← The canon · AItopiaOrAImageddon?

The Dartmouth Summer Research Project on Artificial Intelligence

moment · Proposed by John McCarthy (Dartmouth College), Marvin L. Minsky (Harvard University), Nathaniel Rochester (IBM Corporation) and Claude E. Shannon (Bell Telephone Laboratories); attended by roughly twenty people, of whom three stayed the summer · 1956

Something that happened and changed what people expected next.

Descends from Programming a Computer for Playing Chess, On Computable Numbers, with an Application to the Entscheidungsproblem. Read on: "NEW NAVY DEVICE LEARNS BY DOING; Psychologist Shows Embryo of Computer Designed to Read and Grow Wiser", The Shape of Automation for Men and Management, "Meet Shaky, the first electronic person", "Artificial Intelligence: A General Survey", "Superintelligence: Paths, Dangers, Strategies".

moment is the right kind and proposals.md filed it correctly, but the reason is not the one the filing gives, and getting the reason right is most of this entry's value. The proposal list says the field "got its name from a proposal promising significant progress from ten men in two months," and suggests citing it when "a timeline claim needs its ancestor." Both halves are close enough to be dangerous. The naming is real and is the single most consequential thing that happened. The promise is real and was graded a failure by the man who made it. But the artifact a reading would actually quote is a document dated 31 August 1955 — a year before the moment the id is named for — and the meeting that gives the id its year produced no paper, no proceedings, no agreed programme and no result. What the summer of 1956 produced was a word and a mailing list.

That is still a moment, and it is the canon's founding one. Four other entries already reach for this event as the type specimen of a thing that "visibly happened in public" — clippy-1996, siri-2011 and expert-systems-collapse-1987 all use the identical phrase "Dartmouth, Lighthill, Deep Blue, the LLaMA weights," and logic-theorist-1956 spends a paragraph correcting what people think happened here. The canon has been leaning on this file for a day and a half without having it. But a reading that cites it as the event where AI began will be repeating the thing the scholarship exists to correct, so the sections below are organised around the difference between what the document says, what the meeting did, and what seventy years have made of both.

The tempting re-filing, and why I did not do it. The gradeable content here is a forecast: a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer — made 31 August 1955, due by the end of August 1956, and reported failed by its author. That is a clean prediction, it is the zeroth entry in the series the canon carries as simon-1965, minsky-1970, kurzweil-2005 and amodei-2024-loving-grace, and it is the only one of them whose author graded himself honestly against it. I considered re-filing on that basis and decided against it, because the entry would then be about a 1955 memo and would lose the job proposals.md actually needs it for — being the ancestor at the head of every AI timeline. So the kind stays moment and the prediction is carried inside it, graded to the standard the prediction kind requires: the claim in its own words, the date made, the date due, what happened, and the predictor's own grade alongside an independent one. The value of that grade is that it is the first data point in the base rate, and it is a miss.

Four dates, and the id names the least documented one. The document is dated 31 August 1955. Several accounts give 2 September 1955 as the date it was formally submitted to the Rockefeller Foundation; I could not reconcile the two against an archival record and both are reported below rather than chosen between. The Foundation's grant is dated to late 1955. The workshop itself has no reliable start date in any source I could reach — the plan called for two months from mid-June; the surviving participant record, Ray Solomonoff's notes, begins 22 June and ends with his own last talk on 17 August; Dartmouth's alumni magazine says eight weeks; the Rockefeller Foundation's own seventieth- anniversary retrospective calls it "the five-week 1956 Dartmouth Summer Research Project." A reading that wants to say the workshop is seventy years old this month is on safe ground — the summer of 1956 is not in dispute and Solomonoff's last talk was 17 August — but a reading that gives exact dates should say whose dates they are.

descends_from holds two ids and neither descent is by citation. This needs stating because the canon's standard for the field is a checked citation chain — logic-theorist-1956 earned its single ancestor by my reading a 1954 RAND paper's reference list. The Dartmouth proposal has no reference list. The people it names for prior work are Uttley, Rashevsky and his group, Farley and Clark, Pitts and McCulloch, Holland, von Neumann, Moore, Craik, Lashley and Hebb, and not one of them is a canon entry. So the two ids here are honest about being a different kind of link:

The ancestor this entry most wants is not written. turing-1950 — Computing Machinery and Intelligence — is in proposals.md under interpretation and has no file. It is the reason the four proposers could assume their audience already believed machine intelligence was a coherent research object, and it contains the field's actual first dated forecast (by 2000, five minutes, 30 per cent of interrogators fooled), which is the prediction the Dartmouth one should be graded next to. When that entry exists, this one's descends_from should probably gain it. I have not invented the id in the header, per the rule.

What it is

In the summer of 1955 John McCarthy was a 28-year-old assistant professor of mathematics at Dartmouth College, fresh from co-editing a volume called Automata Studies with Claude Shannon for the Princeton Annals of Mathematics Studies series. He was unhappy with it. His own account, given repeatedly afterwards, is that he was "very much disappointed when most of the papers that we received were, in fact, about automata, and were not in my opinion any contribution to artificial intelligence." He had tried to get the volume renamed Towards Intelligent Automata; Shannon rejected the title as bombast. The lesson McCarthy drew was that the name of a field determines the papers it attracts, and that if he wanted papers about machines thinking, he would have to call the subject something that could not be mistaken for anything else.

On 31 August 1955 he circulated A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence, over four names: J. McCarthy of Dartmouth College, M. L. Minsky of Harvard University, N. Rochester of the IBM Corporation, and C. E. Shannon of Bell Telephone Laboratories. Its first paragraph is the most-quoted paragraph in the history of the field and is worth having in full, because almost every use of it in circulation is a trimmed version that changes what it claims:

> We propose that a 2 month, 10 man study of artificial intelligence be > carried out during the summer of 1956 at Dartmouth College in Hanover, New > Hampshire. The study is to proceed on the basis of the conjecture that every > aspect of learning or any other feature of intelligence can in principle be > so precisely described that a machine can be made to simulate it. An attempt > will be made to find how to make machines use language, form abstractions and > concepts, solve kinds of problems now reserved for humans, and improve > themselves. We think that a significant advance can be made in one or more of > these problems if a carefully selected group of scientists work on it > together for a summer.

Three separate things are doing work in those five sentences and they are usually collapsed into one. The conjecture — that every aspect of intelligence can in principle be so precisely described that a machine can simulate it — is a research posture, hedged with "in principle," and it is not a claim that can come due. The agenda — language, abstraction, problems now reserved for humans, self-improvement — is a list of four hard problems. The forecast — a significant advance in one or more of them, from ten people in a summer — is modest by the standards of what came later and is the only part of the paragraph with a date attached. Most citations of this passage graft the confidence of the third onto the scope of the first.

The body of the proposal breaks the work into seven numbered aspects. In the document's own words and order they are: 1 Automatic Computers, 2 How Can a Computer be Programmed to Use a Language, 3 Neuron Nets, 4 Theory of the Size of a Calculation, 5 Self-Improvement, 6 Abstractions, and 7 Randomness and Creativity. Each is a short paragraph. Section 1 argues that hardware is not the bottleneck:

> If a machine can do a job, then an automatic calculator can be programmed to > simulate the machine. The speeds and memory capacities of present computers > may be insufficient to simulate many of the higher functions of the human > brain, but the major obstacle is not lack of machine capacity, but our > inability to write programs taking full advantage of what we have.

Section 3 is the neural one, and its list of prior workers — "Uttley, Rashevsky and his group, Farley and Clark, Pitts and McCulloch, Minsky, Rochester and Holland, and others" — is the clearest evidence in the document that the proposers knew perfectly well they were not starting from nothing. Section 4 asks for a theory of the complexity of functions in order to have a criterion for efficiency of calculation, which is a request for computational complexity theory a decade before it existed. Section 5 is two sentences long and concedes it has nothing: "Probably a truly intelligent machine will carry out activities which may best be described as self-improvement. Some schemes for doing this have been proposed and are worth further study." Section 7 floats the conjecture that "the difference between creative thinking and unimaginative competent thinking lies in the injection of a some randomness" — the wording is odd, and I flag it as it appears in the Stanford transcription without having checked it against the scan.

After the seven aspects come short individual statements of what each proposer intended to work on. The transcription I was able to read carries the subheadings "Originality in Machine Performance", "The Process of Invention or Discovery" and "The Machine With Randomness", plus an untitled section, and it does not attach names to them. Successive attempts to resolve the attribution against that source returned inconsistent answers, so I am not asserting one; the only individual statement I could tie to its author across independent sources is Shannon's, on the application of information-theory concepts to computing machines and brain models, which discusses von Neumann's and Moore's work on reliable computation from unreliable elements. Separately, "Originality in Machine Performance" is attributed to Rochester in secondary accounts, and Rochester's IBM work simulating Hebbian cell assemblies on the 701 and 704 with John Holland is the obvious fit; a reading that needs the attribution should get it from the scanned original rather than from this file.

The proposal asked the Rockefeller Foundation for $13,500. The estimate is six faculty salaries of $1,200 ($7,200), two graduate-student salaries of $700 ($1,400), eight travel-and-rent allowances averaging $300 ($2,400), secretarial and organisational expense ($850), additional travelling expenses ($600), and contingencies ($550). Those six items sum to $13,000, not the $13,500 the document states. I could not check this against the scanned original — the archive PDF would not yield text — so it may be a transcription slip rather than a 1955 one. It is recorded because it is checkable by anyone with the scan, and because a $500 gap in the founding budget of artificial intelligence is the sort of thing that either is or is not there.

The Foundation awarded $7,500. McCarthy's own retrospective says it "only gave us half the money we asked for," and the halving has propagated into secondary accounts as "$14,000 requested, half awarded"; $7,500 against $13,500 is nearer 56 per cent, and the $14,000 figure does not appear in the document. Appended to the proposal is a distribution list of 47 names — among them John Backus, John Nash, Warren McCulloch, W. Ross Ashby, Donald MacKay, George Miller, R. Duncan Luce, Marcel-Paul Schützenberger, Abraham Robinson and John Kemeny. It is, in effect, the field's first roster, and it is a more accurate description of what the project actually was than the word "conference" is.

What happened in the summer bears almost no resemblance to the plan. McCarthy's plan, as he set it out, had himself, Minsky, MacKay, Solomonoff, Holland and Julian Bigelow present for the whole period; Shannon, Rochester and Selfridge for two weeks at the start and two at the end; and Newell and Simon for two weeks at the end. In the event MacKay and Holland did not come at all. Only three people were there for the whole summer: McCarthy, Minsky and Ray Solomonoff. Shannon and Rochester managed about four weeks, Trenchard More about three, Newell and Simon about two. Solomonoff's records name roughly twenty people who passed through — the three above plus Shannon, Rochester, More, Oliver Selfridge, Julian Bigelow, W. Ross Ashby, Warren McCulloch, Abraham Robinson, Tom Etter, John Nash, David Sayre, Arthur Samuel, Kenneth Shoulders, Alex Bernstein, Herbert Simon, Allen Newell, and the neuropsychologist Peter Milner, identified in the group photograph decades later by Solomonoff's widow Grace Solomonoff. Daily attendance ran between about three and eight people. There was no programme. Participants had the top floor of the Dartmouth mathematics department and talked in the main classroom, mostly about whatever each of them had come already working on.

The one thing everybody remembers being shown was not made there. Newell and Simon came for a few days at the end and described the Logic Theory Machine — the entry logic-theorist-1956 covers it, including the fact that it was not running on a computer that summer and that its famous results are from 1957. McCarthy's word for them, in his own retrospective slides, is "the stars of the show." Alex Bernstein described the chess program he was building at IBM. Minsky proposed a geometry theorem-prover. Trenchard More's notes from the fifth week record that "McCarthy has become increasingly interested in the problem of writing a program that will write programs," which is a fair description of the road that ends at LISP two years later. McCulloch came to argue about whether the brain is a Turing machine. Arthur Samuel — whose checkers work is samuel-checkers-1959, and who had a learning program running in February 1956, before the workshop opened — said afterwards that it was "very interesting, very stimulating, very exciting," which is the kind of thing a person says about a meeting that produced no result.

Two participants' verdicts, both after the fact and both first-hand. McCarthy: "anybody who was there was pretty stubborn about pursuing the ideas that he had before he came, nor was there, as far as I could see, any real exchange of ideas." Solomonoff: "I think McCarthy thought we would all work together and get something really going by the end of the conference" — reported with a laugh at how unrealistic that had been. No proceedings volume was published. No report agreed by the participants exists. The record of the summer of 1956 is individual notes, principally Solomonoff's, which is why a woman going through her late husband's papers is the reason we can now name the man in the photograph.

And yet the naming worked, exactly as intended. McCarthy's stated reason for the phrase was not a definition but a flag. He wanted, as he put it at the 2006 anniversary, "to nail the flag to the mast." His fuller explanation is about a rival field rather than about machines at all: "one of the reasons for inventing the term 'artificial intelligence' was to escape association with 'cybernetics'," he wrote, adding that its concentration on analog feedback seemed misguided and that "I wished to avoid having either to accept Norbert Wiener as a guru or having to argue with him." That sentence is the honest origin of the word this project is named around. It was chosen to mark a break with an existing research community, by a young academic who did not want to fight its founder, and it has been doing definitional work it was never built to do ever since.

Why a reading would cite it

Occasion one: the seventieth anniversary is happening now, and it will be observed with the wrong facts. The workshop ran through the summer of 1956; Solomonoff's last talk was 17 August. A reading taken in the second half of August 2026 sits inside the seventieth-anniversary window of the event itself, and anniversary coverage is a reliable channel for bad history entering the record — the "ten men, two months, and they thought they'd finish it" version, the "birthplace of AI" version, the version where the field predicted human-level machines in a generation. This file exists so that a reading meeting a piece of that coverage can say what the document actually claims, what the meeting actually did, and where the numbers came from, in one sentence, without having to relitigate it. That is the only thing the canon is for: nothing here is evidence, nothing here moves anything, and an anniversary is not a finding.

Occasion two: every timeline claim in the readings has this as its zero point. "Seventy years of AI." "The field has been promising this since the fifties." "The third AI winter." Three canon entries already use Dartmouth as the first item in a list of things the public agreed had happened, and expert-systems-collapse-1987 uses that list to make a point about how visible collapses are compared with quiet ones. When a reading needs to say how long a problem has been open, this is the date it is counting from, and it should be counting from a document dated 1955 whose forecast horizon was eight weeks — not from a mythical conference that set a fifty-year agenda.

Occasion three: it is the head of the prediction base rate. The canon's whole reason for carrying predictions graded is to have a base rate for how wrong AI forecasting runs. proposals.md lists eight of them and the readings will grade more. This is the first, it is short, it is unambiguous, its due date passed seventy years ago, and — uniquely in the series — the person who made it wrote down that it failed and why. That combination makes it the calibration point for the whole kind. When a 2026 forecast is offered with a confident horizon, the useful comparison is not that some forecasts fail; it is that the founding one failed at a horizon of eight weeks, among the best people alive, on a problem they had personally chosen as tractable.

Occasion four: the definitional fights. LENSES.md sets a hard test for the "AI as accelerant" lens — a named model or method actually used in the work, not "AI will revolutionise X" — and that test exists because the word is elastic enough to cover anything. The elasticity is not an accident of later marketing; it is in the word's birth certificate. "Artificial intelligence" was coined as a boundary marker against cybernetics by someone who explicitly did not want to argue about definitions, and the 1955 document defines the field by a list of seven problems, not by a property a system has. When a 2026 dispute turns on whether something "is really AI," the honest historical answer is that the term never carried a definition to be violated. A reading can say that in one clause and move on, which is better than either adjudicating it or pretending the question is new.

Occasion five, and it is a scale anchor rather than an argument. The founding of this field cost $7,500 and involved four proposers, ten invitations and a list of 47 people who might be interested. The readings on this project's own record deal in disclosed figures several orders of magnitude larger — the 16 August 2026 reading records a $63.4bn equity portfolio and a $70bn estimate of off-balance-sheet credit backstops. The concentration lens asks who holds the capability and on what terms; the 1955 answer is that it was held by whoever could get a summer's salary out of a foundation, and that the entire addressable community fit on one page. That contrast is a fact about scale and nothing more. It is not evidence, it does not point up or down, and a reading that uses it to imply a direction is editorialising with a fifty-year-old budget line.

Honest note on the occasion. I read the three digests on disk — 14 August, 15 August midday, 16 August midnight — and none of them cites Dartmouth or would have been improved by citing it. They are about watermarking, datacentre financing, a criminal case, and an equity filing. So this entry is not answering a citation the readings have already needed; it is answering one that proposals.md predicted they would need, and that four other canon files have already reached for rhetorically without a file to point at. That is a weaker warrant than perceptrons-1969 had and I am not going to dress it up. It passes the admission test on occasions one to three, which are real and dated, and I would not have written it on occasion five alone.

What it got right, and what it got wrong

Grading a moment is unusual, but this moment has a written claim inside it with a date on both ends, and the canon's whole doctrine is that an ungraded prediction is an anecdote.

The forecast. Claim made 31 August 1955; due end of August 1956; failed. The words are "We think that a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer." The selected group assembled, the summer passed, and no advance was produced by the meeting. The strongest candidate result of that season — the Logic Theory Machine — was designed and hand-simulated at RAND and Carnegie before the workshop, was described there rather than demonstrated, and published its machine results in 1957. Nothing in the historical record attributes a technical result to the workshop itself.

The self-grade and the independent grade, and they agree, which is the unusual part. McCarthy's own retrospective slides give two reasons the plan failed — "the Rockefeller Foundation only gave us half the money we asked for" and, the one he weights more heavily, "the participants all had their own research agendas" — and concede that his "hope for a breakthrough towards human-level AI was not realized at Dartmouth." Elsewhere he says plainly that not much was really accomplished. What he claims as the outcome is different in kind from what he predicted: "the concept of artificial intelligence as a branch of science," which "inspired many people to pursue AI goals." The independent grade, from the historians, is the same on the technical question and slightly harsher on the framing: the meeting produced no result, published nothing, agreed nothing, and its lasting effects — the name, the network, and the funding lineage that followed — are sociological. The two grades differing is what the prediction doctrine is normally guarding against, and here they do not differ. Record that as the anomaly it is: the founding forecast of AI is the best-behaved item in the base rate precisely because its author refused to regrade it upward.

Right, and it is the whole reason the entry exists: the naming. The proposal's stated purpose in choosing a phrase was to make a field exist by giving it a name that could not be mistaken for automata theory or cybernetics. That worked completely. Seventy years later the word is in statute — the 16 August 2026 reading records a product change driven by the EU AI Act — in company names, in university departments and in the title of this project. No other output of the summer of 1956 comes close. It is also the source of the term's central defect, discussed above: a flag is not a definition.

Wrong, and it is the single most gradeable technical sentence in the document: "the major obstacle is not lack of machine capacity." Section 1 asserts that hardware is adequate and programming is the bottleneck. The canon records the answer at length. alexnet-2012 is a moment whose content is that a decades-old algorithm became decisive when someone ran it on two consumer GPUs; scaling-laws-2020 is the observation that capability is a smooth function of compute, data and parameters over orders of magnitude; and perceptrons-1969 grades a 1988 claim that scaling would not help as the sharpest wrong call in that book. The proposal was not merely wrong about the balance — it was wrong in the direction that made the field's first two decades look like a failure of ideas when a substantial part of it was a shortage of machine. In fairness to the four of them, this is a hard thing to see from 1955, when the machine in question was a JOHNNIAC; and their error is the same one, in mirror image, that perceptrons-1969 grades Minsky and Papert for making thirty-three years later.

Wrong about what a meeting can do. The implicit theory of the proposal is that concentrated collaboration is the limiting resource — McCarthy's own statement of it is "if only we could get everyone who was interested in the subject together to devote time to it and avoid distractions, we could make real progress." What actually happened is that ten specialists in a room continued doing what they had each been doing, which McCarthy described afterwards in almost those words. The finding is not that collaboration is useless; it is that a shared name and a shared mailing list turned out to be worth more than a shared room, and that the durable product of the summer was a distribution list of 47 people.

Right as an agenda, and this is underrated. Take the seven aspects as a prediction of what the field would spend seventy years on and grade them individually.

A note on what this entry must not be blamed for. The confident generation-scale forecasts that the AI winters are usually charged to are later and individual: Simon in 1965, Minsky in 1970, both of them in proposals.md as ungraded candidates. Neither is in the Dartmouth document. Grading the 1955 proposal by the standard of those two is the most common way this entry gets used wrongly, which is the next section.

And the fifty-year check, because someone took it. At the 2006 anniversary conference held at Dartmouth, attendees were polled live by keypad. Asked for the earliest that machines will be able to simulate learning and every other aspect of human intelligence — the 1955 conjecture, put back to the room fifty years on, in nearly the original words — 41 per cent of 123 respondents answered "Never" and another 41 per cent answered "more than 50 years"; 11 per cent said 26–50 years, 2 per cent said 11–25, and 5 per cent said within 10. Sourcing caveat, and it matters: these figures were not published in the conference report; they circulate from the organisers' own data, obtained later by researchers surveying expert opinion, and I could not retrieve either the report or that survey in readable form. Treat the numbers as reported rather than verified until someone reads the primary. What they are useful for, if they hold, is a base rate on the other side: half a century after the founding conjecture, at the founding institution, more than four in ten of the people in the room thought it would never come true — a reminder that expert pessimism is also a forecast with a date on it, and that 2006 was six years before alexnet-2012.

Commonly misused as

The limit kind requires this section; a moment does not. It is here because this entry is misused more consistently than any other proposed item in the canon, and because every misuse below is committed in good faith by people repeating a compressed version they were told.

"AI was born at Dartmouth in 1956." What was born was the name. The research was already running: McCulloch and Pitts had a formal neuron in 1943; Turing published Computing Machinery and Intelligence in 1950; Shannon published the chess paper in 1950 (shannon-chess-1950); Samuel had a checkers program that improved with play in February 1956, before the workshop (samuel-checkers-1959); the Macy conferences on cybernetics had run from 1946 to 1953; Farley and Clark, Rashevsky, Uttley and Ashby were all working on learning machines, and the proposal names most of them. The proposal's own bibliography-by-mention is the best available refutation of the claim its authors are credited with. What 1956 changed is that this work acquired a single label, a shared roster, and a lineage of American funding that flowed to the label — which is a real and large thing, and is not the same thing as the beginning of the research.

"Ten men spent two months and thought they would solve intelligence." The "2 month, 10 man study" is a budget line, and the budget is in the document: six salaries at $1,200 and two at $700. Ten was the number the money bought. It was never ten men for two months in any case — three people were there the whole summer, daily attendance ran three to eight, and the proposal's forecast is explicitly "a significant advance in one or more of these problems," not a solution to any of them. The sentence that people are remembering is a request for eight weeks of summer salary, and the mockery it attracts is aimed at a claim nobody made.

"Dartmouth predicted human-level AI within a generation." It did not. No horizon longer than one summer appears in the document, and no claim about human-level machines appears at all — the conjecture is that intelligence can in principle be described precisely enough to simulate, which is a statement about describability with no date attached. The generation-scale predictions belong to Simon in 1965 ("machines will be capable, within twenty years, of doing any work a man can do") and Minsky in 1970 ("in from three to eight years"), neither of which is in canon/ yet and both of which are the correct targets for that criticism. Attaching them to 1955 launders two individual overclaims into an institutional one and makes the founding document look sillier than it is.

"The proposal defined artificial intelligence." It named it. There is no definition of the term anywhere in the document; there is a list of seven problems and a conjecture about describability. McCarthy's stated purpose was to plant a flag against cybernetics, and he chose a phrase specifically to avoid an argument with Norbert Wiener rather than to demarcate a subject matter. Seventy years of "is that really AI" arguments run on the fact that the word was never asked to do this job. When a reading meets one of those arguments, the useful move is to note that the term is a boundary marker of 1955 vintage and get back to what the system actually does.

"Over-promising at Dartmouth caused the AI winter." The funding history does not run through this document. expert-systems-collapse-1987 covers one collapse in detail and perceptrons-1969 covers the misattribution of the earlier one at length, including the Mansfield Amendment, the 1973 Lighthill report, and the finding that the neural-network research community had largely emptied before the book everyone blames appeared. The Dartmouth proposal asked for $13,500 and made an eight-week claim. Nothing in the winter literature traces a funding decision to it.

And the inverse misuse, which the canon should be equally hard on: "seventy years and it still can't X." The Dartmouth date is a favourite prop for the argument that the field is a permanent failure, usually by counting from 1956 to the present and treating the interval as uniformly barren. The interval is not uniform, and the canon's own entries say so with dates: rosenblatt-perceptron-1958, samuel-checkers-1959, eliza-1966, mycin-1976, deep-blue-1997, watson-jeopardy-2011, alexnet-2012, alphago-move-37-2016, transformer-2017, alphafold2-2020, chatgpt-2022. Using a founding date to argue for stagnation is the same move as using it to argue for inevitability, run in the opposite direction, and a reading that takes the needle seriously should decline both. Neither the length of the interval nor the height of the original ambition tells you anything about what happened in the last twelve hours, which is the only question the reading is actually asked.

Sources

Read directly.

Consulted second-hand, or through search summaries, and flagged as such.

On the boundary. This file cites the canon's own shannon-chess-1950, turing-halting-1936, logic-theorist-1956, samuel-checkers-1959, rosenblatt-perceptron-1958, perceptrons-1969, expert-systems-collapse-1987, alexnet-2012, scaling-laws-2020, transformer-2017, chatgpt-2022, rlhf-christiano-2017, mycin-1976, eliza-1966, deep-blue-1997, watson-jeopardy-2011, alphago-move-37-2016 and alphafold2-2020 as related entries, and names turing-1950, simon-1965, minsky-1970 and lighthill-1973 as proposed-but-unwritten. It refers to the project's own digests of 14–16 August 2026 only to record, in section 2, that none of them needed this entry, and to take two disclosed dollar figures from the 16 August reading as a scale contrast. It deposits nothing in the evidence ledger, grades 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.