> LLMs can easily find errors in outputs they themselves just produced.

Really? That must be a very recent development, because so far this has been a reason for not using them at scale. And noone is.

Do you have a source?