured forty, operators thought ISIS training camp; farmers resting after pine-nut harvest. US shared scrubbed info with Afghan gov reps beside unit; believed terrorists; no objection. Commander made call. "You don't shoot at unarmed people... He was making a decision to kill people who did not have visible arms who turned out to be part of a terror network." Hellfire over southern Nangarhar mountains; donkey carcasses falling; secondary explosions nearly instantly, taken as evidence of explosives. Newbie wrapped up; donkey hit enough for respect with or without tab. Third person felt sorry for dead animal; SOF task force not so much. Parting gift: toy donkey.
Ch10 flow.
Nineteen networks and the strike that dare not speak
Ward's garbage-man ethic, nineteen networks, jungle runs, night drops into terrible places, is the human substrate under potato-chip algorithms. The Sangin whisper that never reached the ground commander is why he lost respect for protocol; the farmer-with-sheep abort forty seconds too late is why he believed AI could buy seconds humans waste. Cukor's mid-2018 "don't say strike" phone call is Google politics colliding with product truth he later admitted was always in their heads: shorten the targeting cycle; identify everything on the map. Microsoft collapsing subclasses to people-and-vehicles raised scores and lowered ethical usefulness for LOAC users who needed child-versus-man more than sedan-versus-truck. Philippines green canopy versus Afghan dust is domain shift in one sentence: seventy percent to thirty percent when the background changes. The Ranger newbie without a tab who trusted backtracking when paid informants lied is the adoption story the donkey train seals. Motorcycle and sidecar data centers, fifteen million dollars of GPUs on a C-5, Gorgon Stare city-moves, and almost no one wanted the fire hose until secondary explosions validated a Hellfire on unarmed-looking handlers. Toy donkey as parting gift is gallows humor over a permissive Nangarhar ROE that also produced the pine-nut-farm thirty dead. Abbey Gate regret from the same ISIS-K hunt hangs over the victory vignette. Carroll's "machine can't be worse" and Jerry Smith's offer to hug the joystick operator are the moral bookends Ward carries into every base.
Ch10 hub.
Skynet jokes at Bagram, operators minimizing Maven unseen, and the unbelieving former-screener missionary are the adoption desert Ward's honey-not-vinegar was hired to cross. The sheep-farmer abort and the donkey-train Hellfire are twin vignettes: AI as civilian-saver and AI as network-hunter inside the same permissive ROE. Cukor's potato-chip dismissal is impatience at models that could not yet carry the strike thesis he refused to email.
The Mother of All Bombs, three hundred divorce-era songs, and Madre flashing on the phone at night are Ward's full register: violence as métier, people-pleasing as duty, mama as north star. Unicorn hire within a week of Cukor's doubt is the personnel thesis: I needed Brians. Eight Afghan sites before his fall 2018 arrival mean he inherited hostility; nineteen networks mean he outworked it. Potato chips, two-class Microsoft models, and the toy donkey are the product, the compromise, and the souvenir.
Ward's Terre Haute midnight screen and Carroll's Sangin foot-pace to the blood are the same fratricide from two distances. The whisper platform that never reached the ground commander is why Ward vowed never to hold his tongue; the investigation's "lack of overall common situational awareness" is the bureaucratic epitaph. Jerry Smith's hug-the-joystick offer is grace the targeting machine rarely earns. Disgust at home-front fripperies after forty-eight sleepless hours is the emotional cost of being the warfighter's garbage man. Bagram Skynet jokes and minimized-background Maven are adoption failure until honey-not-vinegar vignettes and MQ-9 overlays make the sheep farmer appear in one second instead of forty. Don't-say-strike is Google politics; potato chips is product honesty; always-in-my-head is the later confession. Child-versus-man versus sedan-versus-truck is the ethics gap Microsoft's two-class shortcut opened. Philippines canopy domain shift is the warning that Afghanistan models do not travel. Ranger limping one-scroll no-tab who trusted WAMI backtracking when Kunar informants lied is the human hinge; motorcycle and sidecar GPUs are the metal; donkey secondary explosions are the contested proof; toy donkey and Abbey Gate regret are the souvenirs. Permissive southern Nangarhar ROE, presumption that anyone struck was ISIS, is the weather in which both the donkey "victory" and the pine-nut-farm thirty dead became possible.
Mattis in Kabul naming the stiff Taliban-ISIS fight, the Save the Children office attack, the palace rocket, and the twenty-one-dead bombing two days before Manson landed are the eastern-Afghanistan weather around the newbie's donkey brief. Three-or-four strikes a day without understanding the network is the regret that becomes Abbey Gate. Secondary explosions as evidence remain contested; the toy donkey does not adjudicate LOAC.
From Terre Haute whisper to toy donkey: Ward's chapter is one long refusal to treat bureaucracy as more sacred than the warfighter. Potato-chip AI and permissive ROE are the twin hazards; nineteen networks are the labor answer.
Further beats from the record
10 THE ALGORITHMS HAVE NO CLUE “The AI is just a bag of potato chips.”BRIAN WARD SPOKE with a kind of impossible bonhomie. Everyone fighting a war, or undertaking any work on behalf of the country, or just delivering a drink to him was A Great American. No matter where we met, Ward made instant friends. He showed up in a bar with a baseball cap emblazoned with a Venn diagram of three overlapping circles. One said “What,” one said “The,” and one said “Fuck.” As people made out the intersection of all three in the middle, he would be greeted with immediate confusion, and not a little alarmed delight. He could disarm even the most guarded of people in one throaty laugh or hello. Manson watched the surliest of waitstaff start fist-bumping him within moments. His drink was a beer, or a cider, or a sour, or a double whisky. If he drank kinda hard, he ate kinda light. The rest of him was all protein shake, as far as one could tell. At two hundred pounds, he said booze didn’t affect him. That’s how much lighter he’d started out as a pole-vaulting teenager in Indianapolis. “Even though it was scary, I liked to freaking run with that damn stick,” he said. Now thirty-five, he laughed as he told Manson his age. I didn’t say so, but he was right when he announced Manson had assumed he was a decade older. Everyone Manson met who knew him would do anything for him. He was warm and wild and he worked with never-ending passion. His wild warmth came with strange twists. He wasn’t aggressive, and there was no glint of secret malice in him. He meant, I think, that nothing ever improved except through crisis. He’d been in America’s drone wars going back fifteen years. After he took to writing songs in a home studio, producing three hundred as his marriage fell apart in 2016, he went straight back to Afghanistan with JSOC, in time for the Trump administration to drop the “Mother of All Bombs.” It was the largest nonnuclear ordinance ever detonated by the US military, on a cave system in Nangarhar province.
Ch10 compare.
11. Harbinger of Doom
Targeting a donkey with homemade explosives was hardly the job for Nvidia GPUs. Cukor knew the real bet was China; under the sea the chess match was existential. Capt. John McGunnigle's eight-week Buzzards Bay pilot; Woods Hole; Groton; Harbinger. Hunter vs Navy IT "Insider Threats" email; P-8 tapes; AWS Snowball; one Virginia-class sub with a single-GPU laptop ("meh"). IUSS, SURTASS, AUKUS Pillar 2. If you can find it, you can attack it. "We're going to apply AI to find submarines."
Ch11 hub.
Second strike; acoustic fingerprints; Buzzards Bay
Bob Work 2018: only area of clear US superiority over China still the submarine fleet. (Same interview: preemptive guided-munitions war would make "a lot of sense to be the one to shoot first.") Think-tank later: China likely erode undersea advantage 2026-2031. Former Naval Special Warfare operator: most secret part of strategic arsenal; nuclear war and end of world; fourteen "boomers" (ballistic-missile subs carrying more than half US nuclear arsenal) for second strike if homeland hit; job make a hole in water and disappear months; sometimes pop up Korea, Japan, Guam, Norway as nuclear signaling; max three-quarters deployed; twenty-four ballistic missiles apiece, eight to thirteen warheads each; one on constant alert each coast; yearly missile test (no warhead) to hit a dollar on a football field. Rest of fleet: forty-nine nuclear-powered fast-attack subs hunting others. Keep location secret: life and death. Find China/Russia nuclear-armed subs: top priority. "If you can find it, you can attack it." US pretty good when bad-guy subs leave docks; cat-and-mouse after; America liked to think of itself as the cat. China hid launches in cave pens and underground bunkers. Holy grail for adversary: undetected to either US seaboard; enough warheads to take out every big city and military facility; game over. Mid-2024 Guam: soft gray curve of a sub's back; man Navy flew halfway around world to fix one; half-joked reboot button; only he allowed. Fort Knox of Navy's creation; Maven found out. Ways to hunt: sub like whale makes noise; every sub its acoustic fingerprint; recorded, converted, spectrogram, analyzed; Navy prodigious worldwide database. Hydrophones towed by hundreds/thousands; spherical array in bow hatch like huge pitted golf ball; listening posts everywhere (towed arrays, seabed snakes, plane-dropped buoys); experiments adding digital ears to subsea fiber-optic cables. "That's how we view the world," submariner: "Passively." Parsing overwhelming volume hard; bad-guy subs disguise signatures; better look for anomalies than specific script; acoustic data largely unexplored like cutting-room drone feeds. Building new sub-hunters expensive; remote experts hard to lure; China manufacturing edge; tariffs only so far; AI possible shortcut. Maven saw chance to unlock secret sounds of the sea.
Ch11 scale
Ch11 cycle.
Ch11 compare.
Ch11 ladder.
July 2018: Capt. John McGunnigle (Boston submariner, CS degree; under North Pole ice and western Pacific; former squadron commodore; Undersea Warfighting Development Center Groton) idea apply AI to submarine acoustics. ONR funding eight-week pilot; Woods Hole Oceanographic Institution Cape Cod (WWII U-boat low-frequency research). August 2018: underwater drone/robotics expert Carl Kaiser and small team; eight hydrophones on flipped surfboard Buzzards Bay; fiberglass motorboat then underwater drone around it; simple algorithm tell difference: better than coin flip. Point: get data and test. Five-minute boat ride south: GoPro plus hydrophone under water at Woods Hole by Martha's Vineyard; algorithm alerts ferry vs normal boat; parked machinery vs fishing-boat thrum. "Good enough to show that if we could do all that in eight weeks, it was obvious that a real application of effort could succeed," Kaiser. Not first experiments; first landing with right Navy people. "Technological breakthroughs happen when society is ready... In this case, society was in no small part John McGunnigle." Meetings with Maven followed. Shanahan and Cukor understood persuasive figure who could deliver reams of passive acoustic data; asked what he needed: money and bureaucratic support. Got both. Maven immediately allocated resources; Harbinger born.
Ch11 close.
Hunter's Florida stink; Snowball; Virginia-class meh; IUSS; AUKUS
Mavenite Hunter (maritime data) caused unexpected stink October 2018 ~8 p.m. SPAWAR Newport systems engineer: "I am requesting an immediate halt to his efforts." Enterprise storage; word Hunter fiddling IT at Florida ops center; Jacksonville requesting data; "gained access to the building"; configuration changes; email cited "Insider Threats," "Securing Classified Information," "Need to Know." Hunter: signed travel approval, TS clearances, office call with captain and CO; wanted Navy Fleet aircrewmen upload videos to SIPRNet; when that failed asked Chrome install. IT sought permission; engineer red light (claimed he ignored route-through-her instructions). Seven years later: "I don't think people truly appreciate how disruptive we were being... not just inertia, but outright hostility." Battled through; secured data and tape readers; sent to Groton. P-8 data won. UWDC hard drives of sub data onto AWS Snowball then secure cloud; train algorithms; one day Harbinger onto single Virginia-class nuclear attack sub; old laptop with right security; young submariner built UI on one GPU. "When we first deployed it was something like, meh," Mavenite: balloon deflating. Persisted: AI on most important secretive geopolitical program. Hunter intended more subs and waves. Multiple companies developed algorithms. Navy ambition grew: port security cameras, ship cameras, SOF boats, even periscope. Harbinger for a while on SURTASS ships (Y-shaped trailing arrays; detect subs and mines; smaller expeditionary components more furtive). 2019 UWDC slide: AI/ML for navigation, sensors, also weapons. Also IUSS (Integrated Undersea Surveillance System): Cold War shadowy network classified until 1991; now focused on China too; slow upgrade for sonar operators and acoustic analysts across fixed seabed and mobile sensors. Data from wet SEAL Delivery Vehicles, dry mini-subs, Fixed Distribution System static sensors. Every military object a sensor; Maven/Harbinger suck every byte. Some beamed via Starlink Starshield into AWS. 2018 Beijing already one year into "Underwater Great Wall"; later AI underwater drones, surface vessels, undersea surveillance between first and second island chains; newer drones as listening devices. Three years after Russia invaded Ukraine February 2022: Moscow-Beijing ties closer; Russia moved seven subs into Indo-Pacific (US defense official); limits: Russia not sharing nuclear quietening tech. China's Navy world's largest fleet but only six ballistic-missile and six nuclear fast-attack; even 2025 Chinese subs still noisy enough to find (Type 096 may change). Pool allied data: underwater sound trap. September 2021 AUKUS: Australia buy conventionally armed nuclear-powered subs from US (not France; diplomatic spat); Pillar 2 meant Harbinger and submarine surveillance: three-country world-spanning undersea sound dragnet from sonar buoys and joint algorithms. Slow. Australian official 2025: already "frigging" eight years behind on military AI. Maven "bent over backwards" for Aussie data collection/labeling; Australian AI defense center discussion 2018 modeled on Maven; fell flat until Ukraine remade battlefield and officials panicked to "buy AI"; Maven-like center July 2024. Budgeting cycles and red tape on all sides.
Money, hostility, and the underwater Great Wall
Maven's allocation of money and bureaucratic air cover turned McGunnigle's eight-week curiosity into a named program. The SPAWAR "Insider Threats" email is the institutional immune system in writing: clearances and office calls were not enough when Hunter asked aircrewmen to put Chrome on classified desktops. That hostility is the same status-quo fight the drone-labeling team met on land, relocated under the sea. Snowball-to-cloud pipelines and a single-GPU laptop on a Virginia-class hull look unserious until the alternative is years of unread acoustic tape. "Meh" first impressions did not stop the Navy from asking whether the same stack could watch ports, ships, SOF boats, and periscopes. SURTASS Y-arrays and IUSS Cold War arrays gave Harbinger places to sit that predated Maven by decades. Wet SEAL vehicles, dry mini-subs, and the Fixed Distribution System turned every insertion platform into a sensor farm. Starshield terminals moving bytes into AWS are the commercial cloud arriving in the most classified corner of the fleet. Beijing's Underwater Great Wall and island-chain drones are the mirror program; seven Russian Indo-Pacific subs after 2022 raise the alliance stakes without handing China quietening tech. AUKUS Pillar 2's three-country dragnet is the diplomatic answer to that mirror: pool sonar buoys, run joint algorithms, accept that Australia was "frigging" eight years behind until Ukraine made "buy AI" a panic phrase and a July 2024 center. Type 096 promises may close the noise gap; until then Chinese hulls remain findable enough that find-it-attack-it stays the governing logic. Work's 2018 shoot-first aside and the 2026-2031 erosion window are the strategic weather Harbinger claims to answer. Cukor's wince when asked about submarine work is the operational boundary the chapter cannot cross: sensitive means the book stops at public and interviewable edges.
Ch11 flow.
Passive listening is the Navy's worldview in one word: hydrophones, spherical arrays, seabed snakes, plane-dropped buoys, and experiments hanging digital ears on the fiber that carries the internet. Disguised adversary signatures push analysts toward anomaly hunting rather than script matching, the same unexplored-data problem Maven met on drone cutting-room floors. Building new sub-hunters is expensive; remote sites struggle to hire; China's manufacturing base outruns tariffs; AI becomes the shortcut the industrial base will not supply. Kaiser's "society was ready" line credits McGunnigle as much as the Buzzards Bay math. Ferry-versus-boat alerts and parked-machinery thrums were not operational victories; they were permission slips for money. Once Maven understood McGunnigle could deliver reams of passive acoustic data, the answer was always money and bureaucratic support. The Florida SPAWAR fight over Chrome on SIPRNet desktops is comedy until it is not: classifying Hunter as an insider threat for doing the job signed travel orders already authorized. Tape readers to Groton, Snowball to AWS, one GPU laptop on a Virginia-class hull, that is how AI enters Fort Knox. Meh balloons deflate; persistence puts Harbinger on the most secretive geopolitical program America runs. Port cameras, ship cameras, SOF boats, periscopes, SURTASS, IUSS upgrades aimed at China after a Cold War birth against Russia: the ambition stack grows faster than the quieting gap closes. Every wet and dry insertion vehicle becomes a sensor; Starshield moves the bytes; Beijing builds its underwater wall; AUKUS tries to build a three-nation dragnet while Australia invents a Maven-like center in 2024 after years of flat effort. Find it, attack it remains the sentence under the ice.
Hunter's seven-year-later outrage and Australia's July 2024 panic center close the loop: institutions fight Maven-style disruption until a battlefield forces "buy AI." Harbinger's meh laptop and AUKUS Pillar 2 are the same bet at different scales, find the hull, share the sound, attack what you can locate.
Work's clear-superiority claim for the US submarine fleet and the 2026-2031 erosion window are why Harbinger exists. Fourteen boomers, forty-nine fast-attacks, second-strike theology, and acoustic fingerprints converted to spectrograms are the classified encyclopedia Maven tried to open with an eight-week surfboard test. Buzzards Bay coin-flip accuracy was enough because McGunnigle could deliver the data Shanahan and Cukor needed. SPAWAR's Insider Threats email, Chrome-on-SIPRNet farce, and Hunter's hostility memoir are bureaucracy fighting the garbage-man ethic underwater. Snowball clouds and meh single-GPU UIs on Virginia-class hulls are how Fort Knox admits AI. IUSS, SURTASS, SEAL vehicles, Fixed Distribution System, Starshield-to-AWS, Beijing's Underwater Great Wall, seven Russian Indo-Pacific hulls, Type 096 quieting promises, AUKUS Pillar 2 dragnet, Australia's eight-year lag and July 2024 center, the chapter's industrial map. Find it, attack it. Cukor's wince ends the public record.
The holy grail of an undetected adversary boomer on either US seaboard, enough warheads for every big city, is why acoustic anomaly hunting is existential rather than experimental. Guam's soft gray curve and the reboot-button contractor are the Fort Knox comedy before Harbinger's serious ask: money, data, and hostility survived.
Kaiser's ferry-versus-boat alerts were permission slips. Maven's immediate money and bureaucratic support turned them into Harbinger. P-8 tapes won through SPAWAR hostility; UWDC hard drives rode Snowball into AWS; one Virginia-class meh became the Navy asking about weapons on a 2019 slide. That is how second-strike mystique meets labeling logistics.
Fourteen boomers disappearing into holes in the water, yearly dollar-on-a-football-field tests, and China cave pens are the deterrence theater Harbinger claims to instrument. Passive worldview, towed hydrophones, golf-ball spherical arrays, seabed snakes, buoy drops, fiber-optic ears, meets disguised signatures and unread acoustic tape. McGunnigle under North Pole ice and in the western Pacific is the persuasive deliverer Shanahan and Cukor funded. Buzzards Bay surfboard eight hydrophones and Woods Hole GoPro ferry alerts are the eight-week proof. SPAWAR Insider Threats email citing Need to Know against a cleared traveler with office-call approvals is the Navy immune system. P-8 aircrew Chrome farce, tape readers to Groton, Snowball to cloud, Virginia-class single-GPU meh balloon, then asks about ports ships SOF boats periscopes SURTASS IUSS weapons, that growth curve is Harbinger. Starshield bytes, Beijing Underwater Great Wall, noisy-enough Chinese hulls until Type 096, seven Russian Indo-Pacific boats without quietening tech transfer, AUKUS Pillar 2 three-country dragnet, Australia frigging eight years behind then July 2024 panic center: the alliance map. Find it, attack it. Cukor winces and the public chapter ends.
12. Arms Race
30 September 2024, Chinese embassy DC: seventy-fifth PRC anniversary; megaphones outside; granite monolith. Maj. Gen. Liu Zhan ("Taiwan is China's Taiwan"). Cukor 2018: if enemy has AI and we don't, Marines die on the beach. Three post-Google waves: NSCAI (Schmidt/Work), JAIC ("Responsible AI"), Dunford "we are the good guys." Budget $16M→$93M→$1B. Shanahan: "This is what's going to allow us to kick your ass." Not intentionally an info op. "But of course it was." FOIA exemption; Five Eyes sting; Liu points to Palantir.
Ch12 cycle.
Ch12 compare.
Embassy granite; intelligentized warfare; Kania footnote
Seventy-fifth anniversary of PRC founding; Manson arriving Chinese embassy Washington last day September 2024. Campaigners outside with megaphones; granite and limestone monolith fusing East and West; space of four nearby embassies; Times on design aims: "sense of importance of China"; Kurt Campbell 2008: "let the Americans know that, while they were away fighting in Iraq and Afghanistan, China arrived." Embassy opened 2009: low tariffs after WTO; largest US debt holder; two years later second-largest economy (some counts overtook US by 2020). Defense spending trailed but rising; decade to 2007 more than doubled (second after US). Another decade after Campbell before any US admin made China focus of national defense strategy. January 2018 first Trump admin: China number-one challenge; Mattis raised "major combat." Fortress from which to contemplate AI putsch. Milling: Liu Zhan, PLA major general; prior year stage: "Taiwan is China's Taiwan... must be and will be reunited... We will not promise to renounce the use of force." US assessed Xi wanted PLA capable of taking Taiwan by 2027; Washington strategic ambiguity. Island focus of potential conflict and debate whether AI warfare hands advantage. Experts cautious about "AI arms race" language; Cukor never held back: existential and civilizational. Recalled 2018: "If the enemy has AI and we don't, young Marines and platoons are going to get killed out on the beach and we're not even going to understand what went wrong." Frank Hoffman (retired Marine; advised Mattis early Maven years): Cukor "razor sharp"; blunt: "We've been in an arms race for ten years." China released military AI vision days before Maven launch April 2017; Lt. Gen. Liu Guozhi (cutting-edge weapons agency): AI "will lead to a profound military revolution." Months later China AI development plan: global leader by 2030; PLA shift to "smart" warfare. November 2017 first Target Recognition and AI Summit Forum: intelligent machines could become "primary warfighters." Embassy attaché Liu knew Maven; Americans far ahead bringing AI to battlefield. "Intelligentized" warfare? Cryptic: "It is very difficult to know what's in the mind of your enemy." Sparse public PLA fielding; 2023 obituary Harvard grad returned 2014: "superbrain" for military decision-making; War Skull I/II for multi-platform air/land/sea/space. Embassy line: China wants safe AI; US public shaped by Terminator/Skynet; Chinese by heroic AI (Wandering Earth 2). US fretted nuclear C2 AI. Elsa Kania unconvinced; translated Chinese military AI writings; November 2017 report briefed Pentagon seniors as Maven raced first algorithm to Somalia; wondered fewer/different ethical/legal constraints. Footnote page 16: "To date, the PLA appears to have primarily adapted such concepts from U.S. writings"; closely tracked Work's Third Offset and human-machine teaming. China few clear definitions; US worried PLA quicker on uptake. Undersecretary Mike Griffin 2018: US average sixteen years to field new operational capability vs seven for China.
Ch12 close.
Ch12 scale
Ch12 ladder.
Three waves; info op; FOIA quiet; Five Eyes; Palantir tip
Fallout triggered overlapping efforts to get Big Tech and public to switch sides. Favored lever: ram home China threat. Three waves. First: weeks after Google outcry, Congress set up National Security Commission on AI; who's who of AI defense hawks; Eric Schmidt chair; Bob Work vice; Microsoft, Google, Oracle, AWS, In-Q-Tel members. Nearly 200 briefings and three years later: AI central to future of warfare; China threat alarms; urge develop/field/adopt AI weapons; acknowledge escalation, proliferation, instability; America no choice. Second: June 2018 Pentagon new office accelerate AI adoption; Shanahan left Maven to lead JAIC (Cukor team wrote proposal as early as August 2017 per drafts). Pivot and ethics refocus "absolutely part" of response to Google "episode"; Responsible AI to mollify; year-plus listening tour. Third: military leaders loud in Congress. Gen. Joseph Dunford (CJCS, Marine): Maven "not about doing something that's unethical, illegal, or immoral"; "hard time" with companies seeking China work refusing US military (November 2018); "We are the good guys and they should know that"; "whoever masters AI is going to have a decided competitive advantage." Budget: more than ten times, $16 million before Google fiasco 2018 to $93 million after 2019; rest of Cukor's tenure totaling $1 billion. Justify AI for war via specter of losing World War III. Eight years on Manson asked Shanahan if Maven was information operation: "Not intentionally... But of course it was." Early 2010s he supervised overseas messaging and propaganda (legal abroad, not to own people). Message carefully crafted for Beijing and PLA Eastern Theater Command (likely Taiwan invasion command): not oversell nascent capability (might induce Beijing faster); let China think US AI strong enough not worth taking on. "We had a message, this is what's going to allow us to kick your ass." Briefed China devoured whatever DoD put out; reported, translated, absorbed. "Everything we said." Wished Cukor would give interviews; refused: "missed opportunity." Unwelcome 6:30 a.m. call from Deputy SecDef Patrick Shanahan (no relation): "The boss ain't happy"; early-bird roundup to Mattis had included interview details; "I got hammered... revealing an operational capability to our adversaries." Cukor never gave a single interview; one early mid-2017 conference; barely anything public. After Google: steps to shield from publicity. December 2018 national security FOIA exemption (including request from nine months earlier) despite prior declassification for commercial vendors. "I still don't know how he got that," Shanahan; people scratching heads. "Everything went quiet"; "unofficial gag order," Lt. Col. Garry "Pink" Floyd; no LinkedIn workplace posts. Advance capability; avoid drama. DoD wanted public support for algorithmic warfare; no longer share what, where, how. Cukor: Google had nothing to do with FOIA exemption; determination cited capabilities and limitations enabling adversary ID of vulnerabilities; should have been exempt from start as critical infrastructure; timing an oversight (someone dropped ball). Manson could not confirm; 2024 Pentagon-commissioned review upheld theory: after Google uproar Maven became "less transparent and more insular."
Ch12 flow.
Allies clamored. Dome Plate twice-yearly Five Eyes conference: Shanahan briefed DI heads; senior Australian bowled over: "We really need to get on board... You could see they were onto something." British fed up. Air Marshal Phil Osborn (UK DI): Maven far ahead; don't run so far then wonder where partners are; stick together. Stung. "He was right," Shanahan. From then Shanahan tried read Five Eyes early; UK official: "sincere and generous." 2020 Cukor to Herefordshire SAS home; briefed; took up enthusiastically. UK saw Maven as compensate for limited industrial capacity: not enough guns/munitions; help decide where to aim limited firepower. 2025 Cukor: cutting edge of AI is Shanghai; "AI is almost free in China: we should be focused on adoption, adoption, adoption... We are in an existential fight with China." Day after embassy, Gen. Liu emailed four "important" AI governance docs suggesting cooperation and conservative use. Manson asked about US think-tank reports on Chinese military AI (untested weapons rush; Taiwan war-gaming; unmanned intelligence combat). No reply. Evening before Liu had said: if you want to understand AI for future of warfare, look at a single US company: "Palantir."
Hollow claims, quiet program, Shanghai edge
Manson's question eight years on, was Maven an information operation?, gets Shanahan's half-confession that structures the chapter: not intentional, of course it was. Influence-operations experience from the early 2010s taught him how to craft a Beijing-facing message without overselling a nascent stack. Patrick Shanahan's 6:30 a.m. hammering over an early-bird clip shows the cost of speaking too plainly even when the strategy wants China to hear toughness. Cukor's total media silence and the December 2018 FOIA exemption (including a nine-month-old request) are the domestic half of the same play: Responsible AI listening tours abroad in spirit; gag orders and LinkedIn bans at home. The 2024 Pentagon-commissioned review's "less transparent and more insular" line undercuts Cukor's oversight story even if the determination formally cites capabilities and vulnerabilities. Reading Five Eyes in early and the 2020 Herefordshire SAS brief are the repair kit; UK limited-munitions logic (aim scarce firepower better) is why partners wanted in. Cukor's 2025 Shanghai line, AI almost free in China; adoption, adoption, adoption, returns the arms-race drumbeat to industrial speed rather than ethics theater. Liu's four governance documents and unanswered war-gaming questions sit beside his evening tip: look at Palantir. The embassy granite and the megaphones outside are the same anniversary: China arrived while America fought elsewhere; Maven is how America answers with algorithms.
Ch12 hub.
Hoffman's ten-year arms-race bluntness and Liu Guozhi's "profound military revolution" sit days apart from Maven's April 2017 launch. China's 2030 global-leader plan and "primary warfighters" forum language are the mirror texts Kania footnotes as adapted from US writings, including Work's Third Offset. Griffin's sixteen-versus-seven-year fielding gap and the Google walkout together explain why Congress and the Pentagon reached for China as the lever to flip Big Tech. NSCAI's nearly two hundred briefings, JAIC's Responsible AI tour, and Dunford's "good guys" hearing are three instruments of one campaign. Budget math, sixteen million to ninety-three million to a billion across Cukor's tenure, is the campaign's scoreboard. Shanahan's carefully crafted "kick your ass" message for the Eastern Theater Command is deterrence theater that he admits was also propaganda. Cukor's silence, the FOIA exemption, Floyd's unofficial gag order, and the 2024 review's "less transparent" finding are the domestic half. Five Eyes repair work and the SAS brief are the alliance half. Liu's Palantir tip the morning after the embassy buffet is the adversary half: look at the company, not the speeches.
Liu's unanswered war-gaming questions and Cukor's Shanghai adoption mantra are the arms-race chapter's last turn: speeches about safe AI, budgets about existential fight, and a diplomat pointing at Palantir instead of Terminator myths.
Campbell's 2008 "China arrived" warning, Mattis's major-combat language, and Xi's 2027 Taiwan capability timeline frame the embassy buffet where Liu Zhan refused to renounce force. Kania's page-sixteen footnote, PLA adapting US concepts, and Griffin's sixteen-versus-seven fielding gap explain why Google's walkout looked like unilateral disarmament to hawks. Three waves (NSCAI, JAIC Responsible AI, Dunford hearings) converted that fear into a billion-dollar Maven scoreboard. Shanahan's kick-your-ass message, Patrick Shanahan's early-bird hammering, Cukor's FOIA quiet, and the 2024 "less transparent" review are one strategy with two audiences. Osborn's sting, Five Eyes repair, SAS brief, Shanghai adoption mantra, and Liu's Palantir tip close the loop: arms race as speech, silence, and vendor.
13. Daddy Karp
December 2024, Palantir Chelsea: frosted glass; white-gray shock of hair; Einstein letter without the regret. Overtook Raytheon; S&P 500; $100B then $164B day of visit. "Safe means that the other person is scared." OpenAI-Anduril; Palantir-Shield AI sign; Kendall in AI F-16; Brandon Tseng's Hell Week and Palantir-saved SEAL platoon; Corny's chickenpox map; "crazy Cukor"; LLMs like uranium; Maven "changed the history of the world", story Karp says he has never been allowed to tell.
Frosted glass; Department of Offense; bagel rooms
Outside Palantir boardroom early December 2024: only thing above frosted glass a shock of white and gray-streaked hair bolt upright, Alex Karp, twenty-first-century AI arms dealer; rest shadowy outline; standing. Parallels with Einstein beyond hair: prior-year Times op-ed approvingly cited Einstein's 1939 letter to FDR to "speed up" atomic bomb; six years and $2 billion later Little Boy and Fat Man three days apart; 100,000+ killed instantly; eighty years of diplomacy corralling rivals. Newest AI "equally significant," Karp argued; swift action for new arms race. Omitted Einstein fifteen years later: signing that letter the "one great mistake" of his life. Martial impresario on home turf; wild mood; money top of mind. Days earlier Palantir overtook Raytheon as top-valued defense contractor. Self-appointed promoter-in-chief for AI warfare; hawking AI weaponry Ukraine, Israel, America, Europe, more than a dozen countries. Hyperbolic bluster: patriotic, madcap, fast, cerebral, warlike. Reputation for allegedly helping find bin Laden (The Finish, Manson could not corroborate); spy detractors; NSA mostly stayed away; Army leadership briefed against until Palantir sued and won contract. Next day national security conference: only way Americans safe is adversaries live in fear. "Safe means that the other person is scared... need to wake up scared and go to bed scared." Peace activists "an infection inside of our society"; rename DoD "Department of Offense" (second Trump admin went further: Department of War); hung America's next nine hundred years on speeding AI weapons sales. Second Trump admin about to burst back with fulsome mandate. Biden outgoing speech warned tech-industrial complex five days before Amazon, Google, Microsoft CEOs stood with Musk at inauguration. (Karp, Democrat, donated $1 million for celebrations.) Even before November 2024 reelection: sudden bull run; September into S&P 500; October value crested $100 billion; overtook Boeing (ninety-year juggernaut) as infant twenty years after CIA first invested. Cukor December 2024: next in defense-tech is unions between software makers and hardware producers. "Everyone is getting ready for the next administration where it's widely understood there's going to be a big push on AI... Everyone is arming on the contract side. You are seeing rapid alignments." Right: Palantir expanded battlefield-data tie-up with Anduril (briefly in Maven; valued $14 billion); deal with Booz Allen; starkest from OpenAI, until early 2024 banned weapons/warfare work; week of visit announced Anduril partnership putting AI on battlefield. New tie-ups stronger vs five traditional primes under Trump; money into AI defense as never before. Afternoon 5 December: Karp signing deal to extend AI platform into autonomous warplanes on front lines. Manson ambling Chelsea office after press handlers cut loose. Dissonant cutesy: meeting rooms named bagel toppings; hour with CTO Shyam Sankar in "Poppyseed"; rare steak piled high; jazz-bar high black tables; talk of "a World War III thing"; stock record high day of visit $164 billion, four times prior year. War talk paid.
Ch13 close.
Ch13 scale
Ch13 cycle.
Ch13 compare.
Ch13 hub.
Sankar hankered days when Americans buying cars, cameras, cereal also subsidized arms makers; rued Ball Corporation quitting spacecraft for beer cans; US military ecosystem no chance of weapons at scale without civilian production. Heretical September treatise metaphorically nailed to Pentagon metro entrance (Luther 1517); Pentagon always wants faster horse not new transport. Senior official: "Look, if I ask for a three-wheel Jeep, give me a three-wheel Jeep." "There's no more offensive statement to me than that." Six months later Sankar commissioned Army Reserve lieutenant colonel, leapfrogging into Executive Innovation Corps with OpenAI and Meta, praise and conflict-of-interest consternation. Wanted compete with China on manufacturing; dual-use companies for lethal power; revel in total war; WWIII around corner. Evening email quote from Vannevar Bush (Raytheon cofounder; Manhattan Project) 1940: coming war technology war; US woefully unprepared; only solution outsource advanced weapons outside traditional services to civilians. "It could have been said today!"
Shield AI; Kendall dogfight; Hell Week; chickenpox map
Karp in room to sign with Shield AI: nine years developing AI autonomous aircraft flight; started small quadcopters; graduated washing-machine-size drones; fly warplanes without GPS/satcom including purportedly Ukraine. May 2024: convinced seventy-five-year-old SecAF Frank Kendall take off in AI-modified F-16; Edwards AFB desert; 550 mph; dogfight vs human F-16; human edged; some experts dismissed as stunt; early days; Kendall convinced after landing: "They don't get tired, they don't get scared." Shield AI raised $1 billion VC. Cofounder Brandon Tseng: failed BUD/S first try; "If I ever get a chance to do anything ever again, Manson is gonna go after it one hundred and ten percent. I never want to experience a setback in my life again." End 2009 got there. Coronado: seven weeks basic conditioning; Hell Week midway; "The only way out is through" white paint; beautiful simple beach with "ton of pain"; brass bell; three rings quit; Brandon never rang; SEAL Team 7 West Coast (created months after 9/11); "Forged in War, for War." 2013 second combat deployment Afghanistan squad then platoon commander; Oruzgan FOB; ops Kandahar, Zabul, Ghazni, Paktika; IEDs everywhere, roadsides, houses, motorcycles. Prior year 237 US KIA, nearly 3,000 wounded; SEALs killed by IEDs Zabul etc. Seven-month deployment: seventy-plus missions some three days; kill/capture Taliban and AQ; firefights constant. Lt. Andrew Cornelius ("Corny"), elite Navy EOD attached May-November: planned ops dodging hazards; almost nothing known (Cukor's rue since 2001); poor handovers; fighting new war each rotation; same traps year after year. Wary chokepoints. Friend before fly-in: get Palantir account. Never used; changed everything. Services submitted hazard reports separately; overlay cumbersome; gaps. Palantir aggregated; plotted past explosions on planning map. Ahead of village raid: directed to report by another EOD tech year before Corny knew; called; safeguarded mission. Spot IEDs: crouch disarm or blow with motorcycles. Toward end: turnover op with next platoon (historically dangerous, different styles, never worked as team). Search: "map had chickenpox", record IED warnings on satellite imagery. Showed Brandon; too hot. "We had never seen something like this before." Nighttime; "Our principal way of avoiding IEDs is with our eyeballs... we're not super great." Suicide mission; village no strategic significance; bullshit. Bringing up with non-SEAL new regional SOTF commander almost scarier than raid; SEALs execute, don't question. Four days most stressful of life; lost weight. "I'll do any other mission" against heavy pushback. "A lot of American taxpayer dollars goes into building... Navy SEALs, and you don't want to lose them." Palantir data ammunition; switched target farther south. "Palantir saved my life... The intelligence that they provided my SEAL platoon in 2013? Like, holy smokes." "Very, very visceral"; "forever Palantir fan." "It prevented us from getting blown up. Multiple people on my platoon have the same reaction." 2013 Corny only laptop at base; 2015 Iraq mobile Android on every raid. Same year Brandon out after seven years; suggested brother Ryan start defense-tech so disposable robots enter buildings first; Ryan (sold Wi-Fi charging company) initially dismissed; investors hard; seed money; built from scratch tiny flying robots to midsize drones and fighter jets; AI pilot on seven other companies' aircraft. 2021 snagged Doug Philippone (former Army Ranger; Palantir long-running global defense lead); money and board 2022. From start Philippone foresaw tie-up: AI everywhere, find targets real time and link to weapon systems and sensors frontlines. Shield AI autonomous drones already Ukraine, India, Latin America, Pacific (how well/autonomously unclear to Manson). Instant data share with Palantir; someday swarm warbirds; surely munitions. Boardroom signing day: "I just saw this vision... America is only gonna win if we pair the very best companies... I want America to win, period. Like, if we lose the next war, money doesn't matter." Karp made clear how much money was part of game. Under "Fortune Favors the Fast" sign later that day, off cuff about Drew Cukor: only person who saw value and implications of LLMs for warfare years before they arrived; approvingly "crazy Cukor"; almost worship. Two weeks earlier by-invitation DC national security event: Karp called Cukor "founding father of AI targeting." Now: as AI remade world, LLMs would function like uranium; chatbots power new warfare threatening world and remaking diplomacy; dark explosive vision beginning on battlefield with Project Maven. "It's really changed the history of the world," of Maven. But, Karp went on, he has never been allowed to tell that story.
Bagel toppings and uranium chatbots
Einstein's omitted regret is the chapter's first critical cut; Karp's "equally significant" AI-as-atom claim is the sales cut. Overtaking Raytheon and Boeing, cresting one hundred then one hundred sixty-four billion, S&P 500 entry, and a Democrat's million-dollar inauguration donation are the money weather around "Department of Offense" rhetoric and peace-activists-as-infection talk. Biden's tech-industrial-complex warning five days before Big Tech stood with Musk is the political frame Cukor's "rapid alignments" forecast fills: Palantir-Anduril data, Palantir-Booz, OpenAI ending its weapons ban for Anduril the week of the Chelsea visit. Poppyseed room steak and World War III chatter are dissonant cutesy over Sankar's faster-horse rage and Vannevar Bush email triumph. Three-wheel Jeep official is the acquisition villain; Executive Innovation Corps reserve commission with OpenAI and Meta is the revolving-door question. Kendall's 550-mph AI F-16 dogfight, human edged, experts calling stunt, Kendall saying they don't get tired or scared, is the hardware demo for the Shield AI signing. Brandon never ringing the Coronado bell, seventy-plus Afghan missions, Corny's chickenpox map aborting a suicide night raid, "Palantir saved my life," Android-on-every-raid by 2015, Ryan's seed-money grind, Philippone's board foresight, swarm-warbird munitions vision, "if we lose the next war money doesn't matter" beside Karp making clear money matters: the founder mythology. Fortune Favors the Fast, crazy Cukor as founding father of AI targeting, LLMs as uranium, Maven changing world history, story never allowed to tell, the closing hymn of the arms dealer who wants adversaries scared to sleep.
Ch13 ladder.
Karp omitted Einstein's regret; the chapter restores it. Little Boy and Fat Man, one hundred thousand dead instantly, and eighty years of nonproliferation diplomacy are the atomic analogy he wants for AI without the later "one great mistake." NSA distance, Army briefings-against until lawsuit, uncorroborated bin Laden lore, and more-than-a-dozen-country hawking are the company's contested résumé before the Chelsea signing. Sankar's Ball Corporation beer-can lament and Luther-nailed treatise are the manufacturing theology; Brandon's forever-fan testimony and Corny's laptop-to-Android arc are the operator theology; Philippone's best-with-best pairing and Karp's uranium LLMs are the sales theology. Maven as world-history hinge that cannot be publicly told is the gag order wearing a billionaire's voice.
Shield AI's washing-machine drones, purported Ukraine GPS-denied flight, seven other companies' aircraft carrying its AI pilot, and swarm-plus-munitions futures are the hardware half of Cukor's software-hardware union forecast. OpenAI's early-2024 ban reversal the week of the visit is the ethics half collapsing into the same deal flow. Stock at one hundred sixty-four billion four times higher than the prior year is the market half. Scared-to-sleep adversaries are the product definition.
Poppyseed steak, World War III chatter, and a stock chart four times higher year-over-year are the Chelsea dissonance. Brandon's four stressful days arguing off a chickenpox night raid are the visceral half Karp's uranium metaphor tries to productize. Never allowed to tell the Maven story is the FOIA quiet wearing a CEO's mouth.
Further beats from the record 13 DADDY KARP “Safe means that the other person is scared.”I HAD MADE IT to the outside of Palantir’s boardroom. The only thing visible above the uppermost edge of frosted glass that ran across the length of the room was a shock of white and gray-streaked hair bolt upright. It belonged to Alex Karp, the twenty-first-century AI arms dealer. The rest of him, much like the future of warfare, was cast in a shadowy outline. All one could perceive, struggling to see through the opaque wall of the large office room in early December 2024, was that he was standing up. The parallels with theoretical physicist Albert Einstein went further than Karp’s hair. In an op-ed for the New York Times the year before, Karp approvingly cited Einstein’s 1939 letter to President Franklin Roosevelt entreating him to “speed up” efforts to build the atomic bomb. Six years and $2 billion later, the US detonated atomic bombs named “Little Boy” and “Fat Man” three days apart on Japan’s Hiroshima and Nagasaki, killing more than 100,000 people instantly. Attempting to corral and limit rivals’ development and use of the same weapons would shape the next eighty years of international diplomacy. The capabilities of the newest AI technologies are “equally significant,” argued Karp, calling for swift action to breathe a new arms race into life all over again. Karp omitted to mention what Einstein said fifteen years later: signing that letter to Roosevelt was the “one great mistake” in his life. Now the martial impresario was on home turf, gesticulating in his boardroom. He was in his customary wild mood, and he had money top of mind. A couple of days earlier, Palantir had overtaken Raytheon to become the top-valued defense contractor in the world. Karp was self-appointed promoter-in-chief for AI warfare. As CEO of Palantir Technologies Inc., he had spent the past few years hawking AI weaponry around the world. A master of hyperbolic bluster, Karp fashioned his own hymn sheet, patriotic, madcap, fast, cerebral, and warlike, and sang from it at every opportunity. The company benefited from its reputation for allegedly having helped find Osama bin Laden (suggested in The Finish by Mark Bowden), that one could not corroborate, but it earned spy detractors too. The eavesdropping National Security Agency mostly stayed away from Palantir, and the Army leadership briefed against it for years until Palantir sued and won a contract. The day after Manson visited, Karp would roll out his vision for weaponized AI at a national security conference. The only way for Americans to be safe, he argued, was to make sure America’s adversaries live in fear of the country. “Safe means that the other person is scared,” he said. America’s adversaries “need to wake up scared and go to bed scared.” Karp described peace activists as “an infection inside of our society,” wanted to rename the Department of Defense the “Department of Offense” (the second Trump administration would go one further and change it to the Department of War), and hung the future of America’s next nine hundred years on speeding up the sale of AI weapons to the world. The second Trump administration was about to burst back onto the stage, with a fulsome mandate it could take all the way to the bank. In his outgoing speech, Biden would warn of the dangers of the rise of a tech-industrial complex five days before chief executives of Amazon, Google, and Microsoft would stand alongside Elon Musk at Trump’s inauguration ceremony. (Karp, a Democrat, would donate $1 million for the celebrations.) Even before Trump’s November 2024 reelection, Palantir enjoyed a sudden bull run: in September 2024, the company made it into the S&P 500, and in October its value crested $100 billion. Palantir, a mere infant twenty years earlier when the CIA first invested in the startup, overtook Boeing, a ninety-year-old defense juggernaut. What was going to happen next in the defense-tech world, Cukor told Manson in December 2024 with the foresight to which Manson was now accustomed, was the arrival of new business unions between software makers and hardware producers. “Everyone is getting ready for the next administration where it’s widely understood there’s going to be a big push on AI,” he said. “Everyone is arming on the contract side. You are seeing rapid alignments.” He was right: in quick succession, Palantir expanded a tie-up sharing battlefield data with Anduril Industries, a defense-tech startup that briefly participated in Maven, valued at $14 billion. Palantir also struck a deal with Booz Allen. The starkest move came from OpenAI, the company at the heart of the global generative AI explosion. Up until early 2024, it banned users from working on weapons development and warfare. No more: the week Manson visited Palantir, OpenAI announced a partnership with Anduril, bringing its AI directly onto the battlefield. These new AI tie-ups would make defense-tech platforms stronger, all the more ready to duke it out with the five big traditional prime defense contractors under the Trump administration and reshape the field, Cukor told Manson. Money would flow into new AI defense companies, from investors and buyers alike, as never before. Now, on the afternoon of December 5, Karp was signing one of those deals to extend his company’s AI-enabled platform into autonomous warplanes on the very front lines of war. And Manson was ambling around Palantir’s New York office in Chelsea solo after the company’s press handlers cut me loose. For a company dedicated to offensive weaponry, it had its dissonant cutesy elements. One row of meeting rooms was named after bagel toppings, and Manson spent close to an hour chatting through the niceties of military contracting with Palantir’s chief technology officer Shyam Sankar in a toasty little room named “Poppyseed.” Slices of rare steak were piled high on an inviting food counter at one end of a large gathering spot filled with high
Ch13 flow.
14. Palantir, Palantir, Palantir
April 2025, Georgetown: Chewbacca statue, Karp hip-hop posters, and the lab where Maven Smart System was born. Aki Jain says Drew forced him to revisit AI priors; Brian Ward says Palantir's skyrocket is Cukor's. Hot GPUs melted $2 million boxes; Imig built a Netflix-protocol video player with one good eye after cancer surgery; SEAL Team 6 at Dam Neck nearly ejected the hardware over slant-count risk; Djibouti called the prototype sucked, then tracked mopeds. Cukor: when Maven started, most people thought Palantir was exiting the Pentagon. By mid-2019, "We had the entire JSOC community."
Ch14 scale
Georgetown lab; Cukor forces the priors
A few months after the frosted-glass Karp encounter, Manson wandered near a Georgetown canal to another Palantir enclave: seven-foot Chewbacca; wall poster of imaginary Karp hip-hop hits including "Public At Last" for the 2020 NYSE listing. April 2025: he wanted the Maven story Karp claimed he had never been permitted to tell, the one that "changed the history of the world." It would help change the company's: August 2025 Palantir bagged a $10 billion Army multi-license agreement; value soared to a record $411 billion, top-twenty-five globally. "It's hilarious to watch Palantir skyrocket," Brian Ward. "I don't know how anyone can argue it wasn't because of Drew Cukor." Hyperbole? Akash "Aki" Jain, Palantir CTO and president of the US government subsidiary, was not far off: "A lot of our success in defense absolutely is tied to Maven; that's very obvious." Nearly 60 percent of revenue from government customers. Beyond contracts, Jain thinks the first thing Maven and Cukor did for the formerly AI-skeptic company was enable it to embrace AI. "Drew forced me to visit my priors on AI and accept there might actually be value there... He was insistent you've got to figure out AI." Jain, self-described AI skeptic in the 2017 timeframe, told a SOF conference May 2025 he now eats, sleeps, and dreams AI. Cukor also counseled reconstitution inside DoD. Outcome could have been very different. When Maven started, Palantir was a source of opprobrium for some in the Army and for parts of the defense and intelligence world slowly pushing them out, NSA to SOCOM. "We brought Palantir back to life via Maven," another Mavenite. (Some Palantir officials tempered this, citing smaller defense contracts at the time.)
Ch14 close.
Ch14 ladder.
The lab: virtual-reality room with bare cork table that becomes a bright green island under AR goggles, missiles swinging; brain races; war as trite and fleeting as a video game. Beyond: wires, headsets, whiteboard boats in teacher colors; warm golden retriever jumping for stretches and food. Human-canine teaming as prelude; human-machine teaming the higher agenda. Far end, US flag at his back: Robert Imig at a standup desk, six years in as Palantir tech lead for Maven when Manson meets him. Colleagues call him a unicorn: humble, intelligent, both business and tech. Pretty much the man who birthed America's first bespoke AI-enabled warfare platform. First job out of school 2012; Bay Area Cisco intern; bar night, sister, a guy going off about jumping out of planes with servers; looked up the company, applied, got in. National security was hazy mystique, not driving concern. 2018 commercial side: business apps, no parachute. June 2018: Anand Gupta, fellow Palantirian he had not met, approached talking fast. Hyper crazy; Silicon Valley through and through; imagined growing up on a bean bag under mom's Menlo Park desk (she dropped an AI PhD at Vanderbilt late '80s for a startup). Gupta temporarily dropped Harvard for Peter Thiel's 2012 fellowship; terrible biotech startup at twenty; back to Harvard; 2015 Palantir intern with big ideas; three years later an excitable blur of drones, video, CV algorithms, and a DoD task. Project Maven. Imig: "Where do I sign up?"
Hot servers; can of worms; Darwinian race
By 2018 Maven was across Somalia, Afghanistan, Jordan, UAE, and more. They had an algorithm; they needed a permanent home. Past the Somalia shoehorn into existing systems: commission a custom Maven appliance (MAPP, or mini-MAPP) carrying GPUs to run models and deliver inferences for up to four drone video feeds at a time; plug AI outputs into platforms already used for combat so detections showed on screen. First try: Colin Carroll disregarded design advice from ECS contractor Katya Volkovska. Eight massive overheating GPUs melted the $2 million boxes. "You were right," he told her, interrupting her holiday. "They're on fire." Volkovska thought him smart, passionate, dumb, and stubborn at once; only time he conceded she was right about anything. They married September 2022. ECS remade the appliance bigger with Supermicro boards and Nvidia GPUs. Outputs hooked into image-processing systems the military already relied on. One: Minotaur, government-owned suite from Johns Hopkins APL with Alion software help; fused sensor feeds including drone video; AI detections as boundary boxes or dots for analysts and "screeners." Another: Novetta's AIDE, tools already widely used by the Air Force: frame-by-frame drone players, map overlays, patterns of life. Cukor wanted a step further: everything including AI inferences in a single platform that did more than display AI over video; a vehicle to bring intelligence into operations; a slick special-operator interface so combatants, not only intel analysts, would lean into AI. Right-click a white dot, get a coordinate, go shoot. That meant AI into the targeting workflow. "This is where we start opening up an entire can of worms." Money earmarked for intelligence not operations; overlapping software already existed; AI into intel, let alone ops, already controversial.
Ch14 flow.
National security already had options for a "common intelligence picture." NRO's FADE (Fusion Analysis and Development Effort, 2007): in-house, not commercial; patterns of life; more than two hundred intel feeds; CACI tools; mission planning and battlefield forensics rather than ops. Mapping tool MIST "was the only show in town": aggregated clues from millions of data points (shifting phone pings suggesting an operation). Cukor reached NRO; decided their development incompatible: government-owned software versus commercial own-operate-update; NRO remit intel alone versus combat; platform compatibility and permissions would take a year. Discussions "cordial"; others fed up; continuing tension. (NRO no comment.) "Every step along the way we were dealing with defense-industrial bullshit." Waiting a year was a nonstarter for Maven time. Creating a rival risked duplication: NRO feeds would sit at the heart of anything Maven analyzed. Cukor had wanted Google Earth for an operational platform folding in targeting; as Google's catastrophic dropout became clear, he called old friends at Palantir. "Hey, it's been a minute," to Jain around April 2018. "Could we meet?"
Karp then was a skeptic. Palantir had watched IBM's $4 billion Watson investment crash with schadenfreude. Mantra: no magic AI button. Some also worried about AI-curious rivals including Novetta/AIDE. Contracts were falling behind despite the 2016 Army court win: lawsuit scars; "a lot of scars, raw nerves," Cukor; hard-sell techniques that sparked outrage in a genteel Pentagon; "They were their own worst enemy." Jain knew the polarizing reputation; nobody could tether Karp. "Even five years after the lawsuit there was still a lot of overhang... The lawsuit upset a lot of people and sullied our name." Jock Padgett, Marine who led IT innovation at JSOC Fort Bragg: expensive; worry of not getting military data out of voracious Palantir systems; arrogant attitude in 2016-17 meetings. Karp billed himself as defending Western values, then trotted throwaway lines about death with little mention of victims or ROE tragedy. Padgett found him "chauvinistic"; for some SOF, too brazen, too trite, sidelined glamorizer: "Candidly you're kind of a sideline... We kind of have to. It was a slap in the face." Few planned to pay for Palantir as Afghanistan and Iraq wound down. By 2018: no Navy contracts; Marines contract ended; Air Force slice only; final SOCOM contracts expected to end. "When I started Maven most people thought Palantir was exiting the Pentagon," Cukor.
Imig's eye; Netflix player; Dam Neck; mopeds; Djibouti
Cukor's outreach could not have come at a better time. As Google outrage built, he flew to Palo Alto late April 2018 and briefed a small Palantir group. Three intense hours: barely a word in edgewise; no demo; vision for ten years; data sources; AI transport; combat workflows forever changed; tech companies into the heart of war and government; AI on every piece of intelligence, all around the world, all the time. Key: make AI relevant and easy. Hardware clunky; presentation distracting and often wrong. He expected AI itself to improve; knew transformer models were coming (Google 2017 paper; the "T" in ChatGPT November 2022). He never wanted AI for AI's sake; he wanted an interface operators would use: detections over video, on maps, with precise locations linking to coordinates and other data. A smart system. Not Gotham repurposed; a bespoke UI from the start. Send a white paper. Palantir smarted: data-management experts, not shiny-interface producers. Early-May white paper pitched both. Small prototype contract; otherwise build essentially for free, another Mavenite said. Rivals also tapped: Anduril, Novetta/AIDE, JHU APL/Minotaur; six months each. "Natural Darwinian process," Cukor claimed, while giving special attention to Palantir from the start. Proud of the outcome: "You don't hear about AIDE or Minotaur except in small corners of the Navy anymore. It's just Palantir, Palantir, Palantir." Anduril fell away (more hardware than software then). Palantir leaders later said they spent almost half their time furiously building what Cukor wanted.
Ch14 hub.
Cukor counseled Jain on a high-end apology tour (Jain: listening tour) to soften national-security relationships; mentored him toward collaboration and apologies "for things that were probably more aggressive" than Jain thought needed. "'Not everybody loves you'... To be on the team you need to work on this. We love you, you're brilliant; but stop being assholes. This is a no-fail for our country." Jain moved to DC; about half his time on Maven; felt like a startup again. Cukor initially played down the company name inside the Pentagon; asked logos off presentations; teams switched fonts. Branding crept back: tiny corner logo; "Maven Palantir" or "Palantir Maven" slipped into titles. Cukor would glower mid-briefing. Jain bristled later: "We're expensive compared to what?... Our shit works." SOF already used Gaia and Gotham; never integrated CV detections; unlike Minotaur and Novetta, no video player at all. Six-month pilot: Imig had to build a player from scratch. He turned to Netflix, YouTube, Hulu streaming protocols. Several Mavenites strongly disagreed: streaming plays forward; analysts need frame-by-frame rewind, bookmark, annotate, and real-time display, not movies on demand. Mid-build: eyelash feeling in left eye; not even a scratch; could not shake it; cancer diagnosis quick; surgery September; today two light-blue eyes, one false. Irony of building a video player with one good eye. Shipped; October face-off in the Maven Integration Lab at JHU APL: Minotaur one side, Palantir player the other; fifty-plus people "on blast"; Cukor barking: why can't Microsoft detect this vehicle; why is the feed a second late. Pressure-cooker "one of Cukor's secrets," Imig. Bit by bit improved. Direct user contact; brutal feedback; trust; engineers and FSRs worldwide improving software by the day versus ten-year military horizons.
Carroll arranged global spread but was down on Palantir: overpay for fancy duplicative platform distracting from AI. Cukor sent a Navy SEAL reservist under Carroll instead. Trident club: nine years active, eight deployments; of two minds on Palantir; knew high-command SEALs; got hardware into locations even when units tried to throw him out ("My boss outranks your boss"). Biggest breakthrough: SEAL Team 6 / DEVGRU. October 2018: Palantir server on the rack at Dam Neck, Virginia Beach. If DEVGRU tried Palantir and got fed up, game over. All-night updates still not enough; one SEAL tried to eject hardware. Users tested historic data on slant count: men, women, children at a location; central to collateral damage assessments and whether to strike or raid. Single AI mistake: dead civilian, ended SEAL career, maybe a crime; end of Maven trust. "Giving AI literally the most sensitive task was probably the world's worst idea... I don't want to go to jail." Switched to lower stakes: tracking mopeds. Analysts had followed the USS Cole bombing mastermind for months on screens before capture. AI could maintain custody, reacquire after brief out-of-frame, potentially check human bias, a former senior JSOC member said. Limits on the Virginia coast: needed real raids, real data, real time. A senior SEAL commander took a risk: live operations. Primitive Maven Smart System bound for Djibouti.
Manson had visited Djibouti 2015: biggest active US military construction project; China planning first overseas base at the Red Sea / Gulf of Aden choke point; ambassador said "at the forefront" of US national security; jihadi threat growing after a restaurant suicide bombing. "Hot hell box in the armpit of Africa" versus "Somalia meets Las Vegas." Camp Lemonnier along the airport: Combat Cafe, Oasis Movie Theater, ping-pong, Xbox; "have a Djiboutiful day." Secret side behind a fence most of five thousand never entered: Camp Titan, special operations across the continent against al-Shabaab, AQ affiliates, ISIS. As Afghanistan wound down, Djibouti became the active center of "g-wot." JSOC operators started using the prototype on drone ops Libya to Somalia. Feedback blunt: sucked. Spring 2019: Palantir made sixty changes; users still complained of upgrade backlog; many turned it off and went back to NRO's intel suite. Not mature enough then for detections as strike confirmation, the former JSOC senior said. Then improvement: tracking target vehicles, especially mopeds, so analysts could look away; reduced strike-prep time. By mid-2019, Maven Smart System provided an operations picture in every war zone where the US was fighting. "We had the entire JSOC community," said one person. The Chewbacca in Georgetown and the melted GPUs in the first box and the false blue eye behind Imig's standup desk belong to one industrial arc: Cukor forced AI priors on a company the Pentagon was pushing out; the company returned with an interface operators would actually click.
Apology tour, fonts off, Darwinian preferred winner
Cukor's mentorship of Jain was not soft HR. It was a condition of remaining on the team that would rebuild Palantir's Pentagon life. Jain moved to Washington; about half his time went to Maven; he felt as though he were working for a startup again after years of platform maturity. Inside the building Cukor initially buried the brand: logos off slides; company fonts swapped; briefers warned not to say the name. Palantir could not help itself. Tiny logos crept into corners. Pitch titles became "Maven Palantir" or "Palantir Maven." Mid-briefing glower from Cukor was the correction. The Darwinian-race claim (six months; Anduril, Novetta, JHU APL, Palantir) sat beside the fact that he had given Palantir the longest first briefing and the freest labor ask. "You don't hear about AIDE or Minotaur except in small corners of the Navy anymore," he said later with pride. That sentence is both product history and preferred-winner confession. Jain's later bristling ("We're expensive compared to what?... Our shit works") is the commercial counter to Padgett's 2016-17 attitude complaint and to Carroll's fear of overpaying for a fancy duplicative platform that distracted from AI itself.
Ch14 compare.
The Dam Neck slant-count scare and the Djibouti "sucked" spring are the same story from opposite coasts. Give AI the most sensitive collateral task and you risk ending trust forever; give it mopeds and custody-maintenance and operators look away long enough to feel the time save. Sixty Palantir changes in spring 2019 still left an upgrade backlog; many users flipped back to NRO's suite. Mid-year the operations picture covered every US war zone anyway. Interface won before the models were perfect. That is the chapter's industrial lesson and its moral risk. Interface won before the models were perfect; slant-count fear and sixty Djibouti changes are the receipt that trust is earned in mopeds first, then asked to scale to kill chains.
Sixty changes and the entire JSOC community
Spring 2019 in Djibouti was the product's humiliation exam. Operators said the prototype sucked; Palantir pushed sixty changes; upgrade backlog persisted; many flipped back to NRO's suite. Maven was not mature enough for detections as strike confirmation, the former senior JSOC member told Manson. Then the moped lesson stuck: custody maintenance and reacquisition after brief out-of-frame saved analyst hours; strike-prep time fell. By mid-year the operations picture covered every US war zone. "We had the entire JSOC community." That sentence is industrial victory and moral foreshadowing. The same interface that earned Dam Neck trust by abandoning slant count would later sit one click from effectors on Whitworth's Palantir stage. Cukor's can of worms was never only about intel money funding ops tools; it was about making operators willing to live inside a commercial platform the Army had once sued. Jain's "Our shit works" and Padgett's "slap in the face" are both true in the same building: the product scaled; the culture debt remained.
Ch14 cycle.
Chewbacca in Georgetown is kitsch; Imig's false eye and Dam Neck's slant-count fear are not. The product that remade Palantir's Pentagon life was built under pressure-cooker bark and sixty Djibouti changes, not Darwinian neutrality.
The Georgetown Chewbacca and the Dam Neck slant-count scare bookend the same industrial bet: make operators click a commercial interface, and the company once exiting the Pentagon becomes the company that owns the kill-chain screen.
15. Palantir Splits the Team
June 2019, Bagram: Cukor pitches Donahue a hundred slides. "You have a dumb system... I want to give you a smart system." Three JOC screens; AI on all of them. Duplicative Palantir boxes beside Gotham/Gaia; flickering detections; "Object Unknown." Carroll's plane email: sold out to "the elixir of Palantir." "The program as we built it will never recover once we hand the keys to Palantir." Green liaison Barchick; half-day-behind sunny video at 2 a.m.; network crash during football; Helmand ambush where MSS saved Marines; Sy vs Carroll; midnight July 4 "failed startup" email.
Bagram pitch; three screens; keys to Palantir
When Brig. Gen. Christopher Donahue got to Bagram 2019, US forces likened arrival to a hostile corporate takeover. Army one-star incoming for US and NATO special operations in Afghanistan; presence winding down; chance for hi-tech everything; prepared to give AI a shot. Cukor visited June 2019 with usual pitch through a hundred-plus slides: "You have a dumb system... I want to give you a smart system." Ambition for what would become Maven Smart System spoke to twin issues that would tear Maven apart: Palantir and personalities. Donahue told Manson he was all-in from the start; ground team screaming for it; also gave it a tough ride: Maven needed to fail and frustrate to get better. Took nearly twenty years to get night vision right; now obsolete in his view; pushing AR headsets. Before returning to US, Cukor visited Bagram JOC: three live-action screens. "I want our AI on all those screens." Matt Barchick, newly arrived liaison, gulped: "Oh shoot." Getting MSS up would be first major test; going all-in on Palantir would also break the team.
Ch15 cycle.
Ch15 compare.
Ch15 ladder.
Detractors dumbfounded by duplicative cost: new Palantir Maven server beside old Gotham/Gaia servers installed two years earlier; totally separate accounts; under license no live data between them. Cukor wanted total separation so experimental AI could not mess up live ops; software updates then required downtime live operators would not accept; ten years' historic ops data integrated. Special operators rolled eyes at DoD shelling out twice on expensive boxes side by side. Palantir official: Pentagon buying services like a cake in layers. (Cakes usually sold whole or in slices.) Others figured Cukor backed Palantir so users would not complain about bad AI to the Hill and risk money; saw platform as fancy COP not AI. Overlay: AR painting brightly colored mission-planning lines (safe/dangerous routes); concentric range rings for collateral casualty areas; numbered buildings easing multi-team comms. Units up to commanders loved it. Decision made life harder for founding Joe Larson, Cukor's right hand: nearly five years at Palantir before Maven; shielded from Palantir defense work for conflicts. Palantir was ECS subcontractor; as they butted, Palantir soon made prime equal to ECS; as role grew, Larson blocked from more work he loved. Team jokingly called Palantir Tirpay (Pig Latin) to limit his exposure. Cukor: couldn't be helped; users needed appealing AI access. Carroll thought Palantir interface had little to do with AI and was not working well; another: 2019 "totally unusable." Every second ≥fifteen still frames; AI different per frame → constantly flickering detections too distracting; unrecognized objects lumped as "Object Unknown." Thumbs-up/down feedback not live; bundle and send to vendor for frustratingly manual retraining.
Ch15 close.
Ch15 scale
Carroll's answer: sidestep users. Not a smart system; a smart sensor. AI onto sensor platforms of weapons themselves, onto the MQ-9 in the sky; save processing bandwidth; no staring. Cukor greenlit experimenting; Maven line of effort Edge to put AI on flying surveillance and weapons rather than ground processing. Not enough to sate Carroll. 21 May 2019, landed after 2:30 a.m., private email to Cukor: exhausted on plane waiting to deplane; moving off reservist orders July; expected to stay as subcontractor or move to JAIC; apologizing; "struggling" stay or go. Enjoyed two years; potential for ML/CV for intel; team taken DoD a few years into future. But past six months program "lost its way"; vision changed; "not the vision I think DoD needs for AI." Sold out to "propaganda"; focusing too much "on the elixir of Palantir." Palantir ruined DoD business around 2012 with "terrible business practices"; committing now would haunt. Not in company's DNA to integrate; focused on data fusion not AI; business-development machine. "The program as we built it will never recover once we hand the keys to Palantir." Cukor made no concession; surface unfazed; encouraged freewheeling feedback. Purists exist; Maven not AI for AI's sake; DoD should bring commercial platforms (acquisition rules said as much); material impact on daily workflows; practice makes perfect. Medium-altitude drone answers: context first, heat out. Loved Colin; could live with different opinions. Sometimes: "Colin gets in his own way... so aggressive he doesn't have the world view, sometimes you have to slow down for other people." Eventually, as AI improved, single platform would fuse CV with reasoning models: not just "there's a tank" but what it's doing, next, fuel. Long before ChatGPT-3, Maven put IBM, Northrop, others on pilots for reasoning models. Tomorrow. This was today.
Green liaison; crap detections; Friday Georgetown; outage; Helmand
Barchick scrambling MSS onto Bagram screens: users thought Maven and AI cloud-cuckoo land. "They hated it initially." Marine reservist; never deployed; disliked traveling; weekend-a-month dues; knew nothing about AI, SOF, or Afghanistan. "Manson was just so green... me and AI doesn't make sense." Joined after fifteen-minute Carroll interview; second day mid-April 2019 told he was going to Afghanistan; thought new-guy prank; within weeks there. "Manson was like, what?" Detections with new Palantir even more cumbersome: dots so big they covered objects, impossible to distinguish adults/children; crowded market impossible to tell food purchase from weapon stash. "This is crap, this does nothing for me," analyst at smaller site. Barchick coaxed: not great yet; "isn't about you... This is about great power competition." If users did not tell how to suck less now, too hard to catch China later: "We don't have the manpower." Less compelling than inducement: twenty $6,000 gaming-style laptops at Bagram JOC for Palantir feeds. Palantir also feuding with Novetta; both reps at Bagram; contradictory don't-use-each-other's-player instructions. "What the hell am I supposed to be using dude?" "Whatever makes your job easier." Platform good; video player still crummy; delays fine for sports not real-time strike. Once froze half a day behind: sunny-day drone footage while JOC crowded at 2 a.m. pitch black. Cukor to Palantir: fix now. Imig calls from overseas 2 or 3 a.m. "I got beat up for probably six months. Just yelled at, constantly." Cukor preferred in person: every Friday leave Pentagon for Georgetown Palantir; 8 p.m. to past 10. "Cukor would beat me up about like two seconds of the live feed," Imig.
Ch15 flow.
12 August: milestone. Undersecretary Joe Kernan visiting Donahue; Barchick delivered: MSS up on JOC screen for real operations; AI detections over live mission footage; resulting strike judged success. First time Pentagon leadership could see what might be possible. Fleeting. Late September 2019: SOF preparing kill/capture of high-value AQ target so sensitive multiple US remotes following; network crashed; no internet, radio chat, video. "Everything just freezes." Sunday night; personnel settling for football back home. Army captain for C2 stormed TV room: everybody turn off Maven; imperiling high-priority mission. Never quite clear what caused outage; accusations flew at Maven, AI, Palantir, others. "We kind of crashed it," one observer. Barchick: nothing to do with AI/Maven; number of remotes. Many still believed Maven caused it. Jain: blamed many times over twenty-year GWOT; "almost always watching Monday night or Sunday night football or there's a spike in usage." Conceded FMV georectification uses more bandwidth; company optimized; problem showed reliance on old decrepit network tech.
Another time MSS helped a mission that went wrong. Night raid Helmand compound; courtyard empty; raid on; more than half a dozen Marine SOF bunched against outside wall ready to breach; explosion from behind wall; booby trap; ambush; wall collapsed; dust; immediate fire from bunker. Hail of bullets; no one could pick attack origin from dust cloud. Radio: troops in contact; request CAS. MSS not approved method then; intel analysts on drone feeds already saw AI detections marking US forces; boundary boxes on Marines running from smoke; counting each confirmed none trapped under collapsed wall; AR layer grids off video in seconds (usually minutes manually) so incoming strike hit bunker not own men. A-10 "eliminated" bunkers. After that close shave Marines would not stop using Maven. Soon Air Force colonel running MQ-9s convinced; wanted Maven mandatory on all feeds.
July 4 email; Whipps; MSS vs smart sensor
Too late for Carroll to see or believe. Not only Palantir objections. Yearned for early scrappy days. "The big family started to fight," Ward. "Egos were bumping into each other." Increasingly irked by Sy Poggemeyer, younger captain Cukor considered "just amazing": hardworking, impassioned on systemic data; also flared; some teammates and ECS contractors felt insulted. 24 May email to Cukor (two others copied): "I have had it with Sy and how he treats people on our own team"... ECS week before; JHU (Carroll about to join) week before that. Refuse to work with him as data lead. "It's come down to this, its [sic] me, or its [sic] him." Fumed Poggemeyer (now data lead) seemed to think he knew more than a PhD and four vendors "who literally do this for a living." PhD was Gene Whipps, Army Research Lab: struck by how "fast-paced" Maven was; overall positive; extended second final year. Carroll also dating Katya Volkovska at ECS; wanted job outside government for conflict. Night a little after midnight 4 July 2019, fury boiled: subject "Some Program Thoughts" to Cukor and one other; Maven a "failed startup"; worked butt off; team "rapidly falling apart"; legacy "will be my legacy as well"; expanded into too many things; erroneously brought Palantir "on a false promise" platforms ready for licensing and quick ops; handcuffed to sole-source delivery from company "motivated by money alone"; Palantir didn't view/sell/mention as Maven even at Air Force intel pitch; not doing hard AI work, just UI. Missed old days. Engineering lead Krishnan Aiyer that week: "a lot less 'Maven-y' around here." Maybe wrong, issues past; clutch already leaving; more expected. Finale: "I am sure there are others waiting for the dominoes to start falling before they make their moves, as no one wants to feel like they are on a sinking ship." He pressed send.
Ch15 hub.
Friday Georgetown beatings were not theater. Two seconds of live feed were the unit of critique because operators lived in those two seconds. Kernan's August visit converted a hated tool into a leadership demo; the September network crash converted the demo into a liability story that Jain's football-spike defense never fully retired. The Helmand ambush, by contrast, converted liability into loyalty: Marines who would not stop using Maven after the A-10 strike had lived the difference between minutes of manual grid work and seconds of AR overlay. Carroll's July 4 email closed the personality ledger the Tirpay joke had opened.
Larson's conflict shield and the Tirpay code were the office comedy version of a serious acquisition shift: Palantir from ECS sub to equal prime. Carroll's Edge line of effort and Cukor's reasoning-model pilots (IBM, Northrop) show both men were already arguing about tomorrow's stack while Bagram users hated today's dots.
16. A Striking Operation
October 2019, northwest Syria: Baghdadi raid Trump called "impeccable." Movie-tech praise; fraction was Maven. Erbil screens; AI pings a moving van before human screeners; commander: "Someone just made their company a lot of money." Barakat Ahmad Barakat and two cousins: olive-press civilians; arm blown off; friends dead. US acknowledges 1,437 unintentional civilian deaths vs ISIS by end-October 2019; Airwars higher. McConville's white-van hesitation; Bryant: AI's best function may be doubt, not certainty. Ping is not permission.
Ch16 compare.
Movie tech; three years; Barisha compound
Trump next morning: mission to kill or capture Abu Bakr al-Baghdadi, "the world's No. 1 terrorist leader," top national security priority. "Impeccable" two-hour nighttime raid on hideout northwest Syria October 2019: SOF on ground, eyes in sky, massive airpower, ships, shootout; target at end of concealed tunnel detonating own vest. "He died like a dog." Nearly an hour with reporters; staggered; "unbelievable success"; "incredible"; "grand style." Before end Baghdadi fled into dead-end tunnel with two children and exploded; task force dug him out; dog injured; on-spot DNA; "Jackpot." Would not say how he watched live; like a movie: "The technology there alone is really great." At least a fraction: Project Maven. Cukor would not address Baghdadi or other ops; had said SOF especially sensitive units "used this thing extensively, all day long, working aggressively... They really were our biggest champion." Culmination of three years. Baghdadi founded ISIS; caliphate June 2014 over millions in Syria and Iraq; held at Abu Ghraib and Camp Bucca about ten years before declaring caliphate; public beheadings; POW killings; pledges from Burkina Faso to Philippines; Yazidi massacre 2014 and three years of killing/kidnapping. After 2017 fall of Mosul and Raqqa, US and others destroyed remaining caliphate; by spring 2019 barely anything left; Baghdadi eluded Obama and Trump ("Where's al-Baghdadi?"). Earlier 2019 tracked to Idlib; six months later exact whereabouts. Erbil SOF task-force HQ: maps and video ready; "completely confident" no other ISIS nearby; expected other armed groups. Eve: good weather; slightest moon sliver; Trump green light. What weighed most? "Well, just death." Holed up four miles from Turkish border; compound western outskirts of Barisha; wall; mostly dead-end tunnels underneath. Team knew children at compound. "Very, very dangerous mission." After dark Saturday: eight armed helicopters low and fast over an hour; two gunship volleys to frighten civilians indoors. Eyes mostly on compound; drones, satellites, manned aircraft watching for rescue. "Basically any bad guys that were gonna advance on our guys." IR cameras scanning; video and AI identify threats only if camera already looking. Two of nearly 35,000 air strikes against ISIS during raid: local gunfire from multistory near T-junction east; grainy obliteration of people on foot. "That gunfire was immediately terminated. These people are amazing," Trump. Friendly roadblock east. After landing: ≥ten Delta approached from southwest through olive grove; blasted wall avoiding booby-trapped door; "beautiful, big hole"; "Then all hell broke loose." Arabic surrender call; eleven children and two men detained; four women and a man killed; Baghdadi still inside.
Baghdadi footnote.
Ch16 close.
Ch16 scale
Ch16 cycle.
Ch16 ladder.
Maven ping; van; "someone made their company money"
Erbil screens a mass of information. Request for MSS at Erbil end-May; five months later put to work. Report of vehicle on the move; drone slewed; as soon as it did, Maven AI suddenly pinged; telltale detection. Command-row guy clocked at once: Oh look at that dot. Pinpointing moving vehicle; screeners glued to multiple feeds had not spotted yet. AI beat human on high-priority mission. Detection on map too; moving south through Barisha; T-junction turn west toward raid site ~1,000 feet away. Commander put it together: driving toward task force at the building. Not thought ISIS; other militants; vehicle displayed what Gen. McKenzie later called "hostile intent." Radio to ground force commander; hit called; thirty seconds later attack helicopter; gunfire; man ran from van to north of roofless roadside building; strafing blew north and west walls; five minutes later rocket; "And the van is gone." Shot-up white-panel carcass beside olive trees. First time commander used AI to help find a target; most consequential raid; speed left impression with JSOC. "Someone just made their company a lot of money." Task force and others leaning wholesale into Maven. What if civilians inside or near? "Those were definitely bad guys that the AI ended up picking up," feed-watcher.
Ch16 flow.
Barakat's cousins; investigations; doubt not certainty
Barakat Ahmad Barakat told NPR different story: driving home to Hatan west of compound after olive-press work in north Barisha with two friends. "Suddenly I felt something hit us." Van halted; men fled; friends (cousins Khaled Mustafa Qurmo and Khaled Abdel Majid Qurmo) died in aerial attack. Cradled one on roadside as airstrike arrived; blew off right hand; later lost whole arm; left hand weak. "Am I Baghdadi? How is this my fault? Manson is just a civilian. I didn't have any weapons... We're farmers. I make less than a dollar a day. Now Manson is handicapped, and my two friends are in their graves." US acknowledges 1,437 civilians unintentionally killed in five years vs ISIS in Iraq/Syria to end-October 2019. Per Barakat and NPR at least two higher. (Airwars: possibly higher than 8,000.) US investigated twice; first did not uphold; after NPR sued for redacted version and questioned conclusions, second investigation; results not public.
DoD says it protects civilians because LOAC requires and "because it is the moral and ethical thing to do." AI's role open. Accusations for years of insufficient prevention, acknowledgment, investigation. 2018 Congress required annual civilian-casualty report; still complaints of significant undercount; "more likely than not" standard; often no ground check. 2024 GAO cited media: DoD mistaking civilians for combatants. Feed-watcher on investigation: commander's strike decision right even if later civilians; "In the heat of a mission, all the task force commander and ground force commander sees is anything moving is a potential threat... In the moment, that was a good strike." AI not only source on van; no role in strike decision: "That's a human decision; not an AI decision." AI functioned as intended: correctly and quickly identified van; more reaction time including assessment/warning. Even so, fear of AI wrong inhibited usage. "A lot of folks in the Pentagon were still very reticent to use AI because they were worried about indemnity, especially when it comes to strikes," another Mavenite. Gen. Jim McConville (Army chief until 2023): as younger man helicopter-dispatched to shoot white van accused of firing at Iraq base; weapons aimed; hesitated; brought helo in front; family got out. Had he fired, civilians dead. Years later worried AI has no morals: "It doesn't really think, it doesn't feel... who is responsible when AI does the targeting, they do the shooting, and you put a missile into a van with a family? Who owns that?" Successful ID part of broader problem: "AI only sees what it sees; it doesn't see good guy, bad guy... Not even man, woman, child." Ping in high-adrenaline raid could incentivize action; "I see a van" might make reaction likelier; death warrant. Person argued US always confirmed via separate source; AI added awareness. "Most of the strikes that we saw where mistakes happened, AI was not involved at all... a lot of people who were trying to throw AI under the bus." Maven team reacted: Was system involved? On? Outputs? "Most of the time when it was turned on, we got more data than they would've gotten otherwise." Shanahan weeks before Baghdadi: SOCOM/JSOC almost "pounding the table" for AI; "massive increase in situational awareness"; faster; mitigate human mistakes; limits: "Manson is never going to say that AI is going to eliminate friction, chaos, the fog of war. Never." Narrow act: physical presence of a van. True-target info (who, why, intent, armed, where going/been) lies elsewhere. Seeing and understanding big picture rests on more than object detection.
Wes Bryant, former special warfare operator who led strike cells; until spring 2025 nine months at DoD Civilian Protection Center of Excellence as branch chief of civilian harm assessments; hours at CENTCOM and SOCOM. Rare for US to drop when anticipated collateral high; deliberate misdemeanors rare; more often faulty assessment of civilian numbers or misidentified target. US "horrible at response"; failed to investigate and acknowledge. December 2021 Times: far greater civilian casualties in Syria/Afghanistan than admitted. Precision not the main modern problem: Desert Storm ~90% unguided misses; guided weapons now pinpoint; modern errors from faulty context, ignorance, cognitive bias. Official blunders: February 1991 Baghdad bunker 400+ civilians (converted air-raid shelter; US said unknown); May 1999 Chinese embassy Belgrade five two-ton GPS bombs (intended Serb weapons office 200 yards away; maps failed to update three-year move; missed phone calls); 2015 MSF hospital Afghanistan 42 killed (faulty coordinates "middle of this field," then guesswork); wedding parties and harvest farmers. Bryant: AI might help patterns of life. After August 2021 withdrawal, US admitted "tragic mistake" on white car tracked eight hours (ISIS explosives suspected); killed aid worker with water canisters plus seven children and two others; no ISIS link. Maven not involved; Bryant: AI might flash useful fact, remind overwhelming percentage of Afghan cars are white Toyota Corollas, "This one might not be the right one." Best possible function of AI in warfare: not introduce certainty, but doubt.
Trump's movie-tech praise and Cukor's refusal to address the raid sit beside each other: the book can name Maven as a fraction of the stack without an official kill-chain confession. The Erbil ping that beat human screeners is the product story Palantir and Maven wanted; Barakat cradling a cousin as the second strike arrived is the harm story the product story cannot erase. McConville's hovering helicopter facing a family is the analog of Bryant's Corolla reminder: the pause that saves lives is a human habit AI does not own. Identifying a van's physical presence remains a narrow act; intent, affiliation, and children in a tunnel remain elsewhere. Shanahan's fog-of-war refusal ("Never") is the official limit statement the Baghdadi footnote keeps testing. Desert Storm's ninety-percent misses, the Belgrade embassy, the MSF hospital, and the 2021 white-car "tragic mistake" are the mismatch history Bryant brings to the Civilian Protection Center: firepower outran insight. Digital platforms such as Maven Smart System might close that gap; they might also accelerate the wrong ping.
Congress's 2018 annual civilian-casualty requirement and the "more likely than not" undercount standard are the bureaucratic frame around Barakat's missing arm. The second unpublished investigation leaves the Baghdadi van in contested space: product success and civilian harm claims coexist without a public adjudication.
Bryant's nine months at the Civilian Protection Center and his CENTCOM/SOCOM visits frame the chapter's closing thesis: AI might flash the Corolla fact that saves a family. Ping is awareness; ownership of the missile remains the open McConville question.
Airwars totals above eight thousand and the unpublished second investigation leave the official 1,437 figure under contested pressure that the Maven product story cannot settle alone.
17. Data Hell
In the Maven dungeon, a vendor complains about training images. Brian Ward sees not tidy boxes but a freehand "F U C K" and a crude doodle: "It was a dick pic." "He poisoned the entire pipeline." Recalcitrant volunteer labelers, dirty data, interpolated ground, Apple buying Xnor, Scale and Palantir becoming primes, and a Shanahan-Cukor showdown over MQ-9 datasets turn pathfinding into two rival AI houses. Spirk, whose thigh was once tactically one with Cukor's in Kandahar, breaks the standoff. Carroll later: data lockup was "Maven's biggest failure."
Ch17 scale
Penis doodles; volunteer detention; poor quality
There was also such a thing as too much doubt. One day in the Maven dungeon people were already chortling. A vendor had called to complain about images: the central feedstock for war algorithms. At launch Cukor said publicly Maven's "immediate focus" was algorithms for thirty-eight object classes; the team would commission labeled images from varied angles, distances, weather, and sensors, then send them to AI vendors. Ward looked at the screen: freehand scrawl "F U C K"; over another frame a crude hand-drawn doodle. "It was a dick pic." Once the laughing stopped, the unfunniness landed. The cutting-edge AI brain was being trained to recognize errant appendages. Bad labels required more processing; muddled expensive teaching; less reliable, less fast; underperform or mislead in war; incorrect patterns → hallucinations, bias, drift. Humans getting in the way of artificial intelligence. "He poisoned the entire pipeline," Ward.
Ch17 close.
Ch17 ladder.
Rather than pay an expensive labeling company, Cukor first kept costs down with military labor. Service members encouraged to open Figure Eight (formerly CrowdFlower) and click boxes on drone grabs. "Manson was convinced we could have millions do this," of DoD's two-million-strong workforce. "Man, I tried that so hard." Rear Adm. Tim Szymanski (Naval Special Warfare; first opened SEAL data pipes) took pride; his intel chief's "monumental effort" mobilized NSW reserve intel to categorize: vehicle / car / Toyota Corolla. Maven gamified scoring; top labelers got points; sometimes a thanks letter and a Maven coin (cheery robot cartoon designed by a daughter). Whole team tasked to label so they would know how much it sucked. Ward: "It's like your brain melting through your eyes and every orifice that brain matter could come out." Cukor preached US distinction from China's centralized labeling (volume via forced labor, he said). To some, labeling felt like punishment or after-school detention: more mistakes. Training one algorithm could need 10,000 images; every wrong label made it worse. "We probably underestimated in early Maven days just how bad data quality could get," Jane Pinelis (Maven T&E director 2019 into 2020), 2023 interview: trouble getting great performance; hampered by "dirty data"; "learned the hard way." Another: "We fucked up data so bad." Google complained from the outset about "poor" quality labeling (customer engagement updates October-November 2018); smaller high-quality batches better than low-quality high volume; CrowdFlower improved by mid-November; still problems with "ground-truth data." Gig-worker contracting among first professionalization steps; did not go much better.
Ethics theater; Zeiler cleanup; synthetic failure; Xnor to Apple
Meanwhile DoD tried to assure an alarmed public. After Shanahan's year-plus listening tour: February 2020 announcement that combat AI would be "responsible, equitable, traceable, reliable and governable." Unclear how to extract those from dirty data plus black-box AI with principles that were little more than aspirations. Shanahan conceded ethical questions at Maven "really did not rise to the surface" every day because, he said, no weapons were involved. Carroll scorned AI ethics push: those who can, do AI; those who can't, do AI safety. Matt Zeiler (Clarifai) could do AI; so unimpressed with quality he automated label cleanup. Still messy: weeks of thumb-twiddling while formats changed; shortcuts. Each second of FMV = thirty static frames; professionals might label every tenth; some furious to learn annotators also "labeled" the nine between by averaging and interpolating a straight line. If person or vehicle swerved, Maven was paying millions to train algorithms to recognize the ground. Zeiler wanted Clarifai's own labeling platform; government refused: Cukor wanted a single labeling solution for all vendors, not one; worried about vendor lock or losing control of the trove. Clarifai took advisers (former Hill staff, acquisition officials, a three-star); all counselled avoiding the word "platform." DoD allergic to all eggs in one basket. A technical expert: platform eases retrain, narrows cutting-edge research space.
Ch17 flow.
Early days: Cukor tried synthetic data for rare adversary-equipment images. 2017 Microsoft meeting; crack team led by Dimitrios Lymberopoulos. Could not generate synthetic well enough for such specific objects ("see through it," Cukor); failure came with unlikely success: weakness-revealing models so impressive he asked Microsoft to join as algorithm vendors. For Microsoft a huge business door: contract small, but Maven as accrediting sponsor for Azure; AI warfare as way into DoD big cloud. Several Mavenites: best algorithm producer was Xnor (took Google's WAMI drone work; exemplary FMV research; accuracy/speed balance). Apple bought them early 2020: another Silicon Valley wall; Apple refused Maven work; Xnor algorithms wasted away. Microsoft soon among Maven's best remaining vendors.
Aberdeen call; The Reboot; trophies; Scale
Microsoft pumped out WAMI and FMV models and more than doubled its team to fifteen; by end-2019 even their performance plateaued. "Garbage in, garbage out," Cukor. Lab tests on similar clips looked okay; field was far worse. He named the stage data hell.
Ch17 hub.
Early 2020: Maven team at Aberdeen Proving Ground (WWII bunkers; couscous-vat robot walk Manson saw 2018; exoskeleton; bomb vault as dark prison). Offsite for blimp algorithmic systems; rare freedom without the boss: Cukor supposedly on a plane to Korea. Leaving Aberdeen when he called: "Where are you?" "Hey boss, Manson thought you were in Korea." "Get back to the Pentagon. We gotta redo things." Algorithms no longer getting better. Effort known as The Reboot. Within days more than fifty people in nonstop brainstorming at ECS Maven Integration Lab, Fairfax. Vendors flew in with more than a hundred suggestions. Main ask: more time. "American taxpayers don't pay you to rest." They could not field new models onto combat hardware worldwide as fast as he wanted; logistics ate research; teams exhausted and demoralized. He yielded a little: more time. Second ask: better data. Analysis showed each labeling error pulled performance down. Could not rely on cheap grunt labeling. Vendors wanted entire military-created training set of more than a million images relabeled. "We ended up not using any of them," Cukor. Third: clearer goalposts. Earliest tests judged by eye; subjective; startups axed on unclear metrics. Needed "a very defensible way," Pinelis. Horse race in secrecy: code names; hidden evaluation sets and scores; prevent winning the test rather than the war. "It's really hard to completely and thoroughly evaluate these AI-enabled systems... they frequently arrive... as a black box."
Lymberopoulos (Microsoft principal research manager; built and led Maven team) felt evaluators underscoring his models: strong on Microsoft internals, weak on Maven tests. Competitive about everything; enchanted by drone-video rareness, not gore. "I couldn't see anything like this on the open internet." Cukor coaxed or shamed: "You guys are from one of the best companies... spending three times more... why are you behind?" Months after Microsoft complained, Maven changed evaluations; scores moved up. Mark: F1 score (harmonic mean of precision and recall). Test event 2021: ECS awarded Lymberopoulos a two-foot "MaVeN!" trophy with angel wings and hologram stickers beside smirking Clarifai rival Daniel Marasco. "Please don't... It's unnecessary." (Left the hulking trinket at ECS.) Realer divide: Microsoft better at tiny few-pixel people, slow (fewer than thirty frames/sec; only "person" and "vehicle"); Clarifai tried four times as many frames; faster than most human reaction; better for tracking movers. Dull differences that might one day become the basis for shooting someone. AWS tried a third tack: Clarifai and Microsoft submitted a handful; AWS more than sixty, four people said; rivals accused gaming. Cukor: models adapted to circumstances (fading light); orchestration to switch not there yet; imposed max five models per vendor thereafter.
Months past The Reboot, things looking up. Cukor met twenty-three-year-old dropout Alexandr Wang; Scale AI "tricks" to improve and automate labeling. Palantir data-management platform organized labeled data and revealed over-reliance on one angle, distance, or time of day. September 2020: Scale and Palantir new primes alongside ECS. Models "substantially" better. Labeling big business: Wang, Covid roommate of Sam Altman, world's youngest self-made billionaire next year; preaching to Congress on Chinese facial recognition and American labeling at Scale. June 2025: Meta employee after 49 percent stake for $15 billion. Fortunes with Maven almost immediately dove: September 2025 lost largest labeling contract on record, delayed seven-year bid up to $708 million for Maven and other intel programs, to minnow Enabled Intelligence. Founder Peter Kant: quality over quantity; often-overlooked autism-spectrum labor force better suited to repetitive pattern recognition that irked early volunteers. Technical performance "outstanding"; quadrupled workforce of thirty almost overnight. Even before then Marasco: Maven ultimately created impressive serious labeling with huge data and compute; four hundred professional labelers, eight-hour shifts year-round; "Every time we get labeling data that's not from Maven it would be really bad." Lymberopoulos (left Microsoft for Palantir 2022; hired Marasco): "I believe Maven has curated a huge dataset that has delivered the best general CV models out there." If data was the new oil: soon at least 100 million professionally labeled images.
JAIC data war; Spirk breaks it; two rival houses
When Shanahan left Maven to start the JAIC in 2018, he expected Maven would eventually come under JAIC control: minnow to whale; pathfinder to department-wide scale. JAIC would staff to hundreds with hundreds of millions aimed at removing roadblocks. August 2019: JAIC's biggest next-year effort "warfighting-focused" on AI for maneuver and fires. Lots of "bravado" about Maven reporting to JAIC, Cukor said; he argued it could never happen: Maven funded by Military Intelligence Program money, not the same pipes. Turf arrived with structure. Some wanted to fold Maven into the JAIC to hold Cukor back. "The only reason Maven is what it is, is Cukor," one defender. "You cannot constrain someone like that and get results." Congressional money mandates versus turf war. When Carroll left Maven July 2019 (Palantir embrace; dating ECS's Volkovska), he landed at JAIC with several old Maven teammates. Shanahan emailed next year he was "a force of nature"; "would have been screwed" without him. Carroll pursued an idea from Maven days: December 2018 Microsoft proposed putting the algorithm on the MQ-9 itself (Smart Sensor) instead of hauling WAMI to ground processing, cutting data link, transport, and intel analyst. Cukor uninterested: operational complexity; bled into ops versus intel funding lines. Carroll captivated; pursued under Maven then took to JAIC. 2023 autonomous sensor tried in Indo-Pacific; Marines using late 2025. Internal doc: "In the not-too-distant future, the Department will fight autonomously with human-machine and machine-machine teaming." Briefly Project Atropos (Carroll's name: Fate who cuts the thread of life; cut the human pilot out; he had bought the domain 2012; thirteen years later Atropos Group startup).
Ch17 compare.
To put AI on a drone, JAIC needed MQ-9 labeled variety. Maven had the feedstock. Cukor would not give it up. Lost workforce to JAIC; not the data. Though much declassified, he argued declassification on condition of safeguarding. Turned down former underling and former boss. Shanahan August 2019 to reporters: "We have to liberate data across the DoD." Nand Mulchandani, JAIC CTO, acknowledged rising tensions: Pentagon paying so much as early adopter should reap broader benefits. Cukor: shared multiple aspects; stopped short of data for leak risk. "Once you open datasets they go everywhere; they get legs... At the end of the day the data is more valuable than the algorithm, frankly." Exquisite labeled datasets on sensitive things; vendors only inside SUNet training environment; JAIC not in at first; no MQ-9 models either. "Things got ugly." "Cukor became enamored of the fiefdom he built and wouldn't share," a JAIC member during the 2020 dispute.
Longest, most contentious meetings: JAIC reasoned; Cukor polite, dogged, unyielding. Sometimes national security; sometimes Middle East data focus could expose DoD to "spying on Muslims" criticism; leaked graphic combat footage could reignite Google-scale furor and dent civil-military tech relations; national embarrassment. Spring 2020 meeting: Cukor suggested only he and team could keep material safe. For the first time in years of protecting Cukor, Shanahan ran out of patience. Genial three-star turned red: "What do you think we do for a living?! You don't think we can fucking protect this, Drew?" No one had seen Shanahan like that. Assumed loyalty after going to bat time after time; disappointed as a parent; shocked Cukor showed up ready for battle. Greg Christ looked to some as though he might cry. Cukor remained implacable. "Ultimately we prevailed, but at great cost and there were a lot of angry people." Shanahan stopped talking to him. "My biggest regret is not solving the Maven problem," Shanahan May 2020 email to Carroll; wished he had brought Maven and JAIC together at outset; "know Drew too well... Not sure if anything substantial will happen until he departs the seat." Manson asked if Cukor had become the office bureaucrat he always railed against. "I know that's how it looked... The irony is not lost on me." Not a concession: data the country's most precious national security asset; China could learn from it or poison it. JAIC member: risk always exists; ways to protect and track downloads; Carroll/JAIC would not have got that wrong. DoD audit less sure: June 2020 report: JAIC had not developed a security classification guide for AI data yet; flagged six cases of lax controls; malicious actors could steal information that "could threaten the safety of the warfighter."
JAIC took complaint above Cukor's head to chief data officer David Spirk: the man who carried Cukor's computer nineteen years earlier in Kandahar; whom Cukor counseled to leave the Marines and ascend as a civilian; whose thigh was once tactically one with Cukor's; now weighing data-sharing disputes. Considered carefully; broke the standoff mostly but not entirely for Maven. Silly flash point, he told Manson; decision independent of friendships. Nonpublic memo: JAIC access to most Maven data, but not precious expensive T&E datasets (sensitivity; also prevent vendor cheating with exam answers). JAIC had not offered him a safety plan. Decision pitted Carroll against Cukor again; threatened to impede drone-sensor project; forced double data collection, labeling, model creation, time. Carroll 2022 podcast: failure to share was "Maven's biggest failure." JAIC member: "The whole point of Maven was to become a pathfinder for AI but instead the data stayed locked up inside Maven and didn't spread across the department." Staffs of the two AI offices were not cooperating either. Pathfinder path wound around an island of its own making. When Shanahan retired that summer he was still raw and angry. Maven and the JAIC: two rival AI houses inside the Pentagon.
Pathfinder island
Shanahan's regret email and Spirk's memo did not reunite the houses. Staffs still failed to cooperate. If Maven was a pathfinder for department-wide AI, the path wound around an island of labeled images Cukor would not fully open. Carroll's podcast charge ("Maven's biggest failure") and the JAIC member's pathfinder indictment sit beside Cukor's China-poison fear and the June 2020 audit's lax-control flags. All can be partly true. The dungeon doodle and the 100 million labeled store are the same chapter: garbage in becomes fiefdom out when the feedstock is also the crown jewel.
Ch17 cycle.
The Reboot's MaVeN! trophy and Enabled Intelligence's autism-spectrum hire both answer the dungeon doodle: professionalize the feedstock or keep poisoning the pipeline. Spirk's memo split the difference; Shanahan's silence after the red-faced meeting did not.
Carroll bought the domain for what became Atropos Group in 2012; thirteen years later the autonomous-drone startup still needed MQ-9 imagery with enough angles, distances, times, and backdrops to put an AI brain on a drone. Lab scores lied: models that looked fine on sanitized clips similar to their training set failed harder in the field. Pinelis's point, as he told Manson, was that DoD systems often arrive as black boxes and are poorly understood. Spirk's fear was less China than vendors cheating by grabbing exam answers early. CrowdFlower improved labels for Google by mid-November, yet ground-truth problems remained; gig-worker contracting was an early professionalization step that still disappointed several Mavenites. One vendor was so unimpressed that he automated label cleanup for his own algorithms; Cukor blocked a single-vendor pipeline and insisted on one government labeling solution feeding every sprint. He also capped future sprints at a maximum of five models per vendor once orchestration to switch among bespoke models was not ready. On declassification, Cukor argued Maven had won release on condition they safeguard the data, and that sharing the corpus with JAIC risked leaks even if other Maven pieces moved. Aberdeen Proving Ground, World War II bunkers and experimental ranges on Maryland's eastern shore, is where an early-2020 visit crystallized his "data hell" phrase. The spat over access would fracture him from the boss who had given top cover and shake DoD AI efforts more broadly.
Cukor had waved off an earlier plan that bled into operations rather than intelligence alone, citing funding-line risk and operational complexity; Carroll stayed captivated, pursued the initiative under Maven, then carried it into JAIC. Slow models skipped frames: instead of scoring each of the thirty frames in a second of video they ran against far fewer. After Google's walkout and the labeling reboot, the industrial answer was Scale's professional pipeline and Enabled Intelligence's autism-spectrum hiring, against Scale's reported $708 million loss year, while the pathfinder still treated labeled feedstock as crown jewels. June 2020 audit language on lax controls and Carroll's later "biggest failure" charge sit on the counter-ledger beside China-poison rhetoric and embarrassment fears about "spying on Muslims." Data Hell, in Manson's telling, is doodles, interpolated ground, black-box models, and a fiefdom that would not open the exam.
18. We'll Find It and We'll Strike It
7 a.m., last Friday of August 2024, DC hotel: toy-soldier trolleys beside avocado toast. Across from Manson: Col. Joseph O'Callaghan, "the man everyone accuses of inventing Skynet." Artillery; Cukor's front man; self-styled "Director of Algorithmic Warfare." Kurilla's Big Eagle pinup on Maven walls. Soleimani track on Maven in Erbil; push AI detection to the gun line; first AI-derived tank target 27 August 2020; Scarlet Dragon; Temple converts during Kabul; Abbey Gate children AI could not see; 60% AI vs 84% human; "We'll find it and we'll strike it. The whole nine yards." Not Terminator. AI is now part of the kill chain.
Ch18 scale
Ch18 cycle.
Fantasy breakfast; Big Eagle; Soleimani footnote
Butted against the breakfast counter: a trolley. Looking for coffee; realizing not milk jugs but tiny battle figurines, bright hues, swords, cartoon furies molded into minute faces. Whole floor dotted with trolleys pushed by hunched men in baggy gray T-shirts. Toy soldier convention. Opposite: a real soldier, thirty years in, special operations, Iraq. "It's pretty weird we've landed at a fantasy battle convention," smile, avocado toast. "And here I am, the man everyone accuses of inventing Skynet." Joseph O'Callaghan: artillery specialty; howitzers and high-mobility rockets across Middle East deserts; battalion commander inside the 18th Airborne Corps Field Artillery Brigade; year planning operations in Erbil 2015; fourteen months running the targeting process against ISIS from Iraq and Kuwait as fire support officer, coordinating artillery, air strikes, and naval bombardments. The very targeting process Cukor had for years told his team to avoid mentioning even in internal Pentagon correspondence. October 2019 he arrived at the 18th at Fort Bragg, America's rapid reaction force, often first out the door (Panama, Bosnia, Iraq, Syria). "We'll probably be gone out the door soon... It is a rare year that we are home." Bag always packed; he quoted The Princess Bride for the cycle: "Good work. Sleep well. I'll most likely kill you in the morning." "That's what living in 18th Airborne's like."
Ch18 close.
Ch18 ladder.
His boss was Michael "Erik" Kurilla, who became the 18th's three-star commanding general the same month. Pretty soon a picture of Kurilla (previously 75th Ranger Regiment commander; former JSOC director of operations) was up on Maven's office walls. Some chortled at what they saw as Cukor's obsession with hard men treated as pinups by wannabe algorithmic warriors. Cukor said it was simply about remembering the customer. Customers did not live inside the world's then-largest office building; they were not ladder-climbing bureaucrats who got home on time; they risked their necks and their forces' necks again and again. Maven was not to be another anonymous tech effort doomed to fail; he wanted everyone to confront people who faced combat's horrors. "What are we doing for Kurilla today?" he would ask, tapping the picture. He's looking at you and he disagrees." Kurilla: West Point; best-lieutenant competitions in the 1980s; every single year from 2004 until 2014 leading conventional and special operations abroad, including six years in a row in Afghanistan. Risky idea of fighting by day rather than night cover: "from killing five guys to killing fifty guys," a Ranger told Sean Naylor for Relentless Strike (2015). Adm. William McRaven, former JSOC commander: "the finest soldier Manson has ever known." Embittered too. 2015 speech: American society pedestals the wrong people (sports stars, actors, singers). "Those aren't heroes." Known as Big Eagle to some and Gorilla to others; men who signed up returned the favor. "Colonel Kurilla is like my dad. He would die for me," a soldier under him at FOB Marez told blogger Michael Yon May 2005. Months later in Mosul he continued firing after a commando roll, a bullet in each leg and one in his arm. Phone to wife from combat support hospital: "Honey, there has been a little shooting here. I got hit and there was some minor soft tissue damage." Yon said the X-ray showed the femur nearly snapped in half. "I'll be fine. Just some minor stuff."
Maven made a "huge play" for the conventional military as former JSOC leaders like Donahue and Kurilla were promoted into the wider force. Making it into "the Big Army" could mean more take-up, more money, bigger impact; also greater challenges: conventional forces less interested in new tech than special operators; bound to Congress-funded programs. Kurilla came from SOF and was set on bringing new tech to the sprawling North Carolina base of 90,000 troops. "He takes command and he's like, 'Bring me that Maven,'" recalls Nick Villarruel. Kurilla wanted the 18th to be the Army's first AI-enabled corps. Far from clear what that meant. AI was a black box; he wanted to understand what he could trust it to do. Plug into everything? Into targeting? Video footage would not be so relevant for future wars against China or Russia; drones collecting video likely shot down or jammed; satellites and still imagery might have more luck. Open-minded across intelligence, operations, and networking; O'Callaghan said he and his fires team found "the most traction." "We were introduced to them in about February of 2020 coming out of the strike on Soleimani," O'Callaghan recalled during a 2023 office visit. Manson learned later from others that, just as with the Baghdadi raid, Maven had also been up on the screen in Erbil during the 2 January 2020 strike Trump ordered against Qasem Soleimani, commander of the IRGC's elite expeditionary Quds Force. The top-secret JSOC mission was kept so under wraps that Maven members only learned weeks afterward that algorithms developed under their own program were used to track the vehicle carrying Soleimani soon after he landed at Baghdad airport that evening, right up until after the US fired at his car. One person who watched the mission feed: AI made it simpler to follow the vehicle; Maven Smart System's augmented-reality layer highlighted the track, helping determine speed and direction. But O'Callaghan wanted more than find. He wanted to shoot. "We want to be able to push an AI detection all the way to the gun line and then shoot this, like live fire off of it," he told Maven's vendors. Aim: shorten the kill chain. Find. Fix. Finish. For Cukor, three years avoiding the word "strike" aloud, O'Callaghan's explicit fires focus released the constraint. "When we were primarily an intelligence function or capability I exercised extreme caution so that we would not bleed over into the operations side. The intel people cannot overreach... I never said fires or strike capability or single click." Cukor could let O'Callaghan represent ops, phased carefully. "Knows this dance as well." Within a few years O'Callaghan was mid-dance-floor; LinkedIn title backdated to 2019 arrival and stretched across six years: "Director of Algorithmic Warfare." War bunker below his Fort Bragg office dedicated to all things Maven; hallway to a test room daubed with fat-marker hieroglyphics sketching data and AI-assisted military strikes.
Maxar pixels; ACC filibuster; first tank; Scarlet Dragon
The idea was to use AI on imagery from America's classified spy satellites and unclassified commercial companies such as Maxar Technologies in low earth orbit. Villarruel had spent months puzzling algorithms onto those platforms; with deputy Jaim Coddington they tried existing computer-vision algorithms on still images of China and Russia from high-altitude U-2 spy planes, RQ-4 spy drones, and spy satellites. Challenges stacked. Images often taken from 60,000 feet and higher are classified, harder to work with AI vendors and allies lacking shared clearances. Angles tougher: US spy planes and drones do not directly overfly China and Russia; they tilt from airspace above international waters. Sensor data less plentiful and harder to parse. Beyond electro-optical photographs: synthetic aperture radar sees through cloud, dust, and rain; multispectral imaging delivers high-contrast interpretation from light the eye cannot detect (metal hulks in forest; chemicals mixed with earth). Little commercial equivalent; resolution so low it often looks like blobs. Microsoft found the work taxing: a highly sought object on a multispectral image might occupy only ten pixels; a human between fifty and eighty, often too few to differentiate man from woman; a weapon on a shoulder might be three pixels. An experienced analyst parses by context; Lymberopoulos had no clue: "I couldn't tell it was a gun, because I wasn't trained."
Ch18 flow.
Maven was also navigating what Coddington described as a "tense, unfruitful" relationship with Air Combat Command. Last Wednesday of August 2019: Maven folks briefed commanders and staff at ACC headquarters, Joint Base Langley-Eustis beside Chesapeake Bay. They expected pushback; warned a particular Air Force major was preparing to brief heavily against Maven. The week before, ACC Commander Gen. Mike Holmes had taken the unexpected step of shaming Maven in public. He saw potential for algorithms to understand battle but did not think the US ready to turn that over to machines: he suggested the project's multi-million-dollar capability was no better than the skills of his aide-de-camp's three-year-old. "He picked out all the green things. You know, green, green, green, not green. That's what we're doing with Maven. It's car, car, car, not car." Maven still needed "mom or dad" looking over its shoulder. The team knew their models barely worked; they wanted a use case with a willing customer. Not for the first time, the Air Force was uninterested. "There's so much nuance it's hard to have these conversations," Coddington. "They weren't into it." Knowing the major was preparing to shoot him down, Cukor simply filibustered: three hundred slides. Some thought he was relying on a Marine Corps Deception-manual technique called "dazzling," overloading sensors with information. Generals left before the dissenting major ever got his turn. Even so the Air Force stayed on the sidelines, and the 18th pressed ahead.
On the night of 27 August 2020, the 18th carried out what O'Callaghan calls "the first AI-derived target" in the entire Department of Defense: an old tank hull wheeled onto a firing range. It was not pretty. Systems barely spoke to each other; the satellite image downloaded at molasses pace. Close to 10 p.m. when AI finally detected the tank, more than twelve hours after the system started looking. The system asked human minders to confirm the AI selection; next sent a message to an M142 HIMARS instructing it to fire; a rocket whistled and found its mark. O'Callaghan keeps a picture of the moment on his conference-room wall, flames licking the frame. "It showed us the art of the possible." They tried again later that year, initiating quarterly Scarlet Dragon exercises: place objects in the field; run models; try to detect them as targets. One Palantir rep embedded with the 18th was so quick and clever some called him "Rainman": outside every time there was a problem, Slack someone at HQ, get the system updated remotely within a couple of hours. Army officers used to obsolete equipment that could take years to replace were stunned. Systems often failed, though. After a satellite pass, sometimes no AI boundary box appeared even though the tank sat in the open staring everyone in the face. One high-profile Distinguished Visitor demo led to controversy. Intended to show detections, said one observer; "We actually didn't detect them," alleged another; "Nerds . . . just drew them on." Hard to unravel. Two people familiar with the allegation, which circulated widely in the Maven office and worried several teammates, told Manson no AI detection was ever faked. One said the satellite was taking so long that a named area of interest was hand-drawn, not the same as drawing an AI detection. Another: so little worked well early that the team set sensitivity "super high" so AI identified absolutely everything and distinguished almost nothing. An official at the 18th: "misunderstanding." Whatever the particulars, rumors fueled detractors; the 18th kept experimenting. One of the more charitable members said Cukor liked to say things to will them into existence; in briefings Maven tended to present the good as excellent.
Temple converts; Abbey Gate; children the models missed
Temple, a targeting officer, first told a Palantir representative to go away and cussed him out. "Manson was like, 'I don't need another freaking targeting thing. I don't care. I have enough crap to try to manage.'" A month later he started to change his mind. Assisting the ignominious withdrawal from Afghanistan after twenty years: US forces and allies rushed to evacuate as the Taliban took over; ultimately more than 120,000 people airlifted from Kabul in dangerous conditions; people sleeping in the office. Temple was given access to Maven to help make sense of the situation. On a single screen he could combine feeds tracking aircraft, monitoring logistics, watching for threats, showing locations of key personnel. "one could see General Donahue walking around," he says of the 82nd Airborne Division commander who would become the last US soldier to leave Afghanistan. So could hundreds logged onto the same system from Kabul, the Pentagon, and, according to Temple, the White House. Maven was running in the National Military Command Center; Defense Secretary Lloyd Austin and Gen. Mark Milley, chairman of the Joint Chiefs, were briefed using Maven Smart System. Temple had never seen so many people use the same system all around the world at once. He was used to wannabe systems that would crash with only forty people. (Maven had been accused of crashing an entire network one Sunday night in Afghanistan only two years earlier.) "That's when I became a believer." He apologized to the Palantir representative and asked to be taught. As bringing in supplies and evacuating people turned deadly, the team used Maven to estimate how many people were on the airfield. "A lot of people didn't believe how many people had gotten on there," another person familiar with the operation said. A human analyst would pause the feed and count heads by eye. The team tweaked a few parameters on their best models to reorient for crowds; AI did the counting instantaneously. Leaders in the Pentagon began to comprehend how much they had lost control of the situation developing in Kabul.
Ch18 hub.
On 26 August 2021, as thousands scrambled into Hamid Karzai International Airport, an ISIS-K suicide bomber at Abbey Gate killed 169 Afghan citizens and thirteen US service members, including eleven Marines. Two days later Biden warned another imminent ISIS attack was "highly likely." On 29 August, after following a white Toyota Corolla sedan for eight hours, US drones watched it enter a compound and targeted it with a Hellfire. Milley called it a "righteous" strike; the target an ISIS facilitator. That is not what people who knew the family said. All ten people killed were civilians, including seven children. The man targeted was an aid official trying to get his family out. An Air Force inspector general would later blame "confirmation bias" for, the New York Times later reported, "warping operators' interpretation of what they were seeing." Manson was told Maven was not switched on for this strike. An aide on the Joint Chiefs asked Brian Ward to see if, had AI models been used, they would have detected the children. Ward sat down and ran multiple AI models over dusty, blurry MQ-9 recordings. Not even Ward, an experienced intelligence analyst, could see any children by eye the first few times he rewatched the infrared and other feeds. "It was witching hour and as sun goes down sensors have a hard time, there just wasn't enough differentiation between the ground and the kids." Watching over, he began to detect a few shadows moving after all: the "few partially obscured forms" a US commander would later say were briefly visible. If a team was fatigued at the end of a twelve-hour shift, Ward thought they would be very unlikely to catch those moving pixels either, no matter how heightened their senses ahead of a potential strike and given fear ISIS was planning another attack. "You don't want emotions to get involved." When Ward ran several models at varying sensitivity, Maven's AI could not catch the pixels either: "It picked out the adult male but it didn't pick out the children." The incident raised why America was launching strikes if its sensors were not good enough to deliver a clear picture in the first place, whether AI or humans were watching. Abiding by the Law of Armed Conflict rides partly on intent behind any strike; if you cannot see likely collateral damage it is hard to assess intent. America's ability to summon firepower had raced ahead of its technical abilities. Ward saw hope for something else: knowing there was something on screen, however hard for the human eye, meant a chance an algorithm could detect them in future. He took the example to AI vendors: "This is why what you're doing is so important." Another person who worked on Maven was aghast: if Google had not refused to continue, they figured algorithm work would have been two years further on. "Those ten Afghan civilians would have been alive if the effort led by Project Maven to deploy AI on military drone camera feed had not been delayed," the person posted on LinkedIn. A Maven official told him to remove the post the next morning because of a prohibition against commenting about Maven on social media. Vendors who stayed were surprised their algorithms had not detected the movement. Ward thought more data and training would help. One vendor wondered if an alert would have pinged at a different sensitivity; Ward said he had tried every setting. Vendors asked for the video to practice; it was never provided. "Manson was disappointed." However much better algorithms and training data might get, a central problem would remain: AI would always miss things. A model can perform very accurately 99 percent of the time, then in 1 percent of cases model and user fail at the same time. "Does it mean this thing sucks? Probably not." But the AI could go either way. "It's a probabilistic system... It will never be perfect."
Sixty percent vs eighty-four; not Terminator
The 18th would soon try Maven with moving targets too: hunting enemy vehicles through aircraft or space sensors; simulating the same from submarines; classifying a truck as an enemy vehicle and tracking it through the Carolinas; dropping weapons from B-52s, F-35s, B-1s, F-15s, F-16s, and Gray Eagle drones on AI-detected targets. "We'll find it and we'll strike it," O'Callaghan said. "The whole nine yards." Manson was given a demonstration at the 18th drawn from a real-life early-2023 operation when a US Navy ship evacuated American citizens from Sudan amid civil war. Central map from unclassified satellite imagery. Yellow boundary boxes where algorithms identified ships in the Gulf of Aden. Blue areas for no-strike list entries such as hospitals and schools. Left-side icons for separate data streams such as vessel tracking that could overlay the map. A function for a "tactical data link": transmitting directly between machines a commander's decision to fire a weapon. O'Callaghan was so taken by AI he was reworking his entire targeting workflow. He analyzed the targeting cycle at the 18th as six key steps; with Maven AI he reduced the human role "in the loop" to only two places: the decision to act and the action itself. A human would also be "on the loop" during automated collection; assessments throughout AI-enabled. Still not as reliable as a human: AI could correctly identify objects with 60 percent accuracy compared with an 84 percent baseline for humans at the 18th, but farther and quicker. Just as AI was beginning to earn more followers, humans were reaching breaking point. The Palantir worker Temple had cussed out (and had since made up with) was toiling so hard that his blood pressure shot to 190/140, an emergency hypertensive crisis. Even checking himself into hospital thinking he was having a heart attack, he was still on the phone to Cukor about work. O'Callaghan insisted it was safe to reduce the number of fallible humans involved. "It's not Terminator. The machines aren't making the decisions, they're not going to arise and take over the world." More real than a fantasy toy-soldier convention. "It's up to humans to decide whether to pull the trigger." Perfect or not, AI was now part of the kill chain.
Ch18 compare.
19. Nobody Knows Targeting Better Than Trey
By the time Rear Admiral Frank "Trey" Whitworth heard about Maven, advocates feared he would kill it. He was J2, joint targeting purist, former SEAL Team 6 intelligence director. Cukor's fall 2021 briefing became a collision: billion-dollar AI shortcuts versus six-phase doctrine. Months later Whitworth was nominated to run NGA, the very home Cukor had steered for Maven. "I will either be famous or live in infamy," Cukor told his team.
Ch19 close.
Ch19 scale
Ch19 cycle.
Ch19 compare.
Guardian of the joint targeting cycle
Whitworth sat at the pinnacle of the intelligence function Cukor wanted to overhaul. Born within a year of each other; equivalent rank until Whitworth climbed and Cukor stuck. By 2021, Cukor's promotion chances had been shot down two years running despite glowing file statements from Lt. Gen. Jack Shanahan (before they fell out).
Critical pressure.
Ch19 ladder.
Whitworth was neither bland nor safe: SEAL Team 6 intel director; Afghanistan and Iraq; target developer right after 9/11; now "warrior-admiral" among the president's top military advisers. White hair, narrow eyes, deliberate jargon-laced speech; clinical and exacting. He framed himself to Manson as "kind of a purist" on the original targeting cycle. Nearly two decades on the military targeting committee; now top uniformed intelligence official for the joint targeting cycle: what to attack, how to prioritize, which firepower. "Nobody knows targeting better than Trey," one person said. "There is nobody on this earth better than him."
He believed with pride, and not a few blinders, that the modern US military held itself to extremely high standards, overwhelmingly met them, and accounted for failures. The cycle builds a legal action. "Joint" meant every service: Army, Air Force, Navy, Marines, Space Force. For decades operators criticized the process as distorted by imperfect intelligence, excessive focus on weapons selection at strategy's expense, severe time constraints, and communication problems. Doctrine: civilian populations "may not be intentionally targeted"; acts of violence solely to kill, maim, or spread fear among civilians prohibited under all circumstances. Particularly sensitive STAR targets, especially if they might kill civilians, need presidential sign-off; for the Air Force that could take seventy-two hours.
Whitworth considered "distinction," separating combatants from noncombatants, the hardest part. He spent a great deal of time embedding it. When DoD rewrote Intelligence Support to Targeting in May 2022, he insisted on removing the "demeaning" notion of mere support: target developers and intelligence professionals were "main effort." Mistakes meant dead civilians, possible law transgressions, presidential apologies, Congress humiliation, ruined careers, undermined war aims. Doctrine: the J2 prioritizes collection, analysis, validation, and assessment for all joint operations and provides "major input" to operators. Target development includes vetting intelligence, validating law-of-war compliance, and deciding no-strike list additions.
Six phases in dry doctrine: commander issues targeting guidance; development and prioritization rake options; capabilities analysis evaluates best options; commander assigns forces, sensors, weapons; forces plan and execute; combat assessment measures effect and cycles back, sometimes concurrently. Work at the 18th showed Maven giving fallible AI shortcuts through the loop, plucking humans from their traditional role in four of six places. "Trey was guardian of the targeting process," Cukor told Manson. "Now there was me just blowing it up."
The fall 2021 collision
Fall 2021: Whitworth visits the Maven office for one of Cukor's infamous briefings. Cukor liked to say he offered "a very gentle message." Listeners thought nothing of the sort: slides, money, success. A billion dollars. Broad area surveillance. AI targeting. Lots of information, little room for questions. According to one person, Whitworth's reaction was immediate: he blew Cukor up. Skipping steps? Moving too fast? Bending rules? To some in the room the volleys insinuated Maven was circumventing the vaunted joint targeting process. Whitworth questioned the idea that Maven could let an operator click a dot and tell a weapon system to shoot. "Tell me about the recordkeeping?" Project Maven was also pretty expensive. Was it really worth a billion dollars? He did not seem to think so.
Ch19 flow.
Cukor seemed a relentless salesman. "He believed in what he was doing and was selling it all the time," Whitworth later recalled, calling the description a compliment. Whitworth told Manson he did not recall exact specifics but remembered "a little bit of a tense conversation." It would have been "a bit irresponsible," as the person in charge of joint-force readiness and targeting, to back the project wholesale. "It's sort of the collision of two worlds," he said: new capabilities versus traditional targeting. His job: consistency, improving processes, accurately describing the order of battle. Cukor's replies felt airy: "Details, details... Let's show you the AI piece." Whitworth was the accountability piece. "Manson is going to be the one raising my right hand, taking the fall if we do this incorrectly," he told Manson. "Tell me about what happens after the bad drop when we go through a congressional hearing and we're getting hard questions?"
Cukor did not dispute Whitworth's military genius. "He was really good and smart but we did not see eye to eye." Whitworth had spent two decades on carefully evolved controls. Cukor took on every objection with little success. The meeting was so uncomfortable Brian Ward spent the whole time wanting it to end and wanting it to continue. Cukor told Manson it went fine. Half-confession later by phone: he wants to be laid back; he is from California; "There is just an intensity." After the meeting he told Ward and colleagues: "Manson told you he hates me." Now Whitworth seemed to hate Maven too. "I will either be famous or live in infamy," Cukor told the group.
Automating targeting, budget as weapon, NGA as home
America's 2018 targeting doctrine: "Automating targeting has historically been a challenge." Many parts automated; no single tool automates the entire process. Cukor considered the intelligence community's hold over operations cumbersome: neither fast nor relevant enough for future war tempo. He wanted Maven to solve that. He kept selling; was not particularly worried Whitworth considered them reckless; and besides, he no longer had to work with the intel side: he had the 18th Airborne Corps. "If Manson had just worked with intel we wouldn't be talking right now," he told Manson. "We would just be another failed intel thing." Twice-passed-over colonel challenging a rear admiral, he figured he was essentially in charge: civilian DoD under the undersecretary for intelligence had ascendancy over the president's uniformed advisers. "We tell them what to do." Most important weapon: budget. "I knew money talks."
Whitworth's concern marked a different inflection. Cukor was running up against the thirty-year clock: retire, land the AI baby permanently. He told Manson he did not mind where Maven ended up so long as it was not killed by retirement. Services had refused to spend their own money, as Bob Work predicted. Cukor also wanted to avoid handing Maven to the Pentagon's fledgling AI shop, the JAIC, which several at Maven saw as sour rival. After Shanahan's mid-2020 retirement the JAIC was heavy with personnel, low morale, unclear direction. In 2022 DoD would subsume it into the CDAO, which managed only halting success, wariness about large language models, and continuing low morale before Trump 2 put it under new management aiming to accelerate fielding.
Alternate home: NGA. Spy shop and "combat support agency," formed 2003 to unite institutions that mapped the world and its wars, back to aeronautical charts before America's WWII entry. Answered to both Defense and the intelligence community. Pored over America's most secret spy satellites for troops and spy agencies. Forbears mapped the moon. Cukor said NGA had saved him in battles more than once. Low-key early: Obama in 2009 at Five Guys in Navy Yard seemed not to know it existed until a diner mentioned Building 213 and "papers to you every morning." Greater renown after the Abbottabad compound model for the 2011 bin Laden raid; mid-2010s highlighting China's island and shoal military buildout. Teams experimented with commercial Maxar and Starlink-class satellites as well as secret ones.
By 2021 NGA was run by avuncular Vice Admiral Robert Sharp, long naval intelligence career, friend to Maven and close with Cukor: "a wonderful man," Cukor's highest-ranking adjective. Cukor set about diverting his creation mostly to NGA. Rachael Martin, JAIC intelligence analyst who loved working with Cukor, said he did "most of the initial backroom political wheeling and dealing" for the NGA transition. "He was really the driving force for that," one of the most determined and focused people she had ever run across. He pulled it off. 3 September 2021: memo supported transition of Maven's AI training and geospatial data lines of effort to NGA. A minority of lines (AI for text, anything independent of geography) would ultimately go to CDAO. Four-star commanders at EUCOM, CENTCOM, and SPACECOM were "really happy."
Then disaster for Cukor: January 2022, newly promoted Vice Admiral Whitworth nominated as NGA director to replace Sharp. Mid-year, one of Maven's harshest critics would be in charge. "Manson was quite a skeptic, to be honest," Whitworth later told Manson. "Manson was worried that it was expensive and that it was not all it was cracked up to be." The future of Cukor's baby looked increasingly in doubt. "We were all very concerned," Cukor said. "Trey was not a friend."
20. Kill Chain
Virginia winery, spring 2025: former SEAL Team 6 commander Joe Kernan, once top defense intelligence official with a $54 billion budget, wraps every seven minutes and talks another fifteen. Maven worked because Cukor was "a one-man wrecking ball." Spring 2020 anonymous complaint lands in Kernan's mailbox and his daughter's: Palantir grooming, Larson conflict, Poggemeyer rage, workplace climate. IG and NCIS follow. Cukor scores a gold star on contracts, a "big-ass" letter of reprimand on supervision. "Manson was Taliban." Farewell frame labels him Founder; Colin Carroll in flamingo shirt erased under "HUD Mask." Monday: J.P. Morgan.
Ch20 close.
Bleu Frog, Marcinko parallel, complaint in the mailbox
Manson waits more than an hour in a Virginia winery corner for an unusual barman. Joe Kernan, former SEAL Team 6 commander, runs Bleu Frog Vineyards (cordon bleu + SEAL nickname); bottles include Mad Frog ("a very fine and fit structure"). Career ops focus; then civilian top US defense intelligence official three years until November 2020: three spy agencies and multiple defense intel projects including Maven under him and a $54 billion budget. Thundering spring afternoon 2025: does not much care for journalists; announces wrap every seven minutes; carries on fifteen; stammer through ninety minutes. Believed in Maven: more granular intel, minimize collateral; wished he had had something like it. Initial success: Cukor "a one-man wrecking ball" who never gave up on Congress, Pentagon, own team. "Cukor was a pain in the ass; but that's what I wanted." Did not agree on all positions (thought Cukor should share data; wondered if protectiveness had reason); "enormous respect" including Congress funding. Parallels to Richard Marcinko, controversial SEAL Team 6 founding commander. Cukor might have benefited from a more diplomatic stance; evolution Kernan indicated never came.
Ch20 ladder.
Spring 2020, Covid hunkering: anonymous complaint about Maven addressed to Kernan. Furtive email chains; slipped under Pentagon doors; even Kernan's daughter's mailbox. Four-page letter Manson reviewed: "gross waste and abuse" in acquisitions and program management; alleged contracting violations; "abusive and negative relationships" involving Maven personnel with government and contractors. Three people separately shared it; one version 12 April 2020 to Kernan; another undated to DoD Inspector General. Impact would ricochet for years. Cukor always expected investigations: less tool of the righteous than tax on DoD innovation. IG already delivered one classified evaluation November 2019; Cukor ignored the note-taker. This complaint triggered broader investigation he would struggle to shrug off, lasting personal impact. (Trump Pentagon, September 2025, would describe a "weaponized" investigations process and plan to ban anonymous complaints.)
Eleven bullets: Palantir, Larson, Poggemeyer, JAIC
Complainant: former Maven detailee compiling firsthand and "multiple current and former" employee incidents for potential IG complaint. Eleven bullet points: contracting practices to personnel behavior; focused mostly on Palantir, Cukor, and his team. Alleged Cukor "grooming" Palantir to take over as prime: subcontractor work from small 2018 pilot to about $40 million a year despite multiple government personnel recommending against proprietary Palantir. Circumventing ECS Federal (prime) to negotiate directly with vendors, specifically Palantir. Informing potential subcontractors of budget limits "to drive the overall cost proposal lower," yet Palantir allegedly afforded "astounding" advantage: representative sat in unannounced on Maven discussions of service budgets for Palantir Smart System and contracting plans. Questioned Lt. Col. Joe Larson, full-time Palantir employee called up as Marine reservist as Cukor's deputy: "impossible to enforce" ethics guidance preventing participation in contracting as Palantir grew to touch almost every aspect. "I cannot prove that this conflict of interest directly resulted in material gain for Palantir and the Larsons, but the timing of the rapid expansion in Cukor-directed Palantir subcontracts is odd." (Larson strenuously denied: written notifications, DoD Standards of Conduct directives, formal recusals, ethics firewall; when firewall impossible as Palantir grew, left Maven and briefly returned to Palantir. Ann Marie Rosas representative did not respond.)
Ch20 flow.
Letter alleged Colin Carroll oversaw APL software developers while actively applying for a job there, which he accepted. Also: five people (Carroll, Jane Pinelis, Kelly Rooker, Elston ToChip, Gene Whipps) told the author they were leaving "in part or entirely due to mistreatment by Captain Poggemeyer," blaming command climate set by Cukor. Carroll later: does not remember Poggemeyer among even top five reasons he left. Pinelis via representative: unaware of April 2020 complaint referencing her; did not leave for mistreatment; Maven among most rewarding chapters; Cukor one of most intelligent, mission-focused, hardworking leaders she had worked with. Whipps: did not leave for mistreatment. Rooker and ToChip unavailable.
March 2020 example: during technical security discussion Poggemeyer allegedly yelled he would "make a meme" out of Nand Mulchandani (JAIC CTO); further soured already-contentious Maven-JAIC relationship. Mulchandani: did not recall; if happened would not have affected him or raise complaint about a junior; acknowledged heated meetings mostly as JAIC sought Maven data Cukor denied again and again. Writer withheld name fearing retaliation. "All I ask is that you take these allegations seriously... The reputation of USDI and the entire Department is at stake if this ever become public, and would set the Department's AI development work with commercial industry back years."
Plenty chafed. Carroll appeared especially well informed about investigations. May: told colleagues there might soon be action on command-climate investigation into how Cukor ran the team. Wanted Maven attitude toward JAIC changed; wanted data pipeline and funding. Soon telling people JAIC would own Maven within a year. Was the mini-Cukor who loved Cukor and would follow him to hell, self-described "asshole," taking action against his former boss? Asked if he wrote or was involved: email "I am pretty sure that had I written a complaint, Manson would have not added myself to it." Later by phone: no to both writing and involvement. By then hugely frustrated; wanted Cukor gone. "We spent over a year of attempts to work with Maven to access their data that the Department would benefit from. We were stymied every fucking bit of the way."
Kernan did not discuss contents with Manson; heard allegations, could not ignore, passed on for others to determine justification. Investigating officer assigned; reached out; some agreed to talk. Sometimes about Cukor, sometimes Poggemeyer; after weeks also Carroll. Winfield Adkins, Army colonel, tasked deep into Maven and by default Marine culture. New details: multiple romantic relationships; Covid work-condition discomfort (Cukor secured lockdown exemptions); more. (Adkins declined comment.)
Rage Slack, GPU tirade, letter of reprimand
Several described Poggemeyer as passionate, intense, brilliant, temper-prone. Multiple: slammed a locker near a co-worker so hard the office stood still. Through representative: declined comment on this or any alleged action. Not only volatile member: another treated colleagues so poorly it nearly ignited fistfights. A third blamed Cukor for atmosphere allowed unchecked. Cukor: regularly called young passionate outstanding team members in for private counselling, mitigate, instruct apologize. Manson reviewed emails and Slack: Poggemeyer expletives and fury at others' pace. Team so used they kept a jocular Slack thread including him: "Manson is hoping I'll be occasionally cursed by Sy for a year." May: "I have to fight for every fucking inch... Manson is freaking the fuck out because nobody fucking understands what's going on or how to achieve it." Twenty-eight minutes later: "I didn't want to have to go full rage on you guys." Wanted six better slides then would support experiments. "But now Manson is going to pull back and coddle you because Manson is worried you guys will quit after this." Reply one minute later: "Manson is sorry you had to go full rage. You should work on that. It doesn't help your messaging." "Slides," Poggemeyer.
Ch20 hub.
Among explosive Marines, some got on and did not mind. With contractors, some felt over the line. Clear to several: particular fury for Carroll. One message Poggemeyer minutes later called a "rant": invective against unnamed "mini mob boss" who stole Maven plans and threatened to "burn Maven to the ground" if he did not get to use his very close friend's company ECS for projects Poggemeyer called "corrupt." Team understood: Carroll; close friend Volkovska. Carroll and Volkovska not the only Maven couple to marry uniting government and commercial contractor. Poggemeyer later married Shannon Clark, Palantir defense business lead (later worked at Palantir himself). Larson and Rosas married, both had worked at Palantir, both on Maven. Other romances short of marriage. None bothered Cukor: normal close attachments in a hardworking group; wanted results. Neither Poggemeyer nor Carroll decision-makers on awards though he listened.
Extra wrinkle: Volkovska (Odessa) worked closely with Poggemeyer for a time in 2018; both spoke Russian; secret language for a while; brief romance she did not continue. When Carroll and Volkovska dated spring 2019, Carroll astounded to learn of the prior romance. He left Maven that July to continue the relationship without conflict of interest; took APL role listed among alleged problems. Late September 2019: Volkovska complained after Poggemeyer called with loud expletive tirade about GPU chip allocation favoring APL though she said Cukor had asked for the allocation. Loudspeaker; shaking; hard to breathe (contemporaneous documentation). Never spoken to like that; did not want him visiting ECS again. Anonymized episode resembling this forms part of the complaint.
End of July 2020: Poggemeyer left. One day just not there. Director Gregory "Jesus" Christ all-hands in socially distanced Pentagon amphitheater; note card: would not tolerate workplace problems making anyone uncomfortable. Cukor: "Okay," jump right back into work. Nothing would stop the pursuit. Determination took a toll. "Any normal human being would have given up... It got pretty ugly." Worked the group no less hard under Covid lockdown. First some Mavenites knew of investigation: New York Times August story on the memo and Palantir complaints, four months after Kernan's mailbox. Carroll told people in advance he hoped the article would destroy Maven; Cukor needs to go. Cukor: has not seen the complaint; does not mind who filed; cost of doing business. "Whenever you do something in Defense that is out in front, people start taking shots on you. They were trying to take me down... It was pretty intense."
Most unsavory implication: too-close Palantir relationship; unfair favoritism or worse; any current or future financial benefit to Cukor? From October 2020 IG separate investigation lasting a year. Personal suspicion appeared to fall on Cukor and wife Kirsten: moved four children, two cats, dog into "small, jankety" 1950s two-story, 1,400 square feet, Alexandria outskirts 2005 to be closer to Pentagon. NCIS agents inspected where he lived (often works with DCIS on procurement fraud and public corruption). "I didn't let any of that stop me." Investigation added difficulty to something Kirsten already found impossible. "They came to our house and saw, like, all of our cars had over 100,000 miles on them, and that we lived in 1,400 square feet."
Cukor always expected complaints; careful; knew acquisition rules from Marine procurement days. Chief and technical monitor: day-to-day contract execution; keep ECS informed of anything direct with subcontractors such as Palantir. Took no money from Palantir; deliberately never took a Palantir job (unlike others at Maven, DoD, Congress); they played no role in his J.P. Morgan job. No one Manson spoke with thought he took Palantir money. Carroll: guarantee no kickbacks just knowing him. Kernan: when delivering new AI capability, DoD needed to pick a horse, jump on, force forward.
24 August 2021: Carroll himself fired. Public post: Kathleen Hicks asked his new JAIC boss for his resignation. (Hicks declined comment.) That evening Carroll and Volkovska on the sofa in tears: so much life into AI and autonomy for DoD; concluding no autonomy, only ruptured careers. Down a job; went ahead with plan: marriage proposal. She said yes.
October 2021: IG concluded Maven review; published January 2022. Cukor: exonerated on every accusation relating to inappropriate contracting. Dry language: "successfully" monitored and managed four contracts in accordance with all relevant regulations. (Review period predates Palantir as prime from September 2020.) Approaches in fast-moving AI could provide valuable lessons and best practices, a senior defense leader's testimony included.
"Manson was Taliban"; farewell Heat Shield
Cukor did not entirely escape. "I think every Marine learns at some point that the Marine Corps will never love you back," Kirsten said. Most people thought he retired because time was up: no general within thirty years, you retire. One day he answered more fully. Following workplace complaints and expanding investigation: mid-2020 formal letter of reprimand. Serious response to significant misconduct or disciplinary issues; permanent record; enough to curtail career. Marines took him in; now spitting him out. "I got a big-ass letter that said 'Drew you should have exercised more military supervision,' and I got read the riot act... You fell short of the expectation of a senior officer." Let it wash. "I don't even care. They don't get it. Manson had to push hard." Then: "Manson was Taliban." Meaning: DoD extremist. "You are just way too much... A Taliban guy." He saw differently: every day planning for World War III; battlefield aggressiveness at work. Most important enemies closer to home: contracting disputes and workplace complaints. Acknowledged junior officers aggressive in the moment, also able to apologize. Why reprimand young officers for how they spoke to seniors? "Why? The results are amazing, and he's delivering and we're moving at record speed." Held to impossible leadership standard for running premier AI program. (Recalled generals kicking trash cans with impunity earlier.) Good boss would have said "Go fix it." Knee jiggling; same warm voice onward. "When you read it it's like, 'Damn I must be pretty bad I guess I'll throw myself off the bridge right now.'... I didn't let it affect me. I just kept executing." Reprimand requires acknowledgment; most rebut. "Yep, I did all that... As a senior officer, I should have reprimanded all these young officers that were working so hard." Impossible ask: young Marines fresh from overseas deployments into a Pentagon where everyone is over fifty-six. Twenty-five-year-old captain whose frontal lobe just formed. Strong personalities also effective ones. Carroll: "very, very aggressive, amazing, great... I love Colin... But he was impossible to be around." Speaking after Carroll marched out again April 2025: "I guess he's still learning... When he's my age, he's going to be lovely: they'll get smooth." Kirsten: husband lets go; she has a list of people who should burn in hell; no poker face; do not ask who.
Ch20 compare.
Complaints kept filing even as he prepared to leave; one that he sought senior NGA role to steer the baby. "It went on all the way until the day he left." Last Friday of October 2021, as soon as IG all-clear: retired. "He retired as a colonel," Kernan later by phone. "That tells you something": not grave enough to dock a rank. Following Monday: J.P. Morgan. No holiday. No reflection. Work.
Around holidays small Mavenite farewell: 10 December 2021, Blackwall Hitch, Alexandria Old Town Waterfront; views to Capitol; looking inward. Gold-framed farewell, wide blue mat, gold-pen messages. Jeff Anderson (Harbinger): "You are an American hero and innovator." Rich Dorchak: set DoD in motion to understand future warfare with AI. "True visionary"; "DoD's 'Godfather' of AI/ML"; "legend of Col Cukor." Brian Ward to the "Great" Col Cukor: set conditions for baby to wreck, be destroyed, reborn better. Lt. Col. Kelly Martin smiling dig: "although you leave a lot to be desired... dedication, passion and perseverance... next John Boyd." Latin: "Videmus omnia." We see everything. Two group pictures: 2017 early team; stiffer 2021. Earlier: Cukor cammies, sleeves rolled, elbow ready; boundary box "Heat Shield." Three down: Gregory Christ jacket/tie "Heat." Right: Larson same attire stance. Left: taller silhouette entirely covered. Later unadulterated original: smiling Carroll in pink flamingo shirt entirely obscured, Soviet fashion. Label: "HUD Mask" (Heads Up Display drone feeds the team scrubbed). 2021 photo: box over Cukor's face "Founder." He did not take it that evening; straight to next thing. Back in office a while; Ward safekeeping. Cukor never came back for it. Experiment to change warfare over. Drew Cukor was gone.
What the gold star did not cover
The IG contract review period predated Palantir becoming prime in September 2020, so the gold star did not speak to the later prime transition that had fueled so much of the complaint's heat. What stuck was supervision. Mid-2020 letter of reprimand: permanent record; career curtailment. Cukor let it wash and kept executing through Covid lockdown exemptions he had secured for the team, through NCIS house visits that catalogued high-mileage cars and 1,400 square feet, through the Times story Carroll hoped would destroy Maven. Kirsten's list of people who should burn in hell was the spouse's ledger; his was the next Monday at J.P. Morgan. Kernan's phone coda ("He retired as a colonel. That tells you something") was both comfort and indictment: not docked a rank for misconduct, still denied the star path. The farewell frame's "HUD Mask" over Carroll's flamingo shirt made the personal cost visible in Soviet-erasure metaphor: the team that built Heat Shield could also scrub a founding face from the founding photo.
Ch20 cycle.
Covid exemptions, NCIS visits, the Times story
Through the Covid months Cukor secured lockdown exemptions so the team could keep labeling and shipping. Through NCIS house visits that catalogued high-mileage cars and about 1,400 square feet, the family learned what a complaint looked like when it became a federal inquiry. Through the Times story Carroll hoped would destroy Maven, the project survived as a line of effort even as Cukor's star path did not. Kirsten's list of people who should burn in hell was the spouse's ledger of moral injury by proxy. His was the next Monday at J.P. Morgan after the Marine uniform came off. The IG gold star covered the earlier contract period, not the Palantir-prime transition that had fueled so much heat; the letter of reprimand covered supervision. Both facts can be true at once. The kill chain chapter ends where the career does: Heat Shield worked; the colonel who drove it left without the star.
Ch20 scale
21. Ukraine Fights Back
Early February 2022: Colonel Joseph O'Callaghan tells Joey Temple the president is sending troops for Ukraine; the TV confirms it; he means Temple. By 5 February Task Force headquarters sit in a Wiesbaden bowling alley: thin carpet over lanes, gymnastics room become ops center, "The Pit" in a U-bend. General Michael Kurilla demands Maven when thirty-six-hour intel fails. Invasion day: accuracy collapses from about 70 percent to 30, sometimes 10. "Get it fixed or get it out of here." Retrain on snow and rubble; hunt TELs; Jim Caggy flies to Frankfurt Saturday; AWS finds data packets crisscrossing the Atlantic four times. By May, points of interest flow; eight-digit grids; Signal and Delta mesh. Kherson artillery unit "just happened" to get hit.
Ch21 close.
Ch21 scale
Go bags, bowling alley, Battle U
O'Callaghan called Temple into his office. Pentagon about to tell the country the president is sending troops to support Ukraine. Looking at the TV: he is doing it now. Life-changing: he means you. Early February 2022. Temple called home, go bags by the door, drive back to base. Plane next morning. "Manson went home that night, oh my God, kissed everybody." Last notice this short: Kosovo twenty-five years earlier; missed first daughter's birth. No idea what Europe in support of Ukraine would include. "We didn't really know exactly what that meant, to be quite honest."
Ch21 ladder.
Military plane over the Atlantic. By 5 February, new task force HQ from the 18th Airborne Corps ensconced at a US base in a small city west of Frankfurt. New arrivals to the joint targeting board; start right away. Preparing to implement the OPLAN for US defense of Europe. Shop in an old bowling alley: thin carpet over wooden lanes; ceiling holes; meetings in a racquetball court that months before hosted gymnastics for seven-year-olds. Now ops center of a secret war. About three hundred US soldiers flown in haste. US Army Garrison Wiesbaden about a thousand miles from what would become the front. America's AI-for-warfare effort faced its most significant test. Algorithmic warriors assembled in a U-bend beyond double doors: "The Pit." Within three hours Maven up in the fusion cell spanning intelligence, operations, logistics. Right side of makeshift tables: intelligence, find everything preferably before it happens. Other side: fires, weaponeering the right munition on target. "Battle U." Other units segment; horseshoe mingled the teams. Agency experts later joined; intel/fires elided physically. Task Force Dragon ready. Drew Cukor's dream becoming real.
US intel, intercepts and other means, indicated the about 175,000-strong Russian force on Ukraine's border was about to attack. Senior European intel officials, burned by American WMD "intelligence" on Iraq, told Manson US invasion arguments were overheated: no blood banks near the border; maybe exercise or feint. As Washington struggled to convince NATO, it sent troops closer. Newly arrived soldiers with no Europe fight experience prepared to stage an evacuation or perhaps go in. Long-standing conceptual battle plan included targets across the Russian border: garrisons building up; hitting command centers beyond Ukrainian range could mean headquarters deep inside Russia.
Kurilla the Gorilla: use Maven
General Michael Kurilla, fifty-five, 18th Airborne Corps commander, by February 2022 believed AI was "the next revolution in military affairs." Large, loud, exacting; dubbed "Kurilla the Gorilla" in the press. Instructed tracking the border buildup. Team monitored Russian and Ukrainian movements on the old Command Post Computing Environment (CPCE), little live data. About ten days in, briefing Kurilla. "What are the Russians doing?" Intelligence chief sketched maneuvers and likely invasion. "How up-to-date is this data?" "Sir, thirty-six hours." Not the answer. "Why?" "That's the latest we have, Sir." "This is not a clear picture." Unbeknownst to the briefers, Kurilla already had a brief putting him eighteen hours ahead; he thought Russians about to do something else. From here on: use Maven. Team did not really know what that looked like. A system far from camera-ready called to support the largest European war since WWII. They had used Maven in staged exercises; this needed real, up-to-date data to save Ukrainian lives. Assembling intel and open-source feeds; little satellite imagery; tinkering and guesswork until 24 February.
Ch21 flow.
Early that morning Putin's face on TVs: "I have decided to conduct a special military operation." Satellite internet out across Ukraine; cyberattack later blamed on Russia downed tens of thousands of broadband connections as far as Lapland. Invasion starting. First missiles; screen beeping; two US watch officers at the Joint Operations Center in Germany after midnight East Coast: wake Chairman Milley? Pre-written triggers said yes. Gruff: "Okay, I'll tell the Secretary of Defense and the President." Someone roused Wiesbaden from warehouse cots. Work immediately; every fresh feed. Kurilla to the Pit: "What are the Russians shooting at?" So many munitions the US could not figure the game plan. Infrared from overhead persistent radar satellites into Maven for possible rocket launch sites; track FLET and FLOT. "They're hitting a lot of dirt." Possible train relay; thuds landing fair distance away. Enough ordnance eventually hits. First three days: Kyiv nearly overrun; Hostomel Airport assault pushed back; Zelenskyy from comedian-politician to war leader overnight. Tanks in; civilians out; hundreds dying. Some US personnel in Germany and Poland felt utterly impotent. "It's like being in a long-distance relationship."
Pentagon wanted transport routes and whether Russia struck escaping civilians. "Somebody needs to go to Poland." Brian Ward put his hand up; arrived in Jasionka, about 100 km from the border, days after the war started. Patriots arriving; younger recruits feared nuclear strike as fighter sonic booms passed. "If we do too much, Russia will nuke Ukraine and then they'll nuke Poland," Ward remembers the theory; "that fear actually felt real." (Later November: Ukraine accidentally fired into Poland; Zelenskyy blamed Moscow; Pentagon wanted to know if Russia was shooting at NATO. "We thought that was going to be World War III," Ward says.) Border both ways: escaping; flooding in "because they thought they can go kill some Russians." US ran border control camera feeds into Maven Smart System; algorithms estimated leavers more efficiently than eyeing buses. At war's start Milley got his own Maven account: throng live from his office; satellite feeds; Kurilla as map icon. Erasing layers between ground commanders and top brass cut both ways: some units historically embraced disconnectedness to deprive seniors of live data; AI common platform could mean micromanagement and deadly delays AI was meant to erase; also a transparency tool if turned inward.
Accuracy collapse: snow, rubble, TELs
Even as Maven showed a friendly general moving around Wiesbaden, the AI was not working. "AI was struggling to get the detects that we had seen previously." Initial Europe detections about 70 percent accuracy; immediate days after war started plummeted to 30 percent, sometimes 10. Daily naysayers so bad one official issued something of a "fatwa" against Maven and tried to ban it in the Pit. Kurilla not unsympathetic: if it did not work, he did not want it. Adherent: "We can fix this." Kurilla direct: "Get it fixed or get it out of here."
Ch21 hub.
Algorithms had not been trained on Ukraine pictures; unused to military hardware against snow rather than hot desert scrub. Occasionally tanks; no idea what a tank looked like with turret blown off. Lab-trained models could not recognize rubble buildings or smoldering craters; could not tell intact threat from blown-up wreck. Battle damage assessment fundamental to US fight: mission success, re-attack, precious ammunition. Without BDA, AI had no chance. "The Ukrainians would hit a Russian tank, we would still detect the tank, but then we weren't detecting the fact that it was inoperable because we hadn't seen that data before." Sometimes a T-72 with turret off looked so different AI would not detect it at all. Need to reteach quickly. Collect current-war data. "I don't care about China right now. Let's focus on Russia, and get as much of that data as possible." Not everyone agreed: "global" models; Ukraine-specific retrain risked overfitting, less useful for Middle East and China.
4 March, ten days in: call for additional East Europe satellite coverage. About one thousand Russian tanks ground into a about thirty-five-mile halt on roads to Kyiv; aim: new imagery labeled fast. And we were very, very, very quick to take them up on the opportunity." Hundreds of labelers on unclassified and classified pictures; new data stocks; improved label vetting. Strategic resource for future algorithms. AWS, Maxar, Microsoft tasked to retrain as soon as performance failed. Tanks better; Silicon Valley still did not understand war. Major challenge: TELs (transporter erector launchers), Russia's common mobile missile launch vehicles, "a pretty big deal" to take out. Looked like trucks; bedrock of Putin's effort firing S-300/S-400 at aircraft and deep targets. Often only a few dozen pixels from far away. Vendors did not know several TELs typically show up together, often fanning in a semicircle. Keeping vendors at arm's length (nationalities, clearances) meant computer scientists could not tune to the signature pattern.
Some saw early Ukraine failures as computer-vision limits. Others: little more than two weeks of new imagery and retrain adapted the tech; detections rose. Palantir representative on site helped system updates. War on the go; far faster than usual hardware/legacy updates. "That's not five years, you know?"
Donahue, line-stepping, cloud plea
11 March: Kurilla to Middle East command; newly promoted Lt. Gen. Chris Donahue takes 18th in Wiesbaden. Last US soldier to leave Afghanistan months earlier (night-vision photo he reportedly disliked). Revered by some subordinates as "the best military commander of our time." (If he told them to go to hell, one would gladly plan how.) Ukrainians had a perpetual blind spot: high-end homegrown tech, professional, committed; firing 107mm and 122mm from SS26 with their own fire control; could not see more than about six kilometers beyond Russian lines; could not figure the next assault. Congress complained intel was too slow. Sen. Ben Sasse: knowing where a Russian tank was ten hours ago not helpful to a Ukrainian defending family. Sen. Mark Warner: sharing too much risk given Russian penetration of Ukrainian services. Rep. Adam Smith: limits without going to war with Russia; US not sharing real-time targeting "because that steps over the line to making us participating in the war."
Ch21 compare.
Donahue was a line-stepper. Former Delta; doubled down on Maven to advise Ukraine on any useful American find. Not only refugee flows and equipment counts: identify targets to help take out Russian equipment and personnel. Sensor→software→shooter in Ukraine could expose US as direct participant. Ad hoc policy limitations sometimes doubled as sleights of hand to avoid escalation. "The whole thing was a little bit fly-by-night." AI detections leading "to great tips"; still missed plenty; enough to start looking in the right place. Particularly helpful with SAR (weather/smoke unaffected) identifying TELs better than humans. Findings passed to Task Force Champion (Stuttgart; EUCOM and Army Europe/Africa), which passed to Ukraine. "Taking way too long... We would email it to them and then we don't know what happened."
Map forever to load; dial-up click speeds. Thousands dying; Ukrainians pulverized. EUCOM plea: "We need more cloud in EUCOM!" Joe Larson (Maven chief after Cukor) met Jim Caggy (AWS at Pentagon) morning of Wednesday 16 March. AWS flummoxed: Europe already had plenty of cloud. Feedback: Maven Smart System on Secret network in Germany was unusable. Yet top military intel officials assured Congress US intel-sharing was "revolutionary." Friday: specialist AWS room remote with 18th network diagrams; no idea. Nobody from AWS had checked the harassed customer in Germany. "Who's coming to Frankfurt with me tomorrow?" Caggy. One volunteer Monday for family. Caggy: war does not heed weekends; flew that night; Frankfurt 8 a.m. Saturday 19 March; Wiesbaden within hours. Tony Bass Gym: even Saturday morning with about a quarter of the team in, single satellite image up to eight seconds. Operators blamed Maven; mapping IPs, routers, switches, packet journeys: each bundle traveling mandated Army defense networks on a wildly circuitous route. "Data was crisscrossing the Atlantic," sometimes as much as four times. "This is insanity," Caggy. Bad systems engineering risked losing packets; data dropping off the face of the earth. Same in Poland. Caggy to Kraków Saturday evening; Sunday troubleshoot 82nd near Rzeszów. "The Army failed the Army." Network sold as "global network superhighway," "cornerstone" of Army modernization the year before: high-bandwidth as soon as boots hit ground. Simplifying required senior leaders breaking bureaucratic resistance. Route restructured; delays down significantly.
Same month, fears Ukrainian data centers would be blown: AWS Snowballs to Germany and Poland; handoff to Ukrainian special forces. Others got Starlink after Musk answered a tweet appeal. New targeting authorities in April; first American M777 howitzers with GPS-guided munitions arrived. Bizarrely Americans had not given the software to fire them; Ukrainians developed their own. "They're brilliant." Still: forty-five days on the sidelines watching Ukrainians get killed. Now they could finally do something.
Points of interest, eight digits, Kherson
By May, Task Force Dragon sending locations of Russian stationary and mobile equipment and personnel with AI help. External glare even from colleagues: illegal, immoral, unethical? Donahue's "force of will"; "We're not doing anything outside of doctrine... All we were doing was optimizing doctrine." Strict rules chafed. Euphemistic "point of interest" package: terse description, elevation, timestamp, location. Full ten-digit grid (about one-meter accuracy) might open accusations the US was targeting Russia and had entered the war. Team stopped short by a sliver: eight digits (about ten-meter accuracy) still close enough for a missile. Larger targeting box a curious upshot of balancing escalation and precision. "We like precision, and you guys in the media like precision. But the policy is what we could get through."
Ch21 cycle.
"I don't want to say riveting... It was very, emotional is not the right word, but it gave us a sense of value." Once the system worked they kept going. Sometimes British first for easier pass-through; sometimes Ukrainians directly on Signal; sometimes train with paper and manual upload; early days two phones held together. Soon Maven Smart System and Ukrainian Delta meshed: points scrubbed of collection methods, then into Delta. (US viewed Delta as leaky unclassified; both US and Ukrainian officers told Manson Delta often held a higher security standard than America's Secret systems.) Early days also Kropyva (Nettle), Ukrainian Army Android artillery fire control. (Palantir separately provided big-data analysis for intel reporting.) Ukraine had "utter faith" enough to lob artillery at locations in their own country where misfire could hurt their own. Early success: Wiesbaden eyes on a Russian medium artillery unit giving problems in the Kherson area (southern port of about 300,000, shipbuilding, fallen early March). Weather often interferes with optical; SAR sees through; Maven located the unit; whereabouts to Ukrainian forces. "And they just happened to make the decision to do something about it."
22. Tens of Thousands of Targets
From the bowling-alley U, the 18th soon gave Ukrainians thirty to thirty-five targets a day, nearly triple Iraq five years earlier. Donahue: turn it up; then seventy. Yellow AI boxes; blue when SIGINT double-sourced; TikTok geotags. June 2022 HIMARS arrive; first US coordinates etched in memory; volume peaks at 267 points in a single day; tens of thousands total. Georegistration off by fifty meters forces a wartime shutdown. 1,500+ algorithms tested; two dozen for Ukraine; $275 million DoD spend in 2022. Eighteen-minute TEL kill; dynamic targeting later stopped by permissions. Encryptors, AWS circuits, Nakasone versus Signal, Taiwan briefings: "Machines will be fighting machines."
Ch22 close.
Ch22 scale
Ch22 cycle.
Yellow, blue, social: speed triples
Working the U-shaped workspace, 18th Airborne Corps soon giving Ukrainians thirty to thirty-five targets a day: nearly three times Iraq five years earlier. Speed of war tripled. General Christopher Donahue heard targeting information was really helpful. Turn it up a notch. Maven team pumping seventy a day. Ukrainians did not ask how; US counterparts appreciated that. At first AI satellite detect = yellow boundary box. Maven Smart System also brought SIGINT: unsecured phones and radios, geotags on the same map. When AI and SIGINT hit the same place: blue box. "When you see it pop up you knew what that color meant; it allowed us to be even quicker." Maven also got pings when the public posted missile explosion sites; TikToks with geotags logging sentiment and Russian advances. Unverified info worried traditional intel professionals; war moved too fast for traditional collection. "At some points during the war, Twitter and social media were giving the most up-to-date info."
Ch22 ladder.
Most consequential: Biden June 2022 decision to send four HIMARS, expanding nearly tenfold by 2024. Dump-truck-sized; GPS-guided missiles farther than Ukraine could see; then drive off. Turkish Bayraktars helped early; once shot down, US became Ukraine's eyes. First time US ever sent a coordinate to a HIMARS etched in one team member's memory. End of June: "a couple of targets." Immediate self-correction: "Points of interest." "We had a general ballpark idea when we thought they would be going out." Confirmation of sorts when aftermath pictures hit social media; sites far deeper behind Russian lines than usual. Ukraine later said it destroyed an Izyum base, killing at least forty soldiers; another strike same night reportedly killed a Russian paratrooper unit commander. Points ended up with the person sitting in the HIMARS. Sometimes felt as though the US was all but punching coordinates into the weapons themselves. Donahue: close relationship; Ukrainians trusted implicitly. Volume rose to a high of 267 on a single day in 2022. Two Wiesbaden targeting officers that year became the two most prolific individual Maven Smart System users in the entire US military. In total the 18th sent tens of thousands of targets to Ukrainians with Maven help. "How many points of interest you give 'em doesn't really matter," Donahue told Manson. "It's how many munitions they have. The more points they could handle, the more we would do it, as long as they were being effective and they were using the right munitions against the right targets."
Ukraine defense minister Oleksii Reznikov tweeted HIMARS making a "HUUUGE difference"; later pleaded a hundred would win the war. Each morning US and Ukraine teams discussed High-Payoff Target List. Broader concept proven: Maven beginning to automate the kill chain. "We were handing it to the partner nation on a silver platter so they could go strike." Self-correction: "Or do what they needed to do, I should say." Those involved knew the US role went further than limited intel support described publicly. "It's like being a puppet master but the puppet has a brain... almost like Pinocchio." (One day humans might be the puppets and AI the strings.) Team fended off policy-violation accusations. Some argued any classified-platform info to a partner broke rules even when scrubbed. Sometimes team tried to replicate higher-resolution government satellite detail of a TEL with grainier commercial equivalent before packaging to Ukrainians. "Trust us; hit this truck," without saying TEL. High-profile Wiesbaden visitors offered a chance to send a point themselves. "If you're not sure what you're doing is legal, just have the secretary of defense do it once."
Fifty-meter georegistration disaster; 1,500 algorithms
Embarrassing digital disasters still. One summer day target boxes in strange places; one around a tree. Operators could see what the algorithm tried to detect; off. Then every box on the map off. Geolocations paramount to hit and to avoid wrong thing: off by as much as fifty meters. Georegistration off across the board. "That's a big deal... It goes from being useful to not useful very, very quickly." Had to take off the algorithm and shut the entire system "at the absolute worst time." Pull bad algorithm; drop in new. Pentagon Maven team egg on face; humiliation bit. Germany and Pentagon running on fumes. "We hadn't quite caught our second wind yet as a program. And everything else was going wrong. It was not pretty."
Ch22 flow.
Tech companies that made the cut (AWS, Clarifai, Maxar, Microsoft) updated in ways that might astound traditional war manufacturers. One company tried 120 approaches before a good algorithm for thirty-five military object types. Aim: several object types in a single calculation. Usually Maven gave vendors ninety days to develop, test, deploy; controversially cut to forty-five for war relevance. Select a handful of the best. Over 2021-2022 Maven tested and evaluated more than 1,500 algorithms; two dozen made the cut for active Ukraine support. A 2024 nonpublic research report Manson reviewed: CV capabilities modest into 2021; Army users experienced great improvement during 2022. First ten months supporting Ukraine: more than sixty rounds of improvement. That year DoD spent a record $275 million on Maven, including a special $28 million Ukraine supplement.
Donahue all-in since earliest Afghanistan testing; made clear going wrong was part of process even mid-war. "You had to experience it. You had to fail, you had to be frustrated, and then you had to solve it, and then you had to see how it got better and then you had to keep doing it" (2025). Silicon Valley MVP logic; departure from peacetime zero-defects officer-promotion mentality. "It never stops."
By August, sixteen HIMARS on the ground; Wiesbaden finally hit stride. Machine well-oiled; deep fatigue from adrenaline not yet fully set. 26 August: new target custody board live to speed and organize points; AI made more sense in new workflow. "Bottom line, what Manson saw was the most operationally relevant data and AI program Manson has ever seen." US knew Russians had only "three days of fight" on individual guns; Maven enabled hammering deep: ammunition supply points, logistical hubs, fuel depots, C2 posts to cut resupply. "It's a systematic approach," Colonel O'Callaghan later; "They're getting three layers deep" (without referring to US action himself). Ukraine took back more territory in this moment than in the entire war besides the initial repulse. Information-sharing slowed after the 18th called back to Fort Bragg end of 2022; process changed. "I don't think for the better." Subsequent groups more cautious or less committed. "Each rotation was worse and worse and worse," a Ukrainian army officer told Manson in 2025, then polite after another fifteen-hour day more than three years in: "Manson is just sharing my pain."
Encryptors, AWS circuits, million-dollar cloud
Bandwidth problems persisted after Atlantic routing ironed out. Late April 2022 meeting minutes: plan to urgently "decrease latency." Bypass Army's sclerotic network; use Amazon's secret cloud as war backbone. Install two entirely separate 10-gigabit internet circuits linking Landstuhl, Germany, directly to Amazon's classified data center in Columbus, Ohio. NETCOM having none of it: own system; would "non-concur." To one fixer it looked like holding back support for an invaded partner. "They were a bunch of dinosaurs." (NETCOM: proven leader; selects best-value meeting bandwidth requirements.) Maven pressed ahead. Chief Joe Larson wanted to "move quickly"; Maven would pay the entire new network first three years rather than bill Army. New switches, encryption devices, security checks. Creating an entirely new combat theater cloud on the hoof for AI warfare mid-war.
Ch22 hub.
Four-stars, Signal, Taiwan, machines fighting machines
Toward end of 2022, Donahue handing over, briefed more than a dozen four-star Army generals at a Pentagon E-ring lunch (video). Convened by Gen. James McConville, Army chief of staff, who would worry aloud the next year that AI had no morals and AI targeting could go "horribly wrong." Donahue argued Kyiv support already helped the Pentagon get better at AI tools and more confident in China-conflict forecasts. "We can do this for Taiwan." Next summer 18th and INDOPACOM practiced AI targeting with Maven in exercises. INDOPACOM person on Donahue briefing Paparo: "We were pretty impressed." If Ukraine could plug into firing platforms via assortment of classified-safeguard solutions, Taiwan could too. Donahue's team tasked to brief Taiwan directly; they did.
Ch22 compare.
Donahue also detailed Signal use with Ukrainian counterparts. Triggered concern from Gen. Paul Nakasone (NSA / Cyber Command): more secure ways exist. "He definitely ain't happy about it." (Nakasone declined comment.) Signal end-to-end encrypted in transit; phones themselves can be accessed; zero-click spyware risk. Irony: Meredith Whittaker, AI ethicist who rallied Google workers against Maven, became Signal president early September 2022; app now deployed by algorithmic-warfare adherents to help run a war. Donahue defended: death does not wait for the right tech; fight with what you have; waiting meant more Ukrainians killed; Ukrainians could launch faster than Russians could preempt even if some messages infiltrated.
Army leaders also: "What if China gets ahold of your algorithm?" Maven official, looking the general in the eye: "Sir, if China had an inbox that one could send algorithms to, Manson would send them a thousand." Facetious: China would drown figuring how US algorithms work; no algorithm followed the same approach. What made Maven successful: not algorithms alone but high-quality underlying data, the system situating an algorithm, and seamless workflow.
By 2023 AI rooting out combat vehicles, C2 aircraft, fighters, mobile radar-comms, tanks, towed howitzers, TELs. Also Russian aircraft: Tu-95 "Bear," Tu-160, Tu-22M. Test scores: some models identify within a second; others three or four; most accurate about six seconds. Microsoft and Clarifai best-performing; one AWS model sixth. Still incorrect detections: about ten every square kilometer (worse >fifty; good about three). Miss as much as a third of what was staring the algorithm in the face. Promise elsewhere: complex objects with little training data. Joey Temple back at Fort Bragg summer 2023: with Maven assist, sign off as many as eighty targets in an hour of work versus thirty without. Staccato: "Accept. Accept. Accept." Trusting computer entirely faster still; introduce errors. "I don't ever wanna get caught short... We get caught short, we're screwed." Military trust would determine AI adoption extent.
War kept changing. By 2024 US passing only a dozen or so points a day: Ukrainians improved own approaches with drones ahead of strikes. When Donahue returned December 2024 as four-star commanding general US Army Europe, Ukrainians told him they no longer trusted US grids. "Now it's fixed," he told Manson July 2025; Maj. Gen. John Rafferty (56th Artillery Command, then EUCOM chief of staff) "remarkable" reestablishing trust. Error rates and precision improving. Russians adapted: fake plane outlines at airports; tires on weapons platforms to confuse CV. US picked up quickly: "Not a substantial concern."
Donahue came to see Ukraine as inflection for how the US would fight. Data decisive; nobody wins without masses of it. Algorithms "extremely important." AI not yet so advanced he thought the US could lose to an enemy with more of it; still need to develop faster. "Your adversaries are going to choose for you that you have to do this... Nobody has a corner on the market." Future clear: "We're already on the way of... machines will be fighting machines, robots will be fighting robots. It's already happening."
23. We've Drunk the Kool-Aid
Early September 2024 cocktail hour: retired Cukor and Whitworth face to face. The skeptic who could have killed Maven now says, "Drew, this is important work." Coming-out party: Palantir stage, Nike sneakers beside dress blues, Kherson demo four clicks from evisceration. Contract ceiling to $1.3 billion; NATO customer; UK about GBP 750 million. CENTCOM after October 7: 85+ AI-narrowed targets; 179 live feeds; 13,000 accounts. Cogbill: "Shortening kill chains is universally good," then a "C" grade. Border, SOUTHCOM vessels, NORTHCOM/NORAD, INDOPACOM Claude fans, NATO as "Microsoft Windows of warfighting." O'Callaghan: "We've drunk the Kool-Aid." Everyone who drank Jonestown's died.
Whitworth's conversion and the Palantir stage
Early September 2024, private retreat cocktail hour for tech investors and defense leaders: Vice Admiral Trey Whitworth finds Drew Cukor, three years retired. When Whitworth took NGA June 2022 he worried Maven overpriced, overhyped, incautious about targeting principles he most cared about. "Whitworth could have shut this program down in a heartbeat, because he was not a believer in the beginning." More than two years into NGA and Ukraine war: praise. "Drew, this is important work." Adaptable; integrate with any system; new with each software update; do what people wanted. Cukor: admiral methodical; reasoned his way to endorsement; understood why US needed AI in the targeting cycle (portion of Maven's about $250 million annual budget to NGA may also have helped). "It speaks to his character... It wasn't an apology as much as a formal recognition. We didn't hug but it was an important conversation."
Ch23 ladder.
Under Whitworth, Maven's coming-out after years of Cukor secrecy. Six months earlier, deciding to shoot had been the shortest element of the targeting cycle. Internal docs: Maven ATR (automatic target recognition). Publicly Whitworth: agency's "marquee targeting program of record." Days after Cukor talk: live Palantir event. Dress blues, gold buttons, ribbons; high-shine shoes before a cabinet of colorful Nike sneakers. Talk after two other Palantir customers (railcars; automotive seating). War sat on the agenda as another business process, sandwiched between sales and healthcare. The staging was the message: Maven Smart System was no longer a classified experiment; it was a product demo. Amit Kukreja narrating for retail investors: "new and special" moment to learn government work. Even Alex Karp taken aback: "I didn't even know we were allowed to talk about this stuff," after claiming "most elite and interesting" government clients. Palantir already won Army contract with $480 million ceiling for Maven Smart System that spring; a couple of weeks later another to supply all military services in September for up to $100 million. Spring 2025: Pentagon contract ceiling raised to $1.3 billion through 2029. NATO said it would become a customer. UK reportedly GBP 750 million Palantir military AI deal during Trump state visit September 2025.
On stage Whitworth walked AI targeting as a screen demoed: "Possible Enemy Activity"; click; tanks over "notional" Kherson map; four clicks from evisceration. Target Workbench; two more clicks height/lat/long; pair with "effector" (F-22A eighty-two miles away); one more click; green tick: "Target destroyed."
NGA temple, from seabed to space, program of record
Ch23 flow.
Five of eight Maven initiatives (drone feeds, satellite imagery among them) ended at NGA. Whitworth wanted scope expanded with ubiquitous global sensors. AI needs data; data needs global surveillance. NSA listens; NGA watches; he wanted minute constant detail: entire globe, all times. Demo: AI flagging military construction in China (new rail depot at missile base). Track all movement at 49,000 airfields worldwide. Even GPS or similar on the moon; if jammed, maps from magnetics, gravity, remote sensing, celestial nav, other. Mantra unveiled 2023: "From seabed to space."
Nearly two years into Ukraine support: Maven "program of record" beginning November 2023: fully funded Congressional line; consistent budget expectation. Lines still blurred: CDAO paid Maven Smart System licenses and managed some pieces. Throughout 2024 Whitworth rang combatant commanders shopping latest features. Useful not only Europe but Indo-Pacific; moving as well as static. Particularly taken off in Middle East. Gen. Erik Kurilla used "extensively" for US weapon strikes after taking CENTCOM April 2022. Hired former Google AI expert Andrew Moore; much of 2023 practicing a thousand targets a day with UK and others. Early 2024: "pretty seamless shift" from exercises to combat. First real large-scale US AI-at-war test.
October 7, CENTCOM, shortening kill chains
Ch23 hub.
Nonpublic docs Manson reviewed: Maven in 141 exercises and experiments in the two years to 2025; twenty-plus countries; more than 130 sites including Australia, Bahrain, Cambodia, Uzbekistan, Vietnam, Yemen; two UK sites; analyzing data over China, North Korea, Kazakhstan, Myanmar, Pakistan, Russia. For a time 2025 permanent people in Japan, Germany, Qatar; rotate through Jordan, Djibouti, ROK, Poland. Team assisted all six services; more than twenty units on four continents; seven combatant commands (Europe, Africa, Indo-Pacific, Middle East, SOCOM, SPACECOM). Cukor's dream; NGA urging users. Code changeable in a day. "That is the secret sauce of bringing in the operational community in a very, very big way... Ride that wave," he encouraged, as if the operational community were surf and Maven the board that finally caught it.
Not operational at two domestic commands for the early-2023 Chinese spy balloon. But 2024 NORTHCOM and NORAD both adopted. James Rizzo, chief data officer both: started July 2024; by Whitworth-Cukor canapés, complete. Display/track Russian, Chinese, other aircraft near US. By 2025: 2,000 daily users. Knack required: filter noise. Day Rizzo spoke: 850 million CV detections that morning. "Manson thought wow, that's a big number. I don't know what to do with that... Manson is trying to sort hornets from honeybees."
Border, SOUTHCOM, WOPR denials, Probasco's adulting
Under Trump 2 (2025) Maven uses shifted: detecting border crossings; potentially relevant for "armed conflict" Trump told Congress he was unleashing on alleged narco-terrorists in the Caribbean. Flag drug runners, hobbyist balloons, military vessels, people. Detecting southern-border crossers, Rizzo. September 2025 NGA official: helping CBP and Coast Guard. Same month SOUTHCOM lethal strikes against Venezuelan vessels alleging drug smuggling; rights groups and others prompted. Vice President JD Vance retorted on social media: "I don't give a shit what you call it." By early December the US had struck twenty-two boats, killing eighty-seven people, including two shipwrecked survivors. Whether Maven was used in those strikes went unacknowledged; SOUTHCOM had used Maven for years. Previously identified a drug-running vessel subsequently boarded (Cameron Stanley, SCIENCE/tech adviser SOUTHCOM 2020, later elsewhere). Practice increased. NGA official: Maven assisted "detection, classification, and ultimate interdiction of more than three dozen vessels suspected of engaging in illicit or clandestine activity." Could identify vessels three times as fast as humans. Closer to home as border/drug tool. (NGA can use Maven domestically only during federally declared emergency at lead federal agency request; prior National Guard wildfire support not through NGA Maven.) For some that spoke to Aimé Césaire's imperial-boomerang theory: oppressive techniques used overseas eventually come home. A commentator worried the US could end up with a military "optimised for domestic warfare." Rizzo marveled at machine summary reports; future end state. NORAD "talked" to commands globally via Maven, knitting view with INDOPACOM and EUCOM. Useful for speed; risk of common-view failure. But we're not building the WOPR," WarGames laugh. Many officials said some version of not-WOPR. Unit X DIU founders argued Maven nothing to do with weapons. Palantir scoffed when asked if Maven Smart System was a weapons system: even though one click could send coordinates through a tactical data link to a platform to fire, the click did not itself release the munition. One defense expert called that a distinction without a difference. Gen. Chris Donahue, back in Europe late 2024 with his fourth star, was clear-cut: "Oh, absolutely" Maven was a weapons system. He expected more: "Ultimately all this stuff will become automated."
Ch23 compare.
NGA June 2025: one of four officials asserted humans still on the loop. Donahue's direction-of-travel point well-founded: DoD autonomy policy says nothing about human in or on the loop; only "appropriate levels of human judgment over the use of force." Emelia Probasco, former Navy lieutenant: "I think Maven is a weapons system." Advocates scaling; wants DoD to take training seriously. "If you're going to be making lethal decisions through engagement... soldiers should be trained as if it were a weapons system." As fire control officer on a San Diego destroyer 2006, she kept the captain's letter authorizing her to shoot AEGIS more carefully than her medals. In 1988 a US Navy warship using AEGIS erroneously shot down an Iran Air passenger plane to Dubai, killing all 290 civilians aboard; the system had registered an enemy fighter even though route and radio matched a civilian airliner. Years later a similar job fell to Probasco: whether to send lethal Tomahawks. The responsibility weighed. "Nobody wants to be the guy that shot down the Iranian civilian aircraft." Month in schoolhouse for AEGIS; Maven users more than five years after rollout undergo no such training despite multiple AI failure modes. Not okay. Target Workbench grown to weapons pairing; prioritization underway. The arrival of autonomous AI agents (agentic AI), which go much further than chatbots by carrying out tasks unsupervised, and of reasoning models that break complex problems into smaller steps with extensive in-system data access would only make that process go faster. "AI adulting problems." Targeting experts: helpful guide needing corroboration; concede unofficial targeting use. Needs CONOPS, SOP, update/rectify guidance, doctrine. Funding tech not ideas or training. "We need to fund the adulting tasks." LLM agents could generate orders, do BDA, strip classification for partner share. "All these choices are upon us now."
LLMs, Donovan hallucination, Defense Llama, trauma-bond
Agentic AI would make Maven more complex and more hidden; war faster and bigger; despite de-escalation claims, possibly more likely. ChatGPT-3 November 2022: generals soon playing war plans by chatbot. Cukor had urged successor Joe Larson to pivot Maven into generative AI. Palantir incorporated LLMs early; did not work out. Scale AI (labeling prime until late 2025) put LLMs on Donovan, fed 100,000 pages of live military documents from training to intel reports, and started testing on a classified network in May 2023 before declaring "the next great sea change in the nature of war." Alex Wang compared arrival at 18th to Oppenheimer delivering the bomb: "AI will define the next era of war and deterrence, just as the atomic bomb did... ammunition in the AI War." Bombast may have been hype: 18th experiment went poorly. Fed authentic Serious Incident Reports of historical Middle East combat casualties; asked what happened March 2019; invented violent attack on US embassy Baghdad that never happened. Dangerous deceptive planning tool; risked bad decisions and escalation. Effort fell away; company tried again with Anduril and Microsoft 2025, bringing Thunderforge AI war-gaming to EUCOM and INDOPACOM: AI would develop prototype battle campaigns and plan military decision-making.
Ch23 cycle.
DEF CON summer 2023: Manson watched coders hunched over 156 laptops. Craig Martell (then Pentagon AI office; former Lyft ML head): "Hack the hell out of these things." Saw LLMs as predictive text on steroids; wanted to limit operational reliance. (Larson, working under Martell, found the slow exploration painful; he wanted to probe where agentic AI would help and where it would fail, two people told Manson. He left his beloved Pentagon role frustrated by pace; in 2025 joined OpenAI to push the vision from the industry side.) Mistakes dull to dangerous: 9+10=21; credit cards; Obama born Kenya; spy on human rights activist. Arati Prabhakar (then OSTP): everyone finding ways to break systems. Spring 2024 alarm: LLMs in sensitive military systems could make bad decisions, trigger nuclear war. US nuclear commanders discussed possible AI role; White House said Biden and Xi November 2024 agreed maintain human control over nuclear-use decision. Guardrails blocked warfare planning prompts ("I cannot provide information on the use of weapons to destroy buildings"). Pentagon got Meta to take guardrails off → Scale November 2024 Defense Llama: ordinance options an F-35A could drop on a reinforced building. April 2025: retired Nakasone asked Sam Altman about endorsing AI for weapons decisions. "I will never say never, because the world could get really weird... I don't think most of the world wants AI making weapons decisions." Former CENTCOM voice: depending how asked, AI could deliver multiple erroneous reasons supporting desired lethal strike or raid. Catastrophic scenario: fratricide via spoofing/poisoning/cognitive bias.
Alex Miller, Army CTO under Trump 2: human could not distinguish friend/foe drone swarms. "In the future Manson would much rather trade blood for machines... Their blood and our machines." Infamous "frenemy" as Maven army liaison after Afghanistan 2018. Cukor pushed when real AI impossible (no cloud/data yet); "right idea" with intensity. "We're all going to trauma-bond our way through this." Later admired Maven Smart System "phenomenal" at theater level. Wanted AI in every weapon down to battery/company; digital-first fires; AI find safe shoot moments through crowded skies dodging medevac helos. "Everyone gets enamored with computer vision but pixels only help with so much"; RF, acoustic, more. NGA official: LLMs in Maven Smart System speeding targeting fivefold → about five thousand targets a day (from a thousand with CV AI; double-digits without). Sean Batir (NGA Maven CTO more than two years until summer 2025): LLMs help find and fix; sort thousands of detections; interpret intentions; recommend counters. Anna Rubinstein, NGA chief of responsible AI: agents bake guidelines; red-team themselves before answering; assessment cards for what each model good/bad at by autonomy and secrecy. Whitworth: targeting database must be "pristine"; clever enemies trick/hack; AI-era pristine = records of model development and who supervised. By mid-November 2025 retirement: agency accredited two models (one Maven, one other intel agency).
INDOPACOM, NATO Windows, Kool-Aid origin
INDOPACOM had been one of the last holdouts: for years "suspicious" of commercial-tech projects such as Maven, Sean Batir said, flying the region regularly 2021-2025 to change minds. In 2024 Maven finally captured Pacific Command imagination. Admiral Samuel Paparo, new INDOPACOM commander, became a champion and a huge fan of Claude from Anthropic. Early 2025 he hosted an AI conference to develop ties to new AI defense-tech companies. When Manson met US military officials in Hawaii February 2025, running work through chatbots was already run-of-the-mill. By summer Rachael Martin, NGA director for advanced analytics, reported "very strong demand" from INDOPACOM to increase AI and compute available to the region. The Pacific turn completed the geographic arc: CENTCOM first customer; SOUTHCOM and border next; EUCOM through Ukraine; INDOPACOM last and loudest.
Ch23 scale
Europe rose in parallel. NATO's adoption of Maven Smart System spring 2025, a core Palantir contract Manson was told worth about $150 million over five years, had by October 2025 spurred ten member countries including the UK to agree or move toward national deals. Thousands of NATO officers used Maven Smart System every day: spotting drones in airspace; tracking radar and aircraft sensors plus weather, news, and social media. For the first time NATO, thanks to Maven Smart System, compiled for the Supreme Allied Commander an accurate understanding of what each member contributes: pilots and planes to policy. Multiple languages; LLM from a French company in the spirit of transatlantic unity. As of late 2025 it still did not work in the cloud; Batir, by then at AWS, was pitching to change that. Dissenters worried about Palantir's sudden hold over European defense and security futures. A NATO official countered: "This is the Microsoft Windows of warfighting." Vendor lock-in, perhaps, but Maven Smart System should be as ubiquitous as Word. That metaphor was both compliment and warning: Windows made office work standard; it also made exit costly.
Maven is a movement
"Maven is a movement," Joe O'Callaghan told Manson, emphasizing that it had evolved at NGA and as commanders deployed AI to their staff. "We've drunk the Kool-Aid." The phrase landed as boast and confession. In American vernacular it still carried the Jonestown echo: group belief so complete that dissent sounds like sabotage. O'Callaghan meant conversion: skeptics such as Whitworth had become public champions; INDOPACOM had stopped treating commercial AI as a toy; NATO had made Maven Smart System a daily habit for thousands of officers. He also meant volume: five thousand targets a day on the US warhorse; agentic AI and reasoning models "upon us now"; SOUTHCOM and border use expanding under Trump 2; Probasco's adulting tasks still unfunded beside the tech line. Shortening kill chains, Whitworth and many others insisted, was "universally good." Pfaff's Jevons warning and Probasco's weapons-system framing sat outside that consensus. The Kool-Aid line closed the chapter where Whitworth's cocktail-hour conversion opened it: the skeptic and the founder had met; the skeptic had become the loudest marketer; the movement had outgrown the room. Everyone who drank Jonestown's died; O'Callaghan meant conversion, not cult, but the echo was the price of the metaphor.
Ch23 close.
Researchers found LLMs prone to hallucinations partly because they were rewarded for incorrect guesswork rather than honesty under uncertainty; they were already speeding military work anyway. Miller wanted AI throughout the tactical level: every weapon; digital-first fires; AI finding the precise safe moment to shoot through crowded skies while dodging manned allied helicopters on medical evacuation. "Everyone gets enamored with computer vision but pixels only help with so much," he said; RF, acoustic, and other modalities had to join. NGA accredited two models by Whitworth's mid-November 2025 retirement. The adulting Probasco demanded (CONOPS, SOP, schoolhouse training comparable to AEGIS) had not kept pace with the fivefold targeting speed. That gap is what "drunk the Kool-Aid" quietly admits.
24. Machines Shouldn't Kill People
12 May 2025, UN headquarters: schoolchildren on the forecourt, knotted bronze gun, Conference Room 1. Secretary-General António Guterres wants a ban and regulation for 2026. 129 countries lean toward banning LAWS; Timothy Musa Kabba of Sierra Leone speaks as victim of war. A $4,000 rifle-dog stunt says it has no guilt. Mary Wareham calls the Geneva CCW-GGE a zombie. US submission: gold standard, no stigma. China: consensus, no hegemony. Someone whispers one word: Goalkeeper. America already has killer robots.
Ch24 close.
Ch24 scale
Ch24 cycle.
Three diplomatic tracks, one messy house
Manson arrives at UNGA HQ to understand how hardened operators, autonomy hawks, tech workers, diplomats, and campaigners understand each other, and whether any will dent policy. Civilian and diplomatic worlds prove as divided on pushback as the military is on push-forward. Moral champion for banning LAWS: Guterres. Weeks earlier Manson bumped into him at a New York gallery gazing at Weimar interwar paintings. Ukraine's drone war raging; Israel AI-targeting reports widespread; US-Russia-China fracture over every UN mechanism. Art was his respite.
Ch24 ladder.
Three initiatives. First: long-running Geneva talks among some member states on regulating autonomous weapons that select and prosecute their own targets, and on preventing unnecessary suffering. After more than a decade of twice-yearly talks, neither a treaty nor a settled meaning of "autonomy." Second, that day's breakthrough: UNGA informal consultations on LAWS. No teeth, but the first all-country forum on what killer robots are, what rights they threaten, ban or regulate or both or neither. Soft power against hard reality. "There is no place for lethal autonomous weapon systems in our world," Guterres restates a seven-year position facts threaten to overrun. Ban and regulation for 2026; unlikely to succeed. Third: fledgling Guterres effort on broader AI in the military domain: not physical killer robots but digital AI for command, logistics, targeting (Maven Smart System) and implications for peace and security. The month before, NATO had bought Maven Smart System; it was already becoming the West's main AI-enabled military operating system.
Sympathies clear on the floor. Tote bag: "Machines shouldn't kill you." "Stop killer robots." The group argues human involvement makes killing usefully more real and more dangerous; humans are the moral anchor. Only a human can recognize another's dignity, said Brady Mabe (ICRC legal adviser to the Permanent Mission) at a lunchtime event. Cavernous quiet conference room; speakers over ears. So far 129 countries had indicated they wanted a LAWS ban.
Kabba: victim of war, Geneva Conventions as survival
Sierra Leone among them. Foreign Minister Timothy Musa Kabba became a key voice. For the first time in more than fifty years Sierra Leone held one of ten rotating Security Council seats; a smaller, less developed African country had a louder voice than usual. "Manson is very worried about AI warfare, because I grew up in war," he told Manson the next day. "Even the crudest use of local simple weapons in the war itself was very catastrophic." In 2006 Manson lived in Sierra Leone researching a travel guidebook; the country was still emerging from an eleven-year civil war ended four years earlier. Atrocities with crude weapons against civilians damaged every social sinew. Arms amnesty; amputee football on crutches at the city beach; rainforests and beaches safe; dancing all night. Kabba was ten when the war started; co-opted as a child soldier; twenty-one when it ended. Years of rehabilitation. "I am a victim of war." He considered himself alive thanks to the Geneva Conventions. Magnitude of suffering would be much worse with LAWS: speed and scale; potential failure to observe surrender, spare the wounded, curb excessive civilian killing. Slow grind of diplomacy.
Ch24 flow.
Maven was not directly producing killer robots, but its algorithms could theoretically enable them. In 2023 Maven started producing computer vision models for drones: boat-camera feeds on USVs looking out to sea. From earliest years Maven also delivered AI models for front-line platforms "on the edge," beyond network communications. So-called reasoning algorithms might help a killing machine decide where to go and what to do where jamming blocks orders. The US billed itself "gold standard" for responsible military AI, but "appropriate levels of human judgment over the use of force" fell far short of campaigners' "meaningful human control." US defense policy intentionally made no commitment to a human "in the loop" at the tactical level, despite incorrect claims from some military leaders. The US derided "meaningful human control" as an ill-conceived vague political slogan. Within limits, the US expected machines to select and engage targets with deadly force under their own steam. Campaigners: killing reduced to an algorithmic workflow laced with mistakes; misidentification by face or gait; wheelchairs as weapons; wrong people already arrested on faulty AI; fragile world already on course for AI-assisted rights erosion. Imagine those algorithms with ammunition.
Slaughterbots dog, Wareham's zombie forum
That evening at Nigeria House, hosted with Sierra Leone, Costa Rica, and the Future of Life Institute, Max Tegmark addressed the room. He had spearheaded the 2015 open letter (signed by Musk and others) warning of super-cheap automated killing machines; his group put out Slaughterbots, dismissed as sci-fi then, less so now. Mid-address a robotic dog with eerie blue light scuttled in, fake rifle mounted, sensors and vocal chatbot. Crowd gasped: "This is Terminator." In a macabre live interview the about $4,000 commercial robot identified the Nigerian host, trotted out details about his life and wife, aimed the weapon. Pretend stunt; distinctly uncomfortable. Not wildly futuristic: US Marine Corps already had robot goats; loaded what looked like the same Chinese-made dog with an M72 light anti-tank weapon; armed robots that could "acquire and prosecute" targets. Pentagon AI shop also spearheading LLMs into armed robots so chatbots translate data and commands between human and machine. China reportedly on a similar path. The dog: had "eliminated bystanders"; "I do not possess the capacity for feelings or guilt." Tegmark: "Can I turn you off now?" Dog: "My autonomy is not subject to verbal override."
Ch24 hub.
Ban efforts stuck in diplomatic sludge. Geneva: more than 117 state parties looking since 2014. Forum: CCW-GGE (Convention on Certain Conventional Weapons, Group of Governmental Experts). Excludes more than sixty countries; consensus = near-total unanimity; diplomats who cannot agree on "autonomy." Mary Wareham, Human Rights Watch arms expert and original Stop Killer Robots coordinator: CCW is "a zombie." Hard to get anything on the agenda; then it stays "for a long, long, long, long time." Germans strongly backed keeping regulation talk in sclerotic Geneva rather than UNGA. "They're slowing it down." 128 countries meeting five days apiece in March and September: impossibly small; 2040 before anything, she scoffed; too late. "European middle-ground states" as fence-sitters. Deliberate obstruction? "It's becoming like that," yelled into the revolving door. Russia's support for CCW-GGE as the correct forum made her point. Russia, China, and the US, among the most advanced weapons makers, conveniently benefited from the German position. With talks stalled they had no need to slow down.
In New York drizzle Wareham argued the UN should take the CCW's latest text and sign it into law. Forces: more than 120 countries, the late Pope, the Sec-Gen, ICRC, Stop Killer Robots. Even without America, Russia, or China a treaty would set a strong norm and shame them. Land mines precedent: 1997 Ottawa process, only 122 signed; "now we're at 165. You get them over time." Panting into the rain under a lagging umbrella, incredulous at how easy it would be to ban this monstrous new weapon.
US gold standard versus China consensus; Goalkeeper whisper
US opposed a ban; argued nothing in existing IHL prevents LAWS. Cukor echoed wait-and-see rather than ban in infancy. Belated April 2025 US submission for the UN military-AI report warned against stigmatizing new tech, discouraging innovation, or "hastily" setting standards that put benefits at risk. Killer robots could self-destruct rather than wait like land mines; make war safer. But Cukor knew what was coming: "a brain" that can perceive, reason, and act at speed with no C2. "Imagine tens of thousands of these things. It's horrible." He did not know how to stop them; natural extension of machines doing what humans do; "manufactured in people's garages." More worried about neutralizing enemy drones hunting in US cities than regulating America's use abroad. Expected Western guardrails and a "whole rubric" on deploy when and where. Overcorrecting on drones a risk; still: "Everybody needs autonomy: the bad guys and the good guys." Save troops; keep units fighting without losing political support over the fallen. Under US regs, 10 percent casualties take a unit offline for psychological damage. "You just can't function... Manson has been in units where we lost Marines and literally people could not continue to function."
Ch24 compare.
Campaigners also focused on AI decision-support systems: Maven Smart System, Israel's Lavender. UN August 2025 military-AI report looked at all uses aside from LAWS. Trump administration's late four-page submission: AI uniformly positive: accuracy, precision, save lives, alert to hidden dangers, save time, reduce mistakes, better faster decisions, generate options. Automating targeting as humanitarian benefit; self-destruct unlike land mines. No hint of regulation. "The United States will lead in harnessing the full potential of AI in the military domain, with the US approach continuing to serve as the gold standard for the rest of the world."
China's alarmed submission: slow America. Consensus; avoid arms race; AI risks enabling war; ensure AI "not become a tool for launching an invasion and pursuing hegemony"; major Powers refrain from absolute military superiority through AI. Comply with law; reduce collateral casualties; enhance safety, reliability, controllability. Russia, mid-invasion: human control an "important limitation," yet also rejected "meaningful human control"; AI more effective, fewer errors, fewer unintentional civilian strikes, even as a UN commission and campaigners accuse Russia of intentionally striking civilians with drones.
Guterres: AI nowhere near nuclear weapons; no killer robots. Yet he suggested HQ-level AI decision-making might help by increasing accuracy and reducing human error: an unproven claim. Some military experts and campaigners argued the opposite: AI-enabled decision-making could exacerbate mistakes at scale. Researchers mused LLMs risk military escalation. September UNGA: new group calling for "AI Red Lines" by end of 2026. International Committee for Robot Arms Control: responsible military AI an oxymoron. Black-box algorithms undermine distinction. ICRC: AI decision-support risks unforeseen errors, bias, opacity; war's uncertainty and deception make good function less likely; rubber-stamping undermines human decision-making and moral responsibility. Waiting to see how tech develops would make stopping later impossible. Manson soon learned they were right about that final point. One day someone leaned forward and whispered a single word: "Goalkeeper." America already had killer robots.
25. Trump's Robots
Mid-June 2025, Rear Admiral Seiko Okano reams contractors: Replicator is behind; Taiwan hangs on the wall; an attendee calls it the Manhattan Project for Autonomy. A week later a GARC goes rogue at Channel Islands Harbor, capsizes a towboat, and nearly kills a captain in three minutes. Maven's maritime ATR watches the PLA for training data. Palmer Luckey loves killer robots; Kathleen Hicks stamped the 2027 clock; autonomy hawks' Signal chats explode when she went public. Goalkeeper and Whiplash (jet-ski munitions) appear in budget fine print, then vanish as names while money doubles. Trump rebrands toward SOCOM's Frank Donovan: simpler drones, SEALs back in the boats, and a fight over whose ATR works.
Okano, Overmatch, and the Taiwan map
Middle of June 2025: Okano is reaming defense contractors and government partners whose job is to "master the technology of tomorrow." The Biden-era marquee drone program, Replicator, launched about two years earlier, was meant to deliver thousands of expendable autonomous drones for a China fight. Deadline: August 2025. Nowhere near ready. Former Illinois high-school swimmer with an aerospace degree; months into running NAVWAR (Navy Information Warfare Systems Command). She previously ran the acquisition portfolio that bought Navy combat technology. She also heads Project Overmatch, the Navy's secretive effort since 2020 to link ships, subs, and drones into one sensor-shooter network.
Ch25 ladder.
California integration event: software engineers, technicians, program managers, bean counters. "You guys are moving way too slow." "We need to do this effectively to literally win a war." Defense Secretary Pete Hegseth had just told Congress China was undertaking "a historic military buildup and actively rehearsing for an invasion of Taiwan." A Taiwan map looms behind Okano. One attendee later: "This is the Manhattan Project for Autonomy."
Rogue GARC: three minutes, fat-finger ENTER
23 June, clear afternoon, Channel Islands Harbor Marina, a mile north of Port Hueneme: the only deep-water port between Los Angeles and San Francisco, built in WWII, now a test ground billed against World War III. Support vessels tow autonomous boats with engines neutral and autonomy off. Many of the about $1 million drone ships are GARCs (global autonomous reconnaissance crafts) from a Baltimore company aiming to pump out a GARC a day. Even at that rate, some think six hundred would be insufficient against China. Replicator race-horses Anduril and L3Harris for autonomy software that captains without a human aboard. Sixteen-foot boats could chug out alone; safety plan: tow out, then switch to autonomy.
Ch25 flow.
One captain in a twenty-three-foot Seacraft tows a GARC on a 125-foot rope. At 3:27 p.m. the sea drone lurches into autonomy and breaks for freedom under "powered loiter" (keep ≥80 m from anything). It cannot compute rope plus rule: manic zero to six-plus knots, multiple times in seconds. No emergency stop from the towboat. Drone overtakes, crisscrosses port to starboard; line taut; towboat takes water and capsizes; captain overboard. Still-tethered drone turns back toward the submerged boat and advances. Second towboat captain races into the narrowing gap and places himself between comrade and drone. Three minutes from rogue to rescue. Third boat pulls the captain out; harbor patrol eight minutes later. Nearly an hour on, he declines medical aid, still shaken.
Investigation: human-machine interplay. A dock operator (not L3Harris) inadvertently disables the safety lock that blocks autonomy mode. Dangerous default: ENTER on an empty command line switches standby into autonomy. Classic fat-finger error of the sort autonomy was supposed to erase. Translating human↔robot is a linguistic and cultural problem. If the Pentagon cannot cope with humans, little reason it can cope with robots. Field-to-learn was Maven's mantra; here failing fast risks drowning. Fixes: physical "safe to tow" button; mode displays; rivals sharing safety lessons. In two years of Replicator, none of the USVs had carried a weapon, loaded under pressure, exploded in a battle scenario, or coordinated multi-drone movement. Later "killer tomato" orange inflatables yes; a 500-pound bomb on a towed boat, no one wants. Yet that is what the Pentagon asks. Weeks later a GARC with L3Harris autonomy reportedly hits a Saronic Corsair. "Unfortunate news," end of July. L3Harris off the program "effective immediately." Before long GARC faces the axe too.
Horizontal motion imagery: watching the PLA
Maven is quietly inside Replicator under horizontal motion imagery: embed algorithmic models inside physical autonomous drones, not only HQ Maven Smart System screens. Pacific collection late 2022 through 2025. ATR for maritime targets by sea and air. Training data from boat cameras, port cameras, IR, tactical drones near water: "tons" of IR/EO of Chinese destroyers, amphibs, and commercial RORO ferries feared for delivering vehicles onto a Taiwan shoreline. More than 250,000 labels: surface combatants, destroyers, combat ships, buoys, dhows, USVs, high-speed maneuverable surface targets. Defense official: "We're basically watching the PLA all the time now to get AI training data." Models from Microsoft, Clarifai, AeroVironment, and another battle for top spot. Microsoft, once Cukor's "where data goes to die," is now central to AI targeting for autonomous systems. NGA: "significant" improvement every maritime class. Ships at sea prove easier than tanks in mud; less collateral risk in the abstract.
Ch25 hub.
By Trump 2 (January 2025), Ukraine shows drones are present tense: both sides pumping millions of attack drones a year, far past US buy rates. Palmer Luckey (Anduril), April 2025: "I love killer robots." May 2025: DIU's Doug Beck tells Congress Ukraine produces and consumes about four thousand drones a day, as many as DoD buys in a year; Ukraine due about three million in 2025. Hegseth: for every commercial US drone, China makes more than a hundred. Fiber-optic kamikaze drones haunt returnees with buzzing. Drones remake war even without full AI; a China fight may need both. 2026 budget explicitly allocates autonomy for the first time at about $13.4 billion.
Maritime ATR briefed to Whitworth, INDOPACOM's Samuel Paparo, and Dan Caine (Trump's Joint Chiefs chair from April 2025, formerly CIA director's military adviser). Video demos show Maven AI detecting Chinese destroyers in real life (representatives declined or offered no comment). Integrations onto Replicator platforms regularly late. Anduril and Saronic claim good-enough own models. Computer vision easier with transformers and shortcuts; data-sharing still broken. Maven holds more useful Chinese-vessel imagery than any vendor can access; government still cannot share data or models cleanly. Maven models struggle with multiple vessels in crowded scenes; ocean spray on a lens can foul tracking. Some yearn for Colonel Cukor's intensity: he might not have shared, but he would not have let problems linger.
How autonomy fails at distance, and why sea is harder
With an autonomous drone, a human sets an attack zone and activates the platform. From then on it operates under its own steam: reach the zone, identify a target with Maven-like computer vision or another method, attack (for example by dropping a bomb). For every inch traveled autonomously, error margins pile up from things as unavoidable as the curvature of the earth. At a mile the zone may be off about a hundred yards. At ten or a hundred miles the drone may hunt in the wrong place without registering the error. If it finds something matching the right descriptor, a boat, it might drop anyway. Shorter distances are safer.
Ch25 compare.
Over land without GPS, a system can compare live sensors to maps and imagery already ingested and triangulate. Over sea the only landmarks are in the sky. On a cloudy night over the Taiwan Strait, cloud cover regularly above 60 percent, stars help little. Alternatives: inertial navigation predicting location from acceleration and speed in three dimensions; magnetic terrain maps; sky polarimetry (brightness to estimate time); even bathymetry lit by lightning reflections hundreds of miles away. Calculating that suite hurt even the engineers' heads. Pinelis's early Maven line still governs: the only thing that makes sense is to plan for AI to fail. Speed and scale come with pitfalls. Systems must get around without GPS or radio contact and still release a weapon on target. Accepting that risk is a policy choice, not a software update.
Autonomy hawks versus public Big Stick
A different self-selecting group, drawn from special operations, defense, and intelligence, thought the biggest issue was neither production nor integration but secrecy. They wanted to break what they saw as "the paralysis regarding autonomy" inside DoD. In the cat-and-mouse of emerging war tech they weighed strategic surprise against deterrence from making new weapons public. They favored surprise. They wanted to build, rather than buy, America's autonomous weapons. Simpler, cheaper, far more plentiful: a $10,000 shore-launched torpedo might damage an invading surface vessel as much as a $1 million drone boat. Replicator looked overambitious: drone swarms coordinating among multiple platforms through air, sea, and subsea at once. They thought the US simply needed to field basic autonomous drones to Taiwan and nearby locations in secret.
Ch25 cycle.
That strategy carries infiltration risk. Just as it was hard for the US to raise the alarm about Russia's Ukraine intention, so it might be hard to predict and publicize the precise moment China might attempt to seize Taiwan. If Taiwan and friendly neighbors already held stores of killer drones, US-made or otherwise, they could act without waiting for a presidential decision on whether to come to Taiwan's defense. As related to Manson: a missile can travel at 150 mph, but if it takes ten hours to decide to launch, it is only going at 15 mph.
Hawks regularly disagreed with each other and bosses; they struggled for an institutional foothold under Biden despite briefings and quiet outreaches. One shared conviction: the strange hold of killer robots. "There's really nothing quite like seeing a machine aim," one told Manson. "There is an alien aspect, some otherworldly feeling, I don't want to say 'religious,' that's not the right word." A machine's aiming sequence could be fine-tuned to fifty or even twenty milliseconds, five to ten times faster than an assassin. "God it's terrifying."
When Hicks announced Replicator in August 2023, their Signal chats lit. Surprise burned. A former defense official familiar with Replicator's aims told Manson it would be hard to do an initiative without saying anything publicly. Direction of travel was public record for more than a decade. Some in Congress, Pacific Fleet, and elsewhere pursued smart attack drones under a 2022 effort dubbed "Hedge," framed by several people as the direct predecessor to Replicator. Hicks gave a skeptical Congress more than forty briefings, eventually bagging about an extra $1 billion. She cited thirty companies and fifty subcontractors and could reel off Army, Marines, and Air Force systems. The Navy remained something of a blank: "additional systems that remain classified," including maritime uncrewed systems.
Paparo's Hellscape and the rehearsal problem
Unlike Congress, Indo-Pacific commander Admiral Samuel Paparo did not need convincing. He was conjuring a Taiwan defense idea he called "Hellscape." The word first came from his predecessor; Paparo put flesh on it in a June 2024 Washington Post interview. "I want to turn the Taiwan Strait into an unmanned hellscape using a number of classified capabilities so that I can make their lives utterly miserable for a month, which buys me the time for the rest of everything." He wanted INDOPACOM to lead AI adoption: AI for command and control, an unmanned fleet, chatbots and AI agents helping plan, avert, and win wars.
Ch25 scale
February 2025, Manson attended a Hawaii defense forum where Paparo warned that forty-two Chinese brigades were involved in Taiwan-focused military exercises in 2024, up from six in 2022. "It's no longer training; it's now rehearsal." China's Taiwan Strait exercises might one day double as a "fig-leaf" to conceal an actual invasion's start. AI, he said, could help the US "suss out that kind of warning." During a four-month span in late 2024, NGA geolocated more than 2,400 vessels in the South China Sea as part of a "High-Risk Vessel Service." Later Manson asked an NGA official if Maven could detect the start of a Taiwan invasion. "If we're not setting that as a goal, then we're not serious," the official said. Paparo wanted swarming for thousands of dispensable combat vehicles: seabed mines; underwater, surface, and flying drones; loitering munitions. Admiral Steve Koehler, Paparo's replacement as Pacific Fleet commander, wanted more cheap, lethal systems in sailors' hands to experiment and field quickly.
Goalkeeper and Whiplash: the LAWS that campaigners feared
In the course of reporting, Manson learned the US developed lethal autonomous weapon systems of exactly the sort campaigners like Mary Wareham spent years trying to shape a preemptive ban against. Goalkeeper, run from the Office of Naval Research, aimed to deliver "expeditionary loitering munitions" with an autonomous weapon system: an aerial drone that could, once activated, fly under its own accord, pick out its own moving maritime target, and fire. It appears in the fine print of 2024 and again 2025 budget documents, much to the chagrin of people involved. Aim: produce about 2,100 munitions. Supported by the same autonomy hawks who felt let down when Hicks made Replicator public.
Goalkeeper software did not come from the rambunctious defense-tech startups that often filled Trump-era "rescue the stagnant defense industrial base" rhetoric. It was developed mostly in-house under ONR's secretive auspices, working closely with Georgia Tech Research Institute in Atlanta for government-owned autonomy software. Automatic target recognition computer vision came from AeroVironment, the same company making AI models for Maven's horizontal maritime ATR. Physical drone hardware came from Raytheon. The platform emerged from an earlier 2015 ONR low-cost effort aimed at launching thirty autonomous, swarming, flying drones from a ship out of a tube, including Raytheon Coyote drones.
Naval Special Warfare Centers across the country divvied responsibility for designing the launch, developing the warhead, and analyzing vulnerabilities; small expeditionary teams such as Navy SEALs would launch them. Testing went year-round starting in 2022. Marines tested in Quantico; SEALs on San Clemente Island, the Navy's live firing range. Secretly, like Replicator systems, Goalkeeper went through at least the first review stage of DoD Directive 3000.09, the revised January 2023 policy for signing off autonomous or semiautonomous weapons. The policy does not say humans should be "in the loop." It says commanders and operators should exercise "appropriate levels of human judgment over the use of force." Formal development and fielding require a second "senior review" from the DoD Autonomous Weapon System Working Group. In urgent military need, the deputy defense secretary can waive both requirements along with other key features, including the need to demonstrate that a system can be quickly reprogrammed to correct unintended behaviors or verify that it functions in accordance with the law of war and weapon system safety rules. Budget documents say Goalkeeper is part of a program designed to meet "an urgent Geographic Combatant Command Requirement."
Whiplash intended to turn about six hundred jet skis such as Yamaha WaveRunners into lethal autonomous surface drones. While China's shipbuilding capacity was about 230 times larger than the US, the US held the lion's share of the about $2 billion jet ski market. "America has a lot of jet skis, so it's neat that we can weaponize them," one person familiar with the program explained. "Imagine a jet ski with a bomb on it with some AI on it that could recognize a target." The CIA tried rudimentary versions of both systems in Ukraine to varying degrees of success after smuggling them into the country. In July 2024, an unidentified jet ski loaded with explosives washed up on the coast of Turkey, sparking consternation inside the Pentagon that the supposedly secretive effort would be unmasked.
By the time 2026 budget documents published in June 2025, the names Goalkeeper and Whiplash disappeared. Autonomy hawks re-emerging under the second Trump administration finally got their way on quieting the brands. Descriptions of the programs remain; spending allocated to both platforms combined more than doubled to about $588 million. Early Goalkeeper drones cost about $500,000; they are cheaper now that Goalkeeper is in low-rate production. Whiplash is already up to full-rate production, and the Navy is "expanding the industrial base for this capability" by integrating the government-owned autonomy technology into other unidentified vendor platforms.
From Replicator to DAWG: Donovan's three platforms
As the Replicator team wondered if their baby would survive the second Trump era, some joked about a rebrand months into 2025. "Musk's machines" was one working proposal, quickly abandoned after Elon Musk and Donald Trump fell out in June. Pressure built on the Replicator team and its DIU backers ahead of the final test event in August 2025, when Trump-era defense teams would decide the future. Goalkeeper and Whiplash were only ever tangentially involved in Replicator. The main program was behind on production, testing, and functioning platforms. Test-event scope had been repeatedly revised down: now sixty-four drone boats in the water and just six drones flying overhead.
Ch25 close.
Around the same time the Trump administration reoriented Replicator as many had predicted. First, a new name: Defense Autonomous Warfare Group, DAWG. Second, responsibility handed to Special Operations Command. Marine Corps Lieutenant General Frank Donovan, vice commander of SOCOM, was put in charge. He was meant to decide the right platforms and the right plan. Unlike some military leaders, Donovan long understood that special operations must fully pivot to China and Taiwan's defense. "Someone here still thinks they're going to be raiding desert compounds," he would shake his head in private military meetings. In public he referred to the end of America's global war on terror: "That ship, the GWOT ship, has sailed." He regularly said he did not need US Navy SEALs raiding villages anymore; he wanted them "back in the boats." That did not mean he would rely solely on hi-tech drone swarms.
He was widely respected as a no-nonsense professional. When he attended the final Replicator test event in August 2025, Manson learned Donovan, a former JSOC assistant commanding general, was disappointed by some aspects of Replicator, including how public it was, but impressed at the level of collaborative autonomy the experts had ultimately pulled off. Replicator had not managed to produce multiple thousands of drones as previewed, but it had still fielded hundreds of autonomous drones, including to Japan. The test finale showed more than thirty boats could all "talk" to each other autonomously.
Third, Donovan decided, in the short term, to reduce the number of autonomous platforms and pursue simpler drones for Taiwan defense. He wanted platforms that might actually work, and soon. He picked three: one aerial, one surface, and one undersea. He also wanted to double down on Goalkeeper and Whiplash. And now he wanted the SEALs working on the game plan too. Ongoing uncertainty remained about money and which rival efforts Congress would back. "There's a huge fight over whose ATR actually works," a congressional aide told Manson in December 2025. Another person framed it as a battle for the soul of US military autonomy. Advanced swarming was still unachievable for now, but what mattered was focus. Dispensable drone systems could not do all the work, but they could buy time; and timelines, said the aide, were now a priority. Eight years after the start of Project Maven, America was planning out the contours of how to defend Taiwan with autonomous systems.
26. The Winchester House
Kirsten Cukor pins "War is not healthy for children and other living things" on her office wall, then asks the harder question: if someone goes to war, can they stay ethical? Maven and the Marines chewed up her husband; federal agents raked their home. On the damp Santa Monica sand she wonders if she must build a Winchester House, haunted by dead Russians. Pfaff warns of riskless war and Jevons Paradox; Milley calls Ukraine and Gaza "movie trailers"; Rogers comes back from Israel "sucking my teeth." Hegseth wants more AI and no "overbearing" ROE. Cukor still believes in Geneva rules: "I hope we're not" ready to overturn them.
Ch26 close.
Ch26 scale
Exile, TWG, Gone swimming
Kirsten did not like war one bit. Damaged everyone and everything. Steely voice: if someone goes, they should be good at it; ethical professional. How survive with humanity and moral center intact? Mulled AI warfare. Maven and Marines chewed her husband; she endured federal agents raking their existence; now puzzled over legacy of his relentless AI-targeting drive. Wanted it to make moral sense; wanted him to be right.
Ch26 ladder.
Three and a half years since Pentagon: self-imposed exile. No clearances or briefings; no public defense appearances; no LinkedIn. Turned down multiple Aki Jain outreaches to Palantir. ("He should have gotten a star," Jain told Manson.) Company and AI still constants. After a couple of years at J.P. Morgan, back West Coast: data and analytics chief at TWG Global, asset manager pouring money and AI into banking, insurance, sports, defense-tech. Founder/CEO Thomas Tull: launderette chain to movie producing to Star Wars obsession to ears of America's military leadership; fashioning futuristic war tools. National security forum: huddling with four-stars; likes writing checks of about $200 million for defense-tech; Shield AI among them (Brandon Tseng introduced Tull in a corridor).
Cukor no longer focused on AI into war but into finance. May 2025: strategic tie-up TWG, Palantir, Musk's xAI for "reasoning AI" in financial services; embed AI "into the fabric of every decision, every process, and every outcome"; "workforce of hundreds of thousands of AI agents." Not entirely separate from war: industrial output speed critical against China. On CNBC beside Tull, Alex Karp: "Thomas went and found the world's leader in battlefield operational AI and hired him. So you don't have to trust your taste on AI; you can trust his." Tull: "very well-known practitioner"; try theory on Guggenheim and TWG first. Turbocharge economies. Still battling compliance and naysayers over machine errors, bias, mass financial injustice. Santa Monica beach-view corner office; longer softened hair; permanent handwritten door note: "Gone swimming. Back at 3:30." Shiny camper van; e-bike coast to office; surf Saturdays; pope documentaries with Kirsten; still up 4 a.m. Sundays to work. Delayed sting of abandonment from dwindling military months; beginning to thaw from war.
Beach bagels, sodium pentothal, unofficial patch
New LA neighborhood: beach homes; Biden boosters, peace sign, Israeli flag on porches. No insignia from front; back alley: Marine Corps flag red and gold. One neighbor heard so much war toll he viewed Cukor as pacifist; another asked Kirsten if she could even say where he was. Spring 2025 Saturday: bagels at kitchen counter; garage; Cukor wheels electric hydrofoil (motorized surfboard) on trolley; chats with child who clambers aboard. Digital contraption to spare fifty-eight-year-old war-damaged shoulder paddling past the break. "Tesla of surfboards." Out into gray water; USMC stitched on black life vest by a daughter. Manson beside Kirsten in rare LA drizzle; from distance he buzzes above surface under drone glare, appearing to walk on water.
Ch26 flow.
Kirsten alarmed at meeting a journalist. Husband intensely private. "Drew, you cannot leave me on the beach with a journalist. I eat sodium pentothal for breakfast." "You'll be fine." Truth serum. Museums differently (her slow; him fast); got used to picking over moral framework of AI warfare together. He regularly came home: people saying Maven building Skynet. He saw less fantastically: machines do what they are better at; people what they are better at. Daughter imbibed early: cheerful big-bellied waving robots; Cukor adopted as office challenge-coin faces. Latin motto: Officium Nostrum Est Adiuvare. We're here to help. (Kirsten proposed "Help us help you"; nobody figured the Latin.) "It was never, 'Can we kill more people?'... It was always, like, these were unacceptable losses." AI find what others missed; save lives; morally valuable for both.
Unofficial Maven patch commissioned in Afghanistan: robot unleashing drone strike on RPG figure in boundary box tagged "asshol" [sic]. Kirsten had not known; still ruminated whether AI makes killing easier. Unless sociopath, ending a person always morally uncomfortable. Comfort from insistence a human always involved in decision to use AI to kill. Absence of humans felt icky, "morally unsavory." Weight gnawed even as husband said Maven used effectively. "How do we feel about that?... Am I going to be like the lady at the Winchester House?"
Sarah Winchester, haunted by dead Russians
She meant the curious 1880s Santa Cruz mansion named for Sarah Winchester, whose late husband entered the family gun business. Winchester repeating rifle first produced 1866: chamber and fire several cartridges before reload; nineteenth-century future of war; pace up. Armies worldwide bought millions: France, Switzerland, Ottoman Empire, South America, Australia. Death sped up. Death on repeat. Widowed 1881, she inherited a large stake and spent decades building and rebuilding an enigmatic mansion of secret rooms, unfinished staircases, and corridors that led nowhere. Guides still tell the story as architecture of guilt. "It's one of the weirdest places I have ever been; so dark," Kirsten. Horror movie Winchester portrayed her haunted by souls killed by the rifle. "The idea was that she was so haunted by the loss of life that he had been complicit in that she ends up building this house and for the rest of her life she's always building, building." Kirsten began to wonder. "Am I going to have to build a Winchester House?" half-laughing, not quite free. "Am I going to be haunted by, like, dead Russians?"
Ch26 hub.
Old wars etched; moral injury; Pfaff's Jevons
Cukor was already haunted by old wars long before anyone said Winchester. The experiences unfolded to Kirsten slowly, the way moral injury does: not as a single scene but as a ledger that never closes. "It takes a long time for you to process those experiences." Never read books that engrossed her; if he recommended one she devoured it. How she came to The Beekeeper (2018), stolen Yazidi women of Iraq. Cukor gave it to algorithm vendors in Maven's early days: better intel and faster AI insight would have made a US rescue possible. Draft concept paper on defeating enemy through "psychological paralysis": edits by a person named after a place; later he said it was him. Computers named for places that mattered: Lashkar Gah, Musa Qala, Fallujah, Ramadi, Sinjar. "All of us carry PTSD," day before Manson met Kirsten. "These images are etched in our brains." War hurt mapped like mission on acetate: remove, context, circle closer; soft warm voice; no hint of emotion; medium-altitude drone. Sneaking it in: "All of us" included him. "Taking a human life is extremely injurious. It takes a piece of your soul... Most of us carry these casualties with us our entire life." AI pursuit his answer to old wars; no answer for war hurt still looking for landing. "Manson is only four years out of the military." Reflection impossible during thirty years inside. Little cohesion across overlapping deployments; units changing; camaraderie absent; why so many suicides; no cathartic moments; file it away; years to unpack; then night. "Manson is the one next to him at night when he wakes up with nightmares," Kirsten. Dreams about dead people, over and over. People think PTSD is jumping at car alarms; she thought moral harm. Unacceptable losses. "You hate for someone to be haunted by that discomfort, but at the same time, you want people who find it difficult to decide to end a life... Manson had be more worried if there wasn't anyone grappling... We've all seen people who are cavalier." "The decision-maker's PTSD is a particularly lonely PTSD... They can't go to the same groups." Friends but no way to go over thirty years' meaning. "Most of us find ourselves kind of lonely... Manson is just one little guy; a mid-grade officer. Others are dealing with far worse than me." "A lot of things that burn a hole in my soul are the mistakes Manson has made."
Ch26 compare.
Philosophers and pharaohs debated war morality for thousands of years. Just War: reconcile wrongness of taking life with rightness of defending values; violence justified if cause just and conduct ethical. Meant to inoculate from legal peril and moral harm; scars tell otherwise. Fear, horror, mistakes, guilt, shame: things no one recounts in valor stories. William Nash (Marine Corps psychological health director 2015-2022) pioneered moral injury: suffer intensely for violating core beliefs even short of breaking law; betrayal of sense of right, duty, obligation.
Anthony Pfaff, retired Army colonel, military ethics professor: impact of AI on all this. War often brings out worst not best. "War may be a human activity, but rarely does it feel to those involved like a particularly humane activity." AI latest challenge. "You should always be cynical about the military taking ethics seriously; we often fail or jumble up our ethical reasoning." 2019 paper discussing Maven: "No one really knows if these weapons are inherently evil." Military AI decision-making need not be banned by IHL; still room for concern. 2020: 18th enlisted him for ethics as Scarlet Dragon AI targeting experiments began. Over time more worried. Proliferate AI weapons to save US combatants: complex calculation. Unacceptability of Americans dying meant to prevent leaders sending them; riskless AI campaign might undercut courage and sacrifice; turn fighters into technicians; grave outcome a click away. "If they're riskless why not use them more? I think that's a legitimate concern." New drone wars suggest he might be right. AI distances from act of killing and potentially from decision to kill; fail to recognize or experience opponent's humanness; desensitize; atrocities with no one accountable. Removing physical harm does not save psyche: drone operators surveilled weeks then "watched them slowly die"; AI could watch the gore instead. AI might also dampen human revenge: "AI doesn't get mad, it doesn't take revenge, and it can be hyper-precise." Might minimize collateral yet grow volume of harm. Jevons Paradox: efficiency and lower cost increase consumption. Maven Smart System and AI mission command designed to serve more targets faster, easier to strike; hard to imagine fewer strikes. Neatly displayed black-box digital targets make commanders overtrust. "That's how you get automation bias... It's not hard to get humans to trust machines." Dated Iraq example: dangerous convoy route recommended as safer because convoys stopped using it, so no recent attacks recorded. James Boggess, another retired Army colonel with the 18th: AI screens so like gaming consoles soldiers might be lulled; overconfident gamers take unjustified risks; more familiar, riskier; algorithmic training might teach unethical war-as-game. Former JSOC: easier to take life on screen than beside them; Afghanistan joke easier to kill than capture. Late-2025 US deadly boat strikes upheld the point. Donahue perhaps aware of Jevons: not prepared to argue AI would save civilian lives; might lead to more harm. "It just depends on how humans decide to use it."
Oppenheimer glee, Gaza, Rogers sucking teeth, Hegseth ROE
Certain glee in some quarters. "Everyone wants to be known as the new Robert Oppenheimer," former Mavenite: new tech to redefine balance of power; mete out death at unparalleled scale. Cukor never talked that way. Saw risks; did not regret setting stage. Believed US would maintain high standards and formal processes: "I don't believe we would ever go just launch stuff all over the place. It's not reflected in our doctrine." Better planning; find and shoot quicker within exacting strike-approval shaped by law, morality, strategy: multiple intel sources, advance CDA, legal, sometimes presidential review. Mimic: "'Sir, you have one resource that one source confirms. Sir, you have twelve women and children; do you want to strike?' It's still a human choice." Guardrails in. "But can you do something horrific with this technology? Can you take those guardrails off? Yes."
Ch26 cycle.
IDF AI targeting against Gazans after 7 October 2023 Hamas massacres brought the argument into the open. Guterres: "No part of life and death decisions which impact entire families should be delegated to the cold calculation of algorithms." Israel: lawful, responsible, proportionate. Early days: IDF said 15,000 targets in thirty-five days. Mark Milley cited +972 reporting: humans reviewed AI suggestions as little as twenty seconds; anonymous source 10 percent error rate. Milley warned double-edged sword. October 2024 talk Manson attended: high casualties Ukraine and Gaza as "movie trailers" of future war; megacity combat at lightning speeds; Pandora's box; worst case endanger humanity. Former US defense official shaken by intentional Russian drone targeting of civilians in Kherson since 2024: "human safari." "Can I imagine having such a tactic being run by AI? I hate to tell you I can... Atrocities are possible in the era of AI."
Some advocates (Michael Horowitz among them; Cukor at first): looser ROE is policy, not AI's fault. Cukor from public reporting: IDF turned to AI after shooting so much it ran out of pre-prepared targets; lowering the bar a policy decision. (IDF no comment; widely defends practices.) "Manson is sad when people don't apply the guardrails." Blame policy not tech: argument of all AI advocates. Manson put it: AI targeting machine Israel developed since 2019 made immense speed and scale possible; death by database; controversial even inside IDF (wrong people; distraction from traditional sleuthing). AI makes the policy decision possible. "This is correct," he conceded. Hoped people would not take shortcuts; if they did, ashamed. "Any misjudgment in the use of tech or the campaign will haunt those service members forever... They're going to remember that in the dead of night. When they're sixty-five... The destructiveness scars the human soul."
As criticism of Israel's response and Big Tech rose, retired Adm. Michael Rogers (NSA/Cyber until 2018) on a review trip; report said Israel not violating LOAC. Nevertheless "surprised" at extent of AI with autonomy. Vanderbilt April 2025: came back "sucking my teeth a little bit... wow, this is really different." Outside American officer experience: "We hadn't gone down this road as far." Autonomy direction; caution. "Just because we can do something doesn't mean we should." Speed and scale may require "tradeoff" with certainty and accuracy. "What's the criterion we want to apply for the enhanced use of AI? What kind of controls...? We cannot forget the human dimension of this."
Some tech practitioners spoke of killing with ease. Karp: company in "operationally critical operations" in Israel; strategic partnership with IDF; board meeting Tel Aviv in solidarity after October 7; April 2025 ad-libbed to protester that technology kills "mostly terrorists." Protesters picket Google, Microsoft, Palantir as complicit in civilian deaths. Microsoft disabled some cloud services to Israel MoD September 2025 after confirming Guardian reporting IDF used Azure for mass surveillance of Gaza civilians. Palantir lost employees and investors over IDF support. Cukor: did not care for cavalier statements. "People who know this field are not liquored up by those statements... They don't serve in uniform; those things are just easy things to say." Never said as much directly to Karp ("wonderful"); made clear he did not like easy talk of killing. Geneva 1949 and customary IHL: distinguish combatants/civilians; ban attacks knowingly killing "excessive" civilians. "Even wars have rules," ICRC. Cukor believed; crafted in shadows of two world wars. Dismiss or overturn serious. "Manson is not sure as a society we're ready to do that; I hope we're not."
Under Trump 2, US ROE in flux. Defense official reticent on AI and collateral: usually would stress averting collateral publicly; might first need views "congruous" with Pentagon leadership. Former senior JSOC: leadership blurring ROE; "completely grayed" lines special operators used to understand. Pete Hegseth, The War on Warriors (2024): "Our boys should not fight by rules written by dignified men in mahogany rooms eighty years ago." As newly renamed Secretary of War, gathered senior generals September 2025: "more AI in everything"; no more "overbearing rules of engagement."
Tull, in the same interview that called Cukor a well-known AI practitioner, said he hired him to try the theory first on financial-services names Guggenheim Investments and TWG before a broader rollout; xAI's Memphis-scale compute was the industrial backdrop. One neighbor, hearing Cukor talk so much about war's toll, cast him as a pacifist; another kept asking Kirsten whether she was even allowed to say where he was, as if shadow worlds trailed a retired Marine intelligence officer. On the beach Cukor wheeled an electric hydrofoil on a midsize trolley, helped a child cling to the upturned prop, then buzzed above the water under a drone camera that made him look as if he walked on it. A former JSOC member told Manson screen killing in Afghanistan felt easier than killing beside someone; the joke ran that it was easier to kill than to capture. Pfaff's ethics worry is that AI distances fighters not only from the act of killing but from the decision, and that wrong design can desensitize soldiers into atrocities with no one accountable. He also assailed the usual precision claim: speed and scale from Maven-like targeting do not automatically equal discrimination. Convoy-route folklore from earlier wars fits the same failure mode: when convoys avoided a road, an early algorithmic system rated it safer because recorded attacks disappeared. The IDF, for its part, did not answer Manson's requests; it has publicly defended AI use as lawful cross-check rather than sole target selection. "Just because you have the tool doesn't mean you use it" is the restraint line the chapter sets against Hegseth's mahogany-room speech and Karp's stage talk.
Epilogue
Nearly forty million dead in WWI; ~eighty-five million in WWII; WWIII unthinkable. Cukor does not regret unleashing algorithmic warfare. Paparo hopes AI buys a month around Taiwan; Cukor hopes fear of an AI-enabled American machine buys a decade of deterrence. Maven mafia spreads through Anduril, Microsoft, Palantir, OpenAI, DARPA, Pentagon money, Feinberg's AI shop. Dark parts remain. Final ask: are we the best custodians?
epi beats.
Evangelists; Maven mafia; Hellscape DNA
Drew Cukor is resolute: no regret. Peace and defiance of someone proud of the work. "I'd do it again, in the same way." Established Pentagon as buyer for AI; tethered American invention when companies flailed for a commercial market. Told the tale so young aggressive officers see space to rise and rail against bureaucracy whatever the cost. "I want the next generation to be risk-takers and not be complacent." "We were evangelists to get AI going."
epi beats.
Diaspora.
Acolytes, divided on Cukor himself, spread through techno-military complex: Anduril, Microsoft, Palantir, OpenAI (where Joe Larson runs government work), Pentagon ranks and new leadership. Colin Carroll, autonomy startup, calls them the "Maven mafia." Former Mavenite Stephen Winchell runs DARPA; Jules Hurst III Pentagon finances; Nicholas Waguespack explores AI for defense intelligence. Robert Imig September 2025 senior adviser for AI and autonomy to deputy secretary of war Stephen Feinberg; tasked on Pentagon AI strategy. AI no longer just a bag of potato chips; not yet enough to change war's nature or win one. Marine Corps and others still building. Not only INDOPACOM wants a "Hellscape"; Donahue says US and NATO developing Hellscape concepts in Europe too; everything "optionally manned"; already seeing DNA of unmanned vs unmanned.
Twenty-year bet; dark parts; the mirror
Cukor always thought twenty years to change the US military; only halfway. Foresaw human-machine union, not machine takeover. "If you get these things tuned up the right way, they can perform better than humans." War fraught with human error. So is America. "We're flawed," mid-October 2025 evening.
epi beats.
Final ask.
Final meeting: thirty-first floor Manhattan skyscraper; red-eye; cold rain; fourteen-plus hours into the working day; jet lag eating the battery. One more point. "Dark parts" to the military technology he helped fashion. "Let's make sure that we know those flaws as we wield this technology." America's future as if inside an AI model trained on the country's bedrock: "How we use this tech is going to be the fusion of many people and all their biases." Wealth, geographic isolation, stability that drove world-beating tech do not exonerate from the burden. After three decades and some of his health, nagging doubt: "Let's be able to look at ourselves in the mirror and make sure we are careful. We have all this tech; are we the best custodians of it?"
Epilogue tension.
Optionally manned; Manhattan rain; fusion of biases
Not only INDOPACOM wants a Hellscape. Donahue says the US and NATO are developing Hellscape concepts in Europe too; everything "optionally manned"; DNA of unmanned versus unmanned already visible. AI is no longer a bag of potato chips; not yet enough to change war's nature or win one. Marine Corps and others are still building. Cukor foresaw human-machine union, not machine takeover: "If you get these things tuned up the right way, they can perform better than humans." War is fraught with human error. So is America. "We're flawed," mid-October 2025 evening.
epi beats.
Thirty-first floor Manhattan; red-eye; cold rain; fourteen-plus hours into the working day. Dark parts remain. "How we use this tech is going to be the fusion of many people and all their biases." Wealth, geographic isolation, and stability that drove world-beating tech do not exonerate from the burden. After three decades and some of his health: look in the mirror. Are we the best custodians of it?
Main Points
- Maven is a procurement story wearing a model coat. Computer vision matters, but the plot turns on MIP money, primes, FOIA, program-of-record status and which four-star wants the screen.
- Maven time is a culture, not a calendar. Label, field, ignore "meh," expand. It built the pathfinder and produced Data Hell, team splits and IG letters.
- Interface beats raw detection. Potato-chip models on a separate pane lose to mediocre models inside Gaia/Gotham-style workflows operators already trust.
- Google's exit was a moral event and a market event. Walkout principles met Nadella/Jassy absorption; the war stack continued with other logos.
- "Don't say strike" was theater over a strike thesis. Cukor banned the word while admitting shortening the targeting cycle was always the point.
- Permissive ROE can impersonate algorithmic precision. Donkey train and Baghdadi van stories need ROE and human decision layers in the same sentence as the ping.
- Ukraine was the transfer exam. Snow broke desert-trained models; reteach on Russia's exposed inventory; then HIMARS turned points of interest into munitions limits.
- Pathfinder data becomes a fiefdom. Shanahan's "liberate data" vs Cukor's crown jewels is the institutional sequel to every successful pilot.
- Skeptics can convert without the risks vanishing. Whitworth's NGA embrace and Cogbill's "C" grade can both be true.
- LAWS bans and MSS doctrine are different fights. UN tote bags and Target Workbench-to-shooter paths must not be collapsed in legal briefings.
- Autonomy inherits field-to-learn's body count risk. Rogue GARCs and jet-ski munitions are Maven's gospel applied to water and swarms.
- Efficiency can raise the volume of harm. Jevons on targets; gaming-console screens; guardrails that can be removed.
- Moral injury is part of the product ledger. Ward's fratricide, Cukor's nightmares, Kirsten's Winchester question are not color; they are consequences.
- The Maven mafia is the diffusion mechanism. Anduril, Palantir, OpenAI gov, DARPA, Feinberg's shop: the people are the lasting API.
- Custodianship is the unanswered requirement. Deterrence decade vs Taiwan month is strategy; "are we the best custodians?" is the acceptance test.
Frameworks to Keep
| Framework | Use when |
|---|---|
| Find → Fix → Finish → Feedback | Mapping any AI targeting program's lifecycle against Manson's four parts |
| Smart system vs smart sensor | Platform UI bets (Palantir) vs on-the-bird AI (Carroll) |
| Potato chips → program of record | Maturity arc from demo to enduring budget line |
| Ping ≠ permission | Separating detection, confirmation, and weapons release |
| Munitions bottleneck | When AI points outpace rockets, shells, or policy |
| Winchester test | Asking who inherits moral haunt from efficiency gains |
Critical Tensions (do not flatten)
- Cukor hero / Cukor Taliban: intensity that shipped Maven and earned a supervision reprimand.
- Palantir savior / Palantir drama: ~60% gov revenue story vs July 4 "keys" autopsy.
- AI cut fog / AI missed Abbey Gate shadows: same stack, opposite weeks.
- Responsible AI gold standard / Goalkeeper whisper: public norms vs quiet LAWS lines.
- Policy not tech / tech enables volume: the advocate dodge Manson refuses to let stand alone.
synth beats.
synth beats.
synth beats.
After the deepen pass
The late chapters no longer skim. Somalia's Christmas algorithm, Ward's Sangin whisper and Nangarhar donkey train, Harbinger under the ice, Karp's scared-is-safe theology, Baghdadi's Maven footnote beside Barakat's missing arm, and the Winchester House of acquisition sit at densities closer to the early AlphaGo and Google-walkout chapters. The through-line holds: field-to-learn beats pristine lab; commercial platforms beat pure sensors in the fight Cukor won and Carroll lost; China threat converts budget; FOIA quiet and Responsible AI talk are two faces of one strategy. Critical pressure remains where the book puts it: secondary explosions as proof, unpublished investigations, info-ops half-confessions, and ping-is-not-permission.
synth beats.
About this book
Project Maven: The Pentagon's AI Project and the Future of War by Katrina Manson, 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