We've learned a lot through the process of training this model, discovering what scales well and what doesn't.