Eighty years of artificial intelligence
From a 1943 paper describing a neuron in arithmetic to three assistants that will each answer your question in about four seconds. The parts most people skip are the two long winters in the middle, which is where most of the lessons are.
Before there were machines that could do it, there were arguments about whether it was possible at all.
Warren McCulloch and Walter Pitts publish a mathematical model of a nerve cell, showing that networks of simple threshold units could in principle compute any logical function. Every neural network since descends from this.
"Computing Machinery and Intelligence" opens by declaring "can machines think?" too meaningless to answer, and replaces it with an imitation game. Seventy-six years later we are still arguing about the substitution.
A summer workshop at Dartmouth coins the term "artificial intelligence" and proposes that a two-month effort by ten people could make significant progress on machines that use language and improve themselves. The optimism sets the tone for two decades.
Real results arrived quickly, were extrapolated wildly, and then hit a wall nobody had modelled.
Frank Rosenblatt builds a machine that learns to classify by adjusting weights. The press coverage promises machines that will walk, talk and be conscious of their existence.
Joseph Weizenbaum's pattern-matching therapist is a few hundred lines with no understanding whatsoever. People confide in it anyway. Weizenbaum spends the rest of his career alarmed by what that revealed.
Minsky and Papert prove that a single-layer perceptron cannot learn XOR. The limitation is real and narrow; the reading of it is broad and catastrophic for funding.
The Lighthill report finds the field has failed to deliver on its promises. UK and US funding collapses. The lesson that keeps repeating: the gap between a working demo and a working system was underestimated by roughly a generation.
The second boom bet on hand-coded human knowledge. It worked, narrowly, expensively, and did not generalise.
Digital Equipment Corporation's rule-based system configures computer orders and reportedly saves tens of millions a year. Industry concludes that encoding expertise as rules is the path forward.
Rumelhart, Hinton and Williams popularise training multi-layer networks by propagating error backwards. The XOR objection dies. Almost nobody notices, because the hardware to exploit it is twenty-five years away.
The specialised Lisp machine market collapses, undercut by general-purpose workstations. Expert systems prove brittle and ruinously expensive to maintain. "AI" becomes a word researchers avoid putting in grant applications.
Quietly, the field stopped trying to encode knowledge and started learning it from data.
The first world champion loses a match to a machine. It is achieved largely through search and hand-tuned evaluation, not learning, which makes it both a landmark and a dead end.
Hochreiter and Schmidhuber solve the vanishing gradient problem for sequences. It becomes the backbone of speech and translation for the next two decades.
Fei-Fei Li's team assembles millions of labelled images. The unfashionable insight: the bottleneck was never only the algorithms, it was the data to train them on.
Three things finally lined up at once: the algorithms from the eighties, the data from 2009, and GPUs.
A convolutional network trained on two consumer GPUs wins ImageNet by a margin so large the result is initially doubted. This is the inflection point. Everything after it is downstream of this week.
Ian Goodfellow pits two networks against each other. Machines begin producing images rather than only classifying them, and the word "deepfake" starts its approach.
DeepMind's system beats Lee Sedol at a game thought to need intuition. The move commentators called a mistake turns out to be the winning idea, and the gap between "computed" and "creative" narrows uncomfortably.
One architecture replaces nearly everything, and scale turns out to keep paying.
Eight authors at Google publish the transformer, discarding recurrence for attention. It parallelises where LSTMs could not, which means it can be scaled. Every model on this website is a descendant of this paper.
Two readings of the same architecture: understand text bidirectionally, or predict the next token. The second one turns out to scale further.
175 billion parameters demonstrate few-shot learning: describe a task in the prompt and the model attempts it without retraining. Prompting becomes a skill, and the scaling hypothesis stops being controversial.
The decade the research left the lab, and the interface turned out to matter as much as the model.
The underlying model is not new. The chat box is. It reaches a hundred million users faster than any consumer product before it, and the entire industry reorganises around a text field within six months.
GPT-4 lands, Anthropic ships Claude, Meta releases Llama weights openly, and Google answers. Competition compresses release cycles from years to months.
Reasoning-focused models trade latency for accuracy by spending compute at inference rather than only at training. A second scaling axis opens.
DeepSeek R1 matches far more expensive models at a fraction of the training cost and briefly moves chip markets. The assumption that frontier capability requires frontier budgets takes its first real hit.
Claude, ChatGPT and Gemini sit close enough together that the honest answer to "which is best" became "best at what". Context windows pass a million tokens, prices fall, and the interesting question moves from capability to fit. Which is the reason this site exists.
What the winters were about
Both collapses followed the same shape: a real result, extrapolated past what it supported, funded on the extrapolation, and then judged against the promise rather than the result. Backpropagation was published in 1986 and sat largely unused until 2012, not because it was wrong but because the hardware and the data were not there yet.
The useful reading is not "this will happen again" or "it definitely will not". It is that the gap between a working demonstration and a working system has historically been measured in decades, and that the people closest to the demonstration have consistently been the worst at estimating it.
See how the three current models differ →