PyyanThe AI Index

Pyyan / Timeline

1943 to 2026 · 83 years

The history of AI

Every milestone that the field still refers back to, from the first mathematical neuron to the model you used this morning. Two funding winters, one architecture that replaced everything, and a great many people who were told they were wasting their time.

83 milestones25 that changed what followed8 eras83 years
1943 to 1955

Foundations

Before there were machines fast enough to matter, there were the ideas. Almost everything that followed is a footnote to this handful of papers.

  1. 1943A logical calculus of the ideas immanent in nervous activityMcCulloch & PittsThe artificial neuronThe first mathematical model of a neuron. Every network on this site is a descendant of this paper.
  2. 1948A Mathematical Theory of CommunicationClaude ShannonInformation theoryCreated information theory outright, and with it the bit. Gave every later system a way to reason about capacity and noise.
  3. 1948CyberneticsNorbert Wiener
  4. 1950Computing Machinery and IntelligenceAlan TuringThe imitation gameTuring sidestepped “can machines think?” as unanswerable and proposed a test instead. Seventy-six years later the question is still framed his way.
  5. 1951SNARC, the first neural network machineMarvin Minsky
  6. 1952A checkers program that improves by playing itselfArthur Samuel
  7. 1955Logic TheoristNewell & SimonThe first AI programProved theorems from Principia Mathematica. Generally counted the first working artificial intelligence program.
1956 to 1973

The first spring

The field got its name, its first laboratories and its first wave of optimism. Researchers predicted human-level machines within a generation.

  1. 1956The Dartmouth workshopJohn McCarthy“Artificial intelligence” is coinedMcCarthy organised a summer workshop and needed a name for it. The field has been arguing about that name ever since.
  2. 1957The perceptronFrank RosenblattThe first trainable networkA machine that learned to classify by adjusting weights. The direct ancestor of every model here, and the subject of the critique that nearly killed the idea.
  3. 1958LispJohn McCarthyThe language AI research ran on for thirty years. Also gave computing garbage collection and the read-eval-print loop.
  4. 1959Arthur Samuel coins “machine learning”IBM
  5. 1966ELIZAJoseph WeizenbaumThe first chatbotA few hundred lines that reflected your words back as questions. Weizenbaum's own secretary asked him to leave the room so she could talk to it privately, and he spent the rest of his career warning about what that meant.
  6. 1969PerceptronsMinsky & PapertThe critique that caused a winterShowed a single-layer perceptron could not learn XOR. Correct, narrowly read, and widely taken as proof that neural networks were a dead end.
  7. 1970Shakey the robotSRI International
  8. 1972PrologColmerauer & Roussel
  9. 1973The Lighthill reportUK Science Research CouncilFunding collapsesConcluded AI had failed to deliver on its promises. British research funding was cut almost to nothing.
1974 to 1986

Winters and expert systems

Two funding collapses, and in between them the one commercial success of symbolic AI: systems that encoded a human expert's rules and sold well until they did not.

  1. 1974The first AI winter begins1974 to 1980
  2. 1976MYCIN diagnoses infectionsStanford
  3. 1976Computer Power and Human ReasonJoseph Weizenbaum
  4. 1979NeocognitronKunihiko FukushimaConvolution, before convolutionA layered vision network with local receptive fields. Everything LeNet and AlexNet later did was anticipated here.
  5. 1980XCON goes into production at DECExpert systems payConfigured computer orders and reportedly saved tens of millions a year. The proof that AI could be a product.
  6. 1980Searle's Chinese Room argumentJohn Searle
  7. 1982Hopfield networksJohn Hopfield
  8. 1986Learning representations by back-propagating errorsRumelhart, Hinton & WilliamsBackpropagationMade multi-layer networks trainable and answered the Perceptrons critique directly. The algorithm every model on this site is still trained with.
  9. 1987The second AI winter begins1987 to 1993The Lisp machine market collapsed. Researchers began calling their work anything except artificial intelligence.
1989 to 2005

The statistical turn

Quietly, while the field was out of fashion, the methods that would win were being built. The word used was machine learning, not AI.

  1. 1989LeNet reads handwritten digitsYann LeCunConvolutional networks workTrained with backpropagation and deployed to read cheques. Neural networks doing real commercial work, twenty-three years before anyone called it deep learning.
  2. 1992TD-Gammon reaches world-class backgammonIBM
  3. 1995Support vector machinesCortes & VapnikFor most of the next decade, the method that beat neural networks on most problems.
  4. 1997Deep Blue defeats KasparovIBMA machine beats the championThe first time a computer beat a reigning world chess champion in a match. Almost no machine learning involved, and it changed the public conversation anyway.
  5. 1997Long Short-Term MemoryHochreiter & SchmidhuberSequences become learnableSolved the vanishing gradient problem and made sequence learning practical. Ran translation, speech and text for the next twenty years.
  6. 1998PageRankPage & Brin
  7. 2002Roomba shipsiRobotThe first domestic robot to sell in millions.
  8. 2005Stanley wins the DARPA Grand ChallengeStanford
2006 to 2016

Deep learning arrives

Three things landed at once: enough data, GPUs fast enough to use it, and a generation of researchers who had never stopped working on neural networks.

  1. 2006Deep belief networksGeoffrey Hinton“Deep learning” gets its name
  2. 2006The Netflix PrizeNetflixA million dollars for a 10% better recommender. It taught a generation how to do applied machine learning.
  3. 2007CUDANVIDIAThe platform decisionMade GPUs programmable for general computation. Arguably the single most consequential product decision in the history of this field.
  4. 2009ImageNetFei-Fei LiThe benchmark that started itFourteen million labelled images, built on the argument that data, not algorithms, was the bottleneck. She was right.
  5. 2011Watson wins JeopardyIBM
  6. 2011Siri ships on the iPhone 4SOctoberAppleAI in a pocketThe first time hundreds of millions of people spoke to an assistant. Expectations set here took a decade to meet.
  7. 2012AlexNet wins ImageNetSeptemberKrizhevsky, Sutskever & HintonThe moment it turnedWon by such a margin that the entire field changed direction within a year. Trained on two consumer gaming GPUs.
  8. 2013word2vecGoogleWords as vectors, with arithmetic that worked. The idea underneath every embedding and every vector database.
  9. 2014Generative adversarial networksIan GoodfellowMachines that generateTwo networks competing, one making and one judging. Generative modelling's first great leap, and the beginning of the deepfake problem.
  10. 2014Sequence to sequence learningSutskever, Vinyals & Le
  11. 2014Google acquires DeepMindGoogle
  12. 2014Amazon Echo and AlexaNovemberAmazonPut a microphone in the living room and normalised talking to a machine at home.
  13. 2015TensorFlowGoogle
  14. 2015OpenAI foundedDecemberOpenAIFounded as a non-profit to ensure artificial general intelligence benefits everyone. The structure and the mission have both been argued about ever since.
  15. 2015ResNetKaiming HeDepth becomes possibleResidual connections let networks go from tens of layers to hundreds. The most cited paper of the century, and the reason Transformers can stack at all.
  16. 2016AlphaGo defeats Lee SedolMarchDeepMindTen years earlyGo was supposed to be a decade away. Move 37 in game two was a move no human would have played, and it was correct.
  17. 2016PyTorchMetaBecame the language research is written in.
2017 to 2021

The Transformer era

One architecture replaced everything that came before it, and then scaling turned out to keep working long after everyone expected it to stop.

  1. 2017Attention Is All You NeedJuneGoogleThe TransformerRemoved recurrence, made sequence models parallel, and happened to saturate a GPU. Every model on this site is this architecture, stacked deeper.
  2. 2017AlphaZeroDeepMindLearned chess, shogi and Go from the rules alone, with no human games.
  3. 2018GPT-1JuneOpenAIPre-train, then fine-tune
  4. 2018BERTOctoberGoogleLanguage understanding, solved-ishSwept every language benchmark and went into Google Search within a year. For two years, the default answer to any NLP problem.
  5. 2018Turing Award to Hinton, LeCun and BengioACMThe formal acknowledgement that the people who kept working through two winters had been right.
  6. 2019GPT-2, and the decision not to release itFebruaryOpenAIHeld back as too dangerous to publish, which is now remembered as either responsible caution or very effective marketing.
  7. 2020GPT-3JuneOpenAIScale is the whole story175 billion parameters, and the discovery that a big enough model does tasks nobody trained it for. The paper that made scaling laws a strategy.
  8. 2020AlphaFold 2 solves protein structureNovemberDeepMindA fifty-year problemEnded a grand challenge of biology outright, and later won a Nobel Prize. Still the clearest case that this technology does science.
  9. 2021Anthropic foundedAnthropic
  10. 2021CLIP and DALL·EJanuaryOpenAIConnected images and language in one space, which is what made text-to-image possible at all.
  11. 2021GitHub CopilotJuneGitHubAI writes codeThe first AI product a large profession used daily, and the first to raise the licensing questions that are still unsettled.
2022 to 2023

The generative boom

Two years in which the technology left the laboratory entirely. The public arrived, the money arrived, and the regulators started writing.

  1. 2022Chinchilla scaling lawsMarchDeepMindShowed most large models were badly undertrained for their size. Quietly redirected how every lab spent its compute.
  2. 2022Midjourney opens in betaJulyMidjourney
  3. 2022Stable Diffusion released openlyAugustStability AIImage generation escapesLatent diffusion made generation run on a consumer GPU, and releasing the weights put it beyond anyone's control. The moment open weights became a strategy.
  4. 2022ChatGPT30 NovemberOpenAIThe fastest adoption in historyA research preview of a model that had been available for months through an API. The interface was the product, and it reached a hundred million people in two months.
  5. 2023LLaMA leaks, then Llama 2 ships openlyMetaMade capable open weights normal, and created the ecosystem every open model since has been built in.
  6. 2023GPT-4MarchOpenAIThe capability jumpPassed professional examinations, handled images, and set the bar the whole industry then chased for two years.
  7. 2023ClaudeMarchAnthropicIntroduced Constitutional AI, training a model against a written set of principles rather than human preference labels alone.
  8. 2023The pause letterMarchFuture of Life Institute
  9. 2023Mistral 7BSeptemberMistral AIEurope's answer, and proof that a small team could ship a competitive open model.
  10. 2023The EU AI Act is agreedDecemberEuropean UnionThe first comprehensive AI law anywhere. Its high-risk obligations took effect in August 2026.
2024 to 2026

Reasoning, agents and omni

Three new shapes in quick succession: models that spend compute thinking before answering, models that take actions rather than produce text, and models that treat voice, vision and text as one thing.

  1. 2024SoraFebruaryOpenAIVideo generation good enough to make the film industry take it seriously.
  2. 2024Claude 3MarchAnthropic
  3. 2024GPT-4oMayOpenAIThe omni modelOne model for text, vision and audio, with voice latency low enough to interrupt. A new model shape, not a bigger version of the old one.
  4. 2024o1 and reasoning modelsSeptemberOpenAIThinking before answeringSpending compute at inference rather than only at training. Opened a second scaling axis when the first was starting to look expensive.
  5. 2024Nobel Prizes for Hinton, Hassabis and JumperOctoberPhysics and ChemistryHinton for the foundations of machine learning, Hassabis and Jumper for AlphaFold. The field's formal arrival in science.
  6. 2024Model Context ProtocolNovemberAnthropicA standard for toolsAn open protocol for connecting models to tools and data. Adopted across the industry within a year, which almost never happens.
  7. 2025DeepSeek R1JanuaryDeepSeekThe cost assumption breaksFrontier-class reasoning, open weights, trained for a fraction of the assumed cost. It moved markets and reset what a frontier model was expected to cost.
  8. 2025Agents become the productFrom answers to actionsCoding agents, computer use and long-running tasks. The unit of work stopped being a reply and started being a job.
  9. 2025Mixture-of-experts becomes standardFrontier models decouple what they know from what they cost to run. Total parameters stop predicting price.
  10. 2026The EU AI Act's high-risk rules take effect2 AugustEuropean Union
  11. 2026Million-token context becomes ordinaryAnd context rot becomes the thing engineers actually worry about, because recall falls long before the limit.
  12. 2026The current frontierAugustWhere this page endsClaude Opus 5, GPT-5.6, Gemini 3.1 and Grok 4.6 at the closed frontier; DeepSeek, Kimi and Qwen months behind rather than generations. Every one of them is the 2017 Transformer, stacked deeper and trained longer.

Dates are the public announcement or publication, which is often later than the work and always earlier than the general availability people remember. An entry earns a place by being something the field still refers back to, so this is deliberately not a complete list of everything that happened. Corrections and additions are welcome.