[ben_w]

Re #3: They're fairly likely to follow best practices going forward (if you "hold them right" they already can), but bugs are an important thing even when the models are better than any individual human.

Reason being: memetic monoculture.

Everyone has some blind spots, same for LLMs: No matter how much you use any particular model (or ask a particular human), when you hit a blind spot, they can't reflect on it and actually resolve it — even if you ask them to, they'll make a new "solution" with the same flaw, whatever that flaw happens to be.

Code review gets an extra pair of eyes to check for such things (or at least can, I have experienced coworkers who were write-only-read-never for any criticism of their code or architectural choices).

In chess, this is "being a centaur" (human + AI hybrid working better than either alone); people have been arguing that this is no longer true since about 2013, but given that was 16 years after Deep Blue finally beat Kasparov, even that is an economically useful delay for those of us who want to keep getting paid.

(Was the Ask HN itself AI generated? The trouble is, humans copying each other is also a way to get a memetic monoculture, and we also copy LLMs…)