Scaling Laws for Reward Model Overoptimization — AI Alignment Forum