Summing up "Scheming AIs" (Section 5) — LessWrong