Musing on Karpathy's Software 2.0 tweet.

It's conceivable that UIs get generated directly from model to pixels, as we already see with 3D worlds.

Instead of genearting for example HTML to then generate the pixels.

If the models are insanely good, who needs any intermediate representation?

Hallucinations all the way down.

No humans in the loop. Just man and the (single) machine.

The model is a black box, but if the behavior is good enough then it's fine.

But this puts a very, very high bar on quality of the model.