8How does it work?

AI has seen an awful lot

Before a machine can draw, it has to look. How much looking does it need?

Let's start with one that saw very little.

All four were given the same order: draw Poco in the meadow. Only the number of pictures they saw differs — that way you can tell what caused the change.

Leave it on 36 and press Draw it again a few times.

Hardly anything changes. This machine saw 36 pictures in total, and only one of them was Poco in the meadow.

With just the one, it memorised that one picture.

Now switch to 18,000 and press again. Every press is different.

A machine that saw a lot doesn't memorise. It makes a new one each time. That is what you saw in chapter 4 — the picture changing every time.

The pictures a machine looks at are called .

What AI makes has people's work mixed into it. We come back to this in part 3.

A little deeper

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

What a machine with too little data shows is called overfitting. People assume too little training means it cannot do the job, but what really happens is that it memorises.

The machine turns its dials to do well on what it has seen. With few examples it settles into a state that fits those few exactly — like a student who memorised the whole exam range and falls apart on a slightly reworded question.

To show this honestly we trained four separate models: 36, 360, 3,600 and 18,000 pictures. Matching the step count mattered; otherwise the difference would be training time, not data. Measured, the spread among pictures drawn from the same order widened 15.1 → 22.3 → 24.2 → 24.3.

Where children get confused — the pile above holds all six friends, yet the picture below is always Poco. All four machines were given the same order. Keeping the order fixed is what lets you say the only difference is how much they saw.

Try together — switch between 36 and 18,000, redrawing a few times at each. The 36 side barely changes; the 18,000 side differs every press. Ask "which one draws more different things?" rather than "which draws better" — the second question misses what this chapter is about.