The other thing here is we know the human brain learns on far less samples than LLMs in their current form. If there is any kind of learning breakthrough then the amount of compute used for learning could explode overnight