Eg : the concept of "Human" is an abstraction over all the different attributes of all the individuals, to reduce them to the ones they all share. If i say i met a "human being", you can't tell if it's a "he" or a "she", if he has dark of blond hair, etc. What you can tell, is that it has a heart and a brain, stand on at most two legs, and has at most two arms. That could be all you need to deduce a few things, but by saying so, i kept some details hidden from you.
Now, if we get back to rx, its whole purpose is to work with the concept of "streams of events". Noting that many data structure and patterns (such as iterators over arrays, or event source observation) can be thought a specific kind of even streams.
This lets you define stream-combination operations, that work on all those structure. Which may, sometimes, be convenient.
The problem is when people start telling you that the generic tools around Rx Stream (producer, consumer, publish subject, etc) are the best tools to model the behavior of all your application, because "everything is in fact one big giant stream of events".
And then you end up with code monstruosities, because you realize that it's more complicated than that. Maintaining a UI state and mutating it becomes a nightmare, you start wrapping every native structures (such as arrays) into Observers just to perform the equivalent of a join, etc. Either the stdlib and frameworks are "reactive", and then using rx pattern would be invisible, or you're going to have a horrible code.
In some way, it's exactly like generics, or base superclass : you define them in the stdlib, for example to share code between Sets and Arrays, but what the end user is manipulating 99% of the time isn't "Iterable". It's Arrays and Sets.