Summary of the content of the blog found here.

Author seems to claim that due to 1st order approximation of Chebyshev polynomials as the convolution filter, their expressive power might not be good
that's because this approximation, when applied to the special case of regular graphs (2D images, eg), will lead to severely weak models
But the main authors of the GCN paper also note that the Weisfeiler-Lehman algo also doesn't converge on regular graphs!