I think raw base text training data may be decreasingly important vs good RL environments/gyms, and maybe less clear that Google wins by default there?