I could also imagine a sort of two-tier approach, where the on-device model can handle the majority of queries, but recognize when it should pass the query on t

I could also imagine a sort of two-tier approach, where the on-device model can handle the majority of queries, but recognize when it should pass the query on to a larger model running in the cloud.