From HAL to Jarvis, part 4 of 7. A history of artificial intelligence, told as a story, for people who will use it. It began as a talk.

The idea that was declared dead twice

In 1958 Frank Rosenblatt presents the Perceptron, a machine that learns from examples instead of rules. The New York Times, on the strength of a press conference, reports that the Navy expects a machine that will “walk, talk, see, write, reproduce itself and be conscious of its existence.”

In 1969 Marvin Minsky and Seymour Papert publish a book called Perceptrons. They prove that a machine with a single layer can only draw one straight line between yes and no. It cannot even learn “one or the other, but not both”. The money leaves, and a few years later the first winter arrives.

I want to be fair to Minsky and Papert. They were right about the one-layer machine of their day. They were wrong about the idea, because a machine with several layers gets past that limit, and nobody yet knew how to train one. That is the most honourable way to be wrong, and most of us will not manage it.

In 1986, in Nature, David Rumelhart, Geoffrey Hinton and Ronald Williams publish a way to train many layers at once. The idea works. The second winter comes anyway.

So the field had two winters, and this one idea was buried in both.

An idea from 1958 that needed two things the 1980s did not have. Before I tell you what they were, you need the idea itself. This is the only technical passage in the series.

A machine with millions of dials

Imagine a box with a million dials on it. Hundreds of billions, now, but a million will do.

You show the box an example: a photograph. The box produces a guess: “cat”. You tell it the answer. It was a dog.

Now a procedure runs. It nudges every dial, a little, in the direction that would have made the guess slightly better. Then you show it another example. And another. A billion times.

That is a neural network. That is all of it. Three consequences follow, and they explain most of what you will read in the newspapers for the next ten years.

Nobody wrote the rules. The rules are the dial settings. Show it enough examples and they set themselves. Which is, notice, the answer to the rule-writers’ arithmetic: nobody has to write the billion rules if the machine can find them.

Nobody can read them back. A million dial positions are not a document. Which is why the machine can be fluent and confidently wrong at the same time, and why nobody can quite tell you why it said what it said.

And it eats examples. Millions of them. So it needed two things the 1980s did not have. Data: the internet, then everyone’s photographs, then everything ever written. And compute: it turned out that the graphics cards built for video games do millions of small calculations in parallel, which is exactly what turning a million dials takes. The gaming industry paid for the hardware of artificial intelligence. The idea was from 1958. The fuel arrived around 2010.

Trained, not programmed.

Learning is the fair distribution of blame

One more piece, because you will hear the word and I want you to own it. Backpropagation.

The problem in 1969 was this: with several layers of dials, when the answer comes out wrong, which dial do you blame? The dials in the last layer are easy. The dials in the first layer only touched the answer through everything in between.

The answer of 1986: pass the complaint backwards. Start at the last layer. Ask each dial: if you had turned a little, would the answer have been less wrong? Turn it by that much. Hand whatever blame remains to the layer before, and ask again. Layer by layer, back to the first.

Think of a customer complaint travelling back through a company. Finance adjusts a little. Claims adjusts a little. Underwriting. Sales. Each department takes its share of the blame and changes its part slightly. Now do that a billion times. The company becomes very good, and nobody in it can tell you the rules.

The error propagates backwards. Learning is the fair distribution of blame.

The believers

Through the 1990s and the 2000s, while the field called neural networks a dead end, three people kept working on them. Geoffrey Hinton, Yann LeCun, Yoshua Bengio. In those years, putting the word “neural” in your title was a career risk. A small Canadian research programme funded them when few others would.

In 2012 a student of Hinton’s, Alex Krizhevsky, enters a network into the year’s big image-recognition competition, trained on two gaming graphics cards and a million labelled photographs. The error rate on the benchmark drops from twenty-six percent to fifteen in a single entry. Within a year, the whole field switches sides. The winter ends in a season.

In 2018 the three share the Turing Award, computing’s highest prize. In 2024 Hinton shares the Nobel Prize in Physics, for the idea everyone had buried twice.

Being early is indistinguishable from being wrong, until it isn’t. Two winters, one idea, three people who did not stop.

Move 37

In 1997 Deep Blue beat Garry Kasparov at chess by brute force: two hundred million positions a second, scored with rules written by chess masters. Era one at its peak. The machine was told how to play.

Seoul, the tenth of March 2016, game two of five. AlphaGo, a machine that learned the game of Go by studying human games and then playing itself millions of times, is playing Lee Sedol, one of the strongest players alive. On the thirty-seventh move it plays a stone on the fifth line, a place no professional would play. One commentator thinks it is a mistake. By AlphaGo’s own estimate, a human would have played that move one time in ten thousand. Lee Sedol takes almost a quarter of an hour to answer. He loses the game, and that move is why.

Fan Hui, the European champion, who had lost to the machine months earlier and was watching, said: “It’s not a human move. I’ve never seen a human play this move. So beautiful.”

I will not call it creative, because someone will want to argue, and they will have a point. It was the first move the world called creative.

And then the human answered. Game four, move seventy-eight. Lee Sedol plays a stone the machine had rated as nearly impossible. It is remembered as the “divine move”. It is the one game the human wins.

The machine’s move and the human’s answer, in the same week, in the same room. The first move the world called creative, and the last time a human answered one.

Next: Part 5: What a Language Model Is

The series