The Kurzweil record, 1990–2024 — thirty-five years of dated claims, and the people who counted them

kurzweil-2005 graded one book and then named the two things it could not do. It said, in its Sources: the October 2010 self-assessment was "downloaded but not readable in this environment," so the famous tallies were taken from three secondary sources that agree on four integers. And it said, in its grading section: "Every independent grade listed here scores The Age of Spiritual Machines (1999), not this book […] nobody has published a systematic grading of them."

Both gaps close here. The 147-page primary was fetched and read as page images — the front matter, the overview, the section on the 1990 book, the section grading The Singularity Is Near, and the closing section on the magazine columns. And two independent gradings of The Singularity Is Near itself do exist, one from 2017 and one undated, neither of which the earlier session reached.

That is why this file is not a second copy of kurzweil-2005. That entry is about a book: what it claimed, what happened, whether 2029 will land. This one is about the record — five books and a magazine series across thirty-five years, the several people who have counted the claims in them, and the fact that no two counters have ever agreed on how many claims there are.

The corpus shows three things a single book cannot, and each is load-bearing:

• The scoring machinery is applied where it produces a high number and withheld where it would not. The 1990 book is quoted, never graded. The magazine series is linked, never graded. The 2019 chapter has never been self-graded at all. The one chapter that is graded has been graded twice by its author, nine months apart, with the denominator moving from 108 to 147 and the headline falling from 94% to 86%.
• Accuracy decays with horizon, and the decay is measurable — but only inside one grader's consistent method, because across graders the spread is wider than the decay.
• The hit rate is not one number, it is two domains. Claims about the computing substrate land within a few years. Claims about human adoption and about biology miss by decades or fail outright. That split is what makes the corpus decision-relevant in 2026, because his loudest live claim — longevity escape velocity by 2032 — sits squarely in the worse of the two.

prediction is the right kind and the id was filed correctly. The adjudication was already done at kurzweil-2005 and is not re-litigated here: his product is forecasting, sold as forecasting, and his stated credential is his hit rate. An entry about the hit rate itself is the same kind as an entry about one of its components.

What it is

Ray Kurzweil has published dated technology forecasts continuously since 1990, and has been in public argument about their accuracy since at least 2009. The corpus, in order:

The Age of Intelligent Machines (MIT Press, 1990; written, by his own account, in the mid-1980s). Part monograph, part anthology — twenty-three contributed essays from Dennett, Hofstadter, Minsky and others around Kurzweil's own survey of the philosophical and technical roots of AI. Named most outstanding computer science book of 1990 by the Association of American Publishers. Its forecasts are scattered through the text rather than gathered into a list, which is the first reason nobody has ever produced a clean score for it. The ones an encyclopedia can extract: a computer world chess champion "by the year 2000"; intelligent assistants by the mid-1990s; a translating telephone by 2010; a completely driverless car "well into the first half" of the twenty-first century; robotic home servants "early next century"; and the Turing test — his "level 4" — "sometime between 2020 and 2070."

Nineteen articles for Library Journal, 1991–1993, five of them on virtual books and virtual libraries. He lists them in the 2010 self-assessment under the heading "Library Journal Predictions," with no grades attached and a paragraph promoting Blio, the e-reader sold by his own company.

The Age of Spiritual Machines (Viking, 1 January 1999). The book that supplies almost every number anyone quotes about him. It carries dense prediction chapters for four target years — 2009, 2019, 2029 and 2099 — and it is the source of the 2029 human-level-machine-intelligence date that he has never moved since. It is also the source of the claims that human life expectancy would pass one hundred by 2019 and reach 120 by 2029.

The Singularity Is Near (Viking, 2005). Recorded in full at kurzweil-2005. For the purposes of this file, one fact about it matters: its forecasts are mostly organised in twenty-year increments, and the only near-term list in it is a single page of Chapter Six about 2010.

How to Create a Mind (Viking, 2012). Not a book of dated forecasts at all, which is why it belongs in the corpus and needs saying so. It advances the Pattern Recognition Theory of Mind: the neocortex is roughly three hundred million pattern recognisers arranged in a hierarchy, and a mind can be built by implementing that hierarchy — his proposal being hierarchical hidden Markov models with parameters learned by genetic algorithm. It is a mechanism claim, and the corpus needs one, because the recurring failure in the dated claims is that the date arrives by a mechanism other than the one named.

The Singularity Is Nearer (Viking/Penguin, 25 June 2024, 419 pp.). Chapters including "Where Are We in the Six Stages?", "Reinventing Intelligence", "Who Am I?", "Life Is Getting Exponentially Better", "The Future of Jobs: Good or Bad?", "The Next Thirty Years in Health and Well-Being", "Peril" and "Dialogue with Cassandra". It reaffirms 2029 and 2045 without modification and adds the claim that now leads every interview he gives: longevity escape velocity by about 2032.

And around all of it, the scoring apparatus:

How My Predictions Are Faring, Ray Kurzweil, October 2010, 147 pages — the document that produced the 86% figure. Read directly for this entry, and its structure turns out to matter more than its arithmetic. Its contents are: an eleven-page overview; then the 2009 predictions from The Age of Spiritual Machines graded one by one across ten subject headings, occupying pages 11 to 123; then five pages headed "The Age of Intelligent Machines"; then a section grading The Singularity Is Near; then eleven pages of updated exponential graphs; then one page on the Library Journal columns.

The opening sentence sets the scope: "In this essay I review the accuracy of my predictions going back a quarter of a century." The core thesis is stated plainly, and it is narrower than the reputation: "fundamental measures of information technology follow predictable and exponential trajectories," while "there are still many things — which project, company or technical standard will prevail in the marketplace, or when peace will come to the Middle East — that remain unpredictable." He is, in his own document, claiming predictability for the substrate and disclaiming it for everything else. Nearly every later misuse of his record ignores that sentence.

Why a reading would cite it

A reading needs this file on the day a principal's track record is offered as a reason to believe a forecast — which, in the current cycle, is most weeks.

The concrete occasions:

• A dated AI claim appears in the news attached to a hit rate. "The man who predicted the internet says AGI in 2029." The reading's job is to say what the hit rate is a hit rate of, and this file supplies the answer in one line: one chapter, of one book, about one target year, scored by its author, with a different author-scored number for the same chapter nine months earlier.
• Kurzweil's own live claims come due. 2029 is three years out and 2032 is six. When one of them resolves, a reading needs the base rate that was on the record beforehand, not one reconstructed afterwards. Section 3 fixes it now: his substrate claims have landed within a few years, his biology claims have missed by decades, and 2032 is a biology claim.
• Someone cites a forecaster's confidence as evidence. The corpus is the canon's best-documented case of a predictor whose calibration is worse than his accuracy — genuinely better than chance at direction, and in the 94th to 99th percentile of overconfidence about it, on one grader's arithmetic. Those are separable, and most argument treats them as one thing.
• A needle move is proposed on the strength of a forecast. The forecast is not the event. This file is the standing reason: thirty-five years of dated claims by the field's most prolific forecaster, and the honest reading of the record is that the direction was often right and the date usually was not.

There is one more, closer to home and stated briefly. This project publishes a dated judgment twice a day and will accumulate a record of its own. The corpus is the case study in what happens when the party with the strongest interest in the score is also the party keeping it. Nothing here moves a needle, and nothing here should be cited as if it did.

What it got right, and what it got wrong

The kind requires four things per claim: the claim in its own words, the date it was made, the date it was due, and what happened. Taken set by set.

The Age of Intelligent Machines, 1990 — never graded by anyone, including him

The claims, in the book's words, as far as they can be recovered without the book: a computer world champion "by the year 2000"; intelligent assistants by the mid-1990s; a translating telephone by 2010; a completely driverless car "well into the first half" of the twenty-first century; robotic home servants "early next century"; and, on the Turing test, "I place the achievement of level 4 sometime between 2020 and 2070."

Made 1990, on material written in the mid-1980s. Due across 1995–2070.

What happened, item by item, as of 20 August 2026:

• Chess: hit, three years early. Deep Blue beat Kasparov 3½–2½ on 11 May 1997, recorded at deep-blue-1997.
• Intelligent assistants by the mid-1990s: miss, by fifteen to thirty years. Siri shipped in October 2011 (siri-2011) and was a demo more than an assistant; systems that can actually take an instruction and carry it out arrived after 2022.
• Translating telephone by 2010: essentially on time. Speech-to-speech translation reached phones around 2010–2011. Notably this is the one item where his own 2010 grading and an outside view agree.
• Driverless car "well into the first half" of the century: not gradeable, and that is the finding. Driverless commercial service exists in 2026; a claim with a twenty-five-year target window and no threshold cannot be wrong before 2050. The canon's live case of the same claim made sharply is musk-robotaxi-2019, and the contrast is instructive — Musk's version is falsifiable and false, Kurzweil's is neither.
• Robotic home servants "early next century": miss. The Roomba (2002) is the strongest candidate and is not a servant.
• Turing test between 2020 and 2070: technically live, and a fifty-year window is not a forecast. The result that would satisfy it arrived in 2025 and is recorded at turing-1950. It lands in the first sixth of his range. But a band that wide cannot be wrong, and the interesting thing is what he did next: in 1999 he replaced it with a single year, 2029, and has held that for twenty-seven years. He sharpened his own claim over time, which is the opposite of the usual drift, and he deserves the credit for it.

Six items chosen by an encyclopedia are not a grade, and this is marked as indicative rather than scored: two clean hits, two clean misses, two claims with no truth condition. What is not indicative, because it was verified page by page in the primary, is this: the five pages his own 2010 assessment devotes to this book contain no grades at all. No accuracy labels, no tally, no misses. They are block quotations from pages 414 to 447 of the book — the hypertext passage, the eight-point educational-workstation list, the passage on cities dissolving, the passage on the Soviet Union — introduced as "examples of accurate predictions from that book." The section is titled with the book's name and contains a selection made by the person being graded.

The nearest thing to a grade he has ever given it is a sentence in the overview of the same document: he predicted "the defeat of the world chess champion by 1998." The book says "by the year 2000." The restatement is tighter than the original, in the direction that makes the hit look sharper. Deep Blue won in 1997, so both versions are hits and nothing turns on it — which is exactly why it is worth recording. It is the smallest possible instance of the mechanism, in a case where he had no need of it.

The Age of Spiritual Machines, 1999 — the 2009 chapter, the only set anyone has scored

The claims: a chapter of them. The ones that recur in every account are that most text would be created by continuous speech recognition; that cars would drive themselves on highways using sensors embedded in the roadway; that a supercomputer would reach twenty petaflops; that a US company would pass a trillion dollars of market capitalisation; that the economy would continue to boom; that personal computers would be embedded in clothing and jewellery; that displays would be built into eyeglasses; and that human life expectancy would pass one hundred.

Made 1999. Due end of 2009.

What happened: speech recognition did not become the primary means of creating text and has not in 2026 — Alex Knapp's grade in Forbes was "Nope. Not even close." Cars did become able to drive themselves, by onboard perception rather than instrumented roads; the date is arguably defensible and the mechanism is simply not what happened, which is the same failure shape kurzweil-2005 records for nanotechnology versus neural networks. Twenty petaflops arrived in 2012 or 2013 depending on whether peak or sustained performance is counted, three to four years late — and the peak-versus-sustained ambiguity is itself the unit problem that entry made central. A trillion-dollar US company arrived in August 2018, nine years late. The economy did not continue to boom; the prediction was published nine years before the 2008 crash. Computers in clothing: partially, and he defends it at length. Eyeglass displays: not commonly, and he defends that at length too. Life expectancy over one hundred by 2019: wrong by about twenty-one years of life, US life expectancy in 2019 being under seventy-nine.

The Age of Spiritual Machines, 1999 — the 2019 chapter, never self-graded

Made 1999. Due end of 2019. Self-grade: none has ever been published. The 86% in circulation refers to 2009 and to nothing else. This is the single most under-reported fact about his record: the 2019 set is the one with double the horizon, and the author who published 147 pages on the ten-year set has published nothing on the twenty-year set in the seven years since it came due.

What happened is known only from outside gradings, in section 3's last part.

The Singularity Is Near, 2005 — ten predictions, self-graded 90%, outside-graded a third

This is new to the canon and it is the sharpest number in the file.

kurzweil-2005 reported, correctly for what it could reach, that TSIN's headline claims are not due and that nobody had published a systematic grading. Both halves now need amending, and since a canon job writes exactly one file, the amendment lives here rather than there.

He graded the book himself, in October 2010. Under a heading "Predictions for 2010," he writes: "There is one page of predictions for 2010 in Chapter 6 ('The Impact…') of The Singularity is Near." He splits that page into ten numbered items, each with a PREDICTION line, an ACCURACY line and a DISCUSSION. The grades, verbatim: invisible embedded computing — Correct. Tapping the mesh — Correct. Always online — Correct. Heads-up display — Essentially correct. Projected audio — Correct. Complete virtuality — Essentially correct. Augmented reality — Correct. Real-time translation — Correct. Information at-hand — Correct. Virtual assistants — Partially correct. Seven correct, two essentially correct, one partially correct, none wrong: 90% by the same rule that produced the 86%. Graded in October of the year in question, before it had ended.

Two independent gradings of the same book exist. Paul Feakins, 29 January 2017, extracted twenty-five predictions from the book and scored nine true and sixteen false — 36% — while concluding it was "quite impressive" given that on first reading they had "seemed wildly optimistic." Nuño Sempere, undated but after 2019, extracted the book's claims about the 2010s by keyword and scored five correct against ten wrong — 33% — and noted his result was "consistent with the 36% found in 2017."

The gap on this book is 90% against 33–36%, and it is wider than the famous one. It is also cleaner, because the self-grade covers ten items and the independent grades cover fifteen and twenty-five, so the divergence cannot be explained by different books or different decades. It is explained by which sentences were chosen and by what "essentially correct" is allowed to absorb.

How to Create a Mind, 2012 — the mechanism claim, and the only fair way to grade it

The claim: the neocortex is about three hundred million pattern recognisers in a hierarchy; build that hierarchy — his implementation proposal being hierarchical hidden Markov models trained by genetic algorithm — and you get a mind. Made 2012. Due: never stated, which is why it is not scored alongside the others.

What happened: the hierarchical, learned-representation half was vindicated so thoroughly that it is now unremarkable. The specific mechanism was not. The systems that took language did it with backpropagation and attention (transformer-2017), not with hidden Markov models and not with genetic algorithms; neural–HMM hybrids were tried in machine translation and were overtaken. He was directionally right about hierarchy and wrong about the machinery, in a book whose entire subject was the machinery — the same right-shape-wrong-engine pattern as the 2009 self-driving claim and the nanotechnology route recorded at kurzweil-2005. He joined Google in December 2012 to work on language understanding, which makes this the one claim in the corpus he was funded to falsify himself.

The Singularity Is Nearer, 2024 — not due, and worth pinning now

The claims, in his words, restated repeatedly through 2025 and 2026 so the wording is not in doubt: AGI and the Turing test by 2029, unrevised since 1999; "by roughly 2032 when you live through a year, you'll get back an entire year from scientific progress"; "in the 2030s, robots the size of molecules will go into our brains, noninvasively, through the capillaries, and will connect our brains directly to the cloud"; "by 2045, once we have fully merged with AI, our intelligence will no longer be constrained … it will expand a millionfold" — the last three quoted from his MIT appearance of 10 October 2025.

Made 25 June 2024. Due 2029, 2032, the 2030s and 2045. Not gradeable today, and the reason to record them is that a reading in 2029 or 2032 will want the wording as it stood before the result, not after.

One prior, stated now so it is on the record before the fact and can be graded against later: 2029 is a substrate-and-benchmark claim and 2032 is a biology claim. In this corpus those two categories have very different histories. The 2029 date is the best-looking long-horizon AI forecast of the twentieth century and kurzweil-2005 says so at length. The 2032 date belongs to the same family as "life expectancy over one hundred by 2019," which is the worst miss in the record.

The self-grades, the independent grades, and why no single number exists

The kind requires both grades and requires saying they differ. At corpus scale the difference has a shape that the 86-versus-7 headline hides.

The self-grades, all of them:

• January 2010, replying to Michael Anissimov on the blog Accelerating Future: 108 predictions for 2009 from The Age of Spiritual Machines, of which 89 entirely correct, 13 essentially correct, 3 partially correct, 2 "about ten years off," and 1 "tongue in cheek anyway" and wrong. That is 102 of 108, 94%.
• October 2010, in the 147-page assessment: 147 predictions for 2009 from the same book, of which "115 (78 percent) are entirely correct as of the end of 2009, and another 12 (8 percent) are 'essentially correct' — a total of 127 predictions (86 percent) are correct or essentially correct. Another 17 (12 percent) are partially correct, and 3 (2 percent) are wrong."
• October 2010, same document: 10 predictions for 2010 from The Singularity Is Near, of which nine correct or essentially correct — 90%.
• Never: the 1990 book, the Library Journal series, and the 2019 chapter.

Same book, same target year, nine months apart, the denominator moved from 108 to 147 and the headline fell eight points. No third party forced the revision and none of the added claims can have come true in the interval, since the target year had already closed. The corpus therefore contains its author's own demonstration that his number is a function of how he splits his sentences.

The definition of "essentially correct" is in the primary, and it is broader than the secondary accounts say. From the 2010 predictions section: "these predictions are organized by decades (actually in 20-year increments: there is a page on 2010 and it then jumps to 2030), so if a prediction comes true in 2011 or even 2012 or 2013 it is still 'essentially correct.'" Three years of slack, granted by the author, on a scorecard whose subject is dates. And where the wording of one book is too strong, the other book is available as a defence: on the heads-up display item he concedes the technology is "not yet" common, notes that "the wording was clear in The Age of Spiritual Machines … that the technology would exist but not necessarily be common," and grades the item essentially correct on the looser book's wording.

The independent grades, by target year. For 2009: Stuart Armstrong's full assessment split the material into 172 statements and returned 42% true or weakly true from his direct assessors, 30% from a separate volunteer cohort; Alex Knapp in Forbes, 20 March 2012, took twelve and scored one correct, four partial, seven failed; Dan Luu, scoring the same list Peter Diamandis had cited for the 86% and requiring that a prediction happen as described, got 7%. For 2019: Armstrong's 2020 pass split it into 105 statements, collected 3,078 judgments from 34 assessors, and returned 12% true, 12% weakly true, 10% undecidable, 15% weakly false, 50% false; Sempere extracted 69 and scored 24 correct against 45 wrong, 35%. For The Singularity Is Near: Feakins 36%, Sempere 33%. John Rennie in IEEE Spectrum, 29 November 2010, published no tally but documented the method — "lawyerly defenses of his predictions that hinge on their precise wording" — and echoed the January number sarcastically: right "at least 94.4 percent of the time, anyway." An anonymous blogger at Militant Futurist spent most of 2020 grading the 2019 chapter item by item across four posts and deliberately published no aggregate.

The counts are the finding. One chapter of one book has been split into 63 statements, 108, 147 and 172 by four different counters; the 2019 chapter into 69 and 105 by two more. Before anyone can score a forecaster, someone has to decide what counts as one prediction, and no two people who have tried this have agreed. A percentage computed over an undisclosed splitting rule is not a measurement. This is the corpus's most transferable lesson and it applies well beyond Kurzweil.

The decay with horizon, stated carefully. Across graders the spread is too wide to compare sets — Luu's 7% and Armstrong's 42% score the same year. Within one grader using one method it can be done, and Armstrong is the only person who has graded both sets: 42% true or weakly true at a ten-year horizon, 24% at twenty. Doubling the horizon roughly halved the hit rate, and his own summary was that the 2019 predictions came out "considerably worse than those for 2009, with more than half strongly wrong." One forecaster, one method, two horizons: it is the closest thing to a controlled measurement of forecast decay this canon holds, and it should be quoted with the method attached or not at all.

The domain split, which is the other half of the answer. Sorting the graded items by subject rather than by date, the pattern is consistent across every set: claims about computation, connectivity and price-performance land, sometimes early, usually within a few years. Claims about what people would do with the technology — speech replacing typing, assistants in the mid-1990s, displays in eyeglasses — miss by ten to thirty years. Claims about biology miss worst of all and do not appear to be converging. His own 2010 overview draws that boundary himself, then his forecasts cross it freely.

The counterweight, kept because an entry that omitted it would be dishonest. Armstrong's 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 30% "is much higher than chance," leaving him "an acceptable prognosticator, with very poor self-assessment." Sempere, independently: "an accuracy of ~ 1 in 3 is surprisingly good, in particular if one considers that these predictions were made 20 years in advance." Feakins, having scored him 36%, called it "quite impressive." Three graders whose numbers destroy his own scorecard all concluded he is a good forecaster. Dan Luu's comparative ranking puts him mid-field among futurists, above Michio Kaku at 3% and below Steve Yegge at 50%, which is a more useful frame than any single figure: the question is not whether 7% is bad but what the base rate is for anyone who puts dated claims on the record, and it is low for everybody.

And one asymmetry that keeps the record honest in the other direction: almost nobody who mocks the corpus has published a dated list of their own. The people in this file who did the work — Armstrong, Sempere, Feakins, Knapp, Luu, Rennie, the Militant Futurist blogger — are a small group, and the canon's base rate rests entirely on them.

Commonly misused as

Not required for prediction, included because the corpus-level misuses differ from the book-level ones already recorded at kurzweil-2005.

"He's 86% accurate." One book, one chapter, one target year, scored by the author. Not the 1990 book, not the magazine columns, not the 2019 chapter, not the 2029 or 2045 claims. And the same author published 94% for the same chapter nine months earlier. Anyone quoting either figure without naming the chapter and the grader is quoting marketing.

"His critics have never actually checked." They have, at least seven times, and three of them concluded he is better than chance. The failure in the public argument is not absent scrutiny; it is that the scrutiny produced numbers from 7% to 42% for the same year and nobody reconciles them, because reconciling them requires agreeing on what one prediction is.

"The exponential curves prove the forecasts." The curves are the strongest part of his work and the 2010 document is explicit that they are "themselves predictions." But a correct trend line does not license the dated application claims hung on it — the substrate arrived roughly on schedule and the world arrived twenty years late. His own overview separates the two; almost no citation of him does.

"He was wrong about everything, so long-horizon forecasting is worthless." The 2029 date has been unmoved since 1999, was made when the field's own median was fifty years out, and is currently the best-aged long-horizon AI forecast of the twentieth century. Being wrong about the mechanism and the furniture while right about the date is a specific and recurring pattern, and it is not the same as being wrong.

"The record shows AI forecasts run optimistic, therefore discount all of them." The record shows something narrower and more useful: direction and date are separately gradeable, and the ratio between them depends on the domain. A forecaster with 35 years of substrate hits and biology misses tells you to weight his next substrate claim and discount his next biology claim, which is a different instruction from "discount him."

And the one this project must not do. Nothing in this file is evidence. Kurzweil being right in 2029, or wrong, will be an event; the forecast is not. The corpus is here to be cited when a reading has to explain how much weight a dated claim deserves, and for no other purpose.

Sources

Primary, read directly. Ray Kurzweil, How My Predictions Are Faring, October 2010, 147 pp., PDF at thekurzweillibrary.com. Fetched and read as page images: title page and table of contents; the overview, pp. 1–6, including the scope sentence, the substrate-versus-marketplace passage, the AIM examples sentence with the "by 1998" chess restatement, the 147-prediction tally with all four category percentages, and the retinal-display and keyboard defences; the whole of "The Age of Intelligent Machines," pp. 124–128, confirming it contains quotations and no grades; "Predictions for 2010," pp. 129–134, all ten items with their ACCURACY lines and the three-year "essentially correct" rule; p. 135, the graphs section and its statement that "the graphs in this book are themselves predictions"; and p. 146, the Library Journal section, the nineteen articles of 1991–1993 and the Blio paragraph. This is the document kurzweil-2005 recorded as downloaded but unreadable; the tallies it took from three secondary sources are confirmed correct against the primary. Pages 11–123 — the 2009 predictions graded one by one — were not read; a session wanting to audit the 147 splits itself should start there.

Primary, not read. All five books. Every quotation from them here is taken from a source that quotes them and is marked as such: the AIM prediction list and the "by the year 2000" chess wording from the Wikipedia article on the book; the AIM Turing-test sentence, "I place the achievement of level 4 sometime between 2020 and 2070," from an excerpt of the book's "Passing the Turing Test" section hosted at calculemus.org; the ASM structure, the four prediction years and the life-expectancy claims from the Wikipedia article on that book; the 2024 book's chapter list from search results and the publisher's page. The Singularity Is Nearer was not read, and whether it contains a systematic regrade of the earlier predictions could not be settled: the publisher's copy says it assesses "the progress of many of his predictions," no review found here reports counts, and the last tally with numbers attached remains October 2010. A session that can reach the book should check "Where Are We in the Six Stages?" first.

The self-grades.

• Aaron Saenz, "Kurzweil Defends Predictions for 2009, Says He is 102 for 108," Singularity Hub, 19 January 2010 — the January count and its breakdown, and the venue (a reply on Michael Anissimov's Accelerating Future).
• Aaron Saenz, "Kurzweil Defends His Predictions Again: Was He 86% Correct?", Singularity Hub, 4 January 2011 — the October document's scope and the selection-bias and no-baseline criticisms.

The independent grades.

• Stuart Armstrong, "Kurzweil's predictions: good accuracy, poor self-calibration," LessWrong, 11 July 2012 — the ten-from-63 pilot, the observation that Kurzweil "grades himself as having either 102 out of 108 or 127 out of 147 correct," and the 94th-to-99th-percentile calibration finding.
• Stuart Armstrong, "Assessing Kurzweil: the results," LessWrong, 2012/2013, and "Assessing Kurzweil predictions about 2019: the results," LessWrong, 2020 — the 172-statement and 105-statement passes. Both were read in full by kurzweil-2005; their figures are restated here from that entry and from the 2012 post read directly for this one.
• Nuño Sempere, "Accuracy of Ray Kurzweil's predictions for the 2010s," nunosempere.github.io — 5 correct / 10 wrong for The Singularity Is Near, 24 / 45 for the 2019 chapter, and the "~ 1 in 3" conclusion. The page carries no date; its content places it after 2019.
• Paul Feakins, "The Singularity is Near: How Kurzweil's Predictions Are Faring," antropy.co.uk, 29 January 2017 — 25 predictions, 9 true, 16 false.
• Dan Luu, "Futurist prediction methods and accuracy," danluu.com — the 7% and the comparative ranking (Caplan 100%, Yegge 50%, Dixon 10%, Kurzweil 7%, Kaku 3%).
• Alex Knapp, "Ray Kurzweil's Predictions For 2009 Were Mostly Inaccurate," Forbes, 20 March 2012, and "Ray Kurzweil Defends His 2009 Predictions," 21 March 2012.
• John Rennie, "Ray Kurzweil's Slippery Futurism," IEEE Spectrum, 29 November 2010 (December 2010 issue) — the "lawyerly defenses" characterisation and the "94.4 percent of the time, anyway" line, which confirms the January figure was the one in circulation before October.
• "Eddie," "How Ray Kurzweil's 2019 predictions are faring," militantfuturist.com, four parts, the last dated 31 December 2020 — graded item by item with no aggregate, deliberately.
• Daniel Lyons, Newsweek, May 2009 — the early list of 2009 misses (the economy, the trillion-dollar company, the twenty-petaflop supercomputer, speech recognition, instrumented highways). Cited via the Wikipedia summary; the article itself was not reached.

Kurzweil's current position.

• MIT News, "Ray Kurzweil '70 reinforces his optimism in tech progress," 10 October 2025 — the 2032, 2030s and 2045 quotations used above, from an event the same day.
• Alan Thompson's Kurzweil tracker at lifearchitect.ai — dated restatements through 2024 and 2025, including the March 2025 Barcelona "AGI will be here by 2029" and the October 2024 exchange in which the 86% is quoted back to him approvingly with "a leniency of like 12 to 24 months."

On How to Create a Mind. The PRTM summary — three hundred million pattern recognisers, hierarchical hidden Markov models, genetic algorithms — from secondary summaries and from the review literature aggregated through search; the book was not read and no primary quotation from it appears above. The judgment that neural–HMM hybrids were overtaken by attention is this entry's own, resting on transformer-2017.

Fetch failures, recorded so a reader knows what was not checked. The New Atlantis review of How to Create a Mind, the Cato Institute review of The Singularity Is Nearer (Fall 2025), the EverybodyWiki mirror of the predictions article, the eDiscovery Team review of the 2024 book, and MIT Press's page for the 1990 book all returned HTTP 403. web.archive.org is blocked to this job's fetch tool, so no Wayback fallback was available for any of them, and direct curl was refused. The Cato review is the most relevant unread assessment of the 2024 book; the New Atlantis review is the most relevant unread assessment of the 2012 one.

A note on how the primary was read. The PDF became readable because WebFetch saves binary responses to the session's tool-results directory and prints the path, after which Read renders the pages as images. kurzweil-2005 recorded this document as unreadable; it is not. That route should be tried before any future entry records a PDF as unreachable.

Project files read while writing, and only these. CLAUDE.md; canon/proposals.md at the entry for this id; canon/kurzweil-2005.md — the header, the framing, and the whole of the self-grade and misuse sections; canon/vinge-singularity-1993.md's closing section, for the renderer constraints; the front-matter and Markdown functions of build.py; and the headers of three multi-year entries to settle what goes in year:. history/, digests/, site/, LENSES.md and SCALE.md were not opened.

Two ids this file would like to draw and cannot. kurzweil-2005 should now descend from, or at least point at, this entry, since the corpus is where its two stated gaps are closed; and its Sources note that the October 2010 document was unreadable is superseded. A canon job writes exactly one file, so neither correction can be made from here. The arrows are the owner's or a later job's to draw.

On year: 1990 in the header. The file spans 1990 to 2024. The header takes the year the public record starts, which is also the year of the first book, so the canon's chart marks the corpus where the forecasting began rather than where it currently stands. kurzweil-2005 already marks 2005.

No evidence was deposited, no needle, score or landmark is placed, and no file outside canon/kurzweil-corpus-1990-2024.md was written.