The chapter opened with a machine that could not price four bottles of wine, and could not be fixed by trying harder. Every taste weight was being pulled toward two prices at once, so the corrections cancelled and the machine settled into being confidently, permanently wrong.

What was missing was not effort or examples. It was a place to put an idea. Plum is not worth anything on its own. Plum in a cabernet is.

So you added a row in the middle, and the machine got somewhere to hold that thought. But the middle row arrived with a problem attached: the bottle tells you the price, and says nothing whatsoever about what a hidden unit should have sent. There is no correct answer to hand it, and no label to compare against.

The answer was to stop asking for one. A hidden unit never needs to be told what it should have done. It only needs to know how much the final error responds when it moves, and that can be worked out by starting at the error and multiplying, one crossing at a time, all the way back.

That is backpropagation, and it is worth saying plainly what kind of thing it is. It is not intelligence, and it is not insight. It is bookkeeping. It answers one narrow question, how much did this number matter, and it answers it for every weight in one pass. Gradient descent then takes the step. One measures, the other moves.

The reason this matters so much more than a wine merchant's pricing problem is that it removed the ceiling. Before it, a machine could only learn what could be said directly about its inputs. After it, a machine could invent the middle ideas it needed, however many layers deep, and still be told what each of them was worth.

Everything since has been built on that. More layers, more weights, better ways of keeping the walk stable across a hundred crossings, but the same two jobs in the same loop.

The sommelier never could say what made a bottle expensive. She tasted, she guessed, she was told the price, and something in her got better at it. That was the mystery this course started with, and it is still not fully solved in her case.

But you have now built a machine that does the same thing where every step is visible: guess, be wrong, work out what each part had to do with it, and change. No rule written down anywhere. Just being wrong, carefully, over and over, until the connections hold something true.