Timeline
A coarse chronology of key moments in the history of artificial intelligence.
Early ideas & foundations
-
Turing’s “imitation game”
Alan Turing’s paper in Mind reframes machine intelligence as behaviour indistinguishable from a human in dialogue, later popularised as the Turing test.
-
Dartmouth AI workshop convenes
Summer research workshop where John McCarthy proposed the name “artificial intelligence” and a shared research agenda, widely treated as the field’s public debut.
-
Rosenblatt’s perceptron draws attention
Frank Rosenblatt’s trainable single-layer classifier sparks optimism, and later debate, about whether neural networks can scale to general intelligence.
-
ELIZA shows conversational pattern-matching
Joseph Weizenbaum’s ELIZA script demonstrates how simple rules can feel like empathy, raising early questions about illusion, trust, and “understanding” in software.
Commercial AI & winters
-
First AI winter begins
After years of bold claims, the Lighthill Report and DARPA funding cuts cool expectations; neural-network and broad AI projects lose support through the mid-1970s.
-
Expert systems go industrial
Rule-based “expert systems” move from labs into product stories, promising domain capture without full learning, and shaping the 1980s AI business wave.
-
Boltzmann machines bring statistical physics to neural nets
Ackley, Hinton and Sejnowski publish a learning algorithm for Boltzmann machines in Cognitive Science; stochastic binary units and an energy function let networks learn hidden structure, but training needs slow Markov-chain sampling.
-
Backpropagation makes multilayer networks trainable
Rumelhart, Hinton and Williams show in Nature how a network can pass its errors backwards through its layers so that hidden units learn useful features of their own; the method becomes the standard way to train neural networks.
-
Second AI winter begins
After a boom in specialised AI tools, market expectations cool; funding tightens and many commercial projects stall, often dated to the late 1980s “winter.”
-
Convolutional nets read handwritten zip codes
LeCun and colleagues at Bell Labs train a convolutional neural network with backpropagation on US Postal Service digits, an early practical win for local filters in vision.
-
TD-Gammon learns backgammon through self-play
Gerald Tesauro at IBM trains TD-Gammon with temporal-difference learning and self-play, reaching strong backgammon play during the second AI winter.
-
Support vector machines formalise maximum-margin learning
Cortes and Vapnik publish Support-Vector Networks in Machine Learning, linking margin maximisation with kernels for nonlinear classification.
-
Deep Blue defeats Kasparov
IBM’s chess specialist beats world champion Garry Kasparov in a match, an iconic public moment for brute-force search plus hardware, even as broader AI still looked modest.
-
Random forests combine bagged decision trees
Leo Breiman publishes Random Forests in Machine Learning, defining an ensemble of decorrelated trees that becomes a standard classifier before the deep learning revival.
-
Neural probabilistic language models learn word embeddings
Bengio, Ducharme, Vincent and Jauvin publish A Neural Probabilistic Language Model in JMLR, learning distributed word vectors to escape the limits of huge n-gram tables.
-
Stanley wins the DARPA Grand Challenge
Stanford's Stanley completes the 212 km Mojave Desert course in 6 hours 53 minutes, the first Grand Challenge finisher after the 2004 race had no winner.
Deep learning at scale
-
Deep belief networks make depth trainable again
Hinton, Osindero and Teh show how to stack restricted Boltzmann machines into deep belief nets and fine-tune them with backpropagation; a Science paper the same summer demonstrates powerful dimensionality reduction.
-
ImageNet seeds large-scale vision
Fei-Fei Li’s team launches a vast labelled image dataset, later the benchmark that makes modern convolutional networks measurable and competitive.
-
IBM's Watson wins at Jeopardy!
IBM's question-answering system Watson beats record champions Ken Jennings and Brad Rutter on the quiz show Jeopardy!, answering natural-language clues without being connected to the internet.
-
AlexNet wins ImageNet
A deep convolutional network trained on GPUs cuts error rates dramatically on ImageNet, often cited as the spark of the modern deep-learning resurgence.
-
Generative adversarial networks pit two networks against each other
Ian Goodfellow and colleagues in Montreal introduce GANs: a generator produces data, a discriminator tries to tell it apart from real examples, and both improve through the contest; the idea shapes image synthesis for years to come.
-
AlphaGo beats [Lee Sedol](/en/people/lee-sedol/)
DeepMind’s AlphaGo wins four of five games against one of the world’s top Go players, showing that intuition-heavy games are not off limits to learning systems.
Transformers & large language models
-
“Attention is all you need”
The Transformer architecture replaces recurrence with self-attention, becoming the backbone of later large language models and multimodal stacks.
-
GPT-3 scales few-shot prompting
OpenAI’s 175B-parameter language model paper highlights in-context learning: tasks described in natural language without weight updates, reshaping product and research expectations.
-
ChatGPT opens to the public
A conversational wrapper around a large language model becomes a mass-market demo overnight, accelerating debates on safety, work, copyright, and access.
-
GPT-4 ships as a multimodal flagship
OpenAI positions GPT-4 as a safer, stronger successor, vision + text demos reset expectations for capability benchmarks and product roadmaps.
-
Llama 2 weights go open(ish) for research & products
Meta releases Llama 2 under a community license, widening who can fine-tune and ship local or hosted models, and accelerating the “open weights” ecosystem debate.
-
Google launches Gemini 1.0
Google rebrands and ships its multimodal family under the Gemini name, signalling tighter integration across Search, cloud APIs, and consumer hardware stories.
-
EU AI Act adopted
The European Parliament backs the world’s first broad AI rulebook, layering obligations by risk, transparency duties for general-purpose models, and timelines for compliance.
-
OpenAI previews “reasoning” models (o1)
A new line emphasises longer internal deliberation before answering, rekindling public talk about test-time compute, STEM benchmarks, and safety evaluation gaps.
Reasoning, agents & the AI boom
-
Neural networks win the Nobel Prize in Physics
John Hopfield and Geoffrey Hinton are honored for foundational work on artificial neural networks, moving the machinery of modern AI to the center of science.
-
AlphaFold wins the Nobel Prize in Chemistry
A day after the physics prize, the chemistry Nobel honors AlphaFold's protein structure prediction and computational protein design, casting AI as an instrument of discovery.
-
Gemini 2.0 and the "agentic era"
Google releases Gemini 2.0 and puts AI agents at the center, previewing systems that take multi-step actions in a browser and beyond, not just answer questions.
-
DeepSeek-R1 and the efficiency shock
A Chinese startup releases an open-weights reasoning model rivaling the best at a fraction of the reported cost; days later it triggers a record one-day loss in Nvidia's value.
-
The Stargate Project is announced
OpenAI, SoftBank, Oracle and MGX unveil a plan to invest up to $500 billion in US AI data centers, reframing progress as a question of capital and physical infrastructure.
-
The Paris AI Action Summit exposes a governance split
Around 60 countries sign a declaration on inclusive, sustainable AI in Paris; the US and UK refuse, revealing a widening rift over how to govern the technology.
-
AI reaches gold at the International Mathematical Olympiad
Language models from Google DeepMind and OpenAI solve five of six problems at the International Mathematical Olympiad and score 35 of 42 points, gold-medal level, writing proofs in natural language within the exam time.
-
GPT-5 makes reasoning the default
OpenAI releases GPT-5 as a unified system that decides when to answer quickly and when to think longer, and ships it to every ChatGPT user; a bumpy rollout shows how attached people have become to individual models.
-
The New Delhi AI Impact Summit brings the US back to the table
At the first global AI summit held in the Global South, more than 80 countries, including the US, China and the UK, endorse a non-binding declaration on democratising AI, while Washington openly rejects any global governance of the technology.