Stone to Silicon

Timeline · 175 technologies

Artificial intelligence

c. 1305 to September 2026. Each one is placed by when it first worked.

Open the living tree

  1. c. 1305Llull's Ars MagnaRotating paper wheels that combine concepts mechanically, an early dream of reasoning by procedure.
  2. 1774Jaquet-Droz automataClockwork boys that write, draw and play music, programmed by swappable cams.
  3. 1843Lovelace on machine thoughtAda Lovelace sees that a computing engine could manipulate any symbols, not just numbers.
  4. 1943McCulloch-Pitts neuronThe first mathematical model of a neuron: a threshold unit that computes logic.
  5. 1948CyberneticsWiener names the science of feedback and control in animals and machines.
  6. 1949Hebbian learningCells that fire together wire together: the first theory of how networks learn.
  7. 1950Turing testTuring asks 'Can machines think?' and replaces it with a conversation game.
  8. 1954Georgetown-IBM translationAn IBM 701 translates 60 Russian sentences into English in public.
  9. 1956Dartmouth AI workshopA summer workshop names the field 'artificial intelligence' and gathers its founders.
  10. 1956Logic TheoristThe first program built to reason: it proves 38 theorems of Principia Mathematica.
  11. 1956Samuel's checkers learnerA program that improves at checkers by playing itself, and coins 'machine learning'.
  12. 1958PerceptronRosenblatt's machine learns to classify images by adjusting weights from its own mistakes.
  13. 1959General Problem SolverAn attempt at one program that solves any formalised problem via means-ends analysis.
  14. 1960ADALINEWidrow and Hoff train a neuron by gradient descent on squared error.
  15. 1965DendralThe first expert system infers molecular structures from mass spectra.
  16. 1966ELIZAA 200-line pattern matcher playing a therapist convinces users it understands them.
  17. 1966Shakey the robotThe first mobile robot that perceives, plans and acts on its own reasoning.
  18. 1968A* searchAn optimal, efficient path-finding algorithm born from a robot's navigation needs.
  19. 1969Perceptrons (book)Minsky and Papert prove single-layer perceptrons cannot compute XOR.
  20. 1970SHRDLUA program that understands English commands about a world of toy blocks.
  21. 1972MYCINAn expert system that diagnoses blood infections about as well as specialists.
  22. 1972PrologA language where programs are logical facts and the computer searches for proofs.
  23. 1974First AI winterAfter the Lighthill report, funding for AI research dries up in the UK and US.
  24. 1975Statistical speech recognitionHidden Markov models turn speech recognition into a probability problem.
  25. 1980NeocognitronFukushima builds a layered vision network with local features and pooling.
  26. 1980XCONAn expert system that configures computer orders saves DEC about $25 million a year.
  27. 1982Hopfield networkA recurrent network whose memories are energy minima, reviving neural net research.
  28. 1985Bayesian networksPearl gives machines a principled way to reason under uncertainty.
  29. 1985Boltzmann machineA stochastic network that learns hidden features by imitating a probability distribution.
  30. 1986BackpropagationThe chain rule, applied backwards, finally trains networks with hidden layers.
  31. 1987Second AI winterThe Lisp machine market collapses and expert systems prove brittle and costly to maintain.
  32. 1988Temporal-difference learningSutton's method learns predictions from the gap between successive guesses.
  33. 1989Convolutional neural networkLeCun trains a convolutional net by backprop to read handwritten ZIP codes.
  34. 1989Q-learningAn agent learns the value of every action in every state with no model of the world.
  35. 1990Recurrent neural networksNetworks with loops carry memory forward to process sequences like language.
  36. 1990Statistical machine translationIBM learns translation from millions of parallel Canadian parliament sentences.
  37. 1991Mixture of expertsA gating network routes each input to specialist subnetworks.
  38. 1992TD-GammonA neural network teaches itself backgammon to world-class level through self-play.
  39. 1995Support vector machineMaximum-margin classifiers with the kernel trick dominate machine learning for a decade.
  40. 1997Deep BlueIBM's chess machine beats world champion Garry Kasparov in a six-game match.
  41. 1997LSTMGated memory cells let networks remember over thousands of steps.
  42. 1998LeNet-5A convolutional net reads a large share of US bank checks.
  43. 1998MNIST dataset70,000 handwritten digits become machine learning's standard test.
  44. 2003Neural language modelBengio predicts the next word with a network that learns word vectors.
  45. 2005DARPA Grand ChallengeStanford's Stanley drives 212 km of desert with nobody aboard and wins.
  46. 2006Deep belief networksHinton shows deep networks can be trained by stacking layers one at a time.
  47. 2006Monte Carlo tree searchRandom playouts guide search, and computer Go suddenly improves.
  48. 2009GPU-trained neural netsGraphics cards speed neural network training up to 70-fold.
  49. 2009ImageNetFei-Fei Li's team labels 14 million images, and the dataset changes the field.
  50. 2011SiriA voice assistant ships to millions of phones.
  51. 2011Watson wins Jeopardy!IBM's question-answering system beats the two greatest Jeopardy! champions.
  52. 2012Deep learning (AlexNet)AlexNet cuts ImageNet error to 15.3%, and deep learning takes over AI.
  53. 2012Deep speech recognitionDeep networks cut speech recognition errors by about a quarter almost overnight.
  54. 2012Google Brain cat neuronA network trained on YouTube stills learns to detect cats with no labels.
  55. 2013Deep Q-network (Atari)One network learns dozens of Atari games from raw pixels and the score.
  56. 2013Variational autoencoderA network that learns a smooth latent space it can sample new data from.
  57. 2013word2vecWord vectors where king minus man plus woman lands near queen.
  58. 2014Attention mechanismLet the decoder look back at every input word, weighted by relevance.
  59. 2014Generative adversarial networkTwo networks compete, one forging images and one detecting forgeries.
  60. 2014Sequence to sequenceAn encoder LSTM reads a sentence and a decoder LSTM writes its translation.
  61. 2015Diffusion modelsLearn to reverse gradual noising, and you can generate data from noise.
  62. 2015Knowledge distillationA small student network learns from a large teacher's soft predictions.
  63. 2015ResNetSkip connections let networks grow to 152 layers and beat human ImageNet error.
  64. 2015TensorFlowGoogle open-sources its deep learning framework.
  65. 2016AlphaGoDeepMind's AlphaGo beats Lee Sedol 4-1 at Go, a decade ahead of forecasts.
  66. 2016Google Neural Machine TranslationGoogle Translate switches to neural networks and cuts errors by more than half.
  67. 2016PyTorchAn eager, Pythonic deep learning framework that researchers adopt en masse.
  68. 2016WaveNetA network generates raw audio sample by sample, and synthetic voices sound human.
  69. 2017AlphaGo ZeroStarting from only the rules, self-play surpasses every prior Go program in 40 days.
  70. 2017AlphaZeroOne algorithm masters chess, shogi and Go from scratch in under a day.
  71. 2017DeepfakesOpen-source face swapping makes convincing fake video available to anyone.
  72. 2017LibratusAn AI beats top professionals at heads-up no-limit Texas hold'em.
  73. 2017Proximal policy optimizationA simple, stable policy gradient method that becomes RL's default.
  74. 2017RL from human preferencesTrain a reward model from human comparisons, then optimise against it.
  75. 2017Transformer'Attention Is All You Need': drop recurrence, use attention, and train in parallel.
  76. 2018BERTA bidirectional transformer pretrained by filling in blanks sets 11 NLP records.
  77. 2018GPT-1OpenAI pretrains a transformer on books, then fine-tunes it for many tasks.
  78. 2019AlphaStarDeepMind reaches Grandmaster level in StarCraft II.
  79. 2019GPT-2A 1.5-billion-parameter model writes coherent paragraphs and is released in stages.
  80. 2019MuZeroMastering games without being told the rules, by learning its own model.
  81. 2019OpenAI FiveA team of five neural networks beats the Dota 2 world champions.
  82. 2020Denoising diffusion (DDPM)Ho, Jain and Abbeel make diffusion models match GANs on image quality.
  83. 2020Driverless robotaxi serviceMembers of the public hail cars with nobody in the driver's seat.
  84. 2020GPT-3175 billion parameters, and the model learns new tasks from examples in the prompt.
  85. 2020MMLU benchmark57 subjects of exam questions become the scoreboard for language models.
  86. 2020Neural scaling lawsLoss falls as a smooth power law in parameters, data and compute.
  87. 2020Retrieval-augmented generationGive the model a search engine so it can cite documents instead of guessing.
  88. 2020Vision transformerSplit an image into 16x16 patches, treat them as words, and a transformer sees.
  89. 2021CLIPImages and text learn a shared space from 400 million captioned web pictures.
  90. 2021DALL-EType 'an armchair in the shape of an avocado' and get pictures of one.
  91. 2021GitHub CopilotAn AI pair programmer autocompletes whole functions inside the editor.
  92. 2022Chain-of-thought promptingAsking a model to reason step by step sharply improves its math.
  93. 2022ChatGPTA free chat interface reaches 100 million users in about two months.
  94. 2022Chinchilla scalingDeepMind shows models were undertrained: use about 20 tokens per parameter.
  95. 2022Constitutional AIAnthropic trains a harmless assistant using AI feedback guided by written principles.
  96. 2022DALL-E 2Diffusion plus CLIP produces photorealistic images from any description.
  97. 2022InstructGPTRLHF makes a 1.3B model preferred over the 175B GPT-3.
  98. 2022LLM tool useLanguage models learn to call search, calculators and code, then read the results.
  99. 2022MidjourneyAn independent lab's image generator, run through Discord, defines AI art style.
  100. 2022PaLMGoogle's 540-billion-parameter model explains jokes and reasons step by step.
  101. 2022Stable DiffusionAn open-weights image model that runs on a gaming GPU.
  102. 2022WhisperAn open speech recogniser trained on 680,000 hours of multilingual audio.
  103. 2023AI music generationFull songs with vocals from a one-line prompt.
  104. 2023AutoGPTA viral open-source loop lets GPT-4 set its own subgoals and act on them.
  105. 2023ClaudeAnthropic's first public assistant, trained with Constitutional AI.
  106. 2023Claude 2Anthropic's public model with 100K context, then 200K in 2.1.
  107. 2023Commercial text-to-videoRunway's Gen-2 turns text prompts into short video clips for the public.
  108. 2023Gemini 1.0Google DeepMind's natively multimodal model family launches.
  109. 2023GPT-4A multimodal model scores around the top 10% on a simulated bar exam.
  110. 2023LlamaMeta's efficient models leak, and open-weights AI takes off.
  111. 2023Llama 2Meta releases open weights with a commercial licence and a chat model.
  112. 2023llama.cppQuantised large models run on a laptop CPU.
  113. 2023Mistral 7BA Paris startup's 7B open model beats Llama 2 13B on every benchmark tested.
  114. 2023MixtralAn open sparse mixture-of-experts model matches GPT-3.5 at a fraction of the cost.
  115. 2023Vision-language-action modelsRobots controlled by a web-trained vision-language model that outputs actions.
  116. 2024AlphaGeometryA neuro-symbolic system solves olympiad geometry near gold-medallist level.
  117. 2024AlphaProof IMO silverAn AI solves four of six International Math Olympiad problems.
  118. 2024Claude 3 familyOpus, Sonnet and Haiku: Anthropic reaches the frontier with vision.
  119. 2024Claude 3.5 SonnetA mid-size model beats the previous flagship, and becomes a favourite for coding.
  120. 2024Computer use agentsA model looks at screenshots and moves the mouse and keyboard like a person.
  121. 2024DeepSeek-V3A 671B-parameter open MoE model trained for about $5.6 million in GPU time.
  122. 2024Gemini 1.5 long contextA model that reads a million tokens: hours of video or whole codebases at once.
  123. 2024GPT-4oOne model hears, sees and speaks in real time with about human conversational latency.
  124. 2024Llama 3Meta trains on 15 trillion tokens and releases a 405B open frontier model.
  125. 2024Model Context ProtocolAn open standard for plugging tools and data into any AI assistant.
  126. 2024o3 and ARC-AGIA reasoning model scores 87.5% on ARC-AGI, a test built to resist memorisation.
  127. 2024Reasoning models (o1)Models trained with RL to think at length before answering.
  128. 2024Scaling MonosemanticityMillions of interpretable features are extracted from inside a production model.
  129. 2024SoraMinute-long, coherent video from a text prompt.
  130. 2024Test-time compute scalingAccuracy rises smoothly with how long the model is allowed to think.
  131. 2024Veo video modelGoogle DeepMind's high-definition text-to-video model.
  132. 2025Claude 3.7 SonnetThe first hybrid model: one model answers instantly or thinks at length.
  133. 2025Claude 4Claude Opus 4 and Sonnet 4 work autonomously on coding tasks for hours.
  134. 2025Claude CodeAn agent in the terminal that reads, edits, tests and commits across a codebase.
  135. 2025Claude Opus 4.5The first model above 80% on SWE-bench Verified.
  136. 2025Claude Sonnet 4.5An agentic coding model that sustains focus on tasks for over 30 hours.
  137. 2025DeepSeek-R1An open-weights reasoning model matches o1, and the recipe is published.
  138. 2025Gemini 2.5 ProGoogle's thinking model tops the leaderboards with a million-token context.
  139. 2025Gemini 3Google's next frontier model leads on reasoning and multimodal benchmarks.
  140. 2025Gemini RoboticsA frontier multimodal model folds origami and packs lunch boxes with robot arms.
  141. 2025GPT-5OpenAI merges fast answers and deep reasoning behind one router.
  142. 2025Humanity's Last ExamA 2,500-question expert benchmark built after the old ones saturated.
  143. 2025LLMs win IMO goldGeneral-purpose language models solve five of six olympiad problems in natural language.
  144. 2025Qwen3Alibaba's open family with switchable thinking, from 0.6B to 235B parameters.
  145. 2025Stargate compute build-outA planned $500 billion of US AI data centres over four years.
  146. 2025Veo 3 with native audioVideo generation with synchronised dialogue, sound effects and music.
  147. July 2025Grok 4xAI scales reinforcement learning on a 200,000-GPU cluster.
  148. July 2025Kimi K2 open MoEA trillion-parameter open-weight model built for agentic work.
  149. August 2025gpt-oss open-weight modelsOpenAI's first open-weight language models since GPT-2.
  150. September 2025AI sweeps ICPC World FinalsMachines beat every human team at the top college programming contest.
  151. September 2025DeepSeek sparse attentionLearned sparse attention makes long contexts cheap.
  152. September 2025Sora 2Video generation with synchronized sound, shipped as a social app.
  153. November 2025OpenClaw personal agentsAn open-source agent that lives in your chat apps and acts for you.
  154. December 2025GPT-5.2OpenAI's fast answer to Gemini 3, in Instant, Thinking and Pro tiers.
  155. January 2026Claude CoworkClaude Code's agent loop, opened up to people who do not code.
  156. February 2026Claude Opus 4.6Agent teams arrive in Claude Code as Opus leads on real work tasks.
  157. February 2026GLM-5Zhipu's open model tops the open-weight rankings.
  158. February 2026Gemini 3 Deep ThinkExtended reasoning passes 84% on ARC-AGI-2.
  159. February 2026Qwen3.5A natively multimodal open model with only 17B active parameters.
  160. February 2026Gemini 3.1 ProGoogle more than doubles Gemini 3 Pro's abstract reasoning score.
  161. March 2026GPT-5.4A model that uses a desktop better than the average person.
  162. April 2026Muse SparkMeta's superintelligence lab ships its first model.
  163. April 2026Claude Opus 4.7Opus gets sharper on the hardest software engineering tasks.
  164. April 2026GPT-5.5OpenAI pushes into research-level mathematics.
  165. April 2026DeepSeek-V4A 1.6-trillion-parameter open model with a million-token context.
  166. May 2026Gemini 3.5A Flash model overtakes the previous generation's Pro.
  167. June 2026Claude Fable 5 and Mythos 5A Mythos-class model released for general use, with its sibling held back.
  168. July 2026GPT-5.6Three tiers, Sol, Terra and Luna, after a government-restricted preview.
  169. July 2026Kimi K3The largest open-weight model yet, at 2.8 trillion parameters.
  170. July 2026AI perfect scores at IMOOfficially graded 42 out of 42 at the International Math Olympiad.
  171. July 2026Claude Opus 5Near-Fable capability at half the price.
  172. July 2026Gemini Robotics 2One model drives a humanoid from feet to fingertips.
  173. September 2026GPT-6 AstraThe first model pretrained on more than 100,000 GPUs at Stargate Texas.
  174. September 2026Claude Opus 5.5Frontier agentic coding gets about 40% cheaper.
  175. September 2026GPT-6 Sol and LunaGPT-6 capability at half the price of GPT-5.6.

Projections 3

  1. around 2029Autonomous AI researchersAI agents run a research programme end to end, from hypothesis to published result.
  2. around 2032Artificial general intelligenceAI that can do nearly all cognitive work a skilled remote worker can, at comparable cost.
  3. around 2060Whole-brain emulationA human brain's connections are scanned and simulated closely enough to reproduce its function.