From a systems engineering standpoint, the purpose of LLMs is to construct, verify, and "push down" abstractions and deterministic layers. Deterministic layers are able to cope reliably with the law of medium numbers.