Eighty years at a glance
The AI Timeline
1943
- Paper
The first artificial neuron
McCulloch and Pitts model a neuron as a threshold logic unit and show that networks of them can compute logical functions.
1948
- Paper
Information theory
Claude Shannon defines entropy and the limits of communication — later the basis of cross-entropy loss and perplexity.
1950
- PaperLandmark
Turing's imitation game
Alan Turing asks 'Can machines think?' and proposes a conversational test instead — and suggests building machines that learn.
1955
- System
Logic Theorist
Newell, Shaw and Simon build a program that proves theorems from Principia Mathematica by searching through symbolic steps.
1956
- EventLandmark
Dartmouth workshop: AI gets its name
A summer workshop proposed by McCarthy, Minsky, Rochester and Shannon launches 'artificial intelligence' as a field.
1958
- PaperLandmark
The perceptron
Frank Rosenblatt's perceptron learns its weights from examples — the ancestor of every neural network.
1959
- Paper
A checkers program that learns
Arthur Samuel's checkers player improves through experience and self-play, an early success for machine learning.
1966
- System
ELIZA
Weizenbaum's pattern-matching chatbot convinces some users it understands them — an enduring warning about fluent text.
- Setback
ALPAC report on machine translation
A US report concludes machine translation is slower, less accurate and more expensive than human translation; funding is cut.
- System
Shakey the robot (1966–1972)
SRI's mobile robot combines perception, planning and action; the project produces the A* search algorithm.
1968
- Paper
A* search
Hart, Nilsson and Raphael publish A*, which uses a heuristic estimate to find optimal paths efficiently.
1969
- Event
Minsky & Papert's 'Perceptrons'
A rigorous analysis of what single-layer perceptrons cannot compute (such as XOR) contributes to a long decline in neural-network research.
1973
- Setback
The Lighthill report
Asked by the UK government to assess AI, Sir James Lighthill criticizes its failure to meet its promises; UK funding is cut.
1974
- SetbackLandmark
The first AI winter begins (≈1974–1980)
After years of overpromising, funding for AI research dries up in the US and UK.
1976
- Idea
The physical symbol system hypothesis
Newell and Simon argue that symbol manipulation is necessary and sufficient for general intelligence — symbolic AI's manifesto.
1979
- System
MYCIN rivals experts in evaluation
Stanford's rule-based system for diagnosing infections is rated as acceptable as infectious-disease faculty — yet is never used in practice.
1980
- System
XCON goes into use at DEC
A rule-based system configuring computer orders; it grows to about 2,500 rules and is estimated to save DEC $25M a year.
1982
- Event
Japan's Fifth Generation project (1982–1994)
A national project to build logic-programming AI computers; it spurs rival investment abroad but misses its goals.
1986
- PaperLandmark
Backpropagation popularized
Rumelhart, Hinton and Williams show that multi-layer networks can learn useful internal representations with backpropagation.
1987
- SetbackLandmark
Second AI winter begins
The market for specialized LISP machines collapses; expert systems prove costly to maintain. A long downturn follows.
1995
- Paper
Support vector machines
Cortes and Vapnik's support-vector networks become a leading method as statistical machine learning matures.
1997
- Paper
LSTM
Hochreiter and Schmidhuber's long short-term memory lets recurrent networks keep information over long sequences.
- SystemLandmark1997-05
Deep Blue defeats Kasparov
IBM's chess machine beats the reigning world champion 3½–2½ — a triumph of brute-force search, not learning.
1998
- Paper
Convolutional networks read cheques
LeCun and colleagues' gradient-trained convolutional networks recognize handwritten digits in deployed systems.
2003
- Paper2003-02
A neural language model learns word vectors
Bengio and colleagues train distributed word representations with next-word probabilities, sharing evidence across similar contexts.
2006
- Paper
'Deep learning' revival
Hinton and colleagues show that deep networks can be trained layer by layer, reviving interest in many-layered networks.
2009
- Event
ImageNet dataset
A dataset of millions of labelled images is introduced at CVPR 2009; its challenge (from 2010) becomes the field's benchmark.
2011
- System2011-02
Watson wins Jeopardy!
IBM's question-answering system beats champion contestants, combining many retrieval and scoring components.
2012
- PaperLandmark
AlexNet
A deep convolutional network trained on GPUs wins ImageNet with 15.3% top-5 error, more than 10 points ahead — the deep-learning era begins.
2013
- Paper2013-01
word2vec
Fast training of word embeddings on billions of words; vector arithmetic on meaning becomes a popular demonstration.
2014
- Paper2014-09
Sequence-to-sequence learning
Encoder–decoder LSTMs translate whole sentences end to end.
- PaperLandmark2014-09
Attention for translation
Bahdanau, Cho & Bengio let a translation model look back at every source word instead of one compressed vector.
- Paper2014-10
GloVe
Global word co-occurrence statistics become another route to learned word vectors.
2015
- Paper2015-02
Deep Q-networks master Atari
One network learns dozens of Atari games from pixels and score alone — deep learning meets reinforcement learning.
- Paper2015-12
ResNet
Residual connections make networks with over a hundred layers trainable.
2016
- SystemLandmark2016-03
AlphaGo beats Lee Sedol
Neural networks plus tree search defeat one of the world's best Go players 4–1, years earlier than many expected.
2017
- PaperLandmark2017-06
The Transformer
“Attention Is All You Need” removes recurrence entirely. Nearly every modern language model descends from it.
- Paper2017-06
Learning from human preferences
Agents trained from human comparisons instead of hand-written rewards — the foundation of RLHF.
2018
- PaperLandmark2018-10
BERT
Pretrain a Transformer encoder on unlabelled text, fine-tune everywhere: state of the art across NLP benchmarks.
2019
- System2019-02
GPT-2
A 1.5-billion-parameter decoder produces strikingly fluent text; OpenAI initially stages its release, citing misuse concerns.
2020
- PaperLandmark2020-01
Scaling laws
Language-model loss falls as a smooth power law in model size, data and compute — scale becomes something to plan.
- PaperLandmark2020-05
GPT-3 and in-context learning
A 175-billion-parameter model performs new tasks from a few examples in its prompt.
- System2020-11
AlphaFold 2
DeepMind's system predicts protein structures with accuracy competitive with experiments at the CASP14 assessment.
2021
- Paper2021-02
CLIP: images meet language
Contrastive training on image–caption pairs yields a shared image–text space, powering zero-shot vision and text-to-image models.
2022
- Paper2022-01
Chain-of-thought prompting
Asking large models to show intermediate steps markedly improves multi-step reasoning.
- PaperLandmark2022-03
InstructGPT and RLHF
Supervised fine-tuning plus reinforcement learning from human feedback turns a base model into an instruction follower.
- Paper2022-03
Chinchilla: compute-optimal training
For a fixed compute budget, parameters and training tokens should grow together; many large models were undertrained.
- System2022-08
Open text-to-image models
Stable Diffusion's public release puts high-quality text-to-image generation on consumer hardware.
- SystemLandmark2022-11-30
ChatGPT
A conversational assistant built with RLHF reaches a mass audience within weeks; LLMs become a public technology.
2023
- SystemEvolving · as of 2026-09-262023-03
GPT-4
A multimodal model with large gains on professional and academic exams; frontier labs stop disclosing most training details.
- SystemEvolving · as of 2026-09-262023-07
Capable open-weight LLMs
Meta releases Llama 2 with weights available for commercial use; the open-weight ecosystem accelerates.
2024
- EventEvolving · as of 2026-09-262024-08
EU AI Act enters into force
The first comprehensive AI law of a major jurisdiction; obligations for general-purpose model providers apply from August 2025.
- SystemLandmarkEvolving · as of 2026-09-262024-09-12
Reasoning models: OpenAI o1
A model trained to spend more computation 'thinking' before answering improves sharply on math and coding — inference-time compute becomes a lever.
- EventEvolving · as of 2026-09-262024-10
Nobel Prizes for AI research
Physics: Hopfield and Hinton, for foundational work on neural networks. Chemistry: Baker, Hassabis and Jumper, for protein design and structure prediction.
- EventEvolving · as of 2026-09-262024-11-25
Model Context Protocol
Anthropic open-sources a standard for connecting AI assistants to tools and data; it becomes widely adopted for agents.
2025
- SystemEvolving · as of 2026-09-262025-01-20
DeepSeek-R1
An open-weight reasoning model comparable to o1, with a published account of training reasoning through reinforcement learning.
- SystemEvolving · as of 2026-09-262025-02
Agentic coding tools go mainstream
Tools such as Claude Code let developers delegate multi-step software-engineering tasks to models working in a loop with tools.
- EventEvolving · as of 2026-09-262025-07
Gold-medal-level IMO performance
Experimental reasoning models from Google DeepMind (officially graded) and OpenAI (self-reported) score 35/42 at the International Mathematical Olympiad.
2026
- EventEvolving · as of 2026-09-262026-04
A jagged frontier
The Stanford AI Index 2026 reports models matching or exceeding human baselines on PhD-level science and competition math, while still failing simple tasks such as reading analog clocks reliably.