Rohin Shah comments on When to assume neural networks can solve a problem