Modeling the impact of safety agendas — LessWrong