RLHF has the same problem as human reviews of AI code: AI code (and comments) look plausible at first, and if you have 100 other PR's to get to, it looks "good enough" and you approve it. I'm sure the humans doing the "human feedback" part of RLHF at anthropic are just as tired as I am at reading all of it, and start to just approve it when it looks plausible.

It's doubly insidious because it trips up the human brain too: When I'm reading a PR saying "fix lock ordering to avoid deadlocks on user deletion", and there's a comment somewhere in the diff saying "// use the fixed lock ordering here", my brain tends to completely forget the fact that the comment doesn't make any sense in its surrounding context. Because it makes perfect sense in the context of being the human reviewing the diff. But it's a slight bit of mental effort to remind yourself "what is this comment going to look like to someone reading the code after this is merged?"

I wouldn't be surprised whatsoever if the RLHF supervisors forget to apply that extra bit of mental effort and say "yup, this comment looks great", forgetting to check the surrounding code to see if it makes sense on its own.