High-level hopes for AI alignment — LessWrong