You started with four numbers set to 8, meaning nothing at all.

You tasted a bottle, guessed badly, and moved each weight a tenth of the way toward its error. Then you did it again. Eight bottles, over and over, until the guesses stopped being wrong by much.

Nobody wrote the rule those weights ended up holding. It was already in the bottles, and being wrong repeatedly is what pulled it out.

That is the whole of it, and it does not get less true at four hundred billion numbers. It only gets harder to watch.

The next chapter takes the same wine machine and adds one row in the middle. The model becomes able to learn a pattern this one cannot represent at all, and the correction rule you just learned stops working.