AI benchmarks are getting stress-tested and found wanting

The NanoGPT Speedrun Frontier post ran 153 autonomous benchmark runs across 18 frontier models and the discussion immediately zeroed in on what a 'run' even means and how much variance exists between runs. That's the right question: if you can't characterize variance, a benchmark result is noise dressed up as signal. The ProgramBench Vetted thread on reverse engineering from runnable binaries makes the same point from a different angle, explicitly calling out the risk that a benchmark can be hard for the wrong reasons (missing information, shortcuts) rather than measuring the thing you actually care about.

These two threads together show a maturing conversation. A year ago, benchmark scores were taken at face value. Now, the community is asking about methodology, variance, task design, and whether the score reflects real capability or test-set leakage.

The counterpoint is that even flawed benchmarks provide some signal when used comparatively. The danger is when they get cited as ground truth in procurement decisions or marketing.