From the first artificial neuron in 1943 to today's AI agents
As the first computers were being built, neuroscientists and mathematicians tried to describe how the brain works in the language of logic and mathematics.
McCulloch and Pitts model the neuron as a logic circuit.
They treated a nerve cell as a switch that turns on once its input passes a threshold and stays off otherwise, and showed that networks of such neurons could compute any logical operation.
Donald Hebb: connections grow stronger with use.
In The Organization of Behavior, Hebb proposed that the connection between two neurons strengthens when they are active together, an idea usually summed up as "neurons that fire together wire together."
A handful of scientists believed machines could do more than calculate, that intelligence itself could be built. They gave the new field a name.
Alan Turing asks, "Can machines think?"
In "Computing Machinery and Intelligence," Turing proposed the imitation game: if you cannot tell whether you are talking to a person or a machine, why not say the machine thinks?
The term "artificial intelligence" is born.
John McCarthy, Marvin Minsky, Claude Shannon and others gathered for a summer, confident they could make real progress in a few months. That optimism would last for decades.
The first program to prove mathematical theorems on its own.
Built by Newell, Simon and Shaw, it proved 38 of the first 52 theorems in chapter 2 of Principia Mathematica, some more elegantly than the original book.
A machine that learns, and a language for AI.
Frank Rosenblatt built the perceptron, an artificial neuron that learns by adjusting its weights whenever it gets an answer wrong. The same year John McCarthy designed LISP, which would be the standard language of AI research for decades.
Arthur Samuel's checkers program.
At IBM, Samuel wrote a checkers program that improved by playing against itself until it beat its own creator. He called this the ability to learn without being explicitly programmed: machine learning.
Many expected machines that think like people to arrive any day. Then voices pointing out the limits began to grow louder.
The first expert system, built on a chemist's knowledge.
Feigenbaum, Lederberg and colleagues at Stanford wrote a program that inferred molecular structures from mass spectrometry data. Instead of general intelligence, it captured deep expertise in one narrow field.
MIT's chatbot that played a therapist.
ELIZA talked using nothing more than simple pattern substitution, yet people poured their hearts out to it. Its creator, Joseph Weizenbaum, was disturbed by how readily they did.
The first mobile robot to decide its own actions.
SRI's Shakey looked around with a camera, made a plan and moved on wheels. Along the way it gave us A*, the pathfinding algorithm still used in map apps and video games.
Minsky and Papert prove the perceptron's limits.
They showed that a single-layer perceptron cannot solve even a simple problem like XOR. Neural network research was pushed out of the mainstream for years.
The promises were big and the computers were slow. When results fell short, the funding dried up and the first "AI winter" set in.
Write down the rules of logic, and it finds the answer.
A logic programming language born in Marseille. You describe what is true rather than how to compute it, and the computer does the reasoning. It became widely used in AI research in Europe and Japan.
Stanford's system for diagnosing infections.
Using a few hundred rules, MYCIN diagnosed bacterial infections and recommended antibiotics. In evaluations it performed on par with specialists, but questions of responsibility and ethics kept it out of real clinics.
Results fall short, and the money stops.
In 1973 Britain's Lighthill Report concluded that AI had failed to deliver on its promises, and research funding in the UK and US was cut sharply. Real-world problems had far too many possibilities for the computers of the day.
Systems that turned expert knowledge into rules were sold to businesses, and AI briefly came back to life. Meanwhile, in the lab, neural networks were quietly reviving.
Expert knowledge as if-then rules.
DEC's XCON automatically configured computer orders and saved the company tens of millions of dollars a year. Businesses rushed to adopt expert systems, but every rule still had to be written and maintained by hand.
Kunihiko Fukushima's network modeled on the visual cortex.
Inspired by studies of the cat's visual cortex, Fukushima proposed a layered network that builds complex features out of simple ones. It is the direct ancestor of the convolutional neural network (CNN).
A neural network that remembers, explained through physics.
John Hopfield treated a neural network as a physical system that settles into its lowest-energy state, and used it to build associative memory that recalls a whole pattern from a fragment. This work led to the 2024 Nobel Prize in Physics.
A ten-year national bet on AI.
Japan's Ministry of International Trade and Industry poured money into building next-generation computers that would reason with logic. Alarmed, the US and Europe raised their own AI budgets. The project ended without reaching its goals.
A way to train networks with many layers.
A paper by Rumelhart, Hinton and Williams made backpropagation widely known. By passing the output error backward and adjusting each layer's weights, multilayer networks could finally solve problems like XOR.
The market for expert systems and LISP machines collapses.
Cheap PCs and workstations replaced expensive AI hardware, and expert systems that had grown to thousands of rules became impossible to maintain. Investment froze again.
Yann LeCun trains a convolutional network with backpropagation.
At Bell Labs, LeCun trained a convolutional neural network to recognize handwritten zip codes. It was one of the first times a neural network solved a real industrial problem.
Research carried on under names like "machine learning" and "statistical learning" rather than "AI." In the meantime, a machine beat the world chess champion.
A backgammon program that taught itself.
Gerald Tesauro's neural network at IBM played hundreds of thousands of games against itself and approached the level of the world's best players. It found moves no human had considered and changed how experts played.
A classifier that separates data with the widest possible margin.
Cortes and Vapnik's SVM had solid theory and strong results, and for more than a decade it was used more widely than neural networks.
An IBM computer defeats world chess champion Garry Kasparov.
It was less a triumph of learning than of raw computation, searching hundreds of millions of positions per second. Still, a machine beating the best human carried enormous symbolic weight.
A recurrent network with a long memory.
Hochreiter and Schmidhuber used "gates" to solve the problem of recurrent networks forgetting information from far back in a sequence. LSTM became the core of speech recognition and translation in the 2010s.
A convolutional network that read real bank checks.
LeCun's LeNet-5 was used in the US to read handwritten digits on checks. The MNIST dataset released alongside it became the "Hello, World" of machine learning for decades.
The internet piled up data, and graphics chips built for games grew more powerful. The stage was set for neural networks to return.
The first mass-market robot in the home.
iRobot's vacuum cleaner found its way around the house with a few sensors and simple rules. Without any grand intelligence, it sold millions of units and opened up the home robot market.
A self-driving car finishes 212 km of desert.
In the first race in 2004 not a single car reached the finish line. A year later Stanford's Stanley came in first. Members of that team went on to lead Google's self-driving project, now Waymo.
Hinton shows how to train deep networks.
By pretraining one layer at a time, Hinton showed that even deep neural networks could be trained. The name "deep learning" spread from here.
NVIDIA opens up the GPU for general computing.
Graphics chips made for games could now run ordinary programs. With thousands of cores working in parallel, GPUs turned out to be a perfect fit for training neural networks.
Fei-Fei Li's massive image dataset.
More than 14 million labeled images across over 20,000 categories. It became the arena that proved data matters as much as algorithms.
GPUs and big data woke neural networks up. Images, speech, translation, video games and finally Go fell one after another.
A quiz show victory, and a voice assistant in your pocket.
IBM's Watson beat human champions on the US quiz show Jeopardy!, and the same year Apple put Siri into the iPhone.
Google Brain learns a concept without labels.
Trained on 10 million frames from YouTube videos across 16,000 CPU cores, the network developed a neuron that responded to cats, even though no one had ever told it what a cat was.
Deep learning crushes the ImageNet competition.
A CNN trained on two GPUs reached an error rate of 15.3%, far ahead of the runner-up at around 26%. Computer vision moved to deep learning almost overnight.
Words as vectors, meaning as arithmetic.
Tomas Mikolov's team at Google turned words into vectors with hundreds of dimensions. Meaning could now be added and subtracted: king − man + woman ≈ queen.
Reinforcement learning from nothing but the screen.
DeepMind's DQN taught itself dozens of Atari games using only the pixels and the score, and beat human players at many of them. Google bought DeepMind the following year.
Two networks learn by trying to fool each other.
Ian Goodfellow's generative adversarial network pits a generator that makes fake images against a discriminator that tries to spot them. Within a few years they were producing faces hard to tell from real photographs.
Networks that read a sentence and write one.
Seq2seq compressed a sentence into a vector and unpacked it in another language. Attention let the model decide for itself which part of the source to look at while translating.
A 152-layer network beats human accuracy.
Microsoft researchers added "shortcut connections" that skip layers, making very deep networks trainable. ResNet's ImageNet error rate of 3.6% was lower than the often-quoted human level of about 5%.
Lee Sedol is defeated 4–1.
Go, long thought too vast for machines to master, fell to a combination of deep learning and reinforcement learning. Lee's move 78 in game four, the "divine move," won the only official game a human has ever taken from AlphaGo.
Google Translate changes overnight.
Google Translate switched from phrase-based statistical translation to neural translation, and quality improved visibly across many languages.
"Attention Is All You Need."
Dropping recurrent networks that read word by word, the Transformer looked at the whole sentence at once using attention alone, and set new records in translation. Because it parallelizes easily, it was ideal for scaling up.
Give it the rules, and it becomes a master.
With no human games to learn from, only self-play, AlphaZero mastered Go, chess and shogi in hours and beat the strongest program in each.
Large pretrained language models arrive.
OpenAI's GPT-1 and Google's BERT made it standard practice to first train on vast amounts of text, then fine-tune for a specific task.
"Too dangerous to release."
GPT-2 could write long, convincing passages, and OpenAI released it in stages, citing the risk of misuse. The dangers of language models became a public debate for the first time.
Scale the models up, and abilities nobody predicted appear. AI now converses, draws, does science, and has started doing the work itself.
175 billion parameters, and new tasks from a few examples.
Scaling laws showed that performance rises predictably as models, data and compute grow. GPT-3 proved the point: with no extra training, it translated, summarized and wrote code from a handful of examples in the prompt.
A 50-year-old protein-folding problem, solved.
At CASP14, AlphaFold 2 predicted the 3D structures of proteins with accuracy rivaling experiments. More than 200 million predicted structures have since been released for researchers worldwide.
Pictures from words, code from comments.
OpenAI released DALL·E, which turns text into images, and CLIP, which understands images and text together. GitHub Copilot began suggesting the next lines of code inside developers' editors.
DALL·E 2, Midjourney, Stable Diffusion.
Diffusion models that produce photorealistic pictures from a few words arrived in a flood. Stable Diffusion, with its weights released openly, could run on anyone's computer, and fierce debates over copyright and creativity followed.
100 million users in two months.
A language model taught to converse through human feedback (RLHF) was put in a chat window and released on November 30. It became the fastest-growing consumer service in history and set off the AI race.
GPT-4, Claude, Llama, Gemini.
In a single year came GPT-4, which could also understand images, Anthropic's Claude, Meta's openly released Llama and Google's Gemini.
Governments sit down together to discuss AI risk.
The first AI Safety Summit was held at Bletchley Park in the UK, where major countries signed the Bletchley Declaration. The summit has since been held each year in a different country.
A minute of video from a sentence.
OpenAI's Sora stunned viewers by producing remarkably natural video from text alone. After images, generation spread to video, voice and music.
The world's first comprehensive AI law.
The European Union's law, which sorts AI systems by risk and regulates them accordingly, came into force in August. Some uses are banned outright, and strict obligations for high-risk AI are phased in over time.
AI that "thinks" before it answers.
Starting with OpenAI's o1, models appeared that work through a long chain of thought before answering. Performance on math, science and coding jumped, opening a new direction for scaling: spending more compute at inference time.
Physics and chemistry both go to AI research.
The physics prize went to Hopfield and Hinton for laying the foundations of neural networks. The chemistry prize went to David Baker for protein design and to Demis Hassabis and John Jumper for AlphaFold.
An open reasoning model shakes the market.
China's DeepSeek released a reasoning model close to the frontier, weights and all, and said it had been trained at a fraction of the usual cost. The shock was big enough to send NVIDIA's stock sharply down in a single day.
From AI that answers to AI that works.
Coding agents like Claude Code and Codex began reading, editing and testing code on their own. As standards for connecting AI to outside tools, such as MCP, spread, agents moved into browsing, research and workflow automation.
AI solves olympiad problems in natural language.
Models from Google DeepMind and OpenAI wrote out proofs to International Mathematical Olympiad problems in plain language, as human contestants do, and scored at gold-medal level. Just a few years earlier, this had been one of AI's weakest areas.