History of AI

AI is not a sudden invention. It is the result of seventy years of research, marked by cycles of hope and disappointment, up to a recent breakthrough that sped everything up.

The idea of a machine able to reason long predates computing. It is rooted in the formalisation of logic in the 19th century (Boolean algebra, 1854) and then in the founding work of Alan Turing. In 1936, Turing defined a theoretical model of universal computation (the "Turing machine"), which established that a single machine can, in principle, carry out any calculation.

In 1950, in Computing Machinery and Intelligence, he proposed the famous "Turing test" (which he called the imitation game): a machine can be called "intelligent" if a human interlocutor cannot tell it apart from a human in a conversation.

The term "artificial intelligence" was coined in 1956 at the Dartmouth summer conference, organised by John McCarthy, Marvin Minsky, Claude Shannon and Nathaniel Rochester. The starting hypothesis was ambitious: every aspect of intelligence can in principle be described precisely enough for a machine to simulate it. This was the starting point of symbolic AI, which seeks to reproduce reasoning through explicit logical rules.

The first decades were optimistic. In 1957, Frank Rosenblatt invented the perceptron, the first artificial neural network able to learn. But in 1969, Minsky and Papert showed in Perceptrons the mathematical limits of these simple models, which held back connectionist research for a long time.

The history of AI is made of cycles of enthusiasm and disillusion. After unfulfilled promises, funding collapsed: this was the first "AI winter" in the 1970s (accelerated by the 1973 Lighthill report in the United Kingdom).

The 1980s saw a revival with expert systems, programs encoding specialists' knowledge as rules (for example MYCIN, for medical diagnosis). But these costly and rigid systems disappointed in turn: this was the second winter, at the end of the 1980s.

A decisive turning point came in 1986, when David Rumelhart, Geoffrey Hinton and Ronald Williams popularised backpropagation, the algorithm that makes it possible to train multi-layer neural networks. But three factors had to come together for this approach to take off:

The founding moment was 2012: the AlexNet network (Krizhevsky, Sutskever and Hinton) won the ImageNet image recognition competition by a spectacular margin. This was the start of the deep learning revolution. In 2016, DeepMind's AlphaGo program beat the world Go champion, Lee Sedol, in a game long considered out of reach for machines.

In 2017, a Google team published the paper Attention Is All You Need, which introduced the Transformer architecture. Built on a so-called "attention" mechanism, it processes language far more efficiently and in parallel. It is the technical building block behind all of today's large language models. In its wake, OpenAI launched the GPT series (2018-2020) and Google the BERT model.

On 30 November 2022, OpenAI opened ChatGPT to the public. Adoption was meteoric and pushed generative AI into mass culture. GPT-4 followed (2023), then a frantic race between labs.

Two developments marked 2025-2026: the arrival of reasoning models (which "think" longer before answering, such as OpenAI's o series or DeepSeek-R1) and the shift towards agents, systems able to chain actions autonomously.

As a sign of the field's scientific maturity, the 2024 Nobel Prizes rewarded two pieces of AI-related work: John Hopfield and Geoffrey Hinton (physics) for neural networks, and Demis Hassabis and John Jumper (chemistry) for AlphaFold, the protein structure prediction system.