It seems to me a rather big coincidence that statistical physics works so well.
I can see how consistent macroscopic observations can occur just because the microstates that give rise to that behaviour are overwhelmingly more likely than other microstates. And I can see how the word 'likely' here is really just a statement about there being many more ways for particles in a system to share (for instance) energy equally, than for one particle to have most of it, so that given all sorts of complex interactions the system will most often find itself in a more uniform state.
So say a system is in a given microstate at some time $t_1$, at at that time we observe the corresponding macrostate. Now we have phenomenological thermodynamics to tell us how the macroscopic observables will evolve, and what we can expect to observe at time $t_2$. Clearly that means the microstate at $t_1$ evolves in such a way that it gives rise to those observables at $t_2$. But there's many different microstates that could have caused us to observe the original macrostate at $t_1$, so then it must follow that the overwhelming majority of these microstates also evolve to give rise to the macrostates we expect at $t_2$. I don't see a reason why this should necessarily be so.
Is there some mathematics that I'm missing that explains this coincidence? Or is this really just some strange quirk of nature without which macroscopic physics wouldn't even work?
EDIT: It seems I haven't done a great job explaining my confusion. I've found a paper by E.T. Jaynes that touches on these issues in sections 3 and 4. He seems to explain the coincidence using sharply peaked probability distributions, though I don't quite understand how that works, and Jaynes tends to be a little concise in his explanations. It would be great if someone could explain in more detail, or provide some other references.
EDIT: I've removed the introductory part of this post as it created unnecessary controversy because I wasn't being clear. If some comments don't seem to make sense, it's likely because of this.
EDIT: To hopefully put this controversy to rest, quoting myself from the comments:
I think we're arguing on different levels of abstraction. Within 'theories we construct about the universe' I'm making a distinction between two types of use of probability.
The first kind we use to describe flipping a classical coin; we only use probability for this because we lack information to reason about the coin with our other 'theories about the universe' (such as classical mechanics; we don't know all relevant forces in a 'random' coin toss, but we could in principle program a robot to always throw heads).
The second kind of probability we use to describe randomness in our 'theories about the universe' that we cannot explain by means of an underlying theory (such as classical mechanics in the coin toss example) and thus seems fundamental, such as in quantum mechanics.
Again, this applies to the use of probability alone. Since the use of probability in statistical physics seems to be of the first kind (though that's somewhat debatable as it does use quantum physics), I find it strange that everything works out so well, but this may just be a consequence of the Law of Large Numbers.