See e.g. https://
arxiv.org/abs/2212.09993 "Our experiments reveal that while powerful deep models offer reasonable performances on puzzles that they are trained on, they are not better than random accuracy when analyzed for generalization."

→ View original post on X — @fchollet,