Seventy years of trying
From a summer workshop in 1956 to models that write code — including the two winters in between.
The field has been declared dead twice
AI's history is not a steady climb. It runs in cycles: a breakthrough, extravagant promises, funding, disappointment, collapse. The two collapses are known as AI winters, and researchers who lived through them are notably calm about the current excitement.
Knowing the pattern is useful. It tells you which claims to discount, and it explains why progress that looks sudden is usually thirty years of groundwork meeting cheap hardware.
Twelve moments
Enough to follow any conversation about how the field arrived here.
-
1950 Foundations
Turing asks whether machines can think
Alan Turing sidesteps the philosophy and proposes a test instead: if a person conversing by text cannot tell the machine from the human, the question of "thinking" has lost its bite. The paper also answers nine objections he expected to hear, most of which people still raise.
-
1956 Foundations
The field gets its name
A summer workshop at Dartmouth College brings together McCarthy, Minsky, Shannon and Rochester. Their proposal coins the phrase "artificial intelligence" and predicts serious progress in two months. The optimism was off by several decades, but the name stuck.
-
1958 Foundations
The perceptron
Frank Rosenblatt builds a machine that adjusts its own weights to classify simple patterns. It is the direct ancestor of every neural network running today, and the first system that learned from examples rather than following written rules.
-
1966 Foundations
ELIZA, and the first illusion
Joseph Weizenbaum writes a 200-line program that reflects your sentences back as questions. People confided in it. Weizenbaum spent the rest of his career warning that our readiness to see a mind is the real finding.
-
1974 Winter
The first AI winter
Promised results do not arrive. British and American funders read critical reports, conclude the field has oversold itself, and cut the money. Research continues quietly under other names.
-
1980s Winter
Expert systems sell, then stall
Companies buy rule-based systems that encode a specialist's knowledge as thousands of if-then statements. They work in narrow domains and collapse outside them. Maintaining the rules costs more than the systems save, and a second winter follows.
-
1997 Revival
Deep Blue beats Kasparov
IBM's chess machine wins a match against the reigning world champion. It is a triumph of search and specialised hardware rather than learning — Deep Blue understood nothing about chess that its programmers had not put there.
-
2012 Revival
AlexNet, and the deep learning era begins
A deep neural network trained on two gaming graphics cards cuts the error rate on the ImageNet photo-recognition contest by roughly ten points in a single year. The recipe — big network, big dataset, GPUs — becomes the template for everything after.
-
2016 Revival
AlphaGo wins in Seoul
DeepMind's system beats Lee Sedol at Go, a game with more legal positions than atoms in the observable universe, so brute-force search cannot work. Move 37 of game two was a play no human would choose, and it was right.
-
2017 Language
Transformers
A paper titled "Attention Is All You Need" drops the sequential processing that made language models slow and replaces it with attention, which weighs every word against every other word at once. Training becomes parallel, and therefore large.
-
2020–2022 Language
Scale, and things nobody trained for
Making transformers bigger and feeding them more text produces abilities the training never targeted: translation, summarising, arithmetic, code. AlphaFold predicts protein structures at a standard that reshapes biology. Image generators arrive in the same window.
-
Now Language
Ordinary infrastructure
AI stops being a demonstration and becomes plumbing — in phones, clinics, banks and classrooms. The open questions shift from "can it?" to "who is accountable when it is wrong?"
What actually changed in 2012
The ideas behind deep learning were largely in place by 1990. Backpropagation was published in 1986. Convolutional networks read handwritten cheques in the 1990s. Nothing conceptual was missing.
What arrived was scale: the internet produced labelled datasets of millions of images, and gaming graphics cards turned out to be near-perfect for the arithmetic that neural networks need. Old ideas plus new hardware plus enormous data. That is the entire recipe, and it is still the recipe.