Reflection with optimal predictors — AI Alignment Forum