They seemed like strong iterative updates to me? I also appreciate all speech recognition being done on-device