The AI Safety Game (UPDATED) — LessWrong