Many of the 'killer features' of untyped functional languages have migrated their way into today's 'dynamic' languages like ruby, python, etc. You can use a REPL, you don't need to spend 15 minutes compiling, no static typing, etc. So in terms of 'new' things you'll learn from Clojure, you have s-exps and macros. This allows you to manipulate code as data, because the code is data.
The typed functional languages have an advanced type system, as the name indicates. In 'normal' languages, you have what I casually refer to as 'primitive types' Integers, Floats, Doubles, etc, and after that you're on your own. Typed functional languages have ADT's that allow you to model complicated systems via the types. Probably the standard example of a real-life use of this is that you can write code that will throw an error at compile time if you try to inadvertently try to add Feet and Meters. They also have pattern matching, which is pretty handy. Imagine a big if-elsif statement that will throw a compile time error if you handle bank transfers up to a million dollars, and transfers over two million dollars, but not ones that are between one and two-million dollars. You can think of both of these features as typedefs and case/if-elsif statements on crack. That's the kind of stuff you'll get from F#.
I personally think the typed functional languages have much more compelling features in this day and age. Like I said above, the features that I really dug in Lisps have made their way into most dynamic languages, and don't seem as mind-blowing as they once did. But if you really want to go crazy, it's probably worth learning at least one language from each functional sub-family at some point in time.