[
 {
  "id": "mcculloch-pitts-1943",
  "date": "1943-12",
  "title": "McCulloch and Pitts model the first artificial neuron",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "math",
  "concepts": [
   "artificial-neuron"
  ],
  "people": [],
  "one_liner": "A neuron described as simple logic: inputs add up, and it fires past a threshold.",
  "summary": "Warren McCulloch and Walter Pitts showed that networks of very simple on/off units could compute logical functions. It is the first mathematical model of a neuron and the starting point of neural networks.",
  "sources": [
   {
    "title": "A logical calculus of the ideas immanent in nervous activity",
    "url": "https://doi.org/10.1007/BF02478259",
    "kind": "primary"
   }
  ],
  "era": "foundations",
  "precision": "month",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "Crossref: A logical calculus of the ideas immanent in nervous activity [[1943, 12], [1943, 12]]"
   ]
  }
 },
 {
  "id": "turing-1950",
  "date": "1950-10",
  "title": "Turing asks: can machines think?",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "safety",
  "concepts": [
   "turing-test"
  ],
  "people": [],
  "one_liner": "Alan Turing replaces 'can machines think' with a test: can a machine pass as human in conversation?",
  "summary": "In 'Computing Machinery and Intelligence', Turing proposed the imitation game, now called the Turing test, and answered objections to machine intelligence that are still argued today.",
  "sources": [
   {
    "title": "Computing Machinery and Intelligence",
    "url": "https://doi.org/10.1093/mind/LIX.236.433",
    "kind": "primary"
   }
  ],
  "era": "foundations",
  "precision": "month",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "Crossref: I.—COMPUTING MACHINERY AND INTELLIGENCE [[1950, 10, 1], [1950, 10, 1], [1950, 10, 1]]"
   ]
  }
 },
 {
  "id": "dartmouth-1955",
  "date": "1955-08-31",
  "title": "The Dartmouth proposal names the field 'artificial intelligence'",
  "type": "event",
  "labs": [
   "academia"
  ],
  "layer": "learning",
  "concepts": [
   "symbolic-ai"
  ],
  "people": [],
  "one_liner": "McCarthy, Minsky, Rochester and Shannon propose a summer study and coin the term 'artificial intelligence'.",
  "summary": "The proposal for the 1956 Dartmouth summer workshop argued that every aspect of learning could in principle be described precisely enough for a machine to simulate it. The workshop is usually treated as the founding of AI.",
  "sources": [
   {
    "title": "A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence",
    "url": "http://jmc.stanford.edu/articles/dartmouth/dartmouth.pdf",
    "kind": "primary"
   }
  ],
  "era": "foundations",
  "precision": "day",
  "check": {
   "status": "link-ok",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): "
   ]
  }
 },
 {
  "id": "perceptron-1958",
  "date": "1958",
  "title": "Rosenblatt's perceptron learns from examples",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "math",
  "concepts": [
   "perceptron",
   "supervised-learning"
  ],
  "people": [],
  "one_liner": "A machine that adjusts its own weights from labelled examples: the first learning neural network.",
  "summary": "Frank Rosenblatt described the perceptron, a single-layer network with a learning rule that updates weights after each mistake. It could learn simple visual patterns and drew huge public attention.",
  "sources": [
   {
    "title": "The perceptron: A probabilistic model for information storage and organization in the brain",
    "url": "https://doi.org/10.1037/h0042519",
    "kind": "primary"
   }
  ],
  "era": "foundations",
  "precision": "year",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "Crossref: The perceptron: A probabilistic model for information storage and organization in the brain. [[1958], [1958]]"
   ]
  }
 },
 {
  "id": "samuel-checkers-1959",
  "date": "1959-07",
  "title": "Arthur Samuel's checkers program learns to play",
  "type": "paper",
  "labs": [
   "ibm"
  ],
  "layer": "learning",
  "concepts": [
   "machine-learning",
   "reinforcement-learning"
  ],
  "people": [],
  "one_liner": "A program improves at checkers by playing itself, and the phrase 'machine learning' enters use.",
  "summary": "Arthur Samuel at IBM built a checkers player that improved through self-play and stored experience. His paper popularised the term machine learning: getting computers to learn without being explicitly programmed.",
  "sources": [
   {
    "title": "Some Studies in Machine Learning Using the Game of Checkers",
    "url": "https://doi.org/10.1147/rd.33.0210",
    "kind": "primary"
   }
  ],
  "era": "foundations",
  "precision": "month",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "Crossref: Some Studies in Machine Learning Using the Game of Checkers [[1959, 7], [1959, 7]]"
   ]
  }
 },
 {
  "id": "eliza-1966",
  "date": "1966-01",
  "title": "ELIZA, the first chatbot",
  "type": "product",
  "labs": [
   "academia"
  ],
  "layer": "building",
  "concepts": [
   "chatbot"
  ],
  "people": [],
  "one_liner": "A simple pattern-matching program at MIT convinces some users it understands them.",
  "summary": "Joseph Weizenbaum's ELIZA rephrased user input using keyword rules, most famously as a psychotherapist. People attributed understanding to it, an effect now called the ELIZA effect.",
  "sources": [
   {
    "title": "ELIZA: a computer program for the study of natural language communication between man and machine",
    "url": "https://doi.org/10.1145/365153.365168",
    "kind": "primary"
   }
  ],
  "era": "foundations",
  "precision": "month",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "Crossref: ELIZA—A Computer Program For the Study of Natural Language Communication Between Man and Machine [[1966, 1, 1], [1966, 1, 1], [1966, 1, 1]]"
   ]
  }
 },
 {
  "id": "lighthill-1973",
  "date": "1973",
  "title": "The Lighthill report triggers the first AI winter",
  "type": "event",
  "labs": [
   "government"
  ],
  "layer": "safety",
  "concepts": [
   "ai-winter"
  ],
  "people": [],
  "one_liner": "A UK government review concludes AI has failed its promises, and funding is cut.",
  "summary": "James Lighthill's report for the UK Science Research Council criticised the lack of real-world results from AI research. Funding cuts in the UK and later the US followed, a period known as the first AI winter.",
  "sources": [
   {
    "title": "Artificial Intelligence: A General Survey (Lighthill report)",
    "url": "http://www.chilton-computing.org.uk/inf/literature/reports/lighthill_report/p001.htm",
    "kind": "primary"
   }
  ],
  "era": "winters",
  "precision": "year",
  "check": {
   "status": "link-ok",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Chilton::INF::Lighthill Report"
   ]
  }
 },
 {
  "id": "hopfield-1982",
  "date": "1982-04",
  "title": "Hopfield networks store memories",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "architectures",
  "concepts": [
   "hopfield-network"
  ],
  "people": [],
  "one_liner": "A network that settles into stored patterns, recalling a whole memory from a partial cue.",
  "summary": "John Hopfield showed how a recurrent network with symmetric connections can act as associative memory, linking neural networks to physics. This work was part of his 2024 Nobel Prize in Physics.",
  "sources": [
   {
    "title": "Neural networks and physical systems with emergent collective computational abilities",
    "url": "https://doi.org/10.1073/pnas.79.8.2554",
    "kind": "primary"
   }
  ],
  "era": "winters",
  "precision": "month",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "Crossref: Neural networks and physical systems with emergent collective computational abilities. [[1982, 4], [1982, 4], [1982, 4]]"
   ]
  }
 },
 {
  "id": "backprop-1986",
  "date": "1986-10-09",
  "title": "Backpropagation popularised for training multi-layer networks",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "math",
  "concepts": [
   "backpropagation",
   "gradient-descent"
  ],
  "people": [
   "geoffrey-hinton"
  ],
  "one_liner": "A way to share the blame for an error across every layer, so deep networks can learn.",
  "summary": "Rumelhart, Hinton and Williams showed that backpropagating errors lets networks with hidden layers learn useful internal representations. Nearly every neural network trained today still uses this method.",
  "sources": [
   {
    "title": "Learning representations by back-propagating errors",
    "url": "https://doi.org/10.1038/323533a0",
    "kind": "primary"
   }
  ],
  "era": "winters",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "Crossref: Learning representations by back-propagating errors [[1986, 10], [1986, 10]]"
   ]
  }
 },
 {
  "id": "lecun-cnn-1989",
  "date": "1989-12",
  "title": "LeCun applies convolutional networks to handwritten digits",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "architectures",
  "concepts": [
   "cnn",
   "backpropagation"
  ],
  "people": [
   "yann-lecun"
  ],
  "one_liner": "A network with shared filters reads handwritten zip codes for the US postal service.",
  "summary": "Yann LeCun and colleagues at Bell Labs trained a convolutional network with backpropagation to recognise handwritten digits. It is the direct ancestor of the networks that later won ImageNet.",
  "sources": [
   {
    "title": "Backpropagation Applied to Handwritten Zip Code Recognition",
    "url": "https://doi.org/10.1162/neco.1989.1.4.541",
    "kind": "primary"
   }
  ],
  "era": "winters",
  "precision": "month",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "Crossref: Backpropagation Applied to Handwritten Zip Code Recognition [[1989, 12], [1989, 12]]"
   ]
  }
 },
 {
  "id": "q-learning-1992",
  "date": "1992-05",
  "title": "Q-learning gets its convergence proof",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "learning",
  "concepts": [
   "q-learning",
   "reinforcement-learning"
  ],
  "people": [],
  "one_liner": "An agent learns the value of each action by trial and error, with no model of the world.",
  "summary": "Chris Watkins introduced Q-learning in his 1989 thesis, and the 1992 paper with Peter Dayan proved it converges. Deep Q-networks used it to play Atari games in 2013.",
  "sources": [
   {
    "title": "Q-learning",
    "url": "https://doi.org/10.1007/BF00992698",
    "kind": "primary"
   }
  ],
  "era": "winters",
  "precision": "month",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "Crossref: Q-learning [[1992, 5], [1992, 5]]"
   ]
  }
 },
 {
  "id": "td-gammon-1995",
  "date": "1995-03",
  "title": "TD-Gammon reaches expert backgammon by self-play",
  "type": "paper",
  "labs": [
   "ibm"
  ],
  "layer": "learning",
  "concepts": [
   "reinforcement-learning"
  ],
  "people": [],
  "one_liner": "A neural network teaches itself backgammon by playing against itself.",
  "summary": "Gerald Tesauro's TD-Gammon combined temporal-difference learning with a neural network and reached the level of top human players, an early success for self-play reinforcement learning.",
  "sources": [
   {
    "title": "Temporal difference learning and TD-Gammon",
    "url": "https://doi.org/10.1145/203330.203343",
    "kind": "primary"
   }
  ],
  "era": "statistical",
  "precision": "month",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "Crossref: Temporal difference learning and TD-Gammon [[1995, 3], [1995, 3], [1995, 3]]"
   ]
  }
 },
 {
  "id": "svm-1995",
  "date": "1995-09",
  "title": "Support vector machines",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "learning",
  "concepts": [
   "supervised-learning"
  ],
  "people": [],
  "one_liner": "A classifier that draws the widest possible margin between classes.",
  "summary": "Corinna Cortes and Vladimir Vapnik introduced support-vector networks, which dominated many classification tasks for the next 15 years thanks to strong theory and good results on small data.",
  "sources": [
   {
    "title": "Support-vector networks",
    "url": "https://doi.org/10.1007/BF00994018",
    "kind": "primary"
   }
  ],
  "era": "statistical",
  "precision": "month",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "Crossref: Support-vector networks [[1995, 9], [1995, 9]]"
   ]
  }
 },
 {
  "id": "deep-blue-1997",
  "date": "1997-05-11",
  "title": "Deep Blue beats world chess champion Garry Kasparov",
  "type": "event",
  "labs": [
   "ibm"
  ],
  "layer": "infra",
  "concepts": [
   "symbolic-ai"
  ],
  "people": [],
  "one_liner": "IBM's chess computer wins a six-game match against the reigning world champion.",
  "summary": "Deep Blue relied on special chess hardware and brute-force search rather than learning. Its win showed that machines could beat the best humans at a game long seen as a test of intelligence.",
  "sources": [
   {
    "title": "Deep Blue (IBM history)",
    "url": "https://www.ibm.com/history/deep-blue",
    "kind": "primary"
   }
  ],
  "era": "statistical",
  "precision": "day",
  "check": {
   "status": "link-ok",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Deep Blue | IBM"
   ]
  }
 },
 {
  "id": "lstm-1997",
  "date": "1997-11",
  "title": "Long short-term memory (LSTM)",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "architectures",
  "concepts": [
   "lstm",
   "rnn"
  ],
  "people": [],
  "one_liner": "A recurrent network with gates that decide what to remember and what to forget.",
  "summary": "Sepp Hochreiter and Jurgen Schmidhuber solved the vanishing gradient problem for recurrent networks with gated memory cells. LSTMs powered speech recognition and translation until transformers arrived.",
  "sources": [
   {
    "title": "Long Short-Term Memory",
    "url": "https://doi.org/10.1162/neco.1997.9.8.1735",
    "kind": "primary"
   }
  ],
  "era": "statistical",
  "precision": "month",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "Crossref: Long Short-Term Memory [[1997, 11, 1], [1997, 11, 1]]"
   ]
  }
 },
 {
  "id": "lenet-1998",
  "date": "1998-11",
  "title": "LeNet-5 and gradient-based document recognition",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "architectures",
  "concepts": [
   "cnn"
  ],
  "people": [
   "yann-lecun"
  ],
  "one_liner": "The classic convolutional network design, used to read cheques at scale.",
  "summary": "LeCun, Bottou, Bengio and Haffner described LeNet-5 and an end-to-end system for reading documents. The MNIST dataset from this work became the 'hello world' of machine learning.",
  "sources": [
   {
    "title": "Gradient-based learning applied to document recognition",
    "url": "https://doi.org/10.1109/5.726791",
    "kind": "primary"
   }
  ],
  "era": "statistical",
  "precision": "month",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "Crossref: Gradient-based learning applied to document recognition [[1998], [1998]]"
   ]
  }
 },
 {
  "id": "random-forests-2001",
  "date": "2001-10",
  "title": "Random forests",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "learning",
  "concepts": [
   "supervised-learning"
  ],
  "people": [],
  "one_liner": "Many decision trees vote, and the crowd beats any single tree.",
  "summary": "Leo Breiman's random forests combined bagging with random feature selection. They remain a strong default for tabular data, the kind analysts work with every day.",
  "sources": [
   {
    "title": "Random Forests",
    "url": "https://doi.org/10.1023/A:1010933404324",
    "kind": "primary"
   }
  ],
  "era": "statistical",
  "precision": "month",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "Crossref: Random Forests [[2001, 10], [2001, 10]]"
   ]
  }
 },
 {
  "id": "neural-lm-2003",
  "date": "2003",
  "title": "A neural probabilistic language model",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "architectures",
  "concepts": [
   "word-embedding",
   "language-model"
  ],
  "people": [],
  "one_liner": "A neural network predicts the next word and learns word vectors along the way.",
  "summary": "Yoshua Bengio and colleagues trained a network to predict the next word, learning a vector for each word at the same time. This idea of learned embeddings runs through every modern language model.",
  "sources": [
   {
    "title": "A Neural Probabilistic Language Model (JMLR)",
    "url": "https://www.jmlr.org/papers/v3/bengio03a.html",
    "kind": "primary"
   }
  ],
  "era": "statistical",
  "precision": "year",
  "check": {
   "status": "link-ok",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): "
   ]
  }
 },
 {
  "id": "deep-belief-2006",
  "date": "2006-07",
  "title": "Deep belief nets revive deep learning",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "learning",
  "concepts": [
   "deep-learning",
   "unsupervised-learning"
  ],
  "people": [
   "geoffrey-hinton"
  ],
  "one_liner": "Training deep networks one layer at a time finally works.",
  "summary": "Hinton, Osindero and Teh showed a fast greedy way to pre-train deep networks layer by layer. It restarted interest in deep networks and gave 'deep learning' its name.",
  "sources": [
   {
    "title": "A Fast Learning Algorithm for Deep Belief Nets",
    "url": "https://doi.org/10.1162/neco.2006.18.7.1527",
    "kind": "primary"
   }
  ],
  "era": "statistical",
  "precision": "month",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "Crossref: A Fast Learning Algorithm for Deep Belief Nets [[2006, 7], [2006, 7]]"
   ]
  }
 },
 {
  "id": "gpu-training-2009",
  "date": "2009-06",
  "title": "GPUs make large neural networks practical",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "infra",
  "concepts": [
   "gpu"
  ],
  "people": [
   "andrew-ng"
  ],
  "one_liner": "Graphics cards train networks up to 70 times faster than CPUs.",
  "summary": "Raina, Madhavan and Ng showed that graphics processors could speed up unsupervised learning by large factors. GPU training became the norm and later made NVIDIA central to AI.",
  "sources": [
   {
    "title": "Large-scale deep unsupervised learning using graphics processors",
    "url": "https://doi.org/10.1145/1553374.1553486",
    "kind": "primary"
   }
  ],
  "era": "statistical",
  "precision": "month",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "Crossref: Large-scale deep unsupervised learning using graphics processors [[2009, 6, 14], [2009, 6, 14], [2009, 6, 14]]"
   ]
  }
 },
 {
  "id": "imagenet-2009",
  "date": "2009-06",
  "title": "ImageNet: a dataset big enough for deep learning",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "data",
  "concepts": [
   "dataset",
   "benchmark"
  ],
  "people": [
   "fei-fei-li"
  ],
  "one_liner": "Millions of labelled images, organised by the WordNet hierarchy.",
  "summary": "Fei-Fei Li's team built ImageNet with crowd-sourced labels. Its yearly challenge became the benchmark where deep learning first proved itself in 2012.",
  "sources": [
   {
    "title": "ImageNet: A large-scale hierarchical image database",
    "url": "https://doi.org/10.1109/CVPR.2009.5206848",
    "kind": "primary"
   }
  ],
  "era": "statistical",
  "precision": "month",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "Crossref: ImageNet: A large-scale hierarchical image database [[2009, 6], [2009, 6]]"
   ]
  }
 },
 {
  "id": "google-brain-cat-2011",
  "date": "2011-12-29",
  "title": "Google Brain's network learns to spot cats from YouTube",
  "type": "paper",
  "labs": [
   "google"
  ],
  "layer": "learning",
  "concepts": [
   "unsupervised-learning",
   "deep-learning"
  ],
  "people": [
   "andrew-ng"
  ],
  "one_liner": "A huge network trained on unlabelled video frames learns a 'cat detector' on its own.",
  "summary": "Quoc Le, Andrew Ng, Jeff Dean and colleagues trained a billion-connection network on 16,000 CPU cores. Neurons emerged that responded to faces and cats, showing what scale plus unlabelled data could do.",
  "sources": [
   {
    "title": "Building high-level features using large scale unsupervised learning",
    "url": "https://arxiv.org/abs/1112.6209",
    "kind": "primary"
   }
  ],
  "era": "statistical",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Building high-level features using large scale unsupervised learning (v1 2011-12-29)"
   ]
  }
 },
 {
  "id": "alexnet-2012",
  "date": "2012-12",
  "title": "AlexNet wins ImageNet by a wide margin",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "architectures",
  "concepts": [
   "cnn",
   "deep-learning",
   "gpu"
  ],
  "people": [
   "ilya-sutskever",
   "geoffrey-hinton"
  ],
  "one_liner": "A deep convolutional network trained on two GPUs cuts the ImageNet error rate sharply.",
  "summary": "Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton won the 2012 ImageNet challenge with a deep CNN using ReLU, dropout and GPU training. The result convinced the field to switch to deep learning.",
  "sources": [
   {
    "title": "ImageNet Classification with Deep Convolutional Neural Networks (NeurIPS 2012)",
    "url": "https://papers.nips.cc/paper_files/paper/2012/hash/c399862d3b9d6b76c8436e924a68c45b-Abstract.html",
    "kind": "primary"
   }
  ],
  "era": "deep-learning",
  "precision": "month",
  "check": {
   "status": "link-ok",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): ImageNet Classification with Deep Convolutional Neural Networks"
   ]
  }
 },
 {
  "id": "word2vec-2013",
  "date": "2013-01-16",
  "title": "word2vec: words as vectors you can do maths with",
  "type": "paper",
  "labs": [
   "google"
  ],
  "layer": "architectures",
  "concepts": [
   "word-embedding"
  ],
  "people": [],
  "one_liner": "Cheap training gives word vectors where king minus man plus woman lands near queen.",
  "summary": "Tomas Mikolov and colleagues at Google introduced fast methods to learn word embeddings from huge text corpora. Embeddings became a standard building block, and later the core of semantic search and RAG.",
  "sources": [
   {
    "title": "Efficient Estimation of Word Representations in Vector Space",
    "url": "https://arxiv.org/abs/1301.3781",
    "kind": "primary"
   }
  ],
  "era": "deep-learning",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Efficient Estimation of Word Representations in Vector Space (v1 2013-01-16)"
   ]
  }
 },
 {
  "id": "dqn-2013",
  "date": "2013-12-19",
  "title": "Deep Q-networks learn Atari games from pixels",
  "type": "paper",
  "labs": [
   "google"
  ],
  "layer": "learning",
  "concepts": [
   "reinforcement-learning",
   "q-learning",
   "deep-learning"
  ],
  "people": [
   "demis-hassabis"
  ],
  "one_liner": "One network learns to play several Atari games from the screen and the score alone.",
  "summary": "DeepMind combined Q-learning with a convolutional network. The same agent learned multiple games from raw pixels, which led Google to acquire DeepMind in 2014.",
  "sources": [
   {
    "title": "Playing Atari with Deep Reinforcement Learning",
    "url": "https://arxiv.org/abs/1312.5602",
    "kind": "primary"
   }
  ],
  "era": "deep-learning",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Playing Atari with Deep Reinforcement Learning (v1 2013-12-19)"
   ]
  }
 },
 {
  "id": "vae-2013",
  "date": "2013-12-20",
  "title": "Variational autoencoders",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "architectures",
  "concepts": [
   "vae",
   "generative-model"
  ],
  "people": [],
  "one_liner": "A network that learns a smooth space of possibilities and can sample new data from it.",
  "summary": "Kingma and Welling introduced auto-encoding variational Bayes. VAEs became a core generative model and later the image compressor inside latent diffusion models such as Stable Diffusion.",
  "sources": [
   {
    "title": "Auto-Encoding Variational Bayes",
    "url": "https://arxiv.org/abs/1312.6114",
    "kind": "primary"
   }
  ],
  "era": "deep-learning",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Auto-Encoding Variational Bayes (v1 2013-12-20)"
   ]
  }
 },
 {
  "id": "dropout-2014",
  "date": "2014-06",
  "title": "Dropout prevents overfitting",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "training",
  "concepts": [
   "dropout"
  ],
  "people": [
   "geoffrey-hinton",
   "ilya-sutskever"
  ],
  "one_liner": "Randomly switching off neurons during training makes networks generalise better.",
  "summary": "Srivastava, Hinton, Krizhevsky, Sutskever and Salakhutdinov published the full dropout paper. It became one of the most used regularisation tricks in deep learning.",
  "sources": [
   {
    "title": "Dropout: A Simple Way to Prevent Neural Networks from Overfitting (JMLR)",
    "url": "https://jmlr.org/papers/v15/srivastava14a.html",
    "kind": "primary"
   }
  ],
  "era": "deep-learning",
  "precision": "month",
  "check": {
   "status": "link-ok",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Dropout: A Simple Way to Prevent Neural Networks from Overfitting"
   ]
  }
 },
 {
  "id": "gan-2014",
  "date": "2014-06-10",
  "title": "Generative adversarial networks",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "architectures",
  "concepts": [
   "gan",
   "generative-model"
  ],
  "people": [],
  "one_liner": "Two networks compete: one forges images, the other spots forgeries, and both get better.",
  "summary": "Ian Goodfellow and colleagues introduced GANs. They produced the first strikingly realistic synthetic images and led to deepfakes, before diffusion models took over image generation.",
  "sources": [
   {
    "title": "Generative Adversarial Networks",
    "url": "https://arxiv.org/abs/1406.2661",
    "kind": "primary"
   }
  ],
  "era": "deep-learning",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Generative Adversarial Networks (v1 2014-06-10)"
   ]
  }
 },
 {
  "id": "bahdanau-attention-2014",
  "date": "2014-09-01",
  "title": "Attention is introduced for translation",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "architectures",
  "concepts": [
   "attention",
   "rnn"
  ],
  "people": [],
  "one_liner": "Instead of squeezing a sentence into one vector, the model looks back at the words it needs.",
  "summary": "Bahdanau, Cho and Bengio let a translation model learn to align and attend to parts of the input sentence. Attention later became the whole architecture in the transformer.",
  "sources": [
   {
    "title": "Neural Machine Translation by Jointly Learning to Align and Translate",
    "url": "https://arxiv.org/abs/1409.0473",
    "kind": "primary"
   }
  ],
  "era": "deep-learning",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Neural Machine Translation by Jointly Learning to Align and Translate (v1 2014-09-01)"
   ]
  }
 },
 {
  "id": "seq2seq-2014",
  "date": "2014-09-10",
  "title": "Sequence to sequence learning",
  "type": "paper",
  "labs": [
   "google"
  ],
  "layer": "architectures",
  "concepts": [
   "lstm",
   "rnn"
  ],
  "people": [
   "ilya-sutskever"
  ],
  "one_liner": "One LSTM reads a sentence, another writes the translation.",
  "summary": "Sutskever, Vinyals and Le showed that an encoder-decoder LSTM could translate end to end. The encoder-decoder pattern carried straight into the original transformer.",
  "sources": [
   {
    "title": "Sequence to Sequence Learning with Neural Networks",
    "url": "https://arxiv.org/abs/1409.3215",
    "kind": "primary"
   }
  ],
  "era": "deep-learning",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Sequence to Sequence Learning with Neural Networks (v1 2014-09-10)"
   ]
  }
 },
 {
  "id": "adam-2014",
  "date": "2014-12-22",
  "title": "The Adam optimiser",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "math",
  "concepts": [
   "gradient-descent"
  ],
  "people": [],
  "one_liner": "A gradient descent method that adapts its step size per parameter.",
  "summary": "Kingma and Ba introduced Adam, which combines momentum with adaptive learning rates. Its variants, such as AdamW, train most large models today.",
  "sources": [
   {
    "title": "Adam: A Method for Stochastic Optimization",
    "url": "https://arxiv.org/abs/1412.6980",
    "kind": "primary"
   }
  ],
  "era": "deep-learning",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Adam: A Method for Stochastic Optimization (v1 2014-12-22)"
   ]
  }
 },
 {
  "id": "batchnorm-2015",
  "date": "2015-02-11",
  "title": "Batch normalisation",
  "type": "paper",
  "labs": [
   "google"
  ],
  "layer": "training",
  "concepts": [
   "normalization"
  ],
  "people": [],
  "one_liner": "Rescaling each layer's inputs makes deep networks train faster and more reliably.",
  "summary": "Ioffe and Szegedy introduced batch normalisation, allowing higher learning rates and deeper networks. Later transformers use a cousin, layer normalisation.",
  "sources": [
   {
    "title": "Batch Normalization",
    "url": "https://arxiv.org/abs/1502.03167",
    "kind": "primary"
   }
  ],
  "era": "deep-learning",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift (v1 2015-02-11)"
   ]
  }
 },
 {
  "id": "tensorflow-2015",
  "date": "2015-11-09",
  "title": "Google open-sources TensorFlow",
  "type": "product",
  "labs": [
   "google"
  ],
  "layer": "infra",
  "concepts": [
   "deep-learning-framework"
  ],
  "people": [],
  "one_liner": "Google's internal deep learning library becomes free for everyone.",
  "summary": "TensorFlow made production-grade deep learning tools available to all developers and helped spread deep learning beyond research labs.",
  "sources": [
   {
    "title": "TensorFlow: smarter machine learning, for everyone",
    "url": "https://blog.google/technology/ai/tensorflow-smarter-machine-learning-for/",
    "kind": "primary"
   }
  ],
  "era": "deep-learning",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): TensorFlow: smarter machine learning, for everyone",
    "date 2015-11-09 found on page"
   ]
  }
 },
 {
  "id": "resnet-2015",
  "date": "2015-12-10",
  "title": "ResNet: very deep networks with skip connections",
  "type": "paper",
  "labs": [
   "microsoft"
  ],
  "layer": "architectures",
  "concepts": [
   "residual-connection",
   "cnn"
  ],
  "people": [],
  "one_liner": "Shortcut connections let networks go 150 layers deep and still train.",
  "summary": "Kaiming He and colleagues at Microsoft Research won ImageNet 2015 with residual networks. Residual connections are inside every transformer layer today.",
  "sources": [
   {
    "title": "Deep Residual Learning for Image Recognition",
    "url": "https://arxiv.org/abs/1512.03385",
    "kind": "primary"
   }
  ],
  "era": "deep-learning",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Deep Residual Learning for Image Recognition (v1 2015-12-10)"
   ]
  }
 },
 {
  "id": "openai-founded-2015",
  "date": "2015-12-11",
  "title": "OpenAI is founded",
  "type": "event",
  "labs": [
   "openai"
  ],
  "layer": "models",
  "concepts": [
   "agi"
  ],
  "people": [
   "sam-altman",
   "ilya-sutskever"
  ],
  "one_liner": "A non-profit AI lab launches with the goal of ensuring AGI benefits everyone.",
  "summary": "OpenAI was announced with Ilya Sutskever as research director and Sam Altman and Elon Musk as co-chairs, among others. It later created GPT and ChatGPT.",
  "sources": [
   {
    "title": "Introducing OpenAI",
    "url": "https://openai.com/index/introducing-openai/",
    "kind": "primary"
   }
  ],
  "era": "deep-learning",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Introducing OpenAI | OpenAI",
    "date 2015-12-11 found on page"
   ]
  }
 },
 {
  "id": "alphago-2016",
  "date": "2016-01-27",
  "title": "AlphaGo masters Go",
  "type": "paper",
  "labs": [
   "google"
  ],
  "layer": "learning",
  "concepts": [
   "reinforcement-learning",
   "deep-learning"
  ],
  "people": [
   "demis-hassabis"
  ],
  "one_liner": "Neural networks plus tree search beat a Go professional, then world champion Lee Sedol in March.",
  "summary": "DeepMind published AlphaGo in Nature after it beat European champion Fan Hui. In March 2016 it beat Lee Sedol 4 to 1, a decade earlier than many experts expected.",
  "sources": [
   {
    "title": "Mastering the game of Go with deep neural networks and tree search",
    "url": "https://doi.org/10.1038/nature16961",
    "kind": "primary"
   }
  ],
  "era": "deep-learning",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "Crossref: Mastering the game of Go with deep neural networks and tree search [[2016, 1, 27], [2016, 1, 28], [2016, 1, 27]]"
   ]
  }
 },
 {
  "id": "moe-2017",
  "date": "2017-01-23",
  "title": "Sparsely-gated mixture of experts",
  "type": "paper",
  "labs": [
   "google"
  ],
  "layer": "architectures",
  "concepts": [
   "mixture-of-experts"
  ],
  "people": [
   "geoffrey-hinton"
  ],
  "one_liner": "A giant model where each input only wakes up a few expert sub-networks.",
  "summary": "Shazeer, Hinton, Dean and colleagues trained networks with thousands of experts and a learned router. The idea returned in models such as Mixtral, DeepSeek-V3 and many frontier models.",
  "sources": [
   {
    "title": "Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer",
    "url": "https://arxiv.org/abs/1701.06538",
    "kind": "primary"
   }
  ],
  "era": "transformers",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer (v1 2017-01-23)"
   ]
  }
 },
 {
  "id": "rlhf-2017",
  "date": "2017-06-12",
  "title": "Learning from human preferences",
  "type": "paper",
  "labs": [
   "openai",
   "google"
  ],
  "layer": "training",
  "concepts": [
   "rlhf"
  ],
  "people": [
   "dario-amodei"
  ],
  "one_liner": "Humans compare two behaviours, and the agent learns what they prefer.",
  "summary": "Christiano, Leike, Amodei and colleagues trained agents from human comparisons instead of a hand-written reward. This became RLHF, the step that turned GPT-3 into ChatGPT.",
  "sources": [
   {
    "title": "Deep reinforcement learning from human preferences",
    "url": "https://arxiv.org/abs/1706.03741",
    "kind": "primary"
   }
  ],
  "era": "transformers",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Deep reinforcement learning from human preferences (v1 2017-06-12)"
   ]
  }
 },
 {
  "id": "transformer-2017",
  "date": "2017-06-12",
  "title": "Attention Is All You Need: the transformer",
  "type": "paper",
  "labs": [
   "google"
  ],
  "layer": "architectures",
  "concepts": [
   "transformer",
   "attention"
  ],
  "people": [],
  "one_liner": "A network built only from attention, which trains in parallel and scales well.",
  "summary": "Vaswani and colleagues at Google dropped recurrence entirely. The transformer became the architecture behind BERT, GPT, Claude, Gemini and almost every modern model.",
  "sources": [
   {
    "title": "Attention Is All You Need",
    "url": "https://arxiv.org/abs/1706.03762",
    "kind": "primary"
   }
  ],
  "era": "transformers",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Attention Is All You Need (v1 2017-06-12)"
   ]
  }
 },
 {
  "id": "ppo-2017",
  "date": "2017-07-20",
  "title": "Proximal policy optimisation (PPO)",
  "type": "paper",
  "labs": [
   "openai"
  ],
  "layer": "learning",
  "concepts": [
   "reinforcement-learning"
  ],
  "people": [],
  "one_liner": "A stable, simple reinforcement learning algorithm that takes careful small steps.",
  "summary": "Schulman and colleagues at OpenAI introduced PPO. It became the default RL algorithm, including the RL step of RLHF in InstructGPT.",
  "sources": [
   {
    "title": "Proximal Policy Optimization Algorithms",
    "url": "https://arxiv.org/abs/1707.06347",
    "kind": "primary"
   }
  ],
  "era": "transformers",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Proximal Policy Optimization Algorithms (v1 2017-07-20)"
   ]
  }
 },
 {
  "id": "elmo-2018",
  "date": "2018-02-15",
  "title": "ELMo: word meaning depends on context",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "models",
  "concepts": [
   "pretraining",
   "word-embedding"
  ],
  "people": [],
  "one_liner": "The same word gets a different vector depending on the sentence around it.",
  "summary": "Peters and colleagues at the Allen Institute for AI used a pre-trained bidirectional LSTM to produce contextual embeddings, a step toward pre-train then fine-tune.",
  "sources": [
   {
    "title": "Deep contextualized word representations",
    "url": "https://arxiv.org/abs/1802.05365",
    "kind": "primary"
   }
  ],
  "era": "transformers",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Deep contextualized word representations (v1 2018-02-15)"
   ]
  }
 },
 {
  "id": "gpt1-2018",
  "date": "2018-06-11",
  "title": "GPT-1: generative pre-training",
  "type": "model",
  "labs": [
   "openai"
  ],
  "layer": "models",
  "concepts": [
   "pretraining",
   "fine-tuning",
   "llm"
  ],
  "people": [
   "ilya-sutskever"
  ],
  "one_liner": "Pre-train a transformer on lots of text, then fine-tune it for each task.",
  "summary": "OpenAI showed that unsupervised pre-training of a transformer decoder followed by fine-tuning beat task-specific models. It started the GPT line.",
  "sources": [
   {
    "title": "Improving language understanding with unsupervised learning",
    "url": "https://openai.com/index/language-unsupervised/",
    "kind": "primary"
   }
  ],
  "era": "transformers",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Improving language understanding with unsupervised learning | OpenAI",
    "date 2018-06-11 found on page"
   ]
  }
 },
 {
  "id": "bert-2018",
  "date": "2018-10-11",
  "title": "BERT reads in both directions",
  "type": "model",
  "labs": [
   "google"
  ],
  "layer": "models",
  "concepts": [
   "pretraining",
   "transformer",
   "fine-tuning"
  ],
  "people": [],
  "one_liner": "A transformer trained to fill in masked words tops most language benchmarks.",
  "summary": "Devlin and colleagues at Google pre-trained a bidirectional transformer encoder. BERT was soon used in Google Search and set the pre-train then fine-tune pattern for NLP.",
  "sources": [
   {
    "title": "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding",
    "url": "https://arxiv.org/abs/1810.04805",
    "kind": "primary"
   }
  ],
  "era": "transformers",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (v1 2018-10-11)"
   ]
  }
 },
 {
  "id": "gpt2-2019",
  "date": "2019-02-14",
  "title": "GPT-2 and staged release",
  "type": "model",
  "labs": [
   "openai"
  ],
  "layer": "models",
  "concepts": [
   "llm",
   "zero-shot"
  ],
  "people": [],
  "one_liner": "A bigger GPT writes convincing paragraphs, and OpenAI releases it in stages over misuse concerns.",
  "summary": "GPT-2 (1.5 billion parameters) showed zero-shot abilities on many tasks without fine-tuning. Its staged release started the debate about releasing powerful models openly.",
  "sources": [
   {
    "title": "Better language models and their implications",
    "url": "https://openai.com/index/better-language-models/",
    "kind": "primary"
   }
  ],
  "era": "transformers",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Better language models and their implications | OpenAI",
    "date 2019-02-14 found on page"
   ]
  }
 },
 {
  "id": "t5-2019",
  "date": "2019-10-23",
  "title": "T5: every task as text to text",
  "type": "model",
  "labs": [
   "google"
  ],
  "layer": "models",
  "concepts": [
   "transfer-learning",
   "pretraining"
  ],
  "people": [],
  "one_liner": "Translation, summarising and classification all become 'text in, text out'.",
  "summary": "Raffel and colleagues framed all NLP tasks as text-to-text and studied transfer learning at scale. The C4 dataset from this paper was reused widely.",
  "sources": [
   {
    "title": "Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer",
    "url": "https://arxiv.org/abs/1910.10683",
    "kind": "primary"
   }
  ],
  "era": "transformers",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer (v1 2019-10-23)"
   ]
  }
 },
 {
  "id": "scaling-laws-2020",
  "date": "2020-01-23",
  "title": "Scaling laws for language models",
  "type": "paper",
  "labs": [
   "openai"
  ],
  "layer": "training",
  "concepts": [
   "scaling-laws"
  ],
  "people": [
   "dario-amodei"
  ],
  "one_liner": "Loss falls predictably as you add parameters, data and compute.",
  "summary": "Kaplan, McCandlish, Amodei and colleagues found smooth power laws linking model size, data and compute to performance. It justified the race to train ever bigger models.",
  "sources": [
   {
    "title": "Scaling Laws for Neural Language Models",
    "url": "https://arxiv.org/abs/2001.08361",
    "kind": "primary"
   }
  ],
  "era": "transformers",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Scaling Laws for Neural Language Models (v1 2020-01-23)"
   ]
  }
 },
 {
  "id": "rag-2020",
  "date": "2020-05-22",
  "title": "Retrieval-augmented generation (RAG)",
  "type": "paper",
  "labs": [
   "meta"
  ],
  "layer": "building",
  "concepts": [
   "rag",
   "vector-database",
   "embedding"
  ],
  "people": [],
  "one_liner": "Look up relevant documents first, then generate the answer from them.",
  "summary": "Lewis and colleagues at Facebook AI combined a retriever with a generator. RAG became the standard way to ground LLM answers in company documents.",
  "sources": [
   {
    "title": "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks",
    "url": "https://arxiv.org/abs/2005.11401",
    "kind": "primary"
   }
  ],
  "era": "transformers",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (v1 2020-05-22)"
   ]
  }
 },
 {
  "id": "gpt3-2020",
  "date": "2020-05-28",
  "title": "GPT-3 and in-context learning",
  "type": "model",
  "labs": [
   "openai"
  ],
  "layer": "models",
  "concepts": [
   "llm",
   "in-context-learning",
   "few-shot"
  ],
  "people": [
   "dario-amodei"
  ],
  "one_liner": "A 175-billion-parameter model learns new tasks from a few examples in the prompt.",
  "summary": "Brown and colleagues showed that a very large language model can do tasks from examples in the prompt alone, with no fine-tuning. The OpenAI API launched around it and prompt engineering began.",
  "sources": [
   {
    "title": "Language Models are Few-Shot Learners",
    "url": "https://arxiv.org/abs/2005.14165",
    "kind": "primary"
   }
  ],
  "era": "transformers",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Language Models are Few-Shot Learners (v1 2020-05-28)"
   ]
  }
 },
 {
  "id": "ddpm-2020",
  "date": "2020-06-19",
  "title": "Denoising diffusion models",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "architectures",
  "concepts": [
   "diffusion-model",
   "generative-model"
  ],
  "people": [],
  "one_liner": "Learn to remove noise step by step, then generate images from pure noise.",
  "summary": "Ho, Jain and Abbeel showed that diffusion models produce high-quality images. Diffusion became the basis of DALL-E 2, Stable Diffusion, Midjourney and video models.",
  "sources": [
   {
    "title": "Denoising Diffusion Probabilistic Models",
    "url": "https://arxiv.org/abs/2006.11239",
    "kind": "primary"
   }
  ],
  "era": "transformers",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Denoising Diffusion Probabilistic Models (v1 2020-06-19)"
   ]
  }
 },
 {
  "id": "vit-2020",
  "date": "2020-10-22",
  "title": "Vision transformer (ViT)",
  "type": "paper",
  "labs": [
   "google"
  ],
  "layer": "architectures",
  "concepts": [
   "vision-transformer",
   "transformer"
  ],
  "people": [],
  "one_liner": "Cut an image into patches, treat them like words, and a plain transformer sees well.",
  "summary": "Dosovitskiy and colleagues showed transformers can match CNNs on images at scale. It unified vision and language architectures and paved the way for multimodal models.",
  "sources": [
   {
    "title": "An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale",
    "url": "https://arxiv.org/abs/2010.11929",
    "kind": "primary"
   }
  ],
  "era": "transformers",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (v1 2020-10-22)"
   ]
  }
 },
 {
  "id": "dalle-2021",
  "date": "2021-01-05",
  "title": "DALL-E generates images from text",
  "type": "model",
  "labs": [
   "openai"
  ],
  "layer": "models",
  "concepts": [
   "multimodal-model",
   "generative-model"
  ],
  "people": [],
  "one_liner": "Type a description and a model draws it.",
  "summary": "OpenAI's DALL-E showed a transformer generating images from text prompts, such as an avocado armchair. Text-to-image went mainstream with DALL-E 2 and Stable Diffusion in 2022.",
  "sources": [
   {
    "title": "DALL-E: Creating images from text",
    "url": "https://openai.com/index/dall-e/",
    "kind": "primary"
   }
  ],
  "era": "transformers",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): DALL·E: Creating images from text | OpenAI",
    "date 2021-01-05 found on page"
   ]
  }
 },
 {
  "id": "clip-2021",
  "date": "2021-02-26",
  "title": "CLIP connects images and text",
  "type": "model",
  "labs": [
   "openai"
  ],
  "layer": "models",
  "concepts": [
   "multimodal-model",
   "embedding",
   "zero-shot"
  ],
  "people": [],
  "one_liner": "Trained on image and caption pairs, it can classify images by any text label.",
  "summary": "Radford and colleagues trained image and text encoders to match 400 million internet pairs. CLIP embeddings guide image generators and power zero-shot image search.",
  "sources": [
   {
    "title": "Learning Transferable Visual Models From Natural Language Supervision",
    "url": "https://arxiv.org/abs/2103.00020",
    "kind": "primary"
   }
  ],
  "era": "transformers",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Learning Transferable Visual Models From Natural Language Supervision (v1 2021-02-26)"
   ]
  }
 },
 {
  "id": "lora-2021",
  "date": "2021-06-17",
  "title": "LoRA: cheap fine-tuning",
  "type": "paper",
  "labs": [
   "microsoft"
  ],
  "layer": "training",
  "concepts": [
   "lora",
   "fine-tuning"
  ],
  "people": [],
  "one_liner": "Freeze the big model and train tiny add-on matrices instead.",
  "summary": "Hu and colleagues at Microsoft introduced low-rank adaptation. It cut fine-tuning cost by orders of magnitude and made custom models possible on a single GPU.",
  "sources": [
   {
    "title": "LoRA: Low-Rank Adaptation of Large Language Models",
    "url": "https://arxiv.org/abs/2106.09685",
    "kind": "primary"
   }
  ],
  "era": "transformers",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: LoRA: Low-Rank Adaptation of Large Language Models (v1 2021-06-17)"
   ]
  }
 },
 {
  "id": "github-copilot-2021",
  "date": "2021-06-29",
  "title": "GitHub Copilot previews AI pair programming",
  "type": "product",
  "labs": [
   "microsoft",
   "openai"
  ],
  "layer": "building",
  "concepts": [
   "coding-assistant",
   "llm"
  ],
  "people": [],
  "one_liner": "An AI that suggests whole lines and functions inside your editor.",
  "summary": "GitHub and OpenAI launched Copilot as a technical preview, built on OpenAI Codex. It was the first AI developer tool used by millions and started the AI coding tools market.",
  "sources": [
   {
    "title": "Introducing GitHub Copilot: your AI pair programmer",
    "url": "https://github.blog/news-insights/product-news/introducing-github-copilot-ai-pair-programmer/",
    "kind": "primary"
   }
  ],
  "era": "transformers",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Introducing GitHub Copilot: your AI pair programmer - The GitHub Blog",
    "date 2021-06-29 found on page"
   ]
  }
 },
 {
  "id": "alphafold2-2021",
  "date": "2021-07-15",
  "title": "AlphaFold 2 predicts protein structures",
  "type": "paper",
  "labs": [
   "google"
  ],
  "layer": "models",
  "concepts": [
   "deep-learning"
  ],
  "people": [
   "demis-hassabis"
  ],
  "one_liner": "Protein shapes that took years in the lab are predicted in minutes.",
  "summary": "DeepMind published AlphaFold 2 in Nature and released code. It solved a 50-year-old biology challenge and earned Demis Hassabis and John Jumper a share of the 2024 Nobel Prize in Chemistry.",
  "sources": [
   {
    "title": "Highly accurate protein structure prediction with AlphaFold",
    "url": "https://doi.org/10.1038/s41586-021-03819-2",
    "kind": "primary"
   }
  ],
  "era": "transformers",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "Crossref: Highly accurate protein structure prediction with AlphaFold [[2021, 7, 15], [2021, 8, 26], [2021, 7, 15]]"
   ]
  }
 },
 {
  "id": "cot-2022",
  "date": "2022-01-28",
  "title": "Chain-of-thought prompting",
  "type": "paper",
  "labs": [
   "google"
  ],
  "layer": "building",
  "concepts": [
   "chain-of-thought",
   "prompt-engineering"
  ],
  "people": [],
  "one_liner": "Showing the model worked steps makes it reason better.",
  "summary": "Wei and colleagues found that prompting large models with step-by-step examples sharply improved maths and logic. It inspired the reasoning models of 2024.",
  "sources": [
   {
    "title": "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models",
    "url": "https://arxiv.org/abs/2201.11903",
    "kind": "primary"
   }
  ],
  "era": "generative",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (v1 2022-01-28)"
   ]
  }
 },
 {
  "id": "instructgpt-2022",
  "date": "2022-03-04",
  "title": "InstructGPT: following instructions with RLHF",
  "type": "model",
  "labs": [
   "openai"
  ],
  "layer": "training",
  "concepts": [
   "rlhf",
   "instruction-tuning",
   "alignment"
  ],
  "people": [],
  "one_liner": "A small model tuned on human feedback is preferred over a 100 times bigger raw model.",
  "summary": "Ouyang and colleagues fine-tuned GPT-3 on demonstrations and human preference rankings. The recipe became the basis of ChatGPT.",
  "sources": [
   {
    "title": "Training language models to follow instructions with human feedback",
    "url": "https://arxiv.org/abs/2203.02155",
    "kind": "primary"
   }
  ],
  "era": "generative",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Training language models to follow instructions with human feedback (v1 2022-03-04)"
   ]
  }
 },
 {
  "id": "chinchilla-2022",
  "date": "2022-03-29",
  "title": "Chinchilla: train on more data, not just bigger models",
  "type": "paper",
  "labs": [
   "google"
  ],
  "layer": "training",
  "concepts": [
   "scaling-laws"
  ],
  "people": [],
  "one_liner": "For a fixed budget, a smaller model on more tokens wins.",
  "summary": "Hoffmann and colleagues at DeepMind showed most large models were under-trained. Their 70B Chinchilla beat the 280B Gopher, and data-heavy training became the norm.",
  "sources": [
   {
    "title": "Training Compute-Optimal Large Language Models",
    "url": "https://arxiv.org/abs/2203.15556",
    "kind": "primary"
   }
  ],
  "era": "generative",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Training Compute-Optimal Large Language Models (v1 2022-03-29)"
   ]
  }
 },
 {
  "id": "palm-2022",
  "date": "2022-04-05",
  "title": "PaLM scales to 540 billion parameters",
  "type": "model",
  "labs": [
   "google"
  ],
  "layer": "models",
  "concepts": [
   "llm",
   "chain-of-thought"
  ],
  "people": [],
  "one_liner": "A very large Google model shows breakthrough reasoning with chain-of-thought.",
  "summary": "Chowdhery and colleagues trained PaLM on Google's Pathways system. It powered early versions of Bard.",
  "sources": [
   {
    "title": "PaLM: Scaling Language Modeling with Pathways",
    "url": "https://arxiv.org/abs/2204.02311",
    "kind": "primary"
   }
  ],
  "era": "generative",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: PaLM: Scaling Language Modeling with Pathways (v1 2022-04-05)"
   ]
  }
 },
 {
  "id": "stable-diffusion-2022",
  "date": "2022-08-22",
  "title": "Stable Diffusion is released openly",
  "type": "model",
  "labs": [
   "stability",
   "academia"
  ],
  "layer": "models",
  "concepts": [
   "diffusion-model",
   "open-weight-model"
  ],
  "people": [],
  "one_liner": "A high-quality image model anyone can run on a home GPU.",
  "summary": "Stability AI, LMU Munich's CompVis group and Runway released Stable Diffusion with public weights, based on latent diffusion. It started the open image-generation ecosystem.",
  "sources": [
   {
    "title": "Stable Diffusion public release",
    "url": "https://stability.ai/news/stable-diffusion-public-release",
    "kind": "primary"
   },
   {
    "title": "High-Resolution Image Synthesis with Latent Diffusion Models",
    "url": "https://arxiv.org/abs/2112.10752",
    "kind": "paper"
   }
  ],
  "era": "generative",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Stable Diffusion Public Release — Stability AI",
    "date 2022-08-22 found on page",
    "arXiv: High-Resolution Image Synthesis with Latent Diffusion Models (v1 2021-12-20)",
    "paper posted 2021-12-20, after the 2022-08-22 launch: supporting source"
   ]
  }
 },
 {
  "id": "react-2022",
  "date": "2022-10-06",
  "title": "ReAct: reasoning and acting",
  "type": "paper",
  "labs": [
   "google",
   "academia"
  ],
  "layer": "building",
  "concepts": [
   "react-pattern",
   "agent",
   "tool-use"
  ],
  "people": [],
  "one_liner": "The model thinks, takes an action like a search, reads the result, and thinks again.",
  "summary": "Yao and colleagues interleaved reasoning traces with tool actions. The think, act, observe loop is the basic pattern inside most AI agents.",
  "sources": [
   {
    "title": "ReAct: Synergizing Reasoning and Acting in Language Models",
    "url": "https://arxiv.org/abs/2210.03629",
    "kind": "primary"
   }
  ],
  "era": "generative",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: ReAct: Synergizing Reasoning and Acting in Language Models (v1 2022-10-06)"
   ]
  }
 },
 {
  "id": "chatgpt-2022",
  "date": "2022-11-30",
  "title": "ChatGPT launches",
  "type": "product",
  "labs": [
   "openai"
  ],
  "layer": "models",
  "concepts": [
   "chatbot",
   "llm",
   "rlhf"
  ],
  "people": [
   "sam-altman"
  ],
  "one_liner": "A free chat interface to a GPT-3.5 model becomes the fastest growing consumer app of its time.",
  "summary": "OpenAI released ChatGPT as a research preview. It brought large language models to the public and set off the generative AI race across every major tech company.",
  "sources": [
   {
    "title": "Introducing ChatGPT",
    "url": "https://openai.com/index/chatgpt/",
    "kind": "primary"
   }
  ],
  "era": "generative",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Introducing ChatGPT | OpenAI",
    "date 2022-11-30 found on page"
   ]
  }
 },
 {
  "id": "constitutional-ai-2022",
  "date": "2022-12-15",
  "title": "Constitutional AI",
  "type": "paper",
  "labs": [
   "anthropic"
  ],
  "layer": "training",
  "concepts": [
   "constitutional-ai",
   "alignment",
   "rlhf"
  ],
  "people": [
   "dario-amodei"
  ],
  "one_liner": "A model critiques and revises its own answers using a written set of principles.",
  "summary": "Anthropic trained a harmless assistant using AI feedback guided by a constitution instead of human labels for harm. The method shapes how Claude is trained.",
  "sources": [
   {
    "title": "Constitutional AI: Harmlessness from AI Feedback",
    "url": "https://arxiv.org/abs/2212.08073",
    "kind": "primary"
   }
  ],
  "era": "generative",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Constitutional AI: Harmlessness from AI Feedback (v1 2022-12-15)"
   ]
  }
 },
 {
  "id": "msft-openai-2023",
  "date": "2023-01-23",
  "title": "Microsoft extends its OpenAI partnership",
  "type": "event",
  "labs": [
   "microsoft",
   "openai"
  ],
  "layer": "infra",
  "concepts": [
   "llm"
  ],
  "people": [],
  "one_liner": "A multi-year, multi-billion-dollar investment, with Azure as OpenAI's cloud.",
  "summary": "Microsoft announced the third phase of its partnership with OpenAI. It set up OpenAI models inside Azure OpenAI Service, Bing and later every Copilot product.",
  "sources": [
   {
    "title": "Microsoft and OpenAI extend partnership",
    "url": "https://blogs.microsoft.com/blog/2023/01/23/microsoftandopenaiextendpartnership/",
    "kind": "primary"
   }
  ],
  "era": "generative",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Microsoft and OpenAI extend partnership - The Official Microsoft Blog",
    "date 2023-01-23 found on page"
   ]
  }
 },
 {
  "id": "toolformer-2023",
  "date": "2023-02-09",
  "title": "Toolformer teaches models to use tools",
  "type": "paper",
  "labs": [
   "meta"
  ],
  "layer": "building",
  "concepts": [
   "tool-use"
  ],
  "people": [],
  "one_liner": "A model learns by itself when to call a calculator, search or calendar API.",
  "summary": "Schick and colleagues at Meta trained a model to insert API calls into its own text. It anticipated function calling, which every major API added later in 2023.",
  "sources": [
   {
    "title": "Toolformer: Language Models Can Teach Themselves to Use Tools",
    "url": "https://arxiv.org/abs/2302.04761",
    "kind": "primary"
   }
  ],
  "era": "generative",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Toolformer: Language Models Can Teach Themselves to Use Tools (v1 2023-02-09)"
   ]
  }
 },
 {
  "id": "llama-2023",
  "date": "2023-02-24",
  "title": "LLaMA starts the open-weight wave",
  "type": "model",
  "labs": [
   "meta"
  ],
  "layer": "models",
  "concepts": [
   "open-weight-model",
   "llm"
  ],
  "people": [],
  "one_liner": "Meta releases research models whose 13B version beats GPT-3 on many benchmarks.",
  "summary": "Touvron and colleagues released LLaMA to researchers. The weights spread widely and set off a wave of open fine-tunes, local models and tools such as llama.cpp.",
  "sources": [
   {
    "title": "Introducing LLaMA: A foundational, 65-billion-parameter language model",
    "url": "https://ai.meta.com/blog/large-language-model-llama-meta-ai/",
    "kind": "primary"
   },
   {
    "title": "LLaMA: Open and Efficient Foundation Language Models",
    "url": "https://arxiv.org/abs/2302.13971",
    "kind": "paper"
   }
  ],
  "era": "generative",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Introducing LLaMA: A foundational, 65-billion-parameter language model",
    "date 2023-02-24 found on page",
    "arXiv: LLaMA: Open and Efficient Foundation Language Models (v1 2023-02-27)",
    "paper posted 2023-02-27, after the 2023-02-24 launch: supporting source"
   ]
  }
 },
 {
  "id": "claude-2023",
  "date": "2023-03-14",
  "title": "Anthropic launches Claude",
  "type": "model",
  "labs": [
   "anthropic"
  ],
  "layer": "models",
  "concepts": [
   "llm",
   "constitutional-ai"
  ],
  "people": [
   "dario-amodei"
  ],
  "one_liner": "Anthropic's assistant opens to businesses, trained with Constitutional AI.",
  "summary": "Claude and Claude Instant became available through Anthropic's API after a closed alpha with partners.",
  "sources": [
   {
    "title": "Introducing Claude",
    "url": "https://www.anthropic.com/news/introducing-claude",
    "kind": "primary"
   }
  ],
  "era": "generative",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Introducing Claude \\ Anthropic",
    "date 2023-03-14 found on page"
   ]
  }
 },
 {
  "id": "gpt4-2023",
  "date": "2023-03-14",
  "title": "GPT-4",
  "type": "model",
  "labs": [
   "openai"
  ],
  "layer": "models",
  "concepts": [
   "llm",
   "multimodal-model"
  ],
  "people": [],
  "one_liner": "A large step in reasoning, passing many professional exams, with image input.",
  "summary": "OpenAI released GPT-4 with a technical report that withheld architecture details. It scored around the top 10 percent on a simulated bar exam and set the bar for frontier models.",
  "sources": [
   {
    "title": "GPT-4",
    "url": "https://openai.com/index/gpt-4-research/",
    "kind": "primary"
   },
   {
    "title": "GPT-4 Technical Report",
    "url": "https://arxiv.org/abs/2303.08774",
    "kind": "paper"
   }
  ],
  "era": "generative",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): GPT-4 | OpenAI",
    "date 2023-03-14 found on page",
    "arXiv: GPT-4 Technical Report (v1 2023-03-15)",
    "paper posted 2023-03-15, after the 2023-03-14 launch: supporting source"
   ]
  }
 },
 {
  "id": "m365-copilot-2023",
  "date": "2023-03-16",
  "title": "Microsoft 365 Copilot is announced",
  "type": "product",
  "labs": [
   "microsoft"
  ],
  "layer": "building",
  "concepts": [
   "copilot",
   "llm"
  ],
  "people": [],
  "one_liner": "GPT-4 grounded in your emails, documents and meetings, inside Word, Excel and Teams.",
  "summary": "Microsoft announced Copilot for its Office apps, combining LLMs with the Microsoft Graph. It moved generative AI into everyday enterprise work.",
  "sources": [
   {
    "title": "Introducing Microsoft 365 Copilot",
    "url": "https://blogs.microsoft.com/blog/2023/03/16/introducing-microsoft-365-copilot-your-copilot-for-work/",
    "kind": "primary"
   }
  ],
  "era": "generative",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Introducing Microsoft 365 Copilot – your copilot for work - The Official Microsoft Blog",
    "date 2023-03-16 found on page"
   ]
  }
 },
 {
  "id": "fabric-2023",
  "date": "2023-05-23",
  "title": "Microsoft Fabric and Copilot in Power BI are announced",
  "type": "product",
  "labs": [
   "microsoft"
  ],
  "layer": "data",
  "concepts": [
   "microsoft-fabric",
   "copilot-in-power-bi",
   "onelake"
  ],
  "people": [],
  "one_liner": "One analytics platform with a single lake, and Copilot inside Power BI.",
  "summary": "At Build 2023, Microsoft announced Fabric, unifying Data Factory, Synapse and Power BI on OneLake, plus Copilot for Power BI to build reports from natural language.",
  "sources": [
   {
    "title": "Introducing Microsoft Fabric: Data analytics for the era of AI",
    "url": "https://azure.microsoft.com/en-us/blog/introducing-microsoft-fabric-data-analytics-for-the-era-of-ai/",
    "kind": "primary"
   }
  ],
  "era": "generative",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Introducing Microsoft Fabric: The data platform for the era of AI | Microsoft Azure Blog | Microsoft Azure",
    "date 2023-05-23 found on page"
   ]
  }
 },
 {
  "id": "qlora-2023",
  "date": "2023-05-23",
  "title": "QLoRA: fine-tune a 65B model on one GPU",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "training",
  "concepts": [
   "lora",
   "quantization",
   "fine-tuning"
  ],
  "people": [],
  "one_liner": "Quantise the frozen model to 4 bits, then LoRA-tune it cheaply.",
  "summary": "Dettmers and colleagues combined 4-bit quantisation with LoRA, bringing large-model fine-tuning to hobbyists and small teams.",
  "sources": [
   {
    "title": "QLoRA: Efficient Finetuning of Quantized LLMs",
    "url": "https://arxiv.org/abs/2305.14314",
    "kind": "primary"
   }
  ],
  "era": "generative",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: QLoRA: Efficient Finetuning of Quantized LLMs (v1 2023-05-23)"
   ]
  }
 },
 {
  "id": "dpo-2023",
  "date": "2023-05-29",
  "title": "Direct preference optimisation (DPO)",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "training",
  "concepts": [
   "dpo",
   "rlhf"
  ],
  "people": [],
  "one_liner": "Learn from human preferences without a separate reward model or RL loop.",
  "summary": "Rafailov and colleagues at Stanford showed preference tuning can be a simple classification loss. DPO became a popular, cheaper alternative to PPO-based RLHF.",
  "sources": [
   {
    "title": "Direct Preference Optimization: Your Language Model is Secretly a Reward Model",
    "url": "https://arxiv.org/abs/2305.18290",
    "kind": "primary"
   }
  ],
  "era": "generative",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Direct Preference Optimization: Your Language Model is Secretly a Reward Model (v1 2023-05-29)"
   ]
  }
 },
 {
  "id": "llama2-2023",
  "date": "2023-07-18",
  "title": "Llama 2 allows commercial use",
  "type": "model",
  "labs": [
   "meta",
   "microsoft"
  ],
  "layer": "models",
  "concepts": [
   "open-weight-model"
  ],
  "people": [],
  "one_liner": "Open weights with a licence that allows most businesses to build on them.",
  "summary": "Meta released Llama 2 with Microsoft as a launch partner. Commercial open weights made self-hosted LLMs a real option for companies.",
  "sources": [
   {
    "title": "Llama 2: Open Foundation and Fine-Tuned Chat Models",
    "url": "https://arxiv.org/abs/2307.09288",
    "kind": "primary"
   }
  ],
  "era": "generative",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Llama 2: Open Foundation and Fine-Tuned Chat Models (v1 2023-07-18)"
   ]
  }
 },
 {
  "id": "mistral7b-2023",
  "date": "2023-09-27",
  "title": "Mistral 7B",
  "type": "model",
  "labs": [
   "mistral"
  ],
  "layer": "models",
  "concepts": [
   "open-weight-model",
   "small-language-model"
  ],
  "people": [],
  "one_liner": "A 7-billion-parameter model under Apache 2.0 that beats larger Llama models.",
  "summary": "Mistral AI's first model showed that careful training makes small models strong. It ran on laptops and became a favourite base for fine-tunes.",
  "sources": [
   {
    "title": "Announcing Mistral 7B",
    "url": "https://mistral.ai/news/announcing-mistral-7b",
    "kind": "primary"
   }
  ],
  "era": "generative",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Mistral 7B",
    "date 2023-09-27 found on page"
   ]
  }
 },
 {
  "id": "bletchley-2023",
  "date": "2023-11-01",
  "title": "The Bletchley Declaration on AI safety",
  "type": "event",
  "labs": [
   "government"
  ],
  "layer": "safety",
  "concepts": [
   "ai-safety",
   "ai-regulation"
  ],
  "people": [],
  "one_liner": "28 countries and the EU agree that frontier AI risks need international cooperation.",
  "summary": "At the first AI Safety Summit at Bletchley Park, governments including the US, China and the UK signed a shared declaration on frontier AI risks. The UK and US set up AI safety institutes around the same time.",
  "sources": [
   {
    "title": "The Bletchley Declaration",
    "url": "https://www.gov.uk/government/publications/ai-safety-summit-2023-the-bletchley-declaration",
    "kind": "primary"
   }
  ],
  "era": "generative",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): AI Safety Summit 2023: The Bletchley Declaration - GOV.UK",
    "date 2023-11-01 found on page"
   ]
  }
 },
 {
  "id": "openai-board-2023",
  "date": "2023-11-17",
  "title": "OpenAI's board removes Sam Altman, who returns days later",
  "type": "event",
  "labs": [
   "openai",
   "microsoft"
  ],
  "layer": "safety",
  "concepts": [
   "ai-governance"
  ],
  "people": [
   "sam-altman",
   "ilya-sutskever"
  ],
  "one_liner": "Five days of turmoil end with Altman back as CEO and a new board.",
  "summary": "OpenAI's board said it had lost confidence in Altman. After most staff threatened to leave for Microsoft, he returned as CEO on 22 November with a new initial board.",
  "sources": [
   {
    "title": "OpenAI announces leadership transition",
    "url": "https://openai.com/index/openai-announces-leadership-transition/",
    "kind": "primary"
   }
  ],
  "era": "generative",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): OpenAI announces leadership transition | OpenAI",
    "date 2023-11-17 found on page"
   ]
  }
 },
 {
  "id": "mamba-2023",
  "date": "2023-12-01",
  "title": "Mamba: state space models challenge transformers",
  "type": "paper",
  "labs": [
   "academia"
  ],
  "layer": "architectures",
  "concepts": [
   "state-space-model"
  ],
  "people": [],
  "one_liner": "An architecture that handles long sequences in linear time, without attention.",
  "summary": "Gu and Dao introduced selective state space models. Mamba-style layers now appear in hybrid models that mix them with attention for long contexts.",
  "sources": [
   {
    "title": "Mamba: Linear-Time Sequence Modeling with Selective State Spaces",
    "url": "https://arxiv.org/abs/2312.00752",
    "kind": "primary"
   }
  ],
  "era": "generative",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "arXiv: Mamba: Linear-Time Sequence Modeling with Selective State Spaces (v1 2023-12-01)"
   ]
  }
 },
 {
  "id": "gemini-2023",
  "date": "2023-12-06",
  "title": "Gemini: Google's natively multimodal model",
  "type": "model",
  "labs": [
   "google"
  ],
  "layer": "models",
  "concepts": [
   "multimodal-model",
   "llm"
  ],
  "people": [
   "demis-hassabis"
  ],
  "one_liner": "One model family trained from the start on text, images, audio and video.",
  "summary": "Google DeepMind introduced Gemini 1.0 in Ultra, Pro and Nano sizes, replacing PaLM in Bard and later renaming Bard itself to Gemini.",
  "sources": [
   {
    "title": "Introducing Gemini",
    "url": "https://blog.google/technology/ai/google-gemini-ai/",
    "kind": "primary"
   },
   {
    "title": "Gemini: A Family of Highly Capable Multimodal Models",
    "url": "https://arxiv.org/abs/2312.11805",
    "kind": "paper"
   }
  ],
  "era": "generative",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Introducing Gemini: Google’s most capable AI model yet",
    "date 2023-12-06 found on page",
    "arXiv: Gemini: A Family of Highly Capable Multimodal Models (v1 2023-12-19)",
    "paper posted 2023-12-19, after the 2023-12-06 launch: supporting source"
   ]
  }
 },
 {
  "id": "mixtral-2023",
  "date": "2023-12-11",
  "title": "Mixtral brings mixture of experts to open weights",
  "type": "model",
  "labs": [
   "mistral"
  ],
  "layer": "models",
  "concepts": [
   "mixture-of-experts",
   "open-weight-model"
  ],
  "people": [],
  "one_liner": "47B parameters in total, but only about 13B used per token.",
  "summary": "Mistral released Mixtral 8x7B under Apache 2.0. It matched GPT-3.5 on many benchmarks and popularised sparse MoE for open models.",
  "sources": [
   {
    "title": "Mixtral of experts",
    "url": "https://mistral.ai/news/mixtral-of-experts",
    "kind": "primary"
   }
  ],
  "era": "generative",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Mixtral of experts | Mistral AI",
    "date 2023-12-11 found on page"
   ]
  }
 },
 {
  "id": "claude3-2024",
  "date": "2024-03-04",
  "title": "Claude 3 family",
  "type": "model",
  "labs": [
   "anthropic"
  ],
  "layer": "models",
  "concepts": [
   "multimodal-model",
   "llm"
  ],
  "people": [
   "dario-amodei"
  ],
  "one_liner": "Haiku, Sonnet and Opus: three sizes, with Opus leading several benchmarks.",
  "summary": "Anthropic released Claude 3 with vision input and a 200K token context window.",
  "sources": [
   {
    "title": "Introducing the next generation of Claude",
    "url": "https://www.anthropic.com/news/claude-3-family",
    "kind": "primary"
   }
  ],
  "era": "agents",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Introducing the next generation of Claude \\ Anthropic",
    "date 2024-03-04 found on page"
   ]
  }
 },
 {
  "id": "llama3-2024",
  "date": "2024-04-18",
  "title": "Llama 3",
  "type": "model",
  "labs": [
   "meta"
  ],
  "layer": "models",
  "concepts": [
   "open-weight-model"
  ],
  "people": [],
  "one_liner": "Meta's 8B and 70B open models trained on over 15 trillion tokens.",
  "summary": "Llama 3 narrowed the gap between open-weight and closed models and was built into Meta AI across Meta's apps.",
  "sources": [
   {
    "title": "Introducing Meta Llama 3",
    "url": "https://ai.meta.com/blog/meta-llama-3/",
    "kind": "primary"
   }
  ],
  "era": "agents",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Introducing Meta Llama 3: The most capable openly available LLM to date",
    "date 2024-04-18 found on page"
   ]
  }
 },
 {
  "id": "gpt4o-2024",
  "date": "2024-05-13",
  "title": "GPT-4o: real-time voice and vision",
  "type": "model",
  "labs": [
   "openai"
  ],
  "layer": "models",
  "concepts": [
   "multimodal-model"
  ],
  "people": [],
  "one_liner": "One model handles text, audio and images, answering speech in about human response time.",
  "summary": "OpenAI's 'omni' model processed audio natively instead of chaining speech-to-text, and was made available to free ChatGPT users.",
  "sources": [
   {
    "title": "Hello GPT-4o",
    "url": "https://openai.com/index/hello-gpt-4o/",
    "kind": "primary"
   }
  ],
  "era": "agents",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Hello GPT-4o | OpenAI",
    "date 2024-05-13 found on page"
   ]
  }
 },
 {
  "id": "ssi-2024",
  "date": "2024-06-19",
  "title": "Ilya Sutskever founds Safe Superintelligence",
  "type": "person",
  "labs": [
   "ssi"
  ],
  "layer": "safety",
  "concepts": [
   "ai-safety",
   "agi"
  ],
  "people": [
   "ilya-sutskever"
  ],
  "one_liner": "A lab with one product goal: a safe superintelligence.",
  "summary": "After leaving OpenAI in May 2024, Sutskever co-founded SSI with Daniel Gross and Daniel Levy, with offices in Palo Alto and Tel Aviv.",
  "sources": [
   {
    "title": "Safe Superintelligence Inc.",
    "url": "https://ssi.inc/",
    "kind": "primary"
   }
  ],
  "era": "agents",
  "precision": "day",
  "check": {
   "status": "link-ok",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Safe Superintelligence Inc."
   ]
  }
 },
 {
  "id": "claude35-2024",
  "date": "2024-06-20",
  "title": "Claude 3.5 Sonnet and Artifacts",
  "type": "model",
  "labs": [
   "anthropic"
  ],
  "layer": "models",
  "concepts": [
   "llm",
   "coding-assistant"
  ],
  "people": [],
  "one_liner": "A mid-size model that beats the previous top model, plus a side panel for live outputs.",
  "summary": "Claude 3.5 Sonnet raised coding and reasoning scores at mid-tier cost. Artifacts showed generated code and documents in a live window next to the chat.",
  "sources": [
   {
    "title": "Claude 3.5 Sonnet",
    "url": "https://www.anthropic.com/news/claude-3-5-sonnet",
    "kind": "primary"
   }
  ],
  "era": "agents",
  "precision": "day",
  "check": {
   "status": "link-ok",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Introducing Claude 3.5 Sonnet \\ Anthropic"
   ]
  }
 },
 {
  "id": "eu-ai-act-2024",
  "date": "2024-08-01",
  "title": "The EU AI Act enters into force",
  "type": "event",
  "labs": [
   "government"
  ],
  "layer": "safety",
  "concepts": [
   "ai-regulation"
  ],
  "people": [],
  "one_liner": "The first broad AI law sorts AI systems by risk, with obligations phased in over years.",
  "summary": "The EU AI Act bans some uses outright, sets strict rules for high-risk systems, and adds duties for providers of general-purpose AI models from August 2025.",
  "sources": [
   {
    "title": "AI Act enters into force",
    "url": "https://commission.europa.eu/news/ai-act-enters-force-2024-08-01_en",
    "kind": "primary"
   }
  ],
  "era": "agents",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): AI Act enters into force - European Commission",
    "date 2024-08-01 found on page"
   ]
  }
 },
 {
  "id": "o1-2024",
  "date": "2024-09-12",
  "title": "OpenAI o1: models that think before answering",
  "type": "model",
  "labs": [
   "openai"
  ],
  "layer": "models",
  "concepts": [
   "reasoning-model",
   "chain-of-thought",
   "test-time-compute"
  ],
  "people": [],
  "one_liner": "Trained with reinforcement learning to reason in a long hidden chain of thought.",
  "summary": "o1 spent more compute at answer time and jumped ahead on maths, coding and science benchmarks. It started the era of reasoning models.",
  "sources": [
   {
    "title": "Learning to reason with LLMs",
    "url": "https://openai.com/index/learning-to-reason-with-llms/",
    "kind": "primary"
   }
  ],
  "era": "agents",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Learning to reason with LLMs | OpenAI",
    "date 2024-09-12 found on page"
   ]
  }
 },
 {
  "id": "nobel-physics-2024",
  "date": "2024-10-08",
  "title": "Nobel Prize in Physics to Hopfield and Hinton",
  "type": "event",
  "labs": [
   "academia"
  ],
  "layer": "math",
  "concepts": [
   "hopfield-network",
   "backpropagation"
  ],
  "people": [
   "geoffrey-hinton"
  ],
  "one_liner": "Awarded for foundational discoveries that enable machine learning with artificial neural networks.",
  "summary": "John Hopfield and Geoffrey Hinton received the 2024 physics prize, recognising neural networks' roots in physics.",
  "sources": [
   {
    "title": "The Nobel Prize in Physics 2024",
    "url": "https://www.nobelprize.org/prizes/physics/2024/summary/",
    "kind": "primary"
   }
  ],
  "era": "agents",
  "precision": "day",
  "check": {
   "status": "link-ok",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): The Nobel Prize in Physics 2024 - NobelPrize.org"
   ]
  }
 },
 {
  "id": "nobel-chemistry-2024",
  "date": "2024-10-09",
  "title": "Nobel Prize in Chemistry for protein design and AlphaFold",
  "type": "event",
  "labs": [
   "google",
   "academia"
  ],
  "layer": "models",
  "concepts": [
   "deep-learning"
  ],
  "people": [
   "demis-hassabis"
  ],
  "one_liner": "David Baker for protein design; Demis Hassabis and John Jumper for structure prediction.",
  "summary": "Half the 2024 chemistry prize went to Hassabis and Jumper for AlphaFold, an AI system, alongside David Baker.",
  "sources": [
   {
    "title": "The Nobel Prize in Chemistry 2024",
    "url": "https://www.nobelprize.org/prizes/chemistry/2024/summary/",
    "kind": "primary"
   }
  ],
  "era": "agents",
  "precision": "day",
  "check": {
   "status": "link-ok",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): The Nobel Prize in Chemistry 2024 - NobelPrize.org"
   ]
  }
 },
 {
  "id": "computer-use-2024",
  "date": "2024-10-22",
  "title": "Claude computer use",
  "type": "product",
  "labs": [
   "anthropic"
  ],
  "layer": "building",
  "concepts": [
   "computer-use",
   "agent"
  ],
  "people": [],
  "one_liner": "A model that looks at the screen, moves the mouse and types like a person.",
  "summary": "Anthropic released computer use in public beta with an upgraded Claude 3.5 Sonnet, letting agents operate ordinary software through screenshots.",
  "sources": [
   {
    "title": "Introducing computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku",
    "url": "https://www.anthropic.com/news/3-5-models-and-computer-use",
    "kind": "primary"
   }
  ],
  "era": "agents",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Introducing computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku \\ Anthropic",
    "date 2024-10-22 found on page"
   ]
  }
 },
 {
  "id": "mcp-2024",
  "date": "2024-11-25",
  "title": "The Model Context Protocol (MCP)",
  "type": "product",
  "labs": [
   "anthropic"
  ],
  "layer": "building",
  "concepts": [
   "mcp",
   "tool-use",
   "agent"
  ],
  "people": [],
  "one_liner": "One open standard for plugging tools and data into any AI app, like USB-C for AI.",
  "summary": "Anthropic open-sourced MCP to replace one-off integrations. OpenAI, Google and Microsoft adopted it in 2025, and it became the common way to connect agents to tools.",
  "sources": [
   {
    "title": "Introducing the Model Context Protocol",
    "url": "https://www.anthropic.com/news/model-context-protocol",
    "kind": "primary"
   }
  ],
  "era": "agents",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Introducing the Model Context Protocol \\ Anthropic",
    "date 2024-11-25 found on page"
   ]
  }
 },
 {
  "id": "deepseek-v3-2024",
  "date": "2024-12-26",
  "title": "DeepSeek-V3 trains a frontier-class model cheaply",
  "type": "model",
  "labs": [
   "deepseek"
  ],
  "layer": "models",
  "concepts": [
   "mixture-of-experts",
   "open-weight-model"
  ],
  "people": [],
  "one_liner": "A 671B mixture-of-experts model with a reported training cost of about 5.6 million dollars.",
  "summary": "DeepSeek released V3 with open weights. Its efficiency claims (FP8 training, multi-token prediction) surprised the industry.",
  "sources": [
   {
    "title": "DeepSeek-V3 on Hugging Face",
    "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3",
    "kind": "primary"
   },
   {
    "title": "DeepSeek-V3 Technical Report",
    "url": "https://arxiv.org/abs/2412.19437",
    "kind": "paper"
   }
  ],
  "era": "agents",
  "precision": "day",
  "check": {
   "status": "link-ok",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): deepseek-ai/DeepSeek-V3 · Hugging Face",
    "arXiv: DeepSeek-V3 Technical Report (v1 2024-12-27)",
    "paper posted 2024-12-27, after the 2024-12-26 launch: supporting source"
   ]
  }
 },
 {
  "id": "deepseek-r1-2025",
  "date": "2025-01-20",
  "title": "DeepSeek-R1: open reasoning model",
  "type": "model",
  "labs": [
   "deepseek"
  ],
  "layer": "models",
  "concepts": [
   "reasoning-model",
   "open-weight-model",
   "distillation"
  ],
  "people": [],
  "one_liner": "Reasoning close to o1, with open weights and an MIT licence.",
  "summary": "DeepSeek showed that reinforcement learning alone can produce strong reasoning, and distilled it into small models. The release briefly knocked NVIDIA's market value by hundreds of billions of dollars.",
  "sources": [
   {
    "title": "DeepSeek-R1 release notes",
    "url": "https://api-docs.deepseek.com/news/news250120",
    "kind": "primary"
   },
   {
    "title": "DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning",
    "url": "https://arxiv.org/abs/2501.12948",
    "kind": "paper"
   }
  ],
  "era": "agents",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): DeepSeek-R1 Release | DeepSeek API Docs",
    "date 2025-01-20 found on page",
    "arXiv: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning (v1 2025-01-22)",
    "paper posted 2025-01-22, after the 2025-01-20 launch: supporting source"
   ]
  }
 },
 {
  "id": "vibe-coding-2025",
  "date": "2025-02-02",
  "title": "Karpathy coins 'vibe coding'",
  "type": "person",
  "labs": [
   "other"
  ],
  "layer": "building",
  "concepts": [
   "vibe-coding",
   "coding-assistant"
  ],
  "people": [
   "andrej-karpathy"
  ],
  "one_liner": "Building software by describing what you want and accepting what the AI writes.",
  "summary": "In a post on X, Andrej Karpathy described a new way of coding where you 'fully give in to the vibes' and let AI write the code. The term spread fast and was later picked by dictionaries.",
  "sources": [
   {
    "title": "Andrej Karpathy on X",
    "url": "https://x.com/karpathy/status/1886192184808149383",
    "kind": "primary"
   }
  ],
  "era": "agents",
  "precision": "day",
  "check": {
   "status": "unverifiable",
   "checked_on": "2026-10-08",
   "notes": [
    "unverifiable host (X has no free API): https://x.com/karpathy/status/1886192184808149383"
   ]
  }
 },
 {
  "id": "claude-code-2025",
  "date": "2025-02-24",
  "title": "Claude 3.7 Sonnet and Claude Code",
  "type": "product",
  "labs": [
   "anthropic"
  ],
  "layer": "building",
  "concepts": [
   "coding-assistant",
   "agent",
   "reasoning-model"
  ],
  "people": [],
  "one_liner": "A hybrid model with optional extended thinking, plus an agent that codes in your terminal.",
  "summary": "Anthropic released its first hybrid reasoning model and a research preview of Claude Code, a command-line coding agent.",
  "sources": [
   {
    "title": "Claude 3.7 Sonnet and Claude Code",
    "url": "https://www.anthropic.com/news/claude-3-7-sonnet",
    "kind": "primary"
   }
  ],
  "era": "agents",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Claude 3.7 Sonnet and Claude Code \\ Anthropic",
    "date 2025-02-24 found on page"
   ]
  }
 },
 {
  "id": "gemini25-2025",
  "date": "2025-03-25",
  "title": "Gemini 2.5: Google's thinking model",
  "type": "model",
  "labs": [
   "google"
  ],
  "layer": "models",
  "concepts": [
   "reasoning-model"
  ],
  "people": [
   "demis-hassabis"
  ],
  "one_liner": "Gemini 2.5 Pro reasons before answering and leads many leaderboards.",
  "summary": "Google introduced Gemini 2.5 as a thinking model family with a 1 million token context window.",
  "sources": [
   {
    "title": "Gemini 2.5: Our most intelligent AI model",
    "url": "https://blog.google/technology/google-deepmind/gemini-model-thinking-updates-march-2025/",
    "kind": "primary"
   }
  ],
  "era": "agents",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Gemini 2.5: Our newest Gemini model with thinking",
    "date 2025-03-25 found on page"
   ]
  }
 },
 {
  "id": "a2a-2025",
  "date": "2025-04-09",
  "title": "Agent2Agent (A2A) protocol",
  "type": "product",
  "labs": [
   "google"
  ],
  "layer": "building",
  "concepts": [
   "a2a",
   "agent"
  ],
  "people": [],
  "one_liner": "An open protocol for agents from different vendors to talk to each other.",
  "summary": "Google announced A2A with more than 50 partners. MCP connects an agent to tools; A2A connects agents to other agents.",
  "sources": [
   {
    "title": "Announcing the Agent2Agent Protocol (A2A)",
    "url": "https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/",
    "kind": "primary"
   }
  ],
  "era": "agents",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Announcing the Agent2Agent Protocol (A2A) - Google Developers Blog",
    "date 2025-04-09 found on page"
   ]
  }
 },
 {
  "id": "qwen3-2025",
  "date": "2025-04-29",
  "title": "Qwen3 open models",
  "type": "model",
  "labs": [
   "qwen"
  ],
  "layer": "models",
  "concepts": [
   "open-weight-model",
   "reasoning-model",
   "mixture-of-experts"
  ],
  "people": [],
  "one_liner": "Alibaba releases open models from 0.6B to 235B with switchable thinking.",
  "summary": "Qwen3 included dense and mixture-of-experts models under Apache 2.0, and became a common base for local and fine-tuned models.",
  "sources": [
   {
    "title": "Qwen3: Think Deeper, Act Faster",
    "url": "https://qwenlm.github.io/blog/qwen3/",
    "kind": "primary"
   }
  ],
  "era": "agents",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Qwen3: Think Deeper, Act Faster | Qwen",
    "date 2025-04-29 found on page"
   ]
  }
 },
 {
  "id": "build-2025",
  "date": "2025-05-19",
  "title": "Microsoft Build 2025: the open agentic web",
  "type": "event",
  "labs": [
   "microsoft"
  ],
  "layer": "building",
  "concepts": [
   "agent",
   "mcp"
  ],
  "people": [],
  "one_liner": "Microsoft backs MCP across Windows, GitHub and Azure AI Foundry.",
  "summary": "Microsoft announced broad MCP support, the GitHub Copilot coding agent and NLWeb, framing the year around agents that act on behalf of users.",
  "sources": [
   {
    "title": "Microsoft Build 2025: The age of AI agents and building the open agentic web",
    "url": "https://blogs.microsoft.com/blog/2025/05/19/microsoft-build-2025-the-age-of-ai-agents-and-building-the-open-agentic-web/",
    "kind": "primary"
   }
  ],
  "era": "agents",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Microsoft Build 2025: The age of AI agents and building the open agentic web - The Official Microsoft Blog",
    "date 2025-05-19 found on page"
   ]
  }
 },
 {
  "id": "claude4-2025",
  "date": "2025-05-22",
  "title": "Claude 4: Opus 4 and Sonnet 4",
  "type": "model",
  "labs": [
   "anthropic"
  ],
  "layer": "models",
  "concepts": [
   "agent",
   "coding-assistant"
  ],
  "people": [],
  "one_liner": "Models built for long-running coding and agent tasks.",
  "summary": "Anthropic released Claude Opus 4 and Sonnet 4, focused on sustained agentic work, and made Claude Code generally available.",
  "sources": [
   {
    "title": "Introducing Claude 4",
    "url": "https://www.anthropic.com/news/claude-4",
    "kind": "primary"
   }
  ],
  "era": "agents",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Introducing Claude 4 \\ Anthropic",
    "date 2025-05-22 found on page"
   ]
  }
 },
 {
  "id": "gpt-oss-2025",
  "date": "2025-08-05",
  "title": "OpenAI releases open-weight gpt-oss",
  "type": "model",
  "labs": [
   "openai"
  ],
  "layer": "models",
  "concepts": [
   "open-weight-model",
   "reasoning-model"
  ],
  "people": [],
  "one_liner": "OpenAI's first open-weight language models since GPT-2.",
  "summary": "gpt-oss-120b and gpt-oss-20b were released under Apache 2.0 as reasoning models that run on a single GPU or a high-end laptop.",
  "sources": [
   {
    "title": "Introducing gpt-oss",
    "url": "https://openai.com/index/introducing-gpt-oss/",
    "kind": "primary"
   }
  ],
  "era": "agents",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Introducing gpt-oss | OpenAI",
    "date 2025-08-05 found on page"
   ]
  }
 },
 {
  "id": "gpt5-2025",
  "date": "2025-08-07",
  "title": "GPT-5",
  "type": "model",
  "labs": [
   "openai"
  ],
  "layer": "models",
  "concepts": [
   "reasoning-model",
   "llm"
  ],
  "people": [
   "sam-altman"
  ],
  "one_liner": "One system that decides when to answer fast and when to think longer.",
  "summary": "OpenAI released GPT-5 to all ChatGPT users, combining a fast model and a reasoning model with a router.",
  "sources": [
   {
    "title": "Introducing GPT-5",
    "url": "https://openai.com/index/introducing-gpt-5/",
    "kind": "primary"
   }
  ],
  "era": "agents",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Introducing GPT-5 | OpenAI",
    "date 2025-08-07 found on page"
   ]
  }
 },
 {
  "id": "steinberger-openai-2026",
  "date": "2026-02-15",
  "title": "OpenClaw creator Peter Steinberger joins OpenAI",
  "type": "person",
  "labs": [
   "openai"
  ],
  "layer": "building",
  "concepts": [
   "agent",
   "personal-agent"
  ],
  "people": [
   "peter-steinberger",
   "sam-altman"
  ],
  "one_liner": "The developer behind the viral OpenClaw personal agent moves to OpenAI.",
  "summary": "Sam Altman announced that Steinberger would join OpenAI to work on personal agents, after OpenClaw became one of the fastest growing open-source agent projects.",
  "sources": [
   {
    "title": "Fortune: Who is Peter Steinberger",
    "url": "https://fortune.com/2026/02/19/openclaw-who-is-peter-steinberger-openai-sam-altman-anthropic-moltbook",
    "kind": "press"
   },
   {
    "title": "Gigazine: Steinberger joins OpenAI",
    "url": "https://gigazine.net/gsc_news/en/20260216-openclaw-steinberger-joined-openai",
    "kind": "press"
   }
  ],
  "era": "agents",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Who is OpenClaw creator Peter Steinberger? The developer who caught the attention of OpenAI | Fortune",
    "page loads (200): OpenClaw creator Peter Steinberger joins OpenAI, potentially accelerating OpenAI's development of autonomous AI agents - GIGAZINE",
    "date 2026-02-15 found on page"
   ]
  }
 },
 {
  "id": "karpathy-anthropic-2026",
  "date": "2026-05-19",
  "title": "Andrej Karpathy joins Anthropic",
  "type": "person",
  "labs": [
   "anthropic"
  ],
  "layer": "training",
  "concepts": [
   "pretraining"
  ],
  "people": [
   "andrej-karpathy"
  ],
  "one_liner": "OpenAI co-founder and former Tesla AI director joins Anthropic's pretraining team.",
  "summary": "Reports in May 2026 said Karpathy joined Anthropic to work on pretraining for Claude.",
  "sources": [
   {
    "title": "exchange4media: Andrej Karpathy joins Anthropic",
    "url": "https://www.exchange4media.com/people-movement-news/andrej-karpathy-joins-anthropic-154782.html",
    "kind": "press"
   }
  ],
  "era": "agents",
  "precision": "day",
  "check": {
   "status": "verified",
   "checked_on": "2026-10-08",
   "notes": [
    "page loads (200): Former Tesla AI Director Andrej Karpathy joins Anthropic",
    "date 2026-05-19 found on page"
   ]
  }
 }
]