“The same techniques used to train current AI models for code, image input, and voice are expected to scale effectively for computer use capabilities.”