03 · How Models Learn
RLHF
Reinforcement Learning from Human Feedback
The technique that made chatbots feel polished. Humans rank model outputs → those ranks train a “reward model” → the AI is optimized to score highly. This is largely why modern assistants are helpful and refuse harmful requests.
Concrete example
Shown two answers, a person picks the better one; do that at scale and the model learns the qualities people prefer.
Why it matters
The single biggest reason today's chatbots feel aligned rather than like raw text engines.