Well, then let me clear it up. The statistics becomes more precise. As in, for a given prefix A, and token x, the difference between the calculated probability of x following A (P(x|A)) to the actual probability of P(x|A) becomes smaller. Obviously, if you are dealing with a novel problem, then the LLM can’t produce a meaningful answer. And if you’re working on a halfway ambitious project, then you’re virtually guaranteed to encounter a novel problem.