Durability and parallelism aren't that hard, and you won't be needing frontier models for much of this stuff. The parameter efficiency of models is increasing rapidly. I don't see why a (near-)future device wouldn't be capable of running a perfectly capable personal assistant.