My summary of the article would be: Relying on the statistics of a distribution works but relying on a optimization of the same distribution doesn't work.