Nice post! I’m excited to see more attention being paid to multi-agent stuff recently.

A few miscellaneous points:

• I get the impression that the added complexity of multi- relative to single-agent systems has not been adequately factored into folks’ thinking about timelines / the difficulty of making AGI that is competent in a multipolar world. But I’m not confident in that.

• I think it’s possible that conflict / bargaining failure is a considerable source of existential risk, in addition to suffering risk. I don’t really have a view on how it compares to other sources, but I’d guess that it is somewhat underestimated, because of my impression that folks generally underestimate the difficulty of getting agents to get along (even if they are otherwise highly competent).

Neat post, I think this is an important distinction. It seems right that more homogeneity means less risk of bargaining failure, though I’m not sure yet how much.

Cooperation and coordination between different AIs is likely to be very easy as they are likely to be very structurally similar to each other if not share basically all of the same weights

In what ways does having similar architectures or weights help with cooperation between agents with different goals? A few things that come to mind:

• Having similar architectures might make it easier for agents to verify things about one another, which may reduce problems of private information and inability to credibly commit to negotiated agreements. But of course increased credibility is a double-edged sword as far as catastrophic bargaining failure is concerned, as it may make agents more likely to commit to carrying out coercive threats.
• Agents with more similar architectures / weights will tend to have more similar priors / ways of modeling their counterparts and as well as notions of fairness in bargaining, which reduces risk of bargaining failure . But as systems are modified or used to produce successor systems, they may be independently tuned to do things like represent their principal in bargaining situations. This tuning may introduce important divergenes in whatever default priors or notions of fairness were present in the initial mostly-identical systems. I don’t have much intuition for how large these divergences would be relative to those in a regime that started out more heterogeneous.
• If a technique for reducing bargaining failure only works if all of the bargainers use it (e.g., surrogate goals), then homogeneity could make it much more likely that all bargainers used the technique. On the other hand, it may be that such techniques would not be introduced until after the initial mostly-identical systems were modified / successor systems produced, in which case there might still need to be coordination on common adoption of the technique.

Also, the correlated success / failure point seems to apply to bargaining as well as alignment. For instance, multiple mesa-optimizers may be more likely under homogeneity, and if these have different mesa-objectives (perhaps due to being tuned by principals with different goals) then catastrophic bargaining failure may be more likely.

Makes sense. Though you could have deliberate coordinated training even after deployment. For instance, I'm particularly interested in the question of "how will agents learn to interact in high stakes circumstances which they will rarely encounter?" One could imagine the overseers of AI systems coordinating to fine-tune their systems in simulations of such encounters even after deployment. Not sure how plausible that is though.

The new summary looks good =) Although I second Michael Dennis' comment below, that the infinite regress of priors is avoided in standard game theory by specifying a common prior. Indeed the specification of this prior leads to a prior selection problem.

The formality of "priors / equilibria" doesn't have any benefit in this case (there aren't any theorems to be proven)

I’m not sure if you mean “there aren’t any theorems to be proven” or “any theorem that’s proven in this framework would be useless”. The former is false, e.g. there are things to prove about the construction of learning equilibria in various settings. I’m sympathetic with the latter criticism, though my own intuition is that working with the formalism will help uncover practically useful methods for promoting cooperation, and point to problems that might not be obvious otherwise. I'm trying to make progress in this direction in this paper, though I wouldn't yet call this practical.

The one benefit I see is that it signals that "no, even if we formalize it, the problem doesn't go away", to those people who think that once formalized sufficiently all problems go away via the magic of Bayesian reasoning

Yes, this is a major benefit I have in mind!

The strategy of agreeing on a joint welfare function is already a heuristic and isn't an optimal strategy; it feels very weird to suppose that initially a heuristic is used and then we suddenly switch to pure optimality

I’m not sure what you mean by “heuristic” or “optimality” here. I don’t know of any good notion of optimality which is independent of the other players, which is why there is an equilibrium selection problem. The welfare function selects among the many equilibria (i.e. it selects one which optimizes the welfare). I wouldn't call this a heuristic. There has to be some way to select among equilibria, and the welfare function is chosen such that the resulting equilibrium is acceptable by each of the principals' lights.

both players want to optimize the welfare function (making it a collaborative game)

The game is collaborative in the sense that a welfare function is optimized in equilibrium, but the principals will in general have different terminal goals (reward functions) and the equilibrium will be enforced with punishments (cf. tit-for-tat).

the issue is primarily that in a collaborative game, the optimal thing for you to do depends strongly on who your partner is, but you may not have a good understanding of who your partner is, and if you're wrong you can do arbitrarily poorly

Agreed, but there's the additional point that in the case of principals designing AI agents, the principals can (in theory) coordinate to ensure that the agents "know who their partner is". That is, they can coordinate on critical game-theoretic parameters of their respective agents.

Chimpanzees, crows, and dolphins are capable of impressive feats of higher intelligence, and I don’t think there’s any particular reason to think that Neanderthals are capable of doing anything qualitatively more impressive

This seems like a pretty cursory treatment of what seems like quite a complicated and contentious subject. A few possible counterexamples jump to mind. These are just things I remember coming across when browsing cognitive science sources over the years.

• Theory of mind;
• Imagination;
• Folk biology [https://en.wikipedia.org/wiki/Folk_biology];
• "Number sense".

My nonexpert sense is that it is at least controversial both how each of this is connected with language, and the extent to which nonhumans are capable of them.

Maybe pedantic but, couldn't we just look at the decision process as a sequence of episodes from the POMDP, and formulate the problem in terms of the regret incurred by our learning algorithm in this decision process? In particular, if catastrophic outcomes (i.e., ones which dominate the total regret) are possible, then a low-regret learning algorithm will have to be safe while still gathering some information that helps in future episodes. (On this view, the goal of safe exploration research is the same as the goal of learning generally: design low-regret learning algorithms. It's just that the distribution of rewards in some cases implies that low-regret learning algorithms have to be "safe" ones.)

Do you think focusing on s-risks leads to meaningfully different technical goals than focusing on other considerations?

I think it definitely leads to a difference in prioritization among the things one could study under the broad heading of AI safety. Hopefully this will be clear in the body of the agenda. And, some considerations around possible downsides of certain alignment work might be more salient to those focused on s-risk; the possibility that attempts at alignment with human values could lead to very bad “near misses” is an example. (I think some other EAF researchers have more developed views on this than myself.) But, in this document and my own current research I’ve tried to choose directions that are especially important from the s-risk perspective but which are also valuable by the lights of non-s-risk-focused folks working in the area.

[Just speaking for myself here]

I find myself someone confused by s-risks as defined here

For what it’s worth, EAF is currently deliberating about this definition and it might change soon.