Clarification of AI Reflection Problem — LessWrong