Google still has an unbelievable training infrastructure advantage. The second they can figure out how to convert that directly to model performance without worrying about data (as the o1 blog post seemed to imply OAI had) they’ll be kings.