This is an *excellent article* (thanks for pointing it out) and highly recommended. It'd be nice to know what JR and/or other gurus think of his (ingenious whether accurate or not) method of producing a usable superbias for CMOS cameras.
I honestly cannot really comment properly, because none of the example images he shows have full size versions. Only the tiny reproductions in the article. Without looking at the result of the real master minus the superbias at 100% scale, especially with an appropriate stretch applied (something on par with the kind of REAL stretch you might apply to an integrated image), you cannot really know exactly what junk may be left in your data to add noise and maybe even behave as a fixed pattern.
The problem with FPN is it's impact is often "silent", for lack of a better word. Most people wouldn't know that, or how, FPN was affecting their results. The only way to really SEE the impact is to progressively integrate more and more data, and then compare the results of each integration. If FPN is limiting you, this will become evident when you flip through such a progressive set of images. Initially, when flipping through such as tack, you will see that the noise pattern in each successively deeper integration continues to change. Eventually, though, you will start to notice that a particular pattern of noise, and this "pattern" may visually appear to be totally spatially random, changes less and less the more you stack. Eventually it will stop changing at all...no matter how much more you stack. THIS is what FPN does...it puts a cap on your potential SNR.
It's a subtle and largely invisible issue, though, because ... few people (maybe no one at all) ... ever actually try to progressively integrate their data and see what they get as they stack more and more. I've never read of any articles of anyone else doing it, so concretely I think I may be the only nutjob weird enough to even try. Most people...simply integrate teh data they have. So they only see the result of stacking everything. A single sample cannot tell you anything about the progressive change in noise profile as more and more data is integrated, though. Very subtle and very invisible problem.
So despite the article above, which doesn't really seem to properly evaluate the impact on the final result of using a superbias, my strong recommendation is still: NEVER. Use a proper master bias. They really are not that hard. A hundred frames will usually do, unless you are stacking hundreds of lights or more. You can acquire a hundred bias frames in a heartbeat. Integrating them shouldn't take more than a few minutes. Then you are done. You don't have to spend EXTRA time creating a superbias from this, you don't have to waste any braincells worrying if the superbias might be making your data worse. Just stack a hundred biases and your done.