I'm confused as well. This is what their youtube video [1] claims, but to me it seems obvious that a computer simulating another computer would be slower.
For classical computers, one computer simulating another is only polynomially slower. The claim here is that any classical computer will be exponentially slower at simulating this problem. The experiment is just to show that the scaling is really polynomial vs exponential, and the uncertainty comes from the fact that we still can't fit very large N problems into current quantum computers. For small N, the two curves could look similar.