yeah, AI is good at prototyping. it can help find bugs, especially by automating testing, but it can be laboriously slow at actually figuring them out. like, i use claude sonnet 3.7 with junie, and i think that if i had a model that actually knew Go's syntax and grammar rules, it would perform about 10x faster, so often i'm looking in its logs it's saying "oh, we have unused imports, this won't compile" like *duh* it is learning go all over again every time i do something with it. that is a pretty wasteful way to use it.

i also noticed that it doesn't seem to have a principle wired up in it, that if it instruments code with logging that the AI can then accelerate debugging this way. i suggested that in a bug report on junie that i sent out last week after blowing up the i think 128kb token buffer several times on one problem.