1It makes things

AI draws a picture for you

Pick some cards. Who to draw, and where they are.

A picture shows up right away. No person drew it.

Now press Draw one more.

Same cards, different picture. The arms are different. So is where they stand.

does not pull the picture out of a drawer. It makes a new one that existed nowhere before. Making something like this is called .

We trained this machine ourselves. We only showed it six friends, so those six are all it can draw.

It cannot draw a dog or a car. It has never seen one.

How it draws them — that's chapter 7.

A little deeper

For grown-ups, and for children who want more.

What drew this is a diffusion model — the method behind nearly every picture generator today.

It learns backwards. The machine is shown, over and over, how a clean picture turns into pure noise when you sprinkle mess onto it, and it learns how to run that process in reverse. Start from any patch of noise and it can walk back towards a picture. The noise it starts from differs every time, so the result does too.

One thing to be clear about — the pictures on screen were made by this machine, but not at the moment you press. Running the model in a browser takes over a second, so for each of the thirty-six orders we drew eight pictures in advance and show one of those. All eight differ; from the ninth press on, a picture you have seen comes round again.

What children ask most — "a person didn't really draw it?" It did learn from piles of pictures people drew, but this particular one existed nowhere before. Copying and learning are different things, and this is a good moment to say so out loud.

Try together — press Draw one more five times on the same cards. All five differ. Ask "which one is the real Posong?" and the child works it out: all five are real, and all five were newly made by this machine.