“Reinforcement Learning from Human Feedback (RLHF) will predictably fail to scale to superhuman models because it relies on human supervision, which is not viable for systems that humans cannot reliably understand.”