Even really shitty programmers get some sense of the trajectory of the codebase, but this is something that even the best frontier models fail at.

Eh in my experience even good programmers are kinda bad at estimate the future trajectory of the codebase, I've seen experienced people design fancy abstractions only for those abstractions to fall flat on their face when future requirements present themselves.

If you’re only concerned about the “next step”, aka “quality of output”, your error will accumulate and you’ll drift over time. You need to be able to evaluate where you are, so you can correct that error and stay on course. And to do that, you need to be present inside of the code and feel where the pain points are, or your estimation will be so far off that it’s useless.

I feel like this inevitably happens to pretty much every piece of software, hell I'd bet Torvalds regrets a some technical decisions he made with Linux decades ago. Are LLMs worse at this than humans? Probably, but I'm not sure by how much.

I dunno maybe I'm biased because I was always working with fairly crappy and haphazardly written commercial software but I don't see LLMs being that much worse at writing code than humans in general.