Loading slide
We met this idea once already, with the brain: no single neuron holds your grandmother, the memory is spread across the connections. A model works the same way. No single weight holds "Tokyo is the capital of Japan." The fact is smeared across billions of them, each contributing a sliver to countless different outputs.
That sounds like a quirk. It is actually the source of the model's most important power.
A spreadsheet can only give back what was typed into it. Ask it about a city nobody entered, and it has nothing. A model is different. Because it stored a pattern rather than a list of facts, it can answer about things it was never explicitly taught. It learned the shape of how capitals, countries, and language relate, and that shape stretches to fill gaps it never saw during training.
This is generalization, the idea from an earlier chapter: handling examples it was never trained on, like the student who understood the material instead of memorizing the practice exam. Back then we saw that it happens. Now we can see why. Because the knowledge lives in a pattern and not a list, the model has something to stretch. It is not looking up answers. It is extending a pattern.
The same property has an uncomfortable flip side. If no fact lives in one place, no fact can simply be looked up, corrected, or deleted. You cannot open the file and fix the line about Tokyo, because there is no line. And when the pattern stretches too far, the model will extend it confidently into an answer that is just wrong. The same machinery that lets it answer new questions is what lets it make things up. We will come back to that.