Exactly the inverse-scaling point. The weak model gained most because the harness patched gaps it couldn't fix on its own. Cheaper to evolve the runtime than to scale the model, and you keep the smaller footprint at inference.

→ View original post on X — @akshay_pachaar