The Golden Years

From 1956 to the early 1970s, a few dozen researchers built programs that proved theorems, learned checkers, chatted like a therapist and steered a robot. You will see what each one really did, why it dazzled people, and why the confident forecasts of the time fell so far short.

In January 1956, months before the Dartmouth workshop gave the field its name, Herbert Simon walked into a class at the Carnegie Institute of Technology in Pittsburgh and, as he later recalled, told his students that over Christmas he and Allen Newell had invented a thinking machine. It was a bold way to describe a program that did not yet run on any computer. Within a year it did, and it proved theorems from one of the most forbidding books in mathematics. That mixture of genuine achievement and breathtaking confidence defined the next eighteen years: the golden years of artificial intelligence.

Machines that reasoned

The program was the Logic Theorist, written by Newell, Simon and the programmer Cliff Shaw at the RAND Corporation and Carnegie Tech, and shown off at Dartmouth in the summer of 1956 (see Dartmouth 1956). Its task was to prove theorems from chapter 2 of Whitehead and Russell’s Principia Mathematica. It managed 38 of the first 52, and for one of them it found a shorter proof than the authors had printed.

The important idea was not logic itself but how the program looked for a proof. Trying every possible chain of steps would take forever, so the Logic Theorist used heuristics: rules of thumb such as “work backwards from the goal” or “try a substitution that makes this look like something already proved”. Search guided by heuristics became the signature method of what is now called symbolic AI.

Newell, Shaw and Simon generalised the trick in the General Problem Solver (GPS, first described in 1957–59). GPS used means–ends analysis: compare where you are with where you want to be, pick an action that shrinks the difference, and repeat, setting up sub-goals when an action cannot yet be applied. It could solve tidy puzzles such as the Tower of Hanoi. What it could not do was cope with the messy, open-ended problems people face every day — a limitation that would matter a great deal later.

New tools: LISP and learning to play

Symbolic AI needed a language that could juggle symbols and lists as easily as other languages juggled numbers. In 1958 John McCarthy, who had coined the term “artificial intelligence”, began designing LISP at MIT, publishing its theory in 1960. LISP programs could treat code as data and build structures on the fly. It remained the field’s main language for some thirty years, and ideas it pioneered, such as automatic garbage collection, are in most modern languages.

Meanwhile, at IBM, Arthur Samuel had been working on a program that played checkers (draughts). Rather than hand-coding every judgement, Samuel let the program adjust the weights in its scoring function by playing thousands of games, including games against itself, until it played better than he did. His 1959 paper, “Some Studies in Machine Learning Using the Game of Checkers”, popularised the phrase machine learning.

Popular myth The definition so often credited to Samuel — that machine learning gives computers the ability to learn without being explicitly programmed — does not appear in that form in the 1959 paper; it is a later paraphrase of his idea. Likewise, the program’s 1962 win in a single game against the player Robert Nealey was widely reported as if checkers had been conquered. It had not: strong players still beat it.

Learning machines: the perceptron

Not everyone thought intelligence was best built from explicit rules. Frank Rosenblatt, a psychologist at the Cornell Aeronautical Laboratory, took inspiration from the brain. His perceptron (1958) was a simple artificial neural network: inputs fed through adjustable weights into a unit that output yes or no. When it got an answer wrong, its weights were nudged towards the right answer. Rosenblatt proved that if a problem could be solved this way at all, the procedure would find a solution.

He then built it in hardware. The Mark I Perceptron (1960) had a 20-by-20 grid of photocells for an eye and motor-driven potentiometers as its adjustable weights, and it learned to tell simple shapes and letters apart. The press was not restrained. A 1958 New York Times report, based on a US Navy briefing, described the embryo of a computer the Navy expected would be able to “walk, talk, see, write, reproduce itself and be conscious of its existence”.

Oliver Selfridge offered a related vision in “Pandemonium” (presented in 1958, published in 1959): a recognition system made of many little “demons”, each shouting when it spotted its feature — a vertical stroke, a curve — while higher demons listened and the loudest decision won. Layers of simple detectors feeding more abstract ones is, loosely, how today’s image-recognition networks are organised.

Talking, seeing and moving

By the mid-1960s, AI programs had begun to converse, understand and act. Tap through the stages of the most famous of them to see how little was going on inside.

That is ELIZA, written by Joseph Weizenbaum at MIT between 1964 and 1966. Its best-known script, DOCTOR, imitated a non-directive psychotherapist, a clever choice because such a therapist can answer almost anything with a question. ELIZA was the ancestor of every chatbot, and it taught an uncomfortable lesson. People confided in it and insisted it understood them, even when they knew how it worked. This tendency to read understanding into fluent text is now called the ELIZA effect. Weizenbaum was alarmed, and spent much of the rest of his career warning about it.

“…extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people.”Joseph Weizenbaum, Computer Power and Human Reason (1976)

SHRDLU, built by Terry Winograd at MIT around 1968–70, went much deeper, inside a deliberately tiny world. It controlled a simulated arm above a table of coloured blocks and pyramids, and could follow commands such as “pick up a big red block”, answer questions about what it had done and why, and ask what an ambiguous “it” referred to. Within its microworld it really did connect language to meaning. Outside it, it knew nothing.

Shakey, developed at the Stanford Research Institute (SRI) from 1966 to 1972, was the first mobile robot to reason about its own actions. It wobbled through specially built rooms, using a camera and bump sensors to plan routes and push boxes. Shakey’s lasting gifts were software. To plan paths, Peter Hart, Nils Nilsson and Bertram Raphael published the A* search algorithm in 1968: it explores the most promising routes first, using an estimate of the remaining distance, and is still used in map apps and video games. The STRIPS planning system (1971) also came out of the project.

At Stanford, DENDRAL (begun in 1965 by Edward Feigenbaum, Bruce Buchanan, Joshua Lederberg and Carl Djerassi) took a different route. Instead of general cleverness, it packed in chemists’ specialist knowledge to work out molecular structures from mass-spectrometry data, and it performed like an expert. It was the seed of the expert systems boom.

Labs, money and bold promises

The work clustered in a handful of places: MIT (Marvin Minsky and, until 1962, McCarthy), Stanford (McCarthy’s new lab and SRI nearby), Carnegie Tech, later Carnegie Mellon (Newell and Simon), and in Britain, Donald Michie’s group at Edinburgh, whose robot Freddy assembled simple wooden toys. Much of the American money came from the Pentagon’s Advanced Research Projects Agency (ARPA, later DARPA). In 1963 ARPA gave MIT about $2.2 million for Project MAC, which housed the AI group, and under J.C.R. Licklider its funding philosophy was to back talented people rather than tightly defined projects. Few strings, few deadlines.

Generous funding and real progress bred extravagant forecasts. Researchers were not merely optimistic in private; they said it in print.

  1. 1958Simon and Newell: a chess champion within ten yearsIn a 1958 paper they predicted that within ten years a digital computer would be the world’s chess champion. It took until 1997.
  2. 1960Simon: any work a man can do“Machines will be capable, within twenty years, of doing any work that a man can do.” First printed in 1960, the line is usually cited from its 1965 reprint in The Shape of Automation for Men and Management.
  3. 1967Minsky: within a generationIn Computation: Finite and Infinite Machines he wrote that within a generation “the problems of creating ‘artificial intelligence’ will be substantially solved.”
  4. 1970Minsky in Life magazineAn article quoted him predicting a machine with the general intelligence of an average human in three to eight years. Minsky later said he had been misquoted, so treat this one with care.

Why were such clever people so wrong? Part of the answer is that the early successes came in tidy, closed domains such as logic, puzzles, board games and blocks worlds, where the rules are fully known. It was natural to assume that scaling up was mostly a matter of faster machines and more effort. It was not. The real world is open-ended, ambiguous and full of common sense nobody writes down.

Other views
  • The sceptics at the time: the philosopher Hubert Dreyfus argued in a 1965 RAND paper, “Alchemy and Artificial Intelligence”, that human expertise rests on intuition and bodily know-how that rule-following programs could never capture. Many researchers dismissed him; history vindicated part of his critique.
  • The defenders: others point out that the golden years produced lasting tools — search algorithms, LISP, planning, time-sharing computers — and that predictions were partly a way to win funding in a competitive system.
  • Weizenbaum’s view: the question was not only whether machines could do these things but whether they should be trusted with human judgement at all.

What the golden years got right — and wrong

This guide’s framing Two traditions were born side by side in this period. One built intelligence from explicit symbols and heuristic search (Logic Theorist, GPS, SHRDLU, DENDRAL). The other tried to learn it from examples (Samuel’s checkers player, the perceptron). For the next two decades the first tradition dominated, and the second was nearly abandoned. Today’s AI is overwhelmingly the second — but it still uses the first tradition’s search and planning ideas, from A* to the game-tree search inside AlphaGo.

Quick check

What made the Logic Theorist more than brute force?

Heuristic search — rules of thumb that prune the search — was the key idea, not raw speed or learning from data.

Why did ELIZA seem so understanding?

ELIZA matched keywords, swapped pronouns and filled templates. People supplied the understanding themselves — the ELIZA effect.

Which golden-years invention is still in everyday use in navigation software?

A* (1968) finds shortest paths efficiently using an estimate of the remaining distance. It is still a staple of route planning and games.

// Pause and reflect

  1. The golden-years forecasts came from brilliant people. What made their successes in closed “microworlds” such a poor guide to the open world?
  2. The ELIZA effect was noticed in 1966. Where do you see it today, when people talk with modern chatbots?
  3. Should researchers make public predictions to win funding and attention? What are the costs when they turn out wrong?

Your notes are saved privately in this browser.

What happened next

The bills for those promises came due. A damning report on machine translation, a book exposing the perceptron’s limits, and a British government review of the whole field turned optimism into retrenchment. Continue with The First AI Winter.

Key terms on this page

Sources & further reading