Where AI's words came from

Turing proposes a test for thinking machines. Alan Turing introduces the imitation game in his 1950 paper "Computing Machinery and Intelligence," written while he was at the University of Manchester.

McCarthy names the field. John McCarthy, then an assistant professor of mathematics at Dartmouth, picks the name while organizing a group to study thinking machines. He chooses it partly for its neutrality, avoiding cybernetics. The workshop itself runs in the summer of 1956.

Von Neumann's singularity. Stanislaw Ulam recalls a conversation with John von Neumann about accelerating technology "approaching some essential singularity," the first known discussion of a singularity in technological progress. Vernor Vinge popularizes the term in a 1983 op-ed in Omni magazine and his 1993 article "The Coming Technological Singularity."

Arthur Samuel coins it. Samuel, a pioneer of computer gaming and AI, coins the term. His checkers-playing program was among the world's first successful self-learning programs.

Stanford formalizes expert systems. The Stanford Heuristic Programming Project, led by Edward Feigenbaum, formally introduces expert systems around 1965. Early ones like Dendral identify unknown organic molecules. They are built through the 1970s and proliferate in the 1980s.

ELIZA and the ELIZA effect. Joseph Weizenbaum builds ELIZA at MIT. Its best-known script, DOCTOR, simulates a Rogerian psychotherapist. Some users, including his secretary, attribute human-like feelings to it, which comes to be called the ELIZA effect.

First, a planning term. In the 1980s and 1990s, "generative AI planning" names AI planning systems that generate sequences of actions toward a goal. The sense of models that generate text and images comes later, and ChatGPT popularizes it by 2023.

Researchers warn of an AI winter. At the annual AAAI meeting, Roger Schank and Marvin Minsky warn that enthusiasm for AI has spiraled out of control. They describe a chain reaction, like a "nuclear winter," that leads to funding cuts and ends serious research. Three years later the AI industry begins to collapse.

Dechter brings it to machine learning. Rina Dechter introduces "deep learning" to machine learning in a paper on constraint-satisfaction problems. Aizenberg and colleagues bring it to neural networks in 2000.

Mauldin's paper on chatterbots. Michael Mauldin puts the word in the title of his 1994 paper "ChatterBots, TinyMuds, and the Turing Test: Entering the Loebner Prize Competition." "Chatbot" is the later, shorter form.

Gubrud's early use. Mark Gubrud uses the term in a discussion of fully automated military production and operations. Shane Legg and Ben Goertzel re-introduce and popularize "AGI" around 2002.

Deep networks trained one layer at a time. Publications by Geoff Hinton, Ruslan Salakhutdinov, Simon Osindero and Yee-Whye Teh develop deep belief networks for generative modeling. Each is trained one restricted Boltzmann machine at a time, stacking each new one on the last.

An early title use. Thorsten Brants, Jeffrey Dean and three co-authors present "Large Language Models in Machine Translation" at EMNLP-CoNLL 2007 in Prague.

AlexNet wins ImageNet. Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton at the University of Toronto build AlexNet. It wins the ImageNet challenge with a top-5 error rate of 15.3%, more than 10.8 points better than the runner-up.

Two networks trained against each other. Ian Goodfellow and seven co-authors propose training two models at once: a generator that captures the data distribution, and a discriminator.

"Attention Is All You Need". Ashish Vaswani and seven co-authors propose the Transformer, built only on attention and dropping recurrence and convolutions entirely.

BERT, then GPT-2. BERT arrives in 2018 and quickly becomes "ubiquitous." GPT-1 also appears in 2018, but GPT-2 in 2019 draws wide attention after OpenAI says it first judged the model too powerful to release.

Stanford names foundation models. Stanford's Center for Research on Foundation Models calls models like BERT, DALL-E and GPT-3, trained on broad data at scale and adaptable to many tasks, "foundation models."

ChatGPT takes generative AI to the public. OpenAI releases ChatGPT on November 30, 2022, a generative AI chatbot built on large language models. It reaches 100 million users in two months and accelerates the AI boom.