“Next-token predictor” is the wrong mental model for LLMs