Next Token, Every Time

The single most important thing to understand about how a language model works is also the simplest, and it is the thing people find hardest to believe.

It does not plan. It does not draft. It does not think the whole answer through and then write it down. At every step it does exactly one thing: it looks at all the text so far, works out which word is most likely to come next, and produces that one word. Then it does the same thing again. And again. And again.

The entire reply, however long, however coherent, however much it seems to build toward a point, is that one small move repeated. Pick the next word. Append it. Pick the next word.

This sounds far too simple to explain what these systems actually do, and that gap is real. It is worth sitting in. The words come out one at a time, with no plan behind them, and yet the result can feel considered, structured, almost alive. Closing that gap, seeing how something so simple produces something so rich, is how you start to think clearly about what a model is and is not.