Can machines think?
Alan Turing, in the journal Mind, UK
Turing suggested a test: if a machine's typed answers cannot be told apart from a person's, we should take its thinking seriously. The question framed the next 76 years.
- ATAlan Turing
- Mind
The history of artificial intelligence, 1950 to October 2026
76 years of AI. One scroll.
In 1950 Alan Turing asked whether machines could think. Scroll down and watch the answer grow: 8 fields branching and merging, two winters when the money froze, and the people who kept going.
- 55 moments
- 8 fields
- 2 winters
- 1 question
- Games and rewards
- Rules and logic
- Learning from data
- Neural networks
- Vision
- Language
- Images and video
- Assistants and agents
Fields on screen Pick one to follow it
Chapter 11950 to 1973
The dream
A few researchers bet that thinking could be written down as steps a machine could follow.
A summer gives AI its name
Dartmouth College, New Hampshire, USA
Four researchers proposed that every part of learning could be described precisely enough for a machine to copy it. John McCarthy's name for the new field stuck: artificial intelligence.
- JMJohn McCarthy
- MMMarvin Minsky
- NRNathan Rochester
- CSClaude Shannon
- Dartmouth College
- Harvard
- IBM
- Bell Labs
A machine that learns by doing
Frank Rosenblatt, Cornell Aeronautical Laboratory, USA
The perceptron adjusted its own connections until it could sort punch cards marked on the left from those marked on the right. It is the seed of today's neural networks.
- FRFrank Rosenblatt
- Cornell Aeronautical Laboratory
- US Office of Naval Research
Machine learning gets its name
Arthur Samuel, IBM, USA
Samuel's checkers program got better by playing against itself, and his 1959 paper used the phrase machine learning. Learning instead of hand-written rules would win, decades later.
- ASArthur Samuel
- IBM
ELIZA, the chatbot people confided in
Joseph Weizenbaum, MIT, USA
ELIZA matched keywords and turned your words back into questions, like a therapist. People felt understood anyway, which troubled its own creator.
- JWJoseph Weizenbaum
- MIT
A book cools neural networks
Marvin Minsky and Seymour Papert, USA
Their book Perceptrons proved what simple one-layer networks could never learn. Interest and money drifted towards rule-based AI.
- MMMarvin Minsky
- SPSeymour Papert
Logic becomes a programming language
Alain Colmerauer and Philippe Roussel, Marseille, France
In Prolog you write down facts and rules, and the computer works out the answers. It became a favourite tool of rule-based AI, especially in Europe and Japan.
- ACAlain Colmerauer
- PRPhilippe Roussel
- Aix-Marseille University
A report turns off the money
James Lighthill, for the UK Science Research Council
The report judged that AI had failed its grand promises, and Britain cut most university AI funding. With US cuts close behind, the first winter set in.
- JLJames Lighthill
- Science Research Council
Chapter 2About 1974 to 1980
The first winter
About 1974 to 1980, though historians draw the lines differently. Big promises met small computers, and Britain and the US cut funding. Research slowed but never stopped.
No landmark in these years. Funding was thin, but the ideas kept moving in a few labs.
Chapter 31980 to 1986
Rules and experts
AI went to work as long lists of if-then rules, written down from human experts.
A network that sees in layers
Kunihiko Fukushima, Japan
Modelled on the brain's visual system, the neocognitron spotted simple strokes first, then the shapes built from them. It inspired the convolutional networks that read images today.
- KFKunihiko Fukushima
Expert systems go to work
John McDermott, Carnegie Mellon, for DEC, USA
XCON used hand-written rules, about 2,500 in the end, to put together computer orders, and DEC estimated it saved around $25 million a year. Companies rushed to bottle expertise as rules.
- JMJohn McDermott
- Carnegie Mellon
- Digital Equipment Corporation
Japan bets on thinking computers
MITI and ICOT, Japan
A national project, planned for ten years, set out to build parallel machines for logic-based AI. It advanced the research but missed its commercial goals, and closed in 1994.
- MITI
- ICOT
ReadSymbolic AI
A memory borrowed from physics
John Hopfield, in PNAS, USA
Hopfield showed that a network could store patterns and rebuild a whole one from a damaged piece. It helped bring neural networks back, and in 2024 Hopfield shared the Nobel Prize in Physics.
- JHJohn Hopfield
ReadNeural network
Backpropagation brings networks back
David Rumelhart, Geoffrey Hinton and Ronald Williams, in Nature
Their paper showed how to train networks with hidden layers by passing each error backwards through them. The maths was older, from Seppo Linnainmaa in Finland and Paul Werbos in the US; this paper made it famous.
- DRDavid Rumelhart
- GHGeoffrey Hinton
- RWRonald Williams
- Nature
Chapter 4About 1987 to 1993
The second winter
About 1987 to 1993; some say it lasted longer. The rules boom went bust, yet a few people kept teaching networks to learn.
The rules boom goes bust
The AI hardware industry, mostly USA
The market for special Lisp machines collapsed as cheaper desktop computers caught up, and expert systems proved costly and brittle. A second winter began.
A network reads handwritten digits
Yann LeCun and colleagues, AT&T Bell Labs, USA
LeNet learned to read zip codes straight from pictures, trained with backpropagation. Later versions read about one in ten US cheques by 2001.
- YLYann LeCun
- AT&T Bell Labs
ReadConvolutional neural networkImage recognitionBackpropagation
A program teaches itself backgammon
Gerald Tesauro, IBM Research, USA
TD-Gammon played itself over and over, learning from each win and loss, and reached strong master level. Top players copied some of its moves.
- GTGerald Tesauro
- IBM Research
Chapter 51994 to 2011
Learning from data
Instead of writing the rules by hand, let the machine find the patterns in examples.
Learning from data, with solid maths
Corinna Cortes and Vladimir Vapnik, in Machine Learning
Support vector machines, and random forests soon after, worked well on modest computers and came with well-understood maths. For years, methods like these did most of machine learning's practical work.
- CCCorinna Cortes
- VVVladimir Vapnik
A computer beats the chess champion
IBM Deep Blue against Garry Kasparov, New York
Deep Blue won the rematch 3.5 to 2.5. It could check about 200 million chess positions a second.
- GKGarry Kasparov
- IBM
Networks that remember longer
Sepp Hochreiter and Jürgen Schmidhuber, in Neural Computation
Long short-term memory let a network keep track of things it saw many steps earlier. In the mid-2010s it powered voice typing, Google Translate and Siri.
- SHSepp Hochreiter
- JSJürgen Schmidhuber
Deep learning gets a fresh start
Geoffrey Hinton, Simon Osindero and Yee-Whye Teh
They trained deep networks one layer at a time, which made many-layered networks practical again. It helped the phrase deep learning catch on.
- GHGeoffrey Hinton
- SOSimon Osindero
- YTYee-Whye Teh
ImageNet: millions of labelled pictures
Fei-Fei Li and team, Princeton, USA
ImageNet grew to over 14 million images in more than 20,000 categories, labelled by 49,000 online workers in 167 countries. Vision models finally had a big, shared test.
- FLFei-Fei Li
- Princeton
A computer wins Jeopardy!
IBM Watson, USA
Watson beat the quiz show's two greatest champions by searching huge amounts of text for likely answers. Computers could now handle questions in everyday English.
- IBM
An assistant in your pocket
Apple, from an SRI International spin-out, USA
Siri arrived built into the iPhone 4S, so millions could ask a phone questions out loud. It grew out of a US defence research project on digital assistants.
- Apple
- SRI International
Chapter 62012 to 2016
Deep learning
Big datasets, gaming chips and old ideas about neural networks finally clicked.
Deep learning wins on images
Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton, University of Toronto, Canada
AlexNet, trained on graphics chips, cut the ImageNet error rate to about 15%, against 26% for the next best entry. Vision research switched to deep learning soon after.
- AKAlex Krizhevsky
- ISIlya Sutskever
- GHGeoffrey Hinton
- University of Toronto
- ImageNet challenge
Words become points in space
Tomas Mikolov and team, Google, USA
Word2vec turned each word into a list of numbers, so words with similar meanings sit close together. That idea, the embedding, sits inside every language model today.
- TMTomas Mikolov
One program learns Atari games
Volodymyr Mnih and team, DeepMind
It learned to play straight from the pixels on screen, beating a human expert on three of seven games. Deep learning and trial-and-error learning had joined up.
- VMVolodymyr Mnih
- DeepMind
Two networks compete to make images
Ian Goodfellow and team, Université de Montréal, Canada
One network forges examples while the other tries to catch the fakes, and both get better. GANs led to strikingly realistic made-up faces, and later to deepfakes.
- IGIan Goodfellow
- YBYoshua Bengio
- Université de Montréal
Attention: look back at what matters
Dzmitry Bahdanau, Kyunghyun Cho and Yoshua Bengio, Bremen and Montreal
Their translation model learned which words of the original to focus on for each word it wrote. The trick, called attention, became the heart of the transformer.
- DBDzmitry Bahdanau
- KCKyunghyun Cho
- YBYoshua Bengio
- Jacobs University Bremen
- Université de Montréal
Very deep networks finally train
Kaiming He and team, Microsoft Research Asia, China
Shortcut connections let signals skip past layers, so networks with more than a hundred layers could learn. ResNet won the 2015 ImageNet challenge, and its shortcuts are now used almost everywhere.
- KHKaiming He
- XZXiangyu Zhang
- SRShaoqing Ren
- JSJian Sun
- Microsoft Research
AlphaGo beats Lee Sedol at Go
DeepMind, owned by Google, match in Seoul, South Korea
AlphaGo won 4 to 1, helped in game two by a move experts first took for a mistake. More than 200 million people watched.
- LSLee Sedol
- DeepMind
Chapter 72017 to 2021
The transformer age
One new design for reading text, and then a race to make models bigger.
Attention is all you need
Ashish Vaswani and seven co-authors, Google
The transformer dropped word-by-word reading and let every word look at every other word at once, which suits fast parallel chips. Almost every large language model since is built on it.
- AVAshish Vaswani
- NSNoam Shazeer
- NPNiki Parmar
- JUJakob Uszkoreit
Pre-train first, then specialise
OpenAI, USA
OpenAI's first GPT read lots of unlabelled text to learn language in general, then was fine-tuned for each task. This recipe, pre-training, is how chat models start life.
- OpenAI
BERT reads in both directions
Jacob Devlin and team, Google, USA
BERT read text both ways to understand each word from its whole sentence. From 2019 it helped Google Search understand what people meant.
- JDJacob Devlin
- MCMing-Wei Chang
- KLKenton Lee
- KTKristina Toutanova
Bigger models, surprising skills
OpenAI, USA
GPT-3 had 175 billion parameters and could pick up new tasks from a few examples in the prompt. Earlier that year, scaling laws showed that more data, compute and size improve models predictably.
- JKJared Kaplan
- OpenAI
ReadScaling lawsParametersFew-shot promptingLarge language model
AlphaFold predicts protein shapes
John Jumper, Demis Hassabis and team, DeepMind, London
At the CASP14 contest it predicted how proteins fold from their chain of building blocks, a problem biology had faced for 50 years. It led to the 2024 Nobel Prize in Chemistry.
- JJJohn Jumper
- DHDemis Hassabis
- DeepMind
- CASP14
Pictures from a sentence
OpenAI, USA
DALL-E, a 12 billion parameter cousin of GPT-3, drew images from written descriptions. Words and pictures started to share one model.
- OpenAI
Chapter 82022 to 2023
Everyone meets AI
AI stopped being a lab project and became an app on hundreds of millions of phones.
Image generators open to everyone
Midjourney, USA; Stable Diffusion, from LMU Munich, Runway and Stability AI
Midjourney opened its beta in July, and in August Stable Diffusion published its weights so anyone could run it on a home graphics card. Both start from noise and clean it up, step by step, into a picture.
- RRRobin Rombach
- BOBjörn Ommer
- DHDavid Holz
- Midjourney
- LMU Munich
- Runway
- Stability AI
ReadDiffusion modelStable DiffusionMidjourneyOpen-weight model
ChatGPT: everyone meets AI
OpenAI, USA
A chat window on a GPT-3.5 model, tuned with human feedback, reached a million users in five days.
- OpenAI
Meta shares LLaMA with researchers
Meta AI
LLaMA came in sizes from 7 to 65 billion parameters, small enough for university labs to run. It set off a wave of open-weight models anyone could study and adapt.
- Meta
Language models learn to see
OpenAI's GPT-4, as Anthropic's Claude and Google's Bard arrive
GPT-4 took pictures as well as text, and OpenAI reported a top 10% score on a simulated bar exam. The same month Claude and Bard opened, and chatbots became a crowded field.
- OpenAI
- Anthropic
Paris joins in, with open models
Mistral AI, Paris, France
Founded that April by researchers from Google DeepMind and Meta, Mistral released a 7 billion parameter model under a licence free for any use. Strong models no longer came only from US giants.
- AMArthur Mensch
- GLGuillaume Lample
- TLTimothée Lacroix
- Mistral AI
Built to see, hear and read
Google DeepMind
Gemini was trained on text, code, images, audio and video together from the start. It came in three sizes, down to one that runs on a phone.
- Google DeepMind
Chapter 92024 to now
AI that acts
Models learned to reason, to use tools and computers, and to work through whole tasks.
Video from a sentence
OpenAI, USA
Sora's preview made realistic clips up to a minute long from a written prompt. Video models spread fast, though OpenAI itself closed the Sora app in April 2026.
- OpenAI
Models that think before answering
OpenAI's o1, then Google, DeepSeek and others
Trained with reinforcement learning to work through a chain of thought, o1 solved about 74% of a hard maths contest's problems where GPT-4o solved about 12%, by OpenAI's count. Reasoning became the new frontier.
- OpenAI
Nobel Prizes for AI pioneers
Physics and Chemistry, Stockholm
Physics went to John Hopfield and Geoffrey Hinton for the foundations of neural networks. Half of Chemistry went to Demis Hassabis and John Jumper for AlphaFold, the other half to David Baker.
- JHJohn Hopfield
- GHGeoffrey Hinton
- DHDemis Hassabis
- JJJohn Jumper
- DBDavid Baker
- Princeton
- University of Toronto
- Google DeepMind
- University of Washington
AI learns to use a computer
Anthropic's Claude, then OpenAI's Operator in January 2025
Claude could look at a screen, move the cursor, click and type, though at first it was slow and made mistakes. Agents could now work inside apps built for people.
- Anthropic
- OpenAI
A common plug for AI tools
Anthropic; since December 2025, the Linux Foundation's Agentic AI Foundation
The Model Context Protocol gave AI apps one standard way to connect to files, apps and data. A year on it had over 10,000 public servers and moved to a neutral foundation.
- Anthropic
- Linux Foundation
An open reasoning model from Hangzhou
DeepSeek, Hangzhou, China
R1 came close to the leading reasoning models on maths and coding tests, and came with open weights. Its app topped the US App Store, and Nvidia's share price fell about 17% in a day.
- LWLiang Wenfeng
- DeepSeek
Coding agents take off
Anthropic's Claude Code, OpenAI's Codex and others
In February Anthropic previewed Claude Code and Andrej Karpathy coined vibe coding: describe what you want, and let the AI write it. Agents like Claude Code, and OpenAI's Codex soon after, read a codebase, edit files, run tests and fix their own errors.
- AKAndrej Karpathy
- Anthropic
- OpenAI
Gold-medal maths at the Olympiad
Google DeepMind and OpenAI, IMO 2025
An advanced Gemini solved five of six problems, writing its proofs in natural language within the 4.5 hour limit, and IMO coordinators graded it gold. Two days earlier OpenAI had reported a gold-level score for an experimental model, graded by former medallists rather than officially.
- Google DeepMind
- OpenAI
Agents for everyday work
OpenClaw, from Austria; Anthropic's Cowork
OpenClaw, an open-source agent you message from chat apps, had about 247,000 GitHub stars by March, and Anthropic's Cowork brought agents to office work. Security experts warned about agents holding the keys to email and files.
- PSPeter Steinberger
- OpenClaw Foundation
- Anthropic
New Delhi hosts the AI summit
India AI Impact Summit, Bharat Mandapam, New Delhi
88 countries and international organisations endorsed the New Delhi Declaration on AI Impact, which calls for sharing AI's benefits widely. Bengaluru start-up Sarvam showed 30 and 105 billion parameter models with support for Indian languages.
- Government of India
- Sarvam AI
AI finds thousands of hidden bugs
Anthropic's Project Glasswing, with partners across tech
Anthropic said Claude Mythos Preview had found thousands of serious security flaws, including one hidden in OpenBSD for 27 years. It gave the model only to partners defending key software, instead of releasing it publicly.
- Anthropic
- Linux Foundation
- Microsoft
- Apple
A million tokens, open to all
DeepSeek, Hangzhou, China
DeepSeek previewed V4 with open weights and room for up to a million tokens of context, enough for several long books at once. Its largest version, V4 Pro, has 1.6 trillion parameters.
- DeepSeek
Cyber defenders get a new model first
Google's Gemini 4 Argon, via its Fairwind Program
On 30 September Google gave its newest frontier model to vetted cyber defenders first, to refine its guardrails, then made it generally available on 13 October. In April Anthropic had kept Claude Mythos Preview to security partners.
The whole map
You are here. October 2026.
55 moments, 8 fields, two winters. Some lines merged into others, some went quiet, and the newest ones are still being drawn. Tap any dot to jump back to its moment.
- Games and rewards
- Rules and logic
- Learning from data
- Neural networks
- Vision
- Language
- Images and video
- Assistants and agents
Every word on this map, explained
Each chip on the timeline opens a plain-words page with its own picture. Start anywhere.
Checked October 2026. Every date, name and organisation here was checked against the original paper, the organisation's own announcement or a standard reference. Historians date the two AI winters differently, so treat those years as approximate.