Comment by pcwalton

That statement seems extremely hyperbolic. Just today I tried automating some very pedestrian code for a system that wasn't particularly well-documented and ChatGPT 4o hallucinated the entire API. It was deeply frustrating and wasted more of my time than it would have taken to just slog through the documentation.

I won't deny that LLMs can be useful--I still use them--but in my experience an LLM's success rate in writing working code is somewhere around 50%. That leads to a productivity boost that, while not negative, isn't anywhere near the wild numbers that are bandied about.