Counterfactual control incentives — LessWrong