Appendix: how a subagent could get powerful — LessWrong