"The Coming Technological Singularity: How to Survive in the Post-Human Era"

Three files in canon/ asked for this one by name. kurzweil-2005 recorded that The Singularity Is Near opens with a Vinge epigraph and then said flatly: "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." von-neumann-singularity-1958 closed one quarter of that gap and left instructions: "Vinge's thirty years ran out in 2023 (vinge-singularity-1993, proposed, unwritten — it is the gradeable one and someone should write it)." good-intelligence-explosion-1965 closed a second quarter and used Vinge four times as the witness who quotes Good correctly. This file closes the third.

proposals.md filed it as prediction and filed it right, which is worth one paragraph because the rival kind has a real case. Two-thirds of the essay is not forecast: it is an argument about opacity — that past a certain point our models stop working — and an argument of that shape is what this canon files as interpretation, which is how superintelligence-2014 was handled when Bostrom was refused prediction on the ground that a philosopher's product is argument. The refusal does not transfer, for the same reason it did not transfer to Kurzweil. The abstract is two sentences long, and both of them are dated claims. One of them came due in 2023. The essay is cited, quoted and reprinted for those two sentences and almost never for its four-part taxonomy of routes, and the author restated the date under his own name fifteen years later in IEEE Spectrum and said he would stand by it. A forecast that its own author re-signs is a forecast.

A note on the quotations before any of them are used. The primary text is freely readable at the author's university page and explicitly permits verbatim copying. The fetch tool available to this job re-renders a page through a summarising model that declined to reproduce the essay and capped what it returned to short excerpts. Every quotation below therefore reached this file at one remove, through that tool, in fragments obtained across five separate requests against two independent mirrors — the author's own page at SDSU and Hans Moravec's copy at Carnegie Mellon. Where the two mirrors returned the same sentence it is marked as confirmed twice. A session with direct access should expect to sharpen the punctuation and the ellipses, and should not expect the substance to move.

1. What it is

The document

Vernor Vinge, "The Coming Technological Singularity: How to Survive in the Post-Human Era," presented at the VISION-21 Symposium, sponsored by the NASA Lewis Research Center and the Ohio Aerospace Institute, 30–31 March 1993, in Westlake, Ohio. The author's page carries the note that a slightly changed version appeared in the Winter 1993 issue of Whole Earth Review, and the permission line reads: "Verbatim copying/translation and distribution of this entire article is permitted in any medium, provided this notice is preserved."

Vinge was a mathematician and computer scientist at San Diego State University and a science-fiction novelist; A Fire Upon the Deep had been published the previous year and would take that year's Hugo six months after the symposium. Both halves matter to what the essay is, because it is a technical paper written in the register of a manifesto, delivered to an aerospace audience, by a man whose day job was extrapolating futures for a living.

The claim

The abstract is the whole of the famous part, and it is four sentences:

"Within thirty years, we will have the technological means to create superhuman intelligence. Shortly after, the human era will be ended. Is such progress avoidable? If not to be avoided, can events be guided so that we may survive?"

Then, in the body, the sentence that makes it scoreable at all — a window rather than a point:

"I'll be surprised if this event occurs before 2005 or after 2030."

Read the first sentence slowly, because the grading in section 3 turns on it. It does not say superhuman intelligence will exist by 2023. It says the technological means to create it will exist. That is a claim about capability in principle, and it is the sort of claim that cannot be observed until somebody exercises it. The second sentence is the falsifiable one, and it has no mechanism, no threshold and no definition of "shortly."

The four routes

The essay's real contribution — the part that is genuinely original to Vinge and is almost never quoted — is that it does not have one mechanism. It has four, listed as bullets (quoted from the Carnegie Mellon mirror; the SDSU request that covered them returned paraphrases of the same four, in the same order):

• "Computers that are 'awake' and superhumanly intelligent may be developed."
• "Large computer networks and their associated users may 'wake up' as superhumanly intelligent entities."
• "Computer/human interfaces may become so intimate that users may reasonably be considered superhumanly intelligent."
• "Biological science may provide means to improve natural human intellect."

Four named, independently checkable roads. Section 3 checks all four. This is what makes the entry worth its length: unlike the 1958 sentence, and unlike most of what has been written in the genre since, this forecast committed to specifics that could go wrong separately.

The argument

Why it is called a singularity. Vinge's answer is epistemic, not arithmetic: "It is a point where our models must be discarded and a new reality rules." And the driver is recursion: "When greater-than-human intelligence drives progress, that progress will be much more rapid." The consequence, in his most-quoted sentence after the abstract: "From the human point of view this change will be a throwing away of all the previous rules, perhaps in the blink of an eye, an exponential runaway beyond any hope of control."

The hardware premise. The essay does not assert the hardware; it borrows it. It reports a 1992 workshop at Thinking Machines Corporation where a minority thought the largest 1992 machines were within three orders of magnitude of a human brain, and states: "The majority of the participants agreed with Moravec's estimate [16] that we are ten to forty years away from hardware parity." Reference [16] is Mind Children — the book graded in moravecs-paradox-1988, whose own figure is about 100 million MIPS. Ten to forty years from 1993 is 2003 to 2033. Vinge's own 2005–2030 window is Moravec's window with the tails trimmed. The date is borrowed, and the entry it is borrowed from is already in this canon.

Inevitability. The section titled "Can the Singularity be Avoided?" answers no, and the reason is not technological but competitive: "But if the technological Singularity can happen, it will." The mechanism, and this is the sentence most worth carrying out of the essay in 2026: "the competitive advantage — economic, military, even artistic — of every advance in automation is so compelling that passing laws, or having customs, that forbid such things merely assures that someone else will get them first."

The alternative, offered honestly. He also writes down what the world looks like if he is wrong and the hardware curves flatten: "We'd end up with some very powerful hardware, but without the ability to push it further… It would likely be seen as a golden age … and it would also be an end of progress."

Consequences. He does not soften them. "The physical extinction of the human race is one possibility" — and he then argues it is not the worst one, sketching humans persisting as embedded systems, "self-aware daemons in the lower functioning of larger sentients," some used for nothing more than signal processing.

Intelligence Amplification. The essay's second half is a bet on a different road, and it is the bet section 3 grades hardest: "It's very likely that IA is a much easier road to the achievement of superhumanity than pure AI." His reason: "In humans, the hardest development problems have already been solved. Building up from within ourselves ought to be easier than figuring out what we really are." He then does something forecasters almost never do — he lists concrete experiments, including automating human/computer teams on hillclimbing problems, human/computer symbiosis in the visual arts, mobile rather than desk-bound access, symmetric decision-support in which the user guides the program as much as the program informs the user, groupware treating a local network as "a combination organism," and two that are checkable to the letter: allow human/computer teams at chess tournaments, and take the network seriously — "The power and influence of the Internet are vastly underestimated." On direct interfaces he is modest and specific: "Nerve-to-silicon transducers can be made," and even "100 bits per second would be of great use to stroke victims."

The word, and where he got it

Vinge quotes the Ulam memorial passage — the subject of von-neumann-singularity-1958 — and then, in the same breath, tells his readers it does not say what he is about to use it for: "Von Neumann… is still thinking of normal progress, not the creation of superhuman intellect." He quotes I. J. Good's 1965 intelligence-explosion paragraph in full, including the docility clause that later authors drop, and adds: "Good has captured the essence of the runaway, but does not pursue its most disturbing consequences."

So the weld is visible and the welder documented it. good-intelligence-explosion-1965 established that Good's argument does not descend from Ulam's paragraph — Good's three supporting citations are all Good citing Good — and concluded that the pedigree "von Neumann → Ulam → Good → Vinge → Kurzweil" is "broken at its second link." This entry adds the missing half of that finding: Vinge is where the two independent inventions were joined, he joined them deliberately, and he wrote down in the same paragraph that the older one did not mean what he was borrowing it for. Everything downstream inherited the join and dropped the disclaimer.

The word in its modern sense is his own and predates this essay by a decade. The Historical Dictionary of Science Fiction gives as its earliest cite Vinge's "First Word" column in Omni, January 1983: "When this happens, human history will have reached a kind of singularity, an intellectual transition as impenetrable as the knotted space-time at the center of a black hole." The apparent conflict with von-neumann-singularity-1958, which calls Ulam's 1958 sentence the earliest known use, is not a conflict once the senses are kept apart: 1958 is the earliest use of the word for technological acceleration; 1983 is the earliest use of it for machines exceeding us. Both entries are right and neither should be cited for the other's claim.

2. Why a reading would cite it

Four occasions, two of them taken from the reading published at midnight.

1. When a principal announces that the singularity has begun. The 20 August 2026 reading recorded Stripe's investor letter framing 1 January 2026 as "the beginning of the singularity," and declined to deposit it — correctly, under the LENSES.md rule that a principal's forecast is a canon entry and never a deposit. Sam Altman's June 2025 "The Gentle Singularity" opens "We are past the event horizon; the takeoff has started." Both use the word as though it named an observable. This entry supplies the definition its author actually gave: a point where our models must be discarded. A company that believes it is inside the singularity and simultaneously publishes a five-year roadmap and a revenue run rate is asserting the opposite of what the word means. That is not a rebuttal of anything either firm is building; it is the request that the speaker say what they mean in terms that could turn out false.

2. When "the competitive advantage is too great to regulate" is offered as an argument. It is offered constantly — against pause proposals, against export controls, against retention rules, against every instrument this project's policy lens records. The strongest short statement of it is thirty-three years old, is quoted in section 1, and belongs to a man who wanted the danger taken seriously. Set it beside the 1955 von Neumann passage recorded in von-neumann-singularity-1958 (a ban needs the enforcement capacity whose absence made the ban necessary) and beside asilomar-1975 (the case where a moratorium actually held). Three dated statements of the same trade-off, one of which is a counterexample. A reading weighing a voluntary frontier pause has a better argument to answer than the one usually made.

3. When recursive self-improvement is claimed or denied. The 20 August 2026 reading recorded arXiv 2607.27191 — frontier agents given the central research questions of two unpublished NeurIPS 2026 submissions, six days and thousands of dollars of compute, output graded by the papers' own authors, both rejected, one a strong reject, the failures in research judgment rather than engineering. That is a direct measurement against Vinge's core mechanism, which is Good's mechanism: a machine better than humans at designing machines. This entry is where a reading can say what the mechanism claims, whose it is, and that the best current published negative result is aimed at exactly it.

4. As the base rate this kind exists to build. It is the rare AI forecast that is (a) thirty-three years old, (b) dated twice over, with a headline year and an outer bound, (c) decomposed into four named routes that fail separately, and (d) accompanied fifteen years later by the author's own list of falsifiable indicators. With kurzweil-2005 it is one of the two best-specified long-horizon AI forecasts in this canon, and unlike Kurzweil's its headline date has already come due. Grading it is not a gotcha; it is the only way to learn anything from it.

3. What it got right, and what it got wrong

The four things this kind requires, first and plainly.

The claim, in its own words. "Within thirty years, we will have the technological means to create superhuman intelligence. Shortly after, the human era will be ended." Plus the window: "I'll be surprised if this event occurs before 2005 or after 2030." Plus four named routes and, from 2008, six named indicators.

The date it was made. 30–31 March 1993, at VISION-21; the word in this sense is his from January 1983, but the dated forecast is 1993.

The dates it was due. Two clocks, and conflating them is the single commonest error made about this text. The headline clock ran out on 30 March 2023. The author's own outer bound is 2030, restated by him in June 2008. Today is 20 August 2026: three years and nearly five months past the first, about four years and four months short of the second.

What actually happened. Below, route by route, then the indicators he supplied himself.

What it got right

The hardware premise landed, roughly on his borrowed schedule. Moravec's figure, as recorded in moravecs-paradox-1988, is about 100 million MIPS — on the order of 10^14 operations per second — for a human brain, in a window of 2003 to 2033. kurzweil-2005 measured a single consumer accelerator in August 2026 at 1.05 × 10^14 FP32 FLOP/s, and a frontier training cluster is many orders above that. On the unit the estimate was stated in, hardware parity arrived inside the window Vinge quoted, and it arrived without superhuman general intelligence following it. Two cautions, both real: MIPS and FLOP/s are not the same unit, and kurzweil-2005 demonstrated at length that the choice of unit can move a verdict by two orders of magnitude, so nothing finer than "the same order, inside the window" is claimed here; and this comparison is assembled from two canon files, not from a fresh measurement. The direction is not in doubt. The machine that Moravec said would be enough exists, and it was not enough.

The competitive-advantage thesis is the best-aged sentence in the essay. Every mechanism this project's concentration lens records — a $60bn all-stock acquisition of a coding-agent company, a reported second approach within a week, two frontier labs racing toward public listings, roughly $3tn of datacentre capex forecast to 2030 — is the sentence operating. No law and no custom has slowed a frontier capability anywhere, and the argument for why not was written in 1993 in one sentence.

The Internet claim, which reads as banal now and was not in March 1993. NCSA Mosaic shipped about a month after the symposium and the web was a few dozen to a few hundred sites depending on the month; the count is not the point and this entry does not pin it. "The power and influence of the Internet are vastly underestimated" was, at that date, a minority position stated plainly, and it is simply correct — including the specific form he gave it, that the anarchy of the network's development was evidence of its potential rather than a defect in it.

The refusal to make it a curve. Unlike every popularisation that followed, Vinge did not draw an exponential and read a year off it. He used "singularity" in its epistemic sense — models fail, extrapolation stops working — which is the sense von-neumann-singularity-1958 established the term actually carries in mathematics. The man who put the word into general circulation is the one who used it correctly, and everything that has gone wrong with the word since has gone wrong downstream of him.

What it got wrong

1. The headline sentence, at its own deadline. Due 30 March 2023. The human era did not end, and has not. Three years and nearly five months later there is no superhuman general intelligence, no entity that has taken over its own design, and no discontinuity of the kind the essay describes. The deadline did not even pass in a quiet period: GPT-4 shipped on 14 March 2023, sixteen days before the thirty years ran out, and the most capable system anyone had built at the moment the clock expired was a chatbot that could not reliably do arithmetic. Whatever "shortly after" means, it does not mean what the sentence was read to mean.

2. Three of the four routes are dead, and the survivor won by a method he did not name.

Route two — networks and their users "waking up" — has not happened and is no longer a live research programme. Nothing in thirty-three years of the Internet resembles a single superhumanly intelligent entity emerging from the network. The idea has largely dropped out of the literature, and its absence is conspicuous in a document that gave it equal billing.

Route four — biological improvement of natural intellect — has not happened. No pharmacological or genetic intervention produces superhuman cognition; the 2008 restatement's own indicator for this, mind-enhancement drugs enabling real improvements in creativity, has not moved.

Route three — intimate human/computer interfaces — is the one he called the easier road, and it is the one that has moved least. neuralink-symbiosis-2019 records the state as of mid-2026: roughly 150 people worldwide have received a brain implant across all companies, Neuralink itself reports 21 participants, and its flagship published throughput is over 9 bits per second. Vinge's modest 1993 figure — that even "100 bits per second would be of great use to stroke victims" — is an order of magnitude above the best published rate thirty-three years later. Direct nerve-to-silicon interfaces are a real and valuable medical technology and are not on a path to superhumanity.

Route one — an "awake" superhumanly intelligent computer — is the only one still open, and the systems that exist arrived by a route not in the essay. Gradient descent on next-token prediction over internet text, at scale, using an architecture published in 2017. There is no IA in it, no network waking, no biology, and nothing about it was foreseen in a 1993 taxonomy that was otherwise unusually careful to enumerate possibilities.

3. The IA bet was tested in the one domain where it is cleanly measurable, and it lost. This is the entry's sharpest finding, because Vinge proposed the experiment himself: allow human/computer teams at chess tournaments. It was run. Kasparov introduced advanced chess in León in June 1998; the PAL/CSS Freestyle tournaments began in 2005, and the first was won by two American amateurs with ordinary hardware over grandmasters with far better machines — the strongest evidence ever produced for Vinge's thesis, and it is real. Then it eroded. deep-blue-1997 records Kenneth Regan's 2013 study measuring human-plus-engine teams at rough parity with engines alone on error rates, ahead only on move matching, and deliberately declines to settle the question. What has happened since is not in dispute even if the verdict is: neural evaluation overtook conventional engines when AlphaZero beat Stockfish in December 2017, the capability reached commodity hardware with Stockfish 12's NNUE release in September 2020, and the format has since faded from serious play because the human's contribution turned negative — an override of a modern engine is now almost certainly an error. The road Vinge called "much easier" was given twenty-five years, a fair test, enthusiastic participants and a scoreboard, and the human half of the centaur went from decisive to marginal to a liability. Two honest limits: this is a claim about one game and not about IA in general, and deep-blue-1997 explicitly does not settle the centaur question, so this entry does not overrule it. But Vinge named the test, and the test has an answer in the direction opposite to his prediction.

4. His own 2008 indicators, scored — because nobody else has. In "Signs of the Singularity" (IEEE Spectrum, 1 June 2008) Vinge reaffirmed the date — "I'd be surprised if the singularity had not happened by 2030," with the caveat "assuming we avoid the showstopping catastrophes," and "I'll stand by that claim" — and then did the rare and creditable thing of listing what to watch:

• Turing-test performance and its variants. Moved, substantially. turing-1950 records Jones and Bergen's three-party result in PNAS, May 2026: a persona-prompted model judged human 73% of the time in five-minute tests. That is not the eight-hour protocol of the Kurzweil–Kapor wager, and the distinction is kept at kurzweil-2005, but on Vinge's looser indicator this one has moved further than any other.
• A declining ratio of wages to world product. Not checked by this entry. It is a real, published, checkable series and the honest thing is to say that this job did not check it rather than to assert a direction. The most useful single piece of homework this file leaves behind.
• Molecular nanotechnology sustaining hardware progress. No. kurzweil-2005 records the nanotechnology branch of the same era's forecasting as absent rather than late. Hardware progress has continued by other means.
• Neural prostheses restoring lost cognitive function, such as memory. No. See route three above: restoration of motor and communication function for a few hundred people, no memory prosthesis.
• Mind-enhancement drugs producing real improvements in creativity. No.
• Progress blurring the line between neurology and AI research. No, and this is the interesting one. The line got sharper. The systems that work owe essentially nothing to neuroscience, which is the same finding kurzweil-2005 reached from the opposite direction — Kurzweil's route ran through brain scanning and the outcome routed around it.

One of six moved. One is unscored. Four have not moved in eighteen years. That is the closest thing to a scorecard this forecast has, the criteria on it are the forecaster's own, and no published attempt to fill it in was found.

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

This kind requires both, and requires saying that they differ. Here the shape of the difference is unusual and worth the space.

The self-grade is a reaffirmation, and then a silence. In February 2007 Vinge gave a Long Now talk titled "What If the Singularity Does NOT Happen?" in which he said he still believed a Singularity in the next few decades was the most likely outcome, and then laid out three futures in which it does not: a return to nuclear confrontation; a "Golden Age" of stabilised population and enlightened progress; and "The Wheel of Time," repeated catastrophe and recovery, which he called the most probable of the three. In June 2008 he restated 2030 in IEEE Spectrum and stood by it. And then, at the thirty-year mark, nothing. No public grading of the 1993 claim has been found from 2023, and there is a documented reason: Vinge was living with progressive Parkinson's disease and died in La Jolla on 20 March 2024, aged 79. (One outlet gives 21 March; the majority of obituaries give 20 March, and this file uses that.) He was alive for ChatGPT and for GPT-4 and no assessment of either by him surfaced in this research.

That absence is itself the finding, and it generalises. kurzweil-2005 established one failure mode of self-assessment — the forecaster who scores himself generously, 86% against independent recounts of 7% to 42%. good-intelligence-explosion-1965 established a second — Good reversed the valence of his 1965 claim in 1998 and never revisited its schedule. Vinge is the third and least-discussed case: the honest forecaster who names a date, a window and a set of indicators, restates them once, and is then simply not there when the date arrives. Thirty years is longer than a career and, often enough, longer than a life. A field that relies on principals to grade their own long-horizon forecasts will systematically lose the gradings of the longest and most interesting ones. That is an argument for keeping a ledger, and it is the reason this project keeps one.

The independent grades, and they do not agree with each other.

• David Chalmers, Journal of Consciousness Studies, 2010, lists the window accurately — "Vinge (1993) predicts greater-than-human intelligence between 2005 and 2030" — and does not grade it. Recorded here because good-intelligence-explosion-1965 catches the same paper listing Good's expired 2000 date without noticing it had expired ten years earlier. The most careful academic treatment of the singularity catalogues forecasts and scores none of them.
• The LessWrong deep dive, 9 February 2024, six weeks before Vinge's death, is broadly favourable: it credits the inevitability-through-competition argument, the uncontainability argument, the relevance of human/machine collaboration and the rapid dissemination of ideas, and flags direct nerve-to-silicon interfaces as unachieved.
• Mike Kalil, 1 November 2025, grades the four routes directly and reaches the opposite balance: superintelligent computers, failed; networks waking, failed; biological enhancement, failed; computer/human interfaces, partial and mostly medical. Three failed, one partial — which is what section 3 finds independently here.
• David Thorstad, "Against the singularity hypothesis," Philosophical Studies 182:1627–1651 (2024), attacks the growth premise rather than the date: the hypothesis "rests on undersupported growth assumptions," the proportionality claim is supported only locally, and diminishing returns to research make sustained acceleration expensive. superintelligence-2014 and von-neumann-singularity-1958 both already cite him.
• Ted Chiang, "Why Computers Won't Make Themselves Smarter," The New Yorker, 30 March 2021 — published, by coincidence, twenty-eight years to the day after the VISION-21 talk — attacks the recursion premise: intelligence growth is interactive, iterative and cooperative, and plateaus at the individual level.
• The forecasting aggregate, August 2026, which is not a grade but is the relevant comparison. Metaculus's median for AGI stood at roughly 2070 in 2020 and sits near 2033 in 2026, with about a 25% chance by 2029; the AI Impacts survey of 2,778 published researchers gave a 2047 median in 2023 and 2040 in its 2024 update; one combined tracker put 2031 as of 16 August 2026. Vinge's 2030 is at the optimistic edge of that distribution and inside it. Definitions differ enough between these that they should not be averaged, and all three reached this file through search summaries rather than from the platforms directly.

They differ, and here is the shape of it. Score the paths and Vinge is badly wrong: three of four dead, the fourth alive by a method he did not name, and the road he singled out as easier is the one that moved least. Score the date and he is not yet wrong at all, and the professional forecasting community has spent six years moving toward his number rather than away from it. Score the word and he is right in a way nobody else in the literature is, because he used it in its actual mathematical sense. Three defensible verdicts on one document, and which one a citer reports is usually decided by what they wanted to say.

The base rate this contributes, stated plainly. Set this file beside kurzweil-2005 — the canon's two longest-horizon forecasts, graded at length — and they fail in the same shape. Kurzweil named brain scanning and neuromorphic modelling; Vinge named intelligence amplification; neither named the route that worked, and both may yet be roughly right about when. Route and date fail independently, and the record so far is that forecasters are much better at when than at how. A reading that treats a correct date as evidence that the reasoning behind it was sound is making the error both of these entries exist to prevent.

An honest counterweight, because the smug version of this file would omit it. Vinge put a dated, decomposed, falsifiable claim on the record in 1993 and supplied his own indicators in 2008. Most of his critics have never put a dated list on the record at all. He also wrote, in the same essay, the paragraph describing what the world looks like if he is wrong — which is a discipline almost nobody in this genre practises, then or now. On the standard this canon applies to simon-1965, minsky-1970 and lighthill-1973, being wrong in public with a date attached is a contribution.

4. Commonly misused as

Not required for prediction. Included because five of these are in active circulation and two of them are opposites.

1. As the origin of the intelligence-explosion argument. It is not; it is Good's, and Vinge says so and quotes him. good-intelligence-explosion-1965 holds the argument and its defects, including that Good's own supporting citations are three papers by Good. Anyone who needs the argument should cite 1965. Citing 1993 supplies a name for the conclusion, not a reason for it.

2. As von Neumann's idea with a date attached. The word in the machine sense is Vinge's own, from Omni in January 1983. von-neumann-singularity-1958 establishes that the 1958 passage is Ulam's summary of an undated conversation, contains no machine, and was disclaimed by Vinge himself in the act of quoting it. The pedigree everyone recites is assembled, and this essay is where the assembly happened, in public, with the disclaimer printed alongside.

3. As "the singularity has started" / "we are past the event horizon." The live 2026 instances are Altman's June 2025 essay and Stripe's August 2026 investor letter dating it to 1 January 2026. Vinge's definition is "a point where our models must be discarded and a new reality rules" — an assertion that prediction has stopped working. A speaker who claims to be inside it while issuing dated roadmaps is using the word for its prestige and not for its content. The correct response is not derision; it is to ask which of Vinge's four routes is meant and which of his six 2008 indicators has moved.

4. As a vertical asymptote with a year under it. That reading belongs to the popularisers, not to Vinge. He drew no curve, fitted no data and offered no doubling time; his 2005–2030 window is Moravec's borrowed hardware estimate with the tails trimmed, and his own framing is about the failure of extrapolation, not its extension. This canon's one attempt to make the metaphor arithmetic — von Foerster's 1960 "Doomsday: Friday, 13 November, A.D. 2026," due in under three months and wrong by roughly ninety-fold, recorded at von-neumann-singularity-1958 — is a good reason to keep Vinge's version and the curve-fitters' version apart.

5. The reverse misuse: "2023 came and went, so the singularity is nonsense." This entry does not support that either, and it is the more common error in sophisticated company. The author's own outer bound is 2030 and is not due; the independent forecasting aggregate has moved toward it since 2020; and the argument's premises, attacked properly by Thorstad and Chiang, are attacked on their merits rather than on the failure of a headline date. Cite this entry against inflated claims that the transition is under way and against the reflex that a missed deadline settles the question. It grades both, and a reading using it for only one is using half of it.

6. The misuse this project must guard against directly. Nothing in this file is evidence. No date here — not 2023, not 2030, not 2033 — deposits into the ledger or moves the needle, and neither does any of the scoring above. A reading may cite this entry to name the argument behind an inevitability claim, to date and grade a principal's use of the word, to supply the four routes as a checklist, or to ask what would have to be observed for "the takeoff has started" to be true. 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 one man said would happen, thirty-three years ago, and how it has gone.

5. Sources

The primary text, and the exact state of this job's access to it. Vernor Vinge, "The Coming Technological Singularity: How to Survive in the Post-Human Era," VISION-21 Symposium, NASA Lewis Research Center and the Ohio Aerospace Institute, 30–31 March 1993; a slightly changed version in Whole Earth Review, Winter 1993. Two mirrors were used: the author's own page at edoras.sdsu.edu/~vinge/misc/singularity.html (four separate requests) and Hans Moravec's copy at frc.ri.cmu.edu/~hpm/book98/com.ch1/vinge.singularity.html (one request). Neither was read as a document. The available fetch tool re-renders a page through a summarising model, and on the first request that model declined to reproduce the article and thereafter returned excerpts under a short-quote limit. Everything in quotation marks above from the 1993 essay arrived that way, and the provenance is not uniform, so it is itemised. The abstract, the 2005–2030 sentence and the "blink of an eye" sentence came back from both mirrors in the same words and are treated as confirmed. The four route bullets are quoted from Carnegie Mellon; the SDSU request that covered them returned paraphrases, so as quotations they are single-sourced. The competitive-advantage sentence came back from both in two different lengths — a compressed one from Carnegie Mellon, a fuller one from SDSU — and the fuller one is used above, which is by itself a reason to check it. The IA proposals, the "golden age" passage, the Thinking Machines workshop report and the "embedded systems" passage came from a single request each. A session that can read the raw HTML should redo this section — it is the weakest link in the file, and the text is short, free and explicitly copyable.

Vinge's own later positions.

• "Signs of the Singularity," IEEE Spectrum, 1 June 2008. Source for the five scenarios (AI, IA, biomedical, Internet, Digital Gaia), the six indicators scored in section 3, the reaffirmation "I'd be surprised if the singularity had not happened by 2030" with its showstopping-catastrophes caveat, and "I'll stand by that claim." Read through the same summarising fetch tool.
• "What If the Singularity Does NOT Happen?", The Long Now Foundation, 15 February 2007. Source for the three non-Singularity futures and for his stated continuing belief that a Singularity remains the most likely outcome. Read from Long Now's own talk page.
• "First Word," Omni, January 1983 — quoted at one remove, through the Historical Dictionary of Science Fiction entry for "singularity" (sfdictionary.com/view/294), which gives it as the earliest cite for the sense and lists no antedating source. A search result reports the original as Omni vol. 5 no. 4, p. 10; that pagination was not verified and nothing above depends on it. The Singularity Summit Australia page that reproduces the column in full is dead — 2012.singularitysummit.com.au no longer resolves.
• No self-assessment from 2023 was found, and four searches were spent looking. Recorded as not found rather than as not existing. Likewise no comment by him on ChatGPT or GPT-4 surfaced. His 2019 interview with The Next Billion Seconds, republished as a memorial on 22 March 2024, was fetched and the rendering returned no substantive content from the conversation; it is the most likely place for a late position and is worth a retry.

The independent grades.

• David J. Chalmers, "The Singularity: A Philosophical Analysis," Journal of Consciousness Studies 17:7–65 (2010) — the "between 2005 and 2030" characterisation, taken from canon/good-intelligence-explosion-1965.md, which read the pages, rather than from Chalmers directly.
• "[Crosspost] Deep Dive: The Coming Technological Singularity," LessWrong, 9 February 2024.
• Mike Kalil, "The Coming Technological Singularity (Vernor Vinge)," 1 November 2025 — the four-route scoring.
• David Thorstad, "Against the singularity hypothesis," Philosophical Studies 182:1627–1651 (2024) — abstract and summaries only, consistent with how von-neumann-singularity-1958 cites it.
• Ted Chiang, "Why Computers Won't Make Themselves Smarter," The New Yorker, 30 March 2021 — from search summaries and secondary accounts; the article itself was not read.
• Daniel Lemire, "Revisiting Vernor Vinge's 'predictions' for 2025," 4 September 2015 — a grading of a different Vinge work, the 2006 novel Rainbows End, which is set in 2025. Read directly. Not used above except as background, because fiction is not prediction under this canon's rules and because grading a novel's furniture is a different exercise; noted here because a later session looking for more Vinge base rate will find it, and because Lemire's 2015 verdicts (ubiquitous augmented reality unlikely; self-driving imminent; book digitisation exceeding expectations) are themselves now gradeable.
• Forecasting aggregates, all via search summaries and none read at source: Metaculus's AGI median (~2070 in 2020, ~2033 in 2026, ~25% by 2029) via 80,000 Hours and secondary trackers; the AI Impacts survey of 2,778 researchers (2047 median in 2023, 2040 in the 2024 update); a combined forecast of 2031 dated 16 August 2026. Definitions differ across all three — one requires robotic competence — and they are reported as a distribution, not averaged.

The centaur-chess record.

• English Wikipedia, "Advanced chess," for León June 1998, the PAL/CSS Freestyle tournament from 2005 and the amateur "ZackS" win, and for the conflicting later assessments (Cowen 2013 sceptical; Kasparov 2017 still confident; Bridle 2018 asserting the old claim).
• canon/deep-blue-1997.md for Kenneth Regan's 2013 study — human-plus-engine at rough parity with engines alone on error rates — and for that entry's explicit refusal to settle the centaur question, which this file does not overrule.
• AlphaZero's December 2017 result and Stockfish 12's NNUE release of September 2020, and the current practitioner view that a human override of a modern engine is almost certainly an error, from search summaries of chess-technical writing rather than from a primary study. This is the weakest-sourced claim in section 3 and is stated as a direction of travel; a session that wants it firm should look for a post-2020 successor to Regan's measurement.

Fetch failures, recorded so a reader knows what was not checked. Ron Schmelzer, "If The AI Singularity Is Here, Where Is The Evidence?", Forbes, 28 July 2026 — HTTP 403, and it is the most relevant recent independent assessment that this file could not read. A 5 February 2026 close reading of the essay at trance-scripts.com — HTTP 403. 2012.singularitysummit.com.au — DNS failure. Direct curl is not available to this job, so no fetch failure here had a fallback.

Project files read while writing, and only these. CLAUDE.md; the closing section of LENSES.md, for the rule that a principal's forecast is a canon entry and never a deposit; the promoted proposal in canon/proposals.md, added 2026-08-17; canon/kurzweil-2005.md and canon/von-neumann-singularity-1958.md in full; the Vinge passages of canon/good-intelligence-explosion-1965.md, the hardware passages of canon/moravecs-paradox-1988.md, the implant figures in canon/neuralink-symbiosis-2019.md and the advanced-chess passages of canon/deep-blue-1997.md, each located by search rather than read whole; build.py's front-matter parser and Markdown functions; and digests/2026-08-20-00.md for the citation occasions in section 2. README.md was not opened in this session — the rules relied on above are the ones restated in CLAUDE.md and in this job's own prompt.

Two renderer constraints, since build.py was checked anyway and the second is not written down anywhere yet. Quotations here are italic paragraphs rather than blockquotes because md_to_html has no blockquote rule and the marker leaks onto the page as an escaped angle bracket — von-neumann-singularity-1958 found that one. The second is new: bold and italic cannot be nested. Both patterns in inline_md forbid an asterisk inside the emphasised run, so a bolded source line containing an italicised journal title matches neither rule and prints its emphasis markers literally on the page. Several lines in this file's Sources section were written that way and rewritten before saving; at least one entry already in canon/ still contains the pattern. A session writing a new entry should keep the two emphases side by side rather than one inside the other, and a session tidying the canon could fix the existing cases mechanically.

One edge that cannot be drawn from here. kurzweil-2005's descends_from should now include this id — The Singularity Is Near opens on page 22 with a Vinge epigraph and page 24 adds his 1983 Omni piece and his 1986 novel Marooned in Realtime — and a canon job writes exactly one file, so that entry's header cannot be amended from here. The arrow is the owner's or a later job's to draw.

No evidence was deposited, no needle, score or landmark is placed, and no file outside canon/vinge-singularity-1993.md was written.