On decision-prediction fixed points — LessWrong