> What constitutes "enough" is a value judgment that many vary from
> application to application.  For some applications, a much weaker PRNG 
> would suffice, but we decided long ago that we wanted the full power of MT.

I don't really understand for which application 20000 bits of seeding entropy would be required *in practice*. Surely MT has other interesting properties (such as the statistical distribution of the output) than its insanely large cycle length, that make it desirable as a PRNG.

The paper you linked to ("Good Practice in (Pseudo) Random Number Generation for Bioinformatics Applications") doesn't suggest feeding a 20000 bits seed, it actually seems to say that 64 bits is enough for numerical simulations run on large clusters.

While reading 20000 bits off of /dev/urandom might be fast under Linux, it might not necessarily be the case on other systems. It doesn't sound reasonable to read this many data if there isn't a strong reason for doing it.