Evidential Cooperation in Large Worlds: Potential Objections & FAQ
What is this post?
This post is a companion piece to recent posts on evidential cooperation in large worlds (ECL). We’ve noticed that in conversations about ECL, the same few initial confusions and objections tend to get brought up: we hope that this post will be useful as the place that lists and discusses these common objections. We also invite the reader to advance additional questions or objections of their own.
This FAQ does not need to be read in order. The reader is encouraged to look through the section headings and jump to those they find most interesting.
ECL seems very weird. Are you sure you haven’t, like, taken a wrong turn somewhere?
We don’t think so.
ECL, at its core, takes two reasonable ideas that by themselves are considered quite plausible by many—albeit not completely uncontroversial—and notices that when you combine them, you get something quite interesting and novel. Specifically, ECL combines “large world” with “noncausal decision theory.” Many people believe the universe/multiverse is large, but that it might as well be small because we can only causally influence, or be influenced by, a small, finite part of it. Meanwhile, many people think you should cooperate in a near twin prisoners’ dilemma, but that this is mostly a philosophical issue because near twin prisoners’ dilemmas rarely, if ever, happen in real life. Putting the two ideas together: once you consider the noncausal effects of your actions, the world being large is potentially a very big deal.[1]
Do I need to buy evidential decision theory for this to work?
There are some different ways of thinking that take into account acausal influence and explain it in different ways. These include evidential decision theory and functional decision theory, as mentioned in our “ECL explainer” post. Updatelessness and superrationality are two other concepts that might get you all or part of the way to this kind of acausal cooperation.
Evidential decision theory says that what matters is whether your choice gives you evidence about what the other agent will do.
For example, if you are interacting with a near-copy, then the similarity between the two of you is evidence that the two of you make the same choice.
Functional decision theory says that what matters is whether there is a logical connection between you and the other agent’s choices.
For example, if you are interacting with a copy, then the similarity between the two of you is reason to believe there is a strong logical connection.
That said, functional decision theory does not have a clear formalization, so it is not clear if and how this logical connection generalizes to dealing with merely near-copies (as opposed to full copies). Our best guess is that proponents of functional decision theory at least want the theory to recommend cooperating in the near twin prisoner’s dilemma.[2]
Updatelessness strengthens the case for cooperation. This is because updatelessness arguably increases the game-theoretic symmetry of many kinds of interactions, which is helpful to get agents employing some types of decision procedures (including evidential decision theory) to cooperate.[3]
Superrationality says that two rational thinkers considering the same problem will arrive at the same correct answer. So, what matters is common rationality.
In game theory situations like the prisoner’s dilemma, knowing that the two answers, or choices, will be the same might change the answer itself (e.g., cooperate-cooperate rather than defect-defect).
ECL was in fact originally named “multiverse-wide superrationality”.
We don’t take a stance in our “ECL explainer” piece on which of these decision theories, concepts, or others we do not list, is better overall. What matters is that they all (arguably) say that there is reason to cooperate in the near-twin prisoner’s dilemma and generalize from there to similar cases. Arguably, you might even be able to recover some behavior similar to ECL from causal decision theory, at least for ECL with different Everett branches (see Oesterheld, 2017, sec. 6.8).
Okay, I see the intuitive appeal of cooperating in the near twin prisoners’ dilemma, but I am still skeptical of noncausal decision theories. Is there any theoretical reason to take noncausal decision theories seriously?
In brief: Causal decision theory (CDT) falls short in multiple thought experiments.[4] This seems like reason enough to at least take noncausal decision theories seriously. It is worth mentioning, however, that in the academic literature, causal decision theory has a lot of support, and some academics actually take one-boxing in Newcomb’s problem as a decisive point against evidential decision theory.[5]
Details: One objection people could have is that we should not trust our intuitions in the near twin prisoner’s dilemma where causal decision theory arguably falls short. Because this thought experiment is highly contrived and unrealistic, our intuitions might be poorly trained for judging it. In contrast, we have a lot of experience with causation that tells us that whether something causes another thing is extremely relevant for decision-making. In fact, causation seems fundamental to the world: It seems like there is a matter of fact about what causes which other thing, and it seems silly to ignore that when making decisions.
While the debate between causal and noncausal decision theories is not settled, there are many arguments that point against causal decision theory and towards alternatives. We will sketch some prominent arguments against causal decision theory, but mostly advise readers to consult the sources below. Overall, we think the considerations at play do not justify high confidence in causal decision theory. There are of course many arguments in favor of causal decision theory and against any noncausal decision theory, which we don’t focus on here and which we advise readers to seek out if they want to engage with decision theory in detail.
(Note: The remainder of this section was kindly written by Sylvester Kollin.)
Why ain’cha rich?
In Newcomb’s problem, the act of two-boxing has lower average return than one-boxing, even from the causalist’s perspective. And therefore, it seems like the causalist is committed to a foreseeably worse option, which is arguably irrational.
See Lewis (1981) for the canonical reference; Ahmed and Price (2012, §1) and Ahmed (2014, §7.3.1) for a detailed characterization of the argument; and Joyce (1999, §5.1), Bales (2016), Wells (2017) and Oesterheld (2022) for counterarguments.
CDT is exploitable.
Causal decision theory (CDT) relies on some notion of a “causal probability” when making decisions.[6] However, looking ahead, they think about their own future choices as they would about any other proposition, and this yields an inconsistency in diachronic or sequential cases. Oesterheld and Conitzer (2021) use this fact to set up a (diachronic) Dutch book against CDT.[7]
See Spencer (2021), Ahmed (MS) and this post by Abram Demski for additional perspectives and formulations of what is essentially the same argument; Joyce (MS) for a discussion of how convincing the Dutch book argument is; and Rothfus (2021) for a formulation of “plan-based CDT” which avoids some of these problems. Also note that an updateless, cohesive (Mecham, 2010) or resolute (McClennen, 1990; Gauthier, 1997) version of CDT is not exploitable in this way.
CDT offers counterintuitive advice.
While opinions are split regarding what is rational in Newcomb’s problem[8], there are some other decision problems in which CDT clearly recommends the incorrect option. Consider the following decision problem from Ahmed (2014, p. 120)[9]:
Betting on the past. In my pocket (says Bob) I have a slip of paper on which is written a proposition P. You must choose between two bets. Bet 1 is a bet on P at 10:1 for a stake of one dollar. Bet 2 is a bet on P at 1:10 for a stake of ten dollars. So your pay-offs are as in [the table below]. Before you choose whether to take Bet 1 or Bet 2 I should tell you what P is. It is the proposition that the past state of the world was such as to cause you now to take Bet 2.
P | ¬P | |
Take Bet 1 | 10 | -1 |
Take Bet 2 | 1 | -10 |
Ahmed argues that any causal decision theory worthy of the name would recommend taking Bet 1, simply because taking Bet 1 causally dominates taking Bet 2.[10] But it is clearly irrational to take Bet 1: it seems obvious that if an agent decides to take Bet 2, then they should be very confident that P is true, and if they take Bet 1, then they should be very confident that P is false.
See Williamson and Sandgren (2021, forthcoming) for emendations of the standard theory to deal with this case, and Joyce (2016) for a critical discussion of whether this is a genuine decision problem.
Next, consider the following problem from Egan (2007):
The psychopath button. Paul is debating whether to press the “kill all psychopaths” button. It would, he thinks, be much better to live in a world with no psychopaths. Unfortunately, Paul is quite confident that only a psychopath would press such a button. Paul very strongly prefers living in a world with psychopaths to dying. Should Paul press the button?
CDT arguably recommends pressing the button, since this will have no causal effect on whether they are a psychopath or not. But they will instantly regret this: upon pressing the button, they are now confident that they’re a psychopath. Indeed, it seems irrational to press the button.
But note that it is much more contested what CDT actually recommends here than in the previous problem. See, e.g., Arntzenius (2008) and Joyce (2012).[11] See also Williamson (2019) for a defense of CDT’s recommendation.
EDT, medical Newcomb problems and the tickle defense.
One might grant that the recommendations of CDT might be counterintuitive at times, but that noncausal decision theories like evidential decision theory (EDT) also face its fair share of problems. Chief among these worries, at least traditionally, has been so-called “medical Newcomb’s problems”.[12] Consider the following problem from Egan (2007):
The smoking lesion. Susan is debating whether or not to smoke. She believes that smoking is strongly correlated with lung cancer, but only because there is a common cause—a condition that tends to cause both smoking and cancer. Once we fix the presence or absence of this condition, there is no additional correlation between smoking and cancer. Susan prefers smoking without cancer to not smoking without cancer, and she prefers smoking with cancer to not smoking with cancer. Should Susan smoke?
Most people agree that Susan should smoke (and this is indeed what CDT recommends). However, prima facie, it seems like EDT would recommend not smoking since that strongly correlates with having cancer, and having cancer and smoking is dispreferred to not having cancer and not smoking.
This has been forcefully rebutted by, e.g., Eells (1982, §6-8) and Ahmed (2014, §4). In short, the choice to smoke is plausibly preceded by a desire (a “tickle”) to do so, and at that point, smoking does not provide additional evidence of cancer. (This is known as “the tickle defense.”) See Oesterheld (2022) for a review and summary of these arguments.
Ahmed (2014, p. 88) writes (in relation to his discussion of the tickle defense): “The most important point to take forward from this discussion is the distinction between statistical correlations (relative frequencies) and subjective evidential relations in the form of conditional credences. It is the latter, not the former, that drive Evidential Decision Theory.”
Okay, so maybe I am not 100% confident in causal decision theory, but I am also not fully convinced by noncausal decision theory. What now?
Even if one isn’t convinced by the arguments against CDT, or by the arguments for EDT, one should plausibly have some decision-theoretic uncertainty (i.e., not be completely certain in CDT). MacAskill et al. (2021) then make the following argument: (i) we should deal with normative uncertainty by ‘maximizing expected choiceworthiness’; (ii) the stakes are much higher on EDT, due to the existence of correlated decision-makers; and so, (iii) even with a small credence in EDT, one should in practice follow the recommendations of EDT. MacAskill et al. call this the evidentialist’s wager. (See the paper for more details.) This argument could be extended to other decision theories with particularly high stakes. For commentary on this argument taken to its extreme, see the following section.
Is this a Pascal’s mugging?
In short: We don’t think so. As we explain in the above subsection, neither of ECL’s two core ingredients, “the world is large” and “noncausal effects can be action-guiding,” seem all that unlikely to be true.
In more words: There are a couple of cases in which one might think that ECL is a Pascal’s mugging:
1) One has very high credence in causal decision theory, and only takes ECL seriously because of the evidentialist’s wager. That is, one believes that one can only influence one’s causal environment, but one acts as though one has acausal influence because there’s a chance that acausal influence is real, and in those worlds, one has much more total influence.
We think Pascal’s mugging-type objections to lines of reasoning in this reference class are valid, and we share a discomfort with wagering on philosophical worldviews that we think are absurd. However, we do not think very high credence in causal decision theory is justified, which is required for this particular objection to get off the ground.
2) One thinks one’s acausal influence on other agents is small.
However, even if one’s acausal influence over individual agents is small, it would still be true that one’s total acausal influence is very large, without relying on wagers. A first reading of what it means to have a small acausal influence over an agent is: We definitely influence the agent’s action, but only by a small amount. Under this reading, influences still add up to a large effect over many agents.
A second reading is: We have a small probability of acausally influencing a given agent. So, if they only have two choices, like in the prisoner’s dilemma, then we only have a small chance of “changing” their decision. Under this reading, as long as our acausal influences over different agents are somewhat independent, then, given a very large number of agents, we should still have very high confidence that we are influencing a large number of agents. Oesterheld (2017a, p. 20, fn. 14) explains this second reading as follows: “Imagine drawing balls out of a box containing 1,000,000 balls. You are told that the probability of drawing a blue ball is only 1⁄1,000 and that the probabilities of different draws are independent. Given this information, you can tell with a high degree of certainty that there are quite a few blue balls in the box.”
We think belief in small acausal influence only leads to a Pascal’s mugging if you believe you have a low probability of influencing all agents and a high probability of influencing none of them (i.e., your acausal influences are fully correlated). (Apart from your acausal influence on exact copies, perhaps. But since they share your values, your influence on them doesn’t have any action-relevance anyway.) However, we don’t think that’s a likely view unless the belief is a consequence of being very confident in causal decision theory, which we’ve argued against in the previous section.
Should we be clueless about our acausal effects?
Complex cluelessness is a complicated topic that we aren’t experts on, so you should take what we say with a grain of salt. Nonetheless, here is our current view on the issue. In short: We think that cluelessness is a valid concern in general, but that it isn’t of much greater concern to ECL than it is to other cause areas and interventions.
Details: The cluelessness objection to ECL might look something like, “Okay, so if the world is large, then there will be lots of things going on in the world that we aren’t aware of,[13] and some of these things will be very weird to us. Perhaps we have acausal influence over some of these weird things that we don’t understand and this acausal influence has chaotic effects that could be any magnitude of good or bad. Then, it’s a mistake to not take this into account, because this could sign-flip the value of ECL.” While this is a valid concern, it remains regardless of whether you purposefully participate in ECL or choose to ignore ECL. If you believe the world is large and that acausal influence is (or might be) real, our actions have these unpredictable effects either way, whether we are purposefully focusing on them or only focusing on our predictable,[14] causal effects, e.g. by working on improving farm animal welfare.[15]
There is another version of the cluelessness objection to ECL that goes like this: “If the world is large and I have these acausal effects, then I am clueless. In the world where I am clueless, I am paralyzed and don’t know what to do. If the world is not large or I do not have any acausal effects, I am not clueless. I might as well wager on the scenario where I am not clueless and pretend the world is small and void of acausal effects.” This objection does indeed seem specific to ECL and to not necessarily apply to, say, alignment work. The objection comes down to a wager on not having acausal effects. We are personally not very comfortable with this wager and suspect that following the principle to always assume a world where one is not clueless would result in having to make many more uncomfortable wagers.
Does infinite ethics mess with ECL?
Infinite ethics seems even more complicated and confusing than acausal effects and complex cluelessness (see section above), and we haven’t spent a lot of time thinking about it. We will explain our current view on the issue, but you should take it with a grain of salt. Our immediate response here is to say that impossibility results (see Carlsmith, 2022, sec. V; Askell, 2018, sec. 5.1) around how to make choices in an infinite world are “everyone’s problem” and not just for people who consider ECL (Carlsmith, 2022, sec. XV, although Sandberg & Manheim, 2021, argue otherwise).[16]
On the other hand, largeness—perhaps infinite largeness—is one of ECL’s core ingredients. Therefore, it makes sense to inspect how ECL interacts with infinite ethics issues at least a little more closely. Example: If I gain one dollar, then, in an infinite universe with infinite copies of me, my value system gains infinite dollars, according to standard cardinal arithmetic. But if I were to gain two dollars instead, then, in the same infinite universe with infinite copies, standard cardinal arithmetic says this is no better than before—my value system still gains infinite dollars. This is a bizarre conclusion, and it’s problematic for ECL because it leads to different actions (e.g., defect to gain $1 vs. cooperate to gain $10) being evaluated as equally good.
No account of infinite ethics is fully satisfactory,[17] but some of the (clusters of) accounts—arguably the leading ones—rectify the conclusion here and say I’m better off in the second situation. The first of these accounts is based on expansionism/measure theory,[18] the second is about discounting in a semi-principled way (e.g., UDASSA),[19] and the third is Bostrom (2011, sec. 2.4–2.6)’s “hyperreal” approach.
Overall, we agree with those who point out that infinite ethics messes with everything, but we do not think that infinite ethics especially messes with ECL.[20][21]
Is ECL the same thing as acausal trade?
Typically, no. “Acausal trade” usually refers to a different mechanism: “I do this thing for you if you do this other thing for me.” Discussions of acausal trade often involve the agents attempting to simulate each other. In contrast, ECL flows through direct correlation: “If I do this, I learn that you are more likely to also do this.” For more, see Christiano (2022)’s discussion of correlation versus reciprocity, as well as Oesterheld (2017, sec. 6.1).
Is ECL basically a multiverse-wide moral trade?
If you define an agent’s morality as what that agent values, then, yes (kind of). The major caveat is that we can only do things for our ECL cooperation partners in our own light cone. If they value having ice cream themselves, we can’t help with that. The ECL mechanism—direct correlation—might also further constrain the space of possible trades.
In addition, we might not naturally describe many of the things other agents value (and that we want to cooperate with them on!) as moral values. For example, if there are agents out there who meet the conditions for trade and who really value there being lots of green triangles in our light cone because they think that’s funny, then ECL compels us to trade with them by making green triangles. One would be stretching things to call this cooperation a moral trade.
(Note: We are not moral philosophers by training, and it’s possible that some moral philosophers have a different definition of “moral trade” in mind. The EA Forum Wiki’s definition, which is the one we work with, reads: “Moral trade is the process where individuals or groups with different moral views agree to take actions or exchange resources in order to bring about outcomes which are better from the perspective of everyone involved” (link).)
Okay, I’ll grant that your toy setups work. But in the real world, we don’t have two neat actions with one labeled “cooperate.” We have lots of different available actions and compromise points. How is that supposed to work?
At a high level, we’ll note that it is common practice in science and philosophy to draw conclusions about the real world from toy setups. There are people who object to this practice, for instance, those who claim that game theory tells us nothing about the real world, but our overall stance on the issue is “all models are wrong, some are useful.” Our best guess is that the toy models we introduced in our ECL explainer are at least a little bit useful, and we expect more sophisticated models of acausal interactions to be more useful.
Nonetheless, the object-level question is still a good one. Uncertainty over what actions count as cooperations, and which cooperation point(s) to pick, is a real challenge to ECL. For instance, different agents’ understanding of the “default”/”do nothing” action might be quite different.[22] There is some work that tries to understand how agents might approach finding a cooperation point: see this paper by Johannes Treutlein and this post by Lukas Finnveden. On the other hand, we know of unpublished work that argues that the issue of different agents having different options from each other is severe enough to render ECL irrelevant.
We believe these issues only dampen the expected value of acting on ECL, by introducing uncertainty or constraining the set of agents we can cooperate with, instead of posing a fundamental problem to the framework. Do they dampen the expected value enough to make ECL irrelevant? Our current best guess is no. The most important reason for our view is that we are optimistic about the following:
The following action is quite natural and hence salient to many different agents: commit to henceforth doing your best to benefit the aggregate values of the agents you do ECL with.
Commitment of this type is possible.
All agents are in a reasonably similar situation to each other when it comes to deciding whether to make this abstract commitment.
If this is true, then it does not matter that our concrete available actions are extremely different from each other. We can still cooperate by acausally influencing other agents to make this commitment by making this commitment ourselves.
You keep talking about “agents” for us to “do ECL” with… Who exactly are these agents?
The short answer is: Some subset of the naturally evolved civilizations and artificial superintelligences that exist throughout the multiverse. Which subset exactly, though, is complicated and not entirely clear. Fortunately, even if we can only do ECL with a narrow set of agents, it might still be the case that their values are roughly representative of a larger set of agents that it’s easier to draw inferences about (e.g., their values might be representative of all civilizations that arose out of evolutionary processes, plus their digital descendants).
In the simplest model of how ECL might work, all agents who do ECL cooperate with all other agents who do ECL. If this is the case in reality, then one just needs to figure out what kind(s) of agents tend to do ECL in order to cooperate with these agents and thus participate in the ECL trading commons.
However, it seems plausible that not everyone who does ECL has acausal influence over everyone else who does ECL:
For example, maybe there are several different clusters of decision-making processes. Agents might have acausal influence over other agents within their own cluster, but not over agents from other clusters.
Maybe some agents’ epistemic standpoints are too different. For example, if distant superintelligences already know everything about us and our actions, then perhaps from their perspective our actions are already fixed, and so there is nothing left for them to (acausally) influence. See Finnveden (2023) for more detail on when and why knowledge like this might destroy acausal influence.
Updatelessness (Oesterheld, 2016; Soto, 2024) might help with this issue. We expect that anyone who wants to work on ECL will likely have to grapple somewhat with updatelessness or, alternatively, when to stop learning more about other civilizations.
Consequently, agents who do ECL might only be able to cooperate with a subset of the other agents who do ECL. If this is the case, then one needs to figure out what kind(s) of agents tend to do ECL and are similar enough to oneself in terms of decision-making and information.
As a final complication, not everyone we have acausal influence over necessarily has acausal influence over us. If Alice acausally influences Bob, and Bob acausally influences Charlie, and Charlie acausally influences Alice, should Alice try to benefit Bob’s values or Charlie’s values? Working from one of the toy models in our ECL explainer, we currently believe she should try to benefit Charlie’s values. Finnveden (2023) makes a related argument.
In practice, we recommend punting figuring out these details to the future. Humans now should focus on actions that are cooperative with a broad set of agents—i.e., actions that likely benefit the agents we can do ECL with without taking too much away from realizing our own values. At least, this is our best guess as to the present implications of ECL, given our present level of uncertainty. Examples of cooperative actions: making AGIs be cooperative (whether they be aligned or misaligned, and insofar as we build AGI at all); taking a more pluralistic stance towards morality than we otherwise would.
Is ECL action-relevant? Time-sensitive?
We believe ECL is action-relevant. We refer readers to Lukas Finnveden’s recent “Implications of evidential cooperation in large worlds”: rounding off a bunch of nuance, the key implication is that we should be trying hard to ensure that powerful, earth-originating AI systems—whether they be aligned or misaligned—engage in ECL.
Note that having our AIs engage in ECL is not simply punting ECL into the future. Working to increase the chance that future AIs engage in ECL is itself a salient action, in our view. Doing so puts us in the reference class of “civilizations that (genuinely) try to act cooperatively, in the ECL sense”: while not all civilizations in this reference class will succeed in implementing ECL, all will benefit from those that do succeed (to a first approximation, at least).[23]
The argument for time-sensitivity is that we might be able to increase the chance of future AI systems doing ECL in worlds where we cannot do so later, either because we build a misaligned AGI, or because the humans who end up in charge of an intent aligned AGI do not want their AGI to consider ECL, or because we manage to align an AI system’s values with ours but without guarding against grave epistemic errors from the distributional shift to thinking about ECL. This is most plausible on worldviews where views on decision theory, acausal influence, etc., are not convergent—i.e., where it is not the case that future people and/or AGI(s) will arrive at the same conclusions about ECL regardless of what we do this decade. Additionally, we might need to lock in some commitments early to reap the full benefits from ECL (e.g., commit to doing ECL sooner rather than later because our acausal influence might decrease over time—the argument here is similar to the argument for updatelessness).
That said, there are also potential downsides to working on making AI engage in ECL. In particular, this work might trade off with the probability of alignment success. It’s also contested for other reasons, though discussing the full case for and against is beyond the scope of this post.
How big a deal are the implications of ECL? How much does it change the impact of my actions?
This certainly deserves more investigation and its own detailed report. Unfortunately, we do not have anything like that yet. But here are a couple of things to note:
Lukas Finnveden (2023) estimates that ECL increases the case for working on AI systems that benefit other value systems by 1.5x–10x.
Note that this estimate does not include an estimate of the total value of this kind of work. Additionally, there is no comparison here to alignment work.
Paul Christiano says (paraphrased from two emails; all direct quotes but rearranged and in bullet point form for readability):
“I think that ECL or similar arguments may give us strong reasons to be generous to some other agents and value systems, especially other agents who do ECL (or something like that). I think this lines up with some more common-sensical criterion, like: it’s better to be nice to people who would have been nice to you if your situations had been reversed. I think that these may be large considerations.
If I had to guess, I think that figuring out what ECL recommends and then building an appropriate kind of AI might be 20% as good as building an aligned AI in expectation (with a reasonable chance of being close to 0% or 100%, but 20% as a wild guess of an average). [Later addendum by Paul: 20% seems too low.]
There are other ways the ECL stuff can add value on top of alignment.
It’s unclear if it’s tractable without just being able to align our AI (and this would be harder to get buy-in from the broader world). But we haven’t studied it much and there are some obvious approaches that might work, so maybe I’d wildly guess that it’s 20% as tractable as alignment. [Later addendum by Paul: 20% seems too low]
So, [if I had to make up a number right now,] I’d guess that “Figuring out what ECL recommends, how seriously we take that recommendation, and whether it’s tractable to work on” is 5-10% of the value of alignment. [...] By “tractable to work on” I meant actually getting it implemented. [...T]o clarify what the ratio [5-10%] means: I was imagining that an angel comes down and tells us all the ECL-relevant answers, and comparing that to how good it would be if an angel comes down and tells us all the alignment-relevant answers.
Caspar Oesterheld, author of ECL’s seminal paper, estimates that him acting on ECL is roughly as valuable for his values as him not acting on ECL and multiplying his resources (and, by necessary extension, the resources of agents with his values that he has strong acausal influence over) by 1.5 to 5 times. (From conversation.)
In other words, Caspar thinks you should be indifferent between:
A universe in which “your reference class” (i.e. agents that find themselves in a similar decision situation as you, but don’t necessarily share your values) acts on ECL, and
A universe in which your reference class doesn’t act on ECL, but agents in your reference class that share your values get their resources 1.5–5x’ed.
Acknowledgements
We are grateful to Caspar Oesterheld, Sylvester Kollin, Daniel Kokotajlo, Akash Wasil, Anthony DiGiovanni, Emery Cooper, and Tristan Cook for helpful feedback on an earlier version of this post. A special thanks goes to whoever created the great meme we use at the top.
- ^
H/T Daniel Kokotajlo for this framing.
- ^
While we believe proponents of functional decision theory—FDT—“want” for FDT to entail pretty broad forms of cooperation (e.g., the sort of cooperation that drives ECL), Kollin (2022) points out that FDT might not actually entail such cooperation. The implication being that FDT proponents might need to put forward a new decision theory that does have the cooperation properties they find desirable.
- ^
Updatelessness is closely related to updateless decision theory. Updateless decision theory, however, can be a bit of a quagmire because there doesn’t appear to be consensus on what exactly it refers to. Regardless of what is meant by updateless decision theory, the property of updatelessness can be used to augment decision theories like evidential decision theory (see, for example, this comment by Christiano).
- ^
Relatedly, see Christano (2018)’s exploration of why CDT has become the “default” decision theory despite being flawed.
- ^
And, presumably, against noncausal decision theories in general, though noncausal decision theories other than evidential are scarcely discussed in academia—you can find a list of some here. A significant minority of academic philosophers support evidential decision theory.
- ^
See Joyce (1990, §5) for a survey of different versions of CDT, as well as a general formulation.
- ^
Oesterheld and Contizer begin with the following decision problem:
Adversarial Offer: Two boxes, B1 and B2, are on offer. A (risk-neutral) buyer may purchase one or none of the boxes but not both. Each of the two boxes costs $1. Yesterday, the seller put $3 in each box that she predicted the buyer not to acquire. Both the seller and the buyer believe the seller’s prediction to be accurate with probability 0.75.
CDT recommends buying one of the boxes. But then here is the book/pump:
Adversarial Offer with Opt-Out: It is Monday. The buyer is scheduled to face the Adversarial Offer on Tuesday. He also knows that the seller’s prediction was already made on Sunday. As a courtesy to her customer, the seller approaches the buyer on Monday. She offers to not offer the boxes on Tuesday if the buyer pays her $0.20.
Here, CDT recommends paying the seller, despite being able to not pay anything, and not purchase any of the boxes.
- ^
See the 2020 PhilPapers Survey (Bourget & Chalmers), and also Oesterheld’s survey of earlier surveys (2017), for more on the precise numbers involved.
- ^
- ^
Hedden (2023) embraces Ahmed’s point, and makes a distinction between causal and counterfactual decision theory; and argues in favor of the latter, inter alia on the grounds that counterfactual decision theory does not recommend taking Bet 1 in Ahmed’s problem.
- ^
Although, see Ahmed (2014) for an even more damning problem, exploiting the instability of CDT, where the standard responses of Arntzenius and Joyce arguably do not work.
- ^
XOR Blackmail from Levinstein and Soares (2020, §2) is another decision problem which is seen as a counterargument to EDT. No tickle defense is available here. See Ahmed (2021, §3.4) for a discussion and defense of EDT.
- ^
Beyond unawareness, there are other things discussed under the heading of “deep uncertainty”. For example, I can be aware that my actions might have some effect on another agent’s behavior, yet I don’t feel comfortable at all placing sharp probabilities on the various possibilities.
- ^
Of course, many of our strictly causal effects lie beyond the limits of predictability (see chaos theory).
- ^
That is, unless you think our net acausal effects are in expectation zero if we focus on just causal effects, but unpredictable when we focus on acausal effects—this seems like an odd and incorrect position, though. Nevertheless, if one is worried about this, then this might be motivation to do foundational/philosophical work on ECL, or on cluelessness itself.
- ^
A concern we don’t discuss which also points towards infinite ethics being everyone’s problem goes something like: if the universe is infinite, then there is infinite happiness and infinite suffering, and so nothing anyone can do can affect the universe’s total happiness or suffering (Bostrom, 2011).
- ^
Carlsmith (2022): “Proposals for how to do this [compare different infinite values] tend to be some combination of: silent about tons of choices; in conflict with principles like ‘if you can help an infinity of people and harm no one, do it’; sensitive to arbitrary and/or intuitively irrelevant things; and otherwise unattractive/horrifying.”
- ^
This approach has been put forward in at least three different forms. One by Vallentyne and Kagan (1997) and a second by Arntzenius (2014)—these are referred to as “Catching up” and “Expected catching up”, respectively, in MacAskill et al. (2021, pp. 29-32)—and another by Bostrom (2011, sec. 2.3).
- ^
There is no discussion of UDASSA in the academic literature. Nonetheless, it appears to be endorsed by Evan Hubinger, Paul Christiano, and Holden Karnofsky, though Joe Carlsmith’s opinions on UDASSA are more mixed. Karnofsky says:
UDASSA, [the way it works] is you say, “I’m going to discount you by how long of a computer program I have to write to point to you.” And then you’re going to be like, “What the hell are you talking about? What computer program? In what language?” And I’m like, “Whatever language. Pick a language. It’ll work.” And you’re like, “But that’s so horrible. That’s so arbitrary. So if I picked Python versus I picked Ruby, then that’ll affect who I care about?” And I’m like, “Yeah it will; it’s all arbitrary. It’s all stupid, but at least you didn’t get screwed by the infinities.”
Anyway, I think if I were to take the closest thing to a beautiful, simple, utilitarian system that gets everything right, UDASSA would actually be my best guess. And it’s pretty unappealing, and most people who say they’re hardcore say they hate it. I think it’s the best contender. It’s better than actually adding everything up.
- ^
There are some accounts of infinite ethics that do especially mess with ECL, such as Bostrom (2011, sec. 3.2–3.4)’s “causal approach”.
- ^
See also Oesterheld (2017, sec. 6.10)’s brief discussion of ECL and infinite ethics.
- ^
Example: Say, the other agent can pay $1, which results in us getting $10, or they can do nothing (i.e., keep $1 for themself). We can direct-transfer any amount of dollars to them (so they get exactly the amount we pay), including nothing. We don’t know how much exactly we should transfer to maximize the chance of them paying $1, let alone to maximize expected dollars. Another complication is that the other agent’s understanding and our understanding of “do nothing” might be quite different. For example, maybe they have the option to pay $1 and us getting $10 (cooperate), to keep $1 and give us $1 at no cost to them (meaning both of us have $1 in the end), and to keep $1 without affecting our payoff. We might think the default “do nothing, don’t cooperate” action is for them to keep $1 and give us $1 at no cost. They might think the default “do nothing, don’t cooperate” action is for them to keep $1. This means we might be considering the decision from different starting positions, which might lead us to have very different ideas of what actions are equivalent to each other.
- ^
The “genuinely” bit is needed because if we only superficially try to get our advanced AI systems to do ECL, then, according to evidential decision theory, this is evidence that other civilizations like us will also only superficially try to do ECL.
Separately, it’s not as black-and-white as there being a single reference class for civilizations that tried to do ECL. For instance, different civilizations might put different levels of effort into doing ECL. Nonetheless, we think the reference class notion is a useful one.
Note also that the reference class is dynamic rather than static. In joining the reference class, we make it larger not only because we ourselves joined it, but also because we acausally influence other civilizations into joining it.
What are some ideas for how to increase the chance of future AI systems doing ECL?
An obvious approach is to give AIs a decision theory that is likely to recommend ECL, but given how many open problems there are in decision theory (as well as the apparent trajectory of research progress), I think we’re unlikely to solve it well enough in the relevant time-frame to be comfortable with letting AI use some human-specified decision theory to make highly consequential decisions like whether or not to do ECL (not to mention how exactly to do ECL). Instead it seems advisable to try to ensure that AI will be philosophically competent and then let it fully solve decision theory using its own superior intellect before making such highly consequential decisions.
I’m guessing you may have a different perspective or different ideas, and I’m curious to learn what they are.
Thanks! I actually agree with a lot of what you say. Lack of excitement about existing intervention ideas is part of the reason why I’m not all in on this agenda at the moment. Although in part I’m just bottlenecked by lack of technical expertise (and it’s not like people had great ideas for how to align AIs at the beginning of the field...), so I don’t want people to overupdate from “Chi doesn’t have great ideas.”
With that out of the way, here are some of my thoughts:
We can try to prevent silly path-dependencies in (controlled or uncontrolled i.e. misaligned) AIs. As a start, we can use DT benchmarks to study how DT endorsements and behaviour change under different conditions and how DT competence scales with size compared to other capabilities. I think humanity is unlikely to care a ton about AI’s DT views and there might be path-dependencies. So like, I guess I’m saying I agree with “let’s try to make the AI philosophically competent.”
This depends a lot on whether you think there are any path-dependencies conditional on ~solving alignment. Or if humanity will, over time, just be wise enough to figure everything out regardless of the starting point.
One source of silly path-dependencies is if AIs’ native DT depends on the training process and we want to de-bias against that. (See for example this or this for some research on what different training processes should incentivise.) Honestly, I have no idea how much things like that matter. Humans aren’t all CDT even though my very limited understanding of evolution is that it should, in the limit, incentivise CDT.
I think depending on what you think about the default of how AIs/AI-powered earth-originating civilisation will arrive at conclusions about ECL, you might think some nudging towards the DT views you favour is more or less justified. Maybe we can also find properties of DTs that we are more confident in (e.g. “does this or that in decision problem X” than whole specified DTs, which, yeah, I have no clue. Other than “probably not CDT.”
If the AI is uncontrolled/misaligned, there are things we can do to make it more likely it is interested in ECL, which I expect to be net good for the agents I try to acausally cooperate with. For example, maybe we can make misaligned AI’s utility function more likely to have diminishing returns or do something else that would make its values more porous. (I’m using the term in a somewhat broader way than Bostrom.)
This depends a lot on whether you think we have any influence over AIs we don’t fully control.
It might be important and mutable that future AIs don’t take any actions that decorrelate them with other agents (i.e. does things that decrease the AI’s acausal influence) before they discover and implement ECL. So, we might try to just make it aware of that early.
You might think that’s just not how correlation or updatelessness work, such that there’s no rush. Or that this is a potential source of value loss but a pretty negligible one.
Things that aren’t about making AIs more likely to do ECL: Something not mentioned, but there might be some trades that we have to do now. For example, maybe ECL makes it super important to be nice to AIs we’re training. (I am mostly lean no on this question (at least for “super important”) but it’s confusing.) I also find it plausible we want to do ECL with other pre-ASI civilisations who might or might not succeed at alignment and, if we succeed and they fail, part-optimise for their values. It’s unclear to me whether this requires us to get people to spiritually commit to this now before we know whether we’ll succeed at alignment or not. Or whether updatelessness somehow sorts this because if we (or the other civ) were to succeed at alignment, we would have seen that this is the right policy, and done this retroactively.
I’m skeptical about the extent to which these are actually different things. Oesterheld says “superrationality may be seen as a special case of acausal trade in which the agents’ knowledge implies the correlation directly, thus avoiding the need for explicit mutual modeling and the complications associated with it”. So at the very least, we can think of one as a subset of the other (though I think I’d actually classify it the other way round, with acausal trade being a special case of superrationality).
But it’s not just that. Consider an ECL model that concludes: “my decision is correlated with X’s decision, therefore I should cooperate”. But this conclusion also requires complicated recursive reasoning—specifically, reasons for thinking that the correlation holds even given that you’re taking the correlation into account when making your decision.
(E.g. suppose that you know that you were similar to X, except that you are doing ECL and X isn’t. But then ECL might break the previous correlation between you and X. So actually the ECL process needs to reason “the outcome of the decision process I’m currently doing is correlated with the outcome of the decision process that they’re doing”, and I think realistically finding a fixed point probably wouldn’t look that different from standard descriptions of acausal trade.)
This may be another example of the phenomenon Paul describes in his post on why EDT > CDT: although EDT is technically more correct, in practice you need to do something like CDT to reason robustly. (In this case ECL is EDT and acausal trade is more CDTish.)
I’m not sure I understand exactly what you’re saying, so I’m just gonna write some vaguely related things to classic acausal trade + ECL:
I’m actually really confused about the exact relationship between “classic” prediction-based acausal trade and ECL. And I think I tend to think about them as less crisply different than others. I’ve tried to unconfuse myself about that for a few hours some months ago and just ended up with a mess of a document. Some intuitive way to differentiate them:
ECL leverages the correlation between you and the other agent “directly.”
“Classic” prediction-based acausal trade leverages the correlation between you and the other agent’s prediction of you. (Which, intuitively, they are less in control of than their decision-making.
--> This doesn’t look like a fundamental difference between the mechanisms (and maybe there are in-betweeners? But I don’t know of any set-ups) but like...it makes a difference in practice or something?
On the recursion question:
I agree that ECL has this whole “I cooperate if I think that makes it more likely that they cooperate”, so there’s definitely also some prediction flavoured thing going on and often, the deliberation about whether they’ll be more likely to cooperate when you do will include “they think that I’m more likely to cooperate if they cooperate”. So it’s kind of recursive.
Note that ECL at least doesn’t strictly require that. You can in principle do ECL with rocks “My world model says that conditioning on me taking action X, the likelihood of this rock falling down is higher than if I condition on taking action Y.” Tbc, if action X isn’t “throw the rock” or something similar, that’s a pretty weird world model. You probably can’t do “classic” acausal trade with rocks?
Some more not well-in-order not thought-out somewhat incoherent thinking-out-loud random thoughts and intuitions:
More random and less coherent: Something something about how when you think of an agent using some meta-policy to answer the question “What object-level policy should I follow?”, there’s some intuitive sense in which ECL is recursive in the meta-policy while “classic” acausal trade is recursive in the object-level policy. I’m highly skeptical of this meta-policy object-level policy thing making sense though and also not confident in what I said about which type of trade is recursive in what.
Another intuitive difference is that with classic acausal trade, you usually want to verify whether the other agent is cooperating. In ECL you don’t. Also, something something about how it’s great to learn a lot about your trade partner for classic acausal trade and it’s bad for ECL? (I suspect that there’s nothing actually weird going on here and that this is because it’s about learning different kinds of things. But I haven’t thought about it enough to articulate the difference confidently and clearly.)
The concept of commitment race doesn’t seem to make much sense when thinking just about ECL and maybe nailing down where the difference comes from is interesting?
We’ve discussed this before, but I want to flag the following, both because I’m curious how much other readers share my reaction to the above and I want to elaborate a bit on my position:
The above seems to be a huge crux for how common and relevant to us ECL is. I’m glad you’ve made this claim explicit! (Credit to Em Cooper for making me aware of it originally.) And I’m also puzzled why it hasn’t been emphasized more in ECL-keen writings (as if it’s obvious?).
While I think this claim isn’t totally implausible (it’s an update in favor of ECL for me, overall), I’m unconvinced because:
I think genuinely intending to do X isn’t the same as making my future self do X. Now, of course my future self can just do X; it might feel very counterintuitive, but if a solid argument suggests this is the right decision, I like to think he’ll take that argument seriously. But we have to be careful here about what “X” my future self is doing:
Let’s say my future self finds himself in a concrete situation where he can take some action A that is much better for [broad range of values] than for his values.
If he does A, is he making it the case that current-me is committed to [help a broad range of values] (and therefore acausally making it the case that others in current-me’s situation act according to such a commitment)?
It’s not clear to me that he is. This is philosophically confusing, so I’m not confident in the following, but: I think the more plausible model of the situation is that future-me decides to do A in that concrete situation, and so others who make decisions like him in that concrete situation will do their analogue of A. His knowledge of the fact that his decision to do A wasn’t the output of argmax E(U_{broad range of values}) screens off the influence on current-me. (So your third bullet point wouldn’t hold.)
In principle I can do more crude nudges to make my future self more inclined to help different values, like immerse myself in communities with different values. But:
I’d want to be very wary about making irreversible values changes based on an argument that seems so philosophically complex, with various cruxes I might drastically change my mind on (including my poorly informed guesses about the values of others in my situation). An idealized agent could do a fancy conditional commitment like “change my values, but revert back to the old ones if I come to realize the argument in favor of this change was confused”; unfortunately I’m not such an agent.
I’d worry that the more concrete we get in specifying the decision of what crude nudges to make, the more idiosyncratic my decision situation becomes, such that, again, your third bullet point would no longer hold.
These crude nudges might be quite far from the full commitment we wanted in the first place.