This wasn't the Three Amigos fault. They did quite impressive work on the meta-model aspect of UML. It was defeated by two forces. First, the fundamental problem lies with the amorphous nature of software itself. Coming up with a really expressive graphical notation is hard. Most developers know enough to draw boxes for classes and open-arrowheads for inheritance, but don't get much further into the UML specification because it gets quite convoluted (especially if you start looking at the later generations of UML, with Object Constraint Language and its ilk).
The second failure reason is the implicit assumption that you need (nay, must) design all the classes and interactions before you start writing code. Big Design Up Front is a failed technique in almost all software development. The only exceptions are systems that are truly life and death. One of the reasons for the outdatedness of the software on the space shuttle lies with the fact that they have very long iterations. In other words, they are willing to say "once this date passes, we will make no changes to the design of this system. Period." While most business software could make this statement, it ill serves the business. Business processes change like the weather, and you need software that can change just as readily. I don't come to this discussion as a dilettante: for a while, I worked for a company that was a Rational partner. We did the training, and we built software based on the Rational Unified Process. We even had some successes. But it didn't take long for us to realize that the upfront design didn't serve our clients because it hampered the kinds of changes required by their business.
Most developers I know use AML: Arbitrary Markup Language, usually consisting of boxes, circles, and lines. When a given developer writes on a whiteboard, they write in their own version of a diagramming language. It's a shame that we don't have an industry wide diagramming language that everyone feels compelled to use, but that's the reality in most places I've been for the last 5 years. But, having said that, I'm a fan of AML, because it cuts down on irrational artifact attachment: you have nothing except the last 5 minutes invested in the diagram, making it as transient as possible. Transient artifacts are good because you're willing to throw them away, preventing them from becoming a part of the documentation for your project once the actual code has migrated away from that initial stab at design. Out of date documentation is worse than none at all because it actively misleads