If competitors like Deepseek need to distill to build similarly powerful models, that's mostly for evals right? I don't see how it can be cost-effective to distill for a significant amount of training data