Loading slide
Two ideas down, and they fit together. Let us make sure they are solid before the last one, which leans on both.
First: the model writes one word at a time, always forward, never back. It scores the next word, picks one, adds it, and repeats. There is no plan and no draft. That is why it can paint itself into a corner, committing to an opening it then has to live with.
Second: the pick is not always the top-scoring word. A dial called temperature decides how much it gambles. Cold means the safe favourite almost every time; hot means rarer words get a turn. Same scores, different appetite for risk.
Put them side by side and a small mystery resolves. People often assume a long, fluent answer must be the product of careful thought happening somewhere inside the model before it speaks. But there is no inside where that could happen. There is only the next word, scored and chosen, over and over. So where does anything that looks like reasoning actually take place? That is the last idea of the chapter, and it is the most surprising of the three.