In brief,
the paper's says that under certain assumptions the globally optimal
latency would be only 25% better than the selfish result.
If I might, his model says the best is between 0% and a max of 25%
better than the selfish result. Selfish routing may actually get us the
best possible Internet, although that is not proven.
This sounds a lot like the 80-20 rule. Selfish routing gets you 80%.
Infinite knowledge will bring you up to 100% of what is possible but that
presumably costs a very large amount of money.
--Michael Dillon