A framework for thinking about AI power-seeking — LessWrong