Addressing three problems with counterfactual corrigibility: bad bets, defending against backstops, and overconfidence. — AI Alignment Forum