Tag
The author reflects on how AI benchmarks are either saturated at the top or brutally hard, and argues neither captures the real production failure mode — models lacking judgment about whether a task is worth doing. They ask whether anyone has found a way to evaluate judgment before shipping.