By using the models you're training them to be better at these tasks, it's where a good number of improvements have already come from. Frontier labs just feed it back in.