- Renowned CS programs spend less time than you think teaching you how to actually program. Programming is simply a tool to teach things like theory, algorithms, mathematics, logic, artificial intelligence, operating systems, graphics, and so on. I actually took a class on programming C++ and it was one of those independent self-teaching classes. :)
- I think one needs to differentiate between programs that teach programming as a trade and a CS program at a top school. I imagine the experiences are quite different. Much like there would be a difference between someone who learned from Learn C++ in 21 Days vs the hallowed Abelsson and Sussman (Alyssa P Hacker anyone?).
- I think most lower division books can be read by anyone and they would glean the same amount of knowledge. In fact, I would say there isn't much difference if someone went to somewhere like OpenCourseWare and learned it themselves.
- I think you'll find more difficulty at the top end. While it's possible that self-taught programmers would do these exercises, if I ever meet one I'll bow down before him. I'm talking about classes like CS 162 Operating Systems (http://inst.eecs.berkeley.edu/~cs162/sp11/) or CS 170 Efficient Algorithms and Intractable Problems (http://inst.eecs.berkeley.edu/~cs170/sp11/).
- With that said, I personally rarely (if ever) use the things I learned in these classes. Perhaps if I wanted to write my own language, my compilers class would be useful. Or maybe if I got really deep into machine learning, then my AI class would have been more use.
So, I guess with that said, it's really about what doors were open to me on matters that, even when taught, seemed impossible to understand. Doing it self-taught is just that much harder unless you're the guy from Good Will Hunting.
All in all, I'd say that for the majority of people who are not that specialized (like web developers) it doesn't make much of a difference. The people for whom it made a real difference went on to do research in a specific area and are the ones who are hired for some specific knowledge they have (eg. classification algorithms, machine learning algorithms, security, complex statistical models).