RLHF Reinforcement Learning from Human Feedback
The technique that turns a raw LLM into something helpful, harmless and polite: humans rank AI answers and the model learns to produce what they prefer.
Process: show the model two answers to the same question, ask humans which they prefer, train the model to produce the "winning" style more often. Iterated thousands of times.
The reason Claude and ChatGPT feel human rather than autocomplete on steroids. Also the basis for not answering "how do I build a bomb" questions.