My main problem is with "AI is better for visuals than verbals. That is, AI is better for the frontend than the backend, and better for images/video than for text."

It would be nice to have the "flappy bird app coded in an instance" experience, but mine is often more like this one https://youtu.be/EL7Au1tzNxE?si=2HnFl_Ym_vJUy_kL

(Lots of hinting and nudging the model and then the usage limits are reached.)