Next Token, Every Time

Watch a model answer something and it looks like it is telling you what it already worked out. It is not. It produces one piece, looks at everything including the piece it just produced, and produces one more. There is no draft anywhere. Nothing was decided in advance.

Which means it cannot take anything back. Every piece it commits narrows what can sensibly follow, and a model can talk itself into a position in its opening few words and then have to live there. Some of what looks like carelessness is really a corner it painted itself into before it had any idea where the sentence was going.

The piece it produces is not simply the highest-scoring option either. Every token it knows is scored, those scores become chances, and one is drawn. A single setting decides how sharply the chances favour the leader, which is why the same question twice gives you different answers, and why the same setting turned too far produces nonsense.

That setting gets called the creativity dial, which raises a question the chapter cannot settle and does not pretend to. Turning it does not give the machine any opinion about which word would be better. Yet a machine that could only ever take its likeliest option would never produce anything you had not expected, and the shape of that trade turns out to have a strange echo in how people think.

None of this makes the process shallow. A great deal happens inside the machine before each piece appears. But there is no inner monologue behind it, and the words on the page are the only place a thought can be kept, which is why asking for the steps changes what a model can do and why those steps are not a report of how it got there.