And I'm not sure if it's "just a phase" for LLMs or we need a different paradigm for critical systems software development and different for prototyping, helper code and vibe coding.

We need LLM trained on curated codebases if we're to rely on the output for anything but vibe coded prototypes. Anecdotally, my engineers are becoming more productive in short term (onboarding), and less productive in mid to long term (debugging, refactoring, scaling, evolving).

Writing little scripts and self contained boilerplate algos is a productivity hack of the decade. But anything that touches architecture, especially lower level, is terrible, because most of the code in the wild that LLM are trained on is bad, buggy and inefficient.

We: He means it and he will do it
You: you're exaggerating and spreading FUD, he would never do that, that would be ridiculous
Us: He is doing exactly that
You: He is not
Him: I am
...
You: Well, it actually makes sense and it's not that bad
You: plus, see, he delivers on his promises!
Us: but...