The History of AI: Waves of Enthusiasm and Two Winters of Disappointment

A promise on paper

On 31 August 1955 four researchers, John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon, wrote a proposal for a summer study at Dartmouth College. Its title carried the term "artificial intelligence". The proposal rested on the conjecture that every aspect of learning, or any other feature of intelligence, can in principle be described so precisely that a machine can simulate it. It asked for two months and ten researchers, and the workshop took place in summer 1956. An entire field began with a bold guess and a small budget.

Two springs and two winters

Confidence soon grew that machine translation and natural conversation were close. But in 1966 a US advisory committee (ALPAC) concluded that machine translation was more expensive, slower and less accurate than human translation, and the National Research Council ended its support after spending about $20 million. In 1973 Sir James Lighthill's report on AI research appeared in Britain, and a sharp contraction followed; research survived only at a few universities until funding returned in 1983 through the Alvey programme. In 1974 DARPA cancelled a contract worth three million dollars a year for speech understanding research at Carnegie Mellon University.

Historians call this pattern an "AI winter" and date the first to roughly 1974–1980. The usual explanation is that promises outran what the computers and data of the time could deliver, so the reckoning was harsh.

Enthusiasm returned in the 1980s with "expert systems", programmes that encode experts' rules as conditions and conclusions, and with Japan's Fifth Generation project, for which the Ministry of International Trade and Industry (MITI) set aside about $850 million in 1981. Its goals had not been met by 1991, and the project ended on 1 June 1992. In 1987 the market for specialised Lisp machines collapsed. It was an industry worth about half a billion dollars, undercut by cheaper general-purpose workstations that could run the same language. The source dates the second winter to roughly 1987–2000, and historians differ on the exact boundaries of both winters.

What changed?

In May 1997 IBM's Deep Blue beat the world champion Garry Kasparov 3.5 to 2.5, after losing to him 4–2 the year before. It relied on raw computing power, evaluating about 200 million chess positions a second. It was a brilliant result in a narrow domain, not a step towards general intelligence.

The deeper shift came in 2012. In the ImageNet competition the AlexNet network reached a top-5 error of 15.3%, beating its closest rivals by more than 10.8 percentage points, and it was trained on two graphics cards in about five or six days. Three ingredients had come together: large labelled datasets, relatively cheap hardware, and deep networks. Five years later, a study in Nature showed that a deep network trained on 129,450 clinical images classified skin cancer on a par with 21 dermatologists in the test. The authors themselves cautioned that validation in real clinical practice was still needed, and that some image labels had not been confirmed by biopsy. Large language models followed, which we explain in the article "How Do Large Language Models (LLMs) Like ChatGPT Work?".

And tomorrow?

Round dates tempt us to predict. The closest thing we have is a survey conducted in October 2023 of 2,778 researchers who had published at leading venues. Their aggregate forecast gave a 50% chance of "high-level machine intelligence" by 2047, and a 50% chance of full automation of labour by 2116. The figures shift with the wording of the question: with fixed years the date came out roughly twice as far away as with fixed probabilities. The authors admit they do not know what explains the large gap between the two forecasts. These are estimates, not promises, and knowledge of the unseen belongs to God alone.

Three practical lessons

1. When you hear a promise of imminent transformation, ask for the evidence: on what data was the system tested, and by what measure?
2. Distinguish outstanding performance in a narrow task from a flexible, general ability.
3. Do not base an institutional decision on a forecast date, but on what is demonstrated today, while staying ready to adjust course.

The winters taught us that patience combined with taking the proper means serves better than either dazzlement or denial, and that science advances when we measure honestly what we know and admit what we do not.