The history of AI runs from the first artificial neuron in 1943 and the 1956 Dartmouth workshop, through two AI winters, to deep learning in 2012, ChatGPT in 2022 and today's AI agents. This timeline covers 75 milestones in six eras, each with why it mattered and a source.
Pick an era or a type, or search for a name, then copy a link to exactly what you see.
Showing 75 of 75 milestones
Artificial neurons, the Turing test, the Dartmouth workshop and the first programs that reasoned, learned and talked.
Warren McCulloch and Walter Pitts show that networks of simple on/off "neurons" can compute logic, the seed of every neural network. Source for First artificial neuron (opens in a new tab)
Alan Turing's paper "Computing Machinery and Intelligence" proposes the imitation game, later called the Turing test. Source for Turing asks "Can machines think?" (opens in a new tab)
Marvin Minsky and Dean Edmonds build a 40-neuron learning machine from vacuum tubes that learns to run a maze. Source for SNARC, the first neural network machine (opens in a new tab)
Allen Newell, Herbert Simon and Cliff Shaw's program proves 38 of the first 52 theorems in Principia Mathematica, often called the first AI program. Source for Logic Theorist proves theorems (opens in a new tab)
John McCarthy's 1955 proposal coins "artificial intelligence"; the 1956 summer workshop is treated as AI's founding event. Source for Dartmouth workshop names the field (opens in a new tab)
Frank Rosenblatt's perceptron learns to classify inputs from examples, the first trainable neural network. Source for Rosenblatt's perceptron (opens in a new tab)
John McCarthy creates LISP, which becomes the main language of AI research for three decades. Source for LISP programming language (opens in a new tab)
Arthur Samuel's checkers program improves by playing itself; his 1959 paper popularizes the term machine learning. Source for "Machine learning" is coined (opens in a new tab)
SRI's Shakey (1966-1972) is the first mobile robot to reason about its own actions, and gives us the A* search algorithm. Source for Shakey the robot (opens in a new tab)
Joseph Weizenbaum's MIT program imitates a therapist with simple pattern matching, and users still confide in it. Source for ELIZA, the first chatbot (opens in a new tab)
Minsky and Papert prove single-layer perceptrons cannot learn functions like XOR, and neural network research stalls for over a decade. Source for Perceptrons book (opens in a new tab)
Funding cuts, a boom in rule-based expert systems, then a second crash when they proved brittle and costly.
A report for the UK Science Research Council says AI has not delivered the major impact it promised; UK and later US funding is cut, starting the first AI winter (about 1974-1980). Source for Lighthill report (opens in a new tab)
Stanford's rule-based MYCIN (early 1970s) recommends antibiotics about as well as specialists, proving the expert-system idea. Source for MYCIN expert system (opens in a new tab)
The R1/XCON expert system configures VAX computer orders and is reported to save DEC about $40 million a year, sparking an expert-system boom. Source for XCON goes to work at DEC (opens in a new tab)
Japan's ten-year national AI computing program prompts rival funding in the US (Strategic Computing) and UK (Alvey). Source for Japan's Fifth Generation project (opens in a new tab)
Rumelhart, Hinton and Williams show in Nature that backpropagation trains multi-layer networks, the method still used to train modern AI. Source for Backpropagation popularized (opens in a new tab)
Cheaper desktop computers wipe out the specialized AI hardware business, and the second AI winter begins. Source for Lisp machine market collapses (opens in a new tab)
Yann LeCun's team at Bell Labs trains a convolutional neural network with backpropagation to read handwritten zip codes. Source for Neural nets read zip codes (opens in a new tab)
Statistics replaces hand-written rules: Deep Blue, LSTMs, SVMs, ImageNet and the first voice assistants.
Corinna Cortes and Vladimir Vapnik's soft-margin SVM becomes the go-to machine learning method for the next 15 years. Source for Support vector machines (opens in a new tab)
Sepp Hochreiter and Jürgen Schmidhuber's long short-term memory lets networks remember across long sequences; it later powers speech recognition and translation. Source for LSTM networks (opens in a new tab)
IBM's chess computer beats world champion Garry Kasparov 3.5-2.5 in a six-game match, a first for a machine under standard time controls. Source for Deep Blue beats Kasparov (opens in a new tab)
LeCun's LeNet-5 and the MNIST handwritten-digit dataset become the standard first benchmark in machine learning. Source for LeNet-5 and MNIST (opens in a new tab)
iRobot's robot vacuum becomes the first autonomous robot millions of people own. Source for Roomba (opens in a new tab)
Stanford's self-driving car finishes the 132-mile desert course first; in 2004 no car had made it past about 7 miles. Source for Stanley wins the DARPA Grand Challenge (opens in a new tab)
Geoffrey Hinton's team shows how to train deep networks layer by layer, helping revive neural networks under the name deep learning. Source for Deep belief networks (opens in a new tab)
Fei-Fei Li's team releases millions of labeled images; its yearly challenge (from 2010) becomes the proving ground for computer vision. Source for ImageNet dataset (opens in a new tab)
Watson beats champions Ken Jennings and Brad Rutter on TV, answering natural-language clues. Source for IBM Watson wins Jeopardy! (opens in a new tab)
Apple puts a voice assistant in millions of pockets; Google Now, Cortana and Alexa follow. Source for Siri ships on the iPhone 4S (opens in a new tab)
GPUs and big data make neural networks win: AlexNet, AlphaGo and the transformer.
A network on 16,000 CPU cores learns to recognize cats from unlabeled YouTube frames, showing the power of scale. Source for Google Brain's "cat neuron" (opens in a new tab)
Krizhevsky, Sutskever and Hinton's GPU-trained network cuts top-5 error to 15.3% vs 26.2% for the runner-up, starting the deep learning boom. Source for AlexNet wins ImageNet (opens in a new tab)
Google's word embeddings capture meaning as vectors (king - man + woman ≈ queen), a building block for later language models. Source for word2vec (opens in a new tab)
One network learns several Atari games from raw pixels and the score, launching deep reinforcement learning. Source for DeepMind's DQN plays Atari (opens in a new tab)
Ian Goodfellow's GANs pit two networks against each other to generate realistic images, the start of AI image generation. Source for Generative adversarial networks (opens in a new tab)
A voice assistant moves into the home as a smart speaker. Source for Amazon Echo and Alexa (opens in a new tab)
Founded as a nonprofit AI research lab by Sam Altman, Elon Musk, Ilya Sutskever, Greg Brockman and others; it later builds GPT and ChatGPT. Source for OpenAI is founded (opens in a new tab)
Microsoft Research's residual networks train 152 layers deep and win ImageNet 2015 with error below the human benchmark. Source for ResNet (opens in a new tab)
DeepMind's AlphaGo wins 4-1 against one of the world's best Go players, a decade earlier than many experts expected. Source for AlphaGo beats Lee Sedol (opens in a new tab)
Eight Google researchers introduce the transformer, the architecture behind GPT, Claude, Gemini and nearly every modern AI model. Source for "Attention Is All You Need" (opens in a new tab)
Learning only from self-play, AlphaZero masters chess, shogi and Go within a day of training and beats the top programs. Source for AlphaZero (opens in a new tab)
OpenAI shows that pretraining a transformer on lots of text, then fine-tuning, beats task-specific models: the GPT recipe. Source for GPT-1 (opens in a new tab)
Google's BERT tops language benchmarks and soon improves Google Search results. Source for BERT (opens in a new tab)
OpenAI withholds the full 1.5B-parameter model over misuse fears, then releases it in November, starting the debate on releasing powerful models. Source for GPT-2 and staged release (opens in a new tab)
Yoshua Bengio, Geoffrey Hinton and Yann LeCun receive the 2018 Turing Award as the "godfathers of deep learning". Source for Turing Award for deep learning (opens in a new tab)
The first intergovernmental AI standard; later the basis of the G20 AI principles. Source for OECD AI Principles (opens in a new tab)
Large language and image models reach the public: GPT-3, DALL-E, ChatGPT and GPT-4.
A 175-billion-parameter model that writes, translates and codes from a few examples in the prompt; OpenAI sells access through an API. Source for GPT-3 (opens in a new tab)
DeepMind's AlphaFold 2 predicts protein structures about as accurately as lab experiments at the CASP14 competition. Source for AlphaFold 2 solves protein folding (opens in a new tab)
OpenAI generates images from text prompts ("an armchair in the shape of an avocado"). Source for DALL-E (opens in a new tab)
An AI pair programmer autocompletes code in the editor, the first mass-market AI coding tool. Source for GitHub Copilot (opens in a new tab)
Stability AI releases an open-weights image model anyone can run on a home GPU; with Midjourney, AI art goes mainstream. Source for Stable Diffusion (opens in a new tab)
OpenAI's chatbot reaches an estimated 100 million users in about two months, then the fastest-growing consumer app ever, and starts the generative AI race. Source for ChatGPT launches (opens in a new tab)
Meta's model weights spread online within a week, kick-starting the open-weights model ecosystem. Source for Meta's LLaMA (opens in a new tab)
OpenAI's multimodal model scores around the top 10% on a simulated bar exam. Source for GPT-4 (opens in a new tab)
Anthropic, founded in 2021 by former OpenAI researchers, releases its first Claude model the same day as GPT-4. Source for Claude (opens in a new tab)
Thousands of signatories, including Elon Musk and Yoshua Bengio, call for a six-month pause on training models beyond GPT-4. Source for "Pause Giant AI Experiments" letter (opens in a new tab)
President Biden's order requires developers of the most powerful models to share safety test results with the government; it was revoked on January 20, 2025. Source for US Executive Order 14110 (opens in a new tab)
28 countries and the EU sign the Bletchley Declaration on frontier AI risks; the UK and US set up AI safety institutes. Source for Bletchley Park AI Safety Summit (opens in a new tab)
Google DeepMind's natively multimodal model family replaces Bard and PaLM. Source for Google Gemini (opens in a new tab)
Models that think before they answer, use computers and write code on their own.
OpenAI shows minute-long, realistic videos generated from text. Source for Sora video preview (opens in a new tab)
The world's first broad AI law bans some uses and sets risk-based rules, phased in from 2025 to 2027. Source for EU AI Act enters into force (opens in a new tab)
A model trained to think step by step before answering jumps ahead on math, science and coding problems. Source for OpenAI o1 reasoning model (opens in a new tab)
John Hopfield and Geoffrey Hinton win for foundational discoveries that enable machine learning with neural networks. Source for Nobel Prize in Physics for neural networks (opens in a new tab)
Demis Hassabis and John Jumper share the prize (with David Baker) for protein structure prediction. Source for Nobel Prize in Chemistry for AlphaFold (opens in a new tab)
Anthropic's model controls a computer by looking at the screen and moving the mouse, an early general-purpose AI agent. Source for Claude computer use (opens in a new tab)
Anthropic's open standard for connecting AI apps to tools and data is soon adopted by OpenAI, Google and Microsoft. Source for Model Context Protocol (opens in a new tab)
A Chinese open-weights reasoning model rivals o1 at a fraction of the reported cost; Nvidia loses about $589 billion in market value on January 27. Source for DeepSeek-R1 (opens in a new tab)
OpenAI, SoftBank and Oracle announce up to $500 billion for AI datacenters in the US. Source for Stargate announced (opens in a new tab)
Anthropic previews an agentic coding tool that works in the terminal alongside Claude 3.7 Sonnet; coding agents become the breakout AI use case. Source for Claude Code (opens in a new tab)
Anthropic's Claude 4 models can work on long coding tasks for hours with extended thinking and tool use. Source for Claude Opus 4 and Sonnet 4 (opens in a new tab)
An advanced Gemini Deep Think officially reaches gold-medal standard at the International Mathematical Olympiad (35/42); OpenAI reports a gold-level result the same week. Source for AI wins IMO gold (opens in a new tab)
OpenAI's unified model routes between fast answers and deeper reasoning (74.9% SWE-bench Verified). Source for GPT-5 (opens in a new tab)
Google's Gemini 3 Pro takes first place on LMArena (1501 Elo). Source for Gemini 3 (opens in a new tab)
Anthropic releases Claude Opus 5. Source for Claude Opus 5 (opens in a new tab)
OpenAI's first model to reach its internal "Critical" cybersecurity threshold rolls out in phases, cyber partners first. Source for GPT-6 Astra (opens in a new tab)
Google rolls out Gemini 4 Argon globally, including in the Gemini app. Source for Gemini 4 Argon (opens in a new tab)
METR finds Anthropic is likely improving its models about twice as fast as the historical rate, mostly due to AI (with substantial uncertainty). Source for AI speeds up AI research (opens in a new tab)
Choose one of six eras, from the birth of AI in 1943 to today's reasoning models and agents, or view them all.
Show only research breakthroughs, products, policy or industry events.
Type a name, model or year, such as Turing, AlphaGo or 1997.
Copy a link to your filtered view, or copy the milestones as text for notes or slides.
AI became a field at the Dartmouth workshop in the summer of 1956; John McCarthy coined the term "artificial intelligence" in the 1955 proposal for it. Its roots go back further: McCulloch and Pitts described an artificial neuron in 1943, and Alan Turing asked "Can machines think?" in 1950.
John McCarthy is usually called the father of AI because he named the field and organized the 1956 Dartmouth workshop. Alan Turing, Marvin Minsky, Allen Newell and Herbert Simon are also counted among its founders. Geoffrey Hinton, Yoshua Bengio and Yann LeCun are called the godfathers of deep learning.
AI winters were periods when funding and interest collapsed after AI failed to meet expectations. The first ran from about 1974 to 1980, after the 1973 Lighthill report and DARPA cuts. The second ran from about 1987 to the early 1990s, after the expert-system and Lisp machine market crashed.
A common split is: the birth of AI (1943-1972), expert systems and AI winters (1973-1992), machine learning (1993-2011), deep learning (2012-2019), generative AI (2020-2023), and reasoning models and agents (2024 to today).
Three things came together after 2012: GPUs that train neural networks quickly, huge datasets from the internet, and better methods, above all the transformer architecture from 2017. Scaling transformers up led to GPT-3 in 2020 and ChatGPT in 2022.
Milestones and sources were checked on November 14, 2026. Era boundaries are a common convention, not official dates. AI Wins is not affiliated with any company or person listed.
What comes next: compare forecasts for when AGI arrives.
Open toolWatch a tiny neural network learn, step by step.
Open toolPlain-English definitions of the AI terms in this timeline.
Open toolTest what you know about how AI works.
Open toolAI history is being written every week. Get the AI news that matters in the free daily AI Wins digest.
Get the free digestOnly positive AI stories, delivered daily or weekly. Free, no spam, unsubscribe anytime.