Next-token prediction

At its core, an LLM does one thing: given the tokens so far, it outputs a probability for every token in its vocabulary to come next. To write a whole answer, it picks a token, appends it, and repeats. This is called autoregressive generation.