Was there discussion of shrinking model size and the path to self-custodial models? Federated learning as a goal is a given, but I tend to think the priority now is to get really good models down to local size and then supply chain security that includes training data (not sure how). I would love to know if that was discussed.