Training LLMs
RLHF: Teaching Models Manners
8 min+85 XPStep 1/10
Pretraining gives power, not politeness
A base LLM is a next-token machine. Ask it a question and it might answer — or continue with another question, dump a forum thread, or confidently invent a fact. It learned what text looks like, not what humans want.
RLHF — reinforcement learning from human feedback — closes that gap.