Hardening against AI takeover is difficult, but we should try — LessWrong