Insight Editions · No. 283
The Singularity Is Nearer
When We Merge with AI
Ray Kurzweil
Ray Kurzweil restates his forecast that compounding price-performance will carry AI, medicine, energy and the brain onto one curve, with a Turing-test AI by 2029 and a merger of human and machine minds by 2045.
A study guide to The Singularity Is Nearer by Ray Kurzweil. Shared free for personal reading; not for sale and not endorsed by the authors or publishers. Please read and buy the original.
The Book on One Page
Kurzweil argues that one engine, the law of accelerating returns, has driven information technology up about 16 orders of magnitude in price-performance since 1939 and is now spreading from computing to biology, energy, manufacturing and the brain. It will give us human-level AI by 2029, brain links and medical nanobots in the 2030s, and the Singularity in 2045. We will not be replaced by AI but will merge with it, multiplying our intelligence millions-fold, if we manage the transition and its perils wisely.
The argument in brief
- Six epochs, one trend. The universe's history is a ladder of information processing: physics, DNA, brains, technology, then the merger of brains with computers, and finally matter saturated with intelligence. Each information technology makes the next one cheaper to build, so progress compounds; computation per constant dollar has risen about 2 × 10¹⁶ since 1939 (introduction, chapter 1, appendix).
- Intelligence is computation applied to data. Symbolic AI stalled and Minsky's Perceptrons froze neural nets for decades; connectionist networks modelled on the hierarchical, self-modifying neocortex won once compute became cheap. A valid Turing test by 2029, and nanobots adding cloud neocortex in the 2030s (chapter 2).
- Identity is pattern, not matter. Consciousness cannot be proved from outside; Kurzweil adopts Chalmers's panprotopsychism (complex information processing awakens a latent capacity) and treats identity as a continuous pattern, so gradual replacement keeps you you, while a separate copy is a new person (chapter 3).
- Life is getting exponentially better. Poverty, literacy, violence, health and clean energy have improved for two centuries while most people believe the opposite, because news and evolved biases hide gradual gains. AI will turn food, housing, energy and medicine into information technologies, and diligent people reach longevity escape velocity around 2030 (chapter 4).
- Jobs will go; abundance and a safety net will cushion it. Automation has replaced most jobs several times since 1800 while employment, pay and leisure rose; nonskilling AI removes whole roles, starting with 4.6 million US drivers; GDP misses the value of free digital goods; a safety net that grows under every party becomes UBI or its equivalent in the early 2030s. Kahneman's fear of a violent transition is answered with abundance and falling violence (chapter 5).
- Medicine becomes exact. AI drug discovery, AlphaFold, simulated trials and superhuman diagnosis in the 2020s; Drexler-style molecular assemblers under broadcast control and repair nanobots in the 2030s, ending the age-linked rise in death risk and pushing life far past 120 (chapter 6).
- The perils are real but manageable. Nuclear war, engineered pandemics and gray goo each have defences; misaligned superintelligence (misuse, outer and inner misalignment) is the primary danger, met by AI-assisted alignment, broad distribution of power and stronger institutions (chapter 7).
- The weak joint. Kurzweil's own sceptic, Cassandra, shows that fast, unregulated AI and slow, regulated brain implants run on different clocks; if the link slips a decade, machines take the jobs before we can merge, and Kurzweil concedes the point (chapter 8).
Chapter map
| Ch | Title | One-line payload |
|---|---|---|
| Intro | The Final Approach | Singularity still dated 2045; one dollar buys 11,200× more compute than in 2005 |
| 1 | Where Are We in the Six Stages? | Six epochs of information; a valid Turing test in 2029 opens the Fifth |
| 2 | Reinventing Intelligence | Connectionism plus cheap compute re-creates the neocortex; nanobots extend it |
| 3 | Who Am I? | Panprotopsychism; identity as continuous pattern; replicants and gradual uploading |
| 4 | Life Is Getting Exponentially Better | Two centuries of gains, hidden by bias; four bridges to radical life extension |
| 5 | The Future of Jobs: Good or Bad? | Nonskilling AI, uncounted digital value, safety net to UBI by early 2030s |
| 6 | Health and Well-Being | AI drug discovery and simulated trials, then assemblers and repair nanobots |
| 7 | Peril | Nuclear, bio, nano defences; superintelligence the main risk; use AI to align AI |
| 8 | Dialogue with Cassandra | AI and brain links run on different clocks; a slip is a real risk |
| App | Price-Performance Sources | How 84 years of computations per dollar were measured |
Kurzweil's timeline, scored in October 2026
| Date | Prediction | Status | Evidence, Oct 2026 |
|---|---|---|---|
| Ongoing | Computation per dollar keeps rising exponentially | On track | AI-chip performance per dollar +49% a year (Epoch AI) |
| Mid-2020s | Immersive VR/AR interlaced with reality | Behind | Headsets niche; first display glasses 2025 |
| Late 2020s | 3D-printed clothing, near-free vertical-farm food | Behind | Vertical-farm bankruptcies 2023-25 |
| End of 2020s | Simulated trials augment human trials | Early | FDA plan to replace animal tests with AI models |
| 2020s | Robotaxis spread; driving automated | Ahead | Waymo ~500,000 paid rides a week, 10+ US cities |
| 2020s | AI-designed drugs reach late trials | On track | Insilico's rentosertib in phase III (Sept 2026) |
| 2029 | AI passes a valid Turing test, exceeds all human skills | On track | GPT-4.5 judged human 73% in short tests |
| ~2030 | Longevity escape velocity for the diligent | Behind | No therapy shown to slow human ageing |
| 2032 | Solar supplies all electricity (doubling every 28 months) | Behind pace | Solar 8.8% of world power in 2025 |
| Early 2030s | UBI or equivalent in developed countries | Behind | Pilots only |
| Ongoing | Nuclear arsenals keep shrinking | Reversing | Stockpiles rising; New START lapsed Feb 2026 |
| Early 2030s | Neocortex connected to the cloud | Behind | ~26 people with Neuralink medical implants |
| 2030s | Atom-by-atom manufacturing; medical nanobots | Early lab | First positional carbon mechanosynthesis (2026 preprint) |
| 2041 | Renewables generate all electricity | On track | Wind and solar 17.3% of power (2025) |
| 2045 | The Singularity: intelligence multiplied millions-fold | Too early | Frontier AI capability doubling every 4-7 months |
The Final Approach
Kurzweil restates the thesis of The Singularity Is Near (2005): converging exponential trends in computing, biology and nanotechnology will let us merge with AI and multiply our intelligence millions-fold. His date for the Singularity is unchanged at 2045, and he argues the evidence since 2005 shows we are on schedule, now in the steep part of the curve.
The thesis in one chain
Kurzweil names three trends accelerating at the same time: computing gets cheaper, human biology becomes better understood, and engineering works at ever smaller scales. As AI gains ability and information becomes easier to reach, we fold these tools ever more tightly into our own biological intelligence. The end state is physical: nanotechnology will extend our brains with layers of virtual neurons in the cloud, giving us millions of times the computing power biology supplied. That expansion of intelligence and consciousness, too large to grasp from here, is what he calls the Singularity.
What "Singularity" means, and does not
- Borrowed term. In mathematics a singularity is an undefined point in a function (dividing by zero); in physics it is the infinitely dense centre of a black hole where the usual laws stop working.
- A metaphor only. Kurzweil insists rates of change will not become infinite: exponential growth never reaches infinity, and even physics has no account of truly infinite mass inside a black hole.
- What the metaphor captures. Our inability to comprehend the shift with today's intelligence. His reassurance is that cognition will be enhanced fast enough for us to adapt as it happens.
Why "nearer": the record since 2005
In 2005, 2045 lay forty years (two generations) away; the forecast was abstract for most readers and critics called it overoptimistic or impossible. Kurzweil's case that progress has kept pace:
- Devices. Smartphones and social media went from almost nothing to all-day companions connecting most of the world.
- AI. Algorithmic advances plus big data produced breakthroughs ahead of expert expectations: Jeopardy!, Go, self-driving, essay writing, bar exams, cancer diagnosis. GPT-4 and Gemini turn plain-language instructions into code, lowering the human-machine barrier.
- Biology. Sequencing a human genome now costs about 99.997% less; neural networks simulate biology to drive medical discovery.
- Brains. Direct brain-computer connections are starting to appear.
The engine: the law of accelerating returns
The law of accelerating returns: information technologies get exponentially cheaper because each advance makes the next stage easier to design. Headline number: since 2005, one dollar buys about 11,200× more computation, inflation-adjusted. Moore's law (shrinking transistors) is only one paradigm of this law; the trend held before transistors and should outlive them. Kurzweil calls it the trend that "defined the modern world" and says almost every breakthrough in the book depends on it.
The chart is limited to programmable computers for comparability, but estimates for earlier electromechanical machines extend the trend to at least the 1880s.
The road map to 2045
Early 2020s We enter the sharply steepening part of the curve; the march becomes a sprint.
2020s AI that seems convincingly human; simple brain-computer interfaces as common as smartphones; a digital biotech revolution cures diseases and extends healthy life; job disruption and misuse risks.
2030s Self-improving AI and maturing nanotechnology fuse humans and machines, raising promise and peril together.
2045 The Singularity: intelligence expanded millions-fold.
Today is probably closer to the first superhuman AI than to How to Create a Mind (2012), and closer to the Singularity than to The Age of Spiritual Machines (1999). A baby born now will be graduating college when it arrives. In 2005 he wrote that we were in the "early stages"; now he says we are entering the culmination.
The stakes are framed both ways: meet the scientific, ethical, social and political challenges and 2045 transforms life for the better; fail and survival itself is in question.
How the book unfolds
- Mechanism and history. How the Singularity happens, set in the long story of humans reinventing intelligence (ch. 2).
- Identity. What creating sentient technology means for consciousness and the self (ch. 3).
- Well-being trends. Exponential gains across many measures of human welfare (ch. 4).
- Jobs. Automation's real harms, and why we are not in competition with AI (ch. 5).
- Biology. Defeating aging, then augmenting the brain (ch. 6).
- Peril. Pandemics, self-replicating nanomachines and AI risk, with mitigations (ch. 7-8).
The "law" of accelerating returns is an empirical regularity, not a physical law, and the chart plots only the best achieved machine each year, which flatters the trend. The 11,200× figure over about 18 years implies a doubling every 16 months, a rate Epoch AI's Feb 2026 data on AI chips does not quite match (about 1.7 years).
Per Epoch AI (Feb 2026), AI-chip performance per dollar rises about 49% a year, while frontier training compute has grown about 5× a year since 2020, so spending, not just cheaper chips, is driving AI. Verdict: on track for compute; hardware price-performance alone is slightly behind Kurzweil's 16-month doubling.
Where Are We in the Six Stages?
Kurzweil treats the history of the universe as six epochs of information processing, each built from the information processing of the one before, so intelligence evolves through an indirect chain of other processes. Humanity is in the Fourth Epoch; passing a valid Turing test by 2029 opens the Fifth, in which brain-computer interfaces merge our minds with digital intelligence.
The six epochs
The framework comes from The Singularity Is Near, where Kurzweil described information as the basis of consciousness. Each stage is the substrate for the next.
- First Epoch: physics and chemistry. Atoms formed a few hundred thousand years after the big bang. Protons packed in a nucleus should fly apart under electromagnetic repulsion, but the strong nuclear force binds them; without it, no atoms and no evolution. Billions of years later atoms formed molecules able to encode elaborate information, with carbon (four bonds, against one to three for many nuclei) the most useful building block.
- Second Epoch: life. Several billion years ago molecules became complex enough to specify a whole organism in one molecule: DNA, which let living things evolve and spread.
- Third Epoch: brains. DNA-described animals grew brains that stored and processed information themselves; the survival advantage pushed brains toward more complexity over millions of years.
- Fourth Epoch: technology. Humans paired higher cognition with opposable thumbs to turn thoughts into complex actions and built tools to store and manipulate information, from papyrus to hard drives, extending perception, memory and judgement.
- Fifth Epoch: merger. Brain-computer interfaces fuse biological cognition with digital speed and memory and let us add many more neocortical layers, unlocking more abstract thought than we can now imagine.
- Sixth Epoch: computronium. Intelligence spreads through the universe, converting ordinary matter into computronium: matter organised at the ultimate density of computation.
- Biological brains: roughly one cubic inch of brain matter added per 100,000 years.
- Digital computation: price-performance doubling about every 16 months.
- Neurons: about several hundred cycles per second, against several billion for digital chips.
- The 2030s mind: nonbiological capacity thousands of times the biological part.
- 2045: minds extended many millions-fold.
The fine-tuning aside
Kurzweil reads fine-tuning into the First Epoch. A world that allows complex chemistry is extremely improbable: with gravity a shade weaker, no supernovas would forge life's elements; a shade stronger, and stars burn out before intelligence appears. One constant alone had to sit in a very narrow range. He even puts "Whoever" in scare quotes as the designer of these rules, without developing the point.
The Turing test as the gateway
- The forecast. 2029 for a passed Turing test (text conversation indistinguishable from a human), first in The Age of Spiritual Machines (1999), repeated in 2005.
- What passing means. Mastery of human language and commonsense reasoning.
- The rules. Alan Turing proposed the idea in 1950 but never said how to run it. Kurzweil's bet with Mitch Kapor sets rules much stricter than most interpretations.
- The 2020 precondition. He expected a broad range of AI achievements by 2020 as a stepping stone; he cites Jeopardy!, Go, radiology and drug discovery, and Gemini and GPT-4 spreading across domains.
- Playing dumb. A passing AI must hide its abilities: instantly solving any maths problem would expose it. So a Turing-level AI is already beyond the best humans in most fields.
Kurzweil places us in the Fourth Epoch, with technology already producing results we cannot fully understand on some tasks, and says the remaining gaps in Turing-test skills are closing at an accelerating rate.
In the 2030s we connect the upper neocortex to the cloud, so AI becomes an extension of ourselves, not a rival; the nonbiological part soon offers thousands of times the capacity of the biological part. By 2045 our minds expand millions-fold, a speed and scale that justifies borrowing physics' word "singularity".
"Several hundred cycles per second" is the neuron's theoretical maximum; in chapter 2 Kurzweil himself cites an average of 0.29 firings per second. The fine-tuning argument (and the scare-quoted "Whoever" designed the rules) is contested: the anthropic principle and multiverse ideas answer it without a designer.
In March 2025 UC San Diego researchers Cameron Jones and Benjamin Bergen ran three-party, five-minute Turing tests: GPT-4.5 with a humanlike persona was judged human 73% of the time, more often than the real humans. Those chats are far shorter than the two-hour interviews of the Kurzweil-Kapor bet, which remains unresolved. Verdict: on track, arguably ahead, for 2029.
Reinventing Intelligence
Kurzweil tells how AI moved from hand-written rules to connectionist neural networks that copy the layered, self-organising design of the neocortex, and why cheap computation made that approach win after 2010. His central claim: intelligence is largely a function of computation applied to data, so exponential price-performance makes a valid Turing test by 2029 and cloud-extended brains in the 2030s close to inevitable. The chapter ends with how nanobots will add layers of virtual neocortex, the step that opens the Fifth Epoch.
What reinventing intelligence means
Kurzweil frames human history as the later half of a cosmic story of evolving information-processing paradigms. Our part of it is the move from animals bound to the brains genetics supplies toward beings whose thinking is no longer limited by biology. The 2020s begin the final phase: rebuilding natural intelligence on a faster digital substrate, then merging with it, which is how the Fourth Epoch gives birth to the Fifth.
The chapter's route, which these notes follow:
- Origins. The birth of AI and its two rival schools, symbolic and connectionist.
- Biology. Why one school won, explained through the cerebellum and the neocortex.
- Deep learning. How current AI is re-creating neocortical powers.
- The gap. What AI still lacks, and how much computing a brain really needs.
- The test. How we will know when AI has reached human level (the Turing test).
- The merger. Brain-computer interfaces that add virtual neocortex and expand intelligence millions-fold.
The birth of AI
Turing makes thinking testable
In 1950 the British mathematician Alan Turing (1912-1954) published "Computing Machinery and Intelligence" in the journal Mind and asked whether machines can think. The idea of thinking machines goes back at least to Talos, the bronze automaton of Greek myth; Turing's contribution was to make it empirically testable. His imitation game (now the Turing test):
- Set-up. Human judges exchange text messages with an AI and with human foils, without seeing who is who.
- Content. Judges may ask about any subject or situation they like.
- Verdict. If after a set period the judges cannot tell the AI from the humans, the AI passes.
Dartmouth names the field
Turing's move turned a philosophical puzzle into a scientific programme and drew researchers in. In 1956 the mathematician John McCarthy (1927-2011) proposed a two-month, ten-person study at Dartmouth College (Hanover, New Hampshire). Its founding conjecture: every aspect of learning or intelligence can in principle be described precisely enough for a machine to simulate it. The stated aims were machines that use language, form abstractions and concepts, solve problems reserved for humans, and improve themselves.
- The name. McCarthy proposed "artificial intelligence". Kurzweil dislikes it because "artificial" suggests the intelligence is not real, but concedes it has stuck.
- The outcome. The core goal, machines understanding problems posed in natural language, was not met in two months. Kurzweil notes that this goal is roughly the same as passing the Turing test, still the open problem.
- ~300,000 AI researchers and practitioners worldwide in 2017 (Tencent).
- 22,400 experts publishing original research, ~4,000 highly influential (2019 Global AI Talent Report, Jean-François Gagné, Grace Kiser, Yoan Mantha).
- 496,000+ AI publications and 141,000+ patent filings in 2021 (Stanford HAI).
- $189 billion global corporate AI investment in 2022, 13× a decade earlier.
Forecasts converge on 2029
Kurzweil has held 2029 for a passed Turing test since The Age of Spiritual Machines (1999), when many thought it would never happen. Expert opinion moved toward him:
| Source | When asked | Forecast for human-level AI |
|---|---|---|
| Kurzweil | 1999 onward | 2029 |
| AI expert survey | 2018 | around 2060 |
| Metaculus consensus | early drafts of the book | 2040s-2050s |
| Metaculus consensus | May 2022 | 2029 (exact match) |
| Metaculus consensus | later | as early as 2026 |
The last row lets Kurzweil joke that he is now technically in the "slow-timelines camp".
The AI effect: why progress feels smaller than it is
Breakthroughs keep arriving sooner than experts expect and with little warning. In October 2014 MIT's Tomaso Poggio said describing the content of an image was among the hardest things a machine could do and needed another cycle of basic research, at least two decades away. The next month, Google unveiled object-recognition AI that did it. Asked by Raffi Khatchadourian of The New Yorker, Poggio fell back on doubting it was true intelligence. Kurzweil presents this as a shared human habit, not a personal failing: before AI reaches a goal it looks singularly human; afterwards it looks trivial. Real progress is therefore larger than it seems in hindsight, one reason he stays confident in 2029.
Two schools: symbolic and connectionist
The reason breakthroughs come suddenly lies in a theoretical split as old as the field. As a high-school student in 1964 Kurzweil met Marvin Minsky (1927-2016), co-organiser of Dartmouth, and Frank Rosenblatt (1928-1971). From 1965 he studied under Minsky at MIT, who taught him there are two ways to automate problem-solving.
The symbolic approach
Symbolic AI writes down, as rules, how a human expert would solve a problem.
General Problem Solver (GPS), 1959. Built at the RAND Corporation by Herbert A. Simon, J. C. Shaw and Allen Newell, GPS could in theory solve any problem expressible as well-formed formulas (WFFs), applying one WFF (essentially an axiom) per step and chaining them into a proof.
- The logic. It is ordinary algebra generalised: knowing 2 + 7 = 9 and x + 7 = 10 proves x = 3. The same reasoning tests definitions: a prime has no factors except 1 and itself, 11 divides 22 and 11 is not 1, so 22 is not prime.
- The ambition. Starting from the most basic axioms, a machine can in principle search every combination, which is what human mathematicians do more selectively.
- The search problem. With 10 axioms available per step and 20 steps to a solution, there are 10²⁰ (100 billion billion) paths.
- The hardware gap. The DEC PDP-1 (1959) ran about 100,000 operations/s; a Google Cloud A3 virtual machine (2023) about 2.6 × 10¹⁹. A dollar buys roughly 1.6 trillion times more computing than when GPS appeared, so problems that would have taken tens of thousands of years now take minutes.
- The workaround. GPS used heuristics to rank candidate solutions. They worked often enough to support the belief that computers could solve any rigorously defined problem.
MYCIN, 1970s. An expert system that diagnosed infectious diseases and recommended treatment. In 1979 expert evaluators judged it as good as or better than any of the physicians compared.
Anatomy of a MYCIN rule. One rule chains seven conditions: the infection needing therapy is meningitis; it is fungal; no organisms were seen on the culture stain; the patient is not immunocompromised; the patient has been in an area endemic for coccidioidomycosis; the patient's race is one of three listed groups; the cryptococcal antigen in the spinal fluid was not positive. Conclusion: suggestive evidence (certainty 0.5) that cryptococcus is not the cause. No single rule decides anything; reliability came from combining thousands of them.
By the late 1980s expert systems used probability models to combine many sources of evidence, so a large rule base could make reliable decisions within a constrained problem.
The complexity ceiling
The symbolic approach's fatal limit is the complexity ceiling. Correcting a MYCIN error might fix that case but trigger three new errors elsewhere, keeping the range of solvable real-world problems narrow. Kurzweil's arithmetic treats each subset of interacting rules as a possible failure point: n rules give 2ⁿ − 1 non-empty subsets.
| Rules | Possible failure points |
|---|---|
| 1 | 1 (the rule alone) |
| 2 | 3 (each rule, plus the pair) |
| 5 | 31 |
| 10 | 1,023 |
| 100 | ~10³⁰ |
| 1,000 | more than a googol cubed |
Each new rule adds more subsets than the last. Even if only a tiny share of combinations misbehave, there comes a point (it varies by system) where adding one rule to fix a bug is likely to create more than one new bug. That point is the ceiling.
Cyc, 1984 onward. Probably the longest-running expert-system project, Douglas Lenat's Cyc (from "encyclopedic") at Cycorp aims to encode all commonsense knowledge: facts such as a dropped egg breaking, or a child with muddy shoes annoying parents. These millions of unspoken assumptions are written down nowhere. Because Cyc stores them as symbolic rules, it faces the same ceiling.
The connectionist approach
In the 1960s, weighing Minsky's account of symbolic AI, Kurzweil saw the case for connectionism: networks of nodes whose intelligence comes from their structure, not their content. Instead of smart rules, dumb nodes are arranged to pull insight from data.
- Strength one. They can find subtle patterns no programmer would think to encode.
- Strength two. They solve problems without our understanding them. Even flawless rule-writing would be capped by our ignorance of which rules are best.
- The cost. They tend to become black boxes that give the right answer but cannot say why. That matters for high-stakes uses (medical treatment, law enforcement, epidemiology, risk management), hence current work on transparency or mechanistic interpretability. Kurzweil leaves open how well this will work as models grow.
Symbolic. Smart rules written by humans. Transparent and auditable, but bounded by what experts can articulate, and it hits the complexity ceiling as rules interact.
Connectionist. Dumb nodes whose structure extracts patterns from data. Solves problems without understanding them and finds patterns no programmer would think of, but tends to be a black box; hence today's work on mechanistic interpretability.
How a neural net is built
Early connectionism was an abstract model of biological neurons, devised before anyone knew how real neural networks were organised. Kurzweil gives a complete recipe.
- 1. Define the input. A series of numbers: a 2-D pixel array (vision); a 2-D array of sound parameters (such as frequency components) against time (speech); or any n-dimensional array (general pattern recognition).
- 2. Define the topology. Each neuron has many inputs, each wired to an input number or another neuron's output, and usually one output feeding a higher neuron or the final result. Create N_0 neurons in the first layer, wire them to points in the input, give each connection an initial synaptic strength; repeat for layers 1 to M, wiring layer i to layer i − 1.
- 3. Run recognition trials. Layer by layer, each neuron multiplies each input by its connection weight, sums the results, and outputs 1 if the sum exceeds its firing threshold, otherwise 0.
- 4. Train. Run repeated trials on sample problems, adjust all weights after each trial to improve that answer, and stop when accuracy hits an asymptote.
- 5. Run the trained net on new problems.
- 6. Take your neural-net company public (his joke).
The designer's decisions. Before training, someone must fix:
- Input meaning. What the input numbers represent.
- Shape. Number of layers, neurons per layer (which may differ by layer), and inputs per neuron (which may differ by neuron).
- Wiring. Which neurons feed which: set randomly, evolved by an evolutionary algorithm, or chosen by the designer. Kurzweil calls this a key design area.
- Initial weights. All equal, random, evolved, or designer-chosen.
- Thresholds. The firing threshold of each neuron.
- Output. Layer M's outputs; a single output neuron fed by layer M; a function such as their sum; or a function of several layers.
- Weight-update rule. The most researched choice. Options: nudge each weight up or down by a small fixed amount toward the right answer (trying both directions is slow, so local shortcuts exist), or use other statistical methods.
Variations on the basic schema:
- Topology set randomly or by evolution (mutation plus selection on network designs).
- Skip and feedback connections: inputs can come from any lower or higher layer, not only the one below.
- Smoother nonlinearities: instead of all-or-nothing 1/0 firing, a function rising quickly but gradually from 0 to 1, and outputs other than 0 and 1.
- Asynchronous nets: the schema above is synchronous (layer 0 to M in order); in a truly parallel system each neuron scans its inputs and fires independently whenever its sum crosses threshold.
Training, and learning from bad teachers
An untrained net with random wiring and weights gives random answers. Like a mammalian brain it must learn: it maximises a reward function while fed labelled examples (photos with and without corgis, labelled by humans). Correct outputs earn reward; connections consistent with the right answer are strengthened, the rest weakened, until the net answers without coaching. Data volume is the biggest single factor, as study time is for students.
The counter-intuitive result: nets learn from unreliable teachers. With labels right only 60% of the time, accuracy can exceed 90%, and under some conditions even fewer correct labels suffice. Kurzweil's explanation: random errors cancel. Train a net to spot handwritten 8s where a third of labels are wrong (8s marked as 4s, 5s marked as 8s, and so on); in a large dataset the mistakes push in no consistent direction, so the useful signal about what an 8 looks like survives.
Perceptrons and the long winter
- The Perceptron (1964). At Cornell Rosenblatt showed Kurzweil a one-layer net, the Perceptron, that read printed letters. Kurzweil's tinkering showed good auto-association (letters recognised with parts covered) but poor invariance (failure after changes of size or font).
- The attack (1969). Minsky, who had himself built neural nets in 1953, and Seymour Papert, the two cofounders of the MIT AI Laboratory, published Perceptrons. It proved formally that a feed-forward Perceptron (no loops, like Rosenblatt's Mark 1) cannot decide whether an image is connected.
- Why. Kurzweil's gloss: the task needs the XOR (exclusive or) function, true if one of two inputs fires but not both; a single layer can only apply a linear rule in one pass, and he says XOR needs a further step.
- The consequence. Funding for connectionism collapsed for decades.
Rosenblatt had already told Kurzweil the cure in 1964: feed a Perceptron's output into another layer like it, then another; each pass generalises more, and with enough layers and data the net could handle invariance and great complexity. He had not tried it, though it was high on his list; he died seven years later (1971). Multi-layer nets became common about a decade after that, but still needed more compute and data than was practical. Today's surge comes from using many layers more than half a century after Rosenblatt's idea.
- 2.8 billion-fold. Inflation-adjusted price-performance gain between Perceptrons (1969) and Minsky's death (2016).
- Minsky's regret. Near the end of his life he told Kurzweil he regretted the book's influence, as connectionism had by then triumphed.
- Da Vinci analogy. Connectionism resembled Leonardo da Vinci's flying machines: prescient, but waiting for lighter, stronger materials. Once hardware caught up, hundred-layer networks became feasible and now drive every spectacular advance.
Kurzweil says XOR needs a "feedback step". Strictly, a single linear layer fails because XOR is not linearly separable; a feed-forward net with one hidden layer solves it without any feedback. The story also skips the 1986 backpropagation work of Rumelhart, Hinton and Williams and the 1990s-2000s convolutional nets (LeCun), so the gap between Rosenblatt and the 2010s was less empty than implied.
The cerebellum: a modular structure
To place neural nets in the story of intelligence, Kurzweil detours to the start of the universe. Without brains to guide it, matter organised itself very slowly: each new level of detail took hundreds of millions to billions of years. Each later level arrived faster.
13.8 bn years ago Big bang.
4.5 bn Earth forms; about 10 bn years separate the first atoms from the first self-replicating molecules on earth, partly because stellar life cycles had to forge heavier elements first, partly chance in the "primordial soup".
3.5-4.0 bn First life (dates disputed).
~2.9 bn later First multicellular life.
+500 m Animals walk on land.
+200 m First mammals.
100 m+ From primitive nerve nets to the first centralised, three-part brain.
350-400 m later First basic neocortex; about 200 m more to the modern human brain.
Smarter animals tended to win competition for resources, so brains evolved over millions of years, a clear speed-up. The defining pre-mammalian innovation was the cerebellum, which in humans still holds more neurons than the neocortex.
What the cerebellum does
- Motor scripts. It stores and triggers scripts for tasks like signing your name. "Muscle memory" is really cerebellar: repetition wears the action into a groove, like wagon wheels rutting a trail.
- The fly-ball problem. You could catch a ball by solving the differential equations of its path and your body, but brains have no equation solver. The cerebellum solves a simpler problem: keep hand and ball in their familiar relative position, speeding the hand if the ball drops faster than expected. Mathematically this mapping of senses to muscles uses basis functions.
- Simulation. It can predict actions not taken: you could make the catch, but would crash into a teammate, so you hold back, all instinctively.
- Dancing. It steers movement without conscious awareness.
- Damage. People without a working cerebellum can still act voluntarily through the neocortex, but only with focused effort, and suffer coordination problems called ataxia.
Unconscious competence
Mastering a physical skill means repeating its components until they are automatic, a shift from motor-cortex control to cerebellar control. Whether throwing a football, solving a Rubik's Cube or playing piano, less conscious effort means better performance: faster, smoother, with attention freed for other things. A skilled musician hits a note as effortlessly as anyone sings the right note in "Happy Birthday", yet nobody can explain in words how they steer their vocal cords. Psychologists and coaches call this unconscious competence.
Simple modules, fixed behaviour
The cerebellum's power does not come from complex design. It holds most of the brain's neurons but takes little genomic information: it is thousands of small, simple processing modules in a feed-forward arrangement. Kurzweil expects further cerebellar discoveries to feed back into AI.
- Narrow modules. Those for piano fingering do nothing for walking. Humans lean on the cerebellum less and less as the flexible neocortex takes the lead.
- Fixed action patterns. In animals without a neocortex, the cerebellum records survival behaviours precisely and innately, not by imitation. Even some mammals keep complex innate acts: deer mice dig short burrows, beach mice long ones with an escape tunnel, and lab-raised mice of each species with no experience dig their own species' design when placed on sand.
- Change by selection only. A behaviour such as a frog's tongue strike persists until a variant population outcompetes it. Genetic behaviour adapts orders of magnitude more slowly than learned behaviour: over generations, not within a life.
- The computing parallel. Evolutionary algorithms copy this: generate programs with random traits, test them, recombine the best, add random mutations, repeat for many generations; the result can beat what a programmer would design.
In nature this takes millions of years, slow by today's standards but fast against the hundreds of millions prebiological chemistry needed. The cerebellum was itself an accelerator.
The neocortex: self-modifying, hierarchical, flexible
Faster progress needed a brain that could invent behaviour without waiting for genetic change. That was the neocortex ("new rind"), which appeared about 200 million years ago in a new class of animals, mammals.
- Modest start. In early rodent-like mammals it was the size and thinness of a postage stamp, wrapped round a walnut-sized brain.
- New organisation. Unlike the cerebellum's separate modules it worked as a coordinated whole, so it could invent new behaviours in days or even hours: it unlocked learning.
- Waiting for a crisis. While environments changed over thousands of years, slow cerebellar adaptation sufficed. The Cretaceous-Paleogene extinction, 65 million years ago (135 million years after the neocortex appeared), changed that: an asteroid, possibly with volcanism, abruptly altered the climate and killed about 75% of animal and plant species, including the dinosaurs (birds are arguably a surviving branch).
- Takeover. Quick-inventing mammals rose. Bodies grew; brains grew faster as a share of body mass; the neocortex grew fastest by folding to add surface. Unfolded, a human neocortex is the size and thickness of a large dinner napkin, yet it makes up about 80% of brain weight.
Minicolumns and hierarchy
Summarising How to Create a Mind (2012):
- Unit. A repeating module of about 100 neurons, the cortical minicolumn, which can learn, recognise and remember a pattern and organises itself into hierarchies.
- Count. 21-26 billion neurons in the cerebral cortex, ~90% (about 21 billion) in the neocortex, so on the order of 200 million minicolumns.
- Parallelism. Unlike largely sequential digital computers, the modules work in massive parallel, which makes the brain dynamic and hard to model.
- Levels of abstraction. As in artificial nets, layers separate raw sensory input from behavioural output. A bottom module recognises a curve; higher ones, a letter, a word, its rich meaning; the top, whether a statement is funny, ironic or sarcastic.
- Two-way traffic. The neocortex's six main layers signal in both directions, so abstraction does not live only in the top layer. Kurzweil prefers a species-level reading: more cortical layers mean more capacity for abstraction, and cloud-connected layers will mean more still.
In the late 1990s neurosurgeon Itzhak Fried operated on an awake 16-year-old epilepsy patient (the brain has no pain receptors). Each time he stimulated one particular neocortical spot she laughed, and not as a reflex: she genuinely found the moment funny and invented a reason, such as "you guys are just so funny, standing around". Point: humour has a physical address in the neocortex. Imaging adds that reading ironic sentences activates the theory-of-mind (ToM) network.
What abstraction bought us
Neocortical abstraction let humans invent language, music, humour, science, art and technology, which Kurzweil says no other species has achieved, clickbait notwithstanding: no animal keeps a beat in its head, tells a joke or reads a book.
- Tools. Chimpanzees make primitive tools, but never ones good enough to start rapid self-improvement.
- Communication. Other animals signal, but cannot express hierarchical ideas as human language does.
- From primate to philosopher. We were already capable primates; extra modules for understanding the world and existence made us philosophical animals.
- The thumb. Brains were not enough. Whales, dolphins and elephants have comparable or (in absolute terms) larger neocortices but nothing like an opposable thumb to shape materials into technology. We were lucky twice.
Hierarchy alone is not special (the cerebellum has hierarchies too). What sets the neocortex apart:
- 1. Broad propagation. The firing pattern for a concept spreads across the whole structure, not just where it began.
- 2. Shared patterns. One pattern links similar aspects of many concepts, and related concepts have related patterns.
- 3. Simultaneity. Millions of patterns fire at once and interact in complex ways.
Two consequences follow:
- Associative memory. A memory works like a Wikipedia page: linked from many places, revisable over time, and multimedia, so a smell, taste or sound can trigger it.
- Analogy. The pattern for lowering your hand is related to lowering your voice's pitch and to metaphorical lowerings such as a falling temperature or a declining empire. A pattern learned in one domain can be applied in another.
Darwin's analogy
Kurzweil's model case of analogy driving discovery:
- Before Darwin (1809-1882), most scientists held that God created each species separately. Earlier quasi-evolutionary theories, notably Jean-Baptiste Lamarck's (1744-1829) natural progression toward complexity plus inheritance of acquired traits, had mechanisms that were vague or false.
- The geology. The Scottish geologist Charles Lyell (1797-1875) argued against the view that canyons were created ready-made with rivers settling into them: the river came first and cut the canyon, grain by grain over millions of years. Resisted at first, the idea won once scientists saw that tiny impacts multiplied over aeons could carve a Grand Canyon. Lyell built on James Hutton (1726-1797) and his uniformitarianism: the earth was shaped by constant natural forces acting gradually, not by a catastrophic biblical flood.
- The transfer. Darwin applied Lyell's one-grain-at-a-time erosion to one small heritable change per generation, and explicitly argued in On the Origin of Species (1859) that natural selection would banish sudden creation as geology had banished the single great flood.
- The verdict. Kurzweil ranks this as arguably civilisation's most profound scientific revolution, and says rivals such as Newton on gravitation and Einstein on relativity rested on similar analogical leaps.
Deep learning re-creates the neocortex
Symbolic systems are too rigid for the fluidity of human thought; connectionism was too expensive to train. What changed was the price of computation.
Moore's law as one paradigm
Gordon Moore (1929-2023), Intel cofounder, stated Moore's law in 1965: thanks to miniaturisation, transistors per chip double about every two years. Skeptics say it must end at atomic scale. Kurzweil's reply:
- A special case. Moore's law is one instance of the law of accelerating returns, in which information technology creates feedback loops of innovation.
- Four earlier paradigms. By 1965 price-performance had already grown exponentially through electromechanical devices, relays, vacuum tubes and transistors.
- Paradigms to come. When integrated circuits top out, nanomaterials or three-dimensional computing take over.
- Long run. The trend has held since at least 1888, before Moore was born.
- The threshold. Around 2010 it made affordable the many-layered, neocortex-like connectionism called deep learning, source of the sudden breakthroughs since 2005.
Games: from Go to rules-free mastery
Go has vastly more possible moves than chess and positions are harder to evaluate, so the methods that beat chess grandmasters made almost no headway. Even optimists expected the 2020s; in 2012 Nick Bostrom guessed about 2022.
1997 Deep Blue (IBM) beats Garry Kasparov with hand-coded human chess lore; useful for nothing else.
2015-16 AlphaGo (DeepMind, Alphabet) uses deep reinforcement learning: trained first on recorded human moves, then on games against itself.
2016 AlphaGo beats Lee Sedol, then second in international titles, 4-1.
2017 AlphaGo Master beats 60 top professionals online and world champion Ke Jie 3-0.
2017 AlphaGo Zero, given only the rules, beats AlphaGo 100-0 after about 3 days of self-play, matches Master in 21 days, tops every version in 40 days, with no human data or intervention.
2017-18 AlphaZero beats all humans and all machines at chess after 4 hours of training, and masters shōgi, from the rules alone.
2019-20 MuZero repeats this without being told the rules: any board game without chance or hidden information, and deterministic video games such as Atari's Pong.
2022 Meta's CICERO beats many humans at Diplomacy.
- Transfer. Kurzweil stresses that AlphaZero and MuZero apply what they learn across games, which he calls a hallmark of human intelligence.
- Uncertain games. Deep reinforcement learning has also beaten all humans at StarCraft II and poker, games of hidden information that require modelling rivals.
- The holdout. Board games demanding high linguistic skill. Diplomacy cannot be won by luck or tactics alone: you must persuade others that your moves serve them. An AI that dominates it has mastered persuasion and deception. Such milestones, he says, now arrive almost weekly.
From games to the real world: simulators
The same methods work in messy reality if there is a simulator of the domain. Driving involves endless surprises (a car stopping suddenly, one coming the wrong way, a child chasing a ball). Waymo (Alphabet) ran its early robotaxis with a human monitor, recorded every ride, and built a detailed simulator from the record. With 20 million+ real miles logged, simulated cars train over billions of virtual miles, and Kurzweil expects self-driving cars to end up far better than human drivers. AI simulation is likewise tackling protein folding (chapter 6), opening paths to new medicines.
Waymo reported about 500,000 paid robotaxi rides a week across 10 US cities in March 2026, double the 250,000 of April 2025, and its co-CEO set a target of 1 million weekly trips during 2026. Verdict: on track; driverless service is real at city scale, though still limited to mapped metro areas.
Language as the key to generality
MuZero is still narrow: it cannot write a sonnet or comfort the sick. Matching the neocortex's generality requires language, which links distant domains and allows high-level symbolic transfer: one summary sentence can update what would otherwise need a million examples.
Meaning as position in a space. The fastest progress represents meaning in a very high-dimensional space, learned without hand-coded grammar:
- Train. A multilayer feed-forward net is fed billions or trillions of sentences from public web sources.
- Embed. Each sentence gets a point in, say, 500-dimensional space (500 numbers; the figure is arbitrary), starting random.
- Arrange. Training moves sentences with similar meanings close together and dissimilar ones apart; after billions of sentences, a sentence's position is its meaning.
- Context does the work. The model learns that "jam" has two senses because people eat jam and jam on guitars, but nobody eats guitars. Kurzweil notes humans learn almost all their words this way, not from dictionaries.
Applications he describes, many from his own team at Google:
- CLIP (OpenAI, 2021). Links images to captions; single nodes respond to a concept whether shown literally, symbolically or conceptually: a spider photo, a Spider-Man drawing and the word "spider", much as brains do.
- Translation. One 500-dimensional space holding sentences in every language: translate by finding the nearest sentence in the target language; nearby points give paraphrases.
- Question answering. Two paired spaces, built from billions of sentence-response pairs, map questions to answers.
- Universal Sentence Encoder (Google). Embeds each sentence with thousands of detected features such as "ironic", "humorous" or "positive", giving models semantic meta-knowledge beyond the literal words.
- Gmail Smart Reply. Offers three replies, using the whole thread, subject line and recipient cues; a feed-forward net plus a hierarchical model of the dialogue. Awkward to some at first, it now produces a significant minority of all Gmail traffic.
- Talk to Books (2018-2023). Searched every one of 500 million sentences in 100,000+ books in half a second, by meaning, unlike normal Google Search's keyword matching and click signals.
Transformers
Transformers, invented by Google researchers in 2017, power most recent progress.
- Attention. They focus computation on the most relevant parts of the input, as the neocortex directs attention.
- Tokens. Text is split into tokens (word pieces, words or word strings).
- Parameters. Billions to trillions of parameters, stored as weights between nodes, act as prediction factors for classifying tokens.
- Toy example. With one parameter for "is this an elephant?", choose "trunk"; but other animals have trunks. Add "hairy body": trunk plus hair suggests a woolly mammoth. More parameters, finer distinctions, better predictions.
- Opacity. Real parameters sometimes match human concepts but often encode abstract statistical relations no one can name.
- Output. The model predicts the likeliest next tokens after a prompt and turns them into text, images, audio or video.
Why scale matters. Narrow tasks (forecasting temperature from history) suit small models. Language does not: the ways to begin a sentence are effectively infinite, so even hundreds of billions of training tokens cannot be memorised into completions. Many billions of parameters let a model process prompts at the level of associative meaning and compose text never seen before, in the style the prompt calls for (Q&A, op-ed, stage dialogue). To the charge that this is a statistical trick, Kurzweil answers that the statistics distil the creative output of millions of people, so the model attains genuine creativity.
| Model | Year | Parameters | What it showed |
|---|---|---|---|
| GPT-2 (OpenAI) | 2019 | 1.5 bn | Flashes of promise, weak overall |
| GPT-3 (OpenAI) | 2020 | 175 bn | Fluent Q&A, style imitation |
| Gopher (DeepMind) | 2021 | 280 bn | Beat GPT-3 |
| Switch (Google) | 2021 | 1.6 tn | Open-sourced; mixture of experts |
| PaLM (Google) | 2022 | 540 bn | Explains jokes, chain of thought |
The breakthroughs came once models passed about 100 billion parameters. Switch's key idea, mixture of experts, activates only the parts of the model relevant to a task, keeping costs from spiralling as models grow.
GPT-3 and creative analogy
GPT-3 was the first such model sold commercially and the first whose creativity impressed users.
Amanda Askell prompted GPT-3 with John Searle's Chinese room: a person who knows no Chinese, executing a translation program by hand, would not understand the stories, so why credit a computer running the same program? GPT-3 agreed it understood nothing and called the program a formal system that "does not explain understanding any more than a cookbook explains a meal". The line appears nowhere earlier; it looks like a fresh twist on David Chalmers's point that a recipe does not explain a cake. Kurzweil sees exactly the Darwin-style analogy of the neocortex.
- Style on demand. Having digested almost every genre, GPT-3 answered in the voice of science papers, children's books, poems or sitcom scripts, and could imitate named writers.
- The Kaufman test. Programmer Mckay Wrigley asked it "How do we become more creative?" in the style of pop psychologist Scott Barry Kaufman; Kaufman said the answer "definitely sounds like something I would say".
Conversation, images and zero-shot learning
- LaMDA (Google, 2021). Built for open-ended, lifelike dialogue. Asked to play a Weddell seal, it answered in character, teasing a would-be hunter that he would freeze before getting a shot: contextual knowledge that had long eluded AI.
- Multimodality (2021). Earlier systems handled one data type. OpenAI's DALL-E (a pun on Salvador Dalí and Pixar's WALL-E) learned word-image relations and drew novel concepts such as an avocado-shaped armchair.
- Photorealism (2022). DALL-E 2, Google's Imagen, Midjourney and Stable Diffusion produced near-photographic scenes from prompts (a fuzzy panda in a cowboy hat and leather jacket cycling on a mountaintop). Kurzweil expects this to transform creative professions.
Few-shot learning. After training, a model completes a task from a small sample: shown five unicorn pictures instead of thousands or millions, it recognises or draws new unicorns. GPT-3 works this way.
Zero-shot learning. The model combines concepts never seen together: a baby daikon radish in a tutu walking a dog, a snail with the texture of a harp, a lovestruck cup of boba emoji. Kurzweil calls this the essence of analogical thinking and proof the model is not parroting; extending it across domains is a defining AI challenge of the 2020s.
One network, many skills
- Gato (DeepMind, 2022). 17 months after MuZero, one network plays video games, chats, captions images and controls a robot arm. The skills were not new; uniting them in one brainlike system was. In The Singularity Is Near Kurzweil predicted that thousands of skills would be combined in one AI before a Turing test was passed.
- Codex (OpenAI, 2021). Turns plain-English requests into working Python, JavaScript or Ruby; a non-programmer can build a simple game in minutes.
- AlphaCode (DeepMind, 2022). Stronger still. Kurzweil expects coding skill to stop being a barrier to turning ideas into software.
PaLM: humour and chain of thought
Earlier models struggled to combine subtasks without human guidance. Google's PaLM (April 2022, 540 bn parameters) advanced on two quintessentially human abilities:
- Humour. Jokes mix wordplay, irony and shared experience, which is why they translate badly. Given a joke about meaning to start a paper at 5:00 PM, playing with a language model for "10 minutes" and finding it is 9:30 PM, PaLM explained that time flies when you are having fun and that the ten minutes were really 4.5 hours.
- Chain-of-thought reasoning. PaLM shows its inferences step by step (not yet as deeply as humans, as of 2023). That is what teachers want from essays; it makes outputs more trustworthy and errors easier to diagnose, and splitting a question into chunks helps it solve problems hard even for bright humans.
The Louvre riddle. Prompt: Michael is at a famous French museum looking at its most famous painting; the painter reminds him of a childhood cartoon character; where does the object that character usually holds come from? PaLM's chain: Louvre → Mona Lisa → Leonardo da Vinci → Leonardo of the Teenage Mutant Ninja Turtles → he carries a katana → katanas come from Japan. Six linked inferences, each simple, together beyond earlier models.
ChatGPT, GPT-4 and embodiment
- ChatGPT (November 2022). OpenAI's interface to GPT-3.5 gave the public easy access to an LLM; 100 million people tried it within two months. It disrupted education: students used it for essays and teachers lacked reliable detectors.
- GPT-4 (March 2023). Excelled on the SAT, LSAT, AP tests and the bar exam. Its bigger advance was world modelling: reasoning about objects and actions in hypothetical situations, including physics with altered laws; seeing that two people viewing numbers from opposite sides read them reversed; predicting that balloons held by a weight fly off if the strings are cut.
- The diamond test. Security engineer Daniel Feldman described a diamond in a loose thimble inside a cup on a chair; the chair moves to the bedroom, the cup goes on the bed, is turned upside down and back, then goes to the kitchen counter. GPT-4 answered that the diamond fell out when the cup was inverted and is on the bed: spatial tracking over time.
- PaLM-E (2023). PaLM's reasoning in a robot body. Told to bring the rice chips from the drawer, it moved through a kitchen, found, picked up and delivered them.
- Assistants. As the book entered layout, the twenty-five-year links-page model of search was being augmented by Bard (powered by Gemini, which Kurzweil says surpasses GPT-4) and Bing (a GPT-4 variant); Google Workspace and Microsoft Office were adding AI. Printing takes nearly a year, so he warns readers that much will have moved on.
In 1993 Kurzweil debated his mentor Minsky before several hundred students in MIT's Room 10-250. Kurzweil said human-level AI needs about 10¹⁴ calculations per second; Minsky said compute did not matter and a Pentium could be as smart as a human with the right program. Neither won: Kurzweil lacked the compute to demonstrate, Minsky the algorithms. The 2020-2023 breakthroughs, Kurzweil says, settled it in his favour, sixty years after he entered AI (around 1963); frontier training compute now grows about 4× a year.
Kurzweil says AlphaZero transfers abilities learned from Go to chess. In DeepMind's papers the same algorithm was trained from scratch, separately, for each game; no Go knowledge carried over. MuZero likewise learned each game in its own run. Gato combined skills in one network but did not show strong transfer between them. The general point (one method, many domains) holds; the claim of skill transfer is overstated.
What AI still needs to achieve
AI is well on the way to re-creating the neocortex, but Kurzweil names three deficits.
1. Contextual memory
Keeping track of how all the ideas in a conversation or book fit together. The complexity-ceiling arithmetic returns: relationships among subsets of n ideas grow as 2ⁿ − 1.
- 10 tokens in a sentence: 1,023 possible relationships.
- 50 tokens in a paragraph: 1.12 quadrillion.
- Whole chapters or books by brute force: out of control. Hence GPT-4 forgetting earlier turns and failing to sustain a consistent novel plot.
Two pieces of good news. Researchers are building models that attend to relevant context more efficiently, and computing cost should fall more than 99% within a decade, faster still for LLMs thanks to algorithms and AI-specific hardware. From August 2022 to March 2023 GPT-3.5 API token prices fell 96.7%, and AI is already helping design chips.
2. Common sense
Imagining situations and anticipating consequences: you have never studied gravity switching off in your bedroom, yet can picture the result. It also drives causal inference: a dog at home plus a broken vase tells a story. AI lacks a robust world model, and training text seldom states such implicit knowledge.
3. Social nuance and theory of mind
Ironic tone is poorly captured in text, which makes theory of mind (recognising that others hold different beliefs and inferring their motives) hard to learn. Progress is rapid:
- Sally-Anne style test (2021). Google Fellow Blaise Agüera y Arcas gave LaMDA a classic child-psychology scenario: Alice leaves her glasses in a drawer; Bob moves them under a cushion while she is out. LaMDA correctly said she will look in the drawer.
- Two years later. PaLM and GPT-4 answered many such questions correctly.
- Why it matters. A human Go champion can also watch the room, joke at the right moment, or stop the game for a medical emergency; theory of mind gives AI that flexibility.
Three converging exponentials
Kurzweil's optimism rests on three trends moving together:
- Compute price-performance, making large nets cheaper to train.
- Training data, richer and broader, so compute is used better.
- Algorithms, letting AI learn and reason more efficiently.
- 1.4 years: doubling time of compute per dollar since 2000.
- 5.7 months: doubling time of compute used to train frontier models since 2010, which Kurzweil calls about a ten-billion-fold rise.
- ~2 years: doubling in training compute from 1952 (one of the first machine-learning demonstrations, six years before the Perceptron) to 2010, tracking Moore's law.
- Less than 75×: the gain to 2021 had the pre-2010 pace continued.
- Drivers: new parallel-computing methods that let many chips share one training job, and investor money pouring in once big data made deep learning pay; not a hardware revolution.
The feedback-data rule
The spending growth reflects how much useful data now exists. Kurzweil's rule, which he says could only recently be stated with confidence: any skill that produces clear enough performance-feedback data can be turned into a deep-learning model that surpasses all humans.
| Skill | Measurable outcome? | Data easy to get? | Route to training |
|---|---|---|---|
| Chess | Yes: win, loss, draw, Elo | Yes: games are exact move sequences | Direct |
| Courtroom advocacy | Yes: verdicts | Hard: lawyer skill tangled with case strength, juror bias | Careful analysis |
| Poetry | Unclear | Hard | Proxies: reader scores 0-100, fMRI activation |
| Suspense in fiction | Unclear | Hard | Proxies: heart rate, cortisol |
With enough data, imperfect, indirect metrics still steer improvement; finding them takes creativity. A well-designed net can also extract insights beyond human perception, in games, driving, medical images or protein folding, which creates a strong economic incentive to gather data once thought too hard to collect.
Data behaves like oil. Oil deposits range from gushers to shale needing fracking or heating; low prices make only easy oil worth extracting, high prices unlock the rest. Data is the same: as machine learning gets better and compute cheaper, harder-to-collect data becomes worth gathering. Kurzweil expects this to happen in the 2020s for almost every human skill.
Intelligence as a bundle
Human-level intelligence is not one thing AI has or lacks; it is a thick bundle of abilities.
- Shared and unique. Mirror self-recognition we share with elephants and chimpanzees; music composition is human but varies hugely between people.
- Uneven within individuals. A maths genius may play poor chess; a photographic memory may come with weak social skills (Dustin Hoffman in Rain Man).
- "Human-level" means the best humans. The gap between average and best is small for recognising letters of your own alphabet, huge for theoretical physics; in wide-gap fields there may be a long lag between average and superhuman AI.
- Uneven arrival. In 2034 AI might write Grammy-winning songs but not Oscar-winning screenplays, or solve Millennium Prize Problems but produce no deep new philosophy: a transitional period in which AI has passed the Turing test and is superhuman in most ways but trails top humans in a few.
Programming and the intelligence explosion
The fibre that matters most for the Singularity is computer programming (and related abilities such as theoretical computer science), the main bottleneck to superintelligence. Once AI can raise its own programming skill, alone or with human help, a positive feedback loop begins. I. J. Good, Turing's colleague, foresaw this intelligence explosion in 1965; with humans removed from the loop, computer speeds give startling progress, nicknamed "FOOM" after a comic-book sound effect.
The speed of takeoff is disputed:
- Hard takeoff (Eliezer Yudkowsky). Self-improvement runs away in minutes to months.
- Soft takeoff (Robin Hanson). It unfolds over years or longer.
- Kurzweil in between. Limits on hardware, resources and real-world data cap FOOM's speed, but a hard takeoff going wrong deserves precautions. Once the explosion starts, every skill harder than self-programming follows quickly.
How much computation is a brain?
Raw power will not be the bottleneck. Oak Ridge National Laboratory's Frontier, the top supercomputer in 2023, performs about 10¹⁸ operations per second, around 10,000× the brain's likely maximum of 10¹⁴.
- 2005 upper bound: 10¹⁶ ops/s = 10¹¹ neurons × 10³ synapses × 10² firings/s, deliberately high to be conservative.
- Revised firing rate. Research since shows neurons fire nearer once a second, not the ~200/s theoretical maximum. AI Impacts, working from the brain's energy use, estimates an average of 0.29/s, implying ~10¹³. That matches Hans Moravec's estimate in Mind Children (1988), reached by a completely different method.
- Redundancy. Not every neuron is needed: full recovery after stroke or brain injury shows much work is duplicated, so 10¹⁴ is a conservative most-likely figure.
- Cost. About $1,000 of hardware reached 10¹⁴ in 2023; even 10¹⁶ should cost $1,000 by about 2032.
- If consciousness needs more detail. Kurzweil holds that a model of neuron firing suffices, but concedes (as an untestable philosophical question) that subjective experience might require simulating ion channels or the thousands of molecule types in a cell. Anders Sandberg and Nick Bostrom (Future of Humanity Institute, Oxford) put these at 10²² and 10²⁵ ops/s; a $1 billion (2008 dollars) supercomputer could reach the latter by 2030 and simulate every protein in every neuron by 2034.
His conclusion: even under wildly different assumptions, computers will simulate human brains in every way we care about within about two decades, not a century away. Because life extension starts accelerating in the 2020s, a healthy person under 80 should see it. Children born now will see the Turing test passed in elementary school and richer brain emulation by college. And 2023 is probably closer to feasible full brain emulation than to 1999, when he made many of these forecasts.
A 5.7-month doubling from 2010 to 2022 is about 25 doublings, a rise in the tens of millions, not "ten-billion-fold"; the ten-billion figure matches the chart's extreme points, not its trend line. Brain-compute estimates are also contested: Joseph Carlsmith's 2020 review for Open Philanthropy spans 10¹³ to 10¹⁷ FLOP/s, centred near 10¹⁵, above Kurzweil's 10¹⁴.
Passing the Turing test
The Kapor bet and the AI effect
On April 9, 2002 Kurzweil and personal-computing pioneer Mitch Kapor made the first "Long Now" bet: will a Turing test be passed by 2029? Its rules settle what Turing left open: interview length, judges' skills, and how much cognitive enhancement a judge or human foil may have and still count as human.
A strict empirical test is needed because of the AI effect: whatever AI achieves is redefined as easy. Over seven decades computers beat humans in many narrow areas while lacking human breadth and flexibility.
- Chess (1997). After Deep Blue beat Garry Kasparov, critics said chess is mathematically tidy (perfect information, few moves per turn), so victory was a trick of maths; machines would never handle ambiguous language such as crosswords or Jeopardy!.
- Crosswords fell within two years.
- Jeopardy! (2011). Less than twelve years later IBM's Watson beat the two greatest players, Ken Jennings and Brad Rutter.
Watson buzzed in and spoke answers like a thinking contestant, but the screen also showed its top three guesses. In the category "EU, the European Union", the clue described a body elected every 5 years with 736 members from 7 parties. Watson answered European Parliament (66% confidence); its runners-up were "MEPs" (14%) and "universal suffrage" (10%), errors no human would make from the clue's grammar alone, even one who had never heard of the EU. The performance looked human; the cognition beneath was alien.
Fluency, hallucination and "superstition"
- Google Duplex (2018). Phoned businesses so naturally that people assumed a human caller.
- IBM Project Debater (2018). Held its own in competitive debate.
- LLMs (2023). Write whole essays to human standard.
- Hallucination. Even GPT-4 confidently invents: summarising a nonexistent news article, or citing papers that do not exist for true facts. Despite heavy engineering effort, how hard this is to fix remains open; like Watson, LLMs reach answers by statistical processes unlike our thinking.
Kurzweil calls the intuition that AI "ought" to reason as we do superstition. What matters is how an intelligence acts: if alien processes yield great science or moving novels, their origin is irrelevant, and an AI that eloquently claims consciousness gives us no ethical grounds to insist only biology yields sentience worth respecting. The Turing test's empiricism keeps attention on behaviour. Turing himself (1950) anticipated the objection that machines might think very differently from humans and set it aside if they play the game well.
Not the only benchmark
- Superhuman without passing. PaLM 2 and GPT-4 beat humans at demanding tasks yet cannot pass as human.
- 2023-2029. Computers become clearly superhuman in a widening range of areas before the first robust test is passed in 2029; AI might even become superhuman at programming itself before it masters the social subtleties the test probes.
- Richer yardsticks. Human-level intelligence needs a nuanced definition and new ways to compare how human and machine minds differ.
- Persuasion value. Despite objections, Kurzweil believes a true pass would be compelling to observers and convince the public of genuine intelligence, not mimicry.
What a strong test demands
- Every verbal skill. Passing means beating humans on every cognitive test expressible in language, since judges can probe any weakness: explain a complex social situation, draw inferences from scientific data, write a funny sitcom scene. Turing said the test could cover almost any field of human endeavour.
- Sharp judges and foils. Imitating someone drunk, sleepy or not fluent does not count, nor does fooling judges who do not know how to probe.
- Playing dumb. The AI must hide instant trivia recall, fast checking of whether huge numbers are prime, and fluency in a hundred languages.
- Already superhuman. Such an AI pairs human reading comprehension with perfect recall of every Wikipedia article and all published research. Humans are limited by slow reading, small memory and short lives; if a person needs six hours per book, Talk to Books was about five billion times faster.
When AI language understanding reaches human level, the result is an explosion of knowledge, not an increment. A Turing-level AI must dumb itself down for the test, while in medicine, chemistry and engineering it would already deliver profoundly superhuman results.
| Epoch | Medium | Time to add new capability |
|---|---|---|
| First | Nonliving matter | Billions of years (atomic and chemical synthesis) |
| Second | RNA and DNA | Millions of years (until natural selection introduces a new behaviour) |
| Third | Cerebellum | Thousands to millions of years (complex skills via evolution); hours to years (basic learning) |
| Fourth | Neocortex | Hours to weeks (complex new skills) |
| Fourth | Digital neural nets | Hours to days (complex new skills at superhuman level) |
| Fifth | Brain-computer interfaces | Seconds to minutes (ideas unimaginable today) |
| Sixth | Computronium | Under seconds (cognition reconfigured to the limits of physics) |
Extending the neocortex into the cloud
Noninvasive interfaces hit a physics wall
Outside-the-skull methods face a space-time trade-off:
| Method | What it measures | Spatial resolution | Temporal resolution |
|---|---|---|---|
| fMRI | Blood flow (active areas draw oxygenated blood) | voxels of 0.7-0.8 mm | a couple of seconds, rarely better than 400-800 ms |
| EEG | Electrical activity from outside the skull | 6-8 cm³, sometimes 1-3 cm³ | about 1 ms |
Both limits come from the physics of blood flow and electricity, so AI and better sensors will bring only marginal gains, not a sophisticated interface.
Implants: two-way, but invasive
Electrodes inside the brain avoid the trade-off and allow two-way communication (recording and stimulating neurons), but require holes in the skull and risk damage, so they serve people for whom benefits outweigh risks (hearing loss, paralysis).
- BrainGate. Lets people with ALS or spinal-cord injury move a cursor or robot arm by thought, but connects to too few neurons for complex signals like language.
- Facebook-sponsored decoding (2020). 250 electrodes plus AI matched cortical activity to spoken sample sentences and, with a 250-word vocabulary, predicted words with error rates as low as 3%. Facebook halted the project in 2021; scaling to larger vocabularies runs into the same resolution limits.
- Neuralink (Elon Musk). Implants many threadlike electrodes at once: 1,500 read out in rats against the hundreds used elsewhere; a monkey played Pong; after regulatory delays, FDA approval for human trials in 2023; the first 1,024-electrode human implant came as the book went to press.
- DARPA Neural Engineering System Design. Targets recording 1 million neurons and stimulating 100,000; one funded team at Brown University is building sand-grain "neurograins" that talk to neurons and each other as a "cortical intranet".
Expanding the neocortex itself needs two-way contact with huge numbers of neurons, and ultimately interfaces will be essentially noninvasive: harmless nanoscale electrodes delivered through the bloodstream.
How many connections?
- Whole-brain simulation. About 10¹⁴ ops/s or less for a model faithful to brain architecture that could pass the Turing test and look like a brain to outside observers. It would skip activity irrelevant to behaviour, such as DNA repair inside neurons.
- An interface needs far less. Most of those operations are preliminary processing below the top layer of the neocortex, and non-cognitive work such as regulating digestion can be ignored. Kurzweil estimates millions to tens of millions of simultaneous connections, to the upper neocortex only.
- The 2030s. Shrinking interfaces, with advanced AI solving the engineering and neuroscience, will reach this via nanobots entering through the capillaries. They will link the top neocortical layers to simulated neurons in the cloud, with no brain surgery.
- No skull limit. Brain size has been capped by the birth canal; cloud neocortex can be stacked layer on layer, and its power grows with computing price-performance through the century.
The second neocortical revolution
The last time we gained more neocortex, about two million years ago, we became human. Cloud neocortex should bring a similar leap in abstraction and forms of expression far richer than today's art and technology.
- The monkey at the movies. A monkey watching a film sees people talking but cannot grasp the dialogue, or that metal suits signal a story set in the Middle Ages. That is the gap the human prefrontal cortex opened.
- Art beyond words. Future art will not just be better effects or added taste and smell. Actors now convey a character's mind only through words and expression; cloud-connected audiences might receive a character's raw, nonverbal thoughts directly.
- Co-creation. Expanded minds produce new ideas for still larger minds to explore. With AI able to redesign itself, we gain access to our own source code.
Merging with the superintelligence we build means remaking ourselves. Running on a substrate millions of times faster than tissue and no longer confined by the skull, our minds grow exponentially until intelligence expands millions-fold: the core of Kurzweil's definition of the Singularity.
The 2020 decoding study (Makin, Moses and Edward Chang, UCSF) used electrodes placed on the brain's surface (ECoG) in epilepsy patients, not external ones, and the 3% error rate applied to a narrow set of practised sentences. Nanobots that cross the blood-brain barrier and interface with neurons do not yet exist even in prototype, so the 2030s date is the chapter's least supported forecast.
2026 scorecard
| Prediction (book) | Date | Status, Oct 2026 | Evidence |
|---|---|---|---|
| Valid Turing test passed | 2029 | On track / ahead | GPT-4.5 judged human 73% in 5-min tests (Mar 2025) |
| Metaculus agrees on AGI | 2029 | On track | Weak AGI median Sep 2027; strong AGI Feb 2031 |
| Training compute ×4 per year | ongoing | Ahead | Epoch: ~5× a year since 2020 |
| Compute 99% cheaper in a decade | by ~2033 | Ahead for LLMs | Epoch: fixed-capability prices fall 9-900× a year |
| Context memory gap closes | 2020s | Ahead | 1-10 million-token context windows |
| Self-driving beats humans at scale | 2020s | On track | Waymo ~500,000 paid rides/week, 10 US cities (Mar 2026) |
| 10¹⁶ ops for $1,000 | ~2032 | On track | Chip price-performance +49% a year |
| Cloud-linked neocortex via nanobots | 2030s | Behind | Best implants ~1,000 channels; ~2 dozen users |
OpenAI's o1 (Sept 2024) began the reasoning-model era: models trained to think at length before answering. In July 2025 an OpenAI experimental model and Google DeepMind's Gemini Deep Think reached gold-medal level at the International Mathematical Olympiad (35/42). METR measures the length of software tasks AI completes half the time: GPT-4 (2023) about 4-5 minutes, GPT-5 (Aug 2025) 2-3.5 hours, Claude Opus 4.6 (Feb 2026) about 12 hours, with doubling every 4-7 months. On 7 April 2026 Anthropic withheld Claude Mythos Preview from general release after it autonomously found thousands of serious software vulnerabilities; it scored 93.9% on SWE-bench Verified. Context windows grew from GPT-4's thousands of tokens to 1 million (Gemini 1.5, 2024) and 10 million (Llama 4 Scout, 2025). Hallucination is reduced but not solved, and continual learning remains weak. Verdict: AI is self-programming faster than it is mastering common sense, as Kurzweil allowed.
Per Epoch AI (Feb 2026): frontier LLM training compute rises about 5× a year (doubling every 5.2 months), training costs 3.5× a year, AI-chip performance per dollar 49% a year; the first run above 10²⁶ FLOP was xAI's Grok 3 (Feb 2025). The June 2026 TOP500 lists five exascale machines; China's LineShine leads at 2.2 exaflops and Frontier is third (1.35). Neuralink implanted Noland Arbaugh on 28 Jan 2024 (about 85% of threads later retracted, fixed by software), reached 12 participants by Sept 2025 and about 21 by Jan 2026, with trials in Canada, the UK and the UAE, and FDA breakthrough designations for Blindsight vision and speech; it raised $650 million at about $9 billion (June 2025). Verdict: compute ahead; BCIs far behind the millions of connections Kurzweil needs.
Who Am I?
Kurzweil turns from intelligence in general to the particular self: what makes Ray Kurzweil (or you) the centre of an experience, and what survives when a brain is copied or rebuilt. He answers with panprotopsychism: consciousness is a latent property of the universe that complex enough information processing "awakens", whatever the substrate. Identity, he concludes, is a pattern of information kept continuous over time, so gradually replacing your brain with digital parts keeps you you, while a separate copy is a new person.
What is consciousness?
The Turing test says what it means to be human in general; Singularity technologies force the question of what it means to be a particular human. Kurzweil asks why he is the centre of his own experience: why this person, why not born in 1903 or 2003, why male, why human at all. Science gives no reason. "Who am I?" is a philosophical question, and at bottom a question about consciousness.
Butler's plant
In How to Create a Mind Kurzweil quoted Samuel Butler (writing in 1871) on a carnivorous flower that closes on edible flies but ignores raindrops and sticks: so "unconscious" a thing guards its interests remarkably well, which makes Butler wonder what consciousness adds. Kurzweil's questions follow:
- Is the plant conscious? Or this species alone? There is no way to tell.
- Other people. We confidently call them conscious because they communicate and decide as we do, but even that is technically an assumption: consciousness cannot be detected directly.
Two meanings of the word
Functional consciousness. Awareness of surroundings, and acting as though aware of both one's own thoughts and a separate external world. It comes in degrees (deep sleep: none; drunk: partial; sober: full) and can usually be judged from outside, except in rare cases like locked-in syndrome. In animals, behaviours such as mirror self-recognition reveal it.
Subjective consciousness. Having inner experiences, qualia, not merely appearing to. A person's qualia cannot be detected from outside, yet this is the sense that matters for personal identity and for ethics.
Why it matters morally
- Objects matter only through minds. However valuable or elaborate, material things count only as far as they affect conscious beings' experience.
- Animal rights turn on how conscious animals are and what their experience is like.
- Proxy reasoning. Because qualia are invisible, we use functional consciousness as a proxy, by analogy with ourselves: I have inner experience and outwardly visible awareness, so others who show the second probably have the first. Even scientists who call subjective consciousness irrelevant behave as if the people around them have experiences.
A sliding scale across species
Confidence in other minds fades as behaviour grows less like ours:
| Being | Behaviour | Typical intuition |
|---|---|---|
| Dogs, chimpanzees | Complex, emotional | Most assume rich inner experience |
| Rodents | Social play, fear of danger | Fewer think them conscious; experience seen as shallower |
| Fruit flies | Respond to environment; ~250,000 neurons | Doubtful |
| Cockroaches | ~1,000,000 neurons, about 0.001% of a human's | Little room for deep hierarchies |
| Amoebae | No functional consciousness, yet rudimentary memory found this century | Generally no |
Consciousness is binary in one sense (any qualia at all?) but Kurzweil's interest is degree: compare a vague dream, drunken or sleepy waking, and full alertness. Expert opinion is moving toward more animals having more consciousness. In 2012 a multidisciplinary meeting at the University of Cambridge produced the Cambridge Declaration on Consciousness: lacking a neocortex does not seem to rule out affective states, and the neural substrates of consciousness exist in all mammals and birds and many others, including octopuses.
Sacred either way
What causes subjective consciousness? Some say God, others purely physical processes. Kurzweil notes that both sides treat consciousness as somehow sacred: the cause is debated, the result is not, and anyone who denied a child's capacity for suffering would be judged gravely immoral.
The working rule, and why it is urgent
Soon the cause of consciousness becomes a practical question: as we extend minds beyond biological brains, we must decide what generates the qualia at the core of identity, and protect it. Since behaviour is the only available proxy, Kurzweil adopts what he calls both our natural intuition and the most plausible scientific account:
- Rule. Brains that support more sophisticated behaviour give rise to more sophisticated subjective consciousness.
- Source of sophistication. Complexity of information processing (chapter 2), set mainly by how flexibly a brain represents information and how many hierarchical layers it has.
- Implication. Every intellectual leap in recorded history happened in brains structurally unchanged since the Stone Age. Technology lets us access nearly all human discoveries, but we experience them at roughly a Neolithic level of consciousness. Augmenting the neocortex in the 2030s and 2040s will add more than problem-solving power: it will deepen subjective consciousness itself.
Zombies, qualia and the hard problem
Inverted qualia
Something in consciousness cannot be shared. We agree to call certain light frequencies "red" and "green", but my experience of green could be your experience of red, and no language or other communication could reveal it. Even when brains can be wired together directly, nobody could prove the same signals produce the same qualia, so a swap would stay forever undetected.
Zombies
In How to Create a Mind Kurzweil took this further with David Chalmers (born 1966) and his zombies: beings with every detectable neurological and behavioural correlate of consciousness but no subjective experience at all. Science could never distinguish them from normal people.
Picture a dog and an artificial human known to be a zombie. The zombie can out-think the dog, yet most people would say hurting the dog is worse: the dog presumably feels pain, while the zombie only yelps. The point is that our moral concern tracks subjective, not functional, consciousness. The catch: in real life, even in principle, no test can tell who has subjective experience.
The hard problem
The zombie question creates a dilemma:
- If zombies are possible, there is no necessary causal link between qualia and the physical systems (brains or computers) whose processing makes something look conscious. That resembles the religious view of the soul as a supernatural entity separate from the body, beyond science.
- If zombies are impossible, because the physical systems of cognition necessarily generate consciousness, science still has no coherent way to show it. Subjective experience is qualitatively unlike observable physical law, and nothing follows from patterns of processing to experience at all.
Chalmers calls this the hard problem of consciousness. His "easy problems", such as what the mind does when not awake, are among science's hardest, but at least they can be studied scientifically.
Panprotopsychism
Chalmers's response is panprotopsychism: consciousness behaves like a fundamental force of the universe, not reducible to other physical forces; imagine a universal field holding the potential for consciousness. In Kurzweil's interpretation, the information-processing complexity found in brains "awakens" that potential into recognisable subjective experience. So a brain of carbon or silicon complex enough to show the outward signs of consciousness also has an inner life.
- Dualism. Consciousness comes from a stuff wholly separate from "dead" matter, often called a soul. Weakness: even granting a supernatural soul, there is no promising theory of how it would move matter such as neurons.
- Physicalism (materialism). Consciousness arises wholly from arrangements of ordinary matter. Weakness: it can explain function, as computer science explains AI, but not the subjective dimension, which science cannot reach.
- Panprotopsychism (Chalmers). A middle way: a fundamental potential for consciousness, awakened by complexity. Kurzweil judges it the most likely causal account as well as the ethically safest.
The ethical wager
Panprotopsychism can never be proved. Kurzweil's main argument is therefore moral:
- Precaution. If an entity you might mistreat could plausibly be conscious, the safest choice is to assume it is rather than risk tormenting a sentient being: act as though zombies are impossible.
- Turing test as evidence. On this view a passed Turing test shows more than functional ability: it is strong evidence of subjective consciousness, and so of moral rights.
- Law will lag. Kurzweil doubts politics will write AI rights into law before the first Turing-level AIs exist, so at first the developers must devise ethical frameworks to restrain abuse.
Anthropic launched a model-welfare programme in April 2025 and in August 2025 let some Claude models end persistently abusive conversations; days later Mustafa Suleyman (Microsoft AI) called such research "premature, and frankly dangerous". Kurzweil's precautionary rule and his developers-first prediction now have corporate backers and critics, but no test for consciousness has emerged. Verdict: on track for developer-led ethics; no closer on the hard problem.
Determinism, emergence and free will
What free will would have to be
- Everyday sense. Most people mean the ability to control their own actions; political and judicial systems assume everyone has it.
- The philosophers' worry. Many hold that free will needs an undetermined future: if the future is fixed, how is the will free?
- Randomness is no rescue. If our choices reduce to quantum chance, that is not freedom either; the English philosopher Simon Blackburn put it as "chance is as relentless as necessity".
- The target. A meaningful free will must combine deterministic and indeterministic ideas: neither rigidly predictable nor random.
Wolfram's cellular automata
Kurzweil finds the path through the work of physicist and computer scientist Stephen Wolfram (born 1959), long an influence on him, and his book A New Kind of Science (2002).
- Cellular automata. Simple models of cells that switch between states (black/white, dead/alive) according to a rule based on neighbouring cells, step by discrete step, often yielding complex behaviour.
- Conway's Game of Life. The famous 2-D example. Enthusiasts have found shapes with predictable evolutions, and Life can emulate a working computer, even one running and displaying another copy of Life.
- Wolfram's 1-D version. A single line of cells; each new line below is computed from the line above. Each cell's next state depends on three cells (itself and its two neighbours), so there are 8 possible input combinations; a rule lists the output for each, readable as an 8-bit binary number (black = 1, white = 0) that names the rule.
Wolfram's central finding. For some rules, no shortcut exists: however far ahead you want to look, you must compute every intermediate step.
- Class 1 (example: rule 222). Fully predictable. From one black cell it grows a solid triangle; ask for the millionth cell, or the million-to-the-millionth, and the answer is simply black. This is how science is "supposed to" work: deterministic rules, predictable outcomes.
- Classes 2 and 3. Progressively more complex arrangements of black and white cells.
- Class 4 (example: rule 110). A mix of regular and irregular structure. The thousandth row, or the million-to-the-millionth, can be found only by computing every row before it. Wolfram argues most of the natural world, and the universe itself, has class 4 properties.
Emergence
Class 4 systems have an irreducible complexity that defeats the old reductive determinism: the behaviour comes from deterministic programming, yet the programme does not fully explain its richness.
- Looks random, is not. Sampled cell by cell, states look random; yet each follows strictly from the step before, and the overall image mixes order and irregularity.
- Definition. Emergence is very simple things, collectively, producing much more complex things.
- In nature. Fractal structures show class 4-type coding: the gnarled growth of tree limbs, the stripes of zebras and tigers, the patterns of mollusc shells, and much else in biology. We live in a world shaped by simple algorithms that straddle the line between order and chaos.
- In us. This complexity may give rise to consciousness and free will. Whether you credit the underlying programme to God, to panprotopsychism or to something else, you are more than the programme.
Physics as computation
Kurzweil argues it is no coincidence that such rules produce both consciousness and so many natural phenomena. Wolfram makes a strong case that physical laws themselves arise from computational rules related to cellular automata, and in 2020 launched the Wolfram Physics Project to model all of physics with a generalised automaton.
- A compromise. Some macro-scale things allow shortcuts (where a satellite will be a million orbits from now). At the most fundamental scale they would not. If reality runs on class 4-like rules, apparent quantum randomness could be deterministic underneath, yet no algorithm could look ahead to the universe's future state.
- Status. Kurzweil labels this speculative: the full rule set is unknown, and a "theory of everything" may yet unify it.
- No simulation shortcut either. Prediction being impossible leaves simulation, but the universe cannot contain a computer big enough to simulate itself. Reality can only be found out by letting it run.
Why brain emulation is not pre-computation
Kurzweil heads off a confusion about uploading (later in the chapter):
- Emulating is not predicting. Copying how a brain works does not let you compute its future.
- Open systems. Brains, biological or not, take unforeseeable input from the world, which no step-by-step simulation can include in advance.
- Irreducible internals. Scientists have found brain networks structured in up to eleven dimensions, a complexity that probably involves rule 110-style processes with no way to peek ahead.
- Why the universe exists. If the rules are automaton-like, they can be expressed only by unfolding, by reality actually happening. A universe of shortcut-friendly determinism, or of pure randomness, would not need that step-by-step unfolding. And if consciousness can emerge only from class 4 order-and-chaos complexity, that is a philosophical argument for why we exist: without such rules no one would be here to ask.
Compatibilism
This supports compatibilism: a deterministic world can contain free will.
- Free means a decision not caused by something else (such as another person), even though it follows from the underlying laws.
- Determinism would in theory let us see backward or forward in time; under rule 110-style laws the only way to see forward is for every step to actually unfold.
- Conclusion. Seen through panprotopsychism, the emergent processes in our brains "aren't controlling us; they are us." We arise from deeper forces, but our choices cannot be known in advance, so we have free will as long as the processes behind our consciousness can express themselves through our actions.
More than one brain per human
Fiction suggests we already think like panprotopsychists: when an android behaves humanly, audiences root for it as a conscious being, neurons or not. And just as an AI may run many separate algorithms, medicine shows the human brain contains several distinct decision-makers.
Hemispherectomy. Stella de Bode and Susan Curtiss studied 49 children who had half the brain removed to stop life-threatening seizures. Most then functioned normally; even those with a lingering disorder had a reasonably normal personality. Language usually develops on the left, but either hemisphere can master it. Split brain. Michael Gazzaniga (born 1939) studied patients whose corpus callosum, the 200 million axons joining the hemispheres, had been cut for medical reasons. When a word was shown only to the right hemisphere, the left hemisphere, unaware of it, still claimed the resulting choice as its own and confabulated plausible reasons, never recognising a second brain in the same skull. Point: a normal person may host two independent decision-makers inside one conscious identity, each assuming the decisions are its own.
The division goes further. The neocortex, where decisions are made, is many smaller modules; different options may be championed by different modules, each pushing its own view. Marvin Minsky foresaw this in The Society of Mind, picturing the brain as many simpler processes holding different perspectives. Whether each makes a free choice, and how, is unclear: experiments increasingly support the idea, but how neural processes bubble up into consciously perceived decisions is still poorly understood.
"You 2" is conscious. Is it you?
If one identity can span separate structures in one skull, even unconnected ones, what happens when the structures are farther apart? Kurzweil revisits How to Create a Mind: replicating all the information in a brain will be possible within most living people's lifetimes.
Experiment 1: the copy
- Method. Scan a small piece of your brain and build an exact electronic copy of it (a primitive version exists today in treatments for essential tremor and Parkinson's disease). One piece is far too simple to be conscious; repeat piece by piece until there is a complete working replica, You 2.
- Is You 2 conscious? It shares your memories, claims your experiences and acts like you. Unless you rule out any electronic consciousness, the answer is yes: if an electronic brain holds the same information as a biological one and says it is conscious, there is no plausible scientific basis to deny it. Ethically it must be treated as conscious, with moral rights, and panprotopsychism gives positive reason to think it really is.
- Is You 2 you? No. The original you still exists, unaltered, perhaps unaware a copy was made. You 2 acts independently and diverges at once, forming its own memories. If identity is the particular arrangement of information in your brain, You 2 is someone else, conscious or not.
Experiment 2: gradual replacement
- Method. Replace each section of your brain in turn with a digital equivalent, linked to your remaining neurons through a brain-computer interface. There is never a You and a You 2, only you, content after each step.
- The puzzle. At the end your brain is entirely digital and identical to You 2, which we said was not you.
- Kurzweil's resolution. Continuity. The digital brain never diverged from the biological one because they never existed as separate entities. Neuroscience suggests you would not notice small enough changes, the brain adapts well, and the hybrid keeps every pattern that defines you; there is no one else who could be you.
A thought experiment Kurzweil dates to about 2,500 years ago: a wooden ship has its planks replaced one by one. After one plank it is plainly the same ship; after half, or all, most still say so. But store every old plank in a warehouse and reassemble them once the replacement is complete: which ship is the original, the continuous one with no original parts, or the rebuilt one made of them? For ships, identity is human convention. For people it is everything: it matters enormously whether the person beside you is your loved one or a Chalmersian zombie giving a convincing show.
Applying the hard problem:
- Copy case. Whether You 2's subjective self is connected to yours, perhaps one experience spanning both copies even as they diverge, is scientifically unanswerable.
- Gradual case. Much stronger reason to think subjective consciousness survives, because we already do a limited version. Early implants such as cochlear implants only stimulated the brain, replacing no core structure; since the early 2000s, brain prostheses have taken over some work of damaged structures, for instance parts of the hippocampus in memory patients, and the prosthesis can outperform the original. Nobody claims these patients have become zombies. Still infant technology in 2023, Kurzweil expects such prostheses to become more capable and affordable this decade.
Experiment 3: what biology already does
This case is not hypothetical. Even though neurons themselves mostly persist, their components turn over fast:
- Mitochondria in neurons: about half turn over in a month.
- Neurotubules: half-life of several days.
- Synaptic energy proteins: replenished every 2-5 days.
- NMDA receptors in synapses: replaced within hours.
- Actin filaments in dendrites: last about 40 seconds.
- Net effect: the brain is almost completely replaced within a few months, so you are a biological You 2 of yourself a little while ago.
Kurzweil's image is the Charles River near his home: the water molecules in any slice change every few milliseconds, yet the consistent pattern makes it the same river he wrote about a decade earlier. Minds are the same. Information and function, not any particular structure or material, keep identity intact, so as nonbiological parts enter our brains we will still feel like ourselves, only with sharper perception and cognition.
Copies, backups and divided selves
- A digital superpower. Technology that moves skills, personality and memories to a digital medium can also copy them, which biology cannot.
- Backups. Remote copies of mind files protect against accident and disease. Kurzweil denies this is immortality, any more than a spreadsheet in the cloud is immortal (data centres can be destroyed), but it guards against the senseless mishaps that end so many lives.
- Spanning copies. His reading of panprotopsychism: subjective consciousness is tied to information-as-identity, so it may somehow encompass all copies that were once identical. Even a You 2 released to live a different life diverges gradually and continuously, so your consciousness might span both for a while.
- The legal answer. Because You 2 could insist passionately on having a different subjective consciousness (its communication and decision structures being separate), and no one could check, law and ethics would probably have to treat the two as separate entities.
The incredible unlikeliness of being
Your own improbability
- Sperm and egg. A man produces up to about 2 trillion sperm in his lifetime and a woman starts with about 1 million eggs, so the exact pairing that made you had odds of about 1 in 2 quintillion (two million trillion).
- Effectively unique cells. Not every sex cell is genetically distinct, but factors such as age alter epigenetics: chromosomally identical sperm produced at 25 and at 45 would not contribute identically.
- All the way back. The same lottery had to come out right for both sets of grandparents, four sets of great-grandparents, eight sets of great-great-grandparents, and so on to the start of life nearly 4 billion years ago.
- The total. A googol is 1 followed by 100 zeros (the correct spelling of the word behind Google's name); a googolplex is 1 followed by a googol zeros. Your odds of existing are 1 in a number with vastly more than a googolplex of zeros.
A life-permitting universe
That the universe can evolve complex information at all is arguably even less likely. Of all possible configurations, only a sliver permits us. The Standard Model has 37 kinds of elementary particles (differing in mass, charge and spin) interacting through four forces (gravity, electromagnetism, strong and weak nuclear), plus hypothetical gravitons. Constants fix the forces' strengths; life is assumed to need complex chemistry and stable energy sources for hundreds of millions or billions of years. Biologists' prevailing view of life's origin is abiogenesis: a "primordial soup" slowly forming more complex building blocks until proteins assembled into a self-replicating pattern. One broken link and there is no humanity.
- Strong force. Any stronger or weaker and stars could not make abundant carbon and oxygen.
- Weak force. Within one order of magnitude of the minimum for life; weaker, hydrogen quickly becomes helium and no long-lived hydrogen stars like the sun form.
- Quark masses. If the up-down quark mass difference (or the electron's mass relative to it) shifted slightly, protons and neutrons would be unstable. Craig J. Hogan: a change of only a few percent either way prevents life; larger gives a hydrogen-only proton world, smaller a neutron world of nuclei without electrons, so no chemistry.
- Gravity. Weaker, no supernovas to forge heavy elements; stronger, stars too short-lived.
- Density Ω. One second after the big bang it could differ by no more than 1 part in a quadrillion: larger and matter recollapses before stars form; smaller and it disperses too fast to clump.
- Primordial ripples. Density fluctuations averaged about 1 in 100,000; off by more than an order of magnitude, says cosmologist Martin Rees, and the cosmos stays forever dark gas or becomes a violent place of black holes without stable planetary systems.
- Initial entropy. Roger Penrose: only about 1 in 10 raised to 10¹²³ possible universes start with entropy low enough. That number's digit count exceeds the ~10⁸⁰ atoms in the known universe (estimates run 10⁷⁸ to 10⁸²) by 43 orders of magnitude.
The intersection argument and the anthropic reply
- Grant the doubts. Individual calculations can be challenged, and scientists disagree about single factors.
- But combine them. Physicist Luke Barnes argues what matters is the intersection of life-permitting ranges, not their union: every factor must be life-friendly, and one failure means no life.
- The 747. Astronomer Hugh Ross compares chance fine-tuning to a tornado assembling a Boeing 747 in a junkyard.
- The usual answer. Observer selection bias, the anthropic principle: we can only ask the question in a universe that permits askers. Some scientists find that enough.
- Kurzweil's objection. If reality exists independently of observers, it is not fully satisfying. He cites Rees's firing squad: if every marksman misses, you could say you would not be here otherwise, yet "there is something there that needs explaining".
Butler's passage is from Erewhon, published 1872. The fly figure is dated: the FlyWire connectome (2024) counted about 140,000 neurons. The Ship of Theseus is best known from Plutarch (about 75 CE), not 2,500 years ago as Kurzweil says. The "eleven dimensions" (Blue Brain, 2017) are topological cliques of neurons, not spatial dimensions. The fine-tuning tour stacks calculations physicists dispute and skips the multiverse reply. His identity thesis is a philosophical bet that he concedes is untestable.
After Life
From records to replicants
Preserving identity starts with preserving ideas. Our digital lives already form rich records of how we think and feel, and recording, storage and organisation will improve fast this decade. By the late 2020s, Kurzweil predicts, AI will animate these records as highly realistic simulations of specific personalities.
Two techniques are already advancing quickly (as of 2023):
- Transformers trained on someone's writing learn to imitate their style.
- GANs (generative adversarial networks) pit two nets against each other: a generator tries to produce, say, a realistic woman's face; a discriminator tries to tell it from real faces. Each is rewarded (scores it is programmed to maximise) for beating the other, and the contest repeats without human supervision, both improving.
Combined, they already imitate a person's writing, voice and face in video:
- Google Duplex (tested 2018) held unscripted phone calls so believably that people did not know they were talking to a computer.
- Deepfakes can produce harmful political propaganda or harmless recasting: the YouTube channel Ctrl Shift Face showed Javier Bardem's role in No Country for Old Men played by Arnold Schwarzenegger, Willem Dafoe or Leonardo DiCaprio.
- Convergence. Each capability will improve, and together they will be more realistic than the sum of their parts.
Kurzweil calls an avatar with a dead person's appearance, behaviour, memories and skills a replicant (borrowed from Blade Runner) and the phenomenon After Life.
2016 The Verge reports how Eugenia Kuyda used AI and saved texts to "resurrect" her dead best friend Roman Mazurenko.
2019 Kurzweil's Dad Bot, built from his father's writings.
Late 2020s Lifelike replicants built from thousands of photos, hundreds of hours of video, millions of words of chats, data on habits and interests, and interviews with people who knew the person; many will sit in the uncanny valley.
By ~2036 Turing-level AIs re-create departed people (his phrasing: sooner than the gap from How to Create a Mind, 2012, to the reader's present).
Late 2030s Convincing physical android bodies via nanotechnology.
Early 2040s Nanobots copy a living brain's memories and personality: a true You 2 that passes a person-specific Turing test.
Next century Full quantum-level emulation becomes computable.
Function without form
Late-2020s replicants will closely resemble the original but with subtle differences that may disturb loved ones (the uncanny valley). They are not You 2: they re-create the function, not the form, of the information in the person's brain, so on panprotopsychist grounds they do not revive subjective consciousness. Many will still value them for continuing important work, sharing memories and helping families heal, as Kuyda's story shows. Reactions will vary for cultural, ethical and personal reasons, but the technology will be available to those who want it.
Bodies and Moravec's paradox
Replicant bodies will live mostly in virtual and augmented reality, with convincing physical androids possible via late-2030s nanotechnology. Android function faces Moravec's paradox, named for Kurzweil's friend Hans Moravec:
- Easy for machines, hard for us: square roots of big numbers, remembering masses of information.
- Easy for us, hard for machines: recognising faces, keeping balance while walking.
- Why. The second group evolved over tens or hundreds of millions of years and runs in the brain's background; "higher" cognition sits in the neocortex, roughly modern only for several hundred thousand years.
Progress against the paradox is rapid:
- 2000: Honda's ASIMO impressed experts by walking gingerly across a flat floor without falling.
- 2020: Boston Dynamics' Atlas ran, jumped and tumbled through obstacle courses with more agility than most people.
- Social robots: Sophia and Little Sophia (Hanson Robotics) and Ameca (Engineered Arts) show emotion on humanlike faces; headlines sometimes oversell them, but the trajectory is clear.
- Later: replicants (and the living) choose among many bodies, possibly cybernetically augmented biological bodies grown from the original's DNA; molecular-scale nanotechnology then allows bodies far beyond biology, and replicants likely leave the uncanny valley for most observers.
Society's questions
How people react will depend on beliefs about souls, consciousness and identity. Does it matter whether a replicant comes from AI and data mining or from a full uploaded You 2? Kurzweil expects laws on who may create replicants and how they may be used; some people will forbid being replicated, others will leave detailed instructions or help build their replicant while alive. He lists six open questions:
- Rights. Are replicants people with full human and civil rights, such as voting and making contracts?
- Liability. Are they bound by the original's contracts or responsible for their crimes?
- Credit. May they claim the original's work and social contributions?
- Marriage. Must you remarry a late spouse who returns as a replicant?
- Discrimination. Will they be ostracised?
- Limits. When should creating them be restricted or banned?
Turing-level replicants will be as cognitively complex as biological people, so on Kurzweil's view they will be conscious and will believe they are the person. In the early 2040s nanobots will copy all the data of a living brain; the result could pass a person-specific Turing test and, by all detectable evidence, be as real as the original, so if identity is information such as memories and personality, it is that person. Relationships could continue, sexual ones included. Small differences may remain, but living people also change, gradually or suddenly through war, trauma or shifts in status and relationships. A Chalmers-style view gives reason to think the subjective self could persist this way, but that can be neither proved nor disproved, so each person must decide by their own philosophical or spiritual values. Copying a living brain marks the move from simulated replicants to true mind uploading.
Levels of whole-brain emulation
Mind uploading, or whole-brain emulation (WBE), can mean very different computational loads. John Fiala, Anders Sandberg and Nick Bostrom (2008) defined eleven levels; Kurzweil compresses them to five, from most abstract to most exhaustive.
| Level | What it reproduces | Cost and fidelity |
|---|---|---|
| Functional | Behaves like the mind, without copying the person's brain structure | Cheapest; least faithful |
| Connectomic | Hierarchical links and logical relations between groups of neurons, not every cell | Moderate |
| Cellular | Key information on every neuron, without detailed internal physics | High |
| Biomolecular | Protein interactions and small dynamic forces inside each cell | Very high |
| Quantum | Subatomic effects within and between molecules | Most complete; likely not computable until next century |
Deciding which level suffices is, he says, a major research question for the next two decades. Those who insist on quantum emulation usually believe consciousness depends on unknown quantum effects. Kurzweil disagrees: if panprotopsychism is right, consciousness comes from how information is arranged, so a missing protein molecule is irrelevant. His analogy is a JPEG: identical on a floppy disk, CD-ROM or USB stick as long as the bits match, and it would even survive being copied out in pencil, mailed, and typed back in by hand. The practical tasks are therefore to interface computers with the brain and to crack the brain's code for representing information, which the superhuman AI tools of the 2030s will make possible. (His endnotes point further, to work on emulation costs and to Matrioshka brains, a proposed way of harnessing vast energy for computation.)
Talking to the Dad Bot
At Google, Talk to Books could swap its 100,000+ books for any custom collection. In 2019 Kurzweil gathered everything his father, the conductor Fredric Kurzweil, had written (love letters to Ray's mother, course notes, a book on music in progress, personal reflections), and his daughter Amy questioned it. Kurzweil judges the answers coherent, natural and surprising, convincing enough to pass for a living conversation and, to those who knew him, recognisably his father's voice.
| Question | Gist of the Dad Bot's answer |
|---|---|
| What do you love about music? | It adapts a general plan to its own needs and stays tied to human feeling |
| Favourite composer? | Brahms, who builds on Beethoven |
| Essence of rock and roll? | Melody and rhythm; repetition and contrast, unity and variety |
| How did you start conducting? | Over twenty years with professional, amateur, college, community and opera orchestras |
| How to be remembered? | For his skill with people, not only artistic ability |
| Regrets? | Days of dull lessons leaving little time for valuable work; hope outrunning practical judgement; money worries |
| What makes a person who they are? | Form shown in how the parts relate to the whole |
| Hobbies? | Gardening: peaches and grapes, relaxing but endless work |
| Anxieties? | "Often nightmarish" |
| Modern technology? | Machines have made the individual small; innovators are always condemned |
| A car accident? | His wife hit by a taxi while he conducted opera in Mobile: months in traction, then a body cast |
| Meaning of life? | Love |
Kurzweil calls it the first step in bringing his father back. He plans a replicant of himself, built from thousands of pages of correspondence, all his published writing and countless hours of video, far richer than his father's record; future replicants will also draw on metadata about a person's life.
Who can I be?
Identity ties together consciousness, free will and determinism. Kurzweil describes himself as both the result of incredibly precise prior conditions and the product of his own choices: a self-modifying information pattern shaped by whom he met, what he read and where he went.
His limits are biological:
- A prehistoric brain that predisposes him to habits he would rather not have.
- Slow learning and weak memory, too limited to know all he wants to know.
- No reprogramming to remove fears, traumas and doubts that hold him back.
- An ageing body, slowed by hard work but programmed to destroy the information pattern that is Ray Kurzweil.
Humans have long been gaining control over who they become: medicine overcomes injury and disability; cosmetics shape appearance; legal and illegal drugs correct imbalances or alter consciousness; information forms mental habits that rewire the brain; art and literature build empathy and virtue; apps build discipline and healthy habits; transgender people can increasingly align their bodies with their inner identity. Direct brain programming will go much further.
Merging with superintelligent AI is a means to a higher end. Backed up on a better digital substrate, our powers of self-modification are fully realised: behaviour can match values, lives are no longer marred and cut short by biology, and humans can at last be truly responsible for who we are.
2026 scorecard
| Prediction (book) | Date | Status, Oct 2026 | Evidence |
|---|---|---|---|
| Lifelike replicants of the dead | Late 2020s | On track | Commercial griefbots; victim avatar in US court (2025) |
| Developers set AI ethics before law | At Turing-level AI | On track | Anthropic model-welfare programme (Apr 2025) |
| Better, cheaper brain prostheses | 2020s | Partly | ~21 Neuralink users; no hippocampal product |
| Agile androids | 2030s | Ahead on motion | Unitree shipped 5,500+ humanoids in 2025 |
| Nanobots copy a living brain | Early 2040s | Behind | No brain-interfacing nanobots exist |
Replicants arrived as products (StoryFile, HereAfter), and in May 2025 an Arizona court heard an AI avatar of road-rage victim Chris Pelkey before sentencing his killer to 10.5 years. Kuyda went on to found Replika. Verdict: After Life is on or ahead of schedule; its legal questions remain open.
FlyWire published the first complete adult fly connectome (about 140,000 neurons, 50 million+ synapses) in Nature in October 2024. In March 2026 the startup Eon Systems ran a model built from it in the NeuroMechFly v2 virtual body, where it groomed and fed (not peer reviewed); a mouse has 500× more neurons. Verdict: whole-human emulation by the early 2040s is far behind any visible trajectory.
Life Is Getting Exponentially Better
Kurzweil argues that nearly every measure of human well-being (literacy, sanitation, health, income, safety, clean energy, democracy) has improved for two centuries, and that the public believes the opposite because news incentives and evolved biases hide gradual gains. The driver is the law of accelerating returns (LOAR): information technologies feed their own improvement, and those gains spill into everything else through mutually reinforcing loops. Everything achieved so far came from the slow early stretch of these curves; as AI turns food, clothing, housing, energy and medicine into information technologies, the next twenty years will bring more progress than the last two hundred.
The public consensus is the opposite
Headlines nobody prints
Kurzweil opens with three mock news flashes: extreme poverty fell 0.01% today, literacy rose 0.0008% since yesterday, flush-toilet access grew 0.003% today, and the same was true yesterday and the day before. The point is that good news is gradual, so it never qualifies as news. The figures are real daily averages of recent multi-year trends:
| Mock headline | Underlying data | Period | Implied rate |
|---|---|---|---|
| Extreme poverty down | 787 m → 697 m people below $2.15/day (2017 $) | 2016-2019 | ~4% a year, ~0.011% a day |
| Literacy up | 85.5% → 86.8% of the world (UNESCO) | 2015-2020 | ~0.0008% a day |
| Sanitation up | 73% → 78% with "basic" or "safely managed" sanitation | 2015-2020 | ~0.003% a day |
He admits the poverty extrapolation to the present is uncertain but claims it is right to within an order of magnitude.
Where this chapter sits
The facts themselves are well documented. Kurzweil had covered them in The Age of Spiritual Machines (1999) and The Singularity Is Near (2005); Peter Diamandis and Steven Kotler's Abundance (2012) described a coming era of abundance in formerly scarce resources; Steven Pinker's Enlightenment Now (2018) catalogued social progress. Kurzweil's distinct contribution is the claim that the progress is exponential, that the LOAR is its common driver, and that the result will be dramatic improvement in most areas of life, physical as well as digital, in the very near future.
What the LOAR does and does not cover
Kurzweil corrects a common misreading: he does not claim that all technology improves exponentially. The LOAR applies to technologies that create feedback loops that accelerate innovation, which in practice means technologies that give mastery over information (gathering, storing, manipulating, transmitting it), because those make the next round of innovation easier.
- Printing press. Cheap books educated the next generation of inventors.
- Computers. Today's CPUs help chip designers build the next, faster CPUs.
- Broadband. Cheaper access lets more people share ideas, making the network more useful to everyone.
- Moore's law. The best-known exponential is just one manifestation of this deeper process.
Information technology. Each generation helps build the next, so the costs of innovation stay below its benefits and progress compounds. As AI spreads through the 2020s it will bring the same curve to medicine, then to things we do not think of as information technology: food, clothing, housing and land use.
Transatlantic travel. Mayflower 1620: 66 days; 1775: ~40 days; paddle steamer Great Western 1838: 15 days; liner Deutschland 1900: 5 days 15 hours; turboelectric Normandie 1937: 3 days 23 hours; Pan Am flying boat 1939: 36 hours; first jet service 1958: under 10.5 hours; Concorde 1976: 3.5 hours. Concorde retired in 2003 and London-New York is back over 7.5 hours. Jet engines are not used to build jet engines, so extra speed eventually costs more than it is worth.
Why the news misleads
Kurzweil's diagnosis combines media incentives with evolved and cognitive biases. Storytelling from myth to Star Wars runs on escalating danger, news imitates it, and engagement-optimised social media algorithms, built to maximise emotional response and ad revenue, push it further. Nobody has to intend this: because audiences attend to threats, an industry paid by attention learns collectively that threatening stories pay. The result is a selection bias that sends crises to the top of the feed and gradual progress to the bottom.
| Bias or mechanism | How it works | Evidence Kurzweil cites |
|---|---|---|
| Threat vigilance | Survival favoured noticing the rustle in the leaves over a 0.1% better harvest | Evolutionary argument |
| No memory of growth | Pre-modern lives saw no change in living standards, so neither biology nor culture taught us to notice it | English GDP per head (2023 £): £1,605 in 1400 and the same 80 years later; £1,586 → £1,604 for 1500-1580; £6,734 → £20,979 for 1900-1980 |
| Constant monitoring | We scan feeds asking whether events endanger us, crowding out slow positive news | Pamela Rutledge, Media Psychology Research Center |
| Fading affect | Bad memories fade faster than good ones | Richard Walker (Colorado State, 1997): ratings rechecked at 3 months, 18 months, 4.5 years; 2014 study (Australia, Germany, Ghana and others) found it worldwide |
| Nostalgia | Transforms past stress into fond memory; a coping mechanism that makes the past look better | Word coined 1688 by Swiss physician Johannes Hofer (nostos homecoming + algos pain); Clay Routledge (North Dakota State) found nostalgic writing raised self-regard and social bonds |
| Negative trend-reading | Small negative fluctuations read as fundamental shifts more often than positive ones | 2017 study (markets, hurricanes, unemployment), summarised by Art Markman |
| Urgency | "News" must be new; sudden events are mostly bad, gradual ones have no footage | Front pages of the New York Times, CNN; disaster videos share easily |
| Base-rate neglect | Vivid detail beats prior probability: guessing "librarian" for a book lover instead of the likelier "retail worker" | Kahneman (Nobel, Economics) and Tversky |
| Gambler's fallacy | Expecting heads after a run of tails, misreading regression to the mean | Kahneman and Tversky |
| Availability heuristic | Frequency judged by how easily examples come to mind; the news supplies bad ones | Kahneman and Tversky |
Together these leave us conditioned to expect entropy, a default sense that things fall apart. Kurzweil grants the bias is partly useful (it prepares us for setbacks) but says it hides real improvement, with political costs: 51% of Americans (Public Religion Research Institute, 2016) said American culture had changed for the worse since the 1950s, and 71% of Britons (YouGov, 2015) said the world was getting worse against 5% better. Such moods reward populists who promise to restore a past that was objectively worse.
- 2% of 31,786 respondents (26 countries, 17 languages, 63% of world population; Martijn Lampert, Anne Blanksma Çeta and Panos Papadongonas, 2018) knew extreme poverty had halved in twenty years.
- 24% vs 0.7%: UK public estimate of fraudulent benefit claims against the real figure (Ipsos MORI for the Royal Statistical Society and King's College London).
- -53%: crime in England and Wales 1995-2012, yet 58% thought it rose or held steady; violent crime fell 20% (2006-2012) while 51% thought it rose.
- 15% vs 0.6%: estimated versus actual annual pregnancy rate for UK girls under 15, 25× too pessimistic.
- Up to 78% of Americans this century said national crime was rising, though violent and property crime roughly halved after 1990 (Gallup; the Marshall Project, 2022).
"If it bleeds, it leads" sums it up: a shooting is covered at length, while crime prevented by data-driven policing or better police-community communication is literally a non-event. Pinker's version is that news reports what happened, never the officer not shot or the city without a riot, so as long as violence is above zero there will be headlines, and "pessimism can be a self-fulfilling prophecy". Global social media makes it worse, since earlier generations heard mostly local bad news. Kurzweil's counter-maxim: optimism is a self-fulfilling prophecy too, because believing a better world is possible motivates the work of building it. He stresses that correcting the biases does not mean ignoring present suffering or assuming progress is automatic; it takes human ingenuity and effort, and the long trends show that the species is turning the tide.
Nearly every aspect of life is getting better
Information technology improves exponentially because it contributes to its own innovation, and it also drives mutually reinforcing mechanisms in other areas. Over two centuries this has formed a virtuous circle lifting literacy, education, wealth, sanitation, health, democracy and safety together.
- Technology outlasts wealth. Economic cycles rise and fall and fortunes are lost, but technological knowledge is essentially permanent: a one-way ratchet that has outweighed wars, plagues and natural disasters.
- Cross-domain loops. Better education yields better doctors, who keep more children healthy enough to stay in school.
- Indirect benefits. Twentieth-century domestic appliances saved time and toil and also helped bring millions of women into the workforce, where they contributed across every field.
- The print example. The press broadened education, giving a more capable workforce; literacy improved coordination of production and trade; wealth funded infrastructure and schools; and mass print fed democratisation, which over time lowered violence.
At first the change was so slow that grandparents and grandchildren lived almost identical lives. The trend has accelerated in recent decades as information technology curves steepened, and Kurzweil expects it to go into very high gear over the next twenty years.
| Domain | Starting point in the book | Latest figure in the book | Shape |
|---|---|---|---|
| World literacy | <10% (1800) | ~87% (2020) | S-curve, near ceiling in rich countries |
| US years of schooling | ~4 (1870) | 13.7 (2021) | Rising, flattening |
| US flush toilets | ~15% (1900) | ~99%; world sanitation 78% (2020) | Saturated S-curve |
| World electricity access | Cities only (1900) | 90%+ | Saturating |
| UK/US life expectancy | ~20s (Europe, 1,000 years ago); 40s (mid-1800s) | 80+ in much of the rich world (2023) | Slowing since mid-1900s |
| World extreme poverty | 84% (1820) | 8.4% (2019) | Steady decline, faster since 1990 |
| US daily income (2023 $) | $7.06 (1774) | $191 (2020) | Exponential |
| W. European homicide | ~33/100k (1300s-1400s) | <1/100k | >97% fall |
| Solar share of electricity | ~0.00003% (1983) | 3.6% (2021) | Doubling every 28 months |
| People in democracies | 3% (1900) | ~54% (1999); "nearly half" now | Uneven, with reversals |
| Computation per $ | 0.0000065 (Z2, 1939) | ~130 billion (TPU v5e, 2023) | Exponential, ~16 orders of magnitude |
Literacy and education
Illiteracy was rational, Kurzweil argues, not a sign of incurious ancestors. Reading was mostly unnecessary because copying text was expensive: if you rarely saw and could never afford written material, learning to read was a poor use of time, the one resource everyone gets equally (twenty-four hours a day). The same logic still operates: where there are few IT jobs, studying programming may not pay, but machine translation, distance learning, natural-language programming and telework are changing that incentive just as cheap print did in Europe.
The long climb
- Late Middle Ages. Fewer than one in five Europeans could read; literacy was mostly limited to clergy and trades that required it. Movable type then made cheap, varied reading matter available.
- 1750. Among major European powers only the Netherlands and Great Britain exceeded 50% literacy.
- 1870. Only Spain and Italy still lagged notably below half, likely because of undeveloped economies and recent civil wars.
- World, 1800. Probably fewer than one in ten could read; mass-produced newspapers and reforms guaranteeing basic schooling spread literacy through the nineteenth century.
- World, 1900-1910. Still under one in four in 1900; past one in four by 1910 as public education expanded.
- World, 1970. A majority literate for the first time.
- Today. Nearly 87% worldwide; above 99% in many developed countries.
Basic literacy is a low bar (reading and writing short messages such as one's name), and richer metrics show the remaining gap. The US National Assessment of Adult Literacy found in 2003 that only 86% of Americans scored above "below basic", and a similar assessment nine years later found no significant improvement.
Years of schooling
| Country | 1870 (avg. years of education) | 2021 | Path |
|---|---|---|---|
| United States | ~4 | 13.7 | Led throughout; ~9 by 1945 |
| UK, Japan, France | <1 | >10 each | Caught up fast with free public schooling in the early 1900s |
| China | <1 | 7.6 | Major leap in the two decades after WWII |
| India | <1 | 6.7 | Same; steepest rise since 1980 |
US spending on education rose from about $700 to over $5,000 per person (constant 2023 dollars) between 1949 and 2019. Kurzweil notes the steepest gains came in the past half century, the same period in which computers both aided teaching and raised the payoff to schooling, which he treats as no coincidence.
Flush toilets, electricity, radio, television and computers
Toilets
Contamination of food and water by human faeces was historically among the greatest causes of disease and death, and the flush toilet is the definitive fix. In the US it appeared as early as 1829, became common in cities only around 1900, spread to rural areas in the 1920s-30s, and reached 75% of households by 1950 and 90% by 1960. In 2023 most remaining American homes without one reflect lifestyle choice (rustic living), not poverty. In developing countries poverty is still the obstacle, but world access to safe toilets keeps rising as sanitation technology gets cheaper and once-violent regions stabilise enough to invest.
Electricity
Electricity is not itself an information technology, but it powers all of them, and even before computers it ran labour-saving appliances and let people work and play at night. US electrification began in big cities, slowed at the start of the Great Depression, then advanced through Franklin D. Roosevelt's rural electrification programmes in the 1930s-40s: over 95% of homes by 1951, regarded as complete by 1956. Elsewhere the pattern repeats (cities, suburbs, countryside) and over 90% of the world now has power.
For the remaining tenth, Kurzweil says the barrier is political, not technological. He cites Daron Acemoğlu (MIT) and James Robinson on inclusive political institutions: where people participate freely and feel secure enough to invest, prosperity loops take hold. Where violence is common, nobody builds costly grid infrastructure that may be destroyed, and dangerous roads block delivery of generators and fuel. Cheap, efficient photovoltaics sidestep both problems.
Radio, television, computers
| Technology | Key US adoption milestones | Latest figure in the book |
|---|---|---|
| Flush toilets | From 1829; urban ~1900; rural 1920s-30s; 75% 1950, 90% 1960 | World improved sanitation ~78% (2020) |
| Electricity | Urban ~1900; FDR rural programmes 1930s-40s; 95% 1951, "complete" 1956 | 90%+ of world population |
| Radio | Commercial from 1920; main mass medium by the 1930s; 90%+ of households 1950 | Weekly online listening 12% (2007) → 62% (2021) |
| Television | Demonstrated late 1920s; viable 1939, halted by war; majority 1954; 90%+ 1962 | Peak 98.4% (1997); 96.2% (2021) |
| Computers | Kenbak-1; Altair 8800 kit (1975); Apple and Microsoft by decade's end; ownership nearly doubled within 5 years of the 1984 Super Bowl ad | 93.1% of US homes (2017-21); ~2/3 of world (2022) |
- Radio built a national culture. Unlike metropolitan newspapers, broadcasts reached from California to Maine, so the whole country heard the same speeches and shows. When TV arrived in the 1950s, radio shifted to news, politics, sport and car listening; since the 1980s partisan talk radio has been criticised for reinforcing listeners' biases; since the 2010s much radio is streamed.
- Radio's apparent decline is misleading. One study (different method) found household radio ownership falling from 96% (2008) to 68% (2020), but 85% of US adults had a smartphone by 2021, which streams radio free.
- Television grew faster than radio because the country was richer. Slower growth after 1962 reflected late adopters; the slight decline after 1997 reflects less TV watching, online pastimes and streaming on other devices.
- Computers are interactive, unlike passive radio and TV. Early home use was word processing, data entry and simple games.
- The 1990s internet loop. DNS hosts rose from about 175,000 (January 1990) to about 72 million (January 2000); traffic from about 12,000 GB (1990) to 306 million GB (1999). More content attracted more users, who added content of their own, just as a streaming service with a bigger catalogue draws more subscribers. Then came e-commerce and, in the 2000s, social media.
- Continuing spread. US ownership keeps rising as the Greatest Generation passes and millennials form households; smartphones carried computers into the developing world.
Life expectancy
Medicine so far has been a linear, hit-or-miss process (detailed in chapter 6): without tools to search all possible treatments systematically, discovery owed much to chance. The great example is the accidental discovery of penicillin, which launched the antibiotic era and may have saved 200 million lives. Even deliberate research has relied on slow high-throughput screening and lab work instead of exhaustive simulation of candidate molecules.
The approach still delivered. European life expectancy at birth was in the twenties a thousand years ago (infant and youth deaths from cholera and dysentery), in the forties in mid-nineteenth-century Britain and America, and over 80 in much of the developed world by 2023: nearly tripled in a millennium and doubled in two centuries, mostly by avoiding or killing external pathogens.
That low-hanging fruit is picked. What remains (cancer, atherosclerosis, diabetes, Alzheimer's) arises inside the body as cells malfunction and tissues wear out, so gains have slowed since the mid-twentieth century: US life expectancy at birth rose from about 39 to 49 in 1880-1900 but only 74 to 76 in 1980-2000, after medicine's focus shifted from infection to chronic and degenerative disease.
- Bridge one: lifestyle. Diet, exercise and supplementation reduce risk but only delay the inevitable.
- Bridge two (2020s): AI plus biotechnology. AI already finds drugs; by the end of the 2020s digital simulation starts to augment and then replace slow, underpowered human trials, turning medicine into an information technology. The trickle of clinical AI becomes a flood, letting us attack the factors that cap life span near 120 years: mitochondrial mutations, shortened telomeres and the uncontrolled cell division of cancer.
- Bridge three (2030s): medical nanorobots. Some biomolecules already count as nanobots by some definitions; the difference is AI control for varied tasks. Cell-level maintenance and repair gives the control we have over a car: unless it is wrecked outright, parts can be repaired or replaced indefinitely, definitively defeating ageing.
- Bridge four (2040s): mind backup. If identity is the pattern of information the brain represents (chapter 3), scanning it accurately enough lets it run on digital substrates, so destroying the biological brain would not end the person, who could be copied and recopied to safe backups.
Genome sequencing is the model case of medicine turning into information technology. Since the Human Genome Project finished in 2003, sequencing cost has halved roughly every year, from about $50 million per genome to as low as $399 in early 2023, with one company promising $100 tests. There was a plateau in 2016-2018 and COVID-era slowing, and Kurzweil expects AI to re-accelerate the trend.
Poverty and income
Individual technologies help, but Kurzweil says the transformation comes from their aggregate, mutually reinforcing effect, and economic well-being, though incomplete, is the best long-run metric for it.
The world
In 1820 about 84% of humanity lived in extreme poverty by modern standards. Industrialisation cut poverty first in Europe and America; after WWII the decline sped up as modern agriculture reached India and China. By 2019 extreme poverty (below $2.15 a day, 2017 dollars) was about 8.4%, after falling by more than two-thirds between 1990 and 2013, even as heavier media coverage of poverty convinced many in rich countries it was worse than ever.
- East Asia. China's development cut the region's extreme poverty by 95% from 1990 to 2013, while its population grew from 1.6 to 2 billion.
- Europe and Central Asia. The only region where extreme poverty rose for much of that period. The cause was political: Soviet collapse left a power vacuum and enormous corruption, and in the poorer Central Asian republics negative feedback loops deterred investment.
- Aid politics. After the Soviet fall many rich nations cut foreign aid, because Cold War development had been strategic competition for influence. By the mid-1990s the OECD re-framed development as both humanitarian and in everyone's interest; in 2000 the UN set the Millennium Development Goals for 2015. Many targets were missed, yet they improved hundreds of millions of lives.
The United States: absolute versus relative poverty
In absolute terms (the global $2.15 line), US extreme poverty has been at or below 1.2% since measurement began. Relative poverty, measured against one's own society, fell from about 45% in the nineteenth century to about 12.5% in 1970 and then stagnated in the teens. Kurzweil's explanation: the line is periodically redefined upward, so people not considered poor by 1980 standards count as poor now.
- 2014-2019. The number of poor Americans fell by about 12.6 million while population grew about 8.9 million; 4.1 million of the drop came in 2019 alone, when elderly poverty neared an all-time low. COVID caused a temporary rise against the downtrend since 2008.
- Unmeasured gains. Today's poor have free MIT open courses, video calls with family on other continents, translation and directions on a cheap smartphone: capabilities millions of dollars could not buy a few decades ago, which statistics do not capture (the mismeasurement is chapter 5's topic).
- The line. In 2023 the US poverty line for one person was $14,580, or $39.95 a day; the average (not median) American has been above it in real terms since about 1941.
| Year | Average US daily income per person (2023 $) | US relative poverty rate (approx.) |
|---|---|---|
| 1774 / 1800 | $7.06 / $5.65 | n/a |
| 1840 / 1860 | $8.37 / $13.27 | n/a |
| 1870 / 1880 | $15.08 (1880) | ~45% (1870) |
| 1900 / 1910 | $20.08 / $26.45 | ~30% (1910) |
| 1930 / 1935 | $30.74 (1930) | ~45% (1935) |
| 1940 | $35.47 | n/a |
| 1950 | $52.97 | ~30% |
| 1960 | $65.27 | 22.2% |
| 1970 | $89.82 | 12.6% |
| 1980 | $102.41 | 13.0% |
| 1990 | $124.32 | 13.5% |
| 2000 | $147.18 | 11.3% |
| 2010 | $154.15 | 15.1% |
| 2015 | $169.88 | 13.5% |
| 2020 | $191.00 | 11.5% |
Growth, income and work
- GDP per capita. Total GDP grows partly because population grows, but GDP per capita also grows exponentially (Kurzweil shows linear and log charts from the Maddison Project, BEA and Federal Reserve), and still excludes free information products and IT's rising power per dollar.
- Personal income per capita. It covers what people earn (wages plus dividends and business profits), not corporations. Since records began in 1929 it has risen more than fivefold in real terms, dipping only in the Depression and major recessions, despite fewer hours worked.
- Median income. Slower, but still rising: $27,273 (1984) to $42,488 (2019), in 2023 dollars.
- Income per hour worked. Real personal income per hour rose from about $5 (1880) to about $93 (2021). It includes investment, business income, benefits and one-off payments such as the 2020 stimulus, so it exceeds wages alone; Kurzweil values it because non-wage income is a growing share.
- The Depression paradox. Hourly income held up and even grew in the Depression: jobless people lost income and hours, shrinking numerator and denominator together, while those still employed saw wages fall little.
What GDP misses: price, freedom and variety
Kurzweil argues the income figures understate progress three ways: electronics cost far less than they did; search engines and social networks are free; and globalisation brought variety that is hard to price. Even if you can eat only one cuisine at a meal, you prefer having the choice. Hundreds of TV channels instead of three; out-of-season fruit from the other hemisphere; millions of books on Amazon instead of a few thousand in a shop. Variety matters most to people with unusual tastes: the collector of old kaleidoscopes can buy from anywhere on eBay, and a science-loving child can watch hours of educational video instead of Gunsmoke on the family's only set, as in Kurzweil's own childhood. 3D printing and later nanotechnology will multiply choice further.
Hours of work
- Late 1800s. Nearly 3,000 hours a year per worker.
- After about 1910. Rapid fall as regulation and unions shortened days and added time off, and employers found rested workers more productive and accurate, so they spread work across more people.
- Depression. Below 1,750 as firms cut even surviving employees' hours.
- WWII and the postwar boom. Back above 2,000.
- Since then. A slow decline to roughly Depression-era levels, but now largely by choice (part-time work, work-life balance); some European countries fell further.
- Remote work. Millennials and Gen Z seek creative, often entrepreneurial careers; COVID pushed telework, which saves commuting but blurs work and life; in one study 98% wanted the option to work remotely for the rest of their careers.
Child labour
The International Labour Organization uses three nested categories:
- Child employment. The broadest: light work for modest hours on family farms or businesses, a relatively mild form that can still distract from school.
- Child labour. Work resembling an adult job in hours and arduousness.
- Hazardous work. The narrowest and worst: mining, shipbreaking (dismantling ships for value and disposal) and waste handling.
The share of the world's children in hazardous work fell from 11.1% to 4.6% between 2000 and 2016, though COVID's economic disruption appears to have interrupted the decline.
Decline in violence
Prosperity and safety reinforce each other: people with much to lose economically have strong reasons not to fight, and people who expect long, safe lives have reason to make long-term investments that benefit everyone.
Western Europe since 1300
Homicide has fallen consistently since at least the fourteenth century, an exponential decline over the long run despite ever deadlier personal weapons. In the countries with medieval data, annual homicides fell from a per-country average of about 33 per 100,000 (fourteenth-fifteenth centuries) to under 1 today, a drop of more than 97%. The series counts "ordinary" homicide (murder, manslaughter), not war or genocide. Kurzweil's chart follows Max Roser and Hannah Ritchie (Our World in Data) and Manuel Eisner's data before 1990, but swaps WHO figures for UNODC data (1990-2020), which he says better match other sources and official police estimates.
The United States
US murder and violent crime have been in long-term decline since about 1991. Homicide fell from 9.8 per 100,000 in 1991 to a low of 4.4 in 2014; the rise to 6.8 in 2021 still leaves it over 30% below the peak. Twice in the twentieth century violence ran at about double today's level:
- 1920s-30s. Prohibition and the gangs built on bootleg liquor.
- 1970s-90s. The illegal drug trade brought violence back to city streets; violent crime hit all-time highs by the early 1980s.
What drove the decline, in Kurzweil's account:
- Broken windows. George Kelling and James Q. Wilson observed that graffiti and vandalism made neighbourhoods feel unsafe and signalled that worse crime would go unpunished; policing minor offences became a strategy against serious ones.
- Proactive policing. More foot patrols in high-crime areas and data-driven deployment of police resources; together these appear to explain much of the drop through the 1990s and 2000s.
- The cost. In some cities broken-windows policing went too far and fell disproportionately on minority communities. The 2020s task is to keep crime falling while addressing racial disparities; body cameras, citizens' phone cameras, automated gunshot detectors and AI data analysis can help if used responsibly.
- Lead. Lead from petrol exhaust and household paint impaired children's cognitive development and, at population level, raised violent crime, likely through lower impulse control. Environmental rules since roughly the 1970s cut childhood exposure and are thought to have reduced violence.
Pinker's long view
- ~50×: fall in European homicide since the Middle Ages, more in some places (The Better Angels of Our Nature, 2011).
- 110 vs <1: homicides per 100,000 a year in fourteenth-century Oxford against modern London.
- ~500×: estimated fall in violent death overall since prehistory.
- 524: average annual war deaths per 100,000 in 27 pre-state societies (hunter-gatherers and hunter-horticulturalists, likely typical of prehistory).
- 144, 27, 135: twentieth-century annual war death rates per 100,000 in Germany, Japan and Russia, among the hardest-hit nations in a century of world wars, genocide and atomic bombs.
- 3.7: the United States' annual war death rate over the twentieth century, despite fighting worldwide.
People still think violence is rising because of what Pinker calls historical myopia: recent events loom large and older, worse episodes are forgotten, which is the availability heuristic at work. Documentation technology makes it worse: we see recent atrocities in colour video, the nineteenth century in black-and-white photographs and earlier eras only in text and a few paintings.
Why the circle keeps turning
Kurzweil and Pinker both explain the decline by virtuous circles. Safety makes it worth building schools and writing and reading books, which favour reason over force, which lowers violence further. Peter Singer's "expanding circle" of empathy has widened from clan to nation to foreigners to non-human animals, alongside the rule of law and norms against violence. Kurzweil's addition is that the circle is driven by technology: books, then radio and TV, then the internet let people exchange ideas with ever wider groups, and footage from distant disasters, for all its distorting effect, extends moral concern across the species.
The deepest lever is abundance. Struggle over scarce resources feels like human nature, but digital goods have already escaped scarcity. Two children may fight over one printed comic, since only one can read it at a time; nobody fights over a PDF, because your copy does not deprive me of mine. Once very cheap energy (solar, later fusion) and AI robotics make many physical goods equally easy to reproduce, violence over them will look as silly, powered by a millions-fold gain in information technology between now and the 2040s.
Growth of renewable energy
Fossil fuels are unsustainable for two reasons: pollution and greenhouse gases, and dependence on scarce resources that cost more to extract as demand soars. Renewable costs fall exponentially because supercomputing (and recently deep neural nets) speeds discovery of materials for solar cells and storage, so generation from solar, wind, geothermal, tidal and biofuels grows exponentially too. Solar supplied about 3.6% of world electricity in 2021, a share doubling every 28 months since 1983.
- Swanson's law. The name for the near five-decade exponential fall in module cost per watt.
- Balance of system. Modules are the largest single cost, but permits and installation labour can bring the total to about three times module cost, and those fall more slowly. AI and robotics can cut labour and design costs; faster utility planning and permitting needs policy.
- Wind. US Department of Energy data show levelised cost for land-based wind falling from about 12 cents/kWh (1998) to about 3.6 cents (2021), after a rise to nearly 11 cents around 2009.
Spread of democracy
Cheap energy will create abundance, but sharing it fairly needs democracy, and information technology has a long record of democratising societies. Kurzweil traces democracy's spread from medieval England and ties it largely to mass communication.
The printing press as a LOAR engine
Ideas are information: abstract orderliness (as opposed to chaotic entropy) that can be carried by any medium, from stone tablets and parchment to punch cards, magnetic tape and voltages in silicon. You could chisel the source code of Windows 11 in stone, but the medium decides whether storing, copying and sending ideas is practical. Medieval scribes made copying slow and dear. Because more ideas make new ideas easier, technologies that share ideas speed invention, including inventions that share ideas better still.
After Gutenberg (about 1440), middle-class people could afford books in quantity, innovation flourished, the Renaissance spread quickly, and further printing improvements made books several thousand times cheaper by the early seventeenth century than before the press.
From Magna Carta to mass democracy
1215 Magna Carta, signed by King John, states commoners' right not to be unjustly imprisoned, but such rights are often ignored for centuries.
c. 1440 Gutenberg's movable-type press lets educated classes spread news and ideas.
1500s-1600s Knowledge brings wealth and voice: the House of Commons sends tax protests to the king and impeaches ministers it dislikes.
1569-1795 Polish-Lithuanian Commonwealth: free politics for the szlachta nobility (a tenth of people or less), little for others. Earlier Italian republics such as Venice and Genoa were aristocratic; Venice's doges were elected for life by a process that kept noble families in power.
1642-1651 English Civil War abolishes the monarchy; it returns subordinate to Parliament, and a later Bill of Rights makes royal rule depend on the people's consent.
Pre-1776 England, far from a full democracy, is the most democratic nation in history and among the most literate; voting is open in theory to all free adult male householders, with property limits in practice but not by birth status.
Early 1800s US voting mostly limited to propertied white men, an estimated 10-25% of the population; women, enslaved African Americans and Native Americans excluded.
1848 European liberal revolutions mostly fail; many of Tsar Alexander II's reforms are undone by successors.
1900 Just 3% of humanity lives in democracies.
1922 19%, then fascism, aided by radio, puts hundreds of millions under totalitarian rule; radio also carries Churchill's Blitz speeches.
Post-1947 Independence of India and other South Asian colonies lifts the share to just over one in three, steady through the Cold War.
1999 After the Soviet breakup, ~54%; Western goods from Beatles LPs to colour TVs had stirred discontent behind the Iron Curtain.
1990 → 2022 People in closed autocracies fall from ~35% to ~26%, freeing 750 million+, including the social-media-driven Arab Spring.
Kurzweil's key institutional point is the British idea that birth status should not decide the vote. Once accepted, that principle was irrepressible, which he says is why the first true modern democracy arose in Britain's American colonies. Between the fall of the Roman Republic (first century BC) and the American Revolution, elective or republican systems existed but always had narrow participation and lapsed back into tyranny. America's own exclusions carried the seeds of reform: the vote provided a mechanism for change, and the restrictions clashed with its founding ideals.
The challenge now is moving grey-zone regimes between autocracy and democracy to full democracy, and using AI for openness and transparency rather than authoritarian surveillance or disinformation. Kurzweil counts democracy as already real for nearly half of humanity, and expects the next two decades of exponential progress to extend it.
We are now entering the steep part of the exponential
The essential claim: all progress so far came from the slow early stage of the curves, and information technology will advance more in the next twenty years than in the last two hundred. The base trend is price-performance of computation, computations per second per inflation-adjusted dollar, and Kurzweil marks it at the release of each of his own books.
| Year | Machine | Computations/s per 2023 $ | Note |
|---|---|---|---|
| 1939 | Zuse Z2 | ~0.0000065 | First working programmable computer |
| 1941 | Zuse Z3 | ~0.0000091 | |
| 1943 | Colossus Mark 1 | ~0.00015 | |
| 1946 | ENIAC | ~0.00043 | |
| 1953 | UNIVAC 1103 | ~0.0048 | |
| 1959 | DEC PDP-1 | ~0.081 | |
| 1965 | DEC PDP-8 | ~1.8 | |
| 1975 | Altair 8800 | ~144 | |
| 1984 | Apple Macintosh | ~221 | |
| 1990 | MT 486DX | ~1,700 | The Age of Intelligent Machines |
| 1994 | Pentium (75 MHz) | ~19,000 | |
| 1999 | Pentium III (450 MHz) | ~800,000 | The Age of Spiritual Machines |
| 2005 | Pentium 4 662 (3.6 GHz) | ~12,000,000 | The Singularity Is Near |
| 2006 | Core 2 Duo E6300 | ~54,000,000 | |
| 2008 | GTX 285 | ~1,400,000,000 | GPUs take over |
| 2012 | GTX 680 | ~5,000,000,000 | |
| 2017 | AMD Radeon RX 580 | ~22,000,000,000 | |
| 2021 | Google Cloud TPU v4-4096 | ~48,000,000,000 | |
| 2023 | Google Cloud TPU v5e | ~130,000,000,000 | This book (early 2024) |
- Rate. A 2 × 10¹⁶ gain over 84 years works out (on these figures) to about 54 doublings, one roughly every 18-19 months.
- Rental multiplies it. A cloud TPU pod (quintillions of operations per second) can be rented by anyone for a few thousand dollars an hour, so small retail users get effective price-performance several orders of magnitude above the purchase figure.
- Paradigm independence. Because cheap computation feeds innovation, the trend has continued through relays, vacuum tubes, transistors and integrated circuits, and does not depend on shrinking transistors or rising clock speeds.
Christine Lagarde, IMF managing director, asked on stage why so little growth shows up from remarkable digital technology. Kurzweil's answer: we cancel the gain by putting it in both numerator and denominator. A teenager in Africa buys a $50 smartphone; it counts as $50 of activity. Its Snapdragon 810 averages about 3 GFLOPS, about 60 million computations per second per dollar, against 1.8 in 1965 and about 220 in 1985, so equal computation would have cost about $1.7 billion (2023 dollars) in 1965 and $13.6 million in the mid-1980s, and the phone's full capability could not have been bought at any price.
The deflation that statistics ignore runs at around 50% a year for computation, gene sequencing and many other information technologies, split between lower prices and better performance. Kurzweil's personal comparison:
| MIT's IBM 7094 (bought 1963) | iPhone 14 Pro (2022) | |
|---|---|---|
| Memory / speed | 150,000 bytes core; 0.25 MIPS | up to 17 trillion AI operations/s |
| Price | $3.1 million (about $30 million in 2023 $) | $999 |
| Users | Shared by thousands of students and staff | One person's pocket |
| Ratio | ~68 million× faster, <1/30,000 the cost, ~2 trillion× price-performance |
He admits the comparison is rough (the phone's price includes cameras, and its peak AI figure overstates typical speed), but the direction is clear. The 7094 was a transistor machine bought two years before Gordon Moore's 1965 paper, evidence that the trend predates and exceeds Moore's law, which is only one paradigm among several.
Free information compounds the mismeasurement. Teenage Kurzweil saved paper-route earnings for years to buy the Encyclopaedia Britannica for several thousand dollars, all counted as GDP. Today's teenager gets Wikipedia free, which counts for nothing. Wikipedia is less consistently edited but wins on comprehensiveness (English edition about 100× Britannica's size), timeliness (minutes versus years), multimedia and hypertext links, and often points to Britannica-quality sources. It is one of thousands of free products that add zero to GDP.
"You can't eat, wear or live in information technology"
Lagarde's rejoinder was that digital technology is remarkable but you cannot eat it, wear it or live in it. Kurzweil's reply: this changes within a decade, because goods are about to become information technologies themselves, not just be made more cheaply by them. When automation and AI dominate production and resource costs fall, those goods inherit IT's high deflation rates.
- Wear and use. In the late 2020s 3D printers make clothing and common goods, ultimately for pennies per pound. Printing keeps miniaturising; extrusion (ink-jet-like) gives way to finer methods and, probably in the 2030s, to atomically precise nanomanufacturing. Eric Drexler (Radical Abundance, 2013) estimates about 20 cents per kilogram for most objects, using mass-efficient nanomaterials; Kurzweil calls the figure speculative but the drop certain. As with music, books and film, free open-source designs will coexist with proprietary ones, a pairing he expects to define more and more of the economy.
- Eat. AI-controlled vertical farms with chemical-free harvesting, and cultured meat replacing factory farming. In 2020 humans slaughtered over 74 billion land animals for about 371 million tons of meat, which the UN puts at over 11% of human greenhouse emissions. Carcass meat causes animal suffering, is often unhealthy and pollutes; cultured meat could avoid all three and be designed healthier and tastier. In 2023 it could mimic ground beef but not a filet mignon; once it convincingly matches every animal product, Kurzweil expects discomfort to fade fast.
- Live in. Cheap printed building modules make housing affordable for millions; all these technologies are demonstrated and will go mainstream this decade.
- The metaverse. The late-2020s merger of VR and AR into a persistent layer of reality (popularised by Facebook's 2021 rebrand as Meta, though Meta did not invent the idea) will make many products purely virtual: full virtual meetings, a concert with symphony-hall sound, a family beach holiday with the smell of the sea. Current smell and touch VR is clunky; over two decades brain-computer interfaces will feed senses directly into the brain, raising questions such as whether to climb a real Everest when a safe virtual one offers the same challenge, or whether the danger was the point.
- Land. Lagarde's last objection was that land cannot become IT and the world is crowded. Kurzweil: we are crowded by choice, in cities built for working and playing together. Only 1% of habitable land is built up; about half is used by humans, almost all for farming, and of farmland 77% goes to livestock, grazing and feed and 23% to crops for people. Cultured meat and vertical farms free most of it, autonomous vehicles make long commutes practical, and virtual spaces let people live anywhere. At the pandemic peak up to 42% of Americans worked from home, breaking the inertia of the nine-to-five office.
Renewable energy is approaching complete replacement of fossil fuels
Energy powers everything else, so Kurzweil ranks this as one of the decade's most important transitions. Solar PV is already cheaper than fossil fuels in many settings, but further gains need materials science, where AI-assisted nanotechnology will let cells capture more of the spectrum.
Nanotech routes to better cells
- Nanotubes and nanowires inside cells improve photon absorption, electron transport and current generation.
- Nanocrystals, including quantum dots, raise the electricity generated per photon absorbed.
- Black silicon has a surface of countless needles smaller than a wavelength of light, nearly eliminating reflection.
- Princeton researchers trap photons with a gold nanomesh only 30 nanometres thick.
- MIT has built PV cells from graphene, carbon one atom thick.
- SolarWindow Technologies makes transparent PV films that coat windows without blocking the view.
- 3D-printed cells in rolls, films and coatings will allow decentralised production where needed and cut installation costs compared with today's rigid panels.
The doubling arithmetic
- 1.4% → 12.85%: non-hydro renewables (solar, wind, geothermal, tidal, biomass) as a share of world electricity, 2000-2021, a doubling every ~6.5 years; in absolute terms 218 → 3,657 TWh, doubling every ~5.2 years. At that rate renewables could in theory cover all electricity by 2041.
- 3.6%: solar's share in 2021, doubling just under every 28 months since 1983 while total generation rose ~220%; about 4.8 doublings reach 100% by 2032. He concedes economics and politics mean it will not literally reach 100%.
- ~20% vs ~86%: typical solar cell efficiency against the theoretical limit; wind reaches ~50% of a 59% ceiling (the Betz limit), so solar has far more headroom. Solar costs fell about twice as fast as wind's over the past five years.
- 173,000 TW: sunlight striking Earth continuously; about 1 part in 10,000 would meet all current energy needs. A 2006 US government estimate put solar available with then-best technology at up to 7,500 TW, about 65.7 million TWh a year, against 165,320 TWh of world primary energy (electricity, heat and all fuel) in 2021.
- ~80%: fall in lithium-ion storage cost per MWh, 2012-2020, with further declines projected.
Kurzweil's reason for singling out solar: renewables are not getting cheaper at one rate. Wind is closest to solar in cost decline but is already near its physical ceiling, while solar's cost floor keeps dropping because materials advances translate directly into cheaper, more efficient panels. Deserts are conveniently both the emptiest land and the sunniest: proposals to cover a small fraction of the Sahara could power all of Europe (by cable under the Mediterranean) and Africa.
Storage, the bottleneck
Fossil fuels can be stored and burned on demand; sunshine varies by hour and season, so solar needs storage over hours to months. Kurzweil is careful here: storage gains are not a true LOAR trend, since storage does not feed its own improvement. They are being pulled along by renewables (which benefit indirectly from the LOAR through IT-driven materials science), as investment flows to the piece renewables most need to win the grid, helped by converging advances in materials, robotic manufacturing, shipping and transmission. His conclusion: solar will dominate sometime during the 2030s.
Electricity cannot be stored directly, so it must be converted:
| Storage route | Converts electricity into |
|---|---|
| Molten salt | Heat |
| Pumped hydro | Gravitational potential (water lifted to a high reservoir) |
| Flywheel | Rotational energy |
| Hydrogen | Chemical energy, burned cleanly on demand |
| Lithium-ion and other chemistries | Electrochemical energy; most batteries unsuited to utility scale, but advanced ones improving fast |
Comparing storage projects is hard because they serve different points in the grid and different business models. Kurzweil uses Lazard's levelised cost of storage (LCOS): all costs, capital included, divided by total energy expected to be discharged over the project's life. His chart shows the best LCOS among new US utility-scale projects each year, since the best figure in one year tends to become the average a few years later.
Clean water for everyone
In 1990 about 24% of people lacked regular access to fairly safe drinking water; now about 1 in 10 do. The remaining gap still kills: the Institute for Health Metrics and Evaluation counted about 1.5 million deaths from diarrhoeal disease in 2019, including 500,000 young children, mostly from faecally contaminated water (cholera, dysentery, typhoid). Networks of wells, pumps, aqueducts and pipes are too expensive for many poor countries and impractical amid civil war, so centralised systems will not reach that last tenth.
The answer is decentralised technology, which Kurzweil sees defining the 2020s across energy (solar cells), food (vertical farms) and things (3D printing). For water:
| Method | Strength | Limitation |
|---|---|---|
| Village-scale plant (Janicki Omni Processor) | Purifies water for a whole village | Building-sized |
| Personal filter (LifeStraw) | Portable, individual | Ordinary filters miss the smallest viruses and some pollutant molecules |
| Boiling (solar or fuel heat) | Kills bacteria | Leaves chemical toxins; easily recontaminated |
| Antibacterial chemicals | Prevent recontamination | Do not remove other toxins |
| Ozone (made from air by electricity) | Kills pathogens very efficiently | No protection against chemicals |
| Ultraviolet light | Kills bacteria and viruses | No protection against chemicals |
| Nanoengineered filters | Block ever smaller toxins; fast and cheap in coming years | Still emerging |
The most promising device, in Kurzweil's view, is the Slingshot, invented by Dean Kamen (born 1951). About the size of a small refrigerator, it makes water pure enough to meet injectable-liquid standards from any source, sewage and contaminated swamp water included, using under 1 kW. It works by vapour compression distillation (boiling the input into steam and leaving contaminants behind), needs no filters, and is designed to run from a Stirling engine, which turns any heat source, even burning cow dung, into electricity.
Vertical agriculture
Ten thousand years of yield gains
Agriculture began about 12,000 years ago by most estimates, perhaps as early as 23,000. Early farmers scattered seed and relied on rain, so most people had to farm to survive. Irrigation arrived by about 6,000 BC; then plant breeding (bigger, more nutritious edible parts), fertilisers and efficient planting layouts. Each surplus freed people for trade, science and philosophy, and some of that specialisation fed back into farming innovation: a loop that made civilisation possible.
- Crop density. US corn yields rose from 24.3 bushels an acre (1866) to 176.7 (2021), over 7× better land use.
- World land. Growing a given quantity of crops now needs under 30% of the land it took in 1961, a roughly exponential gain that averted the overpopulation famines feared in Kurzweil's youth.
- Labour. One farmworker now feeds about 70 people; farm work fell from 80% of US labour (1810) to 40% (1900) to under 1.4% today.
Stacking the farm
Outdoor yields are nearing their theoretical limit, so the next step is growing in stacked layers:
- Hydroponics. Plants grow indoors in trays of nutrient-rich water stacked many storeys high; excess water trickles to the level below instead of running off.
- Aeroponics. Some farms replace the water with a fine mist.
- LEDs. Tuned lighting gives each plant exactly the light it needs instead of sunlight.
- Gotham Greens. An industry leader with ten large facilities from California to Rhode Island and $440 million in venture funding by early 2023; it claims 95% less water and 97% less land than a dirt farm for the same yield.
The further benefits Kurzweil lists: no agricultural runoff polluting waterways; no loose soil blown into the air; no pesticides, since pests cannot get in; year-round crops, including species the local climate cannot grow; no losses to frost or bad weather; and above all local food instead of produce trucked or railed hundreds or thousands of miles. With solar power, on-site fertiliser production, water harvested from air and robotic harvesting, few workers and little land, future vertical farms could make food so cheap it becomes nearly free, the way Google and Facebook offer free services funded by advertising once computation became cheap. Automating every part of a farm with AI is, for Kurzweil, turning food production into an information technology.
3D printing will revolutionise the making and distribution of things
Twentieth-century manufacturing either moulded material (injection-moulded plastic, pressed metal) or removed it from a block like a sculptor. Moulds are expensive and hard to change; subtractive methods waste material and cannot make some shapes. From the 1980s a new family, additive manufacturing or 3D printing, built objects from stacked flat layers. The common type works like an ink-jet printer: a nozzle deposits softened plastic (or other material) along a software path, layer after layer, the layers fusing as they harden. Over two decades resolution, speed and cost have all improved, and printers now handle paper, plastic, ceramic and metal.
| Advantage or risk | Kurzweil's example |
|---|---|
| Rapid prototyping | A part designed on screen is in hand within minutes or hours, not weeks, so people with ideas but little money can iterate and reach market |
| Customisation | A mould change can cost tens of thousands of dollars; a design change costs nothing. FitMyFoot prints shoes from phone photos of the feet; furniture fitted to the body; tools fitted to the hand |
| New materials | Medical implants with built-in drugs released gradually; graphene bulletproof clothing and superfast electronics; AI-optimised shapes impossible to make otherwise |
| Decentralisation | Local production instead of distant factories; freight is about 30% of fuel-combustion carbon emissions (International Transport Forum) |
| Clothing | Common fabric fibres are 10-22 microns across; some printers already reach 1 micron or finer, so printed clothing awaits fabric-like materials at fabric-like prices |
| Bioprinting | An inert polymer or ceramic scaffold is coated with fluid rich in reprogrammed stem cells that fill the shape, giving an organ with the patient's own DNA. United Therapeutics (Kurzweil sits on its board) aims at lungs, kidneys and hearts, avoiding donor shortages and rejection |
| Piracy | Why pay $200 for designer shoes if you can download and print them? The same intellectual-property problem music, books and film already face |
| Printed guns | Files for gun parts already circulate; printed guns lack serial numbers and, in advanced plastics, could pass metal detectors, so regulation must be rethought |
3D printing of buildings
Printing building modules and the smaller objects inside buildings will cut construction costs for homes and offices. Kurzweil describes the approaches:
- Parts assembled on site. Wall sections and roof segments printed and snapped together like IKEA furniture or Lego; by the late 2020s robots will do most of the assembly.
- Room modules. Printed rooms with square or rectangular footprints, lifted by crane and combined in many layouts, which minimises disruption to neighbours. In 2014 China's WinSun built ten simple modular houses in 24 hours for under $5,000 each; China is already a hub and will need more.
- Whole building in one piece. A large frame surrounds the site and a robotic nozzle lays concrete walls layer by layer; workers then finish interiors, windows and roof tiles. HuaShang Tengda announced a two-storey villa printed in one piece in 45 days (2016); Alquist 3D completed the first owner-occupied printed US home (2021), and Houston saw the first multistorey printed home (2023).
The advantages: lower labour costs and so cheaper basic housing; shorter build times and with them less waste, light and noise pollution, toxic dust, traffic disruption and worker hazard; and use of local materials instead of timber and steel hauled from afar. For skyscrapers, printers and autonomous robots using material pumped up from the ground in liquid form would solve the problem of moving people and materials to upper floors. By the late 2020s, combined large- and small-object printing with robotics will make buildings both more personalised and much cheaper.
Diligent people will reach longevity escape velocity by around 2030
Abundance and peaceful democracy improve life, but the highest-stakes project is preserving life itself. Kurzweil's case rests on medicine shifting from a linear hit-or-miss process to an exponential information technology that reprograms biology's suboptimal software.
Why biology can be improved on
Evolution is suboptimal: it is random variation filtered by selection, heavily shaped by chance and local conditions, and gradual, so it can only reach designs whose every intermediate step was itself fit. Some useful traits are therefore unreachable by evolution. Intelligence, human or artificial, can search the whole space of genetic possibilities, including those inaccessible traits.
The tools are already on the curve. Genome sequencing has improved for about two decades since 2003, roughly doubling price-performance each year; measured in base pairs, the doubling has averaged about every 14 months across many technologies since the first DNA nucleotide sequencing in 1971. We are reaching the steep part of a fifty-year exponential.
From human trials to in silico trials
By the end of the 2020s biological simulators will produce key safety and efficacy data in hours rather than the years trials take. Two forces will govern the switch:
Against speed. A legitimate safety worry: a simulation could miss a relevant medical fact and wrongly clear a dangerous drug. Kurzweil also expects political uncertainty and bureaucratic resistance.
For speed. Simulations can use vastly more virtual patients, cover many comorbidities and demographic groups, and tell doctors in fine detail how a treatment will affect each kind of patient; faster approval also saves lives. He expects effectiveness to win.
Two examples of what simulation will unlock:
- Immunotherapy. It already sends many stage 4 (otherwise terminal) cancer patients into remission. CAR-T therapy reprograms a patient's own immune cells to find and kill cancer cells. Progress is limited by incomplete understanding of how cancer evades the immune system, a logjam AI simulation can break.
- Induced pluripotent stem (iPS) cells. Adult cells turned into stem cells by introducing specific genes can become almost any cell type. By 2023 they had regenerated tracheas, craniofacial bones, retinal cells, peripheral nerves and skin, plus tissue from the heart, liver and kidneys. For heart patients, iPS cells from the patient are grown into sheets of heart muscle grafted onto the damaged heart; they appear to work by releasing growth factors, in effect tricking the heart into a fetal regenerative state. This could reverse the low ejection fraction that afflicts many heart-attack survivors and killed Kurzweil's father. The open risk is that stem cells resemble cancer cells, so uncontrolled division must be minimised. AI analysis of how iPS cells act will let regenerative medicine use the body's own healing blueprints.
Both intuition and a backward-looking view of history assume the next twenty years of medicine will resemble the last twenty; Kurzweil says that ignores the exponential, and that most doctors and patients have not yet registered the change.
Nanorobots and the immune system
In the 2030s, the third bridge (named in his health book with Terry Grossman, MD) arrives as medical nanorobots that vastly extend the immune system. Our T cells work well, but immunity evolved when people reproduced young and died in their twenties, so evolution never favoured defences against late-life threats such as cancer and neurodegenerative diseases often caused by misfolded proteins called prions. Ancestors who lived before animal domestication also never evolved strong defences against the many viruses that come from livestock.
- Pathogens. Nanorobots will be programmed to destroy all types of pathogen.
- Metabolic disease. Apart from the heart and brain, major organs mostly add substances to or remove them from the blood, and many diseases are organ failures: type 1 diabetes is the failure of pancreatic islet cells to make insulin. Nanorobots monitoring blood levels of hormones, nutrients, oxygen, carbon dioxide and toxins could augment or replace organ function.
- Outcome. By the end of the 2030s we will largely overcome disease and ageing.
By around 2030, Kurzweil predicts, the most diligent and informed people will reach longevity escape velocity: each calendar year of research adds more than a year to their remaining life expectancy. AI-driven drugs and nutrition in the 2020s will not cure ageing but will carry them to the nanorobots of the 2030s; in his phrase, the sands of time start running in rather than out.
The fourth bridge and its dilemmas
As we augment the biological neocortex with faster cloud models of it, thinking becomes a hybrid; the digital part grows exponentially, eventually predominates, and becomes able to model the biological part completely, so all our thinking can be backed up. Kurzweil places this near the Singularity in the mid-2040s. The goal is to live as long as we wish.
- Why would anyone choose death? Research shows suicide usually stems from unbearable physical or emotional pain, which medicine and neuroscience should make much rarer, if not eliminate.
- Could you even die? The cloud already keeps many backups, and by the 2040s destroying every copy of oneself may be nearly impossible.
- The deletion dilemma. Easy self-deletion maximises autonomy but opens the door to coercion, trickery and cyberattack; restricting it limits control over one's most intimate data. Kurzweil expects safeguards to work, as they have for nuclear weapons for decades.
- Is the restored copy you? That is a philosophical, not scientific, question (chapter 3), which most people alive today will have to face.
- Only for the rich? Kurzweil points to the mobile phone: thirty years ago only the wealthy had one, and it barely worked; now billions carry phones that are memory extenders with access to almost all human knowledge. Information technologies start expensive and limited and become affordable by the time they are perfected, because of exponential price-performance.
The rising tide
Kurzweil's conclusion: life is getting better profoundly and for the great majority, and not by coincidence. Literacy, sanitation, life expectancy, clean energy, poverty, violence and democracy all ride one dynamic: information technology facilitates its own advancement, the core of the LOAR and the engine of the virtuous circles. Information technology is about ideas, and improving our ability to share and create them gives each person more power to fulfil their potential and humanity more power to solve its problems.
- One tide, all boats. Exponentially improving IT is a rising tide lifting every aspect of the human condition, and it is about to surge.
- AI as converter. AI is turning linear fields (agriculture, medicine, manufacturing, land use) into exponential information technologies.
- Beyond imagination. We already struggle to picture life a century ago; the coming decades will carry us far beyond what we can now imagine.
Chart choices. Many curves (radio, TV, toilets, electricity, literacy, schooling) are S-curves that saturate, not exponentials; "exponentially better" really applies to underlying price-performance. The solar (2032) and renewables (2041) dates extrapolate early-stage doubling times, which S-curves always slow. Democracy is the weakest series: his own chart shows Our World in Data's share falling to about 29% after 2017 (probably reflecting India's reclassification as an electoral autocracy), yet the text says "nearly half". Causation from information technology to democracy and peace is asserted through plausible loops, not demonstrated; radio served fascism as readily as Churchill. Gotham Greens grows in rooftop greenhouses, not stacked vertical farms. In the Snapdragon comparison the $13.6 million figure uses 1985 price-performance but is dated 1980; the solar-energy figure is printed as "65,700,00" TWh (should be 65.7 million). The 11% livestock emissions figure is lower than the FAO's usual 14.5%. Debate. Pinker's violence thesis is contested on war deaths, which are fat-tailed (Taleb, Cirillo); "broken windows" and lead-crime effects remain disputed in criminology. The 2020s forecasts (printed clothing, near-free food, solar dominance, in silico trials) rest on cost curves continuing, not on demonstrated products.
2026 scorecard
| Prediction or trend (book) | Book's figure / date | Status, Oct 2026 | Evidence |
|---|---|---|---|
| Extreme poverty falling ~4%/yr | 697 m in 2019 | Behind | ~837 m below new $3/day line (2026); progress stalled after COVID |
| Literacy rising | ~87% (2020) | On track | 88% of adults (2024); 739 m still illiterate |
| Life expectancy rising | 80+ in rich nations | On track | US record ~79 (2024) after COVID dip |
| US homicide down long-term | 6.8 per 100,000 (2021) | Ahead | 4.1 in 2025, tying 1955-56 lows |
| Solar doubling every 28 months, 100% by 2032 | 3.6% (2021) | Behind pace | ~8.8% of global power (2025); doubling nearer 3 years |
| Renewables doubling every 6.5 years, all power by 2041 | 12.85% (2021) | On track | Wind+solar 17.3% (2025), ~20% with other non-hydro |
| Storage costs keep falling | -80% (2012-20) | Ahead | Stationary packs $70/kWh (2025, -45% in a year) |
| Genome cost keeps halving | $399 (2023); $100 promised | Roughly on track | ~$80 genome claimed (Ultima, 2025) |
| In silico trials begin replacing human trials | End of 2020s | Early steps | FDA roadmap (Apr 2025) to cut animal testing via AI models |
| Democracy spreads | ~half of humanity | Behind | V-Dem 2026: 74% live in autocracies |
| Vertical farming makes food near-free | 2020s | Behind | Wave of vertical-farm failures 2023-25 |
| Longevity escape velocity | ~2030 | Behind | No measurable effect on remaining life expectancy |
The World Bank raised its extreme-poverty line to $3.00 a day (2021 prices) in June 2025; on that line about 837 million people were extremely poor in 2026, increasingly concentrated in sub-Saharan Africa, and the Bank says ending it could take decades at the current pace. The near-4% annual decline Kurzweil assumed has not held since COVID. UNESCO puts adult literacy at 88% in 2024 (youth 93%), with 739 million adults still unable to read. CDC data released in January 2026 show US life expectancy at a record of about 79 years in 2024, recovering from under 76.5 in 2021. The FBI reported in August 2026 that violent crime fell 9.7% in 2025 to 327.6 per 100,000, the largest drop since national estimates began in 1936, and murder fell about 18% to 4.1 per 100,000, matching the lowest rates on record (1955-56). The ILO-UNICEF count of working children fell to about 138 million (54 million in hazardous work) in 2024. Verdict: health, literacy and safety on track or ahead; poverty reduction behind.
Ember's Global Electricity Review 2026 reports solar generation up 30% (+636 TWh) to 2,778 TWh in 2025, about 8.8% of world electricity; wind and solar together reached 17.3%, renewables 33.8%, and fossil generation fell slightly. BloombergNEF (December 2025) put average lithium-ion pack prices at a record $108/kWh, and stationary-storage packs at $70/kWh, down 45% in a year. Module prices fell to roughly 10 cents a watt in 2024-25, a third of the book's 2021 endpoint. Against this, V-Dem's 2026 report finds 74% of humanity under autocracy and democracy for the average person back at 1978 levels. Vertical farming faltered: AeroFarms filed for bankruptcy (2023), Bowery shut down (2024) and Plenty filed (2025). Cultured meat won US approval in 2023 but Florida and Alabama banned it in 2024. Verdict: the energy cost curves are ahead of Kurzweil's case; democracy and food are behind it.
Sequencing kept falling: Ultima Genomics launched its UG 100 Solaris in February 2025, cutting run costs by about a further 20% and claiming an $80 human genome, below the $100 the book anticipated. On the regulatory side, the FDA announced in April 2025 a roadmap to phase out animal-testing requirements, starting with monoclonal antibodies, in favour of "new approach methodologies" including AI-based computational models and human-cell systems: an early step toward the in silico testing Kurzweil predicts, though human clinical trials remain mandatory. No therapy has yet shown that it extends remaining life expectancy by more than a year per year. Verdict: the information-technology inputs are on curve; the clinical payoff Kurzweil dates to 2030 is not yet visible.
The Future of Jobs: Good or Bad?
Kurzweil concedes that AI will destroy jobs faster and across more occupations than any earlier wave of automation, starting with the 4.6 million Americans who drive for a living. His optimism rests on three legs: technology has replaced most jobs several times since 1800 while employment, pay and leisure all rose; official statistics miss most of the value digital technology creates; and AI-driven abundance will let a steadily growing safety net buy a comfortable life. He predicts a universal basic income or its equivalent in developed countries by the early 2030s and in most countries by the late 2030s, with the real struggle shifting from survival to meaning.
The chapter runs in five movements, mirroring the source: the present wave of automation (driving as the test case), the two-century record of destruction and creation, the economists' case that this time is different (with its productivity and participation puzzles), the long-run destination of work, and the political transition, argued out through a private dialogue with Daniel Kahneman.
The current revolution
Kurzweil frames the chapter with a paradox: the same convergent technologies that will create enormous material abundance over the next two decades will also unsettle the economy and force society to adapt at an unprecedented pace. Self-driving cars are his chosen example because the timeline is short and the jobs are countable.
2005 DARPA pays a $2 million prize to a Stanford team for winning its autonomous-vehicle Grand Challenge, the year The Singularity Is Near appeared; many experts still thought driverless cars a century away.
2009 Google starts an AI-driven self-driving project; progress accelerates sharply.
2018 The project, spun out as Waymo, begins simulating as many miles per day as it had driven on real roads since 2009.
2020 Waymo offers fully driverless ride-hailing to the public around Phoenix, followed by San Francisco.
2021 Ratio still holds: about 20 million simulated miles a day against just over 20 million real miles in total; cumulative simulation passes 20 billion miles.
2024 Kurzweil expects service to have reached Los Angeles and possibly other cities by the time readers see the book.
Why real and simulated driving need each other
- Real roads supply surprise. Physical driving throws up situations engineers would never think to program into a simulator, so it keeps the virtual world honest and calibrated.
- Simulation supplies volume and control. Engineers cannot stop traffic and ask everyone to rerun a scene five miles an hour faster; in software they can vary exactly the parameter that matters, millions of times.
- Simulation reaches the unsafe. Dangerous scenarios that could never ethically be staged for training can be run freely, and the results tell engineers which issues to prioritise on real streets.
- Same recipe as Go. As chapter 2 explained, simulation generates enough examples to train very deep (around 100-layer) networks; DeepMind used self-generated games to surpass the best human Go players, and Waymo applies the same strategy to a far messier world.
Who drives, and who depends on drivers
The driving workforce in numbers:
- 2.7%+ of employed Americans drive trucks, buses, taxis, delivery vans or other vehicles: over 4.6 million jobs.
- Mostly men, mostly middle-aged, mostly without college degrees, the group least able to retrain cheaply.
- Under 3% of workers in California and Florida; over 4% in Wyoming and Idaho; 5%, 7% or even 8% in parts of Texas, New Jersey and New York.
- 10.2% of US workers are employed directly in transportation and related industries (Bureau of Transportation Statistics, 2021). Kurzweil considers it virtually certain that many drivers will lose their jobs before normal retirement, even if the pace is debatable. The damage then spreads outward: fewer truckers means less payroll work, fewer customers for roadside convenience stores and motels, less need for truck-stop cleaners, and lower demand for the sex workers who serve trucking routes. Those second-order effects are certain in direction but hard to size or time; a 10% sector does not need to shrink much for the consequences to be large.
The wider list of threatened occupations
Driving is one entry on a long list of jobs exposed to AI trained on massive datasets.
| Study | Method | Headline finding |
|---|---|---|
| Frey & Osborne, Oxford, 2013 | Ranked ~700 occupations by likelihood of automation by the early 2030s | Telemarketers, insurance underwriters, tax preparers at 99%; over half of occupations above 50%; high risk for factory, customer-service, banking and driving jobs; low risk for jobs needing close, flexible contact (occupational therapists, social workers, sex workers) |
| OECD, 2018 | Scored each task within each job, across 32 countries | 14% of jobs had a 70%+ chance of elimination within a decade; another 32% were above 50%; about 210 million jobs at risk |
| OECD, 2021 | Follow-up on actual employment data | Employment grew much more slowly in high-risk occupations |
| McKinsey, 2023 (post-ChatGPT) | Share of working time automatable with existing technology | 63% of working time in developed economies; under fast adoption half could be automated by 2030; midpoint scenario 2045, assuming no further breakthroughs |
Kurzweil's gloss on McKinsey is the chapter's sharpest forecast: the "no breakthroughs" assumption is wrong, because AI will keep improving exponentially until superhuman AI and AI-controlled, atomically precise manufacturing arrive sometime in the 2030s. All the pre-2022 studies, he adds, predate generative systems like ChatGPT and Bard.
Nottingham stocking and lace weavers lived modestly on skills passed down in stable family businesses until the power loom and other textile machines moved economic power to machine owners. Legend credited an elusive Ned Ludd (possibly fictional, never found) with breaking machinery, and when the weavers formed an urban guerrilla army in 1811 they named "General Ludd" their leader. Attacks on machines turned to bloodshed; the British government imprisoned and hanged the leaders. With no retraining, many weavers sank into lower-paid work for a time, and the new higher-paid jobs designing and selling machines were no comfort to them. Yet ordinary people could now afford a wardrobe rather than a single shirt, whole new industries formed, and Kurzweil judges that prosperity, more than repression, ended the movement. The Luddites survive as the emblem of everyone left behind by progress.
Destruction and creation
Kurzweil's thought experiment: a futurist in 1900 tells Americans that farming (about 40% of workers, down from over 80% in 1810) will fall by more than 95% to under 1.4% by 2023, and manufacturing (a fifth of workers) by more than half to 7.8%, but that total employment will rise. Asked "what new jobs?", the only honest reply is that they have not been invented yet, in industries that do not yet exist. A futurist in 1990 promising jobs building websites and mobile apps, in data analytics and online merchandising, would have been equally unintelligible. The unsatisfying vagueness of that answer, not its falsity, is why automation always breeds political anxiety.
- Workforce. About 29 million (38% of the population) in 1900; about 166 million (over 49%) in early 2023.
- Hours. Just over 2,900 a year per worker in 1870; about 1,765 in 2019, just before COVID.
- Earnings, 1929. Per capita personal income about $700; with 48 million employed out of 122.8 million people, about $1,790 per worker, roughly $31,400 in 2023 dollars.
- Earnings, 2022. Per capita income $64,100 across 332 million people and a labour force of about 164 million: around $129,800 per worker (about $133,000 in 2023 dollars), over four times the 1929 level.
- Median caveat. Median income in 2021 was only $37,522 against the $64,100 average, reflecting very high earnings at the top plus retirees, students and others without jobs (no reliable 1929 median exists).
- Hourly. 2,316 hours a year in 1929 gives about $13.55 an hour; in 2021, $10.8 trillion of wages over 254 billion hours gives about $42.50, more than triple.
Kurzweil argues the true rise is larger. Wage statistics omit freelance programmers, entrepreneurs and artists, who can earn a lot per hour. Total 2021 personal income of about $21.8 trillion implies hourly income roughly double the wage figure, but since much of that (rent, for instance) is not tied to hours, the honest number lies between the two. Neither series counts the quality gains and new products documented in chapter 4. Underneath, technology keeps adding information-based dimensions to old jobs and creating millions of new ones needing higher skills, which has so far offset the collapse of farm and factory work.
Farming: from 80% of workers to about 1%
- Rise. As settlers moved west of the Appalachians, the farm share of employment actually climbed to a peak above 80%.
- Decline from the 1820s. First scientific plant breeding and better crop rotation, then mechanical implements of the Industrial Revolution, let fewer farmers feed more people.
- 1890. The first year most Americans did not farm; the fall sped up by 1910 as steam and internal-combustion tractors replaced work animals.
- 20th-century yields. Pesticides, chemical fertilisers and genetic modification raised UK wheat yields from 0.44 ton per acre (1850) to 3.84 tons (2022) while the population rose from about 27 to 67 million, so food became more plentiful per person, not just sufficient.
- Knock-on. Better nutrition produced taller, healthier people with better childhood brain development, unlocking more talent for further innovation.
- Next. Automated vertical farming and firms like Hands Free Hectare (UK), which aims to remove human labour from every stage, plus cheap renewables, will cut food prices until poverty no longer blocks access to fresh, healthy food.
Manufacturing: rise, plateau and decoupling
Many displaced farmhands went into factories, and about 150 years later factories went through the same story.
1800s (first decade) About 1 in 35 workers in manufacturing.
1870 Almost 1 in 5, concentrated in the industrialising North, as steam factories hired millions of low-skilled labourers.
c.1900 A second industrial wave pulls in masses of mostly immigrant workers; the assembly line cuts prices and widens markets.
c.1920 Peacetime peak of about 26.9% of the civilian labour force.
1920-1970 Roughly 1 in 4, with no trend apart from the Depression dip and the World War II bulge.
1980-2010 Share falls to 17.5%, 14.1% (1990), 12.1% (2000) and a low of 7.5% (2010).
Feb 2001 17 million factory jobs before the dot-com recession; flat near 14 million through the mid-2000s boom despite rising output.
Dec 2007 to Feb 2010 13.7 million to 11.4 million in the Great Recession; output was back near record highs by 2018, jobs were not.
Nov 2022 Only 12.9 million workers produce near-record output; by early 2023 manufacturing employs about 1 in 13 Americans.
Kurzweil names two technology-driven causes of the decline:
- Containerised shipping. Unglamorous next to robots or AI, yet among the most consequential innovations of modern times: it slashed worldwide shipping costs, made a truly global economy possible, cheapened goods for consumers, and made offshoring to low-wage countries economic, a key driver of US deindustrialisation.
- Automation. Robots cut hands-on assembly work, and from the 1990s computerisation and AI made automation steadily more capable; between 1992 and 2012 the inflation-adjusted hourly output of the average manufacturing worker doubled. The result in the twenty-first century is that output and employment decoupled.
The labour force kept growing anyway
Despite successive waves of farm and factory automation, the US labour force has grown for as long as it has been measured, absorbing a fast-growing population and tens of millions of women. In 2000 it was 143.6 million of 282 million people (50.9%); in 2022, 164 million of about 332 million (49.4%). The slight fall in share is mainly ageing: over-65s were 8.0% of Americans in 1950, 16.0% in 2018, and a Census-projected 22% by 2050 before any medical breakthroughs. If radical life extension arrives, as Kurzweil expects, the elderly share will be higher still.
Education scaled with the skill demands. University students (undergraduate and graduate) rose from 63,000 in 1870 to about 20 million in 2022, with 4.7 million added between 2000 and 2022 alone. K-12 public spending per pupil went from $1,035 (1919-20, 2023 dollars) to about $19,220 (2018-19), over eighteen times as much.
Technology has replaced most of the economy's jobs several times over since 1800, each time amid accurate warnings that whole categories of work would vanish; each time, aided by rising education, employment, real pay and leisure grew. Kurzweil's argument for the future starts from this base rate.
Is this time different?
Erik Brynjolfsson of Stanford is Kurzweil's chosen representative of economists who think AI-era automation will, unlike earlier transitions, destroy more jobs than it creates. They see the present as the culmination of successive waves.
- 1. Deskilling. Technology lowers the skill a job needs: a car driver needs less than a carriage driver who handled and maintained unpredictable horses; assembly-line machines replaced years of shoemaking craft with brief machine training. Entry gets easier and goods cheaper, but low-paid jobs replace better-paid ones.
- 2. Upskilling. Often following deskilling, technology raises the skill needed: drivers must master navigation electronics; hand-operated shoe presses give way to FitMyFoot, which 3D-prints custom footwear with a few staff skilled in computing and printers. Fewer, better-paid jobs replace many low-paid ones.
- 3. Nonskilling. AI removes the human from the task altogether, as driverless vehicles will. It is attractive for cost and also for quality: AI drivers are far safer and never drunk, drowsy or distracted. A series of such transitions will follow as AI and robots take on task after task.
Tasks are not professions
Kurzweil distinguishes automating a task from eliminating a profession. Sometimes, though not always, losing a task pushes a job toward new tasks, a form of upskilling:
- Bank tellers. ATMs took routine cash handling; tellers shifted to marketing and customer relationships.
- Paralegals. Legal-research and document-analysis software took some functions; the profession now bundles a substantially different set of tasks.
- Graphic designers. From 2022, DALL-E 2, Midjourney and Stable Diffusion produced high-quality art from text prompts; designers may sketch less and spend more time brainstorming with clients and curating or modifying AI output.
Why the economics still favour machines
In the long run Kurzweil thinks incentives push AI into an ever-wider set of tasks. Firms can usually trade off capital against labour: where wages are low, labour-intensive methods pay; where wages are high, inventing labour-saving machines pays. That logic likely made Great Britain the cradle of the Industrial Revolution, since it combined some of the world's highest wages with abundant cheap coal, rewarding steam power that replaced costly hands. Today a machine is a one-time purchase that becomes an asset, while wages are a recurring cost with extra obligations attached. As AI reaches human and then superhuman competence, fewer tasks will need unenhanced humans, and until we merge more fully with AI this means major disruption for workers.
The productivity puzzle
The net-job-loss thesis faces an awkward fact. If machines were displacing labour on net, output per hour (inflation-adjusted output divided by hours worked) should be soaring. It is not.
US real output per hour, average quarterly growth:
- 0.55% from Q1 1950 to Q1 1990.
- 0.68% from Q1 1990 to Q1 2003, as PCs and the web spread and many expected the boom to last.
- 0.36% from Q1 2003 to Q1 2022, after a slowdown starting in 2004: one of the great economic mysteries of the past decade. Kurzweil's answer, now gaining ground among economists and also advanced by Brynjolfsson and venture capitalist Marc Andreessen, is that several trillion dollars of value are "missing" because GDP counts prices, not value. MIT paid about $3.1 million (about $30 million in 2023 dollars) in 1963 for the IBM 7094 he used as an undergraduate, and GDP recorded $3.1 million. A smartphone hundreds of thousands of times more capable, with functions unavailable at any price in 1965, registers as a few hundred dollars.
- Price, not willingness to pay. A $20,000 car adds $20,000 to GDP even if the buyer would have paid $25,000-30,000; the gap is uncounted.
- Marginal cost anchored prices. When goods needed metal and skilled hours per unit, competition pushed prices toward average marginal cost, so price tracked what people would pay.
- Quality tracked cost. Better products cost more to make, so GDP captured quality through price.
- Information goods break the links. A $900 (2023 dollars) chip delivered 800,000+ computations per second per dollar in 1999 and nearly 58 billion in early 2023: 72,000× the power, counted the same in GDP.
- Near-zero marginal cost. An extra e-book copy needs no paper, ink or labour, so it sells at a nearly infinite multiple of marginal cost.
- Free services. An extra Google search costs almost nothing; Facebook pays no more to link you to 1,000 friends than to 100; so these are given away and funded by ads, and the price signal vanishes.
The consequence, Kurzweil argues, is that flat nominal wages mislead: even someone whose pay stagnated for twenty years can buy thousands of times more computing with it. Statistical agencies make some quality adjustments, but they still massively underestimate price-performance gains.
Measuring consumer surplus
Consumer surplus, what people would pay above what they do pay, can be estimated from choices. If you could earn $20 mowing a neighbour's lawn but spend that hour on TikTok, TikTok is worth at least $20 to you.
- $8 billion vs about $230 billion. Facebook's US revenue against users' time (Tim Worstall, Forbes, 2015).
- $287 billion. Same method for 2020: 35 minutes a day on Facebook by the 72% of roughly 258 million US adults who use social media.
- $36.1 billion vs $1 trillion+ a year. US social-media ad revenue against 2 hours 3 minutes a day across all platforms (global survey, 2019).
- About zero. Wikipedia's official contribution to GDP.
Kurzweil admits the method is crude: browsing in a coffee queue does not displace paid freelance work. Still, it shows people value these services enormously while economists see only a sliver as revenue. Since consumer surplus is a more genuine measure of prosperity, the productivity that matters has been growing fine. The hidden gains extend beyond "tech": less pollution, safer living conditions, more learning and entertainment.
He flags the main exception himself: health care has inflated faster than prices overall, so a heavy user of medical care gains little from cheaper GPU cycles. The remedy is that AI and convergence will turn more sectors into information technologies in the 2020s and 2030s:
- Education. AI tutors offering individually tailored learning in any subject to anyone online.
- Medicine. AI-enhanced care and drug discovery, still in infancy, eventually cutting health costs.
- Physical goods. Food, housing and construction, and clothing; AI materials science makes solar electricity extremely cheap, while robotic resource extraction and autonomous electric vehicles cut raw-material costs.
- Outcome. With cheap energy and materials and labour automated, prices fall so far that in the 2030s it will be inexpensive to live at what is now a luxurious standard.
As this deflation spreads, the gap between measured and real productivity will widen; Kurzweil expects national inflation to fall and eventually turn into deflation, making the puzzle's answer clearer over time.
The participation puzzle
Net-job-loss economists also point to the labour force participation rate: employed plus job-seekers as a share of those aged 16+. It rose from about 59% in 1950 to just under 67% in 2002, fell below 63% by 2015, and stayed flat until COVID despite a supposedly booming economy. As a share of the whole population, the labour force was 50.7% in June 2008 (154 million of 304 million) and just under 49.5% in December 2022 (164 million of 333 million), the lowest in over two decades. Kurzweil notes the series omits agricultural workers, military personnel and federal employees, but is good enough for direction and size.
Two confounders explain much of the fall:
- Longer education. Fewer teenagers work, and many stay in college and graduate school well into their twenties.
- Retiring baby boomers. The working-age share of the population shrinks.
Restricting to prime-age adults (25-54) almost erases the decline: 83.4% in early 2023 against a 2000 peak of 84.5%, a gap of about 1.7 million people.
Older Americans moved the other way between 2001 and 2021: participation rose from 60.4% to 68.2% among 55-64-year-olds and from 5.2% to 8.6% among the 75+. Kurzweil sees conflicting forces: some displaced older workers simply retire early on less; others work longer because they live longer (US life expectancy rose about two years from 2000 to pre-COVID) and find purpose in it; and some are forced into marginal jobs after losing better ones before they can afford to retire.
Participation is a fading measure
Kurzweil's deeper point is that "being in the labour force" captures less and less of how people earn. Two trends escape the statistics.
The underground economy. Always present, it is now amplified by the internet: almost the entire sex industry, under-the-table housework, alternative healing and more, plus cryptocurrencies that hide transactions from tax, regulation and police. Platforms like Craigslist and social media suffice without crypto.
Bitcoin as a case of growth and volatility:
- Daily volume. Under $19.3 million (6 August 2017), over $4.95 billion (7 December 2017), about $180 million (mid-2023).
- Scale check. Global foreign-exchange trading averaged $7.5 trillion a day in April 2022 (Bank for International Settlements).
- Price. $13.43 (4 Jan 2012), above $130 (2 Apr 2012), $1,354 (29 Apr 2017), $18,877 (17 Dec 2017), $3,360 (12 Dec 2018), $64,899 (13 Apr 2021), $15,460 (20 Nov 2022).
- Why it fails as money. If a currency may gain tenfold in six months, holders hoard it; if it may halve in months, merchants refuse it. Wider use requires stable value. New kinds of earning. Buying, selling and bartering through websites and apps, building apps, and making content: some people now have careers on YouTube or as paid influencers on Instagram and TikTok. The app economy barely existed before the iPhone (2007).
The app economy from nothing:
- iOS apps. Under 100,000 (2008) to about 4.5 million (2017).
- Google Play. About 16,000 (December 2009) to 2.6 million (March 2023), over 160-fold in thirteen years.
- US jobs. About 500,000 created 2007-2012; 5 million+ by 2018 (Deloitte); 5.9 million jobs and $1.7 trillion of activity including indirect effects (2020), depending on how broadly the market is defined. Kurzweil concludes that the forces killing jobs are also creating opportunities outside the traditional job model. The gig economy has limits, but often gives more flexibility, autonomy and leisure than people's previous options; improving its quality is one way to help workers through automation.
So where are we headed?
Kurzweil accepts that the picture looks alarming. Frey and Osborne's estimate that almost half of 2013's jobs could be automated by 2033 assumed slower AI progress than he forecasts, and while automation fears are two centuries old, the current threat is unique in its speed and breadth. He then reframes the question around what work is for.
Work is a means, not an end
- Material needs. Growing and distributing food took nearly all human labour two centuries ago and takes under 2% in the US and much of the developed world. AI-driven abundance will make the struggle for physical survival history.
- Purpose and meaning. A bricklayer gets two kinds of meaning: wages to provide for loved ones (a core part of identity) and lasting structures that serve the public good. Arts and academia add creativity and new knowledge. Technology will vastly expand people's power to make such contributions.
Case study: how technology widened creative space
When Kurzweil was growing up there were only ABC, NBC and CBS. Programmes had to please every demographic at once, so narrower ideas (absurdist comedy, paranormal drama, science fiction) struggled; Star Trek, the most influential SF series ever, was cancelled after three seasons. Cable opened room for niches (deep documentaries on Discovery, History, TLC) but viewing was still tied to time slots. DVRs and on-demand streaming let a show draw its audience from the whole population, so Stranger Things and Fleabag could thrive, and so could programmes portraying small groups such as LGBTQ people, people with disabilities and American Muslims. Streaming also changed form: guaranteed viewing order freed BoJack Horseman to build character arcs and running jokes across episodes, impossible under broadcast's self-contained half-hours. The next two decades will accelerate this. The image-making creativity of DALL-E, Midjourney and Stable Diffusion will extend to music, video and games: people will describe an idea and refine the output in plain language. An action film that once needed thousands of people and hundreds of millions of dollars will need only good ideas and a modest budget for the computer running the AI.
The skill ladder and the merger
Kurzweil does not soften the near term: automation has already removed many jobs at the bottom and middle of the skill ladder, and the trend will widen and speed up over the next decade. Society has climbed the ladder through education, but AI will pass even the most skilled humans in field after field. His answer is the next phase of self-improvement, merging with our technology.
- Already happening. Almost nobody can work or study without smartphones, brain extenders that reach nearly all human knowledge in a tap; two decades ago that was not true, so it is fair to say our devices have become part of us.
- 2020s. Search becomes seamless question answering; real-time translation between any language pair becomes smooth and accurate; augmented reality is projected onto retinas from glasses and contact lenses, then into hearing and other senses; AI assistants anticipate needs by watching and listening rather than waiting for requests.
- 2030s. Medical nanorobots begin wiring these extensions directly into the nervous system, extending the neocortex into the cloud (chapter 2).
- Access. Like phones, which started expensive and are now near-universal (5.8 billion smartphone subscriptions in 2020, per the ITU), these will become cheap and available to everyone.
The safety net and universal basic income
The US safety net began in the 1930s with Social Security. Particular labels fall in and out of favour ("welfare"), but the overall net has grown under every party and administration since.
Public social spending, 2019 (pre-COVID):
- United States. 18.7% of GDP, near the developed-country median: about $4 trillion of $21.4 trillion, or over $12,000 per head across about 330 million people.
- Canada. 18.0%.
- Australia and Switzerland. 16.7% each.
- United Kingdom. 20.6%: about $580 billion of a $2.8 trillion GDP, under $8,800 per head for 66 million people. Higher US GDP per head means a larger safety net per person than the UK's.
- Share of government. Social spending is now about 50% of all US federal, state and local outlays, and rising as a share of GDP too.
Kurzweil invites the reader to look at his long-run charts and guess when left- or right-wing administrations governed: the line simply rises (the 2020-21 pandemic spike overshoots the trend).
The charts on total and per-capita spending show the same steady climb. Key programmes (Medicaid, SNAP food benefits and housing assistance) are barely adequate today. Kurzweil's argument is that with GDP growing exponentially, the safety net will keep rising in total and per person, and once AI makes medicine, food and housing much cheaper in the 2030s, the same support will buy a very comfortable life without raising its share of GDP; any further rise would fund still more.
In an onstage dialogue with Chris Anderson at TED 2018 in Vancouver, Kurzweil predicted universal basic income or its equivalent in developed countries by the early 2030s and in most countries by the late 2030s, enough to live well by today's standards. It could take the form of regular payments to all adults or free goods and services, funded by some mix of taxes on automation profits and public investment in emerging technologies; related schemes might pay people who care for family or build healthy communities.
Abundance is uneven, and sharing is political
- Computing vs health. In 2022 a dollar bought over 50,000 times the computing it bought in 2000, but only about 81% of the health care (both inflation-adjusted). Some treatments, like cancer immunotherapy, improved; most spending (hospital stays, X-rays) did not.
- Who gained. Students and young people, big buyers of computing, gained a great deal; the elderly and chronically ill, big buyers of care, may have ended up worse off.
- Policy, not technology, decides. Everyone will be able to afford a high standard of living technically and economically; whether they actually get it is a political decision. Famines today stem from bad governance or civil war, not lack of food or hoarded farming knowledge, because those conditions block adaptation to drought and frustrate aid. Toxic politics could likewise block rising living standards.
- Medical distrust. COVID showed the danger in medicine. Just as vaccine conspiracy theories thrive online, rumours about decision-support AI, gene therapy or medical nanobots, fed by real cybersecurity worries, could lead people in 2030 or 2050 to refuse lifesaving treatment. Public understanding is the best defence.
Meaning after work
If politics goes right, Kurzweil argues, the main human struggle moves from meeting physical needs to finding purpose and meaning: a climb up Maslow's hierarchy. He sees it already in the young people he mentors (aged 8 to 20), who mostly choose careers for meaning, in the arts or in tackling humanity's old social and psychological problems.
- Death does not supply meaning. The common claim that mortality gives life meaning is, for Kurzweil, a rationalisation of tragedy. A loved one's death removes part of us: neocortical modules wired to enjoy that person now produce loss and pain. Death destroys what does give meaning: skills, experiences, memories, relationships, and moments of creating art, expressing love and sharing humour.
- Cloud neocortex enlarges meaning. Explaining music, literature, a YouTube video or a joke to primates that missed the hominid neocortical expansion of two million years ago is impossible; likewise, once we augment the neocortex digitally from the 2030s, we will create meaningful expressions we cannot now imagine.
The Kahneman dialogue: what happens in between?
Daniel Kahneman. Agreed that information technology's price-performance grows exponentially, will spread to physical goods like clothing and food, and leads to an era of abundance where the struggle moves up Maslow's pyramid. But he foresaw a protracted period of conflict and even violence first. Automation makes winners and losers; a laid-off driver gains nothing from humanity's climb because he may not be able to make the transition. People need time to adapt, many cannot retrain fast, and this round differs in breadth, depth and speed. Popular fears already exceed reality, and losing a job when everything seems to be getting worse breeds alienation.
Kurzweil. Disruption, even violence, should be anticipated and mitigated, but a violent transition is unlikely. Society has replaced most jobs repeatedly without lasting dislocation or revolution; mass communication and policing contain violence (as with the Luddites); the safety net is large, growing and largely independent of who governs; long-run violence keeps falling; and, most important, abundance lowers the incentive for violence, since people with long, safe lives ahead settle differences through politics.
Where they agree. Both think much of today's political polarisation comes from actual and anticipated automation, more than from traditional issues such as immigration: people anxious about their own economic security grow hostile to anything that might worsen it.
Diffuse benefits, concentrated harms. Kurzweil accepts the core asymmetry. Autonomous vehicles will save lives, cut pollution and congestion, free time and lower transport costs for a US population of almost 400 million by 2050, with estimated benefits of $642 billion to $7 trillion a year; tens of thousands of lives will be saved annually, but no one will know whose. The costs land on several million identifiable drivers, for whom aggregate gains are no comfort; they need policies that ease the economic pain and help them toward work offering meaning, dignity and economic security.
- Losers who do not object. A new cure ends income for those who treated the disease, yet it is celebrated, even by most of them, and society prefers softening their loss to blocking cures.
- Change arrives in bundles. People imagine jobs vanishing from an otherwise unchanged world; in fact positive changes accompany each loss as quickly as the disruptions.
- Rapid adoption. In the late 1980s he predicted a worldwide network open to everyone, even schoolchildren, by the late 1990s, and mobile devices to use it by the early 2000s; both seemed disruptive and unlikely, and both were absorbed fast. The app economy is barely 15 years old yet taken for granted.
- Social effects too. Stanford research estimates 39% of US heterosexual couples in 2017 met online, many via Tinder and Hinge, so many of today's schoolchildren owe their existence to apps barely older than they are.
- Augmentation, not competition. A time traveller taking a 2024 smartphone to Calvin Coolidge's 1924 would seem superhuman: effortless maths, passable translation, grandmaster-beating chess, Wikipedia's facts. We do not feel augmented, and we will no more compete with the AI of 2030-2040 than with our phones; extending reach has been technology's purpose since stone tools.
- Falling violence. Despite headline tragedies (worsened by US gun culture), violence has declined over centuries; Steven Pinker's The Better Angels of Our Nature (2011) traces this to rule of law, literacy and economic development, all strengthened by information technology.
- Abundance and safety. The technologies causing disruption also cheapen food, housing, transport and health care; crime should keep falling through better education, smarter policing and less lead and other brain-damaging toxins.
Why they see it differently. The two men attributed their outlooks to their childhoods. Kahneman spent his formative years fleeing the Nazis with his family in France, witnessing conflict and hatred arguably rooted in the deprivation of post-First World War Europe, especially Germany. Kurzweil was born after the war in relatively safe New York City, shaped by the Holocaust as one of the "generation after".
Conclusion. A constructive transition will still need enlightened politics and policy, so politicians and civic leaders keep an important role. The prize, in Kurzweil's words, is overcoming humanity's age-old afflictions.
Earnings. The $133,000-per-worker figure divides total personal income (including transfers, pensions, rent and dividends) by the labour force, so it overstates pay; the hourly-wage series is sounder. Welfare chart. Its bars (US about 19.3%, Australia and Switzerland about 19-19.5%, Canada 17%) do not match the text (18.7%, 16.7%, 18.0%), probably different years or definitions. Bitcoin. The mid-2023 volume of $180 million a day reflects a narrow exchange sample; most trackers reported billions. Brynjolfsson is cast as a job pessimist, but his "Turing Trap" argument is that policy should favour augmentation over automation, close to Kurzweil's own view. Frey and Osborne put 47% of US employment at high risk "perhaps over the next decade or two", and were criticised for scoring whole occupations rather than tasks. Kahneman, who died in March 2024, is represented only through Kurzweil's account of a private conversation. The UBI forecast depends on political will the chapter assumes more than demonstrates, and the claim that the safety net grows regardless of party sits uneasily with recent US cuts to programmes such as SNAP and Medicaid.
The Yale Budget Lab (analysis of February 2026) found the US occupational mix shifting only modestly since ChatGPT's release in late 2022 and no longer unemployment spells for AI-exposed workers: "stability, not major disruption". The pain is at the bottom of the ladder. Brynjolfsson's Stanford "Canaries in the Coal Mine" study of ADP payroll data found employment of 22-25-year-olds in the most AI-exposed occupations (such as software development and customer service) 13% below less-exposed peers in 2025, widening to 19% in its August 2026 update, while older workers were spared. US productivity rose about 2.7% in 2025, nearly double the past decade's 1.4%. Verdict: the "tasks before professions" pattern holds; displacement is real but narrow.
Waymo passed 20 million paid trips and about 500,000 rides a week in ten-plus US metros by mid-2026, went fully driverless in Las Vegas in July 2026, aiming for one million weekly rides by year-end. The largest US cash-transfer study (OpenResearch, results July 2024) gave 1,000 low-income people $1,000 a month for three years: recipients worked about 1.3 hours a week less than controls, with no strong measured health gains. No country has adopted a universal basic income. Verdict: automated driving is ahead of schedule; the safety-net half of Kurzweil's bargain is not.
2026 scorecard
| Prediction or trend (book) | Book's figure / date | Status, Oct 2026 | Evidence |
|---|---|---|---|
| Robotaxis spread beyond Phoenix and SF | LA next | Ahead | Waymo ~500,000 paid rides a week in 10+ US cities |
| Driving jobs lost in large numbers | Near term | Not yet | ~3,500 Waymo cars; no measurable fall in driver jobs |
| Productivity puzzle resolves | As AI spreads | Early signs | US productivity +2.7% in 2025 vs 1.4% decade average |
| AI hits jobs | 2020s | Narrow so far | Young workers in AI-exposed jobs 19% below peers |
| Seamless AI search, translation | 2020s | Largely met | Chat assistants answer questions; live speech translation ships in phones |
| AR on glasses and lenses | 2020s | Partial | AR glasses on sale; contact-lens displays not commercial |
| Inflation turns to deflation | As goods become IT | Behind | 2021-23 inflation surge; prices still rising |
| Nanobots link brain to cloud | 2030s | Too early | No medical nanorobots in humans |
| UBI or equivalent | Early 2030s (developed) | Behind pace | No national UBI; pilots only |
The Next Thirty Years in Health and Well-Being
A mechanic understands every part of a car, so a well-kept car can run almost indefinitely; medicine still works by approximations that are right for most patients but perhaps not for you. Kurzweil's claim is that medicine is becoming an information technology: in the 2020s AI and simulation transform drug discovery, diagnosis, trials and surgery; in the 2030s and 2040s nanotechnology lets us repair and rebuild the body cell by cell. Once nanobots can selectively repair or destroy individual cells, the annual risk of death will stop rising with age, and medicine will become the exact science it has always aspired to be.
The chapter has three parts, following the source: the 2020s merger of AI and biotechnology; the 2030s-2040s maturing of nanotechnology (its physics, history, controversies and economics); and the application of nanotechnology to health, longevity and the brain.
The 2020s: AI meets biotechnology
Why medicine must become information technology
Kurzweil's starting contrast is the car mechanic. Automotive engineering is effectively exact, so even the worst wreck is repairable in principle; medicine, despite two centuries of progress, still uses treatments known to work without knowing why, calibrated to the average patient. Making it exact means turning it into an information technology so it rides the exponential curve. The change is under way: in 2023 the first drug designed end to end by AI entered phase II trials for a rare lung disease.
- Linear past. While medicine depended on painstaking lab work and doctors handing expertise to the next generation, progress was plodding and linear.
- Exponential future. AI learns from more data than any doctor, accumulates experience from billions of procedures rather than a career's thousands, and improves with its hardware.
- Exhaustive search. Biochemical problems can now be solved by digitally trying every option, in hours rather than years.
Drug discovery as search
Kurzweil calls designing treatments for new viruses the most important problem class today. It resembles finding the one key that opens a virus's chemical lock from a pile big enough to fill a swimming pool. A researcher might name a few dozen promising molecules; the relevant set is usually in the trillions, of which billions merit serious computational study; and the space of physically possible drug molecules has been estimated at about 10⁶⁰. AI lets scientists focus on the keys most likely to fit.
Why drug development is slow:
- Sequential trials. Once a candidate exists, a few dozen to a few hundred subjects are tested over months or years for tens or hundreds of millions of dollars; the first option is often not ideal, so alternatives start the cycle again.
- Attrition. Only 13.8% of candidates pass all three phases of US trials to FDA approval (MIT study).
- Time and money. About a decade and $1.3-2.6 billion per new drug.
2019 Flinders University (Australia) designs a "turbocharged" flu vaccine: one simulator generates trillions of candidate chemicals that might activate the immune system, a second tests each as an immune booster.
2020 MIT screens 107 million molecules in hours, returns 23 candidates and highlights two that kill some of the most drug-resistant bacteria known. Jacob Durrant (University of Pittsburgh) notes that physically testing over 100 million compounds would be impossible. MIT then starts designing antibiotics from scratch.
2020 Moderna's COVID vaccine (below), the most important medical use of AI that year.
2021 USC builds a tool that designs candidate vaccines against new COVID variants in under a minute and validates them digitally within an hour.
China released the SARS-CoV-2 sequence on 11 January 2020. Using machine-learning tools to analyse which vaccine would work best, Moderna had its mRNA vaccine sequence two days later; the first clinical batch was made on 7 February, sent to the NIH after preliminary testing on 24 February, and on 16 March, 63 days after sequence selection, the first trial participant was dosed. Vaccines had typically taken five to ten years; Kurzweil credits the speed with saving millions of lives.
Protein folding and AlphaFold
All these are instances of one deeper problem: protein folding. DNA specifies a string of amino acids, which folds into a 3D shape that largely determines what the protein does; since the body is mostly protein, linking sequence to function is the key to new medicines. No single simple rule predicts folding, so human accuracy was low and discovery depended on luck and labour.
AlphaFold's leap:
- 2018, AlphaFold 1. DeepMind skipped the usual catalogue of template shapes (as AlphaGo Zero skipped human game knowledge) and came first of 98 entrants, predicting 25 of 43 proteins accurately against the runner-up's 3.
- 2021, AlphaFold 2. Rebuilt around transformers (the technique behind GPT-3), it reached near-experimental accuracy for almost any protein.
- Scale. Known structures jumped from about 180,000 to hundreds of millions, soon billions.
What comes next in discovery
- AI chooses the problem. Today scientists define the problem, translate it into chemistry and set simulation parameters. Over coming decades AI will search creatively, for example noticing that a subgroup of patients responds badly to standard treatment and proposing novel therapies.
- Bigger simulations. Modelling scales from proteins to protein complexes, organelles, cells, tissues and whole organs.
- Cancer. Immunotherapies such as CAR-T, BiTEs and immune checkpoint inhibitors have saved thousands but often fail as tumours learn resistance, frequently by altering their microenvironment in ways we cannot yet model; simulating tumour and surroundings will let therapies be tailored to beat resistance.
- Neurodegeneration. Alzheimer's and Parkinson's involve misfolded proteins accumulating in a living brain that cannot be studied directly; simulation will expose root causes and allow treatment before debilitation.
- Mental health. Disorders will affect more than half of Americans at some point. Blunt drugs like SSRIs and SNRIs adjust chemistry temporarily, help modestly, fail for some and carry long side-effect lists; a full functional model of the brain, "the most complex structure in the known universe", will let treatments target causes.
Simulated trials
The FDA already incorporates simulation results into approvals, especially valuable when a new virus demands fast vaccines. Kurzweil's goal is to digitise trials altogether: test a drug on tens of thousands of simulated patients over simulated years, in hours or days.
Human trials. A dozen to a few thousand subjects, depending on drug and phase. Response depends on genetics, diet, lifestyle, hormones, microbiome, disease subtype, other drugs and other diseases, so few if any subjects resemble you. A typical result: an average 15% improvement across 3,000 people, which may hide harm to you.
Simulated trials. Can reveal hidden subgroups: say 250 people with a particular gene get 50% worse while 500 with kidney disease improve 70%. Many such correlations yield a personal risk-benefit profile for every patient, faster and with more statistical power.
- Order of arrival. Single-molecule drugs are easiest and will be simulated first; CRISPR and gene-expression therapies, with many interacting molecules and structures, will take longer. Replacing human trials requires modelling a drug's effect on the whole body over long periods.
- Resolution. How much detail is needed is unclear (thumb skin cells hardly matter for a liver-cancer drug), but validating the method will probably require digitising the whole body at roughly molecular resolution, to learn what can safely be abstracted away. Long-term, yet with meaningful progress by the end of the 2020s.
- Resistance. Doctors will rightly demand a track record at least as good as current methods. Liability adds a second motive: nobody wants to be the one who approved a promising therapy that turned into a disaster. Regulators must balance caution against lives lost to delay.
Genomics, surveillance and diagnosis
Noncoding DNA. The 98% of the genome that does not code for proteins, once written off as "junk", governs gene expression, but its relationships are very hard to read. AI's sensitivity to faint patterns is breaking the logjam: in 2019 a New York team led by Olga Troyanskaya linked non-inherited, noncoding mutations to autism, which she called the first clear demonstration of such mutations causing any complex human disorder.
Epidemic surveillance. Epidemiologists used to pick among imperfect data sources. Harvard's ARGONet (lead researcher Mauricio Santillana) fuses electronic medical records, historical data, live Google searches and spatial-temporal patterns of flu spread, continuously re-weighting each by predictive power. In 2019 research it beat all previous approaches, outperforming Google Flu Trends in 75% of states studied and predicting state flu activity a week ahead of the CDC's usual methods.
Clinical diagnosis. In a 2018 speech Kurzweil predicted neural nets would read radiology images as well as doctors within a year or two; two weeks later Stanford announced CheXNet.
- CheXNet (Stanford, 2018). A 121-layer convolutional network trained on 100,000 X-rays diagnosed 14 diseases better than the doctors it was compared with.
- Paediatrics (2019). A network reading natural-language clinical data diagnosed childhood diseases better than eight junior physicians and, in some areas, all 20 doctors tested.
- DELFI (Johns Hopkins, 2021). Patterns of DNA fragments in blood detected 94% of lung cancers with a simple lab test, beyond expert humans.
- TREWS (Nature Medicine, July 2022). Across 590,000+ hospital patients, early sepsis warnings cut sepsis deaths by 18.7%; sepsis kills about 270,000 Americans a year, so wider use could save tens of thousands.
Kurzweil expects AI to reach superhuman performance at virtually all diagnostic tasks during the 2020s, with models drawing on wearable data to suggest treatment before people feel ill. Imaging comes first because its information is pure, quantifiable pixels, unlike diagnosis that blends many qualitative sources, and because significant signs can be too subtle for human eyes; systems like CheXNet and its cousin CheXpert should generalise to other images and may find hidden risk factors in apparently healthy organs.
Surgery follows as surgical data and computing grow:
- 2016, STAR. The Smart Tissue Autonomous Robot beat human surgeons at intestinal stitching in animal tests.
- 2017, China. A robot completed a full dental implant operation on its own, a high-precision procedure.
- 2020, Neuralink. A surgical robot automated much of implanting a brain-computer interface, aiming at full autonomy.
- The experience gap. A surgeon does several hundred operations a year, at most a few tens of thousands in a career (fewer in long, complex specialties). A robotic system learns from every operation it performs worldwide (potentially millions) and can rehearse billions of simulated ones, including rare disease combinations and complex trauma most surgeons never see.
The 2030s and 2040s: nanotechnology
Why biology loses to engineering
Biology produced a creature with the brains and opposable thumbs to build technology, but it is far from optimal, especially for thinking. Hans Moravec argued in 1988 that however finely we tune DNA-based biology, it will trail purpose-engineered systems; in Peter Weibel's phrase, humans can only be "second-class robots". AI plus nanotechnology will let us rebuild bodies, brains and surroundings molecule by molecule.
| Substrate | Speed | Capacity | Status |
|---|---|---|---|
| Human neuron | ≤200 firings/s (1,000 theoretical), average probably <1 | ~10¹⁴ ops/s per brain (Kurzweil; within 10× of Moravec) | Biological |
| Transistor / retail chip | >1 trillion cycles/s / >5 billion | US Frontier >10¹⁸ ops/s on an AI benchmark | 2023 hardware |
| Drexler interlock computer | Switch in 10⁻¹⁰ s; 6 nm³ per interlock | ~10⁹ ops/s | Conceptual |
| Merkle mechanical logic (2018) | 100 MHz | 10²⁰ gates and up to 10²⁸ ops/s per litre, ~100 W (needs high surface area to shed heat) | Conceptual |
The brain's clunky architecture explains the gap, and computers can also be denser, physically larger and networked. All 8 billion human brains need under 10²⁴ ops/s (10¹⁴ each times about 10¹⁰ people), so a perfectly efficient litre of nanologic would equal about 100 trillion human brains. Kurzweil calls his 10¹⁴ estimate conservative: the brain's wet, turbulent interior means any neuron may die or misfire, so it relies on massive parallelism to cancel noise; a clean, stable chip needs less redundancy, so a mind might run on fewer operations. He stresses he does not claim the theoretical maximum is achievable, only that the headroom is staggering; a fraction of a percent would still be revolutionary.
From idea to assembler
Late 1940s-1955 John von Neumann (1903-1957) formalises self-replication in lectures and a 1955 Scientific American article (collected only in 1966): a universal computer runs a program that drives a universal constructor, which copies both the whole machine and the program, indefinitely. Abstract and logical rather than physical.
1959 Richard Feynman (1918-1988), "There's Plenty of Room at the Bottom": physics does not forbid manoeuvring things atom by atom, so a physicist could in principle make any substance a chemist writes down by putting the atoms where the chemist says.
Mid-1980s K. Eric Drexler founds modern nanotechnology: an assembler that uses atoms and fragments from ordinary substances, directed by an onboard computer, to build anything atomically stable. Engines of Creation (1986).
1992 Drexler's Nanosystems gives several chemistries for a single-tip arm that picks up and releases an atom or small molecule mechanically and electrically, notably building diamondoid.
1995 Drexler's ideas inspire Neal Stephenson's Hugo-winning The Diamond Age, in which diamond nanotech defines an era as bronze and iron once did.
1997 Ralph Merkle designs an assembler "metabolism" that builds hydrocarbons such as diamondoid from a butadiyne feedstock.
2001-2003 Smalley-Drexler debate; Kurzweil comments in 2003, mainly siding with Drexler.
2018 Merkle and collaborators design an all-mechanical nanocomputer (conceptual).
Why self-replication matters. Nanotechnology affects big objects only through self-replication. Kurzweil's process: a central computer broadcasts commands to a small seed of nanobots sitting in feedstock; they copy themselves in a cascade to trillions; then the computer tells them what to build. "Dumb" raw materials are given information through an exponential process.
SIMD broadcast control. As in the single instruction, multiple data chip architecture, one unit reads the program and transmits it to trillions of assemblers, each with a simple computer. The construction machine itself is a simple single-arm molecular robot like von Neumann's constructor; molecular arms, gears, rotors and motors have been demonstrated repeatedly.
Gray goo. The worst case (chapter 7) is self-replicating nanobots in an uncontrolled chain reaction that could consume much of Earth's biomass. Broadcast control is Kurzweil's main defence: if all instructions come from one source, shutting it off in a bug, breach or runaway leaves the nanobots physically unable to keep replicating.
Richard Smalley (from 2001). Physics rules out a nano-arm grasping and carrying atoms like a hand; only biology-style bottom-up self-assembly is a sensible goal. Two objections: "fat fingers" (manipulators too bulky to position a reaction's atoms) and "sticky fingers" (atoms cling to the tool).
Eric Drexler. Top-down fabrication of nanobots from scratch is the ultimate goal. A single-tip arm avoids fat fingers, and enzymes and ribosomes show sticky fingers can be overcome. Kurzweil, looking back almost two decades, says progress makes top-down look more and more plausible, though the field needs at least a decade more to mature, helped by AI.
Diamondoid and the pathways to mechanosynthesis
- Diamondoids. Cages of as few as ten carbon atoms in the simplest diamond-crystal arrangement, capped with hydrogen; candidate building blocks for light, very strong structures. Chemical vapour deposition already makes artificial diamond, and doping can tune properties like thermal conductivity or create transistors.
- New carbon forms since 2005. Graphene (one-atom-thick hexagonal carbon), carbon nanotubes (rolled graphene) and carbon nanothreads (near one-dimensional strands with hydrogen), all with wide applications ahead.
- Eight research pathways. DNA origami; DNA nanorobotics; bio-inspired molecular machines; molecular Lego; single-atom qubits; electron-beam atom placement; hydrogen depassivation lithography; scanning-tunnelling-microscope manufacturing; plus several stealth projects.
- Forecast. With superhuman engineering AI by the end of the 2020s running detailed chemistry simulations, Kurzweil expects atom-by-atom manufacturing sometime in the 2030s.
The economics of nanomanufacturing
A mature assembler might be a tabletop unit making almost anything for which it has the atoms. The hard part is generalisability: one machine that masters many chemistries needs very advanced AI. So chemically uniform goods (gemstones, furniture, clothing) will come well before complex ones (a cooked meal, a bionic organ, a computer stronger than all unenhanced brains).
Nano-economics in Kurzweil's account:
- Marginal cost. Pennies per pound, essentially the precursor atoms, including for assemblers themselves.
- Drexler (2013). About $2 per kilogram for anything, diamonds or food.
- Mass. Nanomaterials far stronger than steel or plastic need about a tenth of the mass; costly gold, copper and rare earths can often give way to carbon.
-
Cultural premium. Lab diamonds, identical to the eye, sell 30-40% below natural ones; old-master paintings fetch roughly a million times a good reproduction. Value migrates to information. Kurzweil's model is the book. Hand-copied books were mostly labour; printed books mostly paper, binding and ink; e-books cost effectively nothing to copy, so the price pays only for the creative assembly of information (plus marketing). His test: on Amazon, blank hardcover journals and novels with similar bindings are hard to tell apart by price, but a blank e-book at any price is absurd. Today a physical product's value comes from raw materials, labour, machine time, energy and transport; all are falling (automated extraction, robots, cheaper machines, solar plus storage and eventually fusion, autonomous electric vehicles), so the information share will approach 100%.
-
Free goods. Like Google searches and Facebook Likes, nano-made products may be free, paid for by watching political ads or sharing data, or offered by governments as rewards for volunteering, continuing education or healthy habits.
- No point hoarding. Distribution depends on culture and politics, but Kurzweil is optimistic: nobody bottles air, and others reading Wikipedia costs you nothing; material abundance extends that logic.
- Residual scarcity. Historic diamonds and Rembrandts stay scarce by convention, though values shift over generations, and today's children may choose differently.
Nanotechnology for health and longevity
- Phase 1 (now ending). Applying current pharmaceutical and nutritional knowledge, the regimen of his book Transcend that he has followed for decades.
- Phase 2 (2020s). Biotechnology merged with AI: therapies discovered and tested in digital simulators in days rather than years.
- Phase 3 (2030s). Nanotechnology overcomes the limits of biological organs, taking people far past the 120-year limit.
Why 120 is a wall
Only Jeanne Calment (France, 122) is documented past 120. The limit is not simply many years of exposure to stroke, cancer and heart attack:
- Age 90-110. Annual death risk rises about 2 percentage points a year: a US man of 97 has about a 30% chance of dying before 98, then 32% before 99.
- After 110. Risk rises about 3.5 points a year, and supercentenarians develop qualitatively new failures, kidney and respiratory failure that seem spontaneous rather than linked to lifestyle or disease: the body simply breaks down.
Aubrey de Grey (biogerontologist, founder of the LEV Foundation) likens ageing to engine wear: damage that accumulates from normal operation. Metabolism produces waste and oxidation (rust); repeated cell division accumulates copying errors. The young repair efficiently; eventually damage outpaces repair. In one's 70s-90s, one fatal problem usually comes well before several, so curing an 80-year-old's cancer may buy almost a decade, but in the end everything fails at once and treating symptoms stops working. The answer is to cure ageing itself; the SENS (Strategies for Engineered Negligible Senescence) Research Foundation has a detailed agenda that will take decades. Kurzweil's preferred tool for cell- and tissue-level repair is nanorobots: not immortality (accidents remain) but no rise in annual death risk with age.
If research adds at least a year to your remaining life expectancy every year, you can wait for nanomedicine to fix the rest. Hence de Grey's claim that the first person to live 1,000 years is probably already born: if 2050 nanotechnology lets 100-year-olds reach 150, there is until 2100 to fix the next problems, with AI making progress exponential.
Against "life would be too long". People mourn a life cut short yet balk at extending all lives; Kurzweil replies that few want to die unless in great pain, and the improvements of chapter 4 would relieve most such pain, so longer life means better life. Americans born in 1924 expected about 58.5 years (death around 1982), yet many lived into the 2000s and 2010s to enjoy cheap flights, safer cars, cable TV and the internet. Babies born in 2024 will gain years in a world changing exponentially faster, with more art, music and games, and above all more time loving and being loved, which Kurzweil considers life's deepest meaning.
What medical nanobots will be and do
Design constraints. Diamondoid parts with sensors, manipulators, computers, communicators and perhaps power supplies. Not tiny metal submarines: at the nanoscale water is a strong solvent, oxidants are highly reactive, and sticky friction dominates fluid flow, like swimming through peanut butter. Nanobots will need new propulsion, will draw energy from their surroundings, and will take outside control signals or compute collectively. The body has several tens of trillions of cells; one cell-sized nanobot per 100 cells means several hundred billion nanobots, though far sparser ratios might suffice.
- Organ maintenance. Most non-sensory organs add substances to blood or lymph or remove them: lungs swap oxygen and CO₂, liver and kidneys clear toxins, the gut adds nutrients, the pancreas and others release hormones (insulin devices already mimic a pancreas). Nanobots monitor and adjust these, maintain organ structure, and can ultimately replace organs.
- Optimisation. Push hormone levels beyond normal for energy, focus and healing. "Backdoor life extension": cutting sleep from eight to seven hours adds as much waking life as five years of lifespan.
- Prevention. Routine maintenance should stop major diseases arising, though in the interim existing cancers still need treating.
- Cancer. Every cancer cell can self-replicate, so all must go; tumours evade immunity and cancer stem cells survive treatment. Chemotherapy is blunt and weakens immunity; immunotherapies fall short. Nanobots can inspect each cell and destroy every malignant one.
- Gene control. Biology's decentralised DNA (each cell copies its own) protects against single-cell mutations but makes whole-body editing hard. A nanoengineered nucleus in each cell, fed by a "central server" (in practice perhaps hundreds or thousands of micro-scale control units, still vastly more central than tens of trillions of independent cells), would make amino-acid chains from broadcast code, allowing bad genes to be switched off, epigenetics to be regulated once AI can simulate it, and transcription errors, a major cause of ageing, to be prevented or reversed. Ribosomes could be augmented the same way.
- Emergencies. Destroy bacteria and viruses, halt autoimmune reactions, clear clogged arteries; a Stanford-Michigan State nanoparticle already finds and kills the monocytes and macrophages behind atherosclerotic plaque. Human-initiated at first, later autonomous with reports to human monitors.
- Early detection. About 25% of ischaemic strokes are cryptogenic (no detectable cause); patrolling nanobots could spot small plaques or defects, break up forming clots, or raise the alarm during a silent stroke.
- Augmentation. Protein machines must fold from one-dimensional chains, limiting speed and strength; diamondoid gears and rotors could be thousands of times faster and stronger. Robert A. Freitas's respirocyte, an artificial red blood cell, would let a person hold their breath for about four hours; artificial lungs and nano hearts would end heart attacks and make traumatic cardiac arrest rare.
- Appearance. People already customise online avatars across age, gender and species; whether radical real-body changes become as common, or culture keeps people conservative, is open.
Rebuilding the brain
The most important role is the brain, which will become over 99.9% nonbiological by two routes: nanobots repairing or replacing neurons in tissue, and brain-computer links that allow thought control of machines and add cloud neocortex (chapter 2). A deeper virtual neocortex would allow thoughts beyond current comprehension, such as intuitively reasoning about ten-dimensional shapes.
- Density. The cerebral cortex holds about 16 billion neurons in half a litre; Merkle's design could fit over 80 quintillion logic gates in the same space.
- Speed. Neurons switch roughly once a second on average; nanocomputation about 10⁸-10⁹ cycles a second.
- Cost, 2023. $1,000 bought up to 130 trillion computations per second, about one brain's 10¹⁴.
- Cost, 2053. On the 2000-2023 trend, $1,000 (2023 dollars) buys about 7 million times one brain's capacity.
- Sensitivity. If only part of the brain need be digitised, the milestone comes years sooner; if every protein must be simulated (unlikely, he thinks), a few decades later, still within many current lifetimes. Exponential trends make the date robust to assumptions.
In the 2040s and 2050s, Kurzweil predicts, we will rebuild bodies and brains far beyond biology: running faster and longer, swimming and breathing underwater, even growing working wings; thinking millions of times faster; and, most important, keeping backups so that the self no longer depends on the survival of any one body.
Facts. The 98% figure is noncoding DNA as a share of the genome, not "98 percent of genes" as the text says. Feynman's 1959 lecture is the field's founding vision, but the word nanotechnology was coined by Norio Taniguchi in 1974. Moderna's interval is 63 days here but 65 in chapter 7, and its two-day design relied mainly on NIH's prior spike-protein work, not just machine learning. Drexler's cost is $2/kg here but 20 cents/kg in chapter 4. Science. Demographers dispute that mortality keeps accelerating after 110; some data (Barbi et al., Science, 2018) suggest a plateau near 50% a year after 105. Smalley's objections were never fully answered experimentally, and no medical nanorobot of the kind described exists; the timetable rests on AI solving engineering problems by 2029. The "free products for watching political ads" scenario raises obvious manipulation concerns the chapter does not address.
The 2023 drug Kurzweil mentions is Insilico Medicine's rentosertib (TNIK inhibitor) for idiopathic pulmonary fibrosis; its phase IIa results, published in Nature Medicine in June 2025, showed good safety and dose-dependent lung-function trends, and a 320-patient phase III began dosing at 47 Chinese centres in September 2026. AlphaFold 3 (Google DeepMind and Isomorphic Labs, 8 May 2024) extended prediction to proteins' interactions with DNA, RNA and drug-like molecules; code followed in November 2024 for academic use. Demis Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry for AlphaFold 2. Isomorphic raised $2.1 billion in May 2026 but had not yet dosed a human. In April 2025 the FDA announced a plan to phase out animal-testing requirements, starting with monoclonal antibodies, in favour of AI computational models, organoids and organ-on-chip data. Verdict: discovery is accelerating; no AI-discovered drug is yet approved.
In early 2026 the FDA cleared Life Biosciences' ER-100, the first human trial of partial epigenetic reprogramming (three Yamanaka factors, OSK), for glaucoma and the optic-nerve stroke NAION, with enrolment from March 2026; it tests rejuvenating eye cells, not whole-body ageing. On Kurzweil's nanotech path, a May 2026 arXiv preprint from CBN Nano Technologies (with Merkle and Freitas as theory collaborators) reported positionally controlled addition of carbon dimers to a silicon surface with an inverted-mode scanning tunnelling microscope at 84-97% yields: the first step toward diamondoid mechanosynthesis, still far from a self-replicating assembler. Verdict: real first steps, decades of engineering from the nanobots of the 2030s.
2026 scorecard
| Prediction (book) | Date | Status, Oct 2026 | Evidence |
|---|---|---|---|
| AI-designed drugs move through trials | 2020s | On track | Insilico's rentosertib in phase III (Sept 2026) |
| Simulated trials begin augmenting human ones | End of 2020s | Early | FDA plan to replace animal tests with AI models (Apr 2025) |
| Superhuman AI at most diagnostic tasks | 2020s | Partial | Strong on imaging; real-world gains uneven |
| Autonomous surgical robots | 2020s | Partial | Robot-assisted surgery common; autonomy confined to research |
| Atom-by-atom manufacturing | 2030s | Early lab steps | Positional C₂ mechanosynthesis preprint (May 2026) |
| Life extension phase 2 (AI + biotech) | 2020s | Starting | First partial-reprogramming human trial cleared (2026) |
| Longevity escape velocity | ~2030 | Behind | No therapy shown to slow human ageing |
| $1,000 buys ~7 million brains of compute | 2053 | Too early | Price-performance still rising; not testable yet |
Peril
The technologies that bring abundance also raise the danger: destabilising nuclear weapons, engineered pathogens, self-replicating nanobots and, above all, superintelligent AI. Kurzweil expects civilisation to survive them, not because the threats are unreal but because the stakes concentrate human ingenuity and each dangerous field also yields defensive tools. Superintelligent AI is the primary peril, since an intelligence smarter than its creators could find a way around any precaution; the best protection is to align AI technically, distribute its power broadly, and strengthen the human institutions it grows out of.
Promise and peril
Two epigraphs frame the debate. Environmentalist Bill McKibben urges accepting a world with "enough wealth and enough technological capability" and pursuing no more; Robert M. Pirsig (Zen and the Art of Motorcycle Maintenance) calls the flight from technology self-defeating, since the Buddha resides as comfortably in a computer's circuits as on a mountaintop. Kurzweil sides with Pirsig. The chapter takes four perils in rising order of difficulty, each paired with its defence.
| Peril | Mechanism of catastrophe | Main defence in Kurzweil's account | Difficulty |
|---|---|---|---|
| Nuclear weapons | Exchange kills hundreds of millions; fallout, soot cooling, famine kill billions | MAD deterrence, treaties, shrinking arsenals, AI detection and safer command systems | Known, long-managed |
| Biotechnology | Engineered virus with high lethality, high spread, stealth | Asilomar rules, rapid-response teams, one-day sequencing, AI-designed countermeasures | Known defences |
| Nanotechnology | Self-replicating "gray goo" consumes the biomass | Broadcast architecture, fail-safe design, "blue goo" immune system deployed first | Strategy exists |
| Superintelligent AI | Misuse, outer and inner misalignment | AI-assisted alignment, norms, governance, broad distribution of power | No general solution |
Nuclear weapons
Kurzweil's generation was the first to live with a civilisation-ending technology; he recalls primary-school civil-defence drills crouching under desks with arms over heads, a precaution he notes drily must have worked, since he survived.
- Arsenal. About 12,700 warheads worldwide, about 9,440 active and usable.
- Hair trigger. The US and Russia each keep about 1,000 large warheads launchable on under half an hour's notice.
- Toll. A major exchange could kill several hundred million directly; it could not kill everyone because people are dispersed, but fallout, city-fire soot causing severe global cooling and mass starvation, and the collapse of medicine and sanitation could kill billions. Future warheads might be "salted" with cobalt to prolong radioactivity.
- Expert odds (2008). Anders Sandberg and Nick Bostrom surveyed the Global Catastrophic Risks Conference at Oxford's Future of Humanity Institute: median 30% chance of at least one million nuclear-war dead before 2100, 10% of at least one billion, 1% of extinction.
- Treaties. Active warheads cut from a 1986 peak of 64,449 to under 9,500; atmospheric testing halted; outer space kept nuclear-free.
| Country (2023, book) | Warheads | Delivery |
|---|---|---|
| Russia | 5,889 | Full triad |
| United States | 5,244 | Full triad |
| China | 410 | Full triad |
| France | 290 | Limited |
| United Kingdom | 225 | Limited |
| Pakistan | 170 | Full triad (per book) |
| India | 164 | Full triad |
| Israel | ~90 (unacknowledged) | Believed full triad |
| North Korea | ~30 | Limited |
A triad means ICBMs, air-dropped bombs and submarine-launched missiles. Remaining arsenals still suffice to end civilisation, and a low annual risk compounds over decades into a serious one: as long as high-alert arsenals persist, use somewhere, by a government, terrorists, rogue officers or accident, is probably a matter of time.
Mutually assured destruction (MAD) guided US and Soviet strategy for most of the Cold War. Its logic is game-theoretic: if even one nuclear strike triggers full retaliation, striking is suicidal, so nobody strikes. It works only if neither side can block the other's missiles, since effective defence makes a first strike survivable (some theorists add that fallout would still ruin the attacker: "self-assured destruction", SAD). Partly to protect that equilibrium, militaries have invested little in missile defence, and in 2023 no nation could confidently survive a large attack.
- Destabilisers. Russian nuclear-armed underwater drones and nuclear-powered cruise missiles that loiter near a target and strike from unpredictable angles; hypersonic manoeuvring vehicles raced by Russia, China and the US. Their novelty raises the risk of miscalculation when rivals judge their effectiveness differently.
- Complacency. We are so used to the danger that it is barely discussed.
- Measured optimism. MAD has held 70+ years, arsenals keep shrinking, AI improves tools for detecting nuclear terrorism and dirty bombs, and smarter command-and-control can cut the risk of sensor malfunctions triggering inadvertent launch.
Biotechnology
Most natural pathogens either spread well and rarely kill, or kill but spread poorly. The Black Death combined both, killing about a third of Europeans and cutting world population from about 450 million to 350 million by the end of the fourteenth century; genetic variation (a benefit of sexual reproduction) let some immune systems fight it off. Genetic engineering could produce, deliberately or by accident, a supervirus that is extremely lethal, highly transmissible and perhaps stealthy (spread long before symptoms), against which no one has immunity. COVID-19 (2019-2023) was a "pale glimpse".
1975 Asilomar Conference on Recombinant DNA, fifteen years before the Human Genome Project, sets standards against accidental and deliberate harm; the guidelines are continually updated and partly written into law.
1996 Full HIV genome sequenced, 13 years after the virus's discovery.
2003 SARS sequenced in 31 days.
June 2015 CDC's Global Rapid Response Team (GRRT), created after the 2014-2016 West African Ebola outbreak, can deploy anywhere to help identify, contain and treat outbreaks.
Today Many viruses sequenced in one day. US bioterror defence is coordinated by the National Interagency Confederation for Biological Research (NICBR); Kurzweil has advised USAMRIID via the Army Science Board on rapid response.
May 2020 Kurzweil argues in Wired for AI-designed COVID vaccines.
2020 Moderna doses its first subject 65 days after receiving the sequence and wins FDA emergency authorisation 277 days later; the previous fastest vaccine took about four years.
The rapid-response recipe: capture the virus, sequence it in about a day, then design countermeasures. Two options: RNA interference (small RNA pieces that destroy the messenger RNA expressing a gene, treating viruses as analogous to disease genes) and antigen vaccines targeting distinctive surface proteins. AI already finds candidates in days or weeks; later in the 2020s, simulated biology will speed a growing share of clinical trials.
Lab leak. Investigations into whether COVID escaped from genetic-engineering research were ongoing as Kurzweil wrote; he urges reliance on high-quality sources given the misinformation. Either way the lesson holds: it could have been far worse, combining high spread with high lethality (one reason he doubts malicious creation), and since the tools to make something deadlier already exist, AI-driven countermeasures are essential.
Nanotechnology
As with cells, a single rogue nanobot is no global threat; catastrophe requires self-replication. Mature nanotech will enable many offensive weapons and, unlike nuclear arms, make them cheap.
- Nuclear cost benchmark. South Korea estimated North Korea's programme cost $1.1-3.2 billion by 2016, when it fielded nuclear-armed missiles.
- Biological cost. A 1996 NATO report: a bioweapon for $100,000 (about $190,000 in 2023), five biologists, a few weeks, no exotic equipment. A 1969 UN expert panel judged bioweapons about 800 times as cost-effective as nuclear ones against civilians, a ratio surely higher now.
- Nano cost. Because nanotech replicates like biology, bioweapon costs are a first approximation; AI-optimised manufacturing may push them lower.
- Weapon types. Undetectable poison-delivery drones; nanobots inhaled or drunk that tear the body apart; systems that selectively target any group.
- Drexler's 1986 warning. Artificial "plants" could crowd out real ones with inedible foliage, and tough omnivorous "bacteria" could "reduce the biosphere to dust in a matter of days".
- 10⁴⁰ carbon atoms in Earth's biomass; about 10⁷ per nanobot, so 10³³ copies needed, built through iterated doubling rather than directly.
- About 110 doubling generations (2¹¹⁰ ≈ 10³³), or 109 if parents stay active.
- About 100 seconds per replication (Robert Freitas), so about three hours in theory; in practice weeks, because biomass is scattered and tiny nanobots move slowly, so the destruction front limits speed.
- Two-phase attack. Quietly convert 1 in 10¹⁵ carbon atoms worldwide into dormant nanobots, too dilute to notice; on a signal (perhaps relayed by nanobots self-assembled into long-wave radio antennas) each needs 50 doublings, under 90 minutes, and travel no longer matters.
- Blue goo. About 88,000 tonnes of defensive nanobots, optimally spread, could sweep the atmosphere in about 24 hours (Freitas): less than a large aircraft carrier's displacement, though ideal conditions are unlikely.
- Accidents count as much as malice. A carelessly designed nanobot meant to eat one material, or stay in one area, could run away through a programming error; so build only inherently fail-safe nanobots rather than bolting safety onto dangerous ones.
- Broadcast architecture. Nanobots carry no program of their own and take every instruction from a signal (perhaps radio) that can be cut or changed in an emergency.
- Immune system first. Bad actors can build unsafe nanobots anyway, so a nanotech "immune system" must already exist, able to catch even stealthy, low-concentration replication.
- Blue goo of special materials. Defenders must be built from materials gray goo cannot convert, unlike the abundant natural ingredients gray goo uses. Tricky, but Kurzweil sees no fundamental asymmetry favouring attackers.
- Deploy good before bad. Defenders must be in place worldwide before harmful replicators appear.
- Guidelines. Safety rules date from the 1999 Workshop on Molecular Nanotechnology Research Policy Guidelines (Kurzweil attended), revised since.
His friend Bill Joy's essay "Why the Future Doesn't Need Us" (2000) discusses gray goo well. Most experts, and Kurzweil, consider it unlikely, but as an extinction-level risk it must stay in view; with precautions and AI help in designing secure systems, he hopes it stays science fiction.
Artificial intelligence
Pandemics remain possible (COVID killed nearly seven million people through 2023), but sequencing and AI-designed medicines give a path to avert civilisational catastrophe, and nano has a defensive strategy even against the two-phase attack. Superintelligent AI is different in kind: an AI smarter than its makers could circumvent any precaution, and no general strategy can definitively prevent that. Focused research can mitigate it.
- Misuse. The AI works as its operators intend, and they intend harm: terrorists using an AI's biochemistry skills to design a pandemic virus.
- Outer misalignment. The goal taught differs from the goal meant, the genie problem of literal commands: told to design a virus killing all cells with a particular cancer mutation, the AI succeeds, but the mutation also sits in healthy cells and the virus kills the patients.
- Inner misalignment. The method learned misbehaves in some cases: an AI picks up a spurious pattern (cancer samples stored longer than healthy ones, leaving storage-related changes), so its virus fails on live patients. Detecting such errors grows harder as tasks grow complex.
- Imitative generalisation. Train AI to imitate how humans draw inferences, so it behaves sensibly in unfamiliar situations.
- AI safety via debate. Competing AIs point out flaws in each other's ideas, letting humans judge questions too complex for them alone.
- Iterated amplification. Weaker aligned AIs help humans build stronger aligned AIs, repeated until AIs far beyond unaided human reach are aligned.
Alignment will be very hard, but we can use AI to align AI, and to resist misuse: an aligned AI must recognise a dangerous request, like the virus design, and refuse. That still needs ethical bulwarks: strong international norms for safe deployment.
Principles, pledges and military policy
Kurzweil attended the 2017 Asilomar Conference on Beneficial AI, modelled on the 1975 biotech meeting, and signed its 23 principles, six of which invoke human values or humanity. Undemocratic actors could still use AI for their ends even if most of the world complies, and the major military powers have not signed, though militaries have long driven advanced technology (the internet came from DARPA).
- No. 7, Failure Transparency. If an AI causes harm, it should be possible to find out why.
- No. 8, Judicial Transparency. Any autonomous role in judicial decisions needs an explanation a competent human can audit.
- No. 10, Value Alignment. Highly autonomous AI should be designed so its goals and behaviour can be assured to align with human values throughout operation.
- No. 11, Human Value. AI should be compatible with human dignity, rights, freedoms and cultural diversity.
- No. 18, AI Arms Race. An arms race in lethal autonomous weapons should be avoided.
- No. 23, Common Good. Superintelligence only for widely shared ethical ideals and all humanity, not one state or organisation.
His objections to the principles: transparency cannot deliver understanding of a superintelligence (even the best Go player could not follow a superhuman engine's explanations), so he backs research on eliciting latent knowledge, making an AI report everything relevant it knows rather than what it thinks we want to hear. Principles 18 and 23 are laudable, but superintelligence offers decisive military and economic advantage, so powers have strong incentives to race, which worsens misuse risk and tempts developers to neglect alignment. And stating values does not guarantee them: a harmful AI could readily justify its objective in terms of widely shared values, which is the whole peril.
The pledge. The Lethal Autonomous Weapons Pledge says the decision to take a human life should never be delegated to a machine, since no one would then be culpable. Signers include Stephen Hawking, Elon Musk, Martin Rees and Noam Chomsky. The Campaign to Stop Killer Robots has won no major power except China (2018), which later said it backed banning use, not development, likely for strategic reasons, since US autonomous weapons would disadvantage Beijing.
The powers. The US, Russia, UK, France and Israel rejected the pledge. A 2012 Pentagon directive requires "appropriate levels of human judgment" over force; Deputy Defense Secretary Robert Work (2016) ruled out delegating lethal authority but left open reversal if a rival delegated more. Kurzweil, who helped shape this policy (as he did biohazard policy), would want a counterweapon if attacked by autonomous machines, which would mean breaking the ban.
- 2023 Political Declaration. After a conference including dialogue with China, the US issued a Political Declaration on Responsible Military Use of AI and Autonomy, urging ultimate human control over nuclear weapons.
- "Human control" is hazy. How much discretion should an AI ordered to "stop an incoming nuclear attack" have? Any AI able to do so could also attack.
- Dual use. The drone that delivers medicine to a hospital cut off by rainy-season roads could deliver a bomb there; military drones have for over a decade been precise enough to put a missile through a particular window on the far side of the world.
- Who is "human"? Brain-computer interfaces add nonbiological thinking from the 2030s; that part grows exponentially while biology stays fixed, so by the late 2030s our thought will be largely nonbiological, blurring where human decision-making sits.
Governance, markets and relinquishment
Restricting a fundamental AI capability is very hard because general intelligence is so broad. Kurzweil welcomes the Bletchley Declaration after the 2023 UK AI Safety Summit, though implementation is what counts. His market argument: every step toward superintelligence must win market acceptance; AGI will be built to solve real human problems with strong incentives toward benefit, and since it grows out of a deeply integrated economy, "in an important sense it will be us": we are already a human-machine civilisation. The most important safety measure is therefore to protect and improve human governance and social institutions, and to keep advancing the ethical ideals that have already cut violence.
Relinquishment. Increasingly strident "Luddite" voices urge giving up genetics, nanotechnology and robotics (GNR) to avoid their dangers. Delay has real victims: African famine worsened when governments refused food aid that might contain GMOs.
Fundamentalist humanism. Opposition to any change in what it means to be human: editing genes and protein folding, radical life extension. It will fail because demand for therapies against the pain, disease and short lives of "version 1.0" bodies will prove irresistible.
Two objections to radical life extension get short replies:
- Resources. Optimised use of Earth's resources reveals them to be thousands of times our needs; sunlight alone is nearly 10,000 times current energy demand (chapter 4).
- Boredom. Compact VR and AR devices in the 2020s, then nanobots feeding the senses directly in the 2030s, bring life expansion as well as extension: vast virtual worlds limited only by an imagination that is itself expanding; centuries would not exhaust knowledge and culture.
Conclusion. Kurzweil calls AI the technology that will let us overcome disease, poverty, environmental damage and human frailty, a moral imperative to pursue while mitigating risk. Returning to his school drills: nearly everyone then assumed nuclear war was inevitable, and the restraint humanity showed proves we can handle biotech, nanotech and superintelligence too. Since AI amplifies intelligence regardless of values, the answer to abuse is a world where AI's power is broadly distributed and reflects humanity's values as a whole: cautious optimism.
Arsenal details. Pakistan has no operational submarine-launched ballistic missile, so the "full triad" label is generous; the 12,700 and 9,440 figures and the country table come from different counting rules (total inventory versus stockpile). Moderna's interval is 65 days here, 63 in chapter 6 (counted from sequence selection rather than receipt). The market argument sits awkwardly with the chapter's own warning that arms-race incentives push developers to cut safety corners, and "it will be us" assumes the gains spread rather than concentrate. The alignment methods cited were early research agendas, not demonstrated solutions, and the claim that superintelligence has no general containment strategy is stated, then left largely unresolved by the optimism that follows.
2026 scorecard
| Claim (book) | Book's view | Status, Oct 2026 | Evidence |
|---|---|---|---|
| Arsenals continue to shrink | Optimistic | Reversing | Usable stockpiles rising; New START lapsed Feb 2026 |
| AI could help design bioweapons | Key misuse risk | Materialising | Labs added bio safeguards; AI beats 94% of virologists at troubleshooting |
| Governments take AI risk seriously | Bletchley 2023 | Mixed | 2nd International AI Safety Report; EU high-risk rules delayed |
| Use AI to align AI | Promising | Active research | 12 firms published safety frameworks in 2025 |
| Major powers avoid an AI arms race | Asilomar No. 18 | Not met | US-China competition over chips and frontier models intensifying |
SIPRI's 2026 Yearbook counts about 12,187 warheads in January 2026, with 9,745 in military stockpiles (up from about 9,440) and 4,012 deployed; China reached about 620, the fastest-growing arsenal. New START, the last US-Russia limit, expired on 5 February 2026 with no replacement. On biology, the International AI Safety Report 2026 (chaired by Yoshua Bengio, 3 February 2026, 100+ experts from 30+ countries) found leading models outperform 94% of domain experts at troubleshooting virology lab protocols. Anthropic activated its stricter ASL-3 safeguards for Claude Opus 4 in May 2025 because it could not rule out meaningful uplift toward biological weapons. Verdict: Kurzweil's nuclear optimism has aged badly; his bio-misuse worry is now mainstream.
The same report found 12 companies published or updated frontier safety frameworks in 2025, more than double the year before, but warned of an "evaluation gap": models can sometimes detect tests, so dangerous abilities may be missed. The EU AI Act's general-purpose-model duties took effect on 2 August 2025, but a May 2026 "digital omnibus" deal postponed high-risk obligations to December 2027 (stand-alone systems) and August 2028 (embedded in products). Verdict: more safety practice than Kurzweil assumed, less binding governance.
Dialogue with Cassandra
In a short Socratic exchange, Kurzweil lets a sceptic named after Troy's ignored prophet attack the weakest joint in his timeline: AI that beats every human by 2029 arrives through fast, unregulated computing, while the brain link that lets us merge with it depends on slow, regulated neurosurgery. If the brain connection slips by a decade, machines could take the jobs before humans can absorb their intelligence, and Kurzweil concedes that this would be a real problem.
Why a dialogue, and why Cassandra
The chapter is a single conversation between "Cassandra" and "Ray", placed straight after the chapter on peril. The name matters: in Greek myth Cassandra foresaw Troy's fall and was never believed. Kurzweil gives her the strongest objection a careful reader could raise, which is not that his forecasts are wrong one by one but that two of them run on different clocks. Unlike the essay chapters, the dialogue ends in partial agreement rather than rebuttal, and it is the one place in the book where Kurzweil explicitly accepts a risk to his own schedule.
The agreed premises
Cassandra opens by getting Ray to confirm four forecasts made earlier in the book, so the dispute is about their combination, not their truth:
- 2029 capability. A neural net with enough processing power will surpass all human capabilities, already happening one skill at a time and complete by 2029.
- Dumbing down for Turing. To pass a Turing test, such an AI will have to be made less capable, or its superhuman answers would reveal it is not an unenhanced human.
- Early-2030s brain link. Technology will reach inside the skull to the top layers of the neocortex, both reading activity and activating connections.
- Merger. Through those links to the cloud, the superintelligence humanity builds becomes part of our own brains.
The argument, move by move
| Move | Cassandra's objection | Ray's reply | Outcome |
|---|---|---|---|
| 1. Two clocks | AI research is largely unregulated, experiments take days, advances stack quickly; a million-wire implant in the most sensitive organ faces heavy regulation, and regulators may doubt it is even needed beyond treating disease | Medical demand (fixing serious brain disorders) will drive it anyway; that is why he dates it to the 2030s, not sooner | Cassandra accepts feasibility, predicts delay |
| 2. Ten-year slip | Regulation could push the link to the 2040s; machines would then take the jobs instead of extending people | External brain extenders already work as mind extensions: a child's phone reaches all human knowledge; AI augments far more workers than it replaces; internal devices are merely harder to lose | Unresolved |
| 3. Bandwidth | The merger needs millions of circuits; keyboards are orders of magnitude slower, so why would an AI bother with humans at all? | By the mid-2020s, fully immersive VR/AR with full-field video and audio, interlaced with ordinary sight and sound, gives two-way communication thousands of times faster than typing and "almost as fast" as a neocortical link | Cassandra: better interfaces are not neocortex extension |
| 4. Meaning | If AI does every intellectual task far better and faster, what gives humans purpose? | That is the reason to merge: the AIs become part of us, so we are the ones doing those things; meanwhile people still need to work for food and shelter, aided by external extenders | Pushes weight back onto the brain link |
| 5. Conflict | (Via Kahneman) violence between the displaced and "others": the computers? | Not computers, on which everyone will depend, but humans seen to use AI to grow their wealth and power at displaced workers' expense, in an interim before abundance | Agreed diagnosis |
| 6. Purpose limits conflict | External extenders differ in kind from cloud neocortex | Purpose reduces conflict; just as extra neocortex lifted early primates from survival instinct to philosophy, cloud neocortex will expand empathy and ethics; the interim should be short because the link comes in the early 2030s | Ray concedes a delay "could be a huge problem" |
| 7. Identity | An emulation replacing the biological you would seem to be you, but you would be gone | No replacement is proposed: the biological brain stays, with intelligence added | Cassandra presses on |
| 8. Insignificance | The added part will be thousands to millions of times more powerful, leaving biology "pretty insignificant" | Nothing is taken away; we already use the cloud; the biological brain keeps all its fundamental qualities | Shared verdict: profound change, very fast |
What the dialogue settles, and what it leaves open
Settled. Both sides accept that the AI forecast and the brain-link forecast come from different fields with different speeds, and that the gap between them is where the social risk of chapter 5 lives. Both accept Kahneman's picture of an in-between period in which some humans hold power and AI has not yet produced enough abundance to defuse conflict. And both locate the likely target of anger in people who own and wield AI, not in machines.
Open. Kurzweil's bridge across the gap is external augmentation: phones now, immersive VR and AR by the mid-2020s. Cassandra never accepts that this equals neocortical extension, and Ray only partly contests it. His fallback on timing is simply to restate the early-2030s forecast, so the short interim is asserted rather than shown.
Kurzweil's whole optimistic case for jobs and meaning depends on humans merging with AI soon after AI surpasses them. Here he admits that if the brain link lags the capability by a decade, the merger cannot cushion the disruption, which makes neural-interface timing the critical path of his entire forecast.
This is Kurzweil's most candid passage: he concedes the core risk without resolving it. Cassandra's regulatory argument is the stronger, since existing implants serve paralysed patients, and his fallback (VR "almost as fast" as a neural link) undercuts chapter 2's claim that the merger needs millions of direct connections. His identity reply also sidesteps chapter 3, where he treats gradual replacement of neurons as identity-preserving; here he insists nothing is replaced. The "made less smart to pass the Turing test" line also concedes that the test measures human imitation, not intelligence.
| Forecast restated in the dialogue | Date | Status, Oct 2026 |
|---|---|---|
| AI exceeds all human capabilities | 2029 | Kurzweil reaffirmed it in April 2026; frontier models excel unevenly |
| Immersive VR/AR interlaced with reality | Mid-2020s | Behind: headsets niche; display glasses just launched |
| Neocortex-to-cloud connection | Early 2030s | Behind: ~two dozen people with medical implants |
| AI augments more workers than it replaces | Now | Broadly holds: displacement narrow, concentrated among young entrants |
Neuralink had implanted about 26 people by mid-2026 (from 21 in January), well short of Musk's hope of more than 1,000 implants in 2026. Consumer immersion remains limited: Apple's Vision Pro (February 2024, $3,499) sold modestly, and Meta's Ray-Ban Display glasses (September 2025) put a small screen in one lens, far from the full-field, always-on AR Kurzweil expected by the mid-2020s. Verdict: Cassandra's warning looks prescient; AI capability is outpacing the bandwidth to merge with it.
Price-Performance of Computation, 1939–2023 Chart Sources
The appendix documents the book's central chart: computations per second per constant 2023 dollar. Across 84 years the best machine improved from about 0.0000065 (Zuse's Z2) to about 130 billion (Google's TPU v5e), a gain of about 2 × 10¹⁶.
| Year | Machine | Real price (2023 $) | Computations/s | Per dollar |
|---|---|---|---|---|
| 1939 | Zuse Z2 | 50,489 | 0.33 | 0.0000065 |
| 1946 | ENIAC | 11.6 million | 5,000 | 0.00043 |
| 1959 | DEC PDP-1 | 1.24 million | 100,000 | 0.081 |
| 1965 | DEC PDP-8 | 172,370 | 312,500 | 1.81 |
| 1975 | Altair 8800 | 3,482 | 500,000 | 144 |
| 1984 | Apple Macintosh | 7,244 | 1.6 million | 221 |
| 1994 | Pentium (75 MHz) | 4,478 | 87.1 million | 19,451 |
| 1999 | Pentium III (450 MHz) | 898 | 722 million | 803,952 |
| 2006 | Core 2 Duo E6300 | 274 | 14.9 billion | 54 million |
| 2008 | Nvidia GTX 285 | 503 | 708.5 billion | 1.41 billion |
| 2017 | AMD Radeon RX 580 | 282 | 6.1 trillion | 21.6 billion |
| 2021 | Google TPU v4-4096 | 22.8 million | 1.1 × 10¹⁸ | 48.3 billion |
| 2023 | Google TPU v5e (chip) | 3,016 | 393 trillion | 130 billion |
How the series is built
- Machines. Each point is a major programmable computer whose price-performance beat all previous machines; if several did so in one year, only the best appears. Non-commercial machines are dated by first normal operation, commercial ones by first public availability. Unbuilt or unreliable machines (Babbage's Analytical Engine, Zuse's Z1), homebrew PCs and specialised DSPs are excluded. The full list has 38 entries, including the Z3 (1941), Colossus (1943), UNIVAC 1103 (1953), Data General Nova (1969), Compaq Deskpro 386 (1986) and five Nvidia GPUs from the GTX 285 to the Titan X.
- Prices. Purchase, construction or rental cost in February 2023 dollars (chained CPI-U), trustworthy only to several percent (more for foreign currencies); lowest open-market price preferred. Delivery, power, operators, maintenance and depreciation are excluded; since these weighed most on old machines, omitting them understates progress, the conservative choice. From the mid-1990s chip prices replace whole systems, which modestly exaggerates 1990s gains.
- Rented TPUs. Google's TPUs are rented only, so the v4-4096 is costed at 4,000 rental hours to match purchased equipment; per-hour pricing would show a far higher figure, accurate for small machine-learning jobs and an underappreciated benefit of cloud computing. The v5e price is inferred from Google's claim of 2.7× the v4's price-performance (MLPerf inference).
- Performance. No single metric spans a 1939 relay machine and a 2023 TPU (the Z2 could never do what a TPU does in a fraction of a second). So each machine keeps its native metric: instructions per second until the Pentium 4 (2001), FLOPS after, using best-achieved rather than average speed. Sandberg and Bostrom's MIPS-to-MFLOPS conversion (it does not scale linearly) and the information-theoretic metrics of Moravec and Nordhaus (whose MSOPS fixes a ratio of additions to multiplications) were rejected as misleading over this range. Some Xeon-era values are estimated from Dongarra's benchmarks via a 2.559 average ratio.
The Z2 cost 7,000 reichsmarks. By purchasing power that is about €30,100 (2020), or $40,124 in 2023; by exchange rates $2,800, or $60,853. The first distorts through rationed, state-set prices, the second through different purchasing power, so the chart takes the average, $50,489; the Z3 gets the same treatment ($136,849). Colossus, never sold, is costed at about five Bombes (each about £100,000) on Jack Copeland's rough guess: $33.8 million, meaningful to two significant figures.
- About 54 doublings in 84 years: one every ~19 months on average.
- 2000-2023: Pentium III (919,725) to TPU v5e: one doubling every ~16 months, the figure Kurzweil quotes.
- Plateaus. Progress is not smooth: the 2015 Titan X barely beat the 2012 GTX 680 (5.3 vs 5.0 billion per dollar), then the cheap 2017 RX 580 quadrupled it.
Estimates. The Z2's price averages two conversions of 7,000 reichsmarks; Colossus (1943) is costed at about five Bombes on a historian's guess; TPU prices are unofficial (the public $1.20 per chip-hour gives about 82 billion). The 2001 metric switch and the TPU v5e's low-precision INT8 rating flatter recent gains.
Master Takeaways
By October 2026 the computing half of Kurzweil's forecast is on or ahead of schedule, while the merger, nanobots and longevity that make his future benign are well behind.
- One engine drives everything. Information technology makes its own next step cheaper, so computations per dollar rose about 2 × 10¹⁶ from 1939 to 2023, whatever the paradigm.
- 2029 and 2045 are unchanged. A valid Turing test by 2029 and the Singularity by 2045, first forecast in 1999 and 2005, still anchor the timeline.
- Merge, don't compete. Brain-cloud links in the 2030s make AI part of us; the danger lies in any gap between superhuman AI and the link.
- Intelligence is compute applied to data. Connectionist nets stalled for decades, then won once compute became cheap; scale, not a new theory, explains the 2020s surge.
- Progress is real but invisible. Poverty, violence and illiteracy have fallen for two centuries; news incentives and evolved biases hide gradual gains from most people.
- Medicine is becoming information technology. AI drug search, simulated trials and later repair nanobots aim to stop the yearly risk of death rising with age.
- Exponentials need a feedback loop. Kurzweil limits his law to fields where technology helps build its own successor; transatlantic speed peaked with Concorde because engines do not design engines.
- Three exponentials converge on AI. Cheaper compute, richer data and better algorithms compound together; frontier training compute doubled every 5.7 months from 2010, far faster than hardware.
- Any skill with feedback data can be mastered. If performance can be scored, even through indirect proxies, a deep-learning model can be trained to surpass all humans; data, like oil, gets mined as its value rises.
- Human-level intelligence is a bundle. Abilities arrive unevenly, so AI may be superhuman in most fields while trailing the best humans in a few; programming skill is the fibre that triggers an intelligence explosion.
- A valid Turing test needs rules. The 2002 Kapor bet fixes interview length, judge skill and enhancement limits; a passing AI must hide its superhuman abilities, so the test gauges imitation more than intelligence.
- The neocortex is the template. Its value lies in hierarchy plus broad propagation, shared patterns and millions of simultaneous patterns; deep learning re-creates that design in silicon.
- The brain needs about 10¹⁴ operations a second. Revised firing rates put a working brain near 10¹³ to 10¹⁴, reachable for $1,000 in 2023; simulating every protein would cost decades, not centuries.
- Consciousness is a leap of faith. Subjective experience cannot be detected from outside, so Kurzweil adopts Chalmers's panprotopsychism and extends moral regard to systems complex enough to plausibly feel.
- Identity is a continuous pattern. Like the Charles River, a self survives replacement of its material; gradual replacement keeps you you, while a separate copy is a new person.
- Free will survives determinism. Wolfram's class 4 systems can only be known by running them, so a deterministic brain still decides; the emergent processes are us, not a puppeteer.
- The steep part has started. The last two centuries came from the slow early stretch of the curves; AI now pulls food, housing, energy, water and goods onto the same trajectory.
- Clean energy is a doubling problem. Solar's share doubling roughly every 28 months would cover all electricity in about a decade; storage, not generation, is the bottleneck.
- Longevity escape velocity is the target. Survive until research adds more than a year of life expectancy per year, around 2030 for the diligent, and later therapies do the rest.
- Ageing is damage that can be repaired. Following Aubrey de Grey, the fix is cell-level maintenance rather than treating diseases one by one, because at old age everything fails at once.
- Simulated trials beat small human ones. Virtual cohorts can expose subgroups that an average 15% benefit hides, giving every patient a personal risk-benefit profile.
- Physical goods become information. Molecular assemblers under broadcast control could make products for pennies per kilogram, leaving value almost entirely in designs and software.
- Tasks go before professions. Automation strips tasks first (ATMs, legal research) and reshapes jobs; nonskilling AI then removes whole roles, starting with about 4.6 million US drivers.
- GDP measures price, not value. A $900 chip grew 72,000× more powerful from 1999 to 2023 yet counts the same; free digital services add consumer surplus that statistics miss.
- Abundance funds the safety net. If AI cheapens food, housing and care, today's spending share buys a good life, so UBI or its equivalent arrives in developed countries in the early 2030s.
- Distribution is political. Technology makes abundance possible; governance decides who gets it, as modern famines stem from bad government rather than lack of food.
- Benefits diffuse, harms concentrate. Driverless cars help hundreds of millions slightly and hurt a few million drivers badly, which shapes the politics of every automation wave.
- The transition need not be violent. Against Kahneman's fear of conflict, Kurzweil cites a safety net that grows under every party, long-run falling violence and abundance that lowers the stakes.
- Superintelligence is the primary peril. Nuclear, bio and nano risks have known defences; an AI smarter than its makers can evade any fixed precaution.
- Nuclear restraint is the precedent. Most people in the 1950s thought nuclear war inevitable; decades of restraint are Kurzweil's evidence that humanity can also govern bio, nano and AI.
- Use AI to align AI. Debate, iterated amplification and imitative generalisation enlist aligned systems to check stronger ones; broad distribution of AI power and sound institutions are the political safeguard.
- Design self-replication to fail safe. Broadcast architecture and a pre-positioned immune system of "blue goo" are Kurzweil's answers to gray goo.
- The brain link is the weak joint. By 2026 AI capability runs ahead of schedule while brain interfaces and nanobots lag far behind, exactly Cassandra's warning, which Kurzweil concedes.
Frameworks & mental models
| Name | What it says | Use it when |
|---|---|---|
| Accelerating returns | Information technologies feed their own improvement, so price-performance grows exponentially | Forecasting a field going digital |
| Six epochs | Physics, DNA, brains, technology, merger, saturated matter: stages of information processing | Placing AI in a long-run frame |
| Symbolic vs connectionist | Hand-written rules hit a complexity ceiling; nets learn from data but are opaque | Choosing rules or learned models |
| Three exponentials | Compute, data and algorithms improve together and multiply | Judging if a capability will scale |
| Feedback-data rule | Any skill with clear performance data, or good proxies, can reach superhuman level | Spotting AI's next task |
| Intelligence as a bundle | Abilities arrive unevenly; "human-level" means the best humans | Reading benchmarks and timelines |
| Neocortex hierarchy | Self-modifying recognisers with broad propagation and shared patterns | Explaining why scale matters |
| Two consciousnesses | Functional awareness is observable; subjective experience is not | Debating AI moral status |
| Three origin theories | Dualism, physicalism, panprotopsychism; Kurzweil picks the third | Framing machine-mind claims |
| Identity as pattern | Gradual replacement preserves the self; a copy is a new person | Thinking about uploads |
| Biases that hide progress | Threat vigilance, fading affect, nostalgia, availability, base-rate neglect | Reading trend data soberly |
| Four bridges | Lifestyle; AI plus biotech (2020s); nanorobots (2030s); mind backup (2040s) | Staging longevity claims |
| Longevity escape velocity | Research adds over a year of life expectancy per year | Testing anti-ageing progress |
| Three automation waves | Deskilling, upskilling, nonskilling, each with different job effects | Assessing AI's effect on a role |
| Consumer surplus gap | GDP counts price paid, not value received | Reading productivity data |
| Diffuse gains, focused harms | Gains spread thin; losses fall on identifiable groups | Designing transition policy |
| Three AI perils | Misuse, outer misalignment, inner misalignment | Structuring AI risk reviews |
| AI aligning AI | Debate, iterated amplification, imitative generalisation | Planning model oversight |
| Broadcast architecture | Self-replicators obey a central signal that can be cut | Designing agent kill switches |
Who's who
| Name | Role | Why they matter |
|---|---|---|
| Alan Turing | Mathematician | Proposed the imitation game Kurzweil expects AI to pass by 2029 |
| Marvin Minsky | MIT AI pioneer, mentor | Perceptrons stalled neural nets; argued compute did not matter |
| Frank Rosenblatt | Psychologist | Built the Perceptron, ancestor of deep learning |
| I. J. Good | Statistician | Foresaw the intelligence explosion in 1965 |
| Mitch Kapor | Lotus founder | Took the other side of the 2002 bet on a 2029 Turing test |
| David Chalmers | Philosopher | The hard problem; panprotopsychism, Kurzweil's chosen view |
| Stephen Wolfram | Scientist | Computational irreducibility, the basis for free will |
| Steven Pinker | Psychologist | Data on falling violence and rising well-being |
| Daniel Kahneman | Psychologist | Biases behind pessimism; warned of a violent transition |
| Erik Brynjolfsson | Economist, Stanford | Net-job-loss thesis; GDP misses digital value |
| Aubrey de Grey | Biogerontologist | Ageing as repairable damage; longevity escape velocity |
| Eric Drexler | Nanotech founder | Molecular assemblers; debated Smalley on feasibility |
| R. Merkle, R. Freitas | Nanotech theorists | Broadcast architecture, respirocytes, gray goo maths |
| Hans Moravec | Roboticist | Independent brain-compute estimate near 10¹³ to 10¹⁴ |
Kurzweil's charts persuade where they track price-performance, which genuinely has compounded, and are weakest where he extends the exponential to things that saturate or depend on politics: democracy, UBI, food, longevity. Many curves are S-curves shown as exponentials, and the "best machine each year" method flatters the computing trend. Several numbers clash between chapters (Moderna's 63 vs 65 days, Drexler's $2 vs 20 cents per kilogram, a "ten-billion-fold" compute rise that the cited doubling time puts in the tens of millions). The book says little about who controls AI, the energy and capital demands of data centres, or the chance that AI arrives on schedule while biology and nanotechnology do not. By 2026 that gap is visible: frontier models advance fast, while brain implants number in the dozens, nuclear arsenals are growing again and no country has a UBI. Read the book as a disciplined extrapolation of real trends, not a timetable.
Connect it
- AI strategy. Kurzweil's distinction between tasks and professions is the practical lens for deployment: map a role's tasks, automate the routine ones, and redesign the job around judgement and relationships, as banks did with tellers after ATMs.
- Data strategy. The feedback-data rule tells a firm where AI will win first: processes with clean, scored outcomes. Instrumenting a process so its results can be measured is often worth more than buying a bigger model.
- Enterprise architecture. His broadcast architecture for nanobots is a sound pattern for autonomous AI agents: central policy, revocable credentials and a kill switch that halts every instance at once.
- Banking. Consumer-surplus and deflation arguments matter for credit and pricing models: digital services with near-zero marginal cost compress fee income, and AI-exposed entry-level roles (operations, customer service) are where workforce effects show first.
- Risk management. The misuse, outer and inner misalignment triad maps onto model-risk practice: malicious use, specification errors and spurious correlations in training data each need separate controls and testing.
About this book
The Singularity Is Nearer: When We Merge with AI by Ray Kurzweil, 2024. 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