Yes I think the whole point of Zoph and Le's research is to empirically demonstrate the desired trade off of compute cycles vs feature engineering