Seeing how monads reintroduce temporal dependence is also an astounding insight, and I can't believe that I've never seen them described that way. That means that the monad has more in common with, say, the mutex, than it does with I/O. I've always tried to approach them with metaphors like sockets and utterly failed to grok anything for more than a couple of days.
I'm thinking that if Haskell is an "impure" functional language, then I would like to see a "pure" imperative language. I envision it as similar to Go but multithreaded and basically every function in the language would be its own process, communicating with sockets (so basically a shell with C-style syntax). The overhead could be largely eliminated with a stackless context switcher. It wouldn't be able to achieve the optimization of FP but it might be more approachable to the mainstream and by extension maybe iterate faster and lead to higher productivity. I’d like to see it take the place of OpenCL/CUDA and SIMD because I find them tedious compared to more elegant languages like MATLAB/Octave.
http://adam.chlipala.net/cpdt/
True purity actually doesn't look much like Go at all. Instead, it means that every expression literally can be evaluated and replaced with its value in a completely arbitrary order. This massively simplifies the meaning of a language and opens it up to bear much more structure (such as that of a mathematical proof).
So, maybe I'm just using the power of wishful thinking with my characterization above.
That said, FP often involves the introduction of monads which restore sequencing (along with many other beneficial effects). A monadic FP computation is very similar to a plain imperative computation... just with all of the arbitrary details automatically selected to be coherent and optimal.
You can even pretty easily embed OO in FP languages (though it gets a little hairy sometimes) since as soon as you can simulate open recursion somehow you can get late-binding as you like. See Oleg's O'Haskell papers for this kind of nonsense.
[0] In this regard, even Haskell is impure since it allows for non-terminating computation which can be seen as an impurity since you cannot expect to just evaluate `let x = x in x` repeatedly and get anywhere meaningful. Other languages like Coq, Agda, Idris are terminating (thus not Turing Complete) and therefore truly pure.