RLHF: How ChatGPT Learned to Be Helpful

Pre-training gives a model knowledge. RLHF (Reinforcement Learning from Human Feedback) gives it alignment — teaching it to be helpful, harmless, and honest.