Does this not imply that its usefulness degrades depending on how much random data you need / the frequency at which you need it? For example if you ask for 8 bits of randomness within the same time-slice (either as a chunk of 8 bits or 8 separate invocations) - then you're only ever going to get 00000000 or 11111111 as your "pseudorandom" values (and this feels like it would be an easily re-producible outcome)