I read it too, and I fail to see the breakthrough. There are plenty of pseudo random number generators, such as the Mersenne Twister, with very long periods, so just occassionally XORing even a poor quality random number into the feedback loop, is enough to make it completely unpredictable.

Mersenne Twister is pretty much the standard for simulating a uniform distribution in a lot of scientific computing. These depend not only upon unpredictability (useful for avoiding biases, and clearly important in the security realm), but also upon properties of the uniform distribution.

But when we test it out, we find it's still not as great as we'd like: look at a histogram of outputs, and you'll see that until you get really large numbers of function calls, the histogram isn't particularly uniform. (In particular, numbers near the bottom and top of the range don't get called quite as often.) This means that simulation properties that rely upon uniform distributions over both long and short time periods may be thrown off, and short- and mid-time simulation results may well stem from the MT rather than from the mathematical model. Moreover, low-probability events may have artificially smaller probabilities in the simulations (because of the non-uniformity of the distributions near the bottom and top ends of the range).

Over very short numbers of function calls (a few hundred to a few thousand), the outputs can even tend to cluster in a small neighborhood. So suppose that you are simulating a tissue with many cells, and calling MT as your uniform distribution to decide if the cells divide or stay dormant (each with independent probability *p*, so each cell divides if PRNG/max(PRNG) < *p*). The math says that for a uniform distribution, you don't need to worry about what order you evaluate your decision across all the cells. But if the PRNG outputs cluster over several sequential calls, then a neighborhood of cells may simultaneously divide if they are all evaluated close to one another sequentially. In analyzing the spatial behavior of such a simulation, you may draw incorrect conclusions in smaller spatial structures that, again, derive from non-uniformity of the PRNG, rather than problems with predictability. (And then you may accept/reject a particular hypothesis or mathematical model pre-maturely.)

So, there's definitely more to it than just unpredictability, depending upon where the code is being used.