I imagine they're using genome size (number of bases in exons) as their measure of complexity, not something ill-defined.

I don't think it's sensitive the way you say, since they're basically just linearly extrapolating the logarithm, but I haven't read the paper. If you somehow manage to shoot high or low by a factor of 100, that's only a temporal error of six or seven doublings. The bigger issue is that they're extrapolating over orders of magnitude that we have no direct evidence about.