"Base models were scoring extremely low on ARC V1 like sub 10% basically... performance of base LLMs on V1 stayed very very low even though in the meantime we had scaled up these models by 50,000x."