LLMs as next-word predictors