[AaronBrown]

There is a minor, but puzzling, error in the article. They give the example of a random number algorithm that generated six-bit random numbers. If you had to use this for a shuffling algorithm, you would just throw away any value it returned greater than 51. You don't need to have bias. With a standard 15-bit random number, you would throw away 8 values and have a perfectly uniform distribution.

My other comment about the analysis is it touts theory over practice. They have eliminated some theoretical problems, and even used "overkill" for what they admit are marketing purposes. In the process of doing that, you can make the system so complicated that it has a bug. Often implementations like this work great when they're built, because you paid top dollar for a motivated programmer to do it right. But when the stuff gets maintained by people with less talent who don't understand it, problems creep in.