It's especially an issue with type-ahead tools that hallucinate function names or introduce subtle bugs: you lose time and get bumped out of the flow as you evaluate the AI's proposal. (Or fix it, if you were unfortunate enough to accidentally hit tab while indenting a line and the AI slipped in a change.)
The agentic tools do better but don't yet have enough context to know that making this change _here_ will break that thing over _there_, so they require a lot of management, which seems to engage a part of the brain than coding.
>“When we watched the videos, we found that the AIs made some suggestions about their work, and the suggestions were often directionally correct, but not exactly what's needed,”
Being aware of this and investigating just sounds like responsible use of AI.
Can't tell if maybe I'm just an idiot or if these tools are actually really useful.
This in particular is very interesting to me. I haven't read the study yet but this makes me consider my own use of AI - I often feel like it is speeding me up, but is it really? Can I measure it in a better way?
If I want to troubleshoot production environment most of the time ai slows me down. It is better if I think and debug than asking the AI
