Here are the slides from the physics colloquium I gave at UC Santa Cruz last week, entitled "Why is the Past Different from the Future? The Origin of the Universe and the Arrow of Time." (Also in pdf.)
The real reason I'm sharing this with you is because this talk provoked one of the best responses I've ever received, which the provokee felt moved to share with me:
Finally, the magnitude of the entropy of the universe as a function of time is a very interesting problem for cosmology, but to suggest that a law of physics depends on it is sheer nonsense. Carroll's statement that the second law owes its existence to cosmology is one of the dummest [sic] remarks I heard in any of our physics colloquia, apart from [redacted]'s earlier remarks about consciousness in quantum mechanics. I am astounded that physicists in the audience always listen politely to such nonsense. Afterwards, I had dinner with some graduate students who readily understood my objections, but Carroll remained adamant.
My powers of persuasion are apparently not always fully efficacious. Also, that marvelous illustration of entropy in the bottom right of the above slide? Alan Guth's office. Update: Originally added as a comment, but I'm moving it up here-- The point of the "objection" is extremely simple, as is the reason why it is irrelevant. Suppose we had a thermodynamic system, described by certain macroscopic variables, not quite in equilibrium. Suppose further that we chose a random microstate compatible with the macroscopic variables (as you do, for example, in a numerical simulation). Then, following the evolution of that microstate into the future, it is overwhelmingly likely that the entropy will increase. Voila, we have "derived" the Second Law. However, it is also overwhelmingly likely that evolving that microstate into the past will lead to an increase in entropy. Which is not true of the universe in which we live. So the above exercise, while it gets the right answer for the future, is not actually "right," if what we care about is describing the real world. Which I do. If we want to understand the distribution function on microstates that is actually true, we need to impose a low-entropy condition in the past; there is no way to get it from purely time-symmetric assumptions. Boltzmann's H-theorem, while interesting and important, is even worse. It makes an assumption that is not true (molecular chaos) to reach a conclusion that is not true (the entropy is certain, not just likely, to increase toward the future -- and also to the past). The nice thing about stat mech is that almost any distribution function will work to derive the Second Law, as long as you don't put some constraints on the future state. That's why textbook stat mech does a perfectly good job without talking about the Big Bang. But if you want to describe why the Second Law actually works in the real world in which we actually live, cosmology inevitably comes into play.