>Sure, but I bet Andrej Karpathy's team is using SOTA and mitigates most of the issues like the ones you mentioned.

Can you give an example of an image classification technique where classification failures are independent, or independent enough?