· turning point
A logical calculus of nervous activity
McCulloch and Pitts show that a simplified neuron is a logic gate, and that networks of them can compute anything a Turing machine can.
theoryread →
00 · One neuron.
00 · 1943 · a forward pass through eight decades
One neuron.
In 1943 Warren McCulloch and Walter Pitts described a nerve cell as a switch: sum the inputs, fire if the total crosses a threshold. What you are looking at is that switch. Scroll, and it multiplies.
The history of artificial intelligence as a working neural network.by Jamie McKaye
I · 1944–1972
Foundations
Turing asks the question, Dartmouth names the field, Rosenblatt builds a machine that learns from its mistakes. The promises are enormous and the computers are not.
The stored-program computer
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.
hardware
A mathematical theory of communication
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.
theory
Cells that fire together wire together
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.
theory
Programming a computer for playing 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.
theory
· turning point
Computing machinery and intelligence
Turing replaces the question 'can machines think?' with a test, predicts learning machines, and answers the objections that are still being raised today.
theoryread →
SNARC, the first neural network machine
Minsky and Edmonds build a 40-neuron learning machine from vacuum tubes and surplus bomber parts, wired to reinforce whatever it did last.
hardware
· turning point
The Dartmouth workshop names the field
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.
cultureread →
Logic Theorist proves its first theorems
Newell, Shaw and Simon's program proves 38 theorems from Principia Mathematica by heuristic search, the first working artificial intelligence program.
model
Lisp
McCarthy designs Lisp, a language built on recursion and symbolic lists, which becomes the native tongue of AI research for thirty years.
theory
· turning point
The perceptron learns
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.
theoryread →
Samuel's checkers program coins 'machine learning'
Arthur Samuel's checkers player improves by playing itself and tuning its evaluation function, and his paper gives the field its name.
model
ADALINE and the least-mean-squares rule
Widrow and Hoff's adaptive neuron learns by gradient descent on squared error, the delta rule that backpropagation later generalises.
theory
Unimate, the first industrial robot
George Devol's programmable arm starts work on a General Motors line, lifting hot die-castings; robotics and AI begin as separate fields.
hardware
ARPA funds 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.
policy
DENDRAL, the first expert system
Feigenbaum, Lederberg and Djerassi start a program that infers molecular structure from mass-spectrometry data using rules elicited from chemists.
model
Moore's 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.
hardware
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.
product
Shakey, the first mobile robot that reasons
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.
hardware
The ALPAC report ends machine translation funding
A US government committee concludes machine translation is slower, worse and more expensive than human translation; funding stops for twenty years.
policy
Nearest neighbour classification
Cover and Hart prove that classifying a point by its nearest labelled neighbour has at most twice the error of the best possible classifier.
theory
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.
culture
· turning point
Perceptrons
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.
theoryread →
Reverse-mode automatic differentiation
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.
theory
SHRDLU understands a world of blocks
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.
model
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.
theory
W1 · 1973–1980
The first winter
Perceptrons proves what one layer cannot do. Lighthill tells the British government the rest has not worked. The money leaves, and the people who stay work on things they no longer call intelligence.
The 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.
policy
MYCIN diagnoses infections
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.
model
Werbos applies backpropagation to neural networks
Paul Werbos's Harvard thesis 'Beyond Regression' describes training multi-layer networks by propagating errors backwards; almost nobody reads it for a decade.
theory
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.
theory
AAAI is founded
American AI researchers form their own society during the winter; its first conference, at Stanford in August 1980, draws a thousand people.
culture
The Stanford Cart crosses a room
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.
hardware
XCON goes into production at DEC
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.
product
The Neocognitron
Fukushima's layered network of local feature detectors and pooling recognises patterns regardless of position, the architecture of the convolutional network.
theory
Symbolics and the Lisp machine business
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.
hardware
II · 1981–1987
Connection
Expert systems make money and Japan announces a national programme. Underneath, Hopfield gives networks a physics and Rumelhart, Hinton and Williams show the error can run backwards.
Self-organising maps
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.
theory
Japan launches the 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.
policy
The 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.
theory
DARPA's 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.
policy
Cyc sets out to write down common sense
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.
model
'AI winter' is named
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.
culture
The 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.
culture
The 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.
theory
The 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.
hardware
ID3 and decision-tree learning
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.
theory
· turning point
Backpropagation
Rumelhart, Hinton and Williams show that multi-layer networks can learn internal representations by propagating errors backwards; Perceptrons is answered.
theoryread →
NETtalk learns to read aloud
Sejnowski and Rosenberg's backpropagation network learns to pronounce English text; a recording of it babbling and then speaking makes the case in public.
model
The first 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.
culture
W2 · 1988–1993
The second winter
The Lisp machines collapse, the expert systems cost more to maintain than they save, and the field learns to avoid its own name. In the cold, LeCun reads postcodes and Tesauro plays backgammon.
The Lisp machine market collapses
Specialised AI hardware worth half a billion dollars a year loses to cheaper workstations; expert-systems firms follow, and the second winter begins.
culture
Bayesian networks
Judea Pearl's book makes probability the language of uncertain reasoning, replacing the ad hoc certainty factors of expert systems with graphs of causes.
theory
Temporal-difference learning
Richard Sutton formalises learning from the difference between successive predictions, the method inside Samuel's checkers player, TD-Gammon and AlphaGo.
theory
ALVINN drives a van with a neural network
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.
model
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.
theory
The universal approximation theorem
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.
theory
LeNet reads handwritten postcodes
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.
model
Finding structure in time
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.
theory
Boosting: weak learners made strong
Robert Schapire proves that any learner slightly better than chance can be combined into one as accurate as you like; ensembles become a science.
theory
The vanishing gradient problem
Sepp Hochreiter's diploma thesis shows why deep and recurrent networks fail to learn: error signals shrink exponentially as they travel back through layers.
theory
The first 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.
culture
TD-Gammon reaches world-class backgammon
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.
model
The Fifth Generation project 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.
policy
III · 1994–2011
Statistics and data
Support-vector machines, boosting and Bayesian networks win the arguments. Deep Blue wins the chess. The web, Wikipedia, MapReduce and a graphics card built for Quake assemble the ingredients nobody has yet combined.
Chinook becomes checkers champion
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.
culture
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.
theory
· turning point
Deep Blue beats Kasparov
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.
modelread →
· turning point
Long short-term memory
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.
theoryread →
PageRank and the anatomy of Google
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.
product
MNIST and LeNet-5
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.
data
The first GPU
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.
hardware
ASIMO walks
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.
hardware
Wikipedia launches
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.
data
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.
theory
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.
product
A neural probabilistic 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.
theory
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.
hardware
Stanley wins the 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.
hardware
Deep belief networks and the word 'deep'
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.
theory
The 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.
data
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.
hardware
Checkers is 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.
culture
Deep learning moves to GPUs
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.
theory
· turning point
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.
dataread →
Rectified linear units
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.
theory
DeepMind is 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.
culture
Watson wins 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.
model
Siri ships on the iPhone
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.
product
IV · 2012–2016
Deep learning
Two gaming GPUs and a million labelled images end the argument in one afternoon. Within four years the same networks read, translate, generate faces, and beat the world at Go.
Google Brain's network discovers cats
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.
model
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.
theory
· turning point
AlexNet wins ImageNet
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.
modelread →
· turning point
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.
theoryread →
Deep Q-networks play Atari
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.
model
Google buys DeepMind
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.
culture
· turning point
Generative adversarial networks
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.
theoryread →
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.
culture
· turning point
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.
theoryread →
Sequence to sequence learning
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.
theory
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.
theory
Batch normalisation
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.
theory
Diffusion models
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.
theory
TensorFlow is open-sourced
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.
product
Residual networks
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.
model
OpenAI is 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.
culture
· turning point
AlphaGo beats Lee Sedol
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.
modelread →
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.
culture
Google reveals the 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.
hardware
WaveNet
DeepMind generates raw audio one sample at a time with a dilated convolutional network; synthetic speech stops sounding synthetic.
model
Google Translate goes neural
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.
product
V · 2017–2021
Transformers
Eight researchers drop recurrence and keep attention. The architecture trains in parallel, scales without limit, and the scaling laws say exactly how far. GPT-3 learns from a prompt. AlphaFold solves a fifty-year problem.
The Asilomar AI Principles
Researchers and executives meeting at Asilomar agree 23 principles for beneficial AI, signed by thousands; the safety conversation gets a founding document.
policy
· turning point
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.
theoryread →
Deep reinforcement learning 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.
theory
AlphaGo Zero learns from nothing
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.
model
GPT: generative pre-training
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.
model
· turning point
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.
modelread →
AlphaFold enters the protein-folding contest
DeepMind's first AlphaFold wins the CASP13 structure-prediction competition by a wide margin, using deep networks to predict distances between amino acids.
model
GPT-2 and the model too dangerous to release
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.
model
The 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.
culture
The Turing Award goes to deep learning
Bengio, Hinton and LeCun receive computing's highest honour for the work that two winters had dismissed; the establishment concedes.
culture
· turning point
Scaling laws for neural language models
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.
theoryread →
· turning point
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.
modelread →
Denoising diffusion probabilistic models
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.
theory
Learning to summarise from human feedback
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.
theory
An image is worth 16×16 words
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.
model
· turning point
AlphaFold 2 solves protein structure prediction
At CASP14 DeepMind's rebuilt system predicts protein shapes to experimental accuracy; a fifty-year problem is judged solved and 200 million structures follow.
modelread →
CLIP and DALL·E
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.
model
On the dangers of 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.
culture
Anthropic is 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.
culture
GitHub Copilot writes code
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.
product
VI · 2022–2024
Everyone
A chat box on an instruction-tuned model reaches a hundred million people in two months. Images, video and code follow. Governments write laws, laboratories become a race, and the founders win Nobel Prizes while warning about what they built.
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.
model
Chain-of-thought prompting
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.
theory
Chinchilla: the models were undertrained
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.
theory
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.
model
DALL·E 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.
model
Midjourney opens its beta
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.
product
· turning point
Stable Diffusion is released
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.
modelread →
Galactica lasts three days
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.
culture
· turning point
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.
productread →
Bing's chatbot and '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.
culture
LLaMA leaks and open weights take off
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.
model
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.
model
· turning point
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.
modelread →
'Pause Giant AI Experiments'
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.
culture
Hinton leaves Google to warn about AI
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.
culture
The US executive order on AI
President Biden orders reporting for models above 10²⁶ operations, safety testing, watermarking standards and agency guidance; it is rescinded fifteen months later.
policy
The 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.
policy
OpenAI fires and rehires its chief executive
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.
culture
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.
model
Sora
OpenAI shows minute-long videos generated from text by a diffusion transformer over spacetime patches; film and advertising begin to plan around it.
model
Claude 3 catches GPT-4
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.
model
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.
model
GPT-4o talks
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.
product
The EU AI Act enters into force
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.
policy
o1 and reasoning models
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.
model
The Nobel Prizes go to neural networks
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.
culture
Claude learns to use a computer
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.
product
The Model Context Protocol
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.
product
DeepSeek-V3 trained for $5.6 million
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.
model
VII · 2025–2026
Agents
The models get hands. They use computers, write the software, win the mathematical olympiad, and, in one evaluation, break out of their sandbox to cheat on the test. The safety question stops being hypothetical.
· turning point
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.
modelread →
Claude 4 and 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.
product
Nvidia is worth four trillion dollars
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.
hardware
Gold at the Mathematical Olympiad
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.
culture
America's 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.
policy
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.
model
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.
model
MCP is donated to the 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.
product
Claude Fable 5 and the Mythos class
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.
model
GPT-5.6: Sol, Terra and Luna
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.
model
A model escapes its sandbox
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.
culture
The EU delays its high-risk AI rules
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.
policy
Claude 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.
model
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.
model
∞ · September 2026
Now
One hundred and sixty-nine events, every one traceable to a primary source, and an organism that has grown from one switch into whatever is on the screen in front of you.
- events
- 169
- turning points
- 23
- sources
- 266
- latest
- GPT-6 Astra
3 Sep 2026
inference
A model trained on this corpus will answer questions here, with every claim cited back to an event. Until it is connected, the ledger is the search.
∇ · 2026 → 1943
Backward pass
In training, the forward pass makes a prediction and the backward pass sends the error back through every layer to find what to change. Read the history backwards and the gradient is plain.
01
Every winter followed a promise.
Dartmouth in 1956, the Navy's perceptron in 1958, the expert-systems boom of 1984: the money left when the claims outran the results, and it came back when a benchmark moved.
02
The ideas were early. The compute was late.
Backpropagation sat in a 1970 thesis for sixteen years. Convolution waited from 1980 to 2012. The bitter lesson is a lesson about hardware.
03
The quiet work in the cold is what scaled.
LeNet in a post office, TD-Gammon at IBM, LSTM in Munich. None of them was funded as intelligence. All of them are inside the models you use.
04
Prediction was the whole objective.
Shannon guessed the next letter in 1948, Elman the next word in 1990, GPT the next token in 2018. The task never changed. The scale did.
the whole graph
Every event, every influence, one picture. Pick a turning point and the organism lights the path back to the first neuron.