This is not necessarily that surprising. Hinton's 'dark knowledge' (not cited in the paper) already showed that a remarkable amount of information is hidden in

This is not necessarily that surprising. Hinton's 'dark knowledge' (not cited in the paper) already showed that a remarkable amount of information is hidden in the classification probabilities emitted by a model, and that one neural net can learn a lot from and reverse-engineer another neural net given just its precise predictions.