One word at a time. No plan, no draft, no going back. The words are the thought, not a record of one.

Once you see this, it changes how you read anything a model writes. Confident prose is not proof of certainty. A long answer is not proof of careful deliberation. A model that "reasons" is one you gave room to write its reasoning down. Underneath it all, the model is doing a single thing on a loop: choosing what comes next, then choosing again.

But that loop has been hiding a question. When the model scores the next word, where do those scores come from? What does the model actually know, and where is that knowledge kept, if not in a list of facts it can look things up in? That is the last chapter of this module, and it is where the model's deepest strength and its most stubborn flaw turn out to be the same thing.