The AI bashing is a response to the "AI is superintelligence and can write a kernel that's better than Linux" mentality, that is tiring, reeks of gross incompetence, and is almost maliciously wrong.

SOTA LLMs can't reliably implement a single, small feature without producing overly complex code. They don't do engineering.

My approach is two-fold: 1. frontload every conversation with a lot of guidance on how to approach things "like an engineer" (those terms are never uttered, but the guidance basically encodes my workflow as an engineer), to focus on only the problem at hand today, not overbuild things, etc, and 2. not trust a single line of code it produces.