The end result could be models trained by models to achieve goals set by models, whose safety is verified only by models. Some fear a disaster.