@zhaisf @geoffreyhinton We have a modern version of the experiment (mnist is tricky: almost everything looks good on it), and i believe a better explanation than "dark knowledge" in our paper: https://arxiv.org/abs/2106.05237