Comment by rytill

LLMs learn a distribution during pre-training, not only an average.

Then, by giving them context or by post-training, you can make them sample non-average parts of the distribution they learned.