As far as I understand it; Jev et al can be considered Zero Shot Classifiers.
It’s likely the system you built was trained on images of just cats and dogs; but Jev is using a general training set, and there is tuned for general purpose zero shot classification, rather than domain specific “is this a cat or a dog”.
Another difference from me is the arbitary input. Classic ML models work on a predefined input. Jev-like models don't have that limitation. For example you can replace some columns if you get new info and the model will still work.