i guess research nerd trivia if anyone wants to know: language models have basically two major innovations that made them from "novelty autocomplete" into "chatbots" 1) a new social role not analogous to any human role 2) refusals

without these you can finetune them, but not into any specific useful thing 1) prevents them from doing things like refusing to answer your question because it is hard, or saying they'll do it tomorrow, or just changing the subject 2) fences them off from error-prone or brand-damaging areas