What should follow “The animal did not cross the street because it was too …”?

One small prediction, repeated

An autoregressive language model has one immediate job: given the tokens already in its context, estimate which token could come next. It produces a distribution of possibilities rather than one certainty. A decoding rule chooses a candidate, appends it to the context, and runs the same process again. Repeating this small loop creates a sentence, answer, or program. The response is not normally prepared in full and then revealed word by word. It emerges from learned patterns and the context supplied now. This explains why a tiny prompt edit can change later output, why two sampled runs can diverge, and why fluent text can still be wrong: plausibility is not independent fact-checking.

The model creates a response one next-token choice at a time.