It's five figures cheaper to ask everyone to FizzBuzz than it is to fire someone even after a week, and God help you if they're in a protected class.
Fizzbuzz wouldn't have filtered me out of getting my first (and current) real software job, but any sort of Google/MS-style interview would have. It may not be the wisest thing to admit to on the internet, but I'm basically in the opposite boat of your general target on HN: I'm fine with selling, but limited (non-EFL) skills to sell!
That, along with seeing numerous senior iOS dev applicants struggle both with Fizzbuzz and with Macs in general, has convinced me that it's more than worth it to test candidates without code portfolios.
If you really want to test how someone thinks over questions like that, actually give them time. Give them the questions before they show up for the interview. Then go over their solutions. They aren't under pressure, they give you a thoughtful answer, and you get a better sense of what they can do in the real world.
I know enough C++ to write a decent GUI for a client (I mostly code in Qt, which is what hooked my resume in the first place). This test covered every possible wacky thing in the language, stuff I have no idea what to do with and is probably never even used in this company's day-to-day work. Nested namespaces? Template specialization?
Why do they do this? I'm totally discouraged from continuing onward with these people. I know I can do a perfectly fine job with this group. I think I'm going to take C++ off my resume and just give up looking for work in that arena.
I've always admired the 37Signals approach of having potential hires do a short contract.
One month of employment should be enough to find out if someone has the practical skills they've expressed in the interview, and catch the 1% that can bullshit their way through.
Of course that 1% becomes a lot bigger number if the interviewer doesn't know what they are talking about. But that's not a problem you can solve with tests.