If you can generate obfuscated chain-of-thought, can you monitor it? — LessWrong