I still don't rely on AI, but I can see that changing relatively soon. It's already pretty good as an assistant to navigate terse documentation and it's getting really good at assisting in diagnosis. Its codegen, however, is still severely lacking for the kind of work I do (that's the most diplomatic way I can put it).

Just for shits and giggles I recently walked it through the last bit of tough debugging I did. At every step it gave 5 or 6 good suggestions on what to check next. Of course, each of those suggestions represented HOURS of testing, but when I fed it the results I found during *my* diagnosis it proceeded correctly to the next, more honed it, round of tests. After about 5 rounds it presented 5 options, one of which was the actual issue. Pretty impressive.

Of course, if I didn't already have expertise in this area it would have taken me weeks to try out all its suggestions and gradually hone in on the actual issue based on its recommendations. So, on its own, or as a replacement for an expert, it's still pretty much useless.

That being said, the next time I have to debug something like this I will probably run my thoughts and observations through an LLM to see if it recommends anything that I'm unaware of.