What you just did with "wibble" has a name, and it is the key that unlocks the whole problem.

Linguists call it the : the meaning of a word is given by the company it keeps. A linguist named John Firth put that same idea in one line back in 1957, decades before neural networks existed: "You shall know a word by the company it keeps." That is exactly the move you made. You knew the word from the company around it. Firth was describing how people understand language. It turned out to describe how a machine could too, in a way he couldn't have guessed.

Here is how that line becomes a method. Watch how it turns into the same guess-and-adjust loop from the learning module, now aimed at language.

Take millions of real sentences and play a simple game with the model: hide a word, show it the neighbours, and make it guess the missing word. At the start it guesses badly, the way every network starts out knowing nothing. Each wrong guess gets nudged a little, just like before. To get good at this game, the model is forced to notice which words keep the same company. If "dog" and "puppy" keep turning up surrounded by the same neighbours, then whenever one of them is the hidden word, guessing the other is a safe bet. So the model learns to treat them as nearly interchangeable, because for the purpose of this game they almost are.

And here is the quiet payoff. The model was never asked to learn meaning. It was only asked to predict neighbours. But to predict neighbours well, it has to place words that share company close together, and the moment it does that, the map we asked for two slides ago builds itself. "Dog" and "puppy" drift into the same neighbourhood, not because anyone defined them, but because guessing their neighbours well requires it. Meaning falls out of prediction as a side effect.

We will see that finished map, and the strange arithmetic it allows, in the final chapter of this module.