Counterfactuals, thick and thin — AI Alignment Forum