I actually learned game theory from Binmore's textbook. So when Opus now said "Binmore says defect in twin PD", I got startled for a minute. Then I went and checked. Here's a full chapter from his "Game theory: a very short introduction":
Fallacy of the twins
Two rational people facing the same problem will necessarily choose the same action. So Alice and Bob will either both play hawk or both play dove in the Prisoner’s Dilemma. Since Alice prefers the outcome (dove, dove) to (hawk, hawk), she should therefore choose dove.
The fallacy is attractive because it would be correct if Alice and Bob were genetically identical twins, and we were talking about what genetically determined behaviour best promotes biological fitness (see Kin selection in Chapter 8). But the relevant game wouldn’t then be the Prisoner’s Dilemma; it would be a game with only one player.
As is commonplace when looking at fallacies of the Prisoner’s Dilemma, we are offered a correct analysis of the wrong game. The Prisoner’s Dilemma is a two-player game in which Alice and Bob choose their strategies independently. The twins fallacy wrongly assumes that Bob will make the same choice as Alice whatever strategy she chooses. This can’t be right, because Bob is supposedly rational and one of his two choices is irrational.
One can modify the assumptions of the fallacy so that Alice and Bob’s strategies coincide only with some sufficiently high probability. The story told to justify such a correlation in their behaviour often kicks up enough dust to obscure the fact that any correlation at all implies that Alice and Bob aren’t choosing independently. But if they don’t choose independently, they aren’t playing the Prisoner’s Dilemma. Even if Alice and Bob’s information were correlated, as hypothesized in Aumann’s notion of a correlated equilibrium, they still wouldn’t play hawk, because hawk is strongly dominated whatever the players may learn about other matters.
As you can see, he says cooperate in the literal twins case. Always check after LLMs!
That said, I agree with Opus that economists haven't cared that much about the perfectly correlated twins case, because it hasn't had much real applications so far. I only got interested in this case because Eliezer's writings made me wonder about cooperation between future AIs. (My first step was reinventing Tennenholtz's program equilibrium, which is well accepted academically and supports cooperation in twin PD.)
Summary: Scott Alexander’s Against SURA post is arguing against a strawman version of unilateral pauses. He admits as much deep into the post. I think his conclusions fail against stronger pause plans for conditional or limited unilateral pauses, where training resumes if the pause does not cause others to pause, and where personnel are redirected toward work on alignment and product, rather than simply fired. In addition, Scott’s post is incorrect about the simplicity of the game theory involved: getting companies to agree to conditional pauses is not strictly easier than convincing them to pause, as the former requires companies to (promise to) take actions based on ambiguous, game-theoretically-informed evidence. That is not the kind of evidence companies tend to act on.
The good version of unilateral pausing is good
Here is the good version of unilateral pausing:
There are multiple places in the post where Scott makes it clear he wants the reader to be thinking of much worse plans than this. Emphasis added:
Again, the correct version of a unilateral pause is a company pivoting to safety + security + product, not “disbanding their entire company immediately”.
This is again arguing against a strawman: OpenAI would not be “out of the industry” if they decided to unilaterally pause. Sam Altman loves to build products! He would very much still be in the industry.
It is possible that Scott will correctly point out that some people do advocate for the extreme version here, but I don’t think he should get credit for this. The title of the post is “Against Stop Unilaterally and Raise Awareness”. I think it is perfectly reasonable for someone to say that SURA-the-phrase applies to my better version. His post is very strongly being interpreted as arguing against unilateral stopping in general. It is being interpreted that way because of his title and framing.
Scott thinks the good version of unilateral pausing is good, or at least exempts it from criticism
Scott appears to agree that the good version of unilateral pausing is good, but attempts to claim that only one group is advocating for this good version, the AI Futures Project.
This should not be a parenthetical. If the goal is clarification for people who haven’t thought about the details much, and Scott believes as I do that there is a good version of unilateral pausing, it should have been in the opening section. It is not. The reader should not be expected to review every section before realizing that the title of the post is wrong.
Oversimplified game theory is bad, and you should feel bad
The core argument behind unilateral pausing is that the game theory is complicated, and it is better to go via deontology and other non-consequentialist analyses if you can’t model the game theory well. Scott reasonably points out that one shouldn’t just say this: he wants details. Unfortunately, in the absence of others providing details, he provides his own game theory and makes basic mistakes, providing evidence for the original why-deontology point.
In particular (emphasis added):
Agreeing to pause conditional on others pausing is not strictly easier than unilaterally stopping. There are three reasons for this: (1) “some definition” hides dragons, (2) promises for future actions can be harder than actions, even if the immediate actions are stronger, (3) the boundary between frontier lab and not is ambiguous.
“Some definition” hides dragons
There are several versions Scott could have meant by “some definition”:
It should be clear that getting a company to agree to the general version is out of the question: it is too wide a promise. Set that aside.
The internal definition version is useless without additional rounds of negotiation. If all companies come up with their own definitions and agree to act on them if others act on their own internally-generated definitions, the immediate outcome is no action. Additional rounds of negotiation then clearly falsify “strictly easier”.
The external definition version is also not strictly easier. It is in practice much harder for a company to agree on an external definition than to take action based on their own internal definition. This is a combination of not-invented-here syndrome and different definitions affecting different companies in different ways due to different team structures and research programs.
I am not guessing here. I have spoken directly to multiple people at all of Anthropic, OpenAI, and DeepMind about unilateral pauses, and they are all sketched out by this question of the precise definition of a pause. One of those people was the CEO of one of those labs. Hiding this complexity behind “some definition” is oversimplifying the game theory, and that results in bad analysis and bad conclusions.
Now of course, unilateral pausing also means definition mismatch and multiple rounds of negotiation. It might be the case that one company pauses according to definition D1, and another company refuses to follow suit unless they strengthen to definition D2. But accepting that we live in a multiround negotiation world is one of the core arguments behind unilateral pauses: the goal is to start a process that can then ratchet up through negotiation.
Promises can be harder than actions
It is not the case that promising to do X in the future based on condition Y is strictly easier than doing X now. This is because
The above are some abstract first principles. But they conform to how AI company executives speak about the issue, and how I’ve seen negotiations break down in the past: Executives are very uncomfortable making conditional promises.
As one example, DeepMind’s attempt to become a company limited by guarantee (CLG) failed in part for similar “reality fails to conform to simplistic game theory” reasons. DeepMind leadership was attempting to guard against “AI capabilities are crazy” worlds where risks are high, and the Google Board was attempting to guard against “AI fizzles and DeepMind wastes a bunch of money” worlds. Simplistically, there was an arbitrage to be made. In practice, negotiations fell through. One should be very careful when assuming that simple theorems of simple game theory apply.
How many frontier labs are there?
Another reason it is not strictly easier to do a Scott-version conditional pause is that it is not clear how many frontier labs there are. In the U.S., the number is somewhere between 2 and 6. There are clearly worlds in which getting 2 companies to do a unilateral and then bilateral action is easier than to get 6 parties to do even a much weaker action, and I claim we are in one of those worlds.
Definitional fairness is less important than survival
As mentioned, when I talk to people at AI companies, they are very concerned that if they stop, other people will stop slightly less, and it will be unfair. We can contrast the following two risks from AI:
The second risk is more important!
Now, one can argue that the first risk is important too, and indeed significant work should go into minimizing the unfairness, since that unfairness does impact pause effectiveness. Too much unfairness means we don’t slow down AI capability progress enough and don’t buy enough time. But I think that fairness is to be achieved over time, after multiple rounds of action, and should not be a precondition for taking the first step.
It may be much easier for an AI company to act on their own definition than by one imposed on them externally. So shifting the conversation to the relative unimportance of the definition used for the first action may be critical to actually causing that action to happen. This applies not just to unilateral or bilateral pauses but also to the grand international coordination that Scott prefers. In that context, it is notable that Scott’s only engagement with definitions is using the phrase “some definition”, handwaving a bit, and adding a footnote about forecaster wording.
Once we will fill in the subtleties around definitions and the details of the game theory, the correct action is to stop.