Detailed · 30 events
A History of Artificial Intelligence
1950s
null0 (Wikimedia Commons, via Flickr) · CC BY-SA 2.0 · Commons ↗ A funding proposal dated 31 August 1955 and signed by John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon put the phrase 'artificial intelligence' into an official research document; the workshop it asked for ran at Dartmouth College in New Hampshire the following summer. The '2 month, 10 man' figure was the plan, not a roster — attendance was fluid and daily sessions drew three to eight people. Its historical weight lies less in any specific technical result than in establishing the field's own name and the researcher network that would carry it for decades.
John C. Hay, Albert E. Murray — Mark I Perceptron Operators' Manual, Cornell Aeronautical Laboratory (Wikimedia Commons) · Public domain · Commons ↗ In July 1958 the US Office of Naval Research held a press conference for the perceptron of Frank Rosenblatt of the Cornell Aeronautical Laboratory: a learning device that weights its inputs, applies a threshold, and adjusts the weights whenever it errs. The demonstration was a simulation running on an IBM 704 that learned, in about fifty trials, to tell cards marked on the left from cards marked on the right. The theory appeared the same year in Psychological Review 65(6), 386–408. The Mark I Perceptron—a separate, purpose-built electromechanical machine whose sensory layer was an array of 400 photocells in a 20×20 grid—was assembled at the same laboratory. It marks the start of the neural-network lineage and remains the structural skeleton beneath the MLP, the convolutional network, and modern deep learning.
Questions this page answers
- What actually ran at the 1958 perceptron demonstration?
- The demonstration was a simulation on an IBM 704 that learned, in about fifty trials, to tell cards marked on the left from cards marked on the right. The Mark I Perceptron, the purpose-built machine with 400 photocells, was separate electromechanical hardware assembled at the same laboratory.
- Where did Frank Rosenblatt work?
- He was a researcher at the Cornell Aeronautical Laboratory. The press conference was held by the US Office of Naval Research, and his theoretical paper appeared the same year in Psychological Review 65(6), 386-408.
1960s
Joseph Weizenbaum (via elizagen.org), Wikimedia Commons · CC0 · Commons ↗ Joseph Weizenbaum wrote ELIZA in MAD-SLIP at MIT and published it in Communications of the ACM in January 1966. ELIZA was the engine; what it said was governed by a swappable 'script', and the famous one—DOCTOR—imitated a Rogerian psychotherapist. It picked keywords out of the input and rephrased them back using decomposition and reassembly rules. In the paper Weizenbaum argued the impression of being understood is the speaker's own contribution—people attribute 'background knowledge, insights and reasoning ability' to their conversational partner—and he later wrote *Computer Power and Human Reason* (1976) to warn against the effect.
Questions this page answers
- Who created ELIZA?
- Joseph Weizenbaum at MIT. He wrote it in MAD-SLIP and published it in Communications of the ACM in January 1966.
- What is the difference between ELIZA and DOCTOR?
- ELIZA is the engine. What it says is governed by a swappable script, and DOCTOR is the best-known one, imitating a Rogerian psychotherapist.
- Did ELIZA understand what people typed?
- No. It picked keywords out of the input and rephrased them back using decomposition and reassembly rules. In the paper Weizenbaum argued that the impression of being understood is the speaker's own contribution.
1970s
University of Edinburgh, School of Informatics, Freddy II Project · CC BY 4.0 · Commons ↗ Commissioned by the British Science Research Council, James Lighthill produced a report dated July 1972 and published in 1973 that concluded 'in no part of the field have the discoveries made so far produced the major impact that was then promised.' The SRC withdrew most of its AI funding, leaving programmes standing only at Edinburgh, Essex and Sussex. The American retrenchment was a separate chain of events: the National Research Council's ALPAC report (1966) ended machine-translation funding, the Mansfield Amendment (1969) obliged DARPA to back mission-oriented work, and in 1974 DARPA cancelled its Speech Understanding Research programme. The decade that followed — lean on both sides of the Atlantic, for its own reasons on each — is known as the first AI winter, the first sustained rebuttal of the optimistic schedules set out by Newell, Minsky, and others.
Questions this page answers
- When was the Lighthill report published?
- It was dated July 1972 and published in 1973, commissioned by the British Science Research Council. The year 1974 is usually given as the start of the first AI winter, not the date of the report.
- Did the Lighthill report cause the American AI winter too?
- No. The US retrenchment was its own chain of events: the ALPAC report of 1966 ended machine-translation funding, the Mansfield Amendment of 1969 obliged DARPA to back mission-oriented work, and DARPA cancelled its Speech Understanding Research programme in 1974.
1980s
Jason Riedy (Wikimedia Commons) · CC BY 2.0 · Commons ↗ Stanford's MYCIN (bacterial-infection diagnosis) and DEC's XCON (automated VAX configuration) showed the appeal of encoding domain expertise as explicit 'if-then' rules. The approach spread rapidly into industry in the early 1980s; Japan's MITI launched the Fifth Generation Computer Project (1982–92). By 1985, corporate spending on AI had passed US$1 billion a year. The maintenance burden and brittleness of rule systems, however, collapsed the boom—the LISP machine market fell apart in 1987—and the second AI winter followed.
Questions this page answers
- What ended the expert-systems boom?
- The maintenance burden and brittleness of rule-based systems. The LISP machine market fell apart in 1987, and the second AI winter followed.
1990s
Christina Xu (Wikimedia Commons, via Flickr) · CC BY 2.0 · Commons ↗ In a six-game rematch played in New York from 3 to 11 May 1997, IBM's chess-specific machine Deep Blue beat the reigning world champion Garry Kasparov 3½–2½. Their first match, in Philadelphia in February 1996, had gone to Kasparov 4–2—though Deep Blue took game 1 there, the first game a computer had won against a reigning world champion under regular time controls. Over the intervening year IBM strengthened the endgame databases and the evaluation function and brought in grandmaster advisers. The 1997 result was the first time a computer won a match against a reigning world champion under standard tournament time controls. Its substance was game-tree search plus a specialised evaluation function, with no modern learning involved.
Questions this page answers
- When did Deep Blue beat Kasparov?
- On 11 May 1997, in the final game of a six-game match in New York, which ended 3.5-2.5. It was the first time a computer won a match against a reigning world champion under standard tournament time controls.
- What happened in their 1996 match?
- Kasparov won it 4-2 in Philadelphia in February 1996. Deep Blue did take game 1 there, the first game a computer had won against a reigning world champion under regular time controls.
2010s
- EVT.007T2iPhone 4S and Siri — Jobs' Last LaunchA History of the iPhoneA History of Mobile Phones and Smartphones
Daniel Voigt Godoy (Wikimedia Commons) · CC BY 4.0 · Commons ↗ Entering as 'SuperVision', Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton of the University of Toronto took the ILSVRC 2012 classification task with a top-5 error rate of 15.3% (16.4% on supplied training data alone). The best result from any other team was 26.2%, from the University of Tokyo's ISI group using conventional methods. Their convolutional network—about 60 million parameters, trained for five to six days on two NVIDIA GTX 580 3GB GPUs—became known as 'AlexNet' after its first author. Computer vision moved from hand-engineered features to deep learning from that point forward.
Immigrant laborer (Wikimedia Commons) · CC0 1.0 (public domain dedication) · Commons ↗ DeepMind's AlphaGo won its five-game series against Lee Sedol, one of the world's top Go players, 4–1, played in Seoul between 9 and 15 March 2016. The combinatorial explosion of Go had long been thought to put a machine victory at least a decade away; the combination of Monte Carlo tree search and deep reinforcement learning broke that wall. In Game 2, AlphaGo's move 37 — a fifth-line shoulder hit the program itself rated a one-in-ten-thousand human choice — was so unlike anything a professional would have played that the Go community discussed it for weeks. In Game 4, the single game Lee won, he answered with move 78, a wedge of equally long odds that commentators named the 'divine move'.
Questions this page answers
- When did AlphaGo play Lee Sedol, and who won?
- The five-game series ran in Seoul from 9 to 15 March 2016, and AlphaGo won it 4-1. Lee Sedol took only game 4.
- Why is move 37 of game 2 famous?
- It was a fifth-line shoulder hit that AlphaGo itself rated a one-in-ten-thousand human choice. In game 4 Lee answered with move 78, of equally long odds, which commentators named the divine move.
Yuening Jia (Wikimedia Commons), DOI:10.1088/1742-6596/1314/1/012186 · CC BY-SA 3.0 · Commons ↗ Ashish Vaswani and seven colleagues at Google Brain and Google Research proposed the Transformer—a sequence-to-sequence architecture built solely on self-attention. Posted to arXiv on 12 June 2017, it was presented that December at NIPS 2017. It displaced the RNN and LSTM, the prevailing NLP architectures, with a structure that parallelised easily and trained efficiently. By 2026 the paper had been cited more than 250,000 times, placing it among the ten most-cited papers of the twenty-first century; every modern large language model (BERT, the GPT family, Claude, Gemini) is a descendant.
Questions this page answers
- When was Attention Is All You Need published?
- It was posted to arXiv on 12 June 2017 and presented at NIPS 2017 that December. The authors were Ashish Vaswani and seven colleagues at Google Brain and Google Research.
Daniel Voigt Godoy (Wikimedia Commons) · CC BY 4.0 · Commons ↗ Jacob Devlin and colleagues at Google AI posted BERT (Bidirectional Encoder Representations from Transformers) to arXiv as 1810.04805. Pretraining a Transformer encoder bidirectionally as a masked-language model set new state of the art on eleven NLP tasks; BERT-Large scored 80.5 on the official GLUE leaderboard against GPT's 72.8. The paper was presented at NAACL 2019, where it won Best Long Paper. It established the pretrain-then-fine-tune paradigm that underlies every modern LLM, and with the autoregressive GPT family forms one of the two great currents of Transformer-based language modelling.
2020s
OpenAI (Wikimedia Commons) · Public domain (below threshold of originality) · Commons ↗ OpenAI's 175-billion-parameter language model. The paper, "Language Models are Few-Shot Learners," went up on arXiv on 28 May 2020; on 11 June an API running models from the same family opened as a waitlisted private beta. At more than a hundred times the scale of GPT-2 (1.5B), it showed 'few-shot' competence—translation, summarisation, question answering, and code produced from a handful of in-prompt examples with no gradient updates. The paper was one of three NeurIPS 2020 Best Paper Award winners, and it pushed the research conversation toward scaling large language models.
OpenAI (Wikimedia Commons) · Public domain (below threshold of originality) · Commons ↗ OpenAI released ChatGPT as a free research preview—a model from the GPT-3.5 series, which finished training in early 2022, fine-tuned for dialogue with RLHF and reachable from any browser. One million users in five days (Sam Altman's figure); an estimated one hundred million monthly active users within two months—the latter a UBS estimate built on Similar Web data and reported by Reuters on 1 February 2023, not a number OpenAI published. It was the fastest ramp any consumer application had then managed (Meta's Threads passed 100 million sign-ups in five days in July 2023), and it moved generative AI overnight from specialist research into homes, schools, and workplaces, forcing strategic pivots at Google, Meta, Anthropic, and Microsoft.
Hstoops / Microsoft (Wikimedia Commons) · Public domain (below threshold of originality; trademark applies) · Commons ↗ Microsoft unveiled 'New Bing'—Bing search with a chat experience built in—as a waitlisted limited preview. On the day, the model was described only as 'a new, next-generation OpenAI large language model that is more powerful than ChatGPT'; Microsoft confirmed it was GPT-4 on 14 March. Microsoft 365 Copilot followed on 16 March, bringing LLMs into the central battlefronts of search and productivity software at once. Coming barely two months after the ChatGPT launch, it triggered the industry-wide pivot to 'how do we integrate an LLM into our existing product?'. Google had announced Bard the day before, on 6 February, and the factual error in its promo clip surfaced on the 8th.
Bubeck, S. et al., "Sparks of Artificial General Intelligence" (arXiv:2303.12712), Microsoft Research — via Wikimedia Commons · CC BY 4.0 · Commons ↗ OpenAI announced GPT-4, a multimodal model taking image and text inputs and producing text output—though image input was a research preview at announcement and did not reach general availability until September 2023. The technical report gives a simulated Uniform Bar Exam score of 298/400, around the top 10% of test takers. That 90th-percentile figure was later shown to be inflated: measured against a July sitting of the same exam GPT-4 falls below the 69th percentile, drops to roughly the 48th percentile against those who passed, and to roughly the 15th percentile on the essays (Martínez, Artificial Intelligence and Law, 2024). Its AP results were uneven—5s in Biology, Environmental Science and Macroeconomics, but 2s in both AP English Language (14th–44th percentile) and AP English Literature (8th–22nd), below a passing grade. Parameter count, hardware, training compute, and dataset construction were all withheld, citing "the competitive landscape and the safety implications of large-scale models"—the generation that marked OpenAI's turn from the publish-and-research posture of GPT-3 to a commercially closed lab. GPT-4 was retired from ChatGPT on 30 April 2025 in favour of GPT-4o, remaining available in the API.
- EVT.022T2iPhone 16 and Apple IntelligenceA History of the iPhoneA History of Mobile Phones and Smartphones
