It was never about wine

Four tastes went into this machine. Nothing about the method cared that they were tastes.

They were four numbers describing one example, with an answer attached. Swap them for something else and every step you have watched still runs.

Make them the brightness of pixels in a photograph, with the answer being what the photograph shows. Make them a stretch of text turned into numbers, with the answer being the word that came next. The counts change from four to millions, and the price becomes something other than a price, but the machine still multiplies inputs by weights, still passes totals through a floor, still guesses, still measures how wrong it was, and still walks that wrongness backwards to find what each weight had to do with it.

The hidden units change job too. Ours ended up standing for wine types because we built them that way by hand. Nobody does that at scale. The middle layers are handed no meanings at all. They are pushed only by the error, and whatever internal signals happen to reduce it are the ones that survive.

What emerges looks a lot like ideas, and in vision networks it has been catalogued in detail. The first layer fills up with edge and colour-contrast detectors. Layers after it build corners and curves out of those edges, then textures, then parts of objects. Nobody wrote down "edge" or asked for one. It appeared because having edges to work with makes the final answer less wrong, exactly as having plummy cabernet to work with made our price less wrong.

They are also messier than our four tidy boxes. A useful idea may be smeared across thousands of units, and one unit may take part in many unrelated ideas. There is usually no single unit you can point at and name.

Then hold the scale in your head for a moment. The machine you turned by hand had four weights, then twenty. A model at the frontier has hundreds of billions of them, arranged in many layers, trained on more text than a person could read in a thousand lifetimes.

Not one of those numbers was set by a person. Every one was moved by the same two jobs you just watched: a walk backwards to find what each weight did to the error, and a small step downhill to change it. Repeated, over an unimaginable number of examples.

That is the honest answer to how these systems are built. Not a rule someone wrote and refined. A very large machine, an enormous number of examples, and an error walked backwards until the connections hold something nobody put there.