Comment by sandos

Sorry, AI does much less than 1% of my work. I work on semi-ossified, old, embedded, safety-critical code. Not exactly AIs forte, sadly.

We were (finally!) given the go for even using AI just before summer vacation this year, and I was very excited, having been obsessed with AI 20 years ago. I was still excited quite a while, until I slowly understood all the limitations that come with LLMs. We can not trust them, and this is one of the fundamental problems: verifying things takes a long time, sometimes its even faster just writing the code yourself.

We do have non-core-product tasks that can greatly benefit from AI, but that is already a small part of our job.

I did find two areas where LLMs are very useful: generatin documentation from code (mermaidjs is useful here) and parsing GDB output!

Seriously, parsing GDB output was like an epiphany. I was for real blown away by it. It correctly generated a very detailed explanation of what kind of overwriting was happening when I happened to use a wild pointer. Its so good at seeing patterns, even combining data from several overwritten variables and parsing what was written there. I could have done it myself, but I seldom do such deep analysis in GDB and it did it in literally 10 seconds. Sadly, it was not that terribly useful this time, but I do feel that in the future GDB+AI is a winning concept. But at the same time, I spend very little time in GDB per year.