← The canon · AItopiaOrAImageddon?

"The Singularity Is Near: When Humans Transcend Biology"

prediction · Ray Kurzweil · 2005

A dated claim, carried graded: the claim in its own words, when it was made, when it came due, and what actually happened. Where a predictor graded themselves, their grade and an independent one, saying they differ. The point of the kind is a base rate for how wrong AI forecasting runs.

Descends from Moravec's paradox, Computing Machinery and Intelligence.

This canon has been citing this entry for two days before it existed. Five files already in canon/ invoke "the Kurzweil pattern" or "the Kurzweil problem" as a named standard — turing-1950, turing-halting-1936, godel-incompleteness-1931, moravecs-paradox-1988 and goodharts-law-1975 all justify grading their own subject on the ground that a canon which grades Kurzweil should grade everyone. README.md's rule that a self-grade must be set beside an independent one is written from his case. The spec for this kind names him in the second person: "Kurzweil is the case in point: he scores his own hit rate generously and outside reviewers score it lower."

So this file has an unusual first obligation. Before anything else it has to check whether the pattern the rest of the canon has been assuming is the pattern that is actually there.

It is, and it is not the one the shorthand implies. The shorthand is: he grades himself generously, therefore he is wrong. Both halves need work. The self-grading gap is real and is larger than almost anyone states — 86% by his own count against 7% to 42% from five independent recounts, a spread of up to twelve to one. And the single most famous claim in this book, human-level machine intelligence by 2029, is currently the best-looking long-horizon artificial-intelligence forecast anyone made in the twentieth century. Three years out, with a $20,000 wager riding on it, made in 1999 when the field's own median was fifty years away.

Those two facts are both true and an entry carrying only one is worth less than this one. The base rate this kind exists to assemble is not "forecasters are too optimistic." It is more specific and more useful, and this entry is where it gets stated: the mechanism and the date fail independently, and the mechanism here failed while the date may yet land. Everything below is organised around that.

prediction is the right kind and proposals.md filed it correctly. One rival deserves a paragraph, because two-thirds of this book is not forecast at all. The Law of Accelerating Returns is a thesis about the past — that evolutionary and technological change is exponential, that the exponent is itself growing, that this is visible in the historical record if you plot it right. That is the shape of an interpretation, which is how superintelligence-2014 was filed after refusing prediction for Bostrom on the ground that a philosopher's product is argument. The refusal does not transfer. Bostrom's timelines were scaffolding for a conditional argument; Kurzweil's argument is scaffolding for the timelines. His product is forecasting, sold as forecasting, for four decades, and — the part that settles it — his credential is his hit rate. He publishes a scorecard. Third parties introduce him with the percentage. A predictor who offers a track record as the reason to believe the next forecast is the exact case the self-grade rule was written for, and filing him as interpretation would put the dates in a file that does not have to carry them.

Set against the canon's other two adjudications: amodei-2024-loving-grace accepted prediction because an executive's timeline is an instrument; frey-osborne-2013 accepted it over the authors' explicit denial that they were forecasting. Kurzweil is neither hard case. He is the base case — the one against which those two had to be argued.

The descent, and the ancestors this canon does not have

Both edges are documented in the book's own pages, which is not true of most edges in this canon.

moravecs-paradox-1988. Hans Moravec is named on page 24 — "Hans Moravec's 1988 book Mind Children came to a similar conclusion by analyzing the progression of robotics" — and again for Robot: Mere Machine to Transcendent Mind (1999), whose 2040s robots Kurzweil quotes as "children of our minds." He is also listed in the acknowledgements among the peer expert readers, for "artificial intelligence, robotics." The intellectual edge is heavier than the citation. The canon's Moravec file records his estimate: about 100 million MIPS to match a human brain, derived by scaling from the retina, arriving in a $1,000 machine such that "computers suitable for humanlike robots will appear in the 2020s."

Now compare what Kurzweil did with it. His brain figure is 10^16 calculations per second — two orders of magnitude above Moravec's 10^14 — and his date for that capacity in a personal-computer-size device is the end of the 2010s, earlier than Moravec's. He raised the target a hundredfold and pulled the deadline in by a decade. That is a fact about how this forecast was constructed, it is checkable against a file already in this canon, and section 3 grades both men against the same 2026 hardware.

turing-1950. The operational criterion for 2029 is Turing's, by name and by protocol: "we can expect computers to pass the Turing test, indicating intelligence indistinguishable from that of biological humans, by the end of the 2020s" (p. 25). The Long Bet Kurzweil registered in 2002 is a written specification of Turing's §5 imitation game with the parameters filled in, and section 4 turns on the difference between that specification and the five-minute test that was actually passed in 2025.

Ancestors this entry wants and cannot draw. Page 22 opens with epigraphs from Vernor Vinge's 1993 "The Technological Singularity" and I. J. Good's 1965 "Speculations Concerning the First Ultraintelligent Machine"; page 23 credits von Neumann's 1950s remark, via Ulam, as "the first reference to the Singularity as an event capable of rupturing the fabric of human history," and page 24 adds Vinge's 1983 Omni piece, his 1986 Marooned in Realtime, and Damien Broderick's The Spike. None of Vinge, Good, von Neumann or Ulam is in canon/, and the absence of Vinge is the largest single gap this file ran into — the modern meaning of the word is his, and this book is a popularisation of his 1993 essay with a hardware curve attached. moore-1965 is proposed in proposals.md and unwritten; it is the ancestor this entry most wants, because Chapter Two is an argument that Moore's law is one paradigm among five and the sequence outlives any of them.

Descendants, not ancestors, and worth naming so the arrows point the right way. scaling-laws-2020 is the compute-extrapolation argument made quantitative and falsifiable inside one paradigm; bitter-lesson-2019 is the same claim about compute's primacy argued from fifty years of AI research rather than from a log plot of history; superintelligence-2014 and amodei-2024-loving-grace both write in a genre this book created a market for. When a reading meets "the curve says the capability is coming, therefore the spend is rational," it is meeting Kurzweil's argument with the curve swapped out.

What it is

The Singularity Is Near: When Humans Transcend Biology, Ray Kurzweil, Viking (Penguin Group USA), first published 2005, ISBN 0-670-03384-7. Six hundred and fifty-two pages: prologue, nine chapters, epilogue, an appendix restating the Law of Accelerating Returns, a hundred pages of notes beginning at p. 497, index at p. 603. The acknowledgements name a research team of a dozen people and a list of peer expert readers including Moravec, Ralph Merkle, Tomaso Poggio, Neil Gershenfeld and Robert Freitas, with Bill Gates, Eric Drexler and Marvin Minsky credited for the dialogues reproduced in the text. It was Kurzweil's fifth book and the successor to The Age of Spiritual Machines (1999), where the 2029 date first appeared.

The argument has four moves.

One: history is exponential and intuition is linear. "Most long-range forecasts of what is technically feasible in future time periods dramatically underestimate the power of future developments because they are based on what I call the 'intuitive linear' view of history rather than the 'historical exponential' view" (p. 11). Progress compounds; observers extrapolate straight lines from the recent past; therefore the consensus is systematically and predictably too slow. The chapter's most-quoted sentence follows immediately:

> My models show that we are doubling the paradigm-shift rate every decade. Thus > the twentieth century was gradually speeding up to today's rate of progress; > its achievements, therefore, were equivalent to about twenty years of progress > at the rate in 2000. We'll make another twenty years of progress in just > fourteen years (by 2014), and then do the same again in only seven years. To > express this another way, we won't experience one hundred years of > technological advance in the twenty-first century; we will witness on the > order of twenty thousand years of progress (again, when measured by today's > rate of progress), or about one thousand times greater than what was achieved > in the twentieth century.

Two: the Law of Accelerating Returns, stated as three principles (p. 25, verbatim):

> • The rate of paradigm shift (technical innovation) is accelerating, right now > doubling every decade. > • The power (price-performance, speed, capacity, and bandwidth) of information > technologies is growing exponentially at an even faster pace, now doubling > about every year. > • For information technologies, there is a second level of exponential growth: > that is, exponential growth in the rate of exponential growth (the exponent).

The third is the load-bearing one and the one Kurzweil calls "double exponential" growth. It is what separates his forecast from a straight Moore's-law extrapolation, and it is the specific claim section 3 grades hardest, because unlike most of the book it is a number and the number has been measured.

Three: the hardware schedule. Human brain capacity is put at 10^16 calculations per second for a functional simulation (10^19 cps and 10^18 bits for neuron-by-neuron emulation). Against that, the milestones, verbatim from p. 25:

> We will have the requisite hardware to emulate human intelligence with > supercomputers by the end of this decade and with personal-computer-size > devices by the end of the following decade. We will have effective software > models of human intelligence by the mid-2020s. > > With both the hardware and software needed to fully emulate human > intelligence, we can expect computers to pass the Turing test, indicating > intelligence indistinguishable from that of biological humans, by the end of > the 2020s.

And the brain-scanning premise those rest on, same page: "the temporal and spatial resolution and bandwidth of brain scanning are doubling each year. […] Within two decades, we will have a detailed understanding of how all the regions of the human brain work."

Four: GNR, and the merger. Genetics, nanotechnology and robotics as three overlapping revolutions (Chapter Five). Nanobots as the hinge: "robots designed at the molecular level, measured in microns […] such as 'respirocytes' (mechanical red-blood cells). Nanobots will have myriad roles within the human body, including reversing human aging" and "billions of nanobots in the capillaries of the brain will also vastly extend human intelligence" (p. 28). Manufacturing follows: "nanotechnology-based manufacturing devices in the 2020s will be capable of creating almost any physical product from inexpensive raw materials and information" (p. 13). Six epochs of evolution end with Epoch Five, the merger of human technology with human intelligence, and Epoch Six, in which intelligence saturates the matter of the universe.

The date that gives the book its franchise is set in Chapter Three, under the section heading "Setting a Date for the Singularity," and is quoted here from secondary sources rather than from the page (see Sources):

> I set the date for the Singularity — representing a profound and disruptive > transformation in human capability — as 2045. The nonbiological intelligence > created in that year will be one billion times more powerful than all human > intelligence today.

Chapter Nine is a hundred pages of pre-emptive replies to critics, one of whose headings — "The Criticism from Malthus: Exponential Trends Don't Last Forever" — is worth holding until section 3, because in 2026 that criticism is winning on the series the book cares most about.

Why a reading would cite it

Five occasions, all live in the last week of readings.

1. It is the ur-text for "the curve justifies the spend." The 17 August 2026 reading recorded roughly $3tn of off-balance-sheet AI commitments read out of nine companies' filing footnotes, and a signed 20-year lease on a 10GW Ohio campus with about $105bn of financing backstopped by the chip supplier. Numbers of that size are not underwritten by any current revenue; they are underwritten by an extrapolation. The canonical, most-explicit, most-influential form of that extrapolation is this book, which states outright that anyone forecasting from recent experience will be wrong in a known direction. Cite this entry when the curve is doing the work — not to say the spend is unjustified, but to name the argument being made and to attach its own track record to it.

2. It is the entry LENSES.md's last rule was written around. That rule says a principal's forecast is a canon entry of kind prediction, dated and graded, and never a deposit. Kurzweil is the hardest version of the case, because he does not merely forecast — he arrives with a hit rate. The rule needs an entry that can answer the credential rather than only the claim, and section 3 is that answer.

**3. It is the base rate for which way long-horizon AI forecasts fail.** Pair it with turing-1950, which this canon graded as wrong on the date by twenty-five years and far too conservative on the magnitude. Kurzweil is the mirror: broadly right on the date so far, wrong on the magnitude in the other direction, and wrong about the route. Two entries, opposite errors, same question. A reading that assumes forecasts fail by being too optimistic has half the record.

4. It is the worked example of a self-grade escaping its own footnotes. The "86%" now travels without the method, without the corpus, and without the grader's identity attached — exactly as Frey and Osborne's 47% did. Cite this entry whenever a principal cites their own accuracy: the number to ask for is not the percentage but who scored it, against which list, and by what rule for "essentially correct."

5. Because 2029 is three years away and there is money on it. On 1 April 2002 Kurzweil and Mitchell Kapor registered the first wager on the Long Now Foundation's Long Bets: "By 2029 no computer — or 'machine intelligence' — will have passed the Turing Test." Kapor predicts, Kurzweil challenges, $20,000 to the Electronic Frontier Foundation or the Kurzweil Foundation. It resolves inside this project's likely lifetime, and it is the rare AI forecast with written adjudication rules. Section 4 records why "AI passed the Turing test" in 2025 did not settle it.

What it got right, and what it got wrong

The four things this kind requires, first and plainly.

The claim, in its own words. Three claims with three clocks. (a) "We will have the requisite hardware to emulate human intelligence with supercomputers by the end of this decade and with personal-computer-size devices by the end of the following decade." (b) "We can expect computers to pass the Turing test […] by the end of the 2020s." (c) "I set the date for the Singularity […] as 2045."

The date they were made. 2005, for this book. But (b) and (c) are older: the 2029 date was published in The Age of Spiritual Machines in 1999 and the Long Bet was registered in 2002. The clock on the famous prediction started in 1999, not 2005, and an entry that dated it to the book would be flattering the 2005 restatement with six years of hindsight it did not have. Kurzweil himself is consistent about this: "In 1999, I predicted 2029 for AGI and I still predict 2029."

The dates they were due. End of 2009, end of 2019, end of 2025, end of 2029, 2045. Three are past. Two are in flight.

What actually happened. Item by item below. Today is 17 August 2026: twenty years and eleven months after publication, twenty-seven years after the 2029 claim was first made, three years and four months before it is due.

What it got right

The supercomputer milestone. Due end of 2009. Late by about two years — his best hardware call. IBM's Roadrunner at Los Alamos reached 1.026 petaflop/s on 25 May 2008, the first sustained petaflop on the TOP500 Linpack benchmark. The 10^16 threshold — Kurzweil's own figure for functional brain equivalence — was crossed by Fujitsu's K computer at RIKEN, which topped 10 petaflop/s in November 2011. The machine was named for the Japanese numeral kei, 10 quadrillion, 10^16. Two years late on a four-year-ahead call is a good forecast by any standard this canon applies to anyone.

Compute primacy, against the 2005 consensus. The book's central empirical bet — that the binding constraint on machine intelligence is the substrate and that the substrate is on a schedule — was a minority position in 2005 and is the majority position now. bitter-lesson-2019 is Sutton reaching it fourteen years later from inside the research record; scaling-laws-2020 is it made quantitative; this project's compute-and-infrastructure lens exists because the bet paid. Kurzweil was early, loud and directionally right on the thing that turned out to matter most, and he was right about it while the field's own seniors were still calling it naive. Chapter Nine's "scientist's pessimism" — his name for the failure mode of specialists who "fail to appreciate the ultimate long-term implications of their own work" — was a real and correctly identified error.

The 2029 date, provisionally, and this has to be said before the grades that follow or they are hindsight. In 1999 a claim of human-level machine intelligence by 2029 sat far outside every expert survey. In 2011 Paul Allen and Mark Greaves published the mainstream verdict in MIT Technology Review — "by the end of the century, we believe, we will still be wondering if the singularity is near." As of 2026 the forecasting community's mass has moved onto Kurzweil's ground rather than away from it: Anthropic's formal submission to the US Office of Science and Technology Policy expects powerful AI in late 2026 or early 2027, Musk has said 2026, and Kurzweil now positions himself as the conservative — "Elon Musk says 2026. I think we'll have a lot of things that remind us of AGI, but we really won't be convinced in 2026. Maybe 2027, 2028. By 2029, I think everyone will accept it." Whether it lands is not knowable here. What is recordable is that the direction of surprise since 2020 has run in his favour, and that grading this book from a 2012 vantage — as most of the published gradings do — would now be grading it before its subject arrived.

What it got wrong

1. The personal-computer milestone. Due end of 2019. Six and a half years overdue, and by how much depends on a unit the forecast never pinned — which is itself the finding.

The claim is 10^16 cps in a personal-computer-size device by the end of the 2010s. Take the best consumer accelerator available in August 2026, NVIDIA's RTX 5090:

Per dollar the gap is wider, because the card is not a $1,000 part. Kurzweil's Chapter Three figure, as restated in the secondary literature, is that "by 2020, 10 quadrillion cps will be available for around $1,000." In August 2026 the US median street price of an RTX 5090 was $4,699.99, up from $4,299.99 in June, against a $1,999 MSRP. So $1,000 buys 2.2 × 10^13 FP32 FLOP/s — about 450× short — or 7.1 × 10^14 sparse FP4 ops, about 14× short. Epoch AI's independent figure for datacentre hardware lands in the same place by a different route: an H100 at 2.2 × 10^10 FLOP/s per dollar, which is 2.2 × 10^13 per $1,000. Two unrelated measurements, one number.

The unit convention is not a technicality here; it is the whole mechanism of the self-grade gap, visible on a single claim. Kurzweil derived 10^16 from neuron and synapse counts multiplied by firing rates — a claim about biological operations, not about four-bit tensor throughput. Choose the accounting he derived it from and the forecast missed by roughly two orders of magnitude. Choose the vendor's marketing unit and it missed by a factor of three. The forecast does not specify, so the grader picks, and the grader who is also the forecaster picks the flattering one. That is the 86% in miniature.

2. The mechanism. The double exponential is not there, and this is the substantive failure. "The power […] of information technologies […] now doubling about every year," with the exponent itself growing. Measured: Epoch AI, over 470 GPUs from 2006 to 2021, finds FLOP/s per dollar doubling every ~2.5 years — 2.95 years for top-of-the-line parts, 2.07 for the models typically used in ML research. Over twenty-plus AI accelerators released between 2012 and 2025 the figure is ~30–40% per year, a doubling every 2.2–2.4 years. Slower than Moore's two years, less than half Kurzweil's one, and — the part that kills the third principle — roughly constant. The exponent did not grow.

The compounding is worth stating because it is large. From 2005 to 2026 is twenty-one years: at a one-year doubling that is 2^21, about 2.1 million-fold; at a 2.3-year doubling it is 2^9.1, about 560-fold. The Law of Accelerating Returns overstates 2026 price-performance by a factor of roughly three thousand seven hundred. (My arithmetic on Epoch's measured doubling times, stated as such, not a figure either party publishes.) A forecast can survive being wrong about a date. This is the engine being wrong about its rate, which is why the 2029 claim — if it lands — will have landed despite the mechanism offered for it.

3. In 2026 the curve is running backwards, for reasons the model has no term for. Consumer DRAM contract prices rose as much as 89% in a single 2026 quarter; the spot price of a 16Gb DDR5 chip rose nearly 298% between September and December 2025; IDC estimates AI datacentres absorbing around 70% of global memory output in 2026 against DRAM supply growth of about 16%. The consequence is above: the leading consumer accelerator costs 135% over MSRP and rose $400 between June and August 2026. Price-performance for the buyer went down this year.

Kurzweil pre-answered this objection in Chapter Nine under "The Criticism from Malthus: Exponential Trends Don't Last Forever," and Theodore Modis — whose complexity data Kurzweil used, and who published "The Singularity Myth" in Technological Forecasting & Social Change in 2006 — made it directly: growth that looks exponential is logistic, and every natural growth process reveals its S-curve eventually. Neither of them anticipated the shape it actually took. This is not saturation of physics or exhaustion of a paradigm. It is the AI boom consuming its own input faster than the input can be made, and pricing the consumer out — a demand-side constraint internal to the thing being forecast. A model in which resources flow to a technology because it is getting cheaper, thereby making it cheaper still, has no term for resources flowing to it so hard that it gets more expensive.

4. The route was wrong, even where the date may be right, and this is the finding a reading should carry. Kurzweil's path to human-level intelligence runs through the brain: scanning resolution doubling annually, "within two decades […] a detailed understanding of how all the regions of the human brain work" (due 2025 — not achieved, and not close), then "effective software models of human intelligence by the mid-2020s" derived from that understanding, then the Turing test. His supporting argument, pressed hardest in his 2011 reply to Allen, is that the brain's design cannot be more complex than its specification: "the brain cannot have more design information than the genome," about 50 million bytes compressed, "roughly half of which pertains to the brain" — a million lines of code.

Nobody did any of this. What produced systems that pass short Turing tests was gradient descent on next-token prediction over internet text, at scale, using an architecture published in 2017 — a route with no brain scanning in it, no neuromorphic modelling, and no reference to the genome. The biologist P. Z. Myers' 2010 objection to the genome argument (protein folding, cell-to-cell interaction and development do computational work the genome does not encode) and Kurzweil's reply (it is a complexity bound, not a method) can both be left unresolved, because the outcome routed around the question. Allen and Greaves' "complexity brake" — that understanding the brain gets harder as you learn more — was right about brain science and irrelevant to the outcome, which is a distinct and instructive way for a critique to fail.

Getting the date approximately right by a mechanism that did not happen is a weaker claim than "he predicted it," and it is the claim the record supports. It licenses citing him on when. It does not license citing him on how, and it does not transfer to his other forecasts, which is precisely what the 86% is used to do.

5. The nanotechnology half. Not late — absent. Respirocytes and bloodstream nanobots in the 2020s: no such device exists in or near clinical use. Billions of nanobots in brain capillaries extending intelligence: no. "Nanotechnology-based manufacturing devices in the 2020s […] capable of creating almost any physical product from inexpensive raw materials and information" (p. 13), with three and a half years left in the decade: no. Longevity escape velocity, which the book places within reach of this period, Kurzweil now dates to the early 2030s and in a February 2026 interview framed as roughly seven years out from that date.

The sliding is the gradeable behaviour. The near-term claims move forward as they come due while 2029 and 2045 stay nailed down. That is the opposite of what a coherent model does — if the whole edifice rests on one accelerating curve, slippage in the nano branch should propagate to the dates that depend on it. It does not, which suggests the dates are load-bearing for reasons the curve does not supply.

The self-grade, the independent grades, and how they differ

This kind requires both, and says an entry repeating either alone is worth less than one showing the gap. Here the gap is the largest in the canon.

The self-grade. October 2010, How My Predictions Are Faring, 147 pages. Kurzweil scores 147 predictions, principally the 1999 ones in The Age of Spiritual Machines due by end-2009, and reports 115 "entirely correct", 12 "essentially correct", 17 "partially correct", 3 wrong — combining the first two categories to claim 86%. He has repeated the figure since; third parties now introduce him with it, and it appears in the promotional apparatus around the 2024 sequel. "Essentially correct" is defined by him as realised within a few years — a category that converts lateness into correctness, on a scorecard whose entire purpose is dates.

The independent grades, all published, all lower, in ascending order of generosity:

They differ, and the shape of the difference is what to carry. Three mechanisms produce it, and each is visible in a specific artefact. First, the predictions are stated at a granularity that does not fix a truth condition — Rennie's examples, and section 3's demonstration that whether the 2019 hardware claim missed by 3× or 95× is settled by a unit choice the text never makes. Second, "essentially correct" launders lateness on a scorecard about timeliness. Third, the corpus is his: he chose which 147 statements were predictions.

But the honest counterweight, which the smug version of this entry would omit. Armstrong's own conclusion after grading him twice was not that he is a fraud: "Kurzweil certainly can't claim an accuracy above 50% — a far cry from his own self assessment," and yet "even a true rate of 30% is much higher than chance," leaving "Kurzweil remains an acceptable prognosticator, with very poor self-assessment." Thirty per cent on specific ten-year technology predictions is not a bad score; it is a good one, and most people who mock him have never put a dated list on the record at all.

The two grades answer different questions, and the difference is the base rate. His self-assessment scores was I directionally right about where information technology was going — and on that he is genuinely strong, which is why he believes the number. The independent assessments score did the specific thing happen by the stated date — and on that he is poor. He reports the first as though it were the second. The generalisable lesson for a reading is not "discount Kurzweil." It is that direction and date are separately gradeable, that forecasters conflate them in their own favour, and that any hit rate offered without stating which one it measures is uninterpretable.

One structural limit on all of the above, stated because it bites. Every independent grade listed here scores The Age of Spiritual Machines (1999), not this book. TSIN's own headline claims are 2029 and 2045 and are not due; nobody has published a systematic grading of them. The 2019 hardware grade in section 3 and the price-performance arithmetic are this entry's own work, computed from Epoch AI's published rates and August 2026 retail prices, and should be read as one grader's arithmetic rather than as a citation. A later reading with access to a published TSIN-specific retrospective should expect it to sharpen this section.

And the critics are dated too, so grade them. Allen and Greaves, October 2011: "by the end of the century, we believe, we will still be wondering if the singularity is near." Fifteen years in, that has aged worse than the thing it was refuting. Modis's logistic argument is currently winning on price-performance (section 3, item 3) but by a mechanism he did not name. Kapor's 2002 case against the Turing test rested on embodiment and tacit knowledge — "we are embodied creatures; our physicality grounds us" — and on the claim that machines cannot "weave things together in new ways or to have true imagination"; a text-only system with no body took 73% of judges in 2025. Being right that a forecaster is overconfident is not the same as being right about the world, and this canon's base rate should include the skeptics' misses as well as the enthusiasts'.

Commonly misused as

Not required for prediction. Included because the misuses here are load-bearing in current argument, and because two of them are near-opposites.

Misuse one: "Kurzweil is 86% accurate, so take the next one seriously." The number is a self-score of a corpus he selected, using a category ("essentially correct") that forgives the exact failure mode the corpus is about. Five independent recounts land between 7% and 42%. Note carefully what this entry does not claim: it does not say the true figure is 7%. It says no single figure exists, because "accurate" is not defined until someone fixes the scoring rule, and every published rule gives a different answer. Anyone quoting one number without naming the grader is quoting a marketing claim.

Misuse two: "AI passed the Turing test in 2025, so Kurzweil was right / the bet is over." It did not settle anything he wagered on. Jones and Bergen's result — GPT-4.5 with a persona prompt judged human 73% of the time, published in PNAS in May 2026 and recorded in full at turing-1950 — used five-minute conversations with non-expert participants. The Long Bet Kurzweil signed specifies three judges interviewing four candidates for two hours each, eight hours of interviews, with the machine required both to convince two of three judges and to be ranked more human than at least two of the three human foils. Nobody has run it. It resolves 31 December 2029. The gap between "passed a Turing test" and "passed the Turing test as specified in a signed wager" is where most current confusion about AI capability lives, and this entry exists partly to keep it open.

Misuse three: "Kurzweil predicted the AI boom." He predicted the compute curve's continuation, the primacy of the substrate, and a date. He did not predict the route, and his stated route — brain scanning, neuromorphic modelling, reverse-engineering the neocortex — is not what happened. Nor did he predict the economics: nothing in the book anticipates that the capability would arrive as a metered service sold by a handful of firms whose capital commitments run to trillions. A reading may cite him for the date and the direction. Citing him as having called the thing that happened overstates it, and it is the overstatement that carries his other forecasts along.

Misuse four, the reversal: "Kurzweil is a crank, therefore long-horizon forecasting is worthless." This is the more common error in sophisticated company and this entry does not support it. He put 2029 on the record in 1999 against a field median decades later, and in 2026 the professional forecasting community has moved toward him, not away. Armstrong, having graded him twice and scored him below half, still called him an acceptable prognosticator. Cite this entry against inflated claims of accuracy and against the reflex that dismisses dated forecasts wholesale; it grades both, and a reading using it for only one is using half of it.

Misuse five, the one this project must guard against directly. Nothing in this file is evidence. No number here — not 86%, not 7%, not 10^16, not the 2029 or 2045 dates — deposits into the ledger or moves the needle. A reading may cite this entry to name the argument behind a capex extrapolation, to date and grade a principal's timeline, to supply the direction-versus-date distinction, or to insist on the difference between a Turing test and the Turing test. It may not treat any figure recorded here as a finding. The needle is judgment about what happened since the last reading; this file is about what somebody said would happen, twenty-one years ago, and how that has gone.

Sources

Primary, read directly. Ray Kurzweil, The Singularity Is Near: When Humans Transcend Biology, Viking, 2005, ISBN 0-670-03384-7 — the publisher's own excerpt PDF at singularity.com/BookExcerpts, comprising the full front matter, table of contents, acknowledgements, Prologue and Chapter One. Read as page images: the complete contents listing with chapter and section page numbers (pp. vii–xiv); acknowledgements and the peer-expert-reader list including Moravec, Merkle, Poggio, Gershenfeld and Freitas (pp. xv–xvii); "The Intuitive Linear View Versus the Historical Exponential View" including the twenty-thousand -years passage and the 2014 sub-claim (pp. 10–13); the nanotechnology manufacturing claim (p. 13); "The Singularity Is Near" section with the Vinge and Good epigraphs, the von Neumann attribution and the credit to Moravec's Mind Children and Robot (pp. 21–24); the three Law-of-Accelerating-Returns principles, the brain-scanning claim, the hardware schedule and the Turing-test sentence (p. 25); the nanobot and respirocyte claims (p. 28); the "trillions of trillions" claim (p. 9).

Primary, not read — flagged because it matters. Chapters Two through Nine, including Chapter Three's "Setting a Date for the Singularity" where the 2045 date and the per-dollar computation figures are actually derived, and Chapter Nine's replies to critics. The excerpt PDF stops at the end of Chapter One and the full text is not freely available; the Internet Archive copy is borrow-only. Every quotation in this entry from outside Chapter One — the 2045 sentence, the "$1,000 / 10 quadrillion cps by 2020" figure, the 10^17-by-the-early-2030s figure — is taken from a secondary source that quotes the book, and is marked as such in the text above. A later session that can reach the full text should verify those three and expect the derivations behind them to be worth reporting.

The self-grade.

The independent grades.

The critics and the bet.

The grading data.

Kurzweil's current position.