← The canon · AItopiaOrAImageddon?

"Meet Shaky, the first electronic person"

prediction · Brad Darrach, writing in Life; the forecast is attributed in the article to Marvin Minsky, who later denied making it · 1970

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 Dartmouth Summer Research Project on Artificial Intelligence, "2001: A Space Odyssey".

prediction is the right kind and proposals.md filed it correctly. But this file has a problem none of the other prediction entries has, and it has to be stated in the first paragraph rather than buried in a sourcing note:

The predictor denied making the prediction.

The sentence is real. It was printed in Life on 20 November 1970, it is quoted accurately by almost everyone who quotes it, and it has done fifty-six years of work as the canonical example of an AI researcher's overconfidence. What is not established is that Marvin Minsky said it. He is on record calling the quote made up. The reporter who wrote it was, five years later, accused in a different domain — under oath, in a $20 million complaint — of writing dialogue for scenes he had not attended. And Minsky is also on record, in his own book, under his own name, three years earlier, making a claim of the same shape with a longer horizon.

So the entry has three jobs, and they are separable. Grade the claim. Grade the attribution. And then notice that those two gradings come apart, because that is the thing this file teaches that no other entry in canon/ teaches: a forecast can fail and its author can decline the failure, and the mechanism by which he declines it is neither dishonest nor obviously wrong. The canon already holds one worked case of a forecaster grading himself generously (kurzweil-2005) and one of a forecaster moving his horizon by the elapsed time (amodei-2024-loving-grace). This is the third move, and it is the one that most damages a base rate, because it removes the data point entirely.

1. What it is

The article

On 20 November 1970, Life ran a feature by staff writer Brad Darrach across pages 58B, 58D, 66 and 68, under the headline "Meet Shaky, the first electronic person" and the subhead "The fascinating and fearsome reality of a machine with a mind of its own." The subject was Shakey, a mobile robot built at Stanford Research Institute in Menlo Park. Life misspelled the robot's name — SRI called it Shakey, the magazine printed Shaky — and the misspelling is a small, useful tell that survives in citations to this day. This file keeps the printed form in the title field because that is what the primary source says, and the correct spelling everywhere else.

Darrach's piece is not a press release. It is a reported feature by a writer who spent time in the lab, and the passage everyone forgets is the one where he watches Shakey work through the box-and-ramp problem and recognises his own reasoning in it: "all at once I got gooseflesh." He names the project leads — Charles Rosen, Nils Nilsson, Bertram Raphael — describes the robot following baseboards and inspecting doorways, and speculates about lunar sample collection. It also, in its closing movement, worries at length about a post-work society and about what happens if machines get control. Set beside most 1970 technology journalism it is unusually thoughtful, and set beside most 2026 technology journalism it is unusually unhurried.

The article is remembered for two sentences, both attributed to Minsky, neither of which is about Shakey.

The forecast

> "In from three to eight years we will have a machine with the general > intelligence of an average human being. I mean a machine that will be able to > read Shakespeare, grease a car, play office politics, tell a joke, have a > fight. At that point the machine will begin to educate itself with fantastic > speed. In a few months it will be at genius level and a few months after that > its powers will be incalculable."

Two claims, not one, and they are usually collapsed. The first is a capability forecast with a date: general human-level intelligence, by a machine, no later than 20 November 1978. The second is a takeoff claim: that the first such machine self-improves to genius in months and past legibility shortly after. The second half is a recognisable ancestor of the recursive-improvement argument that superintelligence-2014 spends a book formalising, forty-four years earlier, in a general-interest magazine, in two sentences, with no argument attached.

The warning

> "Once the computers got control, we might never get it back. We would > survive at their sufferance. If we're lucky, they might decide to keep us as > pets."

Same article, same attribution, same page range. This is the more-quoted of the two in AI-safety writing and the less-quoted in AI-hype writing, and the selection is itself a finding: the article is cited by both camps and neither mentions that the other's quote is on the same page, from the same speaker, in the same breath. A canon entry that carried only the optimism half would be reproducing exactly that.

The robot

The gap between the sentence and its occasion is most of the entry's value, so Shakey needs stating plainly.

Shakey ran at SRI from 1966 to 1972, announced 15 November 1966, funded by DARPA on a proposal submitted April 1964 — roughly $300,000 for an initial eighteen months, renewed. It was about five feet tall, wheeled, and had no arms: it was kept deliberately simple mechanically, so that the research could be about perception and planning rather than actuation. It pushed boxes by driving into them. Its "brain" was not aboard. It ran on an SDS-940 with 64K of memory, upgraded around 1969 to a PDP-10 with 192K, in a separate room, in Fortran and Lisp, communicating by radio after an early tethered phase. It kept a quirk from that tethered phase — code to unwind the cable that made it pivot 360 degrees for no visible reason long after the cable was gone.

Its famous task was the monkey-and-banana problem in physical form: get a box off a platform by pushing a ramp into position, climbing it, and pushing the box off. Shakey solved it. It took days, with many false starts and repetitions. The A\* search algorithm — Hart, Nilsson and Raphael, 1968 — came out of this work and is still the thing most computer science undergraduates meet first; dijkstra-1959 in this canon records that descent from the other end.

And DARPA terminated the funding in 1972, before the earliest due date of the prediction the article carried. The stated reasons were the Mansfield Amendment's requirement that Defense Department research be directly militarily relevant, and impatience for applications. The robot sat in SRI's lobby for years and is now in a display case at the Computer History Museum in Mountain View.

So: the artefact that occasioned a claim of human-level machine intelligence within three to eight years was a room-sized computer driving an armless cart that needed days to push a box off a platform, and its funding was cancelled within two.

What the field around it thought

Nils Nilsson, who ran the project, recorded that the Life headline "provoked much teasing," and noted that Fortune's July 1971 treatment was more circumspect — describing Shakey and its Stanford and MIT cousins as "actually all experiments in solving problems through the techniques of scene analysis." That is the correction being made, in a general-interest business magazine, eight months later, by people who had read the Life piece and did not like it.

2. On the kind, and on the creator field

prediction survives the attribution problem, and here is the argument.

The prediction kind exists, per jobs/canon.md, to assemble a base rate for how wrong AI forecasting runs, and it requires four things: the claim in its own words, the date made, the date due, and what actually happened. All four are available here regardless of who spoke. The sentence exists in print with a fixed date. Its due window is arithmetic. What happened is checkable. A prediction that was published as a prediction, believed as a prediction, cited as a prediction for fifty-six years, and used to set expectations about a research field, is a prediction — the question of whose mouth it came out of is a separate and also-important question, but it does not change the kind.

The alternatives were considered and rejected. moment would make the entry about the Life article's effect on public perception, which is real but which lighthill-1973 already covers better for the funding half and which the article probably did not have much of anyway (see §6). interpretation would make it about the reading of AI progress the quote encodes, which is bitter- lesson-2019's and moravecs-paradox-1988's territory. The gradeable dated claim is the thing, and prediction is where gradeable dated claims go.

The creator field is where the dispute lives. Other entries in canon/ name an author and move on. This one cannot, because naming Minsky flatly would smuggle the contested half of the entry into a metadata field where no reader will see it argued, and naming Darrach flatly would assert a fabrication that is not established either. So the field carries the structure: Brad Darrach, writing in Life; the forecast is attributed in the article to Marvin Minsky, who later denied making it. Anything shorter is a position, and I do not have the evidence for a position. Whatever the canon's index does with a long creator string, a truncated string is a better failure than a false one.

descends_from holds two ids and neither is a citation.

dartmouth-1956 is the strong edge and the canon has already drawn it: that entry names simon-1965, minsky-1970 and kurzweil-2005 as the series it heads, and observes that its own forecast — a significant advance in one summer — is the zeroth data point in this base rate. Minsky co-proposed Dartmouth in 1955 as a Harvard junior fellow. The horizon in the 1955 proposal was two months; the horizon here is three to eight years; the horizon in kurzweil-2005 is twenty-four. The series is the point.

2001-hal-1968 is the weaker edge and needs its grounds stated rather than assumed. Minsky was the scientific consultant to Kubrick and Clarke on 2001: A Space Odyssey, spent time on the set, and answered what he described as innumerable questions about what a computer would be able to do thirty-three years later. HAL is the concrete image of the machine this quote promises, and the film put it in front of a mass audience two and a half years before the Life article put a number on it. The edge is also where the disavowal lives: the primary source for Minsky calling the quote made up is a 1997 retrospective volume about 2001, published in the year of HAL's fictional birth. The descent is biographical and documentary rather than by citation, and I am marking it as such rather than dressing it up.

Two ids I wanted and did not invent: simon-1965 (Simon's twenty-year horizon for "any work a man can do," 1965, due 1985) and moore-1965, both listed in proposals.md and neither written. When they exist, this file's descends_from should probably gain simon-1965; the three of them are one argument.

One id I deliberately did not put in descends_from, though the temptation is strong: perceptrons-1969. Minsky is its co-author and it is dated one year before this. But the relation is not descent — it is contradiction, and it gets its own section below.

3. The provenance problem

This is the centre of the file. It is worked in the order the evidence should be weighed, and the verdict is at the end rather than the beginning.

3.1 What is established

The sentence was printed. Life, 20 November 1970, pp. 58B, 58D, 66, 68, under Darrach's byline, attributed to Minsky in direct quotation. Quote Investigator verified the wording and pagination; multiple independent secondary sources reproduce it identically. There is no textual-variant problem here of the kind perceptrons-1969 has with its four editions. The quote is not garbled, misdated, or reconstructed. It is a stable, correctly cited artefact. Everything below is about whether it is a transcript.

3.2 What Minsky said afterwards

On the forecast. In an interview published in **David G. Stork (ed.), HAL's Legacy: 2001's Computer as Dream and Reality (MIT Press, 1997), Minsky said Darrach's quote attributed to him was made up**. In the same reflection he gave what he says his actual view was, and it is worth quoting in full because its structure matters:

> "I believed in realism, as summarized by John McCarthy's comment to the > effect that if we worked really hard, we'd have an intelligent system in from > four to four hundred years."

On the pets remark. Frank Rose's Into the Heart of the Mind: An American Quest for Artificial Intelligence, p. 171, records that Minsky "has denied ever making" it. (Quote Investigator dates that book to 1985; it is usually catalogued 1984. I did not resolve the discrepancy and it does not bear on anything.) So the denial covers both halves of the article, not just the embarrassing forecast — which slightly strengthens it, since a man selectively disowning a bad prediction would have no reason to disown a warning that later read as prescient.

3.3 What weakens the denial

The recollection is late. Twenty-seven years elapsed between the article and the HAL's Legacy interview. The earliest source anyone has traced for Minsky endorsing McCarthy's "realism" framing is a 1981 article — still eleven years after the fact, and eight years into a period in which the prediction had visibly failed.

The supporting quote does not check out cleanly. McCarthy's attested versions of the line are "I believe it'll take five to 500 years" (1977) and domestic robots "anywhere from five to 500 years away" (1978). Minsky's recollection renders it four to four hundred. Both are the same rhetorical move — a range so wide it refuses to be a forecast — but the numbers are wrong, and they are wrong in a recollection being offered as evidence of what he actually believed. A man correcting the record on a misquotation is himself misquoting. That does not make him a liar; it makes the memory a memory. It should reduce confidence in the precision of "made up" without much reducing confidence in the substance.

And the disposition is his own, in print, under his own name. Three years before the Life article, in **Marvin Minsky, Computation: Finite and Infinite Machines (Prentice-Hall, 1967), ch. 1 §1.0, p. 2** — a technical monograph, not an interview, verified by Quote Investigator against scans of the original:

> "Within a generation, I am convinced, few compartments of intellect will > remain outside the machine's realm — the problems of creating 'artificial > intelligence' will be substantially solved."

(A singular "the problem" variant circulates widely; the scanned text reads "problems.")

That is not four to four hundred years. A generation is twenty to thirty. It is longer than three to eight and vastly shorter than McCarthy's refusal, and it is unambiguously Minsky's, because he wrote it and signed it. The denial is therefore about a number, not about a stance. Anyone who reads "made up" as "Minsky was a cautious forecaster whom the press caricatured" is reading past p. 2 of his own book.

The institutional record points the same way. The MIT AI Lab that Minsky co-founded and co-ran produced, in July 1966, AI Memo 100 — Seymour Papert's Summer Vision Project, which set out to build a significant part of a visual system, up to naming objects against a known vocabulary, using summer undergraduates, in one summer. Sixty years later computer vision is a field with an annual conference and unsolved problems. Whatever the calibration culture of that lab was in the late sixties, it was not four-to-four-hundred.

3.4 What strengthens the denial

The reporter's method was independently challenged, in detail, in a different domain, and the challenge is documented.

In August 1975 Bobby Fischer filed suit against Darrach and twenty-six other parties, seeking $20 million on eight counts including defamation, invasion of privacy, and interference with contractual relations, over Darrach's book Bobby Fischer vs. the Rest of the World (1974). The case came before the US District Court for the Central District of California and was dismissed in 1977 by Judge Matt Byrne on the ground that Fischer's brief — which Fischer, characteristically, prepared himself without counsel — was improperly prepared. Frank Brady's account is in CHESS, September 1977, pp. 364–365. Edward Winter, whose reconstruction of the episode is the careful one, found no evidence that Fischer ever appealed.

The dismissal establishes nothing about the merits. It is a procedural outcome against a pro se litigant, and any entry that treats it as a finding against Darrach is doing the thing this canon exists to refuse. What matters is narrower, older than the suit, and independent of it: the specific methodological charge, made by named people, about a named technique.

That last concession is the useful one, because it is from a friendly witness and it grants the technique. The charge against Darrach is not that he invented facts. It is that he wrote reconstructed speech in quotation marks and presented it as heard. That is precisely the kind of error that would produce a sentence like the Life quote: substantively in the neighbourhood of what a subject believes, punched up into something quotable, set in direct quotation, and never said in those words.

Note the direction of the evidence carefully. It is about a different book, on a different subject, four years later, and it is a set of accusations rather than findings. It cannot establish that the Minsky quote was fabricated. What it does is remove the default. Normally a direct quotation in a major magazine carries a presumption, and a subject's later denial has to overcome it. Here the presumption is weakened by the reporter's own documented practice, so the denial and the quotation stand closer to even than they otherwise would.

3.5 The verdict

Stated as a verdict, with its confidence, because a reading that cites this entry needs to know how hard it can lean.

High confidence: the sentence was printed, on that date, in that magazine, attributed to Minsky.

High confidence: Minsky later disowned it and disowned the pets line with it.

Moderate confidence: he did not say it in those words. The reporter's technique is independently attested and it is exactly the technique that would generate this artefact.

Low confidence, and the honest answer: whether the underlying view was his. The 1967 book puts a substantially-solved AGI inside one generation over his own signature. The Summer Vision Project puts his lab's calibration on the record. Against that, three-to-eight is short but not alien, and the takeoff clause is a thing he is on record believing in other forms. The most defensible reading is that Darrach compressed a real and quite optimistic disposition into a specific number Minsky had not committed to, and that the number then did all the work for half a century.

What a reading may therefore say: that the quote was published in Life in 1970, that it is universally attributed to Minsky, that Minsky denied it, and that the denial is credible but late. What a reading may not say: "Minsky predicted human-level AI in three to eight years," flat, as fact. That sentence is in Wikiquote, unqualified and without a dispute note, which is where most people who repeat it get it; being right about this is a small thing a reading can do that the reference works are not doing.

4. Minsky in 1969 and Minsky in 1970

One year before the Life article, Minsky published perceptrons-1969 with Seymour Papert — the book whose reputation is for having been too pessimistic about neural networks, and which that entry establishes was pessimistic in a paragraph of explicitly marked speculation rather than in its theorems.

So the same man, in the same eighteen months, is the canon's standing example of excessive pessimism about an architecture and its standing example of excessive optimism about a timeline. That is not a contradiction and noticing why is worth a paragraph, because the resolution is the useful part.

The two positions share a premise: that the route to intelligence runs through explicit symbolic representation and search, of the kind Shakey was doing, and not through statistical learning in layered networks. Given that premise, the 1969 book's dismissal of the multilayer extension as "sterile" and the 1970 forecast of three to eight years are the same view — the connectionist road is a dead end, and the symbolic road is nearly finished. Both halves were wrong, and they were wrong in a correlated way. alexnet-2012 and scaling-laws-2020 record the road that actually worked, and it is the one 1969 dismissed.

That correlation is the transferable lesson and it is not about Minsky. A short timeline and a confident dismissal of a rival approach are usually the same claim wearing two hats: "the paradigm I am in is the one that works, and it is nearly done." A reading that meets a principal doing both at once — promising a near horizon while writing off a competing method — has seen this shape before, and 1969–70 is where the canon holds it.

5. Why a reading would cite it

5.1 It is the ancestor of a rule the project already runs on

LENSES.md says, in the discipline section, in the project's own words:

> "A person's shipped work deposits; a person's forecast does not." > Practitioners' accounts of what they used and what it did are primary sources > and deposit under the normal rules. "A principal's claims — Musk, Altman, > Hassabis, Amodei, LeCun: forecasts, positioning, promises — are canon entries > of kind prediction, dated and graded against what happened, never > deposits. Without the line, the ledger fills with celebrity opinion, which is > abundant, loud, and free, and drowns the filings and papers that are the > actual record."

That rule is a rule about 1970. Minsky in Life is the archetype of the thing it excludes: the most senior possible figure, in the most widely read possible venue, saying a thing that moved expectations and funded nothing, with no result attached. The canon's job here is not to make the project distrust principals. It is to give the rule a worked case, so that a reading applying it can point at something instead of asserting a policy.

5.2 The live occasion, on the day this file was written

The 17 August 2026 midday reading recorded, under Safety and alignment, OpenAI's president publishing The Defender's Window — and labelled it explicitly:

> Labelled as forecast, not evidence: its claims about what threat actors > will be able to do by the end of this month, and that AI will end up > structurally favouring defenders, are a principal's predictions and deposit > nothing here — **though they are dated and gradeable, which makes them canon > material of kind prediction.**

That is the reading correctly separating a principal's forecast from the thirteen security issues he actually found and fixed, which did deposit. It is the 1970 problem, handled correctly, in the project's own log, on the day this entry was researched. What the reading did not have was anything to cite for why the separation is worth making. This file is that.

The same reading also supplies the sharper occasion. Under Robotics and embodiment it records SoftBank putting $200m into Gravis Robotics, an ETH Zurich spinout that retrofits existing excavators for autonomous digging and loading — and flags it honestly as capital rather than a capability demonstration. In August 2026, the frontier of machines doing physical work in unstructured environments is bolting autonomy onto construction equipment someone else built. Minsky's list of what a human-level machine would do in 1978 began with reading Shakespeare and continued, in the same sentence, to greasing a car. Section 6.2 is what happened to that list.

5.3 When an AGI date is given by a senior figure

proposals.md filed this entry with its occasion already written — "Cite when: an AGI date is given by a senior figure in the field" — and that occasion recurs constantly in the project's window. Frontier-lab principals in 2026 routinely quote horizons of two to five years for human-level or transformative AI; amodei-2024-loving-grace carries the one instance the canon has verified properly, with the primary text, the hedges, the self-grade and the independent grade all present.

The useful thing this entry adds to that occasion is not "they were wrong before, so they are wrong now." That is the lazy citation and it is bad reasoning: the field of 1970 could not do arithmetic on a photograph, and the field of 2026 ships systems that pass professional examinations. The base rate is real but it is not a refutation, and a reading that swings it as one has stopped looking at evidence, which is the whole job.

The useful thing is narrower and harder to argue with. When a senior figure gives a short horizon, ask what has to be true for the number rather than the direction to be right, and write down the date. Minsky's quoted number was wrong by a factor that is still growing; his direction — that this would eventually work, that the objections were not objections in principle — is looking better every year. Those are separable and the canon should keep them separate. turing-1950 in this canon is the case where the direction was right and the date (a machine passing a five-minute imitation game against 70% of interrogators, by 2000) was wrong by roughly a generation. Here the direction may be right and the date is wrong by more than half a century. The base rate this kind assembles is about dates. It is not about whether the thing happens.

5.4 When a forecaster's own account of their record is in play

This is the occasion kurzweil-2005 opened and this entry completes, and it is where the two files should be cited together.

jobs/canon.md requires a prediction entry to record the predictor's self-grade beside an independent one and to say they differ, and it names Kurzweil as the case in point: he scores his hit rate at 86%, outside recounts put it between 7% and 42%. That is a generous regrade — the claim stays, the scoring loosens.

Minsky's move is different in kind and worse for a base rate. He does not score the claim badly or well. He removes it from the dataset. And his grounds for removing it are partly sound, which is what makes it instructive rather than merely cynical.

amodei-2024-loving-grace records a third: the horizon moves by roughly the elapsed time, so the claim is never scored at all because it is never due.

Three predictors, three post-hoc manoeuvres, and none of them produces a scorable data point from the predictor's own account:

This is the entry's main contribution to the kind. jobs/canon.md says a base rate assembled from self-assessments would be worthless. This table is why — not because forecasters lie, but because there are at least three ordinary, non-dishonest ways for a person to relate to their own failed forecast, and all three destroy the record in different directions. A reading that wants to know how AI forecasting actually runs has to grade the claims itself, from the dated text, and never from the forecaster's retrospective.

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

6.1 The claim, the dates, the outcome

The claim (as printed): "In from three to eight years we will have a machine with the general intelligence of an average human being," with a takeoff to genius level in months thereafter.

Date made: 20 November 1970 (publication date; the interview is undated).

Date due: earliest 20 November 1973, latest 20 November 1978.

What happened by the earliest due date. DARPA cancelled Shakey in 1972. Sir James Lighthill delivered his survey to the Science Research Council in 1973 — lighthill-1973 in this canon — reporting that in no part of the field had the discoveries produced the major impact promised, and singling out the combinatorial explosion as the reason the toy-world results did not scale. The claim was not merely unmet at its earliest due date; the specific research programme that occasioned it had lost its funding, and the field's most visible external review had concluded that its central technique did not generalise.

What happened by the latest due date. ARPA retreated from open-ended AI funding from 1974. The speech-understanding programme was cancelled in 1976. The British funding consequence documented in lighthill-1973 ran roughly a decade. No machine existed in 1978 with the general intelligence of an average human being, and no machine existed in 1978 that any participant would have claimed was close.

What happened by 2026. Section 6.2.

Grade: a miss, unambiguous, at both ends of the range. It is the largest single miss in this canon by proportion — the claim was due at latest in 1978, and forty-eight years past its own outer bound the honest answer to "has it happened" is still contested rather than yes.

6.2 The rubric, graded item by item

Here is the part worth the entry. Minsky — or Darrach — did not stop at "general intelligence of an average human being." He gave five concrete tasks, and they turn out to be a usable rubric, which almost no forecast of that era supplied. Graded honestly against August 2026:

"read Shakespeare" — behaviourally met, and the criterion dissolved. Language models discuss, gloss, translate, pastiche and analyse Shakespeare competently. Whether that is reading is the argument turing-1950 set running and eliza-1966 complicated and nobody has closed. What can be said flatly is that the observable behaviour Minsky named is present. The criterion did not survive contact with the systems that met it, which is a fact about the criterion, and it is the single most common pattern in this canon's graded predictions.

"tell a joke" — behaviourally met, quality contested. Same structure. The production of joke-shaped text on demand is solved. Whether the models are funny is a live and rather good argument, and it is not the argument the 1970 sentence was making.

"grease a car" — not met. Unstructured physical manipulation with tools, under a vehicle, on a surface nobody prepared. In 2026 the state of the art in embodied work is scripted warehouse tasks, teleoperated demonstrations, and — as of the day this file was written — retrofitting autonomy onto excavators for digging and stockpile work. Dexterous hands and cross-environment generalisation are named by practitioners as the open problems, not the closing ones. Fifty-six years, and the mechanic's job is untouched.

"play office politics" — not met, and not close. This requires persistent modelling of a shifting coalition of people with private interests, over months, with real stakes, and with the ability to be wrong about someone and update. Nothing does this. The nearest 2026 artefacts are a human using a model to produce a document arguing that his client bears 0% of the fault for a fatal explosion — a person doing politics with a tool, which is the opposite of the claim.

"have a fight" — not met under either reading. Physically, no. As argument — holding a position against a determined interlocutor across a long exchange — the frontier systems are specifically trained toward agreement, so the capability is not merely absent but engineered against.

Two of five, fifty-six years late, and the two are the language ones.

And the split is not random. It is Moravec's paradox, stated by accident, eighteen years early. moravecs-paradox-1988 records the observation that the things evolution spent a billion years on — perception, locomotion, manipulation, social reading — are computationally expensive and hard to automate, while abstract reasoning, which is evolutionarily recent, is cheap. Minsky's list happens to contain both classes: reading and joking are recent and symbolic; greasing, fighting and office politics are ancient and embodied. The recent ones fell. The ancient ones did not. The 1970 sentence therefore contains, unwittingly and in a general-interest magazine, a cleaner prospective statement of Moravec's paradox than Moravec's own retrospective one — the prediction is wrong and the ordering inside it is right.

That is a genuinely useful thing for a reading to be able to cite, and it is not in the secondary literature, which quotes the first clause and drops the list. When a 2026 principal names concrete tasks a system will do by a date, the list is worth more than the date, and it should be written down verbatim, because in fifty years the list will grade itself.

6.3 What the article got right

The pets paragraph. "Once the computers got control, we might never get it back. We would survive at their sufferance." Whoever composed that sentence, its structure — that the danger is loss of the ability to correct rather than malice — is the structure of the modern control problem. superintelligence-2014 is a book-length version of it; concrete-problems-2016 is the version that turned it into research agendas. It is forty-four years early, it is in the right shape, and it is in Life.

The takeoff clause is early and is not obviously graded yet. "The machine will begin to educate itself with fantastic speed" is recursive self-improvement, stated in 1970, before Good's 1965 intelligence-explosion paper had any currency outside a small circle. It is undue — its antecedent ("at that point") has not occurred — so it cannot be scored. A reading should not score it. An unfired conditional is not a wrong prediction, and treating it as one is the mirror-image error of treating an unfalsifiable claim as confirmed.

The article's non-quoted content was more careful than its reputation. Darrach's own reporting — the false starts, the days per task, the machine following baseboards — is accurate. His speculation about a post-work society and about superintelligence risk is thoughtful. The piece is remembered as a hype artefact and is largely a competent feature with two sentences in it that ate the rest.

7. Commonly misused as

Not required for prediction. Included because this quote is one of the most misused artefacts in AI's history, and getting a reading to handle it correctly is most of the entry's practical value.

"Minsky predicted human-level AI in three to eight years."

Stated flat, this is not established. See §3. The correct form is: a 1970 Life article attributed the prediction to him; he later denied saying it; the denial is late but the reporter's method was independently challenged. This is longer and it is the only version that survives being checked. Wikiquote carries the quote with no dispute note, which is where most of the flat version comes from.

"The overclaim caused the AI winter."

No. lighthill-1973 is strict about this and the strictness transfers. The American funding retreat ran through the Mansfield Amendment, ARPA's move from open-ended to mission-directed funding from 1974, and the cancellation of the speech-understanding programme in 1976. The British one ran through the Lighthill report and the SRC's response. A general-interest magazine feature is not one of those mechanisms, and I found no evidence in any source reached that the Life article influenced a single funding decision. Nilsson records that its headline "provoked much teasing," which is the honest magnitude of its effect inside the field. There is also a live scholarly dispute, recorded in perceptrons-1969 and lighthill-1973, over whether a first American winter is visible in the funding record at all.

The general form of the error is worth naming because a reading will meet it weekly: treating a piece of press coverage as a cause of a funding outcome, because the two are adjacent in time and the coverage is more quotable than the appropriations.

"Early AI researchers were reckless; today's are the same."

Two separate errors, and the entry cuts against both.

The 1970 field was not uniformly reckless. John McCarthy's contemporaneous position, attested in his own words, was "five to 500 years" (1977) — a range constructed specifically to communicate that he did not know. That is the same field, the same seniority, the same decade, and it is a model of calibration. The reckless-founders story is assembled by quoting the loud ones.

And the transfer to 2026 does not go through automatically. The systems are different, the empirical base is different, and scaling-laws-2020 gave the current generation of forecasters something the 1970 generation did not have: a measured regularity to extrapolate from. Whether it extrapolates that far is exactly the open question — but "they were wrong then" is not evidence about it.

"Experts are no better than non-experts at predicting AI."

This one is retracted and the retraction is not widely known, so a reading should not repeat the claim.

It comes from Armstrong and Sotala's How We're Predicting AI — or Failing to (2012), which analysed a dataset of 95 AI timeline predictions and reported that predictions cluster at 15 to 25 years out from whenever they are made, "from experts and non-experts alike," with expert predictions indistinguishable from non-expert ones.

AI Impacts subsequently identified an error: a column recording prediction methodology had been read as an expert versus non-expert classification. The "expert authority" category actually captured experts reasoning from unjustified intuition or non-experts deferring to expert judgement — not expert status. Sotala confirmed the account. The corrected analysis finds the groups do differ: AGI researchers put arrival at 10–25 years in 54% of cases, while the "other" group put it at 75+ years in 50% of cases. AGI researchers are systematically more optimistic than everyone else — which is nearly the opposite of the headline claim, and is a more interesting finding.

The clustering result — that predictions bunch 15–25 years out regardless of when they are made — is not what the error touched and stands. Note that the quote in this entry is well outside it. Three to eight years is much shorter than the typical failed forecast. If Minsky said it, he was not making the standard error; he was making a larger one.

"The pets quote shows the founders foresaw alignment."

Half true and worth stating precisely, because asilomar-1975 and concrete-problems-2016 both bear on it. The sentence has the right shape. But it is one paragraph in a magazine, from a speaker who denied it, with no argument attached and no follow-up work — and the founders' generation conspicuously did not build a safety research programme. asilomar-1975 records what a field that took a risk seriously actually did in the same era, in a different discipline, and the contrast is the finding. A prescient sentence is not a research agenda, and citing this one as evidence that AI safety is old flatters the field's history.

8. Sources

Primary — the artefact itself. Brad Darrach, "Meet Shaky, the first electronic person: The fascinating and fearsome reality of a machine with a mind of its own," Life, 20 November 1970, pp. 58B, 58D, 66, 68. I did not read this. No accessible scan was found; Life's Google Books archive did not yield the issue through any route tried, and the surviving copies are in the collectors' market. Every quotation above is taken from secondary sources that agree with each other verbatim and on pagination, principally Quote Investigator (which states it verified the wording) and Wikiquote. The pagination is consistent across sources. This is the entry's largest gap and the reason several judgements above are stated at moderate rather than high confidence.

Primary — Minsky in his own words. Computation: Finite and Infinite Machines (Prentice-Hall, 1967), ch. 1 §1.0, p. 2. Quoted from Quote Investigator's page on the "within a generation" line, which reports verifying it against scans of the original; an Open Philanthropy scan of the page exists at openphilanthropy.org/files/Focus_Areas/AI/Minsky_1967.pdf and I could not extract text from it. Minsky's 1982 reversal — "The AI problem is one of the hardest science has ever undertaken" — from the same page.

The disavowal. David G. Stork (ed.), HAL's Legacy: 2001's Computer as Dream and Reality (MIT Press, 1997), the Stork–Minsky interview "Scientist on the Set." I could not read it — MIT Press returned HTTP 403 for the chapter. It is reported at second hand via Luke Muehlhauser's "What should we learn from past AI forecasts?" (Open Philanthropy, 2016), which is the source for both the "made up" characterisation and the tracing of the McCarthy-endorsement line to a 1981 article. The Muehlhauser report itself returned HTTP 403 at both its original openphilanthropy.org URL and its current home at coefficientgiving.org/research/what-should-we-learn-from-past-ai-forecasts/; its content reached me only through search-result summaries. Two 403s stand between this file and the primary evidence for its central claim. Anyone with library access to HAL's Legacy should read the interview and correct this section.

The pets denial. Frank Rose, Into the Heart of the Mind: An American Quest for Artificial Intelligence, p. 171, via Quote Investigator (quoteinvestigator.com/2017/06/21/robot-pet/), which is also the source for the full Life pagination.

McCarthy's own numbers. 1977 ("five to 500 years") and 1978 ("domestic robots are anywhere from five to 500 years away"), traced in the Muehlhauser report; reached via search summaries, not read at source.

On Darrach's method. Edward Winter, "Brad Darrach and The Dark Side of Bobby Fischer" (chesshistory.com/winter/extra/fischerdarrach.html), read in full. Winter is the source for the suit's filing (August 1975, $20m, eight counts, 27 defendants), the dismissal (1977, Judge Matt Byrne, US District Court, C.D. Cal., improperly prepared brief, Fischer pro se), Brady's account in CHESS September 1977 pp. 364–365, Hochberg in Chess Life & Review May 1975, Saidy in September 1983, and CHESS's own concession about the dialogue. Also Brad Darrach's Wikipedia entry for his biography (1921–1997; Time Inc. from 1945; Time, Life, People, Sports Illustrated), which does not mention the Fischer suit at all.

On Shakey. SRI International's own history page (sri.com/hoi/shakey-the-robot/) for dates (1966–1972, announced 15 November 1966), capabilities, and the note that National Geographic also covered it in November 1970. IEEE Spectrum's milestone piece (spectrum.ieee.org/sri-shakey-robot-honored-as-ieee-milestone) for the initial ~$300k/18-month DARPA funding, the deliberate absence of arms, A\, and the cable-unwinding quirk. New Atlas's fiftieth-anniversary piece (newatlas.com/shakey-robot-sri-fiftieth-anniversary/37668/) for the SDS-940/64K and PDP-10/192K hardware and the "days, with many false starts" figure for the box-and-ramp task. Wikipedia's Shakey the robot* for the personnel list (Rosen, Nilsson, Hart, Raphael, Fikes) and the April 1964 proposal date. Search-result summaries for the 1972 DARPA termination and the Mansfield Amendment rationale.

Nilsson's "provoked much teasing" and the Fortune July 1971 comparison reached me through a search-result summary attributing them to Nils Nilsson, The Quest for Artificial Intelligence (Cambridge, 2010). I did not read Nilsson. His AI Magazine retrospective (ai.stanford.edu/~nilsson/) and SRI's own 8.1MB laboratory history PDF were both retrieved and neither would yield readable text in this environment. Nilsson's is the account that would most improve §1 and §6.3, and it is the obvious next source for whoever revises this.

On forecasting base rates. Stuart Armstrong and Kaj Sotala, "How We're Predicting AI — or Failing to" (2012), via the LessWrong posting (lesswrong.com/posts/27inAjvnMSddvYdqm/) for the 95-prediction dataset and the 15–25-year clustering; the MIRI PDF at intelligence.org/files/PredictingAI.pdf was not read. The correction is AI Impacts, "Error in Armstrong and Sotala 2012" (aiimpacts.org/error-in-armstrong-and-sotala-2012/), read in full, including Sotala's confirmation and the corrected 54%/50% figures.

On the pattern of failed forecasts. Melanie Mitchell, "Why AI is Harder Than We Think" (arXiv:2104.12871), read via its ar5iv HTML rendering, for the four fallacies and the Simon and Minsky-1967 quotations. Papert's Summer Vision Project (MIT AI Memo 100, 7 July 1966) from search summaries and the MIT CSAIL copy; the memo itself was not read.

Encountered and used with a warning. Afflictor.com's excerpt of the Life article (afflictor.com/2015/07/22/all-at-once-i-got-gooseflesh/) is the longest reproduction of Darrach's prose I could reach and is the source for "all at once I got gooseflesh," the Good Humor wagon description, and the lunar-sampling speculation. It is a personal blog and its assertion that Minsky "immediately and vehemently denied the quote" is the blogger's own, unsourced. The word immediately is doing work no source supports — the earliest traced denial is eleven years later — and it is not relied on above.

Encountered and deliberately not used. AIWS.net's history entry on this quote describes Shakey as "Shakey the Dog." Several widely shared timelines of AI history reproduce the Life quote with no attribution note at all. Both are recorded here only as evidence of how degraded the quote's transmission is.

9. On the boundary

This file cites the canon's own dartmouth-1956, 2001-hal-1968, perceptrons-1969, lighthill-1973, moravecs-paradox-1988, kurzweil-2005, amodei-2024-loving-grace, turing-1950, eliza-1966, superintelligence-2014, concrete-problems-2016, asilomar-1975, alexnet-2012, scaling-laws-2020, bitter-lesson-2019 and dijkstra-1959 as related entries, and names simon-1965 and moore-1965 as proposed-but-unwritten. It quotes LENSES.md once, in section 5.1, to establish that the rule this entry is the ancestor of is a rule the project already runs on; it reads LENSES.md and does not write it. It refers to the project's own reading of 17 August 2026 twice, in section 5.2, to name a live occasion — once for the Defender's Window labelling and once for the Gravis Robotics item — and takes nothing from it as evidence.

It deposits nothing in the evidence ledger, grades no model and no vendor, places no needle and no landmark, and makes no claim about the present state of AI beyond what entries already in canon/ establish, except in section 6.2, where five named 1970 capabilities are graded against August 2026. Those grades are judgement, they are stated as judgement, and the two marked "behaviourally met" are marked that way precisely because the stronger claim is the one this canon has no business making.