Complete History of AI — Part 1: The Pioneer Era (1940–1974)

Summary

This is Part 1 of 6 in the series Complete History of AI: From Turing's Machines to Generative Models. The period between 1940 and 1974 established the theoretical foundations of artificial intelligence, produced the first machines that learned from experience, and culminated in a funding collapse that nearly ended the field before it matured. Understanding this period is not a nostalgic exercise — it is understanding why AI's cycles of hype and disappointment have roots deeper than any product launch.

Context

The context that made AI possible was not technological — it was the Second World War. The urgent need to process information at a scale and speed impossible for humans created the first serious computing laboratories in the world. At Bletchley Park, England, mathematicians and linguists worked under life-or-death pressure to decipher enemy communications. It was in this environment that Alan Turing developed not only practical machines but the theoretical questions that would define the field for decades. In 1943, simultaneously, Warren McCulloch and Walter Pitts published the first mathematical model of artificial neurons — demonstrating that networks of simple on/off elements could, in theory, compute any logical function. Two independent works, in different countries, on the same fundamental question: can machines think?

— 1950-1956: The Question and the Foundation

In October 1950, Turing published "Computing Machinery and Intelligence" in the journal Mind, proposing what would become known as the Turing Test: rather than defining what thinking is — a philosophically thorny question — he proposed an operational test. If a human interrogator could not distinguish a machine's responses from a human's in a text conversation, it would be reasonable to attribute thinking to the machine. The formulation was deliberately pragmatic: it shifted the debate from metaphysics to engineering, establishing a measurable objective. In 1949, Canadian psychologist Donald Hebb had published "The Organization of Behavior," proposing that neurons that fire together strengthen their connections — a principle that decades later would become central to learning algorithms. The field still had no name. That changed in the summer of 1956, when John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon organized the Dartmouth College workshop in Hanover, New Hampshire. It was McCarthy who coined the term "artificial intelligence" during the event. The central proposal was audacious: "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." Optimism was immense. Many present believed machines with human-level intelligence would be developed within a generation.

— 1957-1969: First Machines That Learn

In 1958, two milestones defined the field's direction. McCarthy, now at MIT, developed the LISP programming language — designed not for numerical calculation but for symbolic manipulation, list processing, and knowledge representation. LISP would dominate AI research for decades. That same year, Frank Rosenblatt, a psychologist at the Cornell Aeronautical Laboratory in Buffalo, New York, developed the Perceptron — funded by the US Navy. The Perceptron Mark I learned to recognize visual patterns through photocells connected to adjustable weights. Rosenblatt publicly demonstrated it by teaching it to distinguish cards marked on the left from those marked on the right, proclaiming it "the first machine capable of having an original idea." In 1961, the Unimate, developed by George Devol and Joseph Engelberger, was installed on a General Motors assembly line in New Jersey — the world's first industrial robot. It did not learn, but it automated. In 1964-1966, Weizenbaum at MIT created ELIZA, a program that simulated a psychotherapist through simple pattern matching. The effect was surprising: users attributed genuine understanding to the program and shared deep personal thoughts with it — what Weizenbaum called the "ELIZA effect." In 1969, the Stanford Research Institute presented Shakey — the first mobile robot capable of reasoning about its own actions.

— 1969-1974: Limits Become Visible

In 1969, Minsky and Papert published "Perceptrons," mathematically demonstrating that single-layer neural networks could not learn logical functions like XOR without additional layers. The work was devastating to neural network enthusiasm — and was widely, incorrectly, interpreted as a condemnation of the entire connectionist paradigm. What was a limitation of a specific architecture was read as a fundamental limitation. At the same time, the systems of the era faced real barriers: computers too slow, insufficient memory, natural language processing that stumbled on human language ambiguities. Robots like Shakey only worked in perfectly structured, controlled environments. In 1973, British mathematician Sir James Lighthill delivered to the British government a report that would be read as a verdict. The document harshly criticized AI's progress, arguing that research had produced none of the practical applications it had promised across three domains. The Lighthill Report led to immediate cuts in AI research funding in the UK. DARPA in the US did the same. Researchers left the field. Universities closed AI labs. The First AI Winter had begun.

Comparative Analysis

The central tension of the 1940-1974 period is between the depth of the theoretical foundations established and the shallowness of the practical implementations that followed. Turing and McCulloch built conceptually correct frameworks — artificial neural networks, universal computation, learning through weight adjustment. But available hardware was decades ahead of what was needed for these ideas to work at scale. The result was a generation of researchers making genuinely believed promises that were impossible to fulfill with available resources. The Lighthill Report was not wrong about the facts — it was wrong about the conclusion. The problem was not that AI was impossible. It was that promises had outpaced infrastructure by a margin that the funding of the era could not bridge.

Analysis

Three patterns from the 1940-1974 period persist across all subsequent AI cycles. First: genuine theoretical advances generate expectations of practical application faster than infrastructure supports. Rosenblatt's Perceptron was mathematically correct — computational power to train it at useful scale was missing. Second: the scientific community tends to interpret implementation limitations as fundamental limitations — the Minsky-Papert critique of the Perceptron was read as condemning neural networks in general, when it was a critique of a specific architecture. Third: funding cycles follow promise cycles, not real progress — when promises do not materialize on schedule, cuts are abrupt regardless of the accumulated value of foundational research.

Synthesis

The First AI Winter did not end the field — it redistributed who remained in it. The researchers who survived the 1974 cuts were those who believed in the foundations, not in short-term promises. Geoffrey Hinton, who would later be called "the godfather of deep learning," was a psychology student at Cambridge when the winter arrived — and continued working on neural networks when almost no one else did. This persistence, multiplied across a handful of researchers in different labs, would produce the next revolution. The lesson the 1940-1974 period offers for 2026 is precise: the foundations established by Turing, McCulloch, Pitts, and Rosenblatt took 30 years to find the infrastructure that would make them practical. The foundations established by the 2017 transformer may take less — but the gap between what is theoretically possible and what is practically achievable remains larger than any optimistic prediction suggests. Next: Part 2 — Winters and Renaissance (1974–1993)