Well and my experience of 2 years working full time with AI (and 15 others of engineering) on safety critical adjacent products tells me you're full of it and wrong.

We've had 4 teams on this model and sure it helped in some things (mostly data analysis and scripts) but generating code and doing architecture is utter crap 90℅ of the time. So much so that we've even had juniors noticing some design patterns that, and I quote: "this is one the examples of bad code we were taught in sw design classes'. The worst is how non deterministic they are. The same prompt from different people yields vastly different results

SW engineering is mostly x + y + z Where x=planning, y=writting code and z=reviewing, then rinse and repeat. Llm spedup y but made everything else take so much longer that equation result is much worse. Now reviewing is utter torture and during planning we're dicussing how to mitigate the pitfalls of LLM (like over engineering, too many abstractions) that we spend fewer time on the planning of the engineering itself and more on cuddling this brain addled tool