Yes, it's kinda self-limit itself by the costs of training now. But I think it's inevitable that there will be more efficient training algorithms soon, possibly by orders of magnitude. Probably found with the help of ML, as AI now can be trained for programming and research too.