Counterfactuals for Perfect Predictors — AI Alignment Forum