Chapter 3 of 32

Part 1. How the machine actually learns

The Sunday nobody came.

May, a Sunday morning, the seventeenth.

I opened the class and counted twelve people.

Twelve. On a Sunday when the topic was the most important thing I would teach all term. I sat there and looked at the grid of black squares where faces were supposed to be, and I asked the twelve who showed up whether we should wait a few minutes for the others.

Somebody said we should start.

I want to tell you what I was feeling, because it was not anger. It was closer to the feeling you get when you have cooked for eight people and three arrive. The food is still good. You still have to serve it. But something has gone out of the room and you cannot put it back by pretending.

So I told them the truth. I said today we are going to talk about artificial intelligence, and I do not want you to memorise it, because memorising it will not survive one follow-up question in an interview. I said I want you to actually understand it, and if you understand it, you will be able to talk about it for the rest of your career.

Then I did something I had not planned. Instead of starting with the machine, I started with a child.

I told them to imagine a baby. A year old, maybe a little more. Grew up in a city, all concrete and apartment blocks, and this child has never seen a cat. Somebody brings a picture book. Orange cat with a white chest. Grey cat. Black cat with white on the front, the kind we call a tuxedo at my house, because we feed one in the yard. Page after page of cats, and each time the adult says the word.

Then you take the book away. You show the child a dog and you ask: is this a cat?

And the child says no.

Nobody taught that child the rule. Nobody said a cat has pointed ears and a dog does not. The child watched enough examples that the shape of the thing settled into their head, and when something arrived that did not fit the shape, the child knew.

I asked the twelve where they had learned most of what they know. They said school. I said right, and who was in the room with you. A teacher. Somebody who already knew the answer, standing over you while you learned it.

That is the whole thing, I said. That is supervised learning. You have not been learning about computers for the last five minutes. You have been learning about yourselves, and the machine copied you.

Then I put up an X-ray.

I told them about the doctor who reads those. Elementary school, high school, medical school, residency, then twenty years of practice. Call it forty years to build that person. And what makes them good at the end of it is not the medical school. It is that they have looked at hundreds of thousands of these images. Old patients and young ones, one after another. Somewhere in there the shape of the thing settled into their head, the same way the cat settled into the child's.

Now, I said, take a million labelled scans and feed them to a machine.

The room went quiet in a way that I recognised. It is the sound of somebody doing the arithmetic and not liking where it lands.

One of them asked how much you would charge for that. I said that is exactly the question, and it is why the knowledge business is about to get very rough. Because that machine does not need forty years and it does not need to sleep, and once it exists you can hand it to anybody with a phone.

I taught for another hour. Unsupervised learning, where you dump everything on the table and let the thing sort it out on its own. A four year old does it with a bin of toys. Blocks in one pile, stuffed animals in another, and nobody gave them the categories. Then self-supervised learning, the trick that made all of this possible, where the machine hides a word from itself, guesses it, and checks. Cheap. No adults required. Do it a hundred million times and you have something that can finish your sentences.

At the end I told them to listen to the recording again. Two or three times. In the car, washing dishes, whenever.

Then I closed the meeting and sat there for a minute with the empty grid still on the screen.

I was not upset about the twelve. I was thinking about the other forty. Because the thing I had just explained was going to decide which of them had a career in three years. They had slept through it. And there was no version of the next hour where I could make them care by wanting it more than they did.

That was the Sunday. Twelve people, one X-ray, and a picture book full of cats.