“Reinforcement Learning from Human Feedback (RLHF) and its variants are the primary techniques used by all major labs to align current models like ChatGPT.”