Formally Stating the AI Alignment Problem — LessWrong