Take a breath

Two ideas so far:

One token at a time. The model writes a token, adds it to the text, and looks at the whole thing again before writing the next one. Nothing is drafted ahead and nothing already written gets changed. That is why an early word can box in everything that follows.

The pick is not always the top scorer. Every token in the vocabulary gets a score, the scores become chances, and one is drawn. A setting called temperature decides how sharply those chances favour the leader. Turn it down and the favourite wins nearly every time. Turn it up and the rarer tokens start getting a turn.

Put those together and something follows about how to read what a model writes. A long, fluent answer is a long run of separately chosen tokens. It is not evidence that anything was planned, or checked, or thought through.

Which raises a question the rest of the chapter is about. If there is no plan, what is happening while it writes?