The Dartmouth proposal is dated 31 August 1955, a year before the event it proposes. It was written by four people, it asked for a two-month, ten-man study, and its central promise was that a significant advance could be made in one summer by a carefully chosen group. It did not predict human-level machines in a decade. That prediction came later, from some of the same people, and it has been retro-fitted onto the proposal ever since.
The document, and its date
The full title is A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence. It is a funding application to the Rockefeller Foundation, dated 31 August 1955, and it runs to a handful of typed pages plus individual statements from each author about what they personally intended to work on.
The four signatories, with their affiliations as given on the document:
| Author | Description |
|---|---|
| John McCarthy | Dartmouth College. The organiser, and the person who chose the phrase artificial intelligence. |
| Marvin L. Minsky | Harvard University. Later the founder, with McCarthy, of what became the MIT AI Laboratory. |
| Nathaniel Rochester | IBM Corporation. Designer of the IBM 701 and the reason the project had access to serious hardware. |
| Claude E. Shannon | Bell Telephone Laboratories. By 1955 already the author of the 1948 information theory papers and a 1950 paper on computer chess. |
The money is worth stating precisely, because it is small and the smallness is informative. The proposal asks the Rockefeller Foundation for $13,500 — salaries, travel and expenses for ten people for two months. The Foundation funded it at less than the amount requested, and the workshop went ahead at Dartmouth in the summer of 1956. This was not a moonshot. It was a summer school with a grant that would not cover a fortnight of a modern research programme.
The sentence the field was founded on
One sentence in the proposal does almost all of the work, and it is worth reading with the qualifiers intact:
“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.”
Three qualifiers do the heavy lifting. It is a conjecture, offered as a working assumption rather than a finding. It holds in principle, which says nothing about tractability. And the machine simulates the feature rather than possessing it — which sidesteps, deliberately, the entire question of whether the machine understands anything. Seventy years of philosophical argument about AI is in large part an argument about whether that sidestep is legitimate.
The second load-bearing sentence is the ask: the authors state that they think a significant advance can be made in one or more of these problems if a carefully selected group of scientists works on it together for a summer. That is the claim that was wrong. It was wrong by decades on several of the seven problems, and by more than that on at least one.
The seven problems it listed
The proposal breaks the work into seven headed sections. They are worth reading as a snapshot of what looked hard in 1955, because the ranking is not the one you would guess.
- Automatic computers. The observation that existing machines were fast enough in principle and that the gap was in programming, not in hardware. On the hardware of 1955 this was optimistic; the general point about programming being the bottleneck was right for about thirty years.
- How can a computer be programmed to use a language. Framed around the idea that much human thought consists of manipulating words according to rules of inference and rules of conjecture. This is the symbolic programme in one line.
- Neuron nets. Present from the start, in the founding document of what is usually described as the symbolic tradition. The section notes that some progress had been made on arrangements of hypothetical neurons and asks how to make more.
- Theory of the size of a calculation. A request for a measure of the efficiency of calculation. This is a call for computational complexity theory, several years before the field existed, and it is the one item on the list where the eventual answer was both achieved and devastating: many of the problems on the rest of the list are intractable in the sense this section was groping towards.
- Self-improvement. The argument that a truly intelligent machine would improve itself, and that this should be studied. Every modern discussion of recursive self-improvement is a descendant of this paragraph.
- Abstractions. How a machine could form abstractions from sensory and other data — which is, in retrospect, exactly what representation learning turned out to be.
- Randomness and creativity. The conjecture that creative thinking differs from unimaginative competent thinking by the injection of some randomness, guided by intuition. Anyone who has changed a sampling temperature has an opinion about this.
Two of the seven — abstractions and neuron nets — describe what eventually worked. Both were near the bottom of the field’s priority list for the following thirty years.
Why McCarthy chose the words
“Artificial intelligence” was chosen for the proposal, and it is the first use of the phrase as the name of a research field. McCarthy said in later interviews — this is a paraphrase of his recollection rather than a quotation — that he needed a term that would not commit the project to any of the existing schools, and in particular that would not place it under cybernetics, which would have meant dealing with Norbert Wiener. He also said more than once that he was not especially happy with the phrase; it simply had the advantage of being new.
The alternatives that were live in 1955 tell you what the field was escaping from: cybernetics, automata studies, complex information processing, machine intelligence. Newell and Simon carried on calling their own work complex information processing for years afterwards and did not adopt McCarthy’s term. The name won because it was on the funding document that got funded.
What actually happened in summer 1956
Calling it a conference is the most common factual error about the event. It was a summer research project of roughly six to eight weeks, with no fixed programme and no fixed attendance. People arrived, stayed for a while, and left. Attendance over the period included McCarthy, Minsky, Rochester, Shannon, Ray Solomonoff, Oliver Selfridge, Trenchard More, Arthur Samuel, Allen Newell and Herbert Simon. Solomonoff is generally reported to have been there for most of the period; several others came for days.
There were no proceedings, no papers submitted and no agreed conclusions. The one substantial result presented was not produced at Dartmouth at all: Newell, Simon and J. C. Shaw brought the Logic Theorist, which was already running on a machine at RAND. It proved theorems in propositional logic by heuristic search, and it is reasonably described as the first program that did something people were willing to call reasoning. It worked through the theorems of chapter 2 of Whitehead and Russell’s Principia Mathematica, proving a majority of the first fifty-odd, and for one of them found a proof shorter than the book’s.
The commonly quoted figure for the Logic Theorist is 38 of the first 52 theorems in chapter 2. That number is repeated consistently in the secondary literature and is very probably right; it is given here as a reported figure rather than as something checked against the original program listing.
The workshop’s real product was a social one. It put the ten or so people who would run American AI research for the next twenty years in the same building, and it gave what they were doing a name that funders could put on a line item. That is a substantial achievement and it is not the achievement the proposal promised.
Four things the proposal did not say
- It did not predict human-level AI within a decade. That prediction is real, but it belongs to Simon and Minsky individually in 1957, 1965 and 1967, and is covered on the page about predictions that were confidently wrong. The proposal’s own overreach is smaller and more specific: a significant advance in a summer.
- It did not rule out neural networks. Neuron nets are one of the seven headed sections, written by people who subsequently became identified with the opposing camp. The split between symbolic and connectionist AI hardened later.
- It did not found the field. Turing’s 1950 paper, McCulloch and Pitts in 1943, Shannon’s 1950 chess paper and Wiener’s cybernetics all predate it. What Dartmouth founded was the name and the network.
- It did not define intelligence. The conjecture is explicitly about description and simulation, and the authors were careful about that. Pages that quote the conjecture as a definition of machine intelligence are quoting it against its own grammar.
Top comments (0)