“Reinforcement Learning from AI Feedback (RLAIF) operates on the assumption that for a language model, verifying a solution is significantly easier than generating one.”