In the future, all compute that isn't used by LLMs will be devoted to dumbass architectures that diff the entire model state with the view state to find and apply changes.