Next Token, Every Time

The most important thing to understand about how a language model works is also the simplest.

It doesn't plan. It doesn't draft. It doesn't think ahead and then write. At every step, it does one thing: it looks at everything it has been given, calculates what text is most likely to come next, and produces it. Then it does that again. And again.

The whole output, however long, however coherent, however surprising, is the result of that single operation repeated many times over.

This sounds almost too simple to explain what these systems actually do. The gap between the mechanism and the behavior is real, and it is worth sitting with. The words emerge token by token, with no plan behind them. And yet they can feel considered, structured, alive.

Understanding that gap is how you start to think clearly about what these systems are.