Post

It's a real problem. I end up spending almost as much time corralling them into doing the right thing and verifying the results as I would if I'd written the code myself. I find that dumber models with more explicit instructions tend to be less tedious.

my guess is that while for us, reutilizing code requires way less effort, for them is exactly the opposite and just producing new code is way less effort than trying to understand what's already there, so they'll always resort to that unless you specifically tell them not to

tldr really bad context discovery

they just love hyperspecializing code for one specific prompt and i think that's why i'll never be able to take it seriously

bsky.app/profile/okam...