>                         If you use the async event every time, you should get
> "real" random numbers.  The importance of this property of course dependes on
> the application.

Using related async events for random numbers will usually result in the
numbers themselves being related.  While the time between system startup
and the first keypress may be fairly random, sampling the time between
consecutive keypresses will often yield similar numbers over and over
again.  No matter how precisely you try to time such things, it is just
about impossible to make even the LSB's of a 1MHz counter really "random".
Correlational tendencies frequently leak through, but cannot be reliably
measured or modeled.

A common workaround to this is to try to, e.g., XOR the seed of a PRNG
with the "random" physical input periodically and hope that will produce
random numbers.  While this will often work, it is usually possible for
such systems to get into a bad state where the non-randomness of the PRNG
interacts badly with the non-randomness of the physical source and the
result is a really horrible random number generator.

Therefore, a common recommendation is to use a physical process to seed a
PRNG, but not to use it at all thereafter.  Instead, just choose a PRNG
which has suitable characteristics and use it.