The one-line answer

An LLM is a next-word predictor — at enormous scale.

A large language model takes the text so far and predicts what comes next, one token at a time. Everything else it appears to do — answer, summarise, code — emerges from doing that astonishingly well.