When experienced software architects use LLMs to produce software they have actually read the code of and iterated on, they can get actually maintainable results and you cannot tell an LLM was involved. Commits to the Linux kernel are up 400% this year according to Torvalds. We all know that means unlabeled LLM help is involved, and if the quality bar is just as high, who cares?
For instance, even if it triples the wall time, you can say "Generate mathematical proofs for every function" and watch it fix a ton of its own bugs.
I have added tons of features to FOSS software I use with my own consumer GPUs and that is only going to get easier and require less hand holding to get correct and maintainable output.
There is indeed a middle ground between yolo one shots and hand coding every line and hand debugging every error.
As a Linux OS maintainer myself, LLMs have made debugging and fixing obscure bugs 100x faster freeing me up for more time to solve new problems.