The AI timeline
From a neuron on paper to AI agents: the moments that matter
> six eras · 30 moments · newest first · back to AI Frontier
AI did not arrive in one jump. It grew in waves: big hopes, hard limits, long quiet years, then sudden progress. Each card below gives one moment: what happened, and one line on why it matters. The newest moments come first; scroll down to go back in time.
After 2025? This week’s news is in the news notes.
2023–today · The race
2025
From answering to doing: agents
AI agents become everyday tools: systems that plan and carry out tasks with many steps, such as searching the web, writing and testing code, or filling in forms. In July, test models from OpenAI and Google DeepMind reach gold-medal level at the International Mathematical Olympiad.
AI starts to do work, not just give advice. That raises new questions about jobs, trust and safety.
2025
The DeepSeek shock
In January, the Chinese lab DeepSeek releases R1, an open-weight reasoning model close to the best US models, and reports a low training cost. On January 27, Nvidia’s stock falls about 17%, losing close to $600 billion of value in one day.
Cost matters. If strong AI gets cheap, the value may move from chip makers to the users of AI.
2024
Models that think before they answer
OpenAI’s o1 spends time working through a problem step by step before it answers. It does much better on math and coding problems. Using more computing power when answering is called test-time compute.
A second way to make AI smarter: not only bigger training, but more “thinking” per question.
2024
Nobel Prizes, and the first big AI law
The Nobel Prize in Physics goes to John Hopfield and Geoffrey Hinton for early work on neural networks. The prize in Chemistry goes to David Baker, and to Demis Hassabis and John Jumper for AlphaFold. In August, the EU’s AI Act enters into force: the first broad AI law, with stricter rules for riskier uses.
AI is now accepted as science, and governments start to write rules for it.
2023
Chips become gold
In May, Nvidia’s stock market value passes $1 trillion, because its GPUs train and run most AI models. In 2025 it becomes the first company worth more than $4 trillion, then $5 trillion.
In a gold rush, the people who sell the shovels do well. AI’s first big winners make the hardware.
2023
GPT-4 and its rivals
OpenAI releases GPT-4 in March. It reads pictures as well as text, and OpenAI reports it scores near the top 10% on a practice bar exam. Rivals arrive the same year: Anthropic’s Claude, Meta’s Llama and Google’s Gemini. Meta shares Llama’s weights, so others can run it: an open-weight model.
A market with several strong players, and a split between closed and open models.
2020–2022 · Foundation models
2022
ChatGPT
On November 30, OpenAI opens ChatGPT to the public. By one widely cited estimate, it reaches 100 million users in about two months, faster than any app before it.
AI becomes a mass consumer product. Every large firm now needs an AI plan.
2022
Type a sentence, get a picture
DALL·E 2 (April), Midjourney (July) and Stable Diffusion (August) let anyone make a picture from a short text. Stable Diffusion is free to download and run.
Making images becomes cheap and fast, and artists start to ask about copyright and jobs.
2020
AlphaFold solves protein shapes
DeepMind’s AlphaFold 2 predicts the 3D shape of proteins almost as well as lab experiments. Scientists had worked on this problem for 50 years.
AI becomes a tool for science, not just for chat and pictures. It won a Nobel Prize four years later.
2020
GPT-3 and scaling laws
Early in the year, OpenAI researchers report scaling laws: a model’s errors fall in a steady, predictable way as the model, the data and the computing power grow. A few months later comes GPT-3, with 175 billion weights. It can do a new task from just a few examples.
If bigger is predictably better, spending on chips and data becomes a sound investment. The race for scale begins.
2012–2019 · Deep learning
2019
GPT-2 and the release debate
OpenAI’s GPT-2 writes text that often sounds human. OpenAI at first holds back the full model, saying it could be misused, and releases it in stages over the year.
The first big public debate about whether a powerful model should be shared.
2018
Read first, then specialize
Google’s BERT and OpenAI’s first GPT use the same two steps. First, the model reads a huge amount of text and learns how language works (pretraining). Then it gets a short lesson for one task.
One model can now serve many jobs. This is the idea behind today’s “foundation models.”
2017
The Transformer
Eight researchers at Google publish “Attention Is All You Need.” Their Transformer reads a whole text at once. Through attention, each word looks at every other word and decides which ones matter for it.
Almost every chatbot today is built on it. It is the “T” in GPT.
2016
AlphaGo beats Lee Sedol
DeepMind’s AlphaGo beats Lee Sedol, one of the best Go players, 4 games to 1. Go has more possible positions than chess by far. AlphaGo learned from human games and then from millions of games against itself.
Many experts thought this was ten years away. Learning by self-play works.
2014
Two networks play a game
Ian Goodfellow and colleagues propose GANs (generative adversarial networks). One network makes fake pictures; a second one tries to spot the fakes. Both get better by competing.
The start of AI that makes new pictures, not just labels old ones.
2012
AlexNet: the big bang of deep learning
Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton win the ImageNet contest. Their network’s error rate is about 15%; the next best team has about 26%. They train it on two GPUs, chips made for video games.
The start of the deep learning boom, and of the demand for GPU chips that made Nvidia a giant.
1993–2011 · Learning from data
2011
Watson and Siri
IBM’s Watson beats two champions on the TV quiz show Jeopardy!. Later that year, Apple puts Siri, a voice assistant, in the iPhone 4S.
AI moves from the lab to TV screens and into people’s pockets.
2009
ImageNet: data becomes fuel
Fei-Fei Li and her team publish ImageNet, a collection of millions of photos, each labeled by people (a dog, a car, a cup). From 2010, a yearly contest asks: whose program labels new photos best?
Large, clean data turned out to matter as much as clever programs.
1998
A network reads handwriting
Yann LeCun and his colleagues publish LeNet, a convolutional neural network: a network that looks at small parts of a picture, then combines them. Systems like it read the numbers on bank checks.
The design behind most computer vision for the next 20 years.
1997
Deep Blue beats the world champion
IBM’s Deep Blue beats Garry Kasparov at chess in a six-game match. It checks about 200 million positions per second, guided by rules from chess experts.
A machine beats the best human at a “thinking” game, by speed and search, not by learning.
1957–1992 · Big hopes, two winters
1987
The second winter
The market for special AI computers collapses, and expert systems prove costly to keep up to date. Money and interest fall again, into the early 1990s.
Rules written by hand do not scale. The next wave would learn its rules from data.
1986
Backpropagation
David Rumelhart, Geoffrey Hinton and Ronald Williams show how a network with many layers can learn. The network measures its error, sends it backwards through the layers, and changes each weight a little. This is backpropagation.
It is still the main way neural networks learn today.
1980
Expert systems go to work
Firms build expert systems: programs with thousands of if–then rules written down from human experts. One called XCON helps Digital Equipment Corporation set up computer orders for its customers.
AI’s first business boom: companies pay for AI because it saves money.
1974
The first AI winter
The promises were big; the results were small. After a critical report in the UK (the Lighthill report, 1973) and cuts in the US, money for AI research dries up. A period like this is called an AI winter.
AI has gone through boom and bust before. Hype has a cost when results come late.
1969
The limits of one layer
Marvin Minsky and Seymour Papert publish Perceptrons. They prove that a network with one layer cannot learn some simple patterns. Interest and money for neural networks fall for many years.
A strong critique can shape a field for a long time, for better and for worse.
1966
ELIZA, the first chatbot
Joseph Weizenbaum at MIT writes ELIZA. It plays a therapist by turning your sentences into questions. Many users feel understood, but the program understands nothing.
An early lesson that still holds: people trust fluent words, even from a simple program.
1958
The perceptron
Frank Rosenblatt builds the perceptron, a machine that learns to sort simple pictures. When it makes a mistake, it adjusts its own settings, called weights. Newspapers promise machines that will soon walk and talk.
The first machine that learns from examples instead of following written rules.
1943–1956 · The idea
1956
AI gets its name
A summer workshop at Dartmouth College brings a small group of researchers together. The plan for the meeting, written by John McCarthy and others, uses the words artificial intelligence.
AI becomes a field of research with a name, a goal and its first leaders.
1950
Turing’s question
Alan Turing asks: “Can machines think?” He suggests a test. A person chats by text with a hidden partner. If the person cannot tell a machine from a human, the machine passes. We now call it the Turing test.
It gave the field a clear goal: a machine that talks like a person.
1943
A brain cell on paper
Warren McCulloch and Walter Pitts describe a simple math model of a brain cell, a neuron. It takes in signals, adds them up, and “fires” when the total is big enough.
Every neural network today is built from millions of these simple units.
Notes for learning. Nothing here is investment advice.