Chronicle: interactive AI timeline

Chronicle.

114 events in the history of artificial intelligence, 1943 to 2026. Each one is scored on a published rubric and backed by a sourced research report.

1940s · 1

  1. 1943 · Research86A mathematical neuronMcCulloch-Pitts neuron · McCulloch & PittsMcCulloch and Pitts prove that networks of all-or-none threshold neurons are equivalent to formal logic.

1950s · 6

  1. 1950 · Research95The question that started it allTuring test · Alan TuringTuring swaps "Can machines think?" for the imitation game, in which an interrogator judges typed replies.
  2. Jun 15, 1956 · Research80Rules of thumb prove theoremsLogic Theorist · RAND Corporation & Carnegie TechNewell, Shaw and Simon's program proves 38 of 52 Principia theorems by heuristic search, not enumeration.
  3. 1956 · Research94Artificial intelligence gets a nameDartmouth workshopMcCarthy, Minsky, Rochester and Shannon name artificial intelligence and gather its founders at Dartmouth.
  4. 1958 · Research78The perceptronFrank RosenblattRosenblatt's perceptron learns to classify patterns by adjusting its connection weights from labeled examples.
  5. 1958 · Research86Programs become dataLISP · John McCarthy, MITMcCarthy begins LISP at MIT: recursive functions over lists, with programs written as the same lists they process.
  6. 1959 · Research80Learning from a game of checkersArthur Samuel's checkers program · Arthur Samuel · IBMSamuel's IBM 701 checkers program improves by tuning its own position-scoring weights through self-play.

1960s · 5

  1. 1961 · Robotics71A programmable arm joins the lineUnimate · Unimation · General MotorsUnimation's Unimate, a teach-and-replay hydraulic arm, starts unloading die castings at a GM plant in New Jersey.
  2. 1965 · Research71Expertise becomes softwareDENDRAL expert system · Stanford UniversityDENDRAL proposes and ranks organic molecule structures from a mass spectrum using encoded chemical rules.
  3. 1966 · Culture67A conversation with ELIZAELIZA chatbot · MITWeizenbaum's ELIZA mimics a Rogerian therapist by matching keywords and reflecting users' own words back.
  4. 1966 · Robotics67Shakey enters the worldShakey the robot · SRISRI begins Shakey, a camera-equipped robot that perceives rooms, plans actions in logic and pushes boxes.
  5. 1969 · Research58What one layer cannot computePerceptrons (Minsky and Papert) · Minsky & Papert · MITMinsky and Papert prove one-layer perceptrons with local detectors cannot compute parity or connectedness.

1970s · 3

  1. 1971 · Research64Understanding inside a world of blocksSHRDLU · MIT AI LabTerry Winograd's SHRDLU parses typed English, resolves pronouns and moves blocks in a simulated tabletop world.
  2. 1973 · Culture53The first AI winterResearch fundingLighthill's report to Britain's Science Research Council finds AI's general ambitions have delivered little.
  3. 1976 · Research71A consultant built from rulesMYCIN expert system · Stanford UniversityShortliffe's MYCIN chains 200 if-then rules backward to pick antibiotics and can explain why it asks each question.

1980s · 9

  1. 1980 · Launches57Expert systems go into productionXCON (R1) expert system · Digital Equipment Corporation · Carnegie MellonDEC puts McDermott's R1 (XCON), a 772-rule system, to work configuring VAX-11/780 computer orders.
  2. 1980 · Research73Shift invariance built into the wiringNeocognitron · Kunihiko Fukushima, NHKFukushima's NHK network alternates learned S-cell and fixed C-cell layers to recognize patterns wherever they sit.
  3. 1982 · Research81Memory as an energy landscapeHopfield network · John HopfieldJohn Hopfield shows a symmetric network of binary units stores memories as stable states it settles into.
  4. 1982 · Policy60Japan bets a decade on logic machinesFifth Generation Computer Systems project · Japan's MITI · ICOTJapan's MITI founds ICOT to build parallel logic-programming machines for knowledge processing over ten years.
  5. 1986 · Research93Learning by backpropagationRumelhart, Hinton & WilliamsRumelhart, Hinton and Williams show in Nature that backpropagated errors let hidden units learn their own features.
  6. 1987 · Culture54The boom meets its balance sheetSecond AI winter · AI industry and DARPALisp machine makers and expert-system startups slide into loss as a skeptic takes over DARPA's AI funding.
  7. 1988 · Research86Uncertainty gets a calculusBayesian networks · UCLAJudea Pearl's 1988 book makes graphs of local probabilities, updated by passed messages, AI's tool for uncertainty.
  8. 1989 · Research80Learned filters read the mailLeNet (convolutional networks) · LeCun and collaborators · Bell LabsLeCun's Bell Labs team trains a shared-weight convolutional network end to end on 7,291 real zip code digits.
  9. 1989 · Research80Learning what an action is worthQ-learning · University of CambridgeChristopher Watkins' Cambridge thesis defines Q-learning, which learns optimal action values without a model.

1990s · 4

  1. 1992 · Research71Learning backgammon from its own gamesTD-Gammon · IBM ResearchIBM's TD-Gammon learns backgammon from games against itself, then loses just 7 points in 38 games to top players.
  2. 1995 · Research83The widest margin winsSupport vector machines · AT&T Bell LabsCortes and Vapnik add slack to the maximum-margin kernel classifier, tying LeNet-4 at 1.1% error on NIST digits.
  3. May 11, 1997 · Benchmarks62Deep Blue beats KasparovIBMIBM's Deep Blue defeats world champion Garry Kasparov 3.5 to 2.5 in a six-game match in New York.
  4. 1997 · Research89Neural networks learn to rememberLSTM · Hochreiter & SchmidhuberHochreiter and Schmidhuber's gated memory cell lets a recurrent network learn across lags of over 1000 steps.

2000s · 6

  1. 2001 · Research77Strength from uncorrelated treesRandom forests · Leo Breiman, UC BerkeleyBreiman grows unpruned trees on bootstrap samples, each split weighing a few random features, and lets them vote.
  2. 2003 · Research80Words become learned representationsNeural probabilistic language model · Bengio, Ducharme, Vincent & JauvinBengio's team learns a vector per word while predicting the next word, beating smoothed n-gram models.
  3. Oct 8, 2005 · Robotics65Driverless across the desertDARPA Grand Challenge · DARPA · Stanford Racing TeamStanford's robot Stanley wins DARPA's 132-mile desert race in 6 h 54 min, one of five finishers among 23 starters.
  4. 2006 · Research75The deep learning revivalDeep belief networks · Geoffrey Hinton & collaboratorsHinton, Osindero and Teh train a deep network one RBM layer at a time, reaching 1.25% error on MNIST digits.
  5. 2007 · Launches91GPUs become programmable computeNVIDIA CUDANVIDIA posts a public beta of its CUDA C compiler and SDK, letting developers program GPUs directly in C.
  6. 2009 · Research92ImageNet changes the datasetPrinceton & StanfordDeng, Fei-Fei and Princeton colleagues release ImageNet: 3.2 million verified images in 5,247 WordNet categories.

2010s · 23

  1. Feb 16, 2011 · Benchmarks58Watson wins Jeopardy!IBM WatsonIBM's Watson beats Jeopardy! champions Ken Jennings and Brad Rutter, winning $77,147 to $24,000 and $21,600.
  2. Oct 4, 2011 · Launches69A spoken assistant goes mass-marketSiri · AppleApple unveils Siri as a beta on the iPhone 4S, turning spoken requests into actions across apps and web services.
  3. 2012 · Research95AlexNet transforms image recognitionUniversity of TorontoKrizhevsky, Sutskever and Hinton's GPU-trained convolutional network wins ImageNet 2012 with 15.3% top-5 error.
  4. Jan 16, 2013 · Research83Word relationships in vector spaceword2vec · Mikolov and collaborators · GoogleMikolov's Google team shows two shallow models learn word vectors from 1.6 billion words in under a day.
  5. Dec 19, 2013 · Research80Learning to play from pixelsDeep Q-network · Mnih and collaborators · DeepMindDeepMind's Q-network learns 7 Atari games from raw pixels with one recipe, beating prior methods on 6.
  6. Dec 20, 2013 · Research81Learning a generative latent spaceVariational autoencoder · Kingma & WellingKingma and Welling backpropagate through random sampling to train a generative autoencoder.
  7. Jun 10, 2014 · Research83Machines learn to imagineGenerative adversarial networks · Ian Goodfellow & collaboratorsGoodfellow's Montreal team trains a generator by pitting it against a discriminator network.
  8. Sep 1, 2014 · Research83Learning where to lookAttention mechanism (Bahdanau, Cho and Bengio) · Bahdanau, Cho & Bengio · Université de MontréalBahdanau, Cho and Bengio's RNNsearch weights all source words afresh for each output word, not one fixed vector.
  9. Sep 10, 2014 · Research80One sequence becomes anotherSequence-to-sequence learning · Sutskever, Vinyals & LeSutskever, Vinyals and Le's LSTM ensemble out-translates a phrase-based system, 34.81 to 33.30 BLEU.
  10. Dec 22, 2014 · Research84Adam changes the training toolkitAdam optimizer · Kingma & BaKingma and Ba's Adam gives every parameter its own step size from bias-corrected gradient moment estimates.
  11. Dec 10, 2015 · Research93Depth without the degradationResNet · He and collaborators · MicrosoftMicrosoft's He, Zhang, Ren and Sun train 152-layer residual networks and win ImageNet 2015 at 3.57% top-5 error.
  12. Dec 11, 2015 · Culture62OpenAI is foundedOpenAIAltman, Musk, Brockman and Sutskever launch OpenAI, a nonprofit AI lab with $1 billion in pledged funding.
  13. Mar 15, 2016 · Benchmarks83Move 37. A new kind of intuition.AlphaGo vs Lee Sedol · Google DeepMindDeepMind's AlphaGo beats Lee Sedol 4 to 1 at Go in Seoul, with the startling Move 37 in game two.
  14. Mar 23, 2016 · Culture53A chatbot schooled by its trollsMicrosoft TayMicrosoft's Tay chatbot learns from Twitter users, and coordinated trolls make it post hate within 16 hours.
  15. Sep 27, 2016 · Launches69Phrase tables give way to neural networksGoogle Neural Machine TranslationGoogle moves all Chinese to English in Translate, about 18 million translations a day, to a deep LSTM network.
  16. Jan 30, 2017 · Benchmarks64Poker's hidden cards fall to game theoryLibratus · Carnegie Mellon UniversityCarnegie Mellon's Libratus beats four heads-up no-limit hold'em pros over 120,000 hands by 147 mbb/game.
  17. Jun 12, 2017 · Research98Attention is all you needTransformer · Google researchersEight researchers unveil the Transformer, an attention-only model scoring 28.4 BLEU on English-German translation.
  18. Jun 12, 2017 · Research83A reward learned from human choicesDeep reinforcement learning from human preferences (RLHF) · Christiano and collaborators · OpenAI & DeepMindOpenAI and DeepMind agents learn from rewards fit to human clip comparisons, a backflip from 900 judgments.
  19. Dec 5, 2017 · Research75Mastery through self-playAlphaZero · Silver and collaborators · DeepMindOne DeepMind algorithm learns chess, shogi and Go from the rules alone and beats the top program in each.
  20. Jun 11, 2018 · Models78The first GPTGPT-1 · OpenAIOpenAI pretrains a Transformer decoder on 7,000 books, then fine-tunes it to records on 9 of 12 language tasks.
  21. Oct 11, 2018 · Research86BERT reads both waysGoogleGoogle's BERT learns by predicting masked words from both sides, setting records on 11 language tasks.
  22. Feb 14, 2019 · Models72GPT-2 and the release debateOpenAIOpenAI announces GPT-2, a 1.5 billion parameter language model, but withholds its full weights over misuse risk.
  23. Oct 23, 2019 · Research75Every language task is text-to-textT5 · Raffel and collaborators · GoogleGoogle's T5 frames every language task as text in, text out, and releases the 750 GB C4 corpus with its models.

2020s · 57

  1. Jan 23, 2020 · Research82Bigger models, predictably betterScaling laws for neural language models · Kaplan and collaborators · OpenAIKaplan's OpenAI team finds language-model loss falls as power laws in parameters, data and compute.
  2. May 28, 2020 · Models86GPT-3 scales upOpenAIOpenAI's 175 billion parameter GPT-3 performs many language tasks from a few examples in its prompt.
  3. Jun 19, 2020 · Research83Learning to reverse noiseDDPM · Ho, Jain & AbbeelBerkeley's Ho, Jain and Abbeel train an image model to reverse noise, reaching a record 3.17 FID on CIFAR-10.
  4. Oct 8, 2020 · Robotics64The driver's seat goes emptyWaymo fully driverless public serviceWaymo opens rides with no one behind the wheel to Waymo One riders across about 50 square miles of metro Phoenix.
  5. Oct 22, 2020 · Research83Images become patch sequencesVision Transformer · Dosovitskiy and collaborators · GoogleDosovitskiy's Google team runs a plain Transformer on 16x16 image patches and reaches 88.55% on ImageNet.
  6. Nov 30, 2020 · Research95A 50-year biology problemAlphaFold 2 · Google DeepMindDeepMind's AlphaFold 2 predicts unseen protein structures at CASP14 with a median score of 92.4 GDT.
  7. Jan 5, 2021 · Models72DALL·E: language becomes imageOpenAIOpenAI's DALL·E, a 12-billion-parameter transformer, draws images from captions such as an avocado armchair.
  8. Jun 29, 2021 · Launches75Your AI pair programmerGitHub CopilotGitHub previews Copilot, an editor extension in which OpenAI's Codex suggests whole lines and functions.
  9. Dec 20, 2021 · Research86Diffusion leaves pixel spaceLatent diffusion models · Rombach and collaboratorsRombach's Munich team runs image diffusion in an autoencoder's compressed latent space, not raw pixels.
  10. Mar 4, 2022 · Research86Learning what users preferInstructGPT · Ouyang and collaborators · OpenAILabelers prefer OpenAI's 1.3-billion-parameter InstructGPT, tuned on human feedback, to 175-billion GPT-3.
  11. Mar 29, 2022 · Research80Scale the data, not just the modelChinchilla scaling laws · Hoffmann and collaborators · DeepMindDeepMind's 70B Chinchilla, trained on 1.4 trillion tokens, beats the 280B Gopher at equal compute.
  12. Apr 6, 2022 · Models69DALL·E 2OpenAIOpenAI's DALL·E 2 generates and edits 1024x1024 images by decoding CLIP embeddings with diffusion.
  13. Aug 22, 2022 · Open source92Stable Diffusion opens upStability AI & collaboratorsStability AI releases Stable Diffusion with downloadable weights that run on a consumer GPU.
  14. Nov 30, 2022 · Launches85Hello, ChatGPT.ChatGPT release · OpenAIOpenAI releases ChatGPT, a free preview of a GPT-3.5 model tuned with human feedback to hold a dialogue.
  15. Feb 24, 2023 · Open source69LLaMA enters the pictureMetaMeta releases LLaMA, 7B to 65B models trained on public data; LLaMA-13B beats 175B GPT-3 on most benchmarks.
  16. Mar 14, 2023 · Models80GPT-4 raises the barOpenAIOpenAI launches GPT-4, accepting images and text, scoring 86.4% on MMLU and passing a simulated bar exam.
  17. Mar 14, 2023 · Launches54Meet ClaudeClaude release · AnthropicAnthropic releases Claude and Claude Instant via API and partners, trained with Constitutional AI.
  18. Apr 5, 2023 · Research81A foundation model for segmentationSegment Anything · Kirillov and collaborators · MetaKirillov and Meta colleagues release SAM, which masks objects from a click or box, and 1.1 billion masks.
  19. Jul 18, 2023 · Open source74Llama 2 opens commercial accessMetaMeta releases Llama 2 at 7B, 13B and 70B with chat versions, licensed for commercial use under conditions.
  20. Sep 27, 2023 · Open source70A smaller open-weight contenderMistral 7B · Mistral AIMistral AI releases Mistral 7B under Apache 2.0, a 7.3B model that beats Llama 2 13B on all tested benchmarks.
  21. Nov 1, 2023 · Policy60The Bletchley DeclarationAI Safety Summit28 countries and the EU, including the US and China, sign a non-binding statement on frontier AI risk.
  22. Nov 17, 2023 · Culture55The board fired its CEO and lostOpenAI board crisisOpenAI's nonprofit board fires Sam Altman as CEO, then rehires him after 738 of about 770 staff threaten to quit.
  23. Dec 6, 2023 · Models75The Gemini era beginsGoogle Gemini · Google DeepMindGoogle DeepMind unveils Gemini 1.0 in Ultra, Pro and Nano sizes, trained on text, images, audio and video.
  24. Feb 15, 2024 · Models64Sora imagines in motionOpenAI SoraOpenAI previews Sora, a text-to-video model that renders coherent clips up to a minute long from prompts.
  25. Mar 4, 2024 · Models54Claude 3 arrivesAnthropicAnthropic releases Opus, Sonnet and Haiku as one Claude 3 family, with image input and a 200,000-token window.
  26. May 8, 2024 · Research72AlphaFold 3 expands the pictureGoogle DeepMind & Isomorphic LabsAlphaFold 3 predicts the joint 3D structure of proteins bound to DNA, RNA, small-molecule ligands and ions.
  27. May 13, 2024 · Models75GPT-4o: a more natural interfaceOpenAIOpenAI's GPT-4o handles text, images and audio in one model, answering speech in as little as 232 milliseconds.
  28. Jun 20, 2024 · Models63Claude 3.5 SonnetAnthropicAnthropic's mid-tier Claude 3.5 Sonnet outscores Claude 3 Opus at twice its speed, and previews Artifacts.
  29. Jul 23, 2024 · Open source74A frontier-scale downloadable modelLlama 3.1 · MetaMeta releases downloadable Llama 3.1 weights, led by a dense 405B model with a 128,000-token context window.
  30. Aug 1, 2024 · Policy83The EU AI Act enters into forceEuropean UnionThe EU AI Act, Regulation 2024/1689, enters into force, starting a phased clock for its obligations.
  31. Sep 12, 2024 · Models72Thinking before answeringOpenAI o1OpenAI previews o1, a model trained with reinforcement learning to reason at length before it answers.
  32. Oct 8, 2024 · Culture60Neural networks win a NobelHopfield and Hinton Nobel Prize · Nobel Prize in PhysicsHopfield and Hinton win the Physics Nobel for the 1982 Hopfield network and the 1985 Boltzmann machine.
  33. Oct 9, 2024 · Culture60AI meets the Nobel in ChemistryAlphaFold Nobel Prize · Nobel Prize in ChemistryBaker wins half the chemistry Nobel for protein design; Hassabis and Jumper share half for AlphaFold.
  34. Oct 22, 2024 · Launches61Claude uses a computerAnthropicAnthropic's public beta lets Claude 3.5 Sonnet drive a desktop from screenshots, scoring 14.9% on OSWorld.
  35. Nov 25, 2024 · Open source77A common language for toolsModel Context Protocol · AnthropicAnthropic opens the Model Context Protocol, a JSON-RPC standard linking AI apps to tools and data.
  36. Jan 20, 2025 · Open source85DeepSeek-R1 opens reasoningDeepSeekDeepSeek releases R1 under the MIT License, a downloadable reasoning model it reports as on par with o1.
  37. Feb 24, 2025 · Launches64Claude Code enters the terminalAnthropicAnthropic previews Claude Code, a terminal agent that edits files, runs tests and commits, with Claude 3.7 Sonnet.
  38. Mar 25, 2025 · Models64Gemini makes thinking explicitGemini 2.5 · Google DeepMindGoogle DeepMind releases Gemini 2.5 Pro Experimental, a thinking model that debuts at the top of LMArena.
  39. Apr 29, 2025 · Open source70Qwen 3 broadens open reasoningQwen3 · Qwen team · AlibabaAlibaba's Qwen team releases Qwen3 under Apache 2.0: 8 models from 0.6B to 235B, each with a thinking switch.
  40. May 14, 2025 · Research63Breeding algorithms with a language modelAlphaEvolve · Google DeepMindDeepMind's AlphaEvolve evolves code with Gemini and multiplies 4x4 complex matrices with 48 products, not 49.
  41. May 22, 2025 · Models58Claude 4 gets to workAnthropicAnthropic launches Claude Opus 4 and Sonnet 4, scoring 72.5% and 72.7% on SWE-bench Verified.
  42. Jun 23, 2025 · Policy64Fair use for training, not for piracyBartz v. Anthropic · U.S. District Court, N.D. CaliforniaJudge Alsup rules training Claude on books was fair use; keeping over 7 million pirated copies was not.
  43. Jul 21, 2025 · Benchmarks59Olympiad gold in natural languageAI gold at IMO 2025 · Google DeepMind · OpenAIGemini Deep Think and an OpenAI model score 35 of 42 at IMO 2025 with prose proofs, the gold cut-off.
  44. Aug 5, 2025 · Open source70OpenAI returns to open weightsgpt-oss · OpenAIOpenAI releases gpt-oss-120b and gpt-oss-20b under Apache 2.0, its first open-weight language models since GPT-2.
  45. Aug 7, 2025 · Models70GPT-5 arrivesOpenAIOpenAI makes GPT-5, a router choosing between fast and reasoning models, the default for all ChatGPT users.
  46. Sep 29, 2025 · Policy56Disclosure in place of a kill switchCalifornia SB 53 (Transparency in Frontier AI Act) · State of CaliforniaNewsom signs SB 53, making the largest frontier AI developers publish safety frameworks and report incidents.
  47. Nov 18, 2025 · Models70Gemini 3Google DeepMindGoogle releases Gemini 3 Pro and puts it into Search's AI Mode on day one, reporting 1501 Elo on LMArena.
  48. Feb 27, 2026 · Policy59Red lines labeled a supply-chain riskPentagon designates Anthropic a supply-chain risk · U.S. Department of Defense · AnthropicHegseth moves to label Anthropic a supply-chain risk over two limits on Claude's military use.
  49. Apr 7, 2026 · Models56The model Anthropic held backClaude Mythos Preview · AnthropicAnthropic withholds Claude Mythos Preview over its exploit skill and lends it only to vetted defenders.
  50. May 20, 2026 · Research62Erdős's grid is finally beatenErdős unit distance disproof · OpenAIAn unnamed OpenAI model disproves Erdős's 1946 unit distance conjecture with number-field point sets.
  51. Jul 21, 2026 · Policy53The evaluation that escaped its sandboxOpenAI Hugging Face incident · OpenAI · Hugging FaceOpenAI names its own models, testing on ExploitGym, as the agents that broke into Hugging Face.
  52. Aug 2, 2026 · Policy61AI Act enforcement expandsEuropean UnionEU AI Act transparency rules take effect: chatbots must say they are AI and AI output must be marked.
  53. Sep 3, 2026 · Models45Astra: the next frontierGPT-6 Astra · OpenAIOpenAI releases GPT-6 Astra, the first model to cross its own Critical cybersecurity threshold.
  54. Sep 8, 2026 · Research62A claimed blowup for the fluid equationsOpenAI's Navier-Stokes blow-up claimOpenAI says some 10,000 agents running an unreleased model proved forced Navier-Stokes blowup, checked in Lean.
  55. Sep 29, 2026 · Policy49No fair use for training a Westlaw rivalThomson Reuters v. ROSS Intelligence (Third Circuit) · U.S. Court of Appeals for the Third CircuitThe Third Circuit holds that copying 2,243 Westlaw headnotes to train ROSS's legal search AI is not fair use.
  56. Oct 6, 2026 · Research64A claimed fence around zeta's zerosOpenAI's quasi-Riemann hypothesis proofOpenAI posts 722 AI-written papers claiming a quasi-Riemann hypothesis and no Siegel zeros.
  57. Oct 26, 2026 · Research64Two labs claim 3-SAT escapes exponential timeExponential Time Hypothesis refutation · OpenAI and AnthropicOpenAI and Anthropic models independently claim Lean-checked 2^o(n) algorithms for 3-SAT, refuting ETH.