Scaling Laws for Scalable Oversight — LessWrong