“In 2007, FizzBuzz filtered out people who couldn’t write trivial code. In 2026, AI writes it in under a second. The scarce skill moved: reading code against intent.”
In 2007, a blog post made FizzBuzz famous for a depressing reason: the majority of programming job candidates its author interviewed could not write a simple program. It was never a test of skill. It filtered for the absence of one specific ability — translating a tiny spec into working behavior.
In 2026, AI writes a FizzBuzz-class program in under a second. The filter is gone.
We think the problem it filtered for is not gone. It moved. The scarce ability is no longer producing code; it is telling whether plausible-looking code is correct. Generation became nearly free, verification became the job. As far as we can tell, almost nothing in hiring tests it. Every résumé says “AI-assisted development.” Very few interviews check the candidates' judgement against AI-generated code.
Here is the honest part: the 2007 posts had data behind them. Their authors had sat through the interviews and watched candidate after candidate fail. We have a suspicion, not a dataset. So instead of writing an opinion piece, we built an instrument.
FizzCheck is a five-minute spoken code review. You see a short spec and a short implementation. You talk through whether the code matches the intent: is anything wrong, would you change the implementation, what would you test. No typing. No autocomplete. No model in your ear. At the end you get a scorecard on how you reasoned, not just what you found - quoting your own words back at you.
Like the original FizzBuzz, it is a negative filter, and we are upfront about that. It will not identify great engineers. It is also deliberately small: we built it in two weeks to make one skill legible, and to find out whether the failure mode we suspect is as common as we think.
Try it. Then tell us whether we are wrong.
← FizzCheck— 2026