From bronze giants to chatbots.
People have dreamed of thinking machines for almost 3,000 years. This is how the dream turned into the AI on your phone. You can try three of the old machines yourself along the way.
Dreams of thinking machines
Long before computers, people told stories about machines that think. A few tried to turn thinking into steps a machine could follow.
- c. 700 BC
Talos, the bronze guard
Greek myths told of Talos, a giant bronze man who walked around the island of Crete and threw rocks at ships. It is one of the oldest stories about a made thing that acts on its own.
- c. 1305
Llull's paper wheels
Ramon Llull, a thinker from Majorca, made paper wheels with words written around them. Turning the wheels lined up new statements. He hoped a method like this could settle arguments for good.
- 1670s
"Let us calculate"
Gottfried Leibniz built a machine that could multiply. He dreamed of a language so exact that two people who disagreed could just sit down and say: let us calculate.
- 1843
Ada Lovelace's notes
Ada Lovelace wrote notes on Charles Babbage's planned Analytical Engine. She saw that it could work with symbols and even music, not just numbers. She also wrote that it could never come up with anything truly new. People still argue about that.
- 1854
Logic becomes math
George Boole showed that true-or-false thinking can be written as simple math. Every computer today runs on his idea, in the form of ones and zeros.
- 1921
The word "robot"
Karel Čapek's play R.U.R. opened in Prague and gave the world the word robot. It comes from a Czech word for forced work. The robots in the play rise up against people, a story told many times since.
The first electronic brains
Math, war and the first computers gave people a real question to test: can a machine think?
- 1936
A machine on paper
Alan Turing imagined a simple machine that reads and writes symbols on a long tape. He showed it could carry out any step-by-step method at all. Every computer since is a real version of that idea.
- 1943
A brain cell in math
Warren McCulloch and Walter Pitts wrote down a simple model of a brain cell. It switches on when enough of its inputs push it. Wire enough of them together and they can do logic. Today's neural networks grew from this.
- 1949
Cells that fire together
Donald Hebb suggested that the brain learns by making a link stronger each time two cells fire together. It was one of the first clear ideas of how a machine might learn.
- 1950
The imitation game
Turing asked "can machines think?" and then swapped it for a question he could test. Can a machine chat by typing so well that a judge cannot tell it from a person? We now call this the Turing test.
- 1951
Games become the test
Christopher Strachey wrote a checkers program, and Dietrich Prinz a chess puzzle solver, on early British computers. Games have clear rules and a clear winner. That made them the favorite test for AI for the next 65 years.
- 1952–59
A program that practiced
Arthur Samuel's checkers program at IBM got better by playing against copies of itself. In 1959 he gave this a name: machine learning.
AI gets its name
A summer workshop named the field. Early wins made experts promise human-level AI within 20 years.
- 1956
The Dartmouth workshop
John McCarthy, Marvin Minsky, Claude Shannon and Nathaniel Rochester asked for money for a summer study. In the request, McCarthy used the words artificial intelligence. Their workshop at Dartmouth College is seen as the birth of the field.
- 1956
Logic Theorist
Allen Newell, Herbert Simon and Cliff Shaw wrote a program that proved theorems from a famous math book. For one of them, it found a shorter proof than the book had.
- 1958
The perceptron
Frank Rosenblatt built a machine that learned to sort shapes by seeing examples. The New York Times reported that the Navy expected it to walk, talk and know it exists. It could not. Try the idea yourself below.
try itTrain a perceptron (1958)
The line is the perceptron's guess: blue on one side, orange on the other. Press learn. Each dot on the wrong side nudges the line toward it.
11 of 12 dots are on the wrong side. Press learn.
Load a set:
Or tap the square to add your own dots:
the line: 0.25x − 1.00y + 0.35 = 0
- 1958
Lisp
McCarthy created Lisp, a programming language built for AI work. Many ideas that started in Lisp are now in the languages people use every day.
- 1966
ELIZA, the first chatbot
Joseph Weizenbaum wrote ELIZA, a program that chatted like a therapist. It understood nothing. It spotted key words and turned your own words back into questions. Many people still felt understood, which worried Weizenbaum for the rest of his life.
try itTalk to ELIZA (1966)
Tell it something about your day. Then turn on "show how" to see the trick behind each reply.
ELIZAHow do you do. Please tell me your problem. - 1966–72
Shakey the robot
Shakey, built at SRI in California, was the first robot that could plan its own actions. It took pictures, worked out a plan and rolled from room to room to carry it out. The A* search method, still used in maps and games, came from this project.
- 1969
The book that froze neural networks
Minsky and Seymour Papert published Perceptrons. It proved that a single-layer perceptron cannot learn some simple patterns, like "one or the other, but not both." Money for neural network research mostly dried up for over a decade.
- 1970
SHRDLU and the block world
Terry Winograd's program moved blocks around a pretend table when you typed plain English, like "put the red block on the blue one." It was amazing inside its tiny world. Outside it, the tricks fell apart.
The first AI winter
The promises ran far ahead of the results. The people paying for the research lost patience.
- 1966
Translation falls short
A US report found that machine translation was slower, less accurate and more expensive than human translators. Government money for it was cut.
- 1973
The Lighthill report
James Lighthill told the UK government that AI had not come close to its grand goals. Britain cut most of its AI research money soon after.
- 1970s
Too many choices
Programs that solved toy problems choked on real ones. Each extra step multiplied the number of choices to check, until no computer could keep up. This wall still shapes how AI is built.
- 1974
The money runs out
DARPA, the US military research agency, sharply cut its money for open-ended AI research. The lean years that followed became known as the first AI winter.
The expert systems boom
AI came back as a business: programs packed with rules written down by human experts.
- 1980
XCON earns its keep
At the computer company DEC, a program called XCON worked out the parts for each customer order. It used thousands of if-then rules taken from experts. By the mid-1980s it was said to save DEC about $25 million a year.
- 1982
Japan's big bet
Japan started a ten-year plan to build computers that reason, called the Fifth Generation project. The US and Britain rushed to start AI projects of their own.
- 1982
Networks that remember
John Hopfield showed a network that stores memories and can rebuild a whole pattern from a small piece of it. In 2024 he shared the Nobel Prize in Physics for this kind of work.
- 1984
Cyc, common sense by hand
Doug Lenat began typing everyday knowledge into a computer, one fact at a time. Facts like "water makes things wet." The project ran for decades and showed how much people know without noticing.
- 1986
Backpropagation
David Rumelhart, Geoffrey Hinton and Ronald Williams made famous a way for networks with many layers to learn. Each mistake is sent backward through the network, and every link gets nudged a little. This got past the wall from the Perceptrons book.
The second AI winter
The rule-based systems turned out to be costly to keep running, and the boom ended again.
- 1987
Special AI computers lose
Companies had built costly computers made just for running Lisp. Cheaper desktop computers from Apple and IBM caught up with them. That whole market fell apart in about a year.
- early 1990s
Brittle rules
Expert systems broke on unusual cases, and keeping thousands of rules up to date was slow and costly. Many companies quietly stopped using them.
- 1992
The Fifth Generation ends
Japan's project closed without reaching its goals. For a while, AI was a word that scared away investors. Researchers called their work other names to get it paid for.
Quiet progress
AI kept improving under other names: search, speech, spam filters, recommendations. Then it beat the world champions.
- 1989–98
Networks learn to read
Yann LeCun trained networks that look at small patches of a picture at a time. By the late 1990s, systems like his read the handwritten amounts on millions of US bank checks.
- 1992
TD-Gammon
Gerald Tesauro's backgammon program taught itself by playing against itself. It reached the level of the best human players, and even changed how experts play some moves.
- 1997
Deep Blue beats Kasparov
IBM's Deep Blue beat world chess champion Garry Kasparov in a six-game match. It did not think like a person. It checked about 200 million positions every second.
- 1997
Networks with memory
Sepp Hochreiter and Jürgen Schmidhuber designed a network that can remember things over long sequences, called LSTM. Years later it powered speech recognition and translation on phones.
- 2005
A car drives itself across the desert
In 2004, no robot car finished DARPA's desert race. In 2005, Stanford's car Stanley won it, driving 212 km (132 miles) with no one at the wheel.
- 2009
ImageNet
Fei-Fei Li's team built ImageNet, millions of photos each labeled with what is in them. Its yearly contest gave researchers a fair way to compare their methods. It set up the next big leap.
- 2011
Watson wins Jeopardy!
IBM's Watson beat the two greatest champions of the TV quiz show Jeopardy!. It had to untangle puns and tricky clues, then buzz in faster than the humans.
- 2011
Siri
Apple put Siri, a voice assistant, in the iPhone 4S. Talking to your phone went from science fiction to normal in a few years.
Deep learning takes off
Three things met at last. There was lots of data, fast chips made for games, and an old idea: networks that work a bit like a brain.
- 2012
AlexNet
Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton entered a deep neural network in the ImageNet contest. It got about 15% of its answers wrong, while the next best got 26% wrong. The chips that draw video game graphics made it fast enough to train.
- 2012
The cat detector
A Google network watched millions of frames from YouTube videos. Without being told what a cat was, it ended up with a part that reacted to cats. Nobody wrote a cat rule.
- 2014
Two networks in a contest
Ian Goodfellow set two networks against each other. One makes fake pictures, and the other tries to catch the fakes. Both keep getting better, which led to the first AI-made faces that looked real.
- 2016
AlphaGo
DeepMind's AlphaGo beat Lee Sedol, one of the best Go players alive, four games to one. Go has more possible board positions than there are atoms in the universe. Its move 37 in game two surprised every expert watching.
- 2017
Attention is all you need
Google researchers described the transformer. It reads all the words of a text at once and works out which words matter to each other. Nearly every chatbot today is built on it.
- 2017
AlphaZero
Given only the rules, AlphaZero taught itself chess, shogi and Go by playing against itself. It then beat the strongest programs in all three games.
Machines that write
Feed a transformer a huge share of the internet and teach it to guess the next word. Surprising things happen.
- 2018
GPT and BERT
OpenAI's GPT and Google's BERT learned language by reading huge amounts of text. GPT practiced guessing the next word. BERT practiced filling in missing words. Try a tiny next-word guesser below.
try itGuess the next word
This tiny model read the text below and counted which word comes after which. Pick its next word, or let it pick by chance.
the
After "the", it has seen:
A chatbot does the same job with a far bigger memory. It looks back at thousands of words, not one, and it learned from a huge share of everything people have written online.
- 2019
Too risky to share?
OpenAI first held back the full GPT-2, worried it could flood the web with fake news. It released the model in stages over the year. The fight over how open AI should be had begun.
- 2020
GPT-3
GPT-3 had 175 billion adjustable numbers inside it. Shown just a few examples, it could write essays, simple code and poems, with no extra training.
- 2020
AlphaFold
DeepMind's AlphaFold 2 predicted the 3D shapes of proteins almost as well as lab experiments. Biologists had worked on that problem for 50 years. Demis Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry for it.
- 2021–22
Pictures from words
DALL·E, Midjourney and Stable Diffusion made pictures from a typed description. Stable Diffusion was free to download and run. Artists asked hard questions about their work being used to train these models without asking them.
- 2022
ChatGPT
OpenAI released ChatGPT on 30 November 2022. By one estimate it had 100 million users two months later, faster than any app before it. Suddenly everyone was talking to AI.
AI everywhere
Chatbots became tools millions use at work and school. The questions got bigger too.
- 2023
A year of chatbots
OpenAI's GPT-4 scored well on hard exams like the bar exam. Anthropic released Claude, and Google released Bard, later renamed Gemini. Chatbots learned to read pictures as well as text.
- 2023
Models you can run yourself
Meta shared its Llama models so others could run and change them. Many open models followed, from France's Mistral to China's DeepSeek. AI was no longer only in the hands of a few companies.
- 2024
Nobel Prizes for AI
The physics prize went to Hopfield and Hinton for neural networks. The chemistry prize went partly to Hassabis and Jumper for AlphaFold. Ideas once seen as a dead end now had the top prize in science.
- 2024
The first big AI law
The European Union passed the AI Act. It sorts AI uses by risk, bans a few outright and sets rules for the rest. Its rules come in step by step over several years.
- 2024–25
Models that think first
New models learned to work through a problem in steps before answering. They got much better at math, science and code. DeepSeek showed in 2025 that this could be done for far less money than people thought.
- 2025
From answering to doing
AI agents began to take on whole tasks. They search, write and run code, use other tools and check their own work. For many programmers, an agent that works across a whole project became normal.
- today
The open questions
How do we make AI honest and safe? Who owns the work it learned from? What happens to jobs, and to the energy it uses? Nobody has settled these yet. The rules-versus-examples story is still being written.
Sources and further reading
- Nils J. Nilsson, The Quest for Artificial Intelligence (2009)
- Alan Turing, "Computing Machinery and Intelligence" (1950)
- The Dartmouth proposal (1955)
- Joseph Weizenbaum, "ELIZA" (1966)
- The Lighthill report (1973)
- Krizhevsky, Sutskever and Hinton, the AlexNet paper (2012)
- Vaswani et al., "Attention Is All You Need" (2017)
- Reuters on the UBS estimate of ChatGPT users (2023)
- Nobel Prize in Physics 2024
- Nobel Prize in Chemistry 2024
- The EU AI Act, Regulation (EU) 2024/1689
- Wikipedia, History of artificial intelligence
Updated 22 November 2026. Spotted a mistake? Dates in history are often argued over. Where sources disagree, this page says "about" or "by one estimate."