What they said, except that I wouldn't argue so much with the code bloat — my own body's code has same. Rather, I'd be concerned about the need for better understanding of fitness functions. That is, at least in my own limited experience, the hard part.
But, from a conceptual point of view, it seems to me that GP has over simplified the entire equation in an effort to run before learning to walk. Now that is a hugely oblique and biased statement and based on no serious understanding of the issues, but look at it this way — GP is imitating non-determinanistic systems to achieve specific practical results. Nothing wrong with that, but it would be neat to hear about work that more closely adheres to the model.
Basically, I'd love to see a real code-eat-code world...