I wonder how much of the performance gains can be attributed to their improved dataset rather than their architecture. That would be an expensive experiment.