Where AI came from

Four researchers asked for two months and ten people. The proposal they signed on the last day of August 1955 gave the field its name, and it is careful about what one summer at Dartmouth could deliver. The sentence it rests on is not careful at all. The work would proceed, it says, on the conjecture that every feature of intelligence "can in principle be so precisely described that a machine can be made to simulate it."

Everything since has been evidence about that sentence, and the evidence keeps being read wrong in both directions. Five turns take you from that summer of 1956 to the chatbot on your phone. At each one the question worth asking is not how impressive the thing looked but what it actually settled.

Small worlds, written by hand

For the first twenty-five years or so, making a machine intelligent meant writing down what it should do. Programs proved theorems and answered questions about a tabletop of coloured blocks. Those results were real, and they were real inside worlds made small enough for the rules to cover.

ELIZA is the clearest case, and the part worth having is the choice of subject. Its author picked the psychiatric interview on purpose, as one of the few conversations in which a participant is free to "assume the pose of knowing almost nothing of the real world." A listener who mostly hands your words back is the one partner whose ignorance never shows. Users read a mind into it anyway, and the field's first durable finding was about people rather than programs: the appearance of understanding is cheap to manufacture and expensive to test. Every misread milestone since has some of that in it.

The money, though, was for the promise that the walls would come down soon.

The money leaves, twice

The walls stayed up, and the reckoning came as budgets rather than argument. In 1966 an American committee reviewing machine translation, then the field's flagship application, spent most of its report not on computers but on the economics of translating Russian — what human translators cost, how fast a scientist could simply learn the language. The urgent need the funding rested on did not exist, and American support took about twenty years to return. Seven years later a British survey did the same by another route, rating the practical engineering and the brain-modelling respectable and finding no subject between them to fund. Artificial intelligence, in that reading, was not a field. It was a gap between two other fields.

There is a name for this shape. Expectations climb past what the work supports, someone says so, the money leaves, and the research stops for reasons that have little to do with whether it was working. It is called an AI winter. A winter is a verdict on the promises, not on the work.

It happened again in the 1980s. Researchers warned in 1984 that expectations had run ahead of the results; the industry those expectations were attached to kept growing for two more years before the money turned. The word survives as a standing worry rather than a period. A winter can mean capability stalling, expectations correcting, or capital leaving, and those three have come apart before.

The turn nobody watched

The recovery was not announced. In 2012 a photo-labelling contest published a results table with an odd shape. The entries built the old way, by researchers writing down what the software should look for, were bunched at the top and separated by fractions of a point, which is what a technique looks like once everyone has tried everything. Below them a wide gap, and below that, clear of the whole field, one entry that had been told nothing about what to look for and had worked it out from the photographs.

Nothing there was proof of intelligence. It was proof about one job, done one way. What made it a turn was the width of the margin. A gap that size is not an improvement on the old approach but a verdict that its remaining problems were no longer the interesting ones. Within a few years the same method had taken speech recognition and translation as well.

Two famous games, on opposite sides

The two most public events in this history are board games, filed together and belonging apart. The first needs a step back of fifteen years. Deep Blue, the machine that took the 1997 chess match, was hand-written software of the kind the winters had discredited. People who knew chess told it what a good position looks like, and its edge was that it examined positions at a rate no human approaches. The champion had put his finger on it a year earlier, during a match he went on to win: he was tiring and his opponent was not. So the most famous thing this field has ever done in public was a win for the losing approach. It settled that chess can be searched hard enough to beat anyone alive, and settled nothing that carried to a problem that was not chess.

The Go match of 2016 looked like the same kind of event and was not, because that program had learned what mattered from records of games rather than being told. Within two years a version that had never seen a human game beat the one from the match every game of a hundred, which made the human records a crutch rather than a foundation. Then in 2022 researchers published a way of playing that beats top Go programs almost every time, by walking them into a trap they cannot see, and a person can be taught to run it. Superhuman turned out not to be a number. It was a score against the opponents somebody had thought to try.

The paper that did not know what it was

In June 2017 eight researchers published a new design for trained software. The paper is about machine translation, the same job the 1966 committee had dismissed as a solution in search of a problem. It reports scores for translating English into German and French, it ran on one machine for under four days, and its authors close by planning to try the design on images and audio. Nothing in it anticipates a chatbot.

That design is what nearly everything you use now is built from — at sizes its authors never tested. The public half came five years later: on 30 November 2022, ChatGPT was put behind a text box and opened to anyone, and a technology that had been arriving quietly since 2012 acquired a date. Nothing was solved that week. Something was shown.

The field's own summary of all this has a name, the bitter lesson: across seventy years the methods that won were the general ones able to absorb more computing power, and the work of teaching machines what people already know was overtaken every time. That is the argument under the current spending, and why the bet keeps being placed on bigger rather than cleverer. It is also contested. A rebuttal appeared within a week of its publication in 2019, and the lesson's own author has since put today's chatbots on the wrong side of his own distinction.

So put today's systems where they are: on the learned side of a bet made in a 1955 budget request, running on a design built for translation and scaled past anything its authors tried, not quite four years into the first turn the public has watched from inside. That is a position on a timeline, which is a better thing to hold than a direction. When the next milestone arrives, ask what it settles and about what, and expect the honest answer to be narrower than the headline and, if 2012 is any guide, larger than anyone can see at the time.