In all fairness, I guess it depends on what you need to do. For the vast majority of the day to day coding, it involves more of the architecturing and creative and puzzle solving side that joel talks about.

However, if an abstraction breaks down, you'll need to have a deeper understanding of the innards, which a sound theory can only help the deeper down you go. I think this is even more applicable if you're using something based on math that isn't as mature of a technology (which means they're not a commodity and you'll probably have to fix it if it goes haywire), such as machine learning techniques or computer vision.

How often would that happen? Not very often. But when it does, and you have the background to learn, understand, and fix it, that'll make all the difference between you and the next guy.