But how do you get to the "unlikely" determination? Do you want to get a ballpark reading on "unlikely" within the first 10 minutes of the interview or at the very end after investing an hour?
I'm not suggesting that FizzBuzz is the be-all-end-all but for the folks who already do find useful as a low-investment bs filter, it makes the most sense to use it at the beginning of the process and not at the end. Using it as a "last hurdle" makes no sense if you have a stack of resumes to plow through. If you have prior assumptions that 90% of those candidates are not competent at programming, what's the logic in using FizzBuzz as the last step? You want to identify the 10% as quickly as possible before investing too much extended interview time.
M positions, N > 2*M candidates.
(I actually don't really give these sort of tests at all anymore. Didn't seem that useful in practice.)