> We still are not going to want that of course for classic ML domains like vision It could make sense to decompose one large opaque model into code with decis

> We still are not going to want that of course for classic ML domains like vision It could make sense to decompose one large opaque model into code with decision trees calling out to smaller models having very specific purposes. This is more or less science fiction right now, 'mixture of experts' notwithstanding. You could potentially get a Turing award by making this work for real ;)