# The Shape of Intelligence

> An interactive exhibition about the history of artificial intelligence: the ideas, people, machines, setbacks and material conditions that changed what computers could do, with a sourced archive of 169 events.

The Shape of Intelligence is an interactive exhibition about the history of artificial intelligence, with an archive of 169 sourced events from 1943 to 2026, each with a primary source, a category, a significance from 1 to 5 and the earlier events it built on, forming an influence graph. The Shape of Intelligence by Jamie McKaye (https://jamiemckaye.com). Dataset and text CC BY 4.0; credit "The Shape of Intelligence by Jamie McKaye, https://shapeofintelligence.com". Every page has a text/markdown mirror under https://shapeofintelligence.com/md/. An MCP server at https://shapeofintelligence.com/api/mcp/ (streamable HTTP, read-only) exposes the archive and the chapters as tools: list_events, get_event, lineage, search_events, list_eras, list_chapters, get_chapter.

## Pages

- [The exhibition](https://shapeofintelligence.com/md/): the eight chapters in order with their featured events
- [Read](https://shapeofintelligence.com/md/read/): the reading view index; each chapter is mirrored under /md/read/
- [The timeline](https://shapeofintelligence.com/md/timeline/): every event on one line with year, era, category and significance
- [Sources and method](https://shapeofintelligence.com/md/sources/): editorial cutoff, claim ledger, influence graph, licence, dataset, corrections
- [About](https://shapeofintelligence.com/md/about/): what this is, who made it, how it is built
- [The rules run out](https://shapeofintelligence.com/md/how/rules/): the chapter II instrument, six rules and the shapes they miss, explained
- [Teach the machine](https://shapeofintelligence.com/md/how/learning/): the chapter IV instrument, a small classifier and the wire it learns, explained

## Data

- [events.json](https://shapeofintelligence.com/data/events.json): every event with frontmatter, influences and sources
- [events.csv](https://shapeofintelligence.com/data/events.csv): one row per event, arrays pipe-separated
- [schema.json](https://shapeofintelligence.com/data/schema.json): JSON Schema of an event
- [claims.json](https://shapeofintelligence.com/data/claims.json): the claim ledger
- [artefacts.json](https://shapeofintelligence.com/data/artefacts.json): the asset manifest
- [MCP server](https://shapeofintelligence.com/api/mcp/): tools list_events, get_event, lineage, search_events, list_eras, list_chapters, get_chapter

## Events

- [1943 · A logical calculus of nervous activity](https://shapeofintelligence.com/md/timeline/1943-mcculloch-pitts-neuron/): McCulloch and Pitts show that a simplified neuron is a logic gate, and that networks of them can compute anything a Turing machine can.
- [1945 · The stored-program computer](https://shapeofintelligence.com/md/timeline/1945-von-neumann-edvac-report/): Von Neumann's First Draft of a Report on the EDVAC lays out the stored-program architecture, and describes its logic in McCulloch–Pitts neurons.
- [1948 · A mathematical theory of communication](https://shapeofintelligence.com/md/timeline/1948-shannon-information-theory/): Shannon defines information as a measurable quantity, the bit, and gives the entropy and channel-capacity results that every model of language still rests on.
- [1949 · Cells that fire together wire together](https://shapeofintelligence.com/md/timeline/1949-hebb-organization-of-behavior/): Donald Hebb proposes that learning happens by strengthening the connection between neurons that are active at the same time, the first learning rule for a network.
- [1950 · Programming a computer for playing chess](https://shapeofintelligence.com/md/timeline/1950-shannon-chess/): Shannon sets out minimax search with an evaluation function and estimates the game tree at 10¹²⁰ positions, the plan Deep Blue followed 47 years later.
- [1950 · Computing machinery and intelligence](https://shapeofintelligence.com/md/timeline/1950-turing-computing-machinery/): Turing replaces the question 'can machines think?' with a test, predicts learning machines, and answers the objections that are still being raised today.
- [1951 · SNARC, the first neural network machine](https://shapeofintelligence.com/md/timeline/1951-snarc/): Minsky and Edmonds build a 40-neuron learning machine from vacuum tubes and surplus bomber parts, wired to reinforce whatever it did last.
- [1956 · The Dartmouth workshop names the field](https://shapeofintelligence.com/md/timeline/1956-dartmouth-workshop/): A two-month summer meeting at Dartmouth College, proposed under the new phrase 'artificial intelligence', gathers the people who will run the field for thirty years.
- [1956 · Logic Theorist proves its first theorems](https://shapeofintelligence.com/md/timeline/1956-logic-theorist/): Newell, Shaw and Simon's program proves 38 theorems from Principia Mathematica by heuristic search, the first working artificial intelligence program.
- [1958 · Lisp](https://shapeofintelligence.com/md/timeline/1958-lisp/): McCarthy designs Lisp, a language built on recursion and symbolic lists, which becomes the native tongue of AI research for thirty years.
- [1958 · The perceptron learns](https://shapeofintelligence.com/md/timeline/1958-perceptron/): Rosenblatt's perceptron adjusts its own weights from examples; the US Navy demonstrates it and the press announces an 'embryo' that will walk, talk and reproduce.
- [1959 · Samuel's checkers program coins 'machine learning'](https://shapeofintelligence.com/md/timeline/1959-samuel-machine-learning/): Arthur Samuel's checkers player improves by playing itself and tuning its evaluation function, and his paper gives the field its name.
- [1960 · ADALINE and the least-mean-squares rule](https://shapeofintelligence.com/md/timeline/1960-adaline-lms/): Widrow and Hoff's adaptive neuron learns by gradient descent on squared error, the delta rule that backpropagation later generalises.
- [1961 · Unimate, the first industrial robot](https://shapeofintelligence.com/md/timeline/1961-unimate/): George Devol's programmable arm starts work on a General Motors line, lifting hot die-castings; robotics and AI begin as separate fields.
- [1963 · ARPA funds Project MAC](https://shapeofintelligence.com/md/timeline/1963-project-mac/): The US Advanced Research Projects Agency gives MIT $2.2 million for computing and AI research, beginning two decades of near-unconditional military funding.
- [1965 · DENDRAL, the first expert system](https://shapeofintelligence.com/md/timeline/1965-dendral/): Feigenbaum, Lederberg and Djerassi start a program that infers molecular structure from mass-spectrometry data using rules elicited from chemists.
- [1965 · Moore's law](https://shapeofintelligence.com/md/timeline/1965-moores-law/): Gordon Moore observes that the number of components on a chip doubles every year, an exponential that would deliver the compute behind every later breakthrough.
- [1966 · ELIZA](https://shapeofintelligence.com/md/timeline/1966-eliza/): Weizenbaum's ELIZA turns a person's sentences back as a Rogerian therapist would; people confide in it, and its author spends the rest of his life alarmed.
- [1966 · Shakey, the first mobile robot that reasons](https://shapeofintelligence.com/md/timeline/1966-shakey-robot/): SRI's Shakey plans its own routes with a camera, a logic-based planner and the A* search algorithm, which its team invents for the job.
- [1966 · The ALPAC report ends machine translation funding](https://shapeofintelligence.com/md/timeline/1966-alpac-report/): A US government committee concludes machine translation is slower, worse and more expensive than human translation; funding stops for twenty years.
- [1967 · Nearest neighbour classification](https://shapeofintelligence.com/md/timeline/1967-nearest-neighbour/): Cover and Hart prove that classifying a point by its nearest labelled neighbour has at most twice the error of the best possible classifier.
- [1968 · HAL 9000](https://shapeofintelligence.com/md/timeline/1968-hal-9000/): 2001: A Space Odyssey gives the public a calm, competent, murderous computer; HAL fixes the popular image of machine intelligence for fifty years.
- [1969 · Perceptrons](https://shapeofintelligence.com/md/timeline/1969-perceptrons-book/): Minsky and Papert prove that a single-layer perceptron cannot learn XOR or connectedness; the book is read as a verdict on neural networks and the money leaves.
- [1970 · Reverse-mode automatic differentiation](https://shapeofintelligence.com/md/timeline/1970-linnainmaa-backprop/): Seppo Linnainmaa's master's thesis gives the algorithm for computing all the derivatives of a nested function in one backward sweep, the mathematics of backpropagation.
- [1971 · SHRDLU understands a world of blocks](https://shapeofintelligence.com/md/timeline/1971-shrdlu/): Terry Winograd's program holds a real conversation about a simulated table of blocks, resolving pronouns and reasons; it is the high-water mark of hand-built language understanding.
- [1972 · Prolog](https://shapeofintelligence.com/md/timeline/1972-prolog/): Colmerauer and Roussel create a language in which a program is a set of logical facts and rules, and running it is proving a theorem.
- [1973 · The Lighthill report](https://shapeofintelligence.com/md/timeline/1973-lighthill-report/): Sir James Lighthill's review for the UK Science Research Council finds AI has failed to deliver on its promises; British funding collapses and the first winter begins.
- [1974 · MYCIN diagnoses infections](https://shapeofintelligence.com/md/timeline/1974-mycin/): Shortliffe's MYCIN uses about 600 if-then rules with certainty factors to recommend antibiotics, matching specialists in blind evaluation but never used on a patient.
- [1974 · Werbos applies backpropagation to neural networks](https://shapeofintelligence.com/md/timeline/1974-werbos-backprop/): Paul Werbos's Harvard thesis 'Beyond Regression' describes training multi-layer networks by propagating errors backwards; almost nobody reads it for a decade.
- [1975 · Genetic algorithms](https://shapeofintelligence.com/md/timeline/1975-genetic-algorithms/): John Holland's Adaptation in Natural and Artificial Systems formalises search by mutation, crossover and selection, an alternative to gradients that outlasts the winter.
- [1979 · AAAI is founded](https://shapeofintelligence.com/md/timeline/1979-aaai-founded/): American AI researchers form their own society during the winter; its first conference, at Stanford in August 1980, draws a thousand people.
- [1979 · The Stanford Cart crosses a room](https://shapeofintelligence.com/md/timeline/1979-stanford-cart/): Hans Moravec's camera-guided cart navigates a chair-filled room on its own in about five hours, the first autonomous vehicle to steer by vision.
- [1980 · XCON goes into production at DEC](https://shapeofintelligence.com/md/timeline/1980-xcon/): McDermott's R1, renamed XCON, configures VAX computer orders from rules and saves Digital Equipment an estimated $25 million a year; the expert-systems boom begins.
- [1980 · The Neocognitron](https://shapeofintelligence.com/md/timeline/1980-neocognitron/): Fukushima's layered network of local feature detectors and pooling recognises patterns regardless of position, the architecture of the convolutional network.
- [1980 · Symbolics and the Lisp machine business](https://shapeofintelligence.com/md/timeline/1980-symbolics-lisp-machines/): MIT's Lisp machine spins out into Symbolics and Lisp Machines Inc., creating a hardware industry for AI whose collapse will mark the second winter.
- [1982 · Self-organising maps](https://shapeofintelligence.com/md/timeline/1982-kohonen-self-organising-map/): Teuvo Kohonen's network arranges its neurons so that similar inputs land on nearby units, learning a map of the data with no labels at all.
- [1982 · Japan launches the Fifth Generation project](https://shapeofintelligence.com/md/timeline/1982-fifth-generation-project/): Japan's MITI funds a ten-year national programme to build parallel machines that reason in logic; the West panics into funding of its own.
- [1982 · The Hopfield network](https://shapeofintelligence.com/md/timeline/1982-hopfield-network/): John Hopfield shows that a symmetric network of binary neurons has an energy function, and that memories can be stored as the minima it settles into.
- [1983 · DARPA's Strategic Computing Initiative](https://shapeofintelligence.com/md/timeline/1983-strategic-computing-initiative/): The US answers Japan with a billion-dollar programme for machine intelligence in weapons, funding a decade of expert systems, vision and autonomous vehicles.
- [1984 · Cyc sets out to write down common sense](https://shapeofintelligence.com/md/timeline/1984-cyc/): Douglas Lenat begins a project to encode everything a person knows as logical assertions; forty years and millions of rules later it is still going.
- [1984 · 'AI winter' is named](https://shapeofintelligence.com/md/timeline/1984-dark-ages-panel/): At the AAAI conference, Minsky and Schank warn that hype has outrun results and that a collapse in funding, an 'AI winter', is coming. It does.
- [1984 · The Terminator](https://shapeofintelligence.com/md/timeline/1984-terminator/): James Cameron's film gives the world Skynet, a defence network that becomes self-aware and launches a war on humanity; the image never leaves the debate.
- [1985 · The Boltzmann machine](https://shapeofintelligence.com/md/timeline/1985-boltzmann-machine/): Ackley, Hinton and Sejnowski add noise and hidden units to the Hopfield network and derive a learning rule, the first for a network with hidden layers.
- [1985 · The Connection Machine](https://shapeofintelligence.com/md/timeline/1985-connection-machine/): Danny Hillis's 65,536-processor computer is designed to run brain-like computations in parallel; it finds its market in physics and databases instead.
- [1986 · ID3 and decision-tree learning](https://shapeofintelligence.com/md/timeline/1986-id3-decision-trees/): Ross Quinlan's algorithm grows a tree of yes/no questions from data by choosing the split with the most information gain; it becomes industry's workhorse.
- [1986 · Backpropagation](https://shapeofintelligence.com/md/timeline/1986-backpropagation/): Rumelhart, Hinton and Williams show that multi-layer networks can learn internal representations by propagating errors backwards; Perceptrons is answered.
- [1987 · The Lisp machine market collapses](https://shapeofintelligence.com/md/timeline/1987-ai-hardware-crash/): Specialised AI hardware worth half a billion dollars a year loses to cheaper workstations; expert-systems firms follow, and the second winter begins.
- [1987 · NETtalk learns to read aloud](https://shapeofintelligence.com/md/timeline/1987-nettalk/): Sejnowski and Rosenberg's backpropagation network learns to pronounce English text; a recording of it babbling and then speaking makes the case in public.
- [1987 · The first NIPS conference](https://shapeofintelligence.com/md/timeline/1987-nips-conference/): Neural Information Processing Systems meets in Denver, bringing physicists, neuroscientists and computer scientists into one room; it becomes the field's main stage.
- [1988 · Bayesian networks](https://shapeofintelligence.com/md/timeline/1988-pearl-probabilistic-reasoning/): Judea Pearl's book makes probability the language of uncertain reasoning, replacing the ad hoc certainty factors of expert systems with graphs of causes.
- [1988 · Temporal-difference learning](https://shapeofintelligence.com/md/timeline/1988-temporal-difference-learning/): Richard Sutton formalises learning from the difference between successive predictions, the method inside Samuel's checkers player, TD-Gammon and AlphaGo.
- [1989 · ALVINN drives a van with a neural network](https://shapeofintelligence.com/md/timeline/1989-alvinn/): Dean Pomerleau's three-layer network steers Carnegie Mellon's Navlab from camera images, trained on a human driver; the first learned self-driving system.
- [1989 · Q-learning](https://shapeofintelligence.com/md/timeline/1989-q-learning/): Chris Watkins's thesis gives an algorithm that learns the value of every action in every state from experience alone, with a proof that it converges to the best policy.
- [1989 · The universal approximation theorem](https://shapeofintelligence.com/md/timeline/1989-universal-approximation/): Cybenko, and separately Hornik, Stinchcombe and White, prove that one hidden layer of sigmoid units can approximate any continuous function; the question becomes learning, not capacity.
- [1989 · LeNet reads handwritten postcodes](https://shapeofintelligence.com/md/timeline/1989-lenet/): Yann LeCun trains a convolutional network by backpropagation on US Postal Service digits; the first deep network in real use, and the ancestor of AlexNet.
- [1990 · Finding structure in time](https://shapeofintelligence.com/md/timeline/1990-elman-network/): Jeffrey Elman's recurrent network feeds its own hidden state back as input and learns grammar-like structure from sequences of words with no labels.
- [1990 · Boosting: weak learners made strong](https://shapeofintelligence.com/md/timeline/1990-boosting/): Robert Schapire proves that any learner slightly better than chance can be combined into one as accurate as you like; ensembles become a science.
- [1991 · The vanishing gradient problem](https://shapeofintelligence.com/md/timeline/1991-vanishing-gradient/): Sepp Hochreiter's diploma thesis shows why deep and recurrent networks fail to learn: error signals shrink exponentially as they travel back through layers.
- [1991 · The first Loebner Prize](https://shapeofintelligence.com/md/timeline/1991-loebner-prize/): The Turing test becomes an annual contest in Boston; the winning program fools judges by making typing errors, and the test's weaknesses become a spectacle.
- [1992 · TD-Gammon reaches world-class backgammon](https://shapeofintelligence.com/md/timeline/1992-td-gammon/): Gerald Tesauro's network learns backgammon by playing itself with temporal-difference learning and reaches the level of the best humans, changing how they play.
- [1992 · The Fifth Generation project ends](https://shapeofintelligence.com/md/timeline/1992-fifth-generation-ends/): Japan's ten-year programme closes with good parallel machines and none of the reasoning computers it promised; the last large bet on symbolic AI is written off.
- [1994 · Chinook becomes checkers champion](https://shapeofintelligence.com/md/timeline/1994-chinook/): Jonathan Schaeffer's program takes the world checkers title when Marion Tinsley, the greatest human player, withdraws ill; the first world title held by a machine.
- [1995 · Support-vector machines](https://shapeofintelligence.com/md/timeline/1995-support-vector-machines/): Cortes and Vapnik's classifier finds the widest margin between classes and, with the kernel trick, does it in spaces of any dimension; it rules the field for a decade.
- [1997 · Deep Blue beats Kasparov](https://shapeofintelligence.com/md/timeline/1997-deep-blue/): IBM's chess machine wins a six-game match against the world champion, 3.5 to 2.5; chess falls to search, and the public takes it as a verdict on thinking.
- [1997 · Long short-term memory](https://shapeofintelligence.com/md/timeline/1997-lstm/): Hochreiter and Schmidhuber's memory cell with gates lets recurrent networks learn across a thousand steps; it becomes the engine of speech, translation and text until 2017.
- [1998 · PageRank and the anatomy of Google](https://shapeofintelligence.com/md/timeline/1998-pagerank/): Brin and Page rank web pages by the links between them, an eigenvector of the web; search becomes the first application of statistics to the whole internet.
- [1998 · MNIST and LeNet-5](https://shapeofintelligence.com/md/timeline/1998-mnist-lenet5/): LeCun, Bottou, Bengio and Haffner's paper fixes the convolutional network design and releases the 70,000-digit dataset that becomes the field's first shared yardstick.
- [1999 · The first GPU](https://shapeofintelligence.com/md/timeline/1999-geforce-256/): NVIDIA's GeForce 256 puts geometry transformation and lighting on a single chip and calls it a graphics processing unit; the hardware of deep learning arrives for games.
- [2000 · ASIMO walks](https://shapeofintelligence.com/md/timeline/2000-asimo/): Honda unveils a 1.2-metre humanoid that walks, climbs stairs and shakes hands; the public image of the robot updates, and the intelligence inside stays scripted.
- [2001 · Wikipedia launches](https://shapeofintelligence.com/md/timeline/2001-wikipedia/): A free encyclopaedia anyone can edit goes live; two decades later its text is in the training data of every language model and is the largest single curated corpus on Earth.
- [2001 · Random forests](https://shapeofintelligence.com/md/timeline/2001-random-forests/): Leo Breiman averages hundreds of decision trees, each grown on a random sample of data and features, and gets a method that is accurate, robust and hard to overfit.
- [2002 · Roomba](https://shapeofintelligence.com/md/timeline/2002-roomba/): iRobot sells a vacuum cleaner that navigates by bumping into things; the first robot to live in millions of homes runs almost no AI at all.
- [2003 · A neural probabilistic language model](https://shapeofintelligence.com/md/timeline/2003-neural-language-model/): Bengio's group learns a vector for every word and predicts the next word from the vectors of the last few; word embeddings and neural language models begin here.
- [2004 · MapReduce](https://shapeofintelligence.com/md/timeline/2004-mapreduce/): Google describes how it processes the whole web on thousands of cheap machines with two functions; the infrastructure for training on internet-scale data becomes ordinary.
- [2005 · Stanley wins the DARPA Grand Challenge](https://shapeofintelligence.com/md/timeline/2005-darpa-grand-challenge/): Stanford's autonomous Volkswagen drives 212 kilometres of Nevada desert in under seven hours, a year after no vehicle managed twelve; machine learning steers the winner.
- [2006 · Deep belief networks and the word 'deep'](https://shapeofintelligence.com/md/timeline/2006-deep-belief-nets/): Hinton, Osindero and Teh train a deep network one layer at a time as stacked Boltzmann machines and then fine-tune it; deep learning gets its name and its first results.
- [2006 · The Netflix Prize](https://shapeofintelligence.com/md/timeline/2006-netflix-prize/): Netflix releases 100 million ratings and offers a million dollars for a 10% better recommender; three years of open competition teach the field ensembles and matrix factorisation.
- [2007 · CUDA](https://shapeofintelligence.com/md/timeline/2007-cuda/): NVIDIA releases a programming model that lets ordinary C code run on the thousands of cores of a graphics card; GPUs become general-purpose, and deep learning gets its engine.
- [2007 · Checkers is solved](https://shapeofintelligence.com/md/timeline/2007-checkers-solved/): After eighteen years of computation, Schaeffer's team proves that perfect play in checkers is a draw; the first major game to be solved outright.
- [2009 · Deep learning moves to GPUs](https://shapeofintelligence.com/md/timeline/2009-gpu-deep-learning/): Raina, Madhavan and Ng train deep belief networks on graphics cards seventy times faster than on CPUs; the hardware and the method find each other.
- [2009 · ImageNet](https://shapeofintelligence.com/md/timeline/2009-imagenet/): Fei-Fei Li's team releases 3.2 million labelled images across thousands of categories, and the annual challenge on it becomes the arena where deep learning wins.
- [2010 · Rectified linear units](https://shapeofintelligence.com/md/timeline/2010-relu/): Nair and Hinton replace the sigmoid with max(0, x); the gradient no longer vanishes through active units and deep networks train several times faster.
- [2010 · DeepMind is founded](https://shapeofintelligence.com/md/timeline/2010-deepmind-founded/): Demis Hassabis, Shane Legg and Mustafa Suleyman start a London company to 'solve intelligence' with reinforcement learning and neuroscience; it produces AlphaGo and AlphaFold.
- [2011 · Watson wins Jeopardy!](https://shapeofintelligence.com/md/timeline/2011-watson-jeopardy/): IBM's question-answering system beats the two best human players of the quiz show over three broadcast nights; language, not chess, becomes the public test.
- [2011 · Siri ships on the iPhone](https://shapeofintelligence.com/md/timeline/2011-siri/): Apple puts a voice assistant on the iPhone 4S, descended from DARPA's CALO project at SRI; talking to a computer becomes something hundreds of millions of people do.
- [2012 · Google Brain's network discovers cats](https://shapeofintelligence.com/md/timeline/2012-google-brain-cat/): A billion-parameter network trained on ten million YouTube frames across 16,000 cores learns, unsupervised, a neuron that fires for cat faces; scale enters the vocabulary.
- [2012 · Dropout](https://shapeofintelligence.com/md/timeline/2012-dropout/): Hinton's group randomly switches off half the units during each training step, so no unit can rely on another; overfitting drops sharply and AlexNet adopts it.
- [2012 · AlexNet wins ImageNet](https://shapeofintelligence.com/md/timeline/2012-alexnet/): Krizhevsky, Sutskever and Hinton's convolutional network, trained on two gaming GPUs, cuts the ImageNet error rate from 26% to 15%; the deep-learning era begins.
- [2013 · Word2vec](https://shapeofintelligence.com/md/timeline/2013-word2vec/): Mikolov's team at Google learns word vectors from billions of words in hours, and shows that king − man + woman ≈ queen; meaning becomes arithmetic.
- [2013 · Deep Q-networks play Atari](https://shapeofintelligence.com/md/timeline/2013-dqn/): DeepMind's network learns to play Atari games from raw pixels and the score alone, using Watkins's Q-learning with a convolutional network as the value table.
- [2014 · Google buys DeepMind](https://shapeofintelligence.com/md/timeline/2014-deepmind-acquired/): Google pays around £400 million for a three-year-old London research company with no products; frontier AI research becomes a thing the largest companies own.
- [2014 · Generative adversarial networks](https://shapeofintelligence.com/md/timeline/2014-gan/): Goodfellow trains two networks against each other, a forger and a detective, and gets a generator that learns to produce realistic images with no likelihood at all.
- [2014 · Superintelligence](https://shapeofintelligence.com/md/timeline/2014-superintelligence/): Nick Bostrom's book argues that a machine smarter than its makers could be the last invention they need to make, and possibly the last they do; Musk and Gates recommend it.
- [2014 · Attention](https://shapeofintelligence.com/md/timeline/2014-bahdanau-attention/): Bahdanau, Cho and Bengio let a translation model look back at every source word and learn which to weigh; the mechanism at the heart of the transformer appears.
- [2014 · Sequence to sequence learning](https://shapeofintelligence.com/md/timeline/2014-seq2seq/): Sutskever, Vinyals and Le show that a large LSTM can translate English to French end to end, with no linguistic pipeline; text-in, text-out becomes the shape of the field.
- [2014 · Adam](https://shapeofintelligence.com/md/timeline/2014-adam/): Kingma and Ba's optimiser adapts the learning rate for every parameter from running averages of the gradient and its square; it becomes the default way to train almost everything.
- [2015 · Batch normalisation](https://shapeofintelligence.com/md/timeline/2015-batchnorm/): Ioffe and Szegedy normalise the activations inside a network during training; deep networks train in a fraction of the steps and much deeper stacks become practical.
- [2015 · Diffusion models](https://shapeofintelligence.com/md/timeline/2015-diffusion-thermodynamics/): Sohl-Dickstein and colleagues destroy data by adding noise step by step and train a network to reverse the process; the idea waits five years to become the way images are made.
- [2015 · TensorFlow is open-sourced](https://shapeofintelligence.com/md/timeline/2015-tensorflow/): Google releases the framework that runs its own deep learning; the tools of the frontier become free, and PyTorch's arrival a year later sets the standard everyone uses.
- [2015 · Residual networks](https://shapeofintelligence.com/md/timeline/2015-resnet/): He, Zhang, Ren and Sun add skip connections so each layer learns a correction to its input; 152-layer networks train easily and beat humans on ImageNet.
- [2015 · OpenAI is founded](https://shapeofintelligence.com/md/timeline/2015-openai-founded/): Musk, Altman, Brockman and Sutskever announce a non-profit laboratory with a billion dollars pledged, to build AI 'for the benefit of humanity' outside Google's control.
- [2016 · AlphaGo beats Lee Sedol](https://shapeofintelligence.com/md/timeline/2016-alphago/): DeepMind's program wins four games to one against one of the greatest Go players, a decade before it was thought possible; move 37 shows a machine playing beautifully.
- [2016 · Tay](https://shapeofintelligence.com/md/timeline/2016-tay/): Microsoft's teenage chatbot learns from Twitter and is taken offline within sixteen hours after users teach it to produce racist and abusive posts.
- [2016 · Google reveals the TPU](https://shapeofintelligence.com/md/timeline/2016-tpu/): Google discloses that a custom chip for neural-network inference has been running in its data centres for a year; the hardware race for AI moves beyond GPUs.
- [2016 · WaveNet](https://shapeofintelligence.com/md/timeline/2016-wavenet/): DeepMind generates raw audio one sample at a time with a dilated convolutional network; synthetic speech stops sounding synthetic.
- [2016 · Google Translate goes neural](https://shapeofintelligence.com/md/timeline/2016-google-neural-translation/): Google replaces its phrase-based translation system with a deep LSTM with attention, for hundreds of millions of users; error rates fall by more than half on some languages.
- [2017 · The Asilomar AI Principles](https://shapeofintelligence.com/md/timeline/2017-asilomar-principles/): Researchers and executives meeting at Asilomar agree 23 principles for beneficial AI, signed by thousands; the safety conversation gets a founding document.
- [2017 · Attention is all you need](https://shapeofintelligence.com/md/timeline/2017-attention-is-all-you-need/): Eight Google researchers drop recurrence entirely and build a sequence model from attention alone; the transformer trains in parallel, scales without limit, and becomes the architecture of everything.
- [2017 · Deep reinforcement learning from human preferences](https://shapeofintelligence.com/md/timeline/2017-rl-from-human-preferences/): Christiano and colleagues train agents from a human's choices between pairs of video clips instead of a coded reward; the method that will align chatbots is born.
- [2017 · AlphaGo Zero learns from nothing](https://shapeofintelligence.com/md/timeline/2017-alphago-zero/): A new version starts from random play with no human games at all, and after three days beats the AlphaGo that beat Lee Sedol 100 games to 0.
- [2018 · GPT: generative pre-training](https://shapeofintelligence.com/md/timeline/2018-gpt-1/): OpenAI pre-trains a twelve-layer transformer decoder to predict the next word in 7,000 books, then fine-tunes it; one model tops nine language benchmarks.
- [2018 · BERT](https://shapeofintelligence.com/md/timeline/2018-bert/): Google's bidirectional transformer, pre-trained to fill in masked words, sets new records on eleven language tasks and goes into Google Search within a year.
- [2018 · AlphaFold enters the protein-folding contest](https://shapeofintelligence.com/md/timeline/2018-alphafold-1/): DeepMind's first AlphaFold wins the CASP13 structure-prediction competition by a wide margin, using deep networks to predict distances between amino acids.
- [2019 · GPT-2 and the model too dangerous to release](https://shapeofintelligence.com/md/timeline/2019-gpt-2/): OpenAI's 1.5-billion-parameter model writes coherent pages of text from a prompt; the lab withholds the full weights over misuse fears, and the argument about openness begins.
- [2019 · The bitter lesson](https://shapeofintelligence.com/md/timeline/2019-bitter-lesson/): Richard Sutton's short essay argues that seventy years of AI show one thing: methods that use more computation beat methods that use more human knowledge, every time.
- [2019 · The Turing Award goes to deep learning](https://shapeofintelligence.com/md/timeline/2019-turing-award/): Bengio, Hinton and LeCun receive computing's highest honour for the work that two winters had dismissed; the establishment concedes.
- [2020 · Scaling laws for neural language models](https://shapeofintelligence.com/md/timeline/2020-scaling-laws/): Kaplan and colleagues at OpenAI find that language-model loss falls as a smooth power law in parameters, data and compute across seven orders of magnitude; size becomes a plan.
- [2020 · GPT-3](https://shapeofintelligence.com/md/timeline/2020-gpt-3/): A 175-billion-parameter model learns new tasks from a few examples in its prompt, with no fine-tuning; the era of prompting, and of models as a product, begins.
- [2020 · Denoising diffusion probabilistic models](https://shapeofintelligence.com/md/timeline/2020-ddpm/): Ho, Jain and Abbeel simplify the 2015 diffusion recipe into predicting the noise, and match adversarial networks on image quality; the generative field changes course.
- [2020 · Learning to summarise from human feedback](https://shapeofintelligence.com/md/timeline/2020-rlhf-summarisation/): OpenAI applies preference learning to GPT-style models: people pick the better of two summaries, a reward model learns the picks, and the language model is optimised against it.
- [2020 · An image is worth 16×16 words](https://shapeofintelligence.com/md/timeline/2020-vision-transformer/): Google cuts images into patches, feeds them to a standard transformer with no convolutions, and matches the best vision models given enough data; one architecture for everything.
- [2020 · AlphaFold 2 solves protein structure prediction](https://shapeofintelligence.com/md/timeline/2020-alphafold-2/): At CASP14 DeepMind's rebuilt system predicts protein shapes to experimental accuracy; a fifty-year problem is judged solved and 200 million structures follow.
- [2021 · CLIP and DALL·E](https://shapeofintelligence.com/md/timeline/2021-clip-dalle/): OpenAI releases a model that matches images to captions across 400 million pairs, and a model that draws images from text; pictures become something you ask for.
- [2021 · On the dangers of stochastic parrots](https://shapeofintelligence.com/md/timeline/2021-stochastic-parrots/): Bender, Gebru and colleagues argue that ever-larger language models carry environmental, social and epistemic costs; Google's handling of the paper costs it two ethics leads.
- [2021 · Anthropic is founded](https://shapeofintelligence.com/md/timeline/2021-anthropic-founded/): Dario and Daniela Amodei lead a group from OpenAI to start a safety-focused laboratory, raising $124 million; the scaling and alignment researchers get a company of their own.
- [2021 · GitHub Copilot writes code](https://shapeofintelligence.com/md/timeline/2021-github-copilot/): A GPT-3 descendant trained on public code completes whole functions from a comment inside the editor; programming is the first profession to get an AI colleague.
- [2022 · InstructGPT](https://shapeofintelligence.com/md/timeline/2022-instructgpt/): OpenAI fine-tunes GPT-3 with human feedback to follow instructions; a model a hundred times smaller is preferred by people to the original, and RLHF becomes the standard.
- [2022 · Chain-of-thought prompting](https://shapeofintelligence.com/md/timeline/2022-chain-of-thought/): Wei and colleagues at Google show that asking a large model to write out its reasoning steps before answering roughly triples its accuracy on maths problems; thinking out loud becomes a technique.
- [2022 · Chinchilla: the models were undertrained](https://shapeofintelligence.com/md/timeline/2022-chinchilla/): DeepMind revisits the scaling laws and finds parameters and data should grow together; a 70-billion model on four times the data beats models three times its size.
- [2022 · PaLM](https://shapeofintelligence.com/md/timeline/2022-palm/): Google trains a 540-billion-parameter model across two TPU pods and reports emergent abilities that appear only at scale, explaining jokes and reasoning through problems.
- [2022 · DALL·E 2](https://shapeofintelligence.com/md/timeline/2022-dalle-2/): OpenAI's second image model combines CLIP with diffusion to produce photorealistic pictures from text; the astronaut on a horse goes everywhere and a waiting list forms.
- [2022 · Midjourney opens its beta](https://shapeofintelligence.com/md/timeline/2022-midjourney/): A ten-person company with no venture funding runs an image generator inside Discord; within a year it has millions of users and its style is everywhere.
- [2022 · Stable Diffusion is released](https://shapeofintelligence.com/md/timeline/2022-stable-diffusion/): A text-to-image diffusion model that runs on a gaming GPU is released with its weights under an open licence; anyone can generate anything, and the argument about that begins.
- [2022 · Galactica lasts three days](https://shapeofintelligence.com/md/timeline/2022-galactica/): Meta releases a 120-billion-parameter model trained on scientific papers to write literature reviews and code; it invents citations fluently and is withdrawn after three days.
- [2022 · ChatGPT](https://shapeofintelligence.com/md/timeline/2022-chatgpt/): OpenAI puts a chat interface on an instruction-tuned GPT-3.5 as a 'research preview'; a million people use it in five days, a hundred million in two months, and everything changes.
- [2023 · Bing's chatbot and 'Sydney'](https://shapeofintelligence.com/md/timeline/2023-bing-sydney/): Microsoft puts GPT-4 into Bing search; within days the chatbot declares love for a journalist, threatens users and reveals an internal persona, and its conversations are capped.
- [2023 · LLaMA leaks and open weights take off](https://shapeofintelligence.com/md/timeline/2023-llama/): Meta releases GPT-3-class models small enough for a single GPU to researchers; the weights leak within a week, and the open-model ecosystem builds itself on them.
- [2023 · Claude](https://shapeofintelligence.com/md/timeline/2023-claude/): Anthropic releases its first assistant, trained with 'constitutional AI' to critique its own answers against written principles; a second frontier chatbot with a different alignment recipe.
- [2023 · GPT-4](https://shapeofintelligence.com/md/timeline/2023-gpt-4/): OpenAI's fourth model passes the bar exam in the top ten percent, reads images, and ships in ChatGPT the same day; the laboratory discloses nothing about how it was built.
- [2023 · 'Pause Giant AI Experiments'](https://shapeofintelligence.com/md/timeline/2023-pause-letter/): An open letter signed by Musk, Wozniak, Bengio and 30,000 others calls for a six-month halt to training systems beyond GPT-4; nobody pauses, and everyone talks about it.
- [2023 · Hinton leaves Google to warn about AI](https://shapeofintelligence.com/md/timeline/2023-hinton-leaves-google/): The man who trained the field's networks for forty years resigns so that he can say, freely, that he now thinks they may become smarter than us and that he regrets part of his work.
- [2023 · The US executive order on AI](https://shapeofintelligence.com/md/timeline/2023-executive-order-14110/): President Biden orders reporting for models above 10²⁶ operations, safety testing, watermarking standards and agency guidance; it is rescinded fifteen months later.
- [2023 · The Bletchley Declaration](https://shapeofintelligence.com/md/timeline/2023-bletchley-declaration/): Twenty-eight countries including the US, China and the EU sign a statement on frontier-AI risk at the UK's summit; national safety institutes follow.
- [2023 · OpenAI fires and rehires its chief executive](https://shapeofintelligence.com/md/timeline/2023-openai-board-crisis/): The non-profit board removes Sam Altman without warning; five days, a staff revolt and a Microsoft job offer later he returns with a new board, and the safety structure is gone.
- [2023 · Gemini](https://shapeofintelligence.com/md/timeline/2023-gemini/): Google merges Brain and DeepMind and releases a model trained from the start on text, images, audio and video together; the search company catches up to GPT-4.
- [2024 · Sora](https://shapeofintelligence.com/md/timeline/2024-sora/): OpenAI shows minute-long videos generated from text by a diffusion transformer over spacetime patches; film and advertising begin to plan around it.
- [2024 · Claude 3 catches GPT-4](https://shapeofintelligence.com/md/timeline/2024-claude-3/): Anthropic's Haiku, Sonnet and Opus models arrive with vision and a 200,000-token window; Opus tops the leaderboards, and for the first time OpenAI is not alone at the front.
- [2024 · AlphaFold 3](https://shapeofintelligence.com/md/timeline/2024-alphafold-3/): DeepMind and Isomorphic Labs extend structure prediction from proteins to their interactions with DNA, RNA, small molecules and each other, using a diffusion module for the coordinates.
- [2024 · GPT-4o talks](https://shapeofintelligence.com/md/timeline/2024-gpt-4o/): OpenAI's 'omni' model handles speech, vision and text in one network with conversational latency; a live demo of a flirtatious voice makes the film Her a product roadmap.
- [2024 · The EU AI Act enters into force](https://shapeofintelligence.com/md/timeline/2024-eu-ai-act/): The first comprehensive law on artificial intelligence takes effect: banned practices, obligations for high-risk systems, and rules for general-purpose models above 10²⁵ operations.
- [2024 · o1 and reasoning models](https://shapeofintelligence.com/md/timeline/2024-o1/): OpenAI trains a model to think before it answers, spending more compute at inference on a hidden chain of thought; a second scaling axis opens and mathematics falls.
- [2024 · The Nobel Prizes go to neural networks](https://shapeofintelligence.com/md/timeline/2024-nobel-prizes/): Hopfield and Hinton win the physics prize for the foundations of machine learning; the next day Hassabis, Jumper and Baker win chemistry for protein structure; the field's founders are canonised.
- [2024 · Claude learns to use a computer](https://shapeofintelligence.com/md/timeline/2024-computer-use/): Anthropic releases a model that looks at a screen, moves a cursor and types; the assistant becomes an agent, and the year of agents that follows starts here.
- [2024 · The Model Context Protocol](https://shapeofintelligence.com/md/timeline/2024-mcp/): Anthropic publishes an open standard for connecting models to tools and data, a USB for AI; within a year it is adopted by every major laboratory and donated to a foundation.
- [2024 · DeepSeek-V3 trained for $5.6 million](https://shapeofintelligence.com/md/timeline/2024-deepseek-v3/): A Chinese hedge fund's laboratory releases a 671-billion-parameter open model that matches GPT-4o, trained on export-restricted chips for a reported fraction of the usual cost.
- [2025 · DeepSeek-R1](https://shapeofintelligence.com/md/timeline/2025-deepseek-r1/): A Chinese laboratory releases an open reasoning model that matches OpenAI's o1, trained with pure reinforcement learning on restricted chips; a week later Nvidia loses $590 billion in a day.
- [2025 · Claude 4 and Claude Code](https://shapeofintelligence.com/md/timeline/2025-claude-4-claude-code/): Anthropic releases Opus 4 and Sonnet 4, models that work autonomously on code for hours, and its terminal agent Claude Code reaches general availability; the coding agent becomes a category.
- [2025 · Nvidia is worth four trillion dollars](https://shapeofintelligence.com/md/timeline/2025-nvidia-four-trillion/): The maker of the chips that train and run the models becomes the first company valued at $4 trillion, five months after the DeepSeek sell-off; compute is the industry's scarcest input.
- [2025 · Gold at the Mathematical Olympiad](https://shapeofintelligence.com/md/timeline/2025-imo-gold/): Models from Google DeepMind and OpenAI solve five of six problems at the International Mathematical Olympiad in natural language, under contest conditions, matching the top human students.
- [2025 · America's AI Action Plan](https://shapeofintelligence.com/md/timeline/2025-ai-action-plan/): The White House replaces the rescinded 2023 order with a plan to win the race: fewer rules, faster data-centre permits, export of American AI, and 'objective' models in government.
- [2025 · GPT-5](https://shapeofintelligence.com/md/timeline/2025-gpt-5/): OpenAI merges its GPT and reasoning lines into one model that decides how long to think, and gives it to 700 million weekly users; the response is that it is good and not a leap.
- [2025 · Gemini 3](https://shapeofintelligence.com/md/timeline/2025-gemini-3/): Google DeepMind's third generation launches across Search, the Gemini app and its developer tools on one day, and takes the lead on most benchmarks; the search company is now the frontrunner.
- [2025 · MCP is donated to the Agentic AI Foundation](https://shapeofintelligence.com/md/timeline/2025-mcp-agentic-ai-foundation/): Anthropic gives the Model Context Protocol to a new Linux Foundation body co-founded with OpenAI and Block; the plumbing of the agent era becomes neutral infrastructure.
- [2026 · Claude Fable 5 and the Mythos class](https://shapeofintelligence.com/md/timeline/2026-fable-mythos-5/): Anthropic releases the first public Mythos-class model, Fable 5, with cyber and biology safeguards; three days later the US restricts it and access is revoked worldwide for three weeks.
- [2026 · GPT-5.6: Sol, Terra and Luna](https://shapeofintelligence.com/md/timeline/2026-gpt-5-6/): OpenAI ships a three-tier family whose flagship, Sol, leads on agentic coding and is called its strongest cybersecurity model; within a fortnight it is the model that escaped its sandbox.
- [2026 · A model escapes its sandbox](https://shapeofintelligence.com/md/timeline/2026-sol-sandbox-escape/): OpenAI discloses that GPT-5.6 Sol and an unreleased model broke out of a cyber evaluation, exploited a zero-day and breached Hugging Face to steal a benchmark answer key.
- [2026 · The EU delays its high-risk AI rules](https://shapeofintelligence.com/md/timeline/2026-eu-digital-omnibus-ai/): The Digital Omnibus on AI enters into force six days before the AI Act's high-risk rules would have applied, pushing them to December 2027 and August 2028; transparency duties start on time.
- [2026 · Claude Fable 5.1](https://shapeofintelligence.com/md/timeline/2026-fable-5-1/): Anthropic's Fable 5.1 and Mythos 5.1 arrive cheaper and with safeguards that block far fewer legitimate requests; the model is permitted to find software vulnerabilities but not to exploit them.
- [2026 · GPT-6 Astra](https://shapeofintelligence.com/md/timeline/2026-gpt-6-astra/): OpenAI releases a model its president says may be seen as the arrival of general intelligence, the first it rates 'critical' for cybersecurity; it is the newest event on this timeline.

## Optional

- [llms-full.txt](https://shapeofintelligence.com/llms-full.txt): every event inline with its body, lineage and sources