Thoughts on AI Safety via Debate — LessWrong