I don't understand the premise, maybe I'm missing something.
> LLM-generated code is fuzzy at runtime: non-reproducible, expensive per call, hard to debug.
Wouldn't you have the LLM generate code once, then check in that code. What's not reproducible about that? How is checking in an "E--" script any different from checking in a python script?
Was confused about this as well. The whole thing sounds as if someone confused LLMs writing code and inspecting or executing that code - and then ran with it.
Still, the idea of mostly formal code with dedicated sections that the LLM fills in doesn't sound so bad - but also a bit outdated. It would have been useful a year ago, but not when coding agents generate the entire project in a coffee break.
Edit: The author explains it a bit more in another post: https://news.ycombinator.com/item?id=48863898
If the audience are people who aren't necessarily able to inspect the generated code, but still want more structure and determinism than pure "vibe coding", this makes more sense.