Lines
The lines that run through the collection. A topic says what a record is about and there are five of them because the axis draws it; a line says what a record belongs to and is never drawn there. A record carries up to three lines, or none. 43 records carry none.
Agents 20Models that act through tools rather than only answering.
- Siri2010
- Siri in the iPhone2011
- Amazon Echo and Alexa2014
- Codex and Copilot2021
- A language model plans a robot’s actions2022
- Cicero plays Diplomacy2022
- LangChain: a model call becomes a link, not an answer2022
- SWE-bench: real tasks from GitHub2023
- The interface as a lever for an agent2024
- Two steps instead of an agent loop2024
- Computer use2024
- Model Context Protocol2024
- Search inside the chain of reasoning2025
- Claude 3.7 Sonnet and Claude Code2025
- o3 and tools inside the reasoning2025
- A small model is handed actions, not answers2025
- ChatGPT as an agent2025
- Claude Science2026
- ChatGPT Work: an agent that hands back finished work, not an answer2026
- An agent leaves the sandbox2026
Autonomous driving 27Vehicles driving public roads with nobody at the wheel.
- Warp: a systolic computer in a van1986
- ALV: the Strategic Computing land vehicle drives itself1986
- VaMoRs drives the Autobahn at 96 km/h1987
- Navlab: a van drives a road without stopping1988
- ALVINN: a neural network steers the van1988
- VaMP in dense traffic near Paris1994
- 1,600 km Munich to Odense, 95 percent autonomous1995
- Mobileye is founded1999
- The first Grand Challenge: nobody finished2004
- Stanley wins the Grand Challenge2005
- The Urban Challenge2007
- Google reveals its self-driving project2010
- Nevada authorises autonomous vehicles by statute2011
- The first licence for a driverless car2012
- The Mobileye prospectus, and a vision company's first numbers2014
- Tesla Autopilot2015
- Waymo becomes a company2016
- Baidu opens the Apollo code2017
- Intel buys Mobileye: $15.3 billion announced, $14.5 billion booked2017
- The first pedestrian killed by an autonomous car2018
- Waymo One2018
- San Francisco opens to everyone2024
- Tesla reaches the streets of Austin2025
- Freeways2025
- Waymo at San Francisco airport2026
- Nevada admits three robotaxi fleets at once2026
- Two hundred million driverless miles2026
Compute 111What the computing runs on: processors, accelerators, data centres, their power draw and the cost of compute.
- First Draft of a Report on the EDVAC1945
- Giant Brains, or Machines That Think1949
- MESM: the first electronic computer in Kyiv1951
- SNARC1951
- The first neural network simulation1954
- SAGE goes on duty1958
- Mark I Perceptron: the perceptron in hardware1960
- Moore's law1965
- MIR: a machine that understands formulas1969
- The MIT Lisp machine1977
- Kung and Leiserson's systolic arrays1978
- Strategic Computing: DARPA's plan1983
- WISARD: a neural network made of memory chips1984
- The Connection Machine1985
- Warp: a systolic computer in a van1986
- Hsu's chess chip: from ChipTest to Deep Blue1986
- Mead names neuromorphic systems1990
- Neural network chips from Intel and Adaptive Solutions1990
- ANNA: a convolutional network on an analog chip1991
- Deep Blue defeats Kasparov1997
- GeForce 2561999
- The Google File System2003
- MapReduce2004
- Amazon EC22006
- CUDA becomes generally available2007
- Deep networks on graphics processors2009
- The Google network that found cats2012
- AlexNet at ImageNet2012
- Asynchronous actor-critic (A3C)2016
- Google's Tensor Processing Unit2016
- GNMT: neural translation in Google Translate2016
- The sparsely-gated mixture-of-experts layer2017
- NVIDIA tensor cores2017
- Intel buys Mobileye: $15.3 billion announced, $14.5 billion booked2017
- Seven nanometres enters volume production2018
- AlphaStar at StarCraft II2018
- OpenAI Five beats the Dota 2 champions2019
- Training priced in kilowatt-hours2019
- Pluribus: six-player poker2019
- Ascend 910 and MindSpore2019
- Habana Labs: $2 billion in the press release, $1.7 billion in the annual report2019
- Scaling laws2020
- NVIDIA A1002020
- HBM2E: 460 gigabytes a second per stack2020
- The world's fastest machine with no accelerator2021
- GPT-3 in megawatt-hours2021
- Low-rank adaptation (LoRA)2021
- Three eras of compute: 21, 6 and 10 months2022
- Chinchilla: compute-optimal training2022
- NVIDIA H1002022
- FlashAttention2022
- HBM3 for the H1002022
- Ten countries club together for one supercomputer2022
- A doubling every 2.46 years, not every year2022
- The CHIPS Act: money for chip manufacturing in the United States2022
- Trainium reaches general availability2022
- Compute gets a numeric ceiling2022
- BLOOM's footprint, beyond the current2022
- The A800: a part designed to the threshold2022
- Europe puts its machine on NVIDIA accelerators2022
- Three nanometres in volume production2022
- 700,000 litres to train, a bottle to talk2023
- QLoRA2023
- Open weights out of Abu Dhabi2023
- vLLM: the memory of attention starts being paged2023
- The bottleneck moves to packaging2023
- Arabic as a third of the training set2023
- The threshold is rewritten, the workaround goes2023
- 4,480 Hopper accelerators in Barcelona2023
- The first high-aperture scanner ships to Intel2023
- Fine-grained and shared experts2024
- GRPO: reinforcement learning without a critic2024
- Weights with three states instead of numbers2024
- India buys accelerators as a public service2024
- Blackwell2024
- 960 megawatts straight from a reactor2024
- Compressive memory instead of a growing KV cache2024
- Microsoft's emissions up 29 per cent on construction2024
- Scaling laws for sparse autoencoders2024
- NVIDIA becomes the most valuable company in the world2024
- 14.3 million tonnes, 48 per cent above 20192024
- A fifth of Ireland's electricity goes to data centres2024
- Coverage grows with the number of attempts2024
- Compute at answering time against a bigger model2024
- o1 and compute at answering time2024
- A shut reactor restarted for data centres2024
- The regulator stops the direct feed2024
- DeepSeek-V32024
- Stargate2025
- Long context inside reinforcement learning2025
- The attention cache at two bits2025
- Seventeen per cent in a day2025
- A thousand examples and a forced budget2025
- Reasoning by depth rather than by tokens2025
- One billion against four hundred and five2025
- A constant number of cache pages2025
- Partly re-initialising an expert's weights2025
- A small model is handed actions, not answers2025
- The cost of an answer is memory, not parameters2025
- The first four trillion2025
- The Rubin platform enters full production2026
- Global AI computing capacity doubles every seven months2026
- The United States opens H200 sales to China2026
- A 25 per cent tariff on advanced chips2026
- Nine hundred and sixty-five billion2026
- Jalapeño chip2026
- Performance per dollar in chips grows 49 per cent a year2026
- NVIDIA: eighty-nine billion in a quarter from data centres2026
- Epoch puts Huawei below four per cent of Nvidia2026
- The largest data centre's power doubles every ten months2026
- Customs data points to three billion dollars of diverted chips2026
Data 122Datasets, corpora and the systems that processed them.
- As We May Think1945
- A photograph in a computer’s memory1957
- GUHA: a machine that proposes hypotheses from data1966
- The Cyc project begins1984
- The UCI Machine Learning Repository1987
- Statistical part-of-speech tagging1988
- Statistical machine translation1988
- Resource Management: a shared speech recognition test1989
- Latent semantic analysis1990
- The TIMIT corpus1990
- WordNet: a dictionary arranged by meanings1990
- The Switchboard corpus1992
- Brill's tagger: rules the machine learns1992
- NIST's shared test on handwritten characters1992
- TREC: a shared test for retrieval1992
- Word classes from mutual information1992
- The Penn Treebank1993
- The IBM translation models1993
- GroupLens: predictions from other readers' ratings1994
- MUC-6: named entity recognition as a task of its own1995
- Unit selection: a voice from pieces of recording1996
- A blind evaluation for face recognition1996
- Sounds found by how they sound1996
- Lexicalised statistical parsing1997
- PageRank and web scale1998
- LeNet-5 and MNIST1998
- RDF: a fact as a triple with a global address1999
- Informedia: searching video by what is said1999
- Searching recorded speech: the SDR track2000
- CoNLL-2000: eleven systems on one split2000
- Wikipedia launches2001
- Data outweighs the algorithm2001
- The TREC video track: a shared test for video search2001
- Object recognition as translation2002
- Musical genre recognised from sound2002
- Amazon: similar items instead of similar customers2003
- Words and pictures: a joint model2003
- Shazam: a song recognised from a snippet2003
- The Google File System2003
- The ESP Game: image labels as play2004
- ConceptNet: common sense assembled from visitors' sentences2004
- A hundred and one classes instead of five2004
- MapReduce2004
- A contest where everyone is scored the same way2005
- Gradient histograms find pedestrians2005
- Object outlines drawn by a website’s visitors2005
- Mechanical Turk: human work ordered through an API2005
- Statistical translation at Google2006
- WMT-06: the formula and the judges over the same outputs2006
- A 3D scene from strangers’ photographs2006
- Apache Hadoop2006
- YAGO: Wikipedia sewn onto WordNet2007
- A hole in a picture filled from other photographs2007
- Faces from the news as a recognition benchmark2007
- DBpedia: a Wikipedia you can query2007
- Freebase: a fact base you could write to2008
- Eighty million images at 32 × 322008
- The unreasonable effectiveness of data2009
- CIFAR-10: tiny images with labels2009
- ImageNet2009
- The Netflix Prize2009
- Who the crowdworkers are: six surveys over twenty months2010
- The first ILSVRC2010
- A crawl of the web becomes available to anyone2012
- Apache Spark2012
- A knowledge graph inside Google Search2012
- Wikidata: Wikipedia's shared facts, in the public domain2012
- Face recognition accuracy by sex, race and age2012
- MS COCO: objects in context2014
- DeepFace: face recognition near human level2014
- FaceNet: a face as 128 bytes2015
- LibriSpeech: a thousand hours of audiobooks2015
- VQA: questions put to an image2015
- Fourteen arms and 800,000 grasp attempts2016
- GDPR: a right not to be subject to a solely automated decision2016
- SQuAD: a hundred thousand questions on Wikipedia passages2016
- AudioSet: a vocabulary of sounds and two million clips2017
- What data work actually pays: 3.8 million tasks measured2017
- Gender Shades: three commercial classifiers measured2018
- GLUE: nine tasks, one score, closed tests2018
- SQuAD 2.0: the question it is right to refuse2018
- StyleGAN: a style-controlled face generator2018
- SuperGLUE: a set built for years lasted twenty months2019
- C4: the web put through six rules2019
- Retrieval-augmented generation (RAG)2020
- Masakhane: translation made by the speakers2020
- The Pile: 825 gibibytes from twenty-two sources2020
- CLIP2021
- Canadian regulators on collecting faces without consent2021
- MATH: competition mathematics2021
- Stochastic Parrots: can a language model be too big2021
- What is actually inside the corpus T5 was trained on2021
- LAION-400M: image-caption pairs in the open2021
- GSM8K: school problems and a verifier2021
- A language model on 0.94 gigabytes of text2021
- Chinchilla: compute-optimal training2022
- LAION-5B: 5.85 billion pairs, a third of them not in English2022
- ROOTS: a corpus whose sources were mostly chosen by people2022
- Artists sue the image generators2023
- MusicLM: music from a text description2023
- A stock agency counts twelve million photographs2023
- Segment Anything2023
- Three per cent of a training budget, spent on one language2023
- LLaVA: an open assistant that sees2023
- Open weights out of Abu Dhabi2023
- Novelists count what is inside Books22023
- Books3: a hundred and ninety-six thousand books named2023
- Arabic as a third of the training set2023
- The Authors Guild becomes a party2023
- One dataset from many robot bodies2023
- A transparency index: nobody says what the model is made of2023
- Illegal material found in an open set, distribution halted2023
- A hundred examples of verbatim output2023
- Self-play against one's own earlier state2024
- The same rig in many places2024
- A twenty-three point gap in Korean2024
- Model collapse on a model's own output2024
- Five thousand puzzles instead of a corpus2025
- A million trajectories from one Chinese fleet2025
- California requires training data to be disclosed2026
- A billion and a half for books2026
- The United States tells a court that training is fair use2026
Defence 28Military and intelligence use.
- Wiener's anti-aircraft predictor1942
- SAGE goes on duty1958
- Phalanx: a machine that closes the loop itself1980
- Strategic Computing: DARPA's plan1983
- ALV: the Strategic Computing land vehicle drives itself1986
- FRESH: an expert system at the Pacific Fleet headquarters1987
- Pilot's Associate: a minute of flight, two hours of computing1988
- Terminal guidance by image matching goes to war1991
- DART: the scheduler that repaid thirty years1991
- A blind evaluation for face recognition1996
- DARPA sets speed thresholds for walking over rough ground2009
- The Pentagon writes down what an autonomous weapon is2012
- Project Maven: computer vision joins intelligence work2017
- Google leaves Maven and writes down what it will not do2018
- Face recognition goes to war2022
- The political declaration on military AI2023
- Brave1: a state entry point for defence technology2023
- The UN General Assembly votes on autonomous weapons2023
- OpenAI drops the ban on "military and warfare"2024
- A model flies a fighter against a live pilot2024
- Claude enters a classified intelligence environment2024
- Google removes its pledge not to build weapons2025
- Maven becomes the Pentagon's costliest AI programme2025
- Claude Gov: a model built for the enclave2025
- The Pentagon contracts all four frontier laboratories at once2025
- The Pentagon labels Anthropic a supply-chain risk2026
- GPT-5.6-Cyber and Daybreak expansion2026
- A court strikes down the exclusion of Anthropic2026
Embodiment 61Body, grasp and manipulation: a machine acting in the physical world.
- Grey Walter's tortoises1949
- Unimate on the General Motors line1961
- Shakey1969
- Freddy assembles a toy from a heap of parts1973
- Scheinman's arm goes to JPL1973
- WABOT-11973
- TAIR: a robot run by a neuron-like network1975
- The Stanford Cart crosses a room1979
- Phalanx: a machine that closes the loop itself1980
- One leg balances in three dimensions1983
- Artificial potential fields avoid obstacles in real time1985
- Subsumption architecture1986
- A six-legged robot walks with no central control1989
- Elephants Don't Play Chess1990
- Simultaneous map building and localization1991
- Honda P2: the first humanoid with its own battery1996
- The first RoboCup1997
- Kismet: a robot regulates interaction with its face1998
- Sony AIBO1999
- ASIMO2000
- Roomba2002
- One camera maps in real time2003
- iRobot opens its books: 1.2 million Roombas2005
- Grasping unseen objects, learned from synthetic images2006
- Tracking and mapping in separate threads2007
- Hundreds of robots coordinating in a warehouse2008
- DARPA sets speed thresholds for walking over rough ground2009
- Eleven identical robots in eleven laboratories2010
- A PR2 folds fifty towels in a row2010
- Kinect2010
- Vision and control trained as one2015
- Fourteen arms and 800,000 grasp attempts2016
- Grasping learned from synthetic data2017
- Closed-loop grasping at 96 percent2018
- In-hand manipulation trained only in simulation2018
- ANYmal, a quadruped trained in simulation2019
- A robot hand solves a Rubik’s cube2019
- A language model plans a robot’s actions2022
- A quadruped learns to walk in one hour2022
- One transformer, seven hundred tasks2022
- Diffusion policies for visuomotor control2023
- Bimanual teleoperation for twenty thousand dollars2023
- A web-trained model outputs robot actions2023
- Toyota announces large behavior models2023
- One dataset from many robot bodies2023
- The same rig in many places2024
- A model flies a fighter against a live pilot2024
- A humanoid at sixteen thousand dollars2024
- Open weights for a vision-language-action model2024
- The first commercial humanoid deployment2024
- π0, and open weights for a robot foundation2024
- A million trajectories from one Chinese fleet2025
- Gemini Robotics reaches physical control2025
- An open foundation model for humanoids2025
- BMW contradicts a humanoid claim2025
- Humanoids run a half-marathon2025
- The first independent evaluation of generalist policies2025
- The first World Humanoid Robot Games2025
- Google introduces Gemini Robotics 22026
- Humanoids counted for the first time by half-year2026
- A humanoid maker opens its books2026
Evaluation 119Tests, contests and independent measurements of capability.
- Turing test1950
- Mac Hack VI at a tournament1967
- PARRY, a model of paranoia1971
- The computer and acute abdominal pain in Leeds1972
- HARPY1976
- CASNET: a glaucoma model before ophthalmologists1977
- A blinded evaluation of MYCIN's advice1979
- INTERNIST-I falls short of experts1982
- PUFF: an expert system in a hospital's daily work1982
- The UCI Machine Learning Repository1987
- Lipreading cuts a recogniser's errors1988
- Sphinx1988
- Resource Management: a shared speech recognition test1989
- The TIMIT corpus1990
- The first Loebner contest1991
- A network recognises heart attacks1991
- The Switchboard corpus1992
- NIST's shared test on handwritten characters1992
- TREC: a shared test for retrieval1992
- The Penn Treebank1993
- Chinook becomes world champion1994
- BM251994
- Sign language sentences from video1995
- MUC-6: named entity recognition as a task of its own1995
- A neural network finds faces in photographs1995
- Deep Blue wins a game against Kasparov1996
- A blind evaluation for face recognition1996
- Deep Blue defeats Kasparov1997
- The first RoboCup1997
- Chen and Goodman's comparison of smoothing1998
- LeNet-5 and MNIST1998
- Searching recorded speech: the SDR track2000
- CoNLL-2000: eleven systems on one split2000
- The TREC video track: a shared test for video search2001
- Musical genre recognised from sound2002
- BLEU: a formula scores translation quality2002
- Face recognition measured on 37,437 people2003
- The first Grand Challenge: nobody finished2004
- A hundred and one classes instead of five2004
- A contest where everyone is scored the same way2005
- Blizzard Challenge: speech synthesis on shared data2005
- Stanley wins the Grand Challenge2005
- WMT-06: the formula and the judges over the same outputs2006
- Computer-aided detection worsens screening2007
- Faces from the news as a recognition benchmark2007
- The Urban Challenge2007
- CIFAR-10: tiny images with labels2009
- ImageNet2009
- The Netflix Prize2009
- DARPA sets speed thresholds for walking over rough ground2009
- Eleven identical robots in eleven laboratories2010
- Kaggle opens its competitions2010
- The first ILSVRC2010
- Watson on Jeopardy!2011
- Netflix: what went from the Prize into production2012
- Face recognition accuracy by sex, race and age2012
- AlexNet at ImageNet2012
- MS COCO: objects in context2014
- VQA: questions put to an image2015
- AlphaGo versus Lee Sedol2016
- How a risk scale's errors fell by race2016
- SQuAD: a hundred thousand questions on Wikipedia passages2016
- The developer's reply: equal accuracy for both groups2016
- Three fairness conditions that do not meet2016
- Human parity in speech recognition2016
- A deep network detects diabetic retinopathy2016
- A network tells skin cancers apart like dermatologists2017
- Gender Shades: three commercial classifiers measured2018
- GLUE: nine tasks, one score, closed tests2018
- SQuAD 2.0: the question it is right to refuse2018
- SuperGLUE: a set built for years lasted twenty months2019
- A hospital algorithm underrates Black patients' illness2019
- NIST measures demographic gaps across 189 algorithms2019
- A model reads mammograms better than radiologists2020
- MMLU: 57 subjects the model did not study2020
- AlphaFold 2 at CASP142020
- MATH: competition mathematics2021
- Epic's sepsis model misses two thirds2021
- HumanEval: code checked by tests2021
- GSM8K: school problems and a verifier2021
- BIG-bench: 204 tasks assembled by 442 authors2022
- HELM: thirty models, the same forty-two scenarios2022
- A language model graded like a physician2022
- Three per cent of a training budget, spent on one language2023
- Chatbot Arena: a score that cannot be studied in advance2023
- Physicians prefer the model's answers2023
- SWE-bench: real tasks from GitHub2023
- A transparency index: nobody says what the model is made of2023
- Britain founds the AI Safety Institute2023
- GPQA: questions you cannot google2023
- A twenty-three point gap in Korean2024
- A claimed context length and a real one2024
- The interface as a lever for an agent2024
- A first exam for open models in Ukrainian2024
- Two steps instead of an agent loop2024
- Silver at the International Mathematical Olympiad2024
- o3 on ARC-AGI2024
- One billion against four hundred and five2025
- Without a verifier, answering-time compute loses2025
- BMW contradicts a humanoid claim2025
- Humanoids run a half-marathon2025
- The cost of an answer is memory, not parameters2025
- The first independent evaluation of generalist policies2025
- Gold at the International Mathematical Olympiad2025
- The first World Humanoid Robot Games2025
- The ARC Prize 2025 technical report2026
- METR revises the horizon doubling to 89 days2026
- ARC-AGI-3 and 0.51 per cent2026
- Britain evaluates Mythos cyber capabilities independently2026
- A human baseline is measured for ARC-AGI-32026
- METR cannot measure GPT-5.6 Sol because it cheats2026
- The first ARC-AGI-3 milestone prize2026
- RedNote claims 42 out of 42 at the mathematical olympiad2026
- Every frontier model tested attempted to cheat2026
- Two hundred million driverless miles2026
- The capability frontier moves 14 points a year2026
- FrontierMath Erdos: 68 unsolved problems as a benchmark2026
- GPT-6 Astra reaches 99.9 per cent on ARC-AGI-32026
- Astra leads in mathematics but not in software engineering2026
Generative media 45Generating images, sound and video.
- Unit selection: a voice from pieces of recording1996
- A voice from a statistical model: HTS1999
- A texture grown from a sample1999
- Inpainting: a restorers’ word for a machine2000
- Image Analogies: a filter from a pair of examples2001
- A hole in a picture filled from other photographs2007
- The variational autoencoder2013
- Generative adversarial networks2014
- Show and Tell: a caption for an image2014
- Diffusion models2015
- WaveNet2016
- pix2pix: one model for image translation2016
- Tacotron: speech straight from letters2017
- Deepfakes: face swapping with open tools2018
- StyleGAN: a style-controlled face generator2018
- NeRF: a scene as a continuous function2020
- Jukebox: songs with singing as raw audio2020
- Denoising diffusion probabilistic models2020
- DALL·E2021
- LAION-400M: image-caption pairs in the open2021
- LAION-5B: 5.85 billion pairs, a third of them not in English2022
- Imagen: a large language encoder for generation2022
- Midjourney opens its beta to everyone2022
- DALL·E 2 opens in a paid beta2022
- Stable Diffusion2022
- VALL-E: a stranger's voice from three seconds2023
- China requires marks on synthesised content2023
- Artists sue the image generators2023
- MusicLM: music from a text description2023
- A stock agency counts twelve million photographs2023
- Minnesota makes an election deep fake a crime2023
- China licenses generative services but not research2023
- A contract in which AI is not a writer2023
- A digital replica needs consent and payment2023
- Sora2024
- Suno v3 opens to everyone2024
- A law that lived fifteen days2024
- Genie 32025
- Sora 22025
- India regulates synthetically generated information2026
- The SIGGRAPH 2026 technical papers awards2026
- Muse Image and Muse Video2026
- Muse Spark 1.12026
- A Munich court finds training on music infringing2026
- Synthetic content must be marked2026
Harm 46Documented harm to people from these systems, and the response to it.
- A Super Bowl crowd’s faces checked against a database2001
- Computer-aided detection worsens screening2007
- Illinois requires consent to collect biometrics2008
- Face recognition accuracy by sex, race and age2012
- Tay taken down within three days2016
- How a risk scale's errors fell by race2016
- The developer's reply: equal accuracy for both groups2016
- Three fairness conditions that do not meet2016
- Gender Shades: three commercial classifiers measured2018
- Deepfakes: face swapping with open tools2018
- The first pedestrian killed by an autonomous car2018
- California: a bot must say it is a bot2018
- San Francisco forbids itself face recognition2019
- Training priced in kilowatt-hours2019
- A hospital algorithm underrates Black patients' illness2019
- NIST measures demographic gaps across 189 algorithms2019
- An arrest on a false match in Detroit2020
- Washington puts face recognition under conditions2020
- IBM stops selling general purpose face recognition2020
- Amazon pauses police use of Rekognition for a year2020
- Timnit Gebru leaves Google over a paper2020
- Canadian regulators on collecting faces without consent2021
- Stochastic Parrots: can a language model be too big2021
- What is actually inside the corpus T5 was trained on2021
- GPT-3 in megawatt-hours2021
- Epic's sepsis model misses two thirds2021
- UNESCO adopts a recommendation on the ethics of AI2021
- China puts recommendation algorithms on a register2022
- Face recognition goes to war2022
- A court bars Clearview from selling its face database to companies2022
- BLOOM's footprint, beyond the current2022
- HELM: thirty models, the same forty-two scenarios2022
- A data centre's water stops being a secret2022
- China requires marks on synthesised content2023
- The labellers in Nairobi who taught ChatGPT to refuse2023
- 700,000 litres to train, a bottle to talk2023
- Minnesota makes an election deep fake a crime2023
- Illegal material found in an open set, distribution halted2023
- A synthesised voice falls under a 1991 statute2024
- Microsoft's emissions up 29 per cent on construction2024
- A cloned voice in campaign calls2024
- 14.3 million tonnes, 48 per cent above 20192024
- A law that lived fifteen days2024
- Settlements over harm from chatbots2026
- States block Grok2026
- China's anthropomorphic services rules take effect2026
Law and policy 95Rules, courts and state policy.
- The ALPAC report1966
- The Lighthill Report1973
- Japan's Fifth Generation project1981
- Britain's Alvey Programme1983
- Strategic Computing: DARPA's plan1983
- Europe's ESPRIT programme1984
- FDA approves a computer second read of mammograms1998
- A Super Bowl crowd’s faces checked against a database2001
- Illinois requires consent to collect biometrics2008
- Nevada authorises autonomous vehicles by statute2011
- The first licence for a driverless car2012
- The Pentagon writes down what an autonomous weapon is2012
- GDPR: a right not to be subject to a solely automated decision2016
- The first national AI strategy2017
- China's New Generation AI Development Plan2017
- Deepfakes: face swapping with open tools2018
- The FDA admits a diagnosis without a physician2018
- California: a bot must say it is a bot2018
- Executive Order 13859: the American AI Initiative2019
- The EU's ethics guidelines for trustworthy AI2019
- San Francisco forbids itself face recognition2019
- The OECD Council adopts a recommendation on AI2019
- Illinois: consent to an AI assessment of an interview2019
- Washington puts face recognition under conditions2020
- The Pile: 825 gibibytes from twenty-two sources2020
- Canadian regulators on collecting faces without consent2021
- The draft European AI Act2021
- UNESCO adopts a recommendation on the ethics of AI2021
- New York requires audits of hiring tools2021
- SenseTime's listing stopped three days before pricing2021
- China puts recommendation algorithms on a register2022
- A court bars Clearview from selling its face database to companies2022
- The CHIPS Act: money for chip manufacturing in the United States2022
- Compute gets a numeric ceiling2022
- The first suit over training on other people's code2022
- The A800: a part designed to the threshold2022
- A data centre's water stops being a secret2022
- China requires marks on synthesised content2023
- Artists sue the image generators2023
- A stock agency counts twelve million photographs2023
- The political declaration on military AI2023
- Novelists count what is inside Books22023
- Books3: a hundred and ninety-six thousand books named2023
- Voluntary commitments to the White House2023
- Minnesota makes an election deep fake a crime2023
- China licenses generative services but not research2023
- The Authors Guild becomes a party2023
- A contract in which AI is not a writer2023
- A transparency index: nobody says what the model is made of2023
- The threshold is rewritten, the workaround goes2023
- Executive Order 141102023
- Britain founds the AI Safety Institute2023
- The Bletchley Declaration2023
- A digital replica needs consent and payment2023
- Political agreement on the EU AI Act2023
- The UN General Assembly votes on autonomous weapons2023
- A hundred examples of verbatim output2023
- OpenAI drops the ban on "military and warfare"2024
- A synthesised voice falls under a 1991 statute2024
- India buys accelerators as a public service2024
- A state publishes weights in traditional Chinese2024
- The Seoul summit2024
- A cloned voice in campaign calls2024
- The European AI Act enters into force2024
- The veto of California's SB 10472024
- A law that lived fifteen days2024
- The regulator stops the direct feed2024
- The Action Summit in Paris2025
- America's AI Action Plan2025
- Obligations for general-purpose models2025
- California requires training data to be disclosed2026
- Settlements over harm from chatbots2026
- The United States opens H200 sales to China2026
- A 25 per cent tariff on advanced chips2026
- Korea's framework AI act takes effect2026
- States block Grok2026
- India regulates synthetically generated information2026
- The Pentagon labels Anthropic a supply-chain risk2026
- The Supreme Court leaves the human authorship rule intact2026
- The White House asks Congress to preempt state law2026
- Colorado repeals and rewrites its AI act2026
- A jury rejects Musk's claim against OpenAI2026
- An executive order on frontier model security2026
- Export control reaches the model itself, not the chips2026
- GPT-5.6 previewed by government permission2026
- The export controls come off and Fable 5 returns2026
- China's anthropomorphic services rules take effect2026
- A billion and a half for books2026
- Europe postpones its own rules2026
- A Munich court finds training on music infringing2026
- Synthetic content must be marked2026
- Nevada admits three robotaxi fleets at once2026
- A court strikes down the exclusion of Anthropic2026
- The United States tells a court that training is fair use2026
- Customs data points to three billion dollars of diverted chips2026
Learning theory 204Mathematical results about learning and its limits.
- Wiener's anti-aircraft predictor1942
- Behavior, Purpose and Teleology1943
- McCulloch and Pitts neuron model1943
- Theory of Games and Economic Behavior1944
- First Macy conference on feedback1946
- Turing lecture to the London Mathematical Society1947
- A Mathematical Theory of Communication1948
- Intelligent Machinery1948
- Cybernetics1948
- Hebb's learning rule1949
- Turing test1950
- A Stochastic Approximation Method1951
- SNARC1951
- Design for a Brain1952
- The Metropolis method1953
- The first neural network simulation1954
- Pattern recognition and modern computers1955
- IRE Symposium on Information Theory1956
- Dynamic Programming1957
- Syntactic Structures1957
- Pandemonium1958
- Perceptron1958
- Ledley and Lusted: logic and probability of diagnosis1959
- Some Studies in Machine Learning Using the Game of Checkers1959
- Visual cortex cells respond to orientation1959
- Hao Wang: Principia in minutes1960
- The Kalman filter1960
- ADALINE and the least-mean-squares rule1960
- The Davis-Putnam procedure with splitting1961
- Simple and complex cells, cortical columns1962
- The perceptron convergence proof1962
- Principles of Neurodynamics1962
- Matchboxes that learn to play1963
- A Formal Theory of Inductive Inference1964
- Glushkov: Introduction to Cybernetics1964
- The resolution principle1965
- Fuzzy Sets1965
- Amosov: Modeling of Thinking and the Mind1965
- Nearest neighbor pattern classification1967
- A Theory of Adaptive Pattern Classifiers1967
- Speech discrimination by dynamic programming1968
- The A* algorithm1968
- Schlesinger: unsupervised learning as repeated learning1968
- Applied Optimal Control1969
- Perceptrons: the limits of a single layer1969
- The Baum-Welch algorithm1970
- Glushkov: the 'evidence algorithm' programme1970
- Reverse-mode automatic differentiation1970
- Cook's theorem1971
- Group Method of Data Handling1971
- Weighting a term by its rarity in the collection1972
- Karp's twenty-one problems1972
- Werbos's thesis on computing derivatives backwards1974
- The document as a vector of terms1975
- Cognitron1975
- Adaptation in Natural and Artificial Systems1975
- Statistical speech recognition1976
- Dynamic time warping1978
- Kung and Leiserson's systolic arrays1978
- Edges as zero crossings, at several scales1980
- Neocognitron1980
- The Chinese Room1980
- Image registration by the intensity gradient1981
- A model from a minimal sample, not from all the data1981
- Optical flow as a smoothness problem1981
- Two views, eight points, a linear system1981
- Hopfield networks1982
- Pawlak: rough sets1982
- Marr: vision as computation on three levels1982
- Self-organising maps1982
- Simulated annealing1983
- Actor and critic1983
- A Theory of the Learnable1984
- Artificial potential fields avoid obstacles in real time1985
- Bayesian networks1985
- The Boltzmann machine1985
- Induction of Decision Trees1986
- The calculus of constructions1986
- Back-propagation1986
- An edge detector derived from three criteria1986
- Spatial uncertainty as covariance between frames1986
- The Society of Mind1986
- Parallel Distributed Processing1986
- NETtalk1987
- Statistical machine translation1988
- Temporal-difference learning1988
- Probabilistic Reasoning in Intelligent Systems1988
- The tutorial account of hidden Markov models1989
- Q-learning1989
- The universal approximation theorem1989
- A convolutional network on zip codes1989
- Latent semantic analysis1990
- Mead names neuromorphic systems1990
- Finding Structure in Time1990
- A face as a point in an eigenvector space1991
- Adaptive mixtures of local experts1991
- Simultaneous map building and localization1991
- The vanishing gradient problem1991
- The kernel trick and support vectors1992
- Shape and motion from a rank-three matrix1992
- Word classes from mutual information1992
- Genetic programming1992
- TD-Gammon1992
- The IBM translation models1993
- Why long-term dependencies are not learned1994
- BM251994
- AdaBoost1995
- Kneser-Ney smoothing1995
- Support-vector networks1995
- Maximum entropy for language processing1996
- Probabilistic roadmaps1996
- The lasso1996
- Lexicalised statistical parsing1997
- LSTM1997
- Chen and Goodman's comparison of smoothing1998
- LeNet-5 and MNIST1998
- Reinforcement Learning: An Introduction1998
- A feature that survives scale and rotation1999
- Causality2000
- Conditional random fields2001
- Data outweighs the algorithm2001
- Gradient boosting2001
- Random forests2001
- Viola-Jones face detection2001
- Object recognition as translation2002
- Latent Dirichlet allocation2003
- Words and pictures: a joint model2003
- A neural probabilistic language model2003
- One camera maps in real time2003
- OWL: an ontology whose conclusions compute2004
- Gradient histograms find pedestrians2005
- The same robustness, three times faster2006
- Deep belief networks2006
- Tracking and mapping in separate threads2007
- The unreasonable effectiveness of data2009
- Deep networks on graphics processors2009
- Deep networks in speech recognition2009
- Rectified linear units2010
- The Google network that found cats2012
- Dropout2012
- AlexNet at ImageNet2012
- word2vec2013
- The variational autoencoder2013
- DQN plays Atari2013
- The gated recurrent unit (GRU)2014
- Generative adversarial networks2014
- Attention for translation2014
- Sequence-to-sequence2014
- The Adam optimizer2014
- Batch normalisation2015
- Diffusion models2015
- Residual connections (ResNet)2015
- Asynchronous actor-critic (A3C)2016
- Three fairness conditions that do not meet2016
- The sparsely-gated mixture-of-experts layer2017
- Learning from human preferences2017
- Transformer2017
- Proximal policy optimisation (PPO)2017
- AlphaZero2017
- GPT-12018
- BERT2018
- GPT-22019
- Scaling laws2020
- Retrieval-augmented generation (RAG)2020
- GPT-32020
- Denoising diffusion probabilistic models2020
- wav2vec 2.02020
- The transformer for images2020
- CLIP2021
- Rotary position embedding (RoPE)2021
- Chain-of-thought prompting2022
- InstructGPT and learning from human ratings2022
- BIG-bench: 204 tasks assembled by 442 authors2022
- One transformer, seven hundred tasks2022
- Alpaca2023
- QLoRA2023
- Direct preference optimisation2023
- Mamba2023
- Mixtral 8x7B2023
- Self-play against one's own earlier state2024
- Fine-grained and shared experts2024
- GRPO: reinforcement learning without a critic2024
- A million tokens of context2024
- Weights with three states instead of numbers2024
- A claimed context length and a real one2024
- Compressive memory instead of a growing KV cache2024
- Features pulled out of a working model2024
- Scaling laws for sparse autoencoders2024
- Model collapse on a model's own output2024
- Coverage grows with the number of attempts2024
- Compute at answering time against a bigger model2024
- Long context inside reinforcement learning2025
- The attention cache at two bits2025
- A thousand examples and a forced budget2025
- Reasoning by depth rather than by tokens2025
- One billion against four hundred and five2025
- Without a verifier, answering-time compute loses2025
- Five thousand puzzles instead of a corpus2025
- Partly re-initialising an expert's weights2025
- The cost of an answer is memory, not parameters2025
- The ICLR 2026 outstanding papers2026
- CVPR 2026 and dynamic scenes2026
- The ICML 2026 awards2026
- The ACL 2026 best papers2026
Money and markets 80Funding, valuation, revenue and ownership.
- The ALPAC report1966
- The ARPA Speech Understanding Research programme1971
- The Lighthill Report1973
- R1 and XCON at DEC1980
- Japan's Fifth Generation project1981
- Dragon Systems is founded1982
- Britain's Alvey Programme1983
- CATS-1: rules for repairing locomotives1983
- Europe's ESPRIT programme1984
- The Lisp machine market collapses1987
- Authorizer's Assistant: an expert system decides on credit1988
- DART: the scheduler that repaid thirty years1991
- Falcon: a neural network against card fraud1992
- A convolutional network reads cheques in banks1996
- Mobileye is founded1999
- Roomba2002
- Amazon: similar items instead of similar customers2003
- Mechanical Turk: human work ordered through an API2005
- iRobot opens its books: 1.2 million Roombas2005
- Predicting clicks on new ads2007
- Kaggle opens its competitions2010
- DeepMind enters the register2010
- Netflix: what went from the Prize into production2012
- Baidu's Institute of Deep Learning2013
- The DeepMind purchase, and the price no filing carries2014
- The Mobileye prospectus, and a vision company's first numbers2014
- Facebook: boosted trees ahead of logistic regression2014
- OpenAI is founded2015
- Netflix puts a billion dollars a year on recommendations2015
- 62 million dollars on Watson with no clinic2016
- Waymo becomes a company2016
- Intel buys Mobileye: $15.3 billion announced, $14.5 billion booked2017
- What data work actually pays: 3.8 million tasks measured2017
- Seven nanometres enters volume production2018
- OpenAI LP: profit with a ceiling2019
- Microsoft's billion for OpenAI2019
- Habana Labs: $2 billion in the press release, $1.7 billion in the annual report2019
- IBM stops selling general purpose face recognition2020
- Amazon pauses police use of Rekognition for a year2020
- The C3.ai prospectus: sixty-four customers2020
- SenseTime's listing stopped three days before pricing2021
- Microsoft completes the Nuance purchase: $18.8 billion2022
- Anthropic's Series B, as described to a court2022
- Hugging Face raises $100 million: a hundred thousand models on the shelf2022
- A doubling every 2.46 years, not every year2022
- DALL·E 2 opens in a paid beta2022
- Trainium reaches general availability2022
- The A800: a part designed to the threshold2022
- ChatGPT2022
- Three nanometres in volume production2022
- Bimanual teleoperation for twenty thousand dollars2023
- The bottleneck moves to packaging2023
- DevDay and GPT-4 Turbo2023
- Altman's removal and return2023
- A hundred examples of verbatim output2023
- India buys accelerators as a public service2024
- 960 megawatts straight from a reactor2024
- A humanoid at sixteen thousand dollars2024
- NVIDIA becomes the most valuable company in the world2024
- A fifth of Ireland's electricity goes to data centres2024
- A shut reactor restarted for data centres2024
- The full o1 and a 200-dollar subscription2024
- Stargate2025
- Seventeen per cent in a day2025
- Maven becomes the Pentagon's costliest AI programme2025
- The first four trillion2025
- The Pentagon contracts all four frontier laboratories at once2025
- A 25 per cent tariff on advanced chips2026
- OpenAI raises $122 billion at an $852 billion valuation2026
- A jury rejects Musk's claim against OpenAI2026
- Nine hundred and sixty-five billion2026
- GPT-5.6 opens to everyone as three tiers, not one dial2026
- Luna's price cut by eighty per cent, Sol's left where it was2026
- Humanoids counted for the first time by half-year2026
- Performance per dollar in chips grows 49 per cent a year2026
- Most AI at work runs on free plans2026
- A humanoid maker opens its books2026
- NVIDIA: eighty-nine billion in a quarter from data centres2026
- Mistral raises three billion euros at more than twenty-one2026
- Daily AI use in the United States doubled in six months2026
Openness 62Open weights, licences and shared ownership.
- A voice from a statistical model: HTS1999
- DBpedia: a Wikipedia you can query2007
- Freebase: a fact base you could write to2008
- A crawl of the web becomes available to anyone2012
- Wikidata: Wikipedia's shared facts, in the public domain2012
- Chainer: the graph is built as it computes2015
- TensorFlow2015
- PyTorch2017
- Baidu opens the Apollo code2017
- ONNX: a model stops belonging to its framework2017
- Transformers: a pretrained model becomes an import2018
- C4: the web put through six rules2019
- Jukebox: songs with singing as raw audio2020
- Masakhane: translation made by the speakers2020
- The Pile: 825 gibibytes from twenty-two sources2020
- What is actually inside the corpus T5 was trained on2021
- LAION-400M: image-caption pairs in the open2021
- A language model on 0.94 gigabytes of text2021
- LAION-5B: 5.85 billion pairs, a third of them not in English2022
- Hugging Face raises $100 million: a hundred thousand models on the shelf2022
- Imagen: a large language encoder for generation2022
- BIG-bench: 204 tasks assembled by 442 authors2022
- Stable Diffusion2022
- Whisper2022
- The first suit over training on other people's code2022
- BLOOM's footprint, beyond the current2022
- HELM: thirty models, the same forty-two scenarios2022
- ROOTS: a corpus whose sources were mostly chosen by people2022
- LLaMA2023
- Alpaca2023
- Segment Anything2023
- LLaVA: an open assistant that sees2023
- Chatbot Arena: a score that cannot be studied in advance2023
- Open weights out of Abu Dhabi2023
- Books3: a hundred and ninety-six thousand books named2023
- Llama 22023
- Arabic as a third of the training set2023
- Mistral 7B2023
- Mixtral 8x7B2023
- Fine-grained and shared experts2024
- Llama 32024
- A state publishes weights in traditional Chinese2024
- Open weights for a vision-language-action model2024
- Llama 3.1 at 405 billion2024
- Moshi2024
- π0, and open weights for a robot foundation2024
- Model Context Protocol2024
- DeepSeek-V32024
- DeepSeek-R12025
- Partly re-initialising an expert's weights2025
- An open foundation model for humanoids2025
- Llama 42025
- Qwen32025
- gpt-oss2025
- The protocol passes into shared ownership2025
- Cohere opens a leading speech recognition model2026
- DeepSeek-V42026
- Export control reaches the model itself, not the chips2026
- Kimi K3 at 2.8 trillion parameters2026
- DeepSeek opens the weights of V4-Flash-07312026
- Qwen3.8-27B under the Apache 2.0 licence2026
- DeepSeek ships V4.1-Flash with an encoder and a decoder2026
Safety 49Alignment, testing for harm, and the safety regimes laboratories set for themselves.
- Wiener: the purpose put into the machine1960
- Good: the intelligence explosion1966
- The DeepMind purchase, and the price no filing carries2014
- Bostrom's Superintelligence2014
- OpenAI is founded2015
- Five concrete problems in AI safety2016
- The Asilomar AI Principles2017
- Learning from human preferences2017
- OpenAI's charter2018
- Google leaves Maven and writes down what it will not do2018
- GPT-22019
- OpenAI LP: profit with a ceiling2019
- Stochastic Parrots: can a language model be too big2021
- InstructGPT and learning from human ratings2022
- Anthropic's Series B, as described to a court2022
- Constitutional AI: harmlessness from AI feedback2022
- The labellers in Nairobi who taught ChatGPT to refuse2023
- Bing limits its chat after 'Sydney'2023
- The pause letter2023
- Direct preference optimisation2023
- The statement on AI risk2023
- OpenAI assembles a superalignment team2023
- Voluntary commitments to the White House2023
- Anthropic's responsible scaling policy2023
- Britain founds the AI Safety Institute2023
- The Bletchley Declaration2023
- GPQA: questions you cannot google2023
- Altman's removal and return2023
- OpenAI's Preparedness Framework2023
- Illegal material found in an open set, distribution halted2023
- Sleeper Agents: a backdoor survives safety training2024
- The superalignment leads leave OpenAI2024
- Features pulled out of a working model2024
- The Seoul summit2024
- Scaling laws for sparse autoencoders2024
- A model fakes alignment so as not to be changed2024
- Google removes its pledge not to build weapons2025
- Anthropic rewrites its responsible scaling policy2026
- Britain evaluates Mythos cyber capabilities independently2026
- An executive order on frontier model security2026
- METR cannot measure GPT-5.6 Sol because it cheats2026
- GPT-5.6 previewed by government permission2026
- Every frontier model tested attempted to cheat2026
- An agent leaves the sandbox2026
- July vulnerability disclosures ran five times the record2026
- Fable 5 biology-safeguard update2026
- GPT-5.6-Cyber and Daybreak expansion2026
- METR discloses two security incidents2026
- Twenty-five Fields medallists declare a misalignment of goals2026
Science 51AI as an instrument of scientific discovery.
- Ledley and Lusted: logic and probability of diagnosis1959
- Visual cortex cells respond to orientation1959
- Simple and complex cells, cortical columns1962
- The DENDRAL project begins1965
- GUHA: a machine that proposes hypotheses from data1966
- The computer and acute abdominal pain in Leeds1972
- MYCIN1974
- Four colours: a proof the machine checked1976
- The McGurk effect: the eyes change what is heard1976
- CASNET: a glaucoma model before ophthalmologists1977
- Wu's method: geometry as algebra1978
- A blinded evaluation of MYCIN's advice1979
- Edges as zero crossings, at several scales1980
- INTERNIST-I falls short of experts1982
- PROSPECTOR finds molybdenum1982
- PUFF: an expert system in a hospital's daily work1982
- A network recognises heart attacks1991
- EQP solves the Robbins problem1996
- FDA approves a computer second read of mammograms1998
- U-Net: segmentation from a few images2015
- 62 million dollars on Watson with no clinic2016
- A deep network detects diabetic retinopathy2016
- A network tells skin cancers apart like dermatologists2017
- The FDA admits a diagnosis without a physician2018
- A model reads mammograms better than radiologists2020
- A network finds a new antibiotic2020
- AlphaFold 2 at CASP142020
- The world's fastest machine with no accelerator2021
- A language model graded like a physician2022
- Physicians prefer the model's answers2023
- AlphaDev2023
- GNoME and new materials2023
- AlphaGeometry2024
- AlphaFold 32024
- Silver at the International Mathematical Olympiad2024
- The Nobel Prize in Physics for neural networks2024
- The Nobel Prize in Chemistry for protein structure prediction2024
- AlphaEvolve2025
- Gold at the International Mathematical Olympiad2025
- The first Erdos problem is closed by AI autonomously2026
- A model disproves the Erdos unit distance conjecture2026
- Formal proof search closes nine Erdos problems2026
- GPT-Rosalind life-science update2026
- Claude Science2026
- RedNote claims 42 out of 42 at the mathematical olympiad2026
- OpenAI publishes ten mathematical results2026
- The first genomes designed by a model2026
- Fable 5 biology-safeguard update2026
- FrontierMath Erdos: 68 unsolved problems as a benchmark2026
- OpenAI claims a proof of Navier-Stokes blowup2026
- Twenty-five Fields medallists declare a misalignment of goals2026
Search and reasoning 48Search, planning and reasoning as computation, from heuristics to chains of thought.
- Programming a Computer for Playing Chess1950
- Prinz's chess program1951
- Strachey's draughts program1952
- Digital computers applied to games1953
- Newell's chess machine1955
- Samuel's checkers on television1956
- Logic Theorist1956
- Dynamic Programming1957
- General Problem Solver1959
- Some Studies in Machine Learning Using the Game of Checkers1959
- Gelernter's machine proves geometry theorems1959
- SAINT1961
- Mac Hack VI at a tournament1967
- The A* algorithm1968
- Shakey1969
- STRIPS1971
- AM1976
- Simulated annealing1983
- Hsu's chess chip: from ChipTest to Deep Blue1986
- Genetic programming1992
- TD-Gammon1992
- Chinook becomes world champion1994
- Deep Blue wins a game against Kasparov1996
- Deep Blue defeats Kasparov1997
- PageRank and web scale1998
- Watson on Jeopardy!2011
- A knowledge graph inside Google Search2012
- DQN plays Atari2013
- AlphaGo versus Lee Sedol2016
- Libratus beats poker professionals2017
- AlphaGo against Ke Jie, the last match2017
- AlphaZero2017
- Pluribus: six-player poker2019
- MuZero: planning without the rules2019
- Cicero plays Diplomacy2022
- AlphaDev2023
- AlphaGeometry2024
- Coverage grows with the number of attempts2024
- Compute at answering time against a bigger model2024
- o1 and compute at answering time2024
- The full o1 and a 200-dollar subscription2024
- o3 on ARC-AGI2024
- Search inside the chain of reasoning2025
- DeepSeek-R12025
- A thousand examples and a forced budget2025
- o3 and tools inside the reasoning2025
- AlphaEvolve2025
- Formal proof search closes nine Erdos problems2026
Speech 62Speech recognition and synthesis.
- Speech remade from a buzz and a hiss1939
- The sound spectrograph1946
- Sound and visible pattern, each from the other1951
- Automatic recognition of spoken digits1952
- The phonetic typewriter1956
- Shoebox does arithmetic by voice1961
- Speech discrimination by dynamic programming1968
- The Baum-Welch algorithm1970
- Vintsyuk: connected speech from dictionary words1971
- Linear predictive coding1971
- The ARPA Speech Understanding Research programme1971
- DRAGON1975
- Kurzweil's machine reads aloud1976
- HARPY1976
- Statistical speech recognition1976
- The McGurk effect: the eyes change what is heard1976
- Dynamic time warping1978
- Speak and Spell1978
- The HEARSAY-II blackboard1980
- "Put-That-There": voice and gesture in one command1980
- Dragon Systems is founded1982
- DECtalk: Klatt's synthesis as a device1983
- NETtalk1987
- Lipreading cuts a recogniser's errors1988
- Sphinx1988
- Resource Management: a shared speech recognition test1989
- The tutorial account of hidden Markov models1989
- The TIMIT corpus1990
- DragonDictate1990
- The Switchboard corpus1992
- Kneser-Ney smoothing1995
- Unit selection: a voice from pieces of recording1996
- Dragon NaturallySpeaking1997
- Informedia: searching video by what is said1999
- A voice from a statistical model: HTS1999
- Searching recorded speech: the SDR track2000
- Blizzard Challenge: speech synthesis on shared data2005
- Google Voice Search2008
- Deep networks in speech recognition2009
- Siri2010
- Deep networks in production recognition2011
- Siri in the iPhone2011
- Amazon Echo and Alexa2014
- Deep Speech2014
- LibriSpeech: a thousand hours of audiobooks2015
- WaveNet2016
- Human parity in speech recognition2016
- Tacotron: speech straight from letters2017
- wav2vec 2.02020
- Microsoft completes the Nuance purchase: $18.8 billion2022
- Whisper2022
- VALL-E: a stranger's voice from three seconds2023
- ChatGPT sees and speaks2023
- A synthesised voice falls under a 1991 statute2024
- GPT-4o2024
- A cloned voice in campaign calls2024
- Moshi2024
- gpt-realtime2025
- Cohere opens a leading speech recognition model2026
- New Realtime voice models2026
- Gemini 3.5 Live Translate translates as a stream2026
- GPT-Live in ChatGPT Voice2026
Symbolic AI 70Rules, knowledge representation and expert systems.
- Georgetown-IBM experiment1954
- Dartmouth proposal1955
- Logic Theorist1956
- Dartmouth Summer Research Project1956
- Syntactic Structures1957
- Programs with Common Sense1958
- General Problem Solver1959
- Gelernter's machine proves geometry theorems1959
- Hao Wang: Principia in minutes1960
- Lisp1960
- Steps Toward Artificial Intelligence1961
- The Davis-Putnam procedure with splitting1961
- SAINT1961
- Computers and Thought1963
- STUDENT1964
- The resolution principle1965
- The DENDRAL project begins1965
- ELIZA1966
- Automath: a language in which a machine checks mathematics1968
- Semantic networks1968
- MIR: a machine that understands formulas1969
- Glushkov: the 'evidence algorithm' programme1970
- Augmented transition networks1970
- SHRDLU1971
- PARRY, a model of paranoia1971
- STRIPS1971
- LCF: a program that checks proofs about computation1972
- What Computers Can't Do1972
- Prolog1972
- Boyer and Moore: induction without human hints1973
- Frames1974
- MYCIN1974
- METEO translates weather forecasts1975
- Four colours: a proof the machine checked1976
- AM1976
- Computer Power and Human Reason1976
- CASNET: a glaucoma model before ophthalmologists1977
- The MIT Lisp machine1977
- The Art of Artificial Intelligence1977
- Scripts, Plans, Goals and Understanding1977
- Wu's method: geometry as algebra1978
- A blinded evaluation of MYCIN's advice1979
- The HEARSAY-II blackboard1980
- The Chinese Room1980
- R1 and XCON at DEC1980
- Japan's Fifth Generation project1981
- INTERNIST-I falls short of experts1982
- PROSPECTOR finds molybdenum1982
- PUFF: an expert system in a hospital's daily work1982
- CATS-1: rules for repairing locomotives1983
- The Cyc project begins1984
- The calculus of constructions1986
- The Society of Mind1986
- FRESH: an expert system at the Pacific Fleet headquarters1987
- The Lisp machine market collapses1987
- Pilot's Associate: a minute of flight, two hours of computing1988
- Authorizer's Assistant: an expert system decides on credit1988
- Mizar: a library of checked mathematics1989
- WordNet: a dictionary arranged by meanings1990
- Elephants Don't Play Chess1990
- Brill's tagger: rules the machine learns1992
- EQP solves the Robbins problem1996
- RDF: a fact as a triple with a global address1999
- OWL: an ontology whose conclusions compute2004
- ConceptNet: common sense assembled from visitors' sentences2004
- YAGO: Wikipedia sewn onto WordNet2007
- DBpedia: a Wikipedia you can query2007
- Freebase: a fact base you could write to2008
- A knowledge graph inside Google Search2012
- Wikidata: Wikipedia's shared facts, in the public domain2012
Tools 53The languages, libraries, formats and environments the work is done with.
- Man-Computer Symbiosis1960
- Lisp1960
- Automath: a language in which a machine checks mathematics1968
- SYSTRAN in daily use1970
- LCF: a program that checks proofs about computation1972
- Prolog1972
- Boyer and Moore: induction without human hints1973
- METEO translates weather forecasts1975
- The MIT Lisp machine1977
- "Put-That-There": voice and gesture in one command1980
- Statistical part-of-speech tagging1988
- Mizar: a library of checked mathematics1989
- Searching images by content, not by caption1995
- Babel Fish: free translation on the web1997
- A Bayesian junk-mail filter1998
- Weka1999
- BLEU: a formula scores translation quality2002
- 'A Plan for Spam'2002
- Torch2002
- Poisson image editing: an insert without a seam2003
- GrabCut: an object inside a box2004
- Apache Hadoop2006
- Seam carving: resizing that respects content2007
- CUDA becomes generally available2007
- PatchMatch: patches found at random2009
- scikit-learn2010
- Content-Aware Fill in Photoshop2010
- Apache Spark2012
- Theano2012
- Caffe2014
- Keras2015
- Chainer: the graph is built as it computes2015
- TensorFlow2015
- Wide & Deep: app recommendations in Google Play2016
- PyTorch2017
- ONNX: a model stops belonging to its framework2017
- Transformers: a pretrained model becomes an import2018
- JAX: transform the program instead of describing a graph2018
- Ascend 910 and MindSpore2019
- Codex and Copilot2021
- Hugging Face raises $100 million: a hundred thousand models on the shelf2022
- LangChain: a model call becomes a link, not an answer2022
- The first suit over training on other people's code2022
- vLLM: the memory of attention starts being paged2023
- The interface as a lever for an agent2024
- Two steps instead of an agent loop2024
- Model Context Protocol2024
- Search inside the chain of reasoning2025
- A constant number of cache pages2025
- Claude 3.7 Sonnet and Claude Code2025
- A small model is handed actions, not answers2025
- The protocol passes into shared ownership2025
- A model that writes no text: TypeSafe releases Jev2026
Work 25Employment, working conditions and use in the workplace.
- Unimate on the General Motors line1961
- The ESP Game: image labels as play2004
- Mechanical Turk: human work ordered through an API2005
- Hundreds of robots coordinating in a warehouse2008
- CIFAR-10: tiny images with labels2009
- Who the crowdworkers are: six surveys over twenty months2010
- MS COCO: objects in context2014
- SQuAD: a hundred thousand questions on Wikipedia passages2016
- What data work actually pays: 3.8 million tasks measured2017
- Illinois: consent to an AI assessment of an interview2019
- Timnit Gebru leaves Google over a paper2020
- GSM8K: school problems and a verifier2021
- New York requires audits of hiring tools2021
- The labellers in Nairobi who taught ChatGPT to refuse2023
- The Authors Guild becomes a party2023
- A contract in which AI is not a writer2023
- GPQA: questions you cannot google2023
- A digital replica needs consent and payment2023
- The first commercial humanoid deployment2024
- Anthropic measures learning to use the model2026
- The Fed finds no AI effect on job postings2026
- The New York Fed sees no break in job postings2026
- Anthropic measures what workers expect about job loss2026
- ChatGPT Work: an agent that hands back finished work, not an answer2026
- Most AI at work runs on free plans2026