Imagine an agentic coding system with a good memory implementation. It's prompted to build an app with sophisticated drag-and-drop features. And then another one... After twenty such apps it realises there must be better way and comes up with declarative dragging.

It's pretty clear to me how far current models are from this being real. I'd say it's a good example of how AGI it not upon us, despite what the big labs are claiming.

Even the remarkable mathematical achievements are in some ways brute-forcing in a limited space. The space of "how could we build this better" is in an entirely different category of scale.