Rohin Shah comments on Naive self-supervised approaches to truthful AI