Large language models are trained on tasks like predicting missing words from context, or deciding whether a sentence follows another one. The current models got quite good at this, but ultimately they are producing plausible sequences of sentences.