Milestones of Artificial Intelligence
The history of artificial intelligence does not run as a steady ascent but in waves: grand promises, bitter disappointments, long dry spells, surprising breakthroughs. Eleven stations are enough to grasp the arc.
The founding years
1950: Turing puts the question differently
Instead of asking whether machines can think, Alan Turing asks whether a machine can hold a conversation in such a way that a human cannot tell it apart from a human.1 The question about the inner life is replaced by a question about behaviour. This manoeuvre shapes the debate to this day.
1956: The field gets its name
At a summer workshop at Dartmouth College, John McCarthy proposes the term "artificial intelligence". The participants are convinced they will solve the essential problems within a few years. This misjudgement repeats itself in every generation.
1958: The perceptron learns
Frank Rosenblatt builds a machine that learns from examples instead of following rules: the perceptron, ancestor of all neural networks. The press reports enthusiastically on a machine that will soon walk, talk and be conscious of itself.
1966: ELIZA and human credulity
Joseph Weizenbaum programs a chatbot that imitates a psychotherapist by mirroring statements back as questions. The program is a few pages long and understands nothing. Test subjects confide personal matters to it all the same. Weizenbaum is so alarmed by this that he becomes the field's best-known critic.2
The sobering up
1969: The limits of the perceptron
Marvin Minsky and Seymour Papert show mathematically what simple perceptrons cannot learn in principle.3 Research funding dries up, the field plunges into the first "AI winter". A second follows in the late 1980s, when industry's expensive expert systems fail to keep their promises.
1986: Networks learn in depth
David Rumelhart, Geoffrey Hinton and Ronald Williams popularise a method that also allows multi-layer networks to be trained: backpropagation of errors.4 With that, Minsky's objection is technically answered. All that is missing is the data and the computing power to demonstrate it. It takes 26 years.
The breakthrough
1997: Deep Blue beats Kasparov
An IBM computer wins against the reigning world chess champion. It does so through raw computing power and pre-programmed chess knowledge, not through learning. The event is symbolically more potent than it is technically.
2012: AlexNet wins the image competition
A deep neural network wins the ImageNet competition by such a margin that the whole field switches course within two years. The reason for the success is not new ideas but graphics cards and millions of labelled images. This is where the present begins.
2016: AlphaGo and the unusual move
A system from DeepMind beats one of the world's best Go players. Go was considered out of reach because it cannot be brute-forced. Move 37 of the second game becomes famous: a move no human would have played and that experts at first took for a mistake. It wins the game.
2017: The transformer
A research group at Google presents a new network architecture that processes word relationships across entire texts in parallel.5 It is the technical foundation of practically every language model in use today. The paper is among the most cited in computer science.
2022: AI goes public
With the release of ChatGPT, a language model becomes usable for the first time by millions of people without specialist knowledge. Technically it is an intermediate step; socially it is the real break. Since then it is no longer only the specialists who discuss AI, but schools, courts, newsrooms and legislators.
What the curve shows
Three patterns recur. First: decades often lie between an idea and its effect, because the preconditions are missing. Second: the leaps rarely come from new theory but from more data and more computing power. Third: every wave of euphoria has so far been followed by a sobering up. Anyone who knows the milestones reads today's promises with the composure they call for.