Also. RLHF mean that models spit out according to certain human preference, so it depends what set of humans and in what mood they've been when providing the feedback.
In general, Claude was impressed by what Codex produced and noted the parts where it (i.e. Claude) had missed something vs. Codex "thinking of it".
From a "daily driver" perspective I still use Claude all the time as it has plan mode, which means I can guarantee that it won't break out and just do stuff without me wanting it to. With Codex I have to always specify "Don't implement/change, just tell me" and even then it sometimes "breaks out" and just does stuff. Not usually when I start out and just ask it to plan. But after we've started implementation and I review, a simple question of "Why did you do X?" will turn into a huge refactoring instead of just answering my question.
To be fair, that's what most devs do too (at least at first), when you ask them "Why did you do X" questions. They just assume that you are trying to formulate a "Do Y instead of X" as a question, when really you just don't understand their reasoning but there really might be a good reason for doing X. But I guess LLMs aren't sure of themselves, so any questioning of their reasoning obliterates their ego and just turns them into submissive code monkeys (or rather: exposes them as such) vs. being software engineers that do things for actual reasons (whether you agree with them or not).
