Need systems that can do logical reasoning and planning with gaurauntees, otherwise you won’t really be able to count on the assistant to take care of things you actually care about.
There’s work in inductive logic programming (ML on symbolic data) and neuro symbolic approaches that are very interesting. I think we’ll see new classes of systems in the next 5 years or so that combine much of the reliable decision making work that’s been going on since the beginning of AI with NN learning progress from the last decade