Clear weak spots remain counting, generating alternate images when the training data is thick (full glasses of wine, clocks with oddly set hands), etc. It isn't hard to make them fail. But there is a lot they do very well, and the gains have been pretty quick so far.

→ View original post on X — @emollick