In the summer of 1956, ten researchers gathered at Dartmouth College with a proposal that reads, in retrospect, with the particular poignancy of someone who is right about the destination and wrong about the distance. They believed that if enough talented people worked on the problem with sufficient focus, the core problems of artificial intelligence could be substantially solved in a single summer.
They did not solve them. But the question of whether they were wrong is considerably more interesting than the usual telling suggests. The summer of 1956 did not fail. It named a field, established its first research program, and set in motion a train of consequences that, seventy years later, has arrived exactly where the proposal said it would. The road was longer than anyone expected. The destination was correct.
The Proposal and the Name
The Dartmouth Summer Research Project on Artificial Intelligence was written in 1955 by John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon as a funding request to the Rockefeller Foundation. It is worth reading in full, because the list of problems the authors proposed to solve reads today less like a research agenda from 1955 and more like a table of contents for the past decade of AI progress.
They proposed to work on automatic computers, programming languages capable of handling abstraction, neural networks, the theory of computation, machine self-improvement, and natural language processing. They estimated that significant progress could be made on all of these in a single summer by a group of carefully selected researchers. On the question of whether the summer would be sufficient: they were wrong by approximately sixty years. On the question of whether the problems were solvable: they were right about all of them.
The proposal was also the moment the field got its name. McCarthy chose the phrase “artificial intelligence” deliberately, in the face of reservations from colleagues who found the framing imprecise or provocative.1 He wanted something memorable enough to attract funding and talent, distinct enough from cybernetics and automata theory to carve out its own institutional space, and ambitious enough to signal that the researchers were after something more than incremental engineering.
He succeeded on all three counts. He could not have known that the name would prove so durable that it would survive every hype cycle, every winter, every redefinition for the next seventy years, accumulating expectations it had not earned and fears that had very little to do with the actual work. “Artificial” implies imitation, something made to resemble the real thing without being it. “Intelligence” implies the kind of general reasoning capacity that humans deploy flexibly across domains. Together they set a standard that no system of 1956 could meet, and the gap between the name and the reality became a recurring source of both disappointment and misunderstanding.
The Room Where It Happened
The people who gathered at Dartmouth that summer were not a random sample of mid-century American science. They were an extraordinary concentration of talent, and their distinct intellectual temperaments shaped the field’s fault lines for decades.
McCarthy himself was a mathematician and logician, precise and combative, interested in formal systems and what could be proven about them. He would go on to invent LISP, the programming language that dominated AI research for thirty years, and to found the Stanford AI Laboratory. His vision of intelligence was fundamentally symbolic: thought as the manipulation of formal structures according to explicit rules.
Minsky was more expansive and more willing to speculate. Where McCarthy wanted proofs, Minsky wanted models. He was interested in how the mind worked and willing to build theories before the data fully supported them. His later work on frames and the society of mind anticipated ideas that would not become mainstream for decades, though his confident predictions about near-term machine intelligence would contribute to the eventual backlash. He was one of those figures whose vision outran his timeline by just enough to damage his reputation among people who confused the timeline for the vision.
Shannon, already famous for founding information theory with his 1948 A Mathematical Theory of Communication, was the elder statesman of the group. He was already somewhat skeptical that the summer’s ambitions were achievable on the proposed timeline. His presence lent the project credibility. His skepticism, largely private at the time, was prescient. Shannon spent his later years building mechanical mice that could navigate mazes and unicycles that could balance themselves, apparently more interested in elegant demonstrations of specific capabilities than in grand claims about general intelligence.
Newell and Simon arrived with something most of the other attendees did not have: working software. Their presence, and what they brought with them, changed the terms of the conversation.
The Logic Theorist and the Symbolic Path
Allen Newell and Herbert Simon had been working at the RAND Corporation and Carnegie Mellon on a question that seemed almost presumptuous: could a computer prove mathematical theorems? Not by exhaustive search, not by brute force, but by something resembling the heuristic reasoning a mathematician actually uses, working backwards from a goal, recognizing useful intermediate steps, pursuing promising paths and abandoning unproductive ones.
The Logic Theorist, demonstrated in 1956, proved 38 of the first 52 theorems in Whitehead and Russell’s Principia Mathematica, the foundational text of symbolic logic.2 This was not a parlor trick. It was a working system that took formal symbolic inputs, applied transformation rules, and produced valid proofs, a machine reasoning about mathematics in a way that at least superficially resembled how a human mathematician reasons.
The Logic Theorist established the symbolic AI paradigm that would dominate the field for the next three decades. The core idea is both powerful and, in retrospect, limited: represent knowledge as symbols, represent reasoning as rules for manipulating those symbols, and build systems that apply rules to symbols to solve problems. Expert systems, natural language parsers, game-playing programs, automated planners, the entire first generation of practical AI applications was built on this foundation.
The power of the symbolic approach is that it is interpretable. You can examine a symbolic system and understand exactly why it reached a conclusion. The rules are explicit. The reasoning is traceable. For applications where explainability matters, medical diagnosis, legal reasoning, safety-critical systems, this is not a trivial advantage, and it remains relevant today.
The limit of the symbolic approach is that someone has to write the rules. The world is vastly more complex than any set of explicit rules can capture, and the brittleness of symbolic systems, their tendency to fail when they encounter situations their rules do not cover, became increasingly apparent as the field’s ambitions grew. But in 1956, the limit was not yet visible. The symbolic path looked like the path.
The Other Path
Almost simultaneously, and largely unnoticed by the symbolic camp, a twenty-nine-year-old psychologist at the Cornell Aeronautical Laboratory named Frank Rosenblatt was building something entirely different.
Rosenblatt was not primarily a mathematician or a computer scientist. He was interested in how biological brains learn, and he thought the answer had something to do with the physical organization of neurons and the way connections between them are strengthened or weakened by experience. In 1957, he introduced the perceptron: a mathematical model of a neuron that could learn to classify inputs by adjusting the weights assigned to each input based on whether its previous classifications had been correct.
The mechanism is simple enough to state in a sentence. The perceptron receives a set of inputs, multiplies each by a learned weight, sums the results, and produces an output based on whether the sum exceeds a threshold. If the output is wrong, the weights are adjusted, up for inputs that should have contributed more, down for inputs that contributed too much. Do this enough times with enough examples, and the perceptron learns to classify correctly, without anyone telling it the rules.
This is learning from experience rather than learning from explicit instruction. Nobody tells the perceptron how to classify. It figures it out from examples, by adjusting its internal parameters until its outputs match the desired results. The difference from the symbolic approach is not merely technical. It is philosophical. The symbolic approach assumes that intelligence can be encoded. The perceptron approach assumes that intelligence must be learned.
Rosenblatt’s claims for the perceptron were, characteristically for the era, somewhat exuberant. He suggested it might eventually perceive, recognize, and remember, and that further development might lead to systems capable of walking, talking, and being aware of their own existence.3 This attracted press coverage that made his more cautious colleagues uncomfortable, and it set up the backlash that was coming.
The backlash came from Minsky and Papert, who published Perceptrons in 1969 and demonstrated mathematically that a single-layer perceptron could not solve certain classes of problems, most famously the XOR function. This was true and important and also, in a critical way, incomplete: it applied specifically to single-layer perceptrons, not to multi-layer networks. The mathematical tools needed to train multi-layer networks efficiently, backpropagation, the algorithm for computing how each weight in a deep network should be adjusted, existed in embryonic form but had not been developed into a practical method. That algorithm was worked out rigorously by Rumelhart, Hinton, and Williams in 1986, nearly thirty years later.
The publication of Perceptrons contributed substantially to the first AI winter, cutting off funding for connectionist research for over a decade. Whether Minsky and Papert fully anticipated this consequence is debated. What is not debated is that the most important idea in the history of artificial intelligence was sidelined for a generation by a mathematical critique that was technically correct and strategically premature.
Why the Perceptron Was Right
Here is what the winter obscured: the perceptron was right. Not about everything Rosenblatt claimed for it. About the fundamental approach.
A modern deep neural network is, at its mathematical core, a very large stack of perceptron-like operations. Each layer takes inputs, multiplies them by learned weights, applies a nonlinear function, and passes the result to the next layer. The training process adjusts the weights across all layers simultaneously using backpropagation. Add enough layers, enough parameters, enough training data, and enough compute, and you get systems that recognize faces, translate languages, generate text, diagnose diseases from medical images, and predict the three-dimensional structure of proteins from their amino acid sequences. The last of these, accomplished by DeepMind’s AlphaFold in 2020, solved a problem that had resisted structural biology for fifty years.
The transformer architecture that underlies modern large language models, introduced in the 2017 paper “Attention Is All You Need”, is a sophisticated elaboration of the same foundational insight Rosenblatt had in 1957: that intelligence can emerge from learned numerical weights rather than from explicit symbolic rules, and that the way to build an intelligent system is to show it examples and let it adjust its internal parameters until it gets them right. The specific mechanisms are different. The philosophy is the same.
Rosenblatt died in a boating accident in 1971, on his 43rd birthday, before the vindication of his approach. He did not live to see backpropagation made practical, or the second AI summer, or the deep learning revolution of the 2010s, or the language models of the 2020s. The symbolic researchers won the argument in 1969. The connectionist researchers won the field, definitively, about sixty years after Rosenblatt pointed down the path.
The Hardware Had to Catch Up
The reason the perceptron approach was set aside was not mathematical. It was computational. Training a multi-layer neural network requires performing enormous numbers of arithmetic operations, adjusting enormous numbers of parameters, iterating over enormous numbers of training examples. The computers of 1957 could not do this. The computers of 1987 could barely do it for small networks on simple problems. The computers of 2007, equipped with graphics processing units originally designed to render video game environments, could begin to do it at useful scale. By 2012, a neural network trained on GPUs won the ImageNet visual recognition competition by a margin large enough to end the debate about whether deep learning worked. Everything since has been a consequence of that result.
The Dartmouth researchers could not have foreseen Moore’s Law compounding for sixty years, or the internet generating the training data that large models require, or the specific accident that hardware optimized for matrix arithmetic in video games would turn out to be exactly what neural network training needs. These developments were not predictable in 1956. But they were, in some sense, consequences of the road the Dartmouth summer helped to pave.
The theory was right in 1956. The hardware caught up in 2012. The intervening fifty-six years were not wasted, they produced the symbolic AI tradition, the expert systems industry, decades of research into search and planning and natural language that informed the systems that came after. But the bus, to use a metaphor, had a destination marked on the front from the beginning. The destination was the perceptron, scaled up until it could do everything Rosenblatt said it might, and then some.
We are living in the moment of arrival. Understanding where the journey started, and who was right, and when, and why it took so long, is not merely historical curiosity. It is the context that makes the present moment legible.
Footnotes
-
Shannon himself was skeptical of the name, finding it imprecise. This is worth savoring: the man who had literally demonstrated that logical reasoning could be implemented in electrical circuits, and who had founded the mathematical theory of information, thought the phrase “artificial intelligence” raised more questions than it answered. He was not wrong. The field has been arguing about the definition ever since. ↩
-
One of the Logic Theorist’s proofs was, by some measures, more elegant than the original proof in Principia Mathematica. When Simon sent the result to Bertrand Russell, then in his eighties, Russell wrote back that he was delighted. It is one of the more charming moments in the history of the field, and a reminder that the people who built the foundations were often generous enough to enjoy being surpassed. ↩
-
The New York Times covered the perceptron’s unveiling in 1958 under the headline “New Navy Device Learns By Doing.” The Navy was funding Rosenblatt’s research, which adds a layer of irony to the subsequent history: the foundational technology of modern AI was developed on a military contract, sidelined after an academic critique, and vindicated decades later by researchers with hardware originally designed for rendering video game graphics. The perceptron’s biography is almost novelistically improbable. ↩