It's an old book, I know, and one that many of us have already read. But if you haven't, you should.
If there's anything in the world that deserves to be called a martial art of rationality, this book is the closest approximation yet. Forget rationalist Judo: this is rationalist eye-gouging, rationalist gang warfare, rationalist nuclear deterrence. Techniques that let you win, but you don't want to look in the mirror afterward.
Imagine you and I have been separately parachuted into an unknown mountainous area. We both have maps and radios, and we know our own positions, but don't know each other's positions. The task is to rendezvous. Normally we'd coordinate by radio and pick a suitable meeting point, but this time you got lucky. So lucky in fact that I want to strangle you: upon landing you discovered that your radio is broken. It can transmit but not receive.
Two days of rock-climbing and stream-crossing later, tired and dirty, I arrive at the hill where you've been sitting all this time smugly enjoying your lack of information.
And after we split the prize and cash our checks I learn that you broke the radio on purpose.
Schelling's book walks you through numerous conflict situations where an unintuitive and often self-limiting move helps you win, slowly building up to the topic of nuclear deterrence between the US and the Soviets. And it's not idle speculation either: the author worked at the White House at the dawn of the Cold War and his theories eventually found wide military application in deterrence and arms control. Here's a selection of quotes to give you a flavor: the whole book is like this, except interspersed with game theory math.
The use of a professional collecting agency by a business firm for the collection of debts is a means of achieving unilateral rather than bilateral communication with its debtors and of being therefore unavailable to hear pleas or threats from the debtors.
A sufficiently severe and certain penalty on the payment of blackmail can protect a potential victim.
One may have to pay the bribed voter if the election is won, not on how he voted.
I can block your car in the road by placing my car in your way; my deterrent threat is passive, the decision to collide is up to you. If you, however, find me in your way and threaten to collide unless I move, you enjoy no such advantage: the decision to collide is still yours, and I enjoy deterrence. You have to arrange to have to collide unless I move, and that is a degree more complicated.
We have learned that the threat of massive destruction may deter an enemy only if there is a corresponding implicit promise of nondestruction in the event he complies, so that we must consider whether too great a capacity to strike him by surprise may induce him to strike first to avoid being disarmed by a first strike from us.
Leo Szilard has even pointed to the paradox that one might wish to confer immunity on foreign spies rather than subject them to prosecution, since they may be the only means by which the enemy can obtain persuasive evidence of the important truth that we are making no preparations for embarking on a surprise attack.
I sometimes think of game theory as being roughly divided in three parts, like Gaul. There's competitive zero-sum game theory, there's cooperative game theory, and there are games where players compete but also have some shared interest. Except this third part isn't a middle ground. It's actually better thought of as ultra-competitive game theory. Zero-sum settings are relatively harmless: you minimax and that's it. It's the variable-sum games that make you nuke your neighbour.
Sometime ago in my wild and reckless youth that hopefully isn't over yet, a certain ex-girlfriend took to harassing me with suicide threats. (So making her stay alive was presumably our common interest in this variable-sum game.) As soon as I got around to looking at the situation through Schelling goggles, it became clear that ignoring the threats just leads to escalation. The correct solution was making myself unavailable for threats. Blacklist the phone number, block the email, spend a lot of time out of home. If any messages get through, pretend I didn't receive them anyway. It worked. It felt kinda bad, but it worked.
One obvious "upgrade" to any decision theory that has such problems is to discard all of your knowledge (data, observations) before making any decisions (save for some structural knowledge to leave the decision algorithm nontrivial). For each decision that you make (using given decision algorithm) while knowing X, you can make a conditional decision (using the same decision algorithm) that says "If X, then A else B", and then recall whether X is actually true. This, for example, mends the particular failure of not being able to precommit (you remember that you are on the losing branch only after you've made the decision to do a certain disadvantageous action if you are on the losing branch).
You can claim that you are using such a decision theory and hence that I should find your precommitments credible, but if you have no way of proving this, then I shouldn't believe you, since it is to your advantage to have me believe you are using such a decision theory without actually using it.
From your earlier writings I think you might be assuming that AIs would be intelligent enough to just know what decision algorithms others are using, without any explicit proof procedure. I think that's an interesting possibility to consider, but not a very likely one. But maybe I'm missing something. If you wrote down any arguments in favor of this assumption, I'd be interested to see them.