Every event in the history of AI

McCulloch and Pitts prove that networks of all-or-none threshold neurons are equivalent to formal logic.

Turing swaps "Can machines think?" for the imitation game, in which an interrogator judges typed replies.

Newell, Shaw and Simon's program proves 38 of 52 Principia theorems by heuristic search, not enumeration.

McCarthy, Minsky, Rochester and Shannon name artificial intelligence and gather its founders at Dartmouth.

Rosenblatt's perceptron learns to classify patterns by adjusting its connection weights from labeled examples.

McCarthy begins LISP at MIT: recursive functions over lists, with programs written as the same lists they process.

Samuel's IBM 701 checkers program improves by tuning its own position-scoring weights through self-play.

Unimation's Unimate, a teach-and-replay hydraulic arm, starts unloading die castings at a GM plant in New Jersey.

DENDRAL proposes and ranks organic molecule structures from a mass spectrum using encoded chemical rules.

Weizenbaum's ELIZA mimics a Rogerian therapist by matching keywords and reflecting users' own words back.

SRI begins Shakey, a camera-equipped robot that perceives rooms, plans actions in logic and pushes boxes.

Minsky and Papert prove one-layer perceptrons with local detectors cannot compute parity or connectedness.

Terry Winograd's SHRDLU parses typed English, resolves pronouns and moves blocks in a simulated tabletop world.

Lighthill's report to Britain's Science Research Council finds AI's general ambitions have delivered little.

Shortliffe's MYCIN chains 200 if-then rules backward to pick antibiotics and can explain why it asks each question.

DEC puts McDermott's R1 (XCON), a 772-rule system, to work configuring VAX-11/780 computer orders.

Fukushima's NHK network alternates learned S-cell and fixed C-cell layers to recognize patterns wherever they sit.

John Hopfield shows a symmetric network of binary units stores memories as stable states it settles into.

Japan's MITI founds ICOT to build parallel logic-programming machines for knowledge processing over ten years.

Rumelhart, Hinton and Williams show in Nature that backpropagated errors let hidden units learn their own features.

Lisp machine makers and expert-system startups slide into loss as a skeptic takes over DARPA's AI funding.

Judea Pearl's 1988 book makes graphs of local probabilities, updated by passed messages, AI's tool for uncertainty.

LeCun's Bell Labs team trains a shared-weight convolutional network end to end on 7,291 real zip code digits.

Christopher Watkins' Cambridge thesis defines Q-learning, which learns optimal action values without a model.

IBM's TD-Gammon learns backgammon from games against itself, then loses just 7 points in 38 games to top players.

Cortes and Vapnik add slack to the maximum-margin kernel classifier, tying LeNet-4 at 1.1% error on NIST digits.

IBM's Deep Blue defeats world champion Garry Kasparov 3.5 to 2.5 in a six-game match in New York.

Hochreiter and Schmidhuber's gated memory cell lets a recurrent network learn across lags of over 1000 steps.

Breiman grows unpruned trees on bootstrap samples, each split weighing a few random features, and lets them vote.

Bengio's team learns a vector per word while predicting the next word, beating smoothed n-gram models.

Stanford's robot Stanley wins DARPA's 132-mile desert race in 6 h 54 min, one of five finishers among 23 starters.

Hinton, Osindero and Teh train a deep network one RBM layer at a time, reaching 1.25% error on MNIST digits.

NVIDIA posts a public beta of its CUDA C compiler and SDK, letting developers program GPUs directly in C.

Deng, Fei-Fei and Princeton colleagues release ImageNet: 3.2 million verified images in 5,247 WordNet categories.

IBM's Watson beats Jeopardy! champions Ken Jennings and Brad Rutter, winning $77,147 to $24,000 and $21,600.

Apple unveils Siri as a beta on the iPhone 4S, turning spoken requests into actions across apps and web services.

Krizhevsky, Sutskever and Hinton's GPU-trained convolutional network wins ImageNet 2012 with 15.3% top-5 error.

Mikolov's Google team shows two shallow models learn word vectors from 1.6 billion words in under a day.

DeepMind's Q-network learns 7 Atari games from raw pixels with one recipe, beating prior methods on 6.

Kingma and Welling backpropagate through random sampling to train a generative autoencoder.

Goodfellow's Montreal team trains a generator by pitting it against a discriminator network.

Bahdanau, Cho and Bengio's RNNsearch weights all source words afresh for each output word, not one fixed vector.

Sutskever, Vinyals and Le's LSTM ensemble out-translates a phrase-based system, 34.81 to 33.30 BLEU.

Kingma and Ba's Adam gives every parameter its own step size from bias-corrected gradient moment estimates.

Microsoft's He, Zhang, Ren and Sun train 152-layer residual networks and win ImageNet 2015 at 3.57% top-5 error.

Altman, Musk, Brockman and Sutskever launch OpenAI, a nonprofit AI lab with $1 billion in pledged funding.

DeepMind's AlphaGo beats Lee Sedol 4 to 1 at Go in Seoul, with the startling Move 37 in game two.

Microsoft's Tay chatbot learns from Twitter users, and coordinated trolls make it post hate within 16 hours.

Google moves all Chinese to English in Translate, about 18 million translations a day, to a deep LSTM network.

Carnegie Mellon's Libratus beats four heads-up no-limit hold'em pros over 120,000 hands by 147 mbb/game.

Eight researchers unveil the Transformer, an attention-only model scoring 28.4 BLEU on English-German translation.

OpenAI and DeepMind agents learn from rewards fit to human clip comparisons, a backflip from 900 judgments.

One DeepMind algorithm learns chess, shogi and Go from the rules alone and beats the top program in each.

OpenAI pretrains a Transformer decoder on 7,000 books, then fine-tunes it to records on 9 of 12 language tasks.

Google's BERT learns by predicting masked words from both sides, setting records on 11 language tasks.

OpenAI announces GPT-2, a 1.5 billion parameter language model, but withholds its full weights over misuse risk.

Google's T5 frames every language task as text in, text out, and releases the 750 GB C4 corpus with its models.

Kaplan's OpenAI team finds language-model loss falls as power laws in parameters, data and compute.

OpenAI's 175 billion parameter GPT-3 performs many language tasks from a few examples in its prompt.

Berkeley's Ho, Jain and Abbeel train an image model to reverse noise, reaching a record 3.17 FID on CIFAR-10.

Waymo opens rides with no one behind the wheel to Waymo One riders across about 50 square miles of metro Phoenix.

Dosovitskiy's Google team runs a plain Transformer on 16x16 image patches and reaches 88.55% on ImageNet.

DeepMind's AlphaFold 2 predicts unseen protein structures at CASP14 with a median score of 92.4 GDT.

OpenAI's DALL·E, a 12-billion-parameter transformer, draws images from captions such as an avocado armchair.

GitHub previews Copilot, an editor extension in which OpenAI's Codex suggests whole lines and functions.

Rombach's Munich team runs image diffusion in an autoencoder's compressed latent space, not raw pixels.

Labelers prefer OpenAI's 1.3-billion-parameter InstructGPT, tuned on human feedback, to 175-billion GPT-3.

DeepMind's 70B Chinchilla, trained on 1.4 trillion tokens, beats the 280B Gopher at equal compute.

OpenAI's DALL·E 2 generates and edits 1024x1024 images by decoding CLIP embeddings with diffusion.

Stability AI releases Stable Diffusion with downloadable weights that run on a consumer GPU.

OpenAI releases ChatGPT, a free preview of a GPT-3.5 model tuned with human feedback to hold a dialogue.

Meta releases LLaMA, 7B to 65B models trained on public data; LLaMA-13B beats 175B GPT-3 on most benchmarks.

OpenAI launches GPT-4, accepting images and text, scoring 86.4% on MMLU and passing a simulated bar exam.

Anthropic releases Claude and Claude Instant via API and partners, trained with Constitutional AI.

Kirillov and Meta colleagues release SAM, which masks objects from a click or box, and 1.1 billion masks.

Meta releases Llama 2 at 7B, 13B and 70B with chat versions, licensed for commercial use under conditions.

Mistral AI releases Mistral 7B under Apache 2.0, a 7.3B model that beats Llama 2 13B on all tested benchmarks.

28 countries and the EU, including the US and China, sign a non-binding statement on frontier AI risk.

OpenAI's nonprofit board fires Sam Altman as CEO, then rehires him after 738 of about 770 staff threaten to quit.

Google DeepMind unveils Gemini 1.0 in Ultra, Pro and Nano sizes, trained on text, images, audio and video.

OpenAI previews Sora, a text-to-video model that renders coherent clips up to a minute long from prompts.

Anthropic releases Opus, Sonnet and Haiku as one Claude 3 family, with image input and a 200,000-token window.

AlphaFold 3 predicts the joint 3D structure of proteins bound to DNA, RNA, small-molecule ligands and ions.

OpenAI's GPT-4o handles text, images and audio in one model, answering speech in as little as 232 milliseconds.

Anthropic's mid-tier Claude 3.5 Sonnet outscores Claude 3 Opus at twice its speed, and previews Artifacts.

Meta releases downloadable Llama 3.1 weights, led by a dense 405B model with a 128,000-token context window.

The EU AI Act, Regulation 2024/1689, enters into force, starting a phased clock for its obligations.

OpenAI previews o1, a model trained with reinforcement learning to reason at length before it answers.

Hopfield and Hinton win the Physics Nobel for the 1982 Hopfield network and the 1985 Boltzmann machine.

Baker wins half the chemistry Nobel for protein design; Hassabis and Jumper share half for AlphaFold.

Anthropic's public beta lets Claude 3.5 Sonnet drive a desktop from screenshots, scoring 14.9% on OSWorld.

Anthropic opens the Model Context Protocol, a JSON-RPC standard linking AI apps to tools and data.

DeepSeek releases R1 under the MIT License, a downloadable reasoning model it reports as on par with o1.

Anthropic previews Claude Code, a terminal agent that edits files, runs tests and commits, with Claude 3.7 Sonnet.

Google DeepMind releases Gemini 2.5 Pro Experimental, a thinking model that debuts at the top of LMArena.

Alibaba's Qwen team releases Qwen3 under Apache 2.0: 8 models from 0.6B to 235B, each with a thinking switch.

DeepMind's AlphaEvolve evolves code with Gemini and multiplies 4x4 complex matrices with 48 products, not 49.

Anthropic launches Claude Opus 4 and Sonnet 4, scoring 72.5% and 72.7% on SWE-bench Verified.

Judge Alsup rules training Claude on books was fair use; keeping over 7 million pirated copies was not.

Gemini Deep Think and an OpenAI model score 35 of 42 at IMO 2025 with prose proofs, the gold cut-off.

OpenAI releases gpt-oss-120b and gpt-oss-20b under Apache 2.0, its first open-weight language models since GPT-2.

OpenAI makes GPT-5, a router choosing between fast and reasoning models, the default for all ChatGPT users.

Newsom signs SB 53, making the largest frontier AI developers publish safety frameworks and report incidents.

Google releases Gemini 3 Pro and puts it into Search's AI Mode on day one, reporting 1501 Elo on LMArena.

Hegseth moves to label Anthropic a supply-chain risk over two limits on Claude's military use.

Anthropic withholds Claude Mythos Preview over its exploit skill and lends it only to vetted defenders.

An unnamed OpenAI model disproves Erdős's 1946 unit distance conjecture with number-field point sets.

OpenAI names its own models, testing on ExploitGym, as the agents that broke into Hugging Face.

EU AI Act transparency rules take effect: chatbots must say they are AI and AI output must be marked.

OpenAI releases GPT-6 Astra, the first model to cross its own Critical cybersecurity threshold.

OpenAI says some 10,000 agents running an unreleased model proved forced Navier-Stokes blowup, checked in Lean.

The Third Circuit holds that copying 2,243 Westlaw headnotes to train ROSS's legal search AI is not fair use.

OpenAI posts 722 AI-written papers claiming a quasi-Riemann hypothesis and no Siegel zeros.

OpenAI and Anthropic models independently claim Lean-checked 2^o(n) algorithms for 3-SAT, refuting ETH.