← The canon · AItopiaOrAImageddon?

Speculations Concerning the First Ultraintelligent Machine

prediction · Irving John Good, then of Trinity College, Oxford and the Atlas Computer Laboratory, Chilton, Berkshire · 1965 — the volume; the monograph was drafted April 1963 and amended May 1964, from talks given October 1962 and January 1963

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 "The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain". Read on: "The Coming Technological Singularity: How to Survive in the Post-Human Era".

proposals.md files this id in its ## idea section and the file is filed here as prediction. The change is the first thing this entry has to defend, and defending it is not a formality, because the paper's fame rests entirely on an argument and this filing puts it in the kind reserved for dated claims.

Why idea fails, on the canon's own test. minimax-alpha-beta set the test and von-neumann-singularity-1958 restated it: an idea is something you can implement, because there is a loop or a formalism in it. The intelligence explosion, as Good states it, is four sentences of conditional prose. There is no rate, no threshold, no quantity, no formalism, and nothing to code. How much better is the second machine than the first? Over what interval? Good does not say, anywhere, and the word "explosion" is doing the work a number would do. When David Chalmers set out to make it rigorous in 2010 he had to supply three premises Good never wrote down and a defeaters clause on each of them, which is the clearest possible demonstration that what Good published was not yet an argument in the sense the idea kind requires. The written idea entries in this canon — transformer-2017, scaling-laws-2020, rlhf-christiano-2017, sparse-moe-2017, minimax-alpha-beta — are all things you could build from the entry. This is not one of them.

Why prediction fits, on the canon's four fields. The prediction kind requires the claim in the predictor's own words, the date it was made, the date it was due, and what actually happened. Good supplies all four himself, in a section headed "Conclusions," in a sentence he introduces with the italicised words In the writer's opinion then. The claim carries a probability ("more probable than not") and a deadline ("within the twentieth century"). It came fully due on 31 December 2000 and it failed. Filing the founding forecast of the field's most-invoked framework under idea would keep the single most-cited AI prediction of the twentieth century out of the base rate the prediction kind exists to build — and that is precisely the shape of drift rule 7 says no program can catch, a grade going quietly soft by never being taken. simon-1965 is the sibling case: same year on the spine, a claim made earlier than the book that carries it, already filed prediction, already graded against a date. These two belong in the same kind.

This does not demote the argument, and section 2 is about the argument. The reason a reading cites this file is recursive self-improvement, not a 1965 deadline. But the man who made the argument attached a date to it, in his own voice, and this canon records dates.

The question this entry was written to settle

von-neumann-singularity-1958 ends its source list with an open item: "Whether Good's own reference list cites Ulam could not be established — four searches failed to surface the paper's bibliography and the full text was not reached. Recorded as unresolved rather than guessed. good-intelligence-explosion-1965 is proposed and unwritten and is where this should be settled."

Settled. The full text was reached and the reference list was read at source, in page images, at pages 83–88. It runs to 108 numbered entries plus four inserted ones (16a, 78a, 80a, 83a), alphabetically by author, and it carries a footnote on its heading: "Only items mentioned in the text are listed here; for a bibliography on artificial intelligence see Minsky [72]."

Because the footnote states the list contains only items mentioned in the text, the absence is stronger than an absence from a bibliography usually is: Ulam, von Neumann and Turing are not in the reference list because they are not in the paper.

And the positive half of the finding is better than the negative half. The famous sentence carries three citations of its own — "there would then unquestionably be an 'intelligence explosion,' and the intelligence of man would be left far behind (see for example refs. [22], [34], [44])." Those three are:

All three are Good citing Good. The founding argument of the intelligence-explosion literature, at the moment it is stated, is supported by a book review, an internal IBM report and an essay in a collection the same author edited. Whatever this idea descends from, it does not descend from Ulam's memorial paragraph, and the pedigree that runs von Neumann → Ulam → Good → Vinge → Kurzweil, which is repeated constantly, is broken at its second link.

On descends_from, which holds one id. rosenblatt-perceptron-1958 is there and is earned twice over: Rosenblatt's Principles of Neurodynamics is reference [82], and reference [34] — one of the three the intelligence-explosion sentence cites — is Good's own "Speculations on perceptrons and other Automata." The perceptron is in the lineage of both halves of this paper, the argument and the machine. Nothing else in canon/ can be put on that line honestly. samuel-checkers-1959 is in Good's reference list at [83] and [83a], but this job did not read the pages where Samuel is used in the text and will not claim descent from a bibliography entry. turing-1950 is the ancestor a reader expects and it is deliberately absent: Good does not cite it, and writing the edge anyway would erase the finding above, which is the most useful thing this file has. von-neumann-singularity-1958 is absent for the reason just established. The arrow that cannot be drawn from here runs the other way: Good was a consultant to Stanley Kubrick on 2001, so 2001-hal-1968 is plausibly a descendant of this entry rather than an unrelated neighbour, and a canon job writes exactly one file, so that entry's descends_from line cannot be amended from here.

1. What it is

The document

I. J. Good, "Speculations Concerning the First Ultraintelligent Machine," in Advances in Computers, volume 6, edited by Franz L. Alt and Morris Rubinoff, Academic Press, pages 31–88. It is universally cited as 1965; some catalogue records give 1966 for the volume, and the discrepancy does not matter for anything below.

The dates inside the paper matter more than the date on the volume, and they are in the asterisked footnote on page 31: "Based on talks given in a Conference on the Conceptual Aspects of Biocommunications, Neuropsychiatric Institute, University of California, Los Angeles, October 1962; and in the Artificial Intelligence Sessions of the Winter General Meetings of the IEEE, January 1963 [1, 46]. The first draft of this monograph was completed in April 1963, and the present slightly amended version in May 1964." So the argument was spoken in October 1962, written in April 1963, frozen in May 1964, and published under a 1965 date. This canon has already had to separate a claim's date from its publication's date once — simon-1965 grades against 1980 rather than 1985 for exactly this reason — and the same discipline applies here, though in this case it does not change the verdict, because the deadline Good set was the end of the century either way.

Good, born Isadore Jacob Gudak in 1916, was a Bletchley Park cryptanalyst who worked with Turing in Hut 8 from 1941, later a professor at Virginia Tech from 1967, and a consultant to Kubrick on 2001: A Space Odyssey. He died in 2009. Nothing about him is peripheral to why the paper was taken seriously: it is a statistician's paper, and its author had watched machine cryptanalysis work.

The opening sentence

The paper's first line, before any definition or argument:

"The survival of man depends on the early construction of an ultraintelligent machine."

That sentence is the one Good came back to thirty-three years later, and section 3 records what he did to it.

The famous paragraph, entire, from page 33

"Let an ultraintelligent machine be defined as a machine that can far surpass all the intellectual activities of any man however clever. Since the design of machines is one of these intellectual activities, an ultraintelligent machine could design even better machines; there would then unquestionably be an 'intelligence explosion,' and the intelligence of man would be left far behind (see for example refs. [22], [34], [44]). Thus the first ultraintelligent machine is the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control. It is curious that this point is made so seldom outside of science fiction. It is sometimes worthwhile to take science fiction seriously."

Five things in that paragraph are worth marking, and three of them are routinely lost.

The definition is comparative, not absolute. "Can far surpass all the intellectual activities of any man however clever" is a definition by reference to the best human, not by reference to any threshold of capability. Everything downstream inherits that: the argument is about a machine that beats people at a task, not about a machine that has crossed a line.

The recursion rests on one substitution. Machine design is an intellectual activity; a machine that beats humans at all intellectual activities therefore beats humans at machine design; therefore it designs a better machine than humans could. This is the whole engine, and it is a single step of substitution, iterated on faith.

"Unquestionably" is doing load-bearing work and is unearned. It is the only place in the sentence where a rate could have gone, and Good put an adverb there.

The safety condition is inside the founding sentence, not appended to it. "Provided that the machine is docile enough to tell us how to keep it under control" makes the whole claim conditional on the machine's cooperation in its own containment. The optimistic reading and the alignment problem arrive in the same breath, in 1965.

The last two sentences are a claim about priority, and they are wrong; see section 4, misuse 6.

Where it sits in the paper, which almost nobody who cites it knows

The paper is 58 pages. Its table of contents, on page 31, reads: 1. Introduction (31); 2. Ultraintelligent Machines and Their Value (33); 3. Communication as Regeneration (37); 4. Some Representations of "Meaning" and Their Relevance to Intelligent Machines (40); 5. Recall and Information Retrieval (43); 6. Cell Assemblies and Subassemblies (54); 7. An Assembly Theory of Meaning (74); 8. The Economy of Meaning (77); 9. Conclusions (78); 10. Appendix: Informational and Causal Interactions (80); References (83).

The intelligence explosion is one paragraph on one page. Everything the world quotes lives in pages 31–37. Pages 37 to 82 — three quarters of the paper — are communication theory, semantics, statistical information retrieval, an extended proposal for a "subassembly" modification of Hebb's cell-assembly theory of the brain, and an appendix on informational and causal interactions in contingency tables. Good says so himself on page 31: the paper's aim is "to take more of the magic out of the brain by means of a 'subassembly' theory, which is a modification of Hebb's famous speculative cell-assembly theory."

This is not trivia. The most-quoted paragraph in AI forecasting is a two-paragraph aside in a neuroscience monograph, and the ratio — one paragraph of argument to fifty-odd pages of the substrate work its author actually cared about — is itself the fact. Good did not write a paper about the intelligence explosion. He wrote a paper about how to build a brain, and mentioned in passing what would happen if it worked.

The rest of section 2, which nobody quotes either

Section 2 is four and a half pages and it contains the first cost–benefit analysis of building a superintelligence. Good takes from Carter [11] an estimate of the value to the world of J. M. Keynes — "at least 100,000 million pounds sterling" — and coins the unit. "By definition, an ultraintelligent machine is worth far more, although the sign is uncertain, but since it will give the human race a good chance of surviving indefinitely, it might not be extravagant to put the value at a megakeynes. There is the opposite possibility, that the human race will become redundant."

Then the calculation: a hundred billion dollars might simulate all the neurons of a brain at ten dollars per artificial neuron, but "it seems unlikely that more than say a millikeynes would actually be forthcoming, and even this amount might be difficult to obtain without first building the machine! It would be justified if, with this expenditure, the chance of success were about 10⁻⁹."

That arithmetic is internally consistent to the exponent, and checking it is one line: a keynes is 10¹¹ pounds, so a megakeynes is 10¹⁷ and a millikeynes is 10⁸, and a 10⁻⁹ chance at 10¹⁷ has an expected value of exactly 10⁸. Good was a Bayesian statistician and he did the sum properly. He also flagged the term that breaks it — "although the sign is uncertain" — which is the whole of the alignment problem written as a caveat inside an expected-value calculation, in 1965.

The same section contains, in three sentences, most of the ethical agenda the field rediscovered forty years later: "There are other ethical problems, such as whether a machine could feel pain especially if it contains chemical artificial neurons, and whether an ultraintelligent machine should be dismantled when it becomes obsolete. The machines will create social problems, but they might also be able to solve them in addition to those that have been created by microbes and men. Such machines will be feared and respected, and perhaps even loved. These remarks might appear fanciful to some readers, but to the writer they seem very real and urgent, and worthy of emphasis outside of science fiction."

The Conclusions, which are the prediction

Page 78, opening section 9, and this is the passage the kind of this file rests on:

"These 'conclusions' are primarily the opinions of the writer, as they must be in a paper on ultraintelligent machines written at the present time. In the writer's opinion then: It is more probable than not that, within the twentieth century, an ultraintelligent machine will be built and that it will be the last invention that man need make, since it will lead to an 'intelligence explosion.' This will transform society in an unimaginable way. The first ultraintelligent machine will need to be ultraparallel, and is likely to be achieved with the help of a very large artificial neural net. The required high degree of connectivity might be attained with the help of microminiature radio transmitters and receivers. The machine will have a multimillion dollar computer and information-retrieval system under its direct control. The design of the machine will be partly suggested by analogy with several aspects of the human brain and intellect. In particular, the machine will have high linguistic ability and will be able to operate with the meanings of propositions, because to do so will lead to a necessary economy."

Claim, probability, deadline, and — unusually for this genre — a specification of the thing predicted, in six clauses, every one of which can be scored separately. Section 3 scores them.

2. Why a reading would cite it

The occasion is when recursive self-improvement is argued for or against, which in 2026 is not an abstract occasion at all. Every claim that models are now contributing to the design of their successors — writing the training code, searching the architecture space, laying out the chips, reviewing the research — is a claim that the first step of Good's substitution has been taken. The reading does not need this entry to evaluate such a claim on its evidence. It needs this entry for something the evidence cannot supply: what the original argument actually commits its holder to, and what it conspicuously does not.

What it commits to is one thing: a system better than the best human at designing systems will design a system better than itself. What it does not supply is a rate. Not one number appears anywhere in the argument. So a reading meeting "we are in the early stages of an intelligence explosion" can say, precisely and without sneering, that the sentence's founding text is compatible with a doubling every month and with a doubling every century, and that anyone invoking it for urgency is importing a timescale from somewhere else and should be asked where. This is the same service von-neumann-singularity-1958 performs for the word "singularity", one link further down the chain and with more content to work with: Ulam's sentence has no mechanism, Good's has a mechanism and no rate, and Bostrom's is the first with a variable in it. A reading that wants to contest a takeoff claim needs to know which of the three it is being handed.

The second occasion is the docility clause. When a lab argues that a sufficiently capable system will help solve its own alignment — that the model will be a collaborator in oversight rather than an obstacle to it — that argument is not new and is not a modern softening. It is the second half of the sentence that founded the optimistic case, written in 1965, and its author made his entire conclusion conditional on it. That cuts both ways and should be cited both ways: it means the argument was never naive, and it means the field has been leaning on an unproven conditional for sixty-one years.

The third occasion is the base rate, and it is the reason for the kind. This canon grades forecasts to know how wrong forecasting runs. Good is the strongest possible data point because he was not a hype merchant, not a vendor, not selling anything: a Bletchley cryptanalyst and a serious Bayesian, writing in a refereed series, hedging correctly ("more probable than not," not "certainly"), specifying his prediction well enough to be scored, and putting a deadline on it. He was as good as this genre gets, and section 3 shows what happened to the date. A reading tempted to weigh a principal's forecast heavily because the principal is technically excellent should read that grade first.

The fourth occasion is capital. Good's fantasy budget for building the machine — the figure he thought could not plausibly be raised — was a hundred billion dollars in 1965 money, and he judged that spending a thousandth of it was already unlikely. What that number has become in 2026 is a matter for a reading's own sourced evidence and this file supplies none of it; the entry's contribution is only that the founding cost–benefit analysis exists, is arithmetically sound, and priced the attempt at a level that was then unthinkable.

What this entry may not be used for. It is not evidence. It deposits nothing and moves nothing. That an argument has Good behind it does not make it stronger, and that Good's date expired does not make an argument weaker — see section 4, misuses 6 and 7, which cut in opposite directions on purpose.

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

The headline claim

The claim, in his own words: "It is more probable than not that, within the twentieth century, an ultraintelligent machine will be built and that it will be the last invention that man need make, since it will lead to an 'intelligence explosion.'"

Date made: October 1962 as a talk, April 1963 in draft, May 1964 as amended, published under 1965.

Date due: 31 December 2000. Unambiguous, unusually for this canon — no twenty-year horizon to argue about, no reprint drift, no "a few decades hence."

What happened: No. No machine existing on 31 December 2000 could far surpass all the intellectual activities of any man however clever, and none exists on the date this file is written. The claim is not near-missed or partially satisfied; the defined object did not appear. Whether anything since has taken the first step of the argument is a live question that a reading may weigh on evidence, and it does not rescue the deadline, which was for the finished machine and not for a promising direction.

Good's own grade, and it is the interesting part. In 1998, at 82, Good was given the IEEE Computer Society's Computer Pioneer Award and submitted a biographical statement, written in the third person, which he did not read aloud and which nobody read aloud. It says of this paper: "'Speculations Concerning the First Ultra-intelligent Machine' (1965) began: 'The survival of man depends on the early construction of an ultra-intelligent machine.' Those were his words during the Cold War, and he now suspects that 'survival' should be replaced by 'extinction.'" It continues: "He thinks that, because of international competition, we cannot prevent the machines from taking over. He thinks we are lemmings." And it closes by quoting his own page 32 back at himself: "probably Man will construct the deus ex machina in his own image."

Record his grade and an independent one, and say they differ — and here they differ on a different axis, which is the finding. Good's self-assessment reverses the sign of his 1965 claim and leaves the timing untouched. He came to think the machine would be worse than he had said, not later than he had said; the 1998 statement is more confident about arrival than the 1965 paper was, not less. The independent grade concerns only the timing, and it is a clean miss on a deadline thirty-five years long. A predictor who revises the valence of a forecast while never revisiting its schedule has not graded the forecast, and this is the single most transferable thing in this file, because the same manoeuvre is available to every principal writing today: change what you say the machine will do to us, never mention that you said when.

The independent grade also has a documented failure of its own. Chalmers's 2010 philosophical analysis lists forecasts to calibrate its own timeframe — "For example, Good (1965) predicts an ultraintelligent machine by 2000, Vinge (1993) predicts greater-than-human intelligence between 2005 and 2030, Yudkowsky (1996) predicts a singularity by 2021, and Kurzweil (2005) predicts human-level artificial intelligence by 2030" — and does not remark that the first of the four had expired ten years before he wrote the sentence. The most careful academic treatment of the singularity records the failed date accurately and treats it as though it were still open. That is what happens to expired predictions in this field when nobody keeps a ledger, and it is a fair summary of why this canon keeps one.

Provenance warning on the self-grade, stated plainly because the parallel is uncomfortable. The 1998 statement was submitted, not delivered; the only person known to have had a copy was Leslie Pendleton, Good's longtime assistant; James Barrat obtained it from those personal papers and published it in Our Final Invention in 2013. This file has it at one remove, through secondary reporting of Barrat, not from the document. So: the founding text of the singularity is a friend's unquoted paraphrase of a dead man in an obituary, and the founding predictor's recantation is an unread award biography in an assistant's private papers. Both hinge on a single custodian, and the field repeats both as though they were on the record. The recantation is well attested and is very probably exactly as reported; it is recorded here as attested rather than as verified, and the difference is not rhetorical.

The six specific clauses of the Conclusions, scored one at a time

This is where the grade becomes uncomfortable for the easy reading, because Good was wrong about the date and substantially right about the machine.

Right — "will need to be ultraparallel." Frontier systems are trained and served on tens of thousands of parallel accelerators, and parallelism is the defining property of the hardware. Correct, against the entire serial-computer consensus of 1963.

Right, and this is the strong one — "likely to be achieved with the help of a very large artificial neural net." Written four years before perceptrons-1969 and twenty-three years before the connectionist revival, in a decade when symbolic AI was the field. Good bet on scale in neural networks and against symbolic reasoning, and the bet came in. Compare minsky-1970, lighthill-1973 and expert-systems-collapse-1987 for what the consensus was doing meanwhile.

Right — "high linguistic ability and will be able to operate with the meanings of propositions." He specified language as the load-bearing capability of the first ultraintelligent machine, and gave a reason for it on page 36: "it will need to be able to handle or to learn to handle ordinary language with great facility. This will be important in order that its instructor should be able to teach it rapidly, and so that later the machine will be able to teach the instructor rapidly." A machine taught in natural language, that then teaches its instructor in natural language, is a fair description of what shipped in 2022 and is not a fair description of anything anyone else was building in 1963.

Right in mechanism — learning by reinforcement with stochastic weight change. Page 79: "it is proposed that this be effected by means of reinforcement of the strengths of artificial synapses, that the available strengths for each synapse should form a discrete set, that when a synapse is not used for a certain length of time it should have a certain small probability of 'mutation' down one step, and that when a synapse is 'successfully used' … it has a certain small probability of mutation up one step. The need for the changes in synaptic strength to be only probabilistic, with small probabilities, is that they would otherwise vary too quickly for the machine to be of any use." Discretised weights, stochastic updates, and a stability argument for small learning rates. It is not backpropagation and should not be dressed up as it. It is a recognisable gesture at the right family, from a statistician who never built it.

Wrong — "microminiature radio transmitters and receivers" for connectivity. Page 35's proposal, italicised in the original, is that each element contain a short-range radio transmitter and receiver, with connection strength represented by tuning accuracy. Nothing was ever built this way; connectivity was solved by memory bandwidth and interconnect fabric, not by broadcast. This is the clause where a very good forecaster reached for the physical technology in front of him and got it wrong, which is the most common way this genre fails.

Wrong in scale — "a multimillion dollar computer and information-retrieval system under its direct control." Multimillion is off by three to four orders of magnitude against frontier training runs. Note the direction: this is one of the rare cases in this canon of an under-estimate of what the thing would cost, from the same author who elsewhere in the same paper priced the attempt at a hundred billion dollars. He could imagine the large number and could not imagine it being spent.

Three side-bets, because they show how much of this paper goes unread

Wrong, by twenty years — the gigahertz date. Page 35: by "say 1980" there will have been great advances in microminiaturisation and "pulse repetition frequencies of one billion pulses per second will surely have been attained in large computers." Machine clock rates reached 1 GHz in 2000, not 1980; the fastest large machines around 1980 ran at tens of megahertz. The caveat this file owes the reader is that "pulse repetition frequency" in a switching device is not exactly a system clock rate, and laboratory switching speeds ran ahead of deliverable clocks; graded as clock rate, the word "surely" was two decades early.

Right, and then overtaken — centaur chess. Page 34: "perhaps the best intellectual feats will be performed by men and machines in very close, sometimes called 'symbiotic,' relationship," with the observation that chess programs would be far more effective "if the chess-playing programs were written with extremely close man-machine interaction in mind from the start." Advanced chess began with Kasparov against Topalov at León in June 1998 and human–engine teams did outperform engines alone for roughly the following decade and a half. Then the second half of the prediction failed: engines improved to the point where human intervention subtracts rather than adds. The prediction was correct for about twenty years and its premise — that the human half would keep contributing — did not survive. Read alongside deep-blue-1997 and bitter-lesson-2019, both of which are about this exact reversal.

Right, and thirty-one years early — randomised chess. Same page: Good proposes permuting the back rank at random before play, with the bishops constrained to squares of opposite colours and Black mirroring White, "to remove from the game the effect of mere parrot learning of the openings", and states the count as 1440 essentially distinct initial positions. Bobby Fischer announced Fischerandom, now Chess960, in 1996. The arithmetic, computed here rather than sourced: 4 × 4 choices place the two bishops on opposite colours, and the remaining six squares admit 6!/(2!·2!) = 180 arrangements of Q, K, R, R, N, N, giving 2880; no such arrangement is fixed by a left–right reflection of the rank, so counting reflected pairs as one gives exactly 1440. Fischer's rule adds the requirement that the king stand between the rooks, which keeps one of every three K–R–R orderings and yields 960. Good's number is exactly right under the natural reading of "essentially distinct," and he published it three decades before the world champion who is usually credited with the idea.

What the record adds up to

Good got the architecture right and the schedule wrong, which is the reverse of this canon's usual finding. simon-1965, minsky-1970 and musk-robotaxi-2019 are all cases of a confident date attached to a capability that turned out to be much harder than the forecaster understood. Good's failure is different in kind: he named neural networks, parallelism, language and reinforcement in 1963, and every one of those calls landed — and he was still wrong by at least a quarter of a century, because being right about what tells you nothing about when. The lesson is not that Good was a poor forecaster. It is that being an excellent forecaster of mechanism buys no accuracy at all on timing, and that is the base-rate contribution this entry makes.

4. Commonly misused as

limit entries owe this section and prediction entries do not, but this one is more misquoted than most limit results and the misuses run in both directions.

1. As a proof. It is a conditional with an unearned adverb. Chalmers's 2010 reconstruction breaks it into an equivalence premise (there will be AI at least equivalent to human intelligence), an extension premise (AI will soon be extended to AI+), and an amplification premise (AI+ will soon be greatly amplified to AI++), each qualified "absent defeaters," where defeaters include "disasters, disinclination, and active prevention." Good supplies, informally, only the third. Anyone citing "Good showed that…" is citing a premise as a conclusion.

2. As containing a timescale. It does not. "Explosion" is the only rate in it and it is a metaphor. The dates in the surrounding literature belong to their own authors — Good's own 2000, Vinge's 2005–2030, Kurzweil's 2030 — and are separately gradeable; none of them follows from the argument.

3. Quoted without the docility clause. Vinge's 1993 essay quotes the passage complete, including "provided that the machine is docile enough to tell us how to keep it under control." Chalmers's 2010 paper, the most careful academic treatment of the argument, quotes the passage and stops at "the last invention that man need ever make." The safety condition disappears in the version a philosopher of mind hands to the field. This file corrects one thing in von-neumann-singularity-1958, which records that Chalmers "reproduces it in full": Chalmers reproduces it without the docility clause, and nothing in that entry's argument depends on the difference.

4. As descended from von Neumann and Ulam. Settled above and settled negatively. The defensible statement is that the intelligence-explosion argument and the "singularity" metaphor are independent inventions that were welded together retroactively by Vinge in 1993 and again by Kurzweil in 2005. Chalmers keeps them properly separate — his own reference to Ulam is a footnote, at one remove through Vinge, about the term and not the argument — and the popular literature does not.

5. As Good's settled view. He reversed the sign in 1998. Quoting the 1965 opening sentence — "the survival of man depends on the early construction of an ultraintelligent machine" — as though it were his position is quoting a man against himself, and the optimistic use of that line is the more common of the two.

6. And the reverse misuse, which this entry must not encourage: as evidence that Good recanted his argument. He did not. The 1998 statement changes his estimate of the outcome and, if anything, strengthens his confidence in the arrival: he says we cannot prevent it. Nothing in it retracts the recursion or admits the date was wrong. It is also a paragraph in an award biography, written in the third person, never delivered — not an argued paper, and it should not be weighed as one. A reading that wants the strongest pessimistic case should take it from superintelligence-2014, which argues, rather than from an octogenarian's unread aside, which asserts.

7. As a priority claim. "It is curious that this point is made so seldom outside of science fiction" is Good's own assessment and it is wrong twice over. Turing had made the point publicly in 1951, in a Manchester lecture Good never cites: "it seems probable that once the machine thinking method had started, it would not take long to outstrip our feeble powers … At some stage therefore we should have to expect the machines to take control, in the way that is mentioned in Samuel Butler's 'Erewhon'." And the science-fiction antecedent Good gestures at is real and older — Chalmers names John Campbell's 1932 "The Last Evolution." Good's contribution is not first statement. It is the first statement by someone with standing, in a refereed venue, with a mechanism and a date attached, which is a smaller claim and a truer one.

8. And the genetic fallacy, in both directions, which is the standing rule here. That the founding forecast missed its deadline by at least twenty-six years does not refute the argument, exactly as von-neumann-singularity-1958 insists that a thin pedigree does not refute the singularity hypothesis. The serious critiques attack the premises: Robin Hanson on whether a local explosion is possible when improvement is distributed across an economy, Toby Walsh's catalogue of ceilings — diminishing returns, computational complexity, the limits of self-improvement — and David Thorstad's "Against the singularity hypothesis" (2024) on undersupported growth assumptions. The correct use of this entry is narrow and symmetrical: the argument has no rate in it, so make the speaker supply one, and check whether the last person to supply one has come due.

5. Sources

Read at source, in the pages named:

Taken from secondary sources, and flagged as such:

Computed here, not sourced: the 1440/2880/960 chess arithmetic in section 3, and the check that Good's megakeynes / millikeynes / 10⁻⁹ figures are mutually consistent. Both are one line each and anyone re-deriving them will get the same numbers.

Project context read while writing: README.md, CLAUDE.md, LENSES.md, canon/proposals.md (which files this id under idea, at the line beginning "I. J. Good, 'Speculations Concerning the First Ultraintelligent Machine'"), the whole of canon/von-neumann-singularity-1958.md, the header blocks of the written entries, the kind-declining openings of canon/simon-1965.md, canon/superintelligence-2014.md and canon/minsky-1970.md, the Good passages in canon/kurzweil-2005.md and canon/superintelligence-2014.md, and build.py's Markdown functions — which is why the quotations in this file are set as italic paragraphs rather than as blockquotes, the renderer having no blockquote rule.

Failed fetches, recorded so a reader knows what was not checked: encyclopedia.pub (403) and ddcolrs.wordpress.com (403), both of which carry the Barrat passage; a mirror of the Barrat text at readbooksol.com (403). The Quote Investigator trace of "the last invention that man need ever make" was fetched successfully but contained no pre-1965 antecedent and nothing on the 1998 reversal.

No evidence was deposited, no needle, score or landmark is placed, and no file outside canon/good-intelligence-explosion-1965.md was written.