Ah! Interesting, I thought the capability was capped by parameters, but you're saying you can keep getting more capability from a fixed parameter size by continuing to train past what the chinchilla paper specifies. That's really cool