Comment by daemonologist

Corporate forbids AI tools due to IP concerns. When I Google for syntax or something I do sometimes look at the AI result though.

On personal projects I usually use AI (Zed Zeta) for tab completion, although sometimes I get annoyed by it interfering with the UI or my cursor and turn it off. I will also occasionally feed a bug or error into Gemini if I'm really stuck - this only works occasionally but is worth a shot.

Every couple of months I try the current hotness (Copilot/Cursor/Gemini Code/etc.) for a small or greenfield project; if I stick with the project for more than a few days I inevitably find the AI can't do anything except the most common possible thing, and turn it off.

I think the disconnect is in my ability to explain to the model in English what I want it to create. If it's something common, I can give it the gist and its assumptions as to the rest will be valid, but if it's something difficult my few paragraphs of instruction and outlining probably just doesn't provide enough specificity. This is maybe the same problem low-code tools run into: the thing that fully defines what I want to happen is the code itself; any abstraction you build on top of that will be at least as complex, unless I happen to want all the defaults.

On top of that, as others have said: writing all the code yourself is a good way to ensure you know how it all works, and learn about anything new you need to use. This reality reduces my motivation to rely on AI in the first place, because even if it works for now I'm expecting to get hurt by it down the road.