You don't care at all about (c) maintainability or fragility?
One of the biggest problems I see is LLMs duplicate sources of truth they have no business duplicating... all the time. E.g. if an internal constant in some code is X then they just hard-code X in fifty other places far away and code based on that. And it's not just constants, they often duplicate entire pieces of logic too. So everything is fine and dandy when the code is fresh, but the next time the original is updated, the rest of the code breaks, because nothing forces this update to update too. Of course the garbage tests they produce never test such hidden assumptions either.
And this is before I get to other things, like (d) test costs and (e) the limits of test matrices. That "inline assembly" you just inserted now requires you to test everything on older CPUs that may lack newer instructions. And on other architectures too. That's a huge cost and this kind of thing exponentially grows your test matrix. Or, more likely, you end up with a breakage on some CPU because you didn't care.
If you have unlimited tokens i guess this is pretty much true these days. I've noticed opus write write and remove the same piece of code in a session. They prefer to write a lot, a lot of defensive checks in python etc.
At my company I have around 4000 GitHub copilot credits per month. I cannot afford to just let the LLM finish a whole project by itself using just spec documents. I did that once, ran out of tokens and then found out i couldn't understand what was going on.
This will probably change in the future but right now I need to work with what's available to me