Most benchmarks fail before the model gets a chance
A benchmark can look rigorous and still answer the wrong question. Dan Luu’s long exercise works through AI evals, napkin-math estimates, and winter tires to show where misleading assumptions enter, from underspecified tasks to comparisons that ignore base rates and real-world conditions. The useful takeaway is less a checklist than a habit: interrogate what the measurement excludes before trusting its score.