Emergent Deception and Emergent Optimization — LessWrong