Causality with Tadeg Quillien
2026-04-14 12:00:00.000 +00:00

I'm David Pinsov. I'm David Petrozowski. And this is Evolutionary Psychology, the podcast, Actual Evolutionary Psychology. By actual evolutionary psychologists. Wow, that was worthy of an Oscar. Okay. David today's guest is uh Tidek Quillian and he studies really complicated things that are abstract and difficult to grasp from an adaptationist perspective. Well yeah, you make it sound really exciting and appealing for the listener. Well I I honestly this is if you if you're if you're if you feel like you want a mental Rubik's Cube, this is the episode for you. This this is uh a a mentally challenging episode for sure. It certainly um expanded both of our minds. Um so yeah maybe if you're sleepy this isn't the best time to to dive into this meaty

podcast. I did I did uh uh I did rant about uh what a domain is though so there's yeah we we had some good rants in there about uh simplicity and domain specificity and what it means to be computational. Uh though this this one if if you're you know alert, you got a cup of coffee with you and you're and you think you can tackle this, this is a really um fascinating, insightful, mind expanding episode about the nature of causality, yeah. Uh which touches on literally everything the mind does, which is a point we get to in the episode. We could not do anything if we had no understanding of causality. Uh and and Taydeg just has a really uh elegant, insightful way of thinking about this complex topic.

Um and I I found a lot of value and learned a lot from uh talking to Tay Deg. Yeah. Uh it's a really cool example of cool uh work at the cutting edge of uh computational modeling and um sort of philosophy of causation. Um and uh I think it's a challenge. It's this is you should take this as a challenge to um uh grapple with these issues. These are these are chewy issues to deal with. Absolutely. Very chewy, very crunchy, um, very meaty. Uh I'm just gonna run through the food metaphors. Uh this is great. Um but uh but yeah, without further ado, here is Tidek uh and uh David and I talking about all things easy like causality. Tadek, tell us about where you are and how long you've been there. I'm in Scotland.

I'm at the University of Edinburgh. And it's been about five years now. There's like a storied history of Edinburgh and um computer science. And um yeah, is that true? So the I think the the first center for co cognitive science uh maybe in the world was in Edinburgh. So we had uh people like Nick Chater were here, uh people like Jeff Hinton, Peter Diane. Why Edinburgh and then they chose to dissolve the center for some reason. Why why Edinburgh? I mean not to be US centric about it or continental centric about it, but like why Edinburgh? Um how do you mean why uh such a what is the why the the the the answer I like is that maybe in the it's a city that has a long intellectual history and we have people like um people like David Hume that's yeah I was gonna say David

I was gonna do David Hume joke actually. So I I like to think that that is where the cognitive science energy is coming from. Yeah. David Hume. Have you read David Hume? I have. It's actually I started as a political science major. And there was a political philosophy class I had to take. And almost kind of randomly, I trust you to read a book by David Hume because we had to kind of pick a book and write a report about it. And that's, I think that was my gateway direct towards uh cognitive science. Yeah. Um just to be on this for a second, um, there are two opinions about David Hume, and I think both are right. And I'm curious where you stand. One is that he was not an empiricist, and one is that he was.

Where do you stand? Um I think I I stand with the kind of classical view that he was an empiricist. Um but I mean I can see the the pull of the opposing argument because uh so Immanuel Kant said that he was woken from his dogmatic slumber by reading David Hume, and Kant is famously not an empiricist. Yeah. Um and so I kind of read David Hume as um like someone who pushes empiricism to to its logical limits, and you find that if you push empiricism to its limits, then you're forced to uh quite an extreme degree of skepticism. Uh and then that's a reason in itself to be skeptical of empire of empiricism. For people who might not know, can we just make this kind of grounded and concrete? Like um just like you know, what what was the basic argument This is gonna turn into uh

uh um uh what what do you know about David Hume dial a friend? Uh well I mean at least for for my purposes, yeah. Um what I find interesting is his argument about about causality. Yeah. Um because um so David Hume was concerned with why it is that we can come to have causal knowledge at all. Okay, and so he observed that well, we can come to know that there is an association between uh the presence of clouds in the sky and then rain. And I know day after day we see that when there is when there are clouds in the sky, rain follows. And so we form by by induction this kind of sort of a strong association between those two things. But then it seems that in order to really uh come to the conclusion that there is a

causal relationship between those things, that the the rain causes the that the cloud is the cause of the rain, then there's an extra step that we need to take. And it is not clear that this take can be justified in in a rational way. So the idea is that so I've always seen clouds, then rain, clouds and rain. And then I say, look, clouds cause rain, and somebody asks me, Well, how do you know that? And I say, Well, they've always preceded that. And the argument is that well, yeah, that's your data, but where do you get where do you get the conclusion that that caused that? Maybe it was something else. Yeah, exactly. And it's quite difficult to justify that. Is this just the correlation causation fallacy?

Because you could say, well, no, clouds don't cause rain, rain causes clouds. Or maybe some other thing, maybe uh Zeus getting angry causes both, right? Um right. So how do we know that that it's called the causal arrow is going from clouds to rain? Is it is that it or is is is the the skepticism deeper than that? So I think this might be pushing beyond my ex philosophical expertise, but uh I I would say that maybe the worry is even a little bit deeper. Um the idea is that even if you can um sort of control exclude those alternative explanations. So even if you can definitely exclude that the causal direction is not from the rain to the clouds. Then you still have the question of what justifies me to think that there is a causal

relationship. And uh we're kind of getting into the territory of um it is a bit weird to posit something like a causal relationship. This is not something that we can see. Yes. It's not so there is a data. Yes, yes, there is a lip of safe that we have to take regardless. And I think like this is something I was worried about before we had this conversation, and I want to like clear away because when I when I I remember when I first started learning about like causation and things like that as an early student, I found it very obtuse, like this kind of thing where I'd be like, okay, yeah, whatever. Like, of course, clouds, clouds, rain, whatever. Why guys this is important? And I think it's worth like just taking a step back to just dispelling like why this is not

obtuse and why it's actually really important. And I think the cloud thing is a good example where my understanding is and like why I now find it not at all obtuse, but like it's fundamentally like the same non-bullshitty thing as like kids asking, well, why, why, why? Like that's a simple non-bullshitty thing. And my understanding is that it is kind of like David's thing where it's like clouds rain, clouds rain. You've always seen that relationship, but you don't know. It might be something else. And the problem of induction is really just that it's always correlation. Like that was kind of my read of Hume was like, it's always correlation. And it's weird that people don't represent that. Well, I'm I have a tentative guess that it is a correlation, but I should reserve in my

mind the fact that maybe that's just a correlation. And I don't have this idea that no matter what, and especially in the future, it's going to be the same way. Like I'm really taking a leap of faith in saying the way it is today is the same it's going to be tomorrow because I don't have all the information I need to make that leap of faith. And then therefore you could be wrong. And to my mind, like that's the problem of induction, is it's always correlation because there's an infinite number of causal relationships that the the correlation you've seen so far is not determinative of what the actual underlying causal thing, and that's the obtuse way. But the simple way of putting it is like if you don't know why something happens, you might

be wrong in the future. And that that's my read. And that sort of grounds it in like non-bullshit, like this is a fitness problem because like you gotta know like did the tiger eat the my friend for X reason or Y reason and can I be go back there tomorrow? Like that's not bullshit. That's not like, you know, well is it causation or correlation? But we know it always like the this is like really like um real um stuff here. I just so that's my um yeah clearing of that. Yeah the the often I mean not often hopefully but sometimes it will be a a life or death matter. Yeah, exactly. Yeah. I don't know. Yeah. One one example to you, one reason why it's important to represent causality instead of just a pattern of correlations, is that um storing

having a a predictive model of the world that is just a pattern of correlations um is not a very stable way of representing the world. So I mean you might um notice that there is a high correlation between um there being rain and then the grass in your in your garden being wet. And you store the fact that the correlation is maybe 0.3. So far so good. But maybe you install a sprinkler in your garden and to to to water the grass. Now once you do that, then the correlation between the rain and the grass being wet is going to be much lower. Yeah. And this is not something you know this is not something where you have a nice procedure where we can say, Well, I can just adjust the correlation based on this new information.

Yeah. There's nothing in the correlation that really tells you how you sh you're supposed to do that. Instead, if what's if your representation is a representation of the is a causal relationship between the rain and the and the the grass being wet, then you can incorporate this new information that there is a sprinkler in a much easier way. Yeah. It's like Paul Smaldino calls it uh being brittle, and he uses the example of science, right? So, like if you have this thing of like, I think we use the example, I think Noam Chomsky talks about this, is you could have a model, uh a correlational non-causal model of leaves falling, just describing the data. And that's what like a lot of science is, like, especially social science, it's just

describing the data. But then that's not going to work for bowling balls, how they fall. Um, that's not going to work for a lot of other things. It's very brittle. It only applies to that context. And you don't even know what the context is because you don't have a theory, because you don't have like um coefficients of drag and accelerations of gravity. You just have the descriptions of the relationships. And so the whole replication crisis, or like the this effect holds in this context, but not another one, is the scientific equivalent of exactly what you're talking about. Um that it's this, as Paul Smaldino says, it's this brittle thing that as soon as you try to pick it up and move it, it's like a tower of blocks and it just falls apart.

Yeah, and I mean to be clear, uh having a description is a a good starting point. Sometimes you sometimes you can't really start by having a nice theory, you need to gather a lot of data. Um, but it certainly cannot be the end goal of uh of a science. Um and yeah, and because we we should ideally be extracting the invariants, the things such that if uh if you look at a different context, you know that the invariants are still there and you understand how they work, um, and then you can make new new predictions. So I uh to me this is uh a lovely example of uh instinct blindness and and making the natural seem strange. Uh this idea that causality is not in the input itself, it's something that the mind

constructs in the same way that you know uh the mental states of other people, there's there's nothing in you looking uh in one direction that necessarily says that you are that you know what's over in that direction, right? That's something that I'm projecting onto you. Uh, if you're moving really uh quickly toward a particular object, I'm projecting the fact that you want that object, right? That that's not in the data. Um, in the same way as you know, I'm picking up this coffee mug and drinking it, and when I put it down on the table, there's nothing in the sensory stimulation on the coffee cup that says it's my coffee cup, that it belongs to me. Ownership is another thing that we project onto the world.

And what's uh lovely about your work is that it's we're doing the same thing with causality, right? And in some sense, we're constructing it. Um so to me, that leaves the question of how we are constructing. So, given that all we're getting is just uh, you know, photons hitting the retina, vibrations hitting our eardrums, how how do we take that information and construct causal relationships out of that data? Yeah. And can I add to that question really quick? What is it that we're constructing? Like what is it? Uh software, like what does it mean that we have a causal reper? That would be great. Thank you. Yeah. Yeah. Yeah. Yeah. And that's uh that's that's last question is the first question that we should ask,

obviously. Um so I mean I I take it and and a lot of people who are working in the field of causal cognition take it that the target, the goal is to construct a causal model uh of the world, or at least a causal model of the particular system that we're interested in. And so this causal model uh will be made up of of variables uh and then the causal relationships between them. So um if I have a a causal model of of the rain then maybe one of my variables will be the presence of clouds in the sky. Another variable will be the presence of rain. And then we will have another variable that is is the grass wet, for example. And then we can express the causal relationships between them. So what is the color the causal relationships between the rain and the clouds?

And then what is the causal relationships between the whether the grass is wet uh and and the rain. And there's various formalisms for how you might represent those causal relationships. And we don't need to get into the the details. But basically something that is quite um important and that is kind of foundational to the way that we think about causality is that the relationship between two variables is asymmetric, right? So it's not just that we know that there is an association between uh clouds and rain. It's just that there's this asymmetry that it is the clouds that are causing the rain and and not vice versa. Um and so one way to actually represent that pact concretely is by building into that the model this commitment that if we were to intervene on the on the

clouds, then this will have a downstream effect on rain. So if by some if I had some magical powers and I was able to to prevent the clouds from forming, then there would not be any rain. But if I were to kind of uh if I were to build a giant dome over the city that shelters us from the rain, this would not make the clouds go away. And so that is the way that we kind of build in this important distinction between something like X causing Y and Y causing X. So just to we can keep going on, but just to clarify. So what I'm hearing is because it's hard to talk about this, because this is again where natural language fails us, is that if we snip out the word causal, if we're trying to explain what a causal relationship is and we snip out because we don't want to use the term

in the explanation. It's a relationship that excuse me is um it's a relationship that has like directionality and then also it sort of picks out um what what is likely to if one thing is present and another thing will be present in sort of a compositional order-dependent way. So it's sort of saying, you know, like when this when this bit of the world is there. I don't even know how to talk about it, but yeah, right? Like, is that is that fair? Like it's sort of that, you know, this this is necessary. I don't even like using the word necessary, but it's just it's these placeholders for um you can't change the order, and then if you manipulate this thing, this other thing, they're kind of attached to each other in terms of

how comes of the thing. This this notion of manip manipulability is quite important. So sometimes we talk about interventions of manipulate manipulability, etc. But it's this idea that if if someone could kind of reach into the system from the outside and change the state of change what happens to uh at some node in the network, then you will have particular effects somewhere. Right. I like that we not example yeah, I like the example of moon and tides. Like nobody's ever nobody's ever grabbed the moon. But it but if but the theory is the the causal theory of moon and tides is if somebody grabbed the moon and moved it, the tides would be different. Yes, and it's quite remarkable that we can entertain this despite the obvious imp

impracticality of actually moving the moon. It's something that's quite easy to imagine to think about. And if you were to explain to someone how tides work, then that would be one way to to do that. You can say, well yes, if we move the moon then the tide would happen in a different way. And I guess you're the person that you're trying to teach to would not be bewildered at at the idea. I don't know. Maybe maybe in certain cultures this this is a bit strange to teach things in that way. But I guess the general yeah, the general idea is quite intuitive. So it seems it seems like what you're saying is that causality is ultimately about hypothetical actions. And that if I think A causes B, what that means is just uh I'm just translating that into

the language of hypothetical things I can do. If I add A, I'm gonna get B. If I take away A, uh then B goes away. Right? These are hypothetical, but e even if I'm in incapable of taking away A or adding A, if I you know, if if I don't have magical powers, I can't do anything about the clouds. Uh merely the the ability to entertain uh the the possible action of adding or removing A will have an effect. I'm using up the word effect, but will change B. Um, that is ultimately what causality is. It's framed in terms of hypothetical interventions or actions or impacts you can have on the world, uh, and what is going to be changed in the world as a result. Again, I'm using the word as a result, which gets us back into causality.

But what so it's a it's about hypothetical actions. Right? Yes. Uh yes, yes. That's uh it's a great way of explaining it. Um And I think that the the fact that we are we we are quite uh able to entertain those hypothetical actions. Um I think this is one of the best psychological evidence we have that people are constructing causal models of the world. They're not just passively trying to store patterns of correlations in in that that results from their experience, right? Because I have never experienced a state of the world where the moon has been kind of changed course. But I can imagine a manipulation that would do so. Maybe a better example is I have never imagined the state of the world where the sun

suddenly disappears. But I can quite easily imagine what would happen if a sort of an action, a hypothetical action, was taken and the sun goes out. And this, I mean, if you are if you are looking at the mind from a sort of behaviorist perspective, where all you're doing is keeping track of associations, maybe keeping track of the results of your own actions, you would not really predict that people have the ability to engage in that kind of thinking. And you'd you don't you would even less predict the fact that people actually quite like thinking about those those sorts of things. Yeah. It's also a point like I I don't know who made this point, or maybe other whatever. Um, is uh the associationism never really talked about what is done with those

representations either. Like it was very much uh um explain the data given the study. But but yeah, like the problem, like a good example also is the desert ant where you know it goes out, it has polar light sensitive thing, it does dead reckoning where it encodes distance by steps and uh angle of the sun as it moves. And then it does that integration when it finds prey and goes back to the nest. You know, if you just say, well, it's just associating, um, it's just associating stuff, like there's no notion of like what the association is for. But the second you have this thing of like, well, the whole point of doing the association is creating some representation of where the nest is, That's very much like a model, right?

Um, that's integrating stuff for a purpose and it allows you to predict and go on a novel d dead point uh line to where the nest is so you don't fry yourself in the desert, which is why it's a selection pressure. But no, I like that that this is very evolutionary um um in a way is in is this causal model uh literature evolutionary and what does evolution buy you that may not be in the literature or or is it i don't know yeah is that a coherent yeah no that that that's that's a great question i i think um i would say that the causal cognition literature is implicitly evolutionary. So it definitely has a functionalist band. So people who look at causal cognition, they they ask, well no, what is the purpose of the

system? Why why should we have those computations? And and it's really coming committed to this idea that people are building representations. And so uh there's also this uh very strong commitment to to optimality, at least as a as a sort of methodological heuristic. So because I mean once you have this machinery of causal models, you can ask, well, what is the best way to work with a causal model? So for example, you're trying to learn well, given all the data that I have, what is the probability that um the what is the probability that it is raining right now, given that all I can observe is that the grass is wet. And this is something that you can solve using the mathematics of uh of probability theory.

So it's basically a Bayesian inference problem. So you have to set up um Bayes' theorem, and then you can work out what is the probability of a cause given the effect. Um and so we can look at whether people actually um Compute, engage in causal reasoning in that way. To what extent is it the case that when people sorry, when people make those inferences, there are actually reasoning in the normative way that is described by probability theory. have hypotheses that can be tested. And from my own point of view, as an evolutionary psychologist, I think that the another justification for the optimality analysis is that you should expect to have first order of approximation that people will solve those problems in uh in the

normatively correct way because they have been a relatively high selection pressure for people to really um be good at causal reasoning. So that's the first part of the of the answer. The second part of the answer is that people who do causal recognition, they are often not um they're not quite concerned about domain specificity in the way that most of us would be concerned. So we're not going to to ask well no what are the the principles that people draw on when they think about biology, for example. And one, I guess, one motivation behind this choice is that there are general principles of causality that are supposed to apply across the board. And so this idea that causality is a matter of intervention, and that you need to

engage in some sort of Bayesian reasoning in order to engage in causal reasoning, those are things that will apply regardless of whether you are thinking about psychological reasoning or biological biological reasoning or physical physical reasoning. And so I mean to us at least from the outside this can appear as something that is very domain general. But I would argue that there is a justification to that. Because you can think of the domain as being causality and you have certain invariants of uh of good causal reasoning that will apply across the board. Yeah. Yeah, don't yeah, yeah that's great. Yeah, yeah. It's I mean it's I mean it's interesting. If if um causality is ultimately about things I can do, like what will

happen if I do this, uh well that's basically almost pretty much all of what the brain does is trying to figure out what I should do. What will happen if I do this. Yeah, you can't do it. Like it's a it's a machine for figuring out what to do. And it seems like causality should be literally a part of every system in the mind, because every system in the mind is ultimately about what should I do next, right? What's gonna happen if I do this? What's gonna happen if I do that? Um so I think you could think about it as you know ultra-domain general generality, but I don't think that's a really helpful way to think about it. I think it's it's better to think about in terms of functional specificity.

We have systems that are functionally designed for the purpose of picking out causal relationships. Um, and that should by design affect every part of the mind that is involved in figuring out what to do next. Yeah. And and can I just like elaborate too? Like I think like domain generality is or just the whole issue of domain, like the word domain, is if you look at how people use it, I think it's exceedingly unclear. Um if you look at how Fodor uses it, it's very folk psychological, where like a domain is like a thing out in the world. Like, is it like cheese versus like cows? Right, like that's a domain. There's remarkably little like notion of like once you start thinking about the psychology,

what do you mean by domain at all? And I would put it as like what David just described, which is exactly right, is you have all these systems for representing all the stuff out in the world. And all of those systems are all very specific for making all those representations possible while you're representing objects or the social world or physical relationships or emotion, mental states, like all these things in specialized machinery. once all those very specific systems are all working, they're also generating formats in this like event system that can take many of these inputs in because of all this very specific machinery working. And it's that event system being fed all of this content where

that that has this rule, these rules of sort of like, um, you know, well, if I do this, this is the other thing that's going to come along with it. And then these other things don't matter, which applies to lots of relationships in the world. Yeah, that's common to all of them, but it's that there's more and more specialized machinery. There's even machinery for that job. I mean, it might get replicated across systems, who knows? But I think this is a thing where, like, if you say it's domain general in the sense of it applies to cows and cheese, it sort of sounds like, well, one mechanism does cows and cheese, but it's actually there's a ton of mechan, there's even more mechanisms than just if there was one mechanism for cows and a different mechanism for cheese.

It's actually there's more mechanisms that allow something to be domain general in the in the unclear way than the domain specific way. I don't know if that's helpful for this, but I just always have this rant of like, my God, domain general versus domain specific is like it's it's I think it's even worse than innate and learned um or it's just as bad. Um yeah. Yeah yeah I mean yeah I I agree with that. Um I think to some extent this this was a thing that um it's a it's a nice it's a nice set of concepts to use if you're not already being uh in the in the process of thinking about things in uh in a functional way. Yes, hundred percent. It's kind of a gateway drug. Yeah, it's a gateway functionality.

Where you can use people's intuitive understanding about the world 100%. Yeah, yeah. So uh you mentioned that um uh uh we uh have looked at how well people conform to normative standards of causal reasoning or Bayesian reasoning. And you mentioned that um there are some cases where we fall short of the normative standard. What what are some of those cases and and and what do you think might be going on there? Uh yeah. So there are many things we can could say about that. And this this could be its own discussion. But I think maybe I can talk about two case studies that I think are quite interesting. Um and so the first case study actually illustrates the fact that we are our brain is uh wired up to reason to represent the world in

causal terms. Um so if you if you read psychology from the 70s, 80s, um there's a bit of a pessimistic picture that emerges about uh how good people are at at reasoning. Uh so I know there are studies by people like Tversky and Kahneman. Um, and one classic classic example of that is the um the diagnostic test problem. And so in that problem, you're told that there's this disease that maybe one in a thousand people have. And there's this test to see if you have the disease. And you're told that the test has a certain false positive rate and it has a certain false negative rate. And putting all this information together, you can in principle compute what is the probability that you have the disease, given that you were tested positive to the to the

disease. Um and when you ask participants uh you know what is this probability, um typically they will say something like oh like 80%. They will say something pretty high, even if the numbers are set up in such a way that the answer is actually something like seven percent. And so that is a huge gap between the normative Bayesian answer and the correct answer. Um and so when you look at that, you can maybe conclude, well, it looks like people are actually terrible at probabilistic reasoning. Um, but then if you look at the way this problem is set up, you are never really given an explanation for why there might be false positives. You kind of just told that well, sometimes the test doesn't work very well.

And so it is difficult to kind of for the participants to construct a causal model of the situation that they have to reason about. And so one hypothesis that that is just the explanation for why this is a hard problem. And so some psychologists have done an experiment where modify this problem in a slight way, and the only thing that they do is that they say, well, sometimes people, you know, like I don't know, maybe 70% of people have a benign uh keist that will cause a false positive. And when you reformulate this problem in that way, then people suddenly they they have a sort of a causal explanation that is available to account for the presence of false positives. And now their answers get much closer to the Bayesian optimum.

And so it seems that people are able to engage in in good Bayesian reasoning, but they need to form a causal representation in order to do that. And so yeah. Yeah, sorry to in that example too, just really quick, I think there's also additional work showing that if you do this with frequencies, so you show people this is the number of people who have false positives, this is the number of people who have the correct hits, it also goes away in the sense of the Tversky and Kahneman results was actually very improbable and it was really designed. My friend and colleague Tomas Leharaga has a paper in Psych Bulletin about how people used to see people say that people were good statisticians, and Tversky Kahneman reversed that,

but they were using very abstract, uh non-evolutionarily valid sort of cues of like, you know, not using whole objects or frequencies and sort of giving people uh counterintuitive things. So I just want to say like that's really interesting that even just changing the uh adding the causal model is like another problem with that methodology. So yeah, yeah. Um I mean and and to some extent I I think that First Ken Ganneman um they still did a service to the field by pointing out that there are those cases where people do really really badly, right? Because then see something that is crying for explanation. And then when we explain why it is it that people do very badly, we learn new things about

how the mind works. principle, it's almost um well, it is actually an axiom of causal reasoning that is called the the Markov condition. So it's named after this mathematician uh from the early 20th century called Andrei Markov, who studied very simple systems where uh the state of the system in uh in a particular uh time just depends on the previous state of the system. And this is relevant to causality because if you have a causal system where suppose you have the again the clouds cause the rain and the rain causes the grass to be wet. If that is all there is to your causal system, then the clouds can have an influence on the grass being wet. So if if there are clouds, the the grass are more likely to be wet.

But this the whole causal influence of the clouds can only go uh via the the rain. Okay so the the the only reason that the clouds can make the rain wet is can make the the grass wet is is via the is by making it rain. And so there's um from um like from from the sort of perspective of of causal reasoning um then the the Markov condition is just a fancy way of saying that once you know whether it has rained or not, then anything that you learn about whether the clouds are there or not should not give you any more information about whether the grass is wet. You got the most important piece of information you get all the information right. And sometimes we say that the the rain screens off the influence of the clouds.

And so what this means is that if you asked well, what is the probability that the grass is wet, given that we know that it's it's been raining, you should give some answer, right? And then if later on the experimenter asks you, Well, what is the probability that the grass is wet given that it has rain and there are clouds. You should give the exact same answer. But people don't really do that. People will give a higher probability if you tell them that, in addition to the fact that there is rain, there are also clouds. So it seems like a blatant violation of the rules of how you should reason with causal networks. And so this has gotten people in the field quite interested because is consistent with the fact that people are actually reasoning with causal models.

when they need to engage in probabilistic reasoning, um they make uh use of approximations. Um and the the reason why you need to engage in approximations is that computing a probability will in general be relatively um expensive in computational terms. Um right you need to uh you need to engage in quite complex uh computations um and there are actually shortcuts that people who do computational statistics and machine learning typically use in order to make the computations easier. And it turns out that if you assume that people are engaging in the same kind of shortcuts, then you predict that they will make that kind of mistake. And so the picture that emerges from that is not that people are don't know the rules of

causal models, but they they use some approximations in order to compute things in an efficient way. I'm I'm a little bit confused. I mean I hear what you're saying, but let me just let me be devil's advocate for the um so like with the rain and the clouds, is the is the study designed that like you tell the same person rain and then the same person rain in clouds and which one do you think is more likely that the grass is wet? Um so there are various ways that the experiment has been done. And of course uh when people do those experiments, they they use a more controlled setup where you say, well, there's this machine, and then there's this component of the machine that can influence this component that then can influence that component.

But why is that so why is it wrong for me to be more confidently grass is wet if I say uh clouds uh and rain as opposed to rain? Um so if I ask you, you know, suppose you know that there is rain, yeah, what is the probability that the grass is wet? Um, then the the answer you give to that should be the same as if I ask you, suppose you know that there is rain and there are clouds. Yeah, but I but I would push back on that because like that ignores that like is it's very like um uh I mean I don't know, I'm not saying you're doing this, but I'm saying like the causal model people, like there's a difference between like there's a little bit of rain versus like there's a storm, right? So like there's differing I mean, in some sense, people Yeah.

I mean, just to take a step back, like causality doesn't exist. It's a representation of the mind of like how certain am I likely that this event will produce this event? And if you tell me rain versus rain and clouds, um, how likely is it that the r the water, the grass is drenched? how much water is there? How likely is it that that was rain and not like a bird peeing or something? You know, like that's r all relevant. And if you tell me like there there's a there was a little bit of rain versus like there's rain in clouds, like you just if there's all this common sense knowledge that I think a smart computer would say it's much more likely that whatever you're trying to communicate of like how likely is it the grass is wet, how wet it got, how

confident you were that like, you know, it was a rainstorm. All these things make sense. Yeah. Or or even like the person who's telling you that it's raining without clouds, like maybe you should not trust that person. Maybe they were wrong. Like if you've ever spelt rain when there's no clouds, you're like, what is this? And it does happen, right? Because like the wind is strong. So like I I I don't I mean maybe maybe we're we're hanging a lot on this example, but like it seems what you're saying scientists are wrong. I think that those are those are good objections to sort of the experiment in the way that I described it. Yeah, exactly. Um of course in in in practice when people do those experiments you have a much more

controlled setting where you really make it clear to people that well those are just binary variables and really the only way that this like the the end end variable, the only way that this could be affected is by the state of the previous variable. And so that way you you bypass that kind of concern about well what inferences people are making from the fact that you tell them many things and so on and so forth. But isn't it still rational in the sense of like shouldn't an evolved psychology represent because, like, I mean, from an evolved psychology perspective, it should acknowledge that all it is, all it can do is pick up on correlations, right? Like, like the evolved psychology is is only ever

detecting co occurrences and is using the fact that some co occurrences are actionable for intervention and actually generally produce the right outcome, right? Like, so that's what we mean by causation, but it's all a guess. Um so I it strikes me as like a I mean maybe this is I maybe I should get more concrete, but I don't know. I just I feel like the the scientists are wrong in this case and that it's a smart system to to be uh if more if more predictors are present, you should increase your likelihood. Yeah. I mean I have I have a a version of arguments that I've made in in print, which is that um so I mean I I buy the this idea that I described that kind of if people engage with in in shortcuts in their computations, then they will make the kind of mistakes that have been

observed. Um but a different way that those kinds of mistakes will be produced is that if people don't completely buy the story that has been given to them by the experimenter. So maybe they think, well, you know, the experimenter tells me what those variables are causally related, but I don't know, maybe they're not, or maybe they're related but in a different way than has been told to me. And so if that's what you think, then the smart move is to just maintain those different hypotheses in mind and then and then to average over them. So you're not you're telling me that people aren't assuming that experimenters and psychology experiments are infallible and uh omniscient. What if you let people do the experiment themselves?

Like, has anybody ever done that? Where like you actually give people the ability to actually test the relationships themselves? Like the experience, the experiential format of teasing apart what causes what, as opposed to somebody telling you, because actually, thank you. I f I felt like I was humming and hawing, but that's exactly what I was doing. I was like, I don't believe you. That's exactly what I was doing. And and and what what we what we've have shown in in that paper is that well if you if you assume that people have this uncertainty and that they are averaging over that uncertainty, then you predict that the that they will make the kind of quote unquote mistakes that have been observed in these experiments.

But to answer your question, I don't think that that experiment has been run where you really allow participants to manipulate the causal structure themselves. That's it, yeah, it would be interesting to perform that experiment, that's for sure. Yeah. So is your take that um this idea that humans are bad at causal reasoning or they're irrational, uh, that we deviate from normative standards. Is your take that this idea is overblown and that actually, if you dig into the details and operationalize things in a sensible way and in an ecologically valid way, that actually people are pretty good at causal reasoning and and and fit pretty closely to what you might expect uh an optimal system to look like?

Yeah, yeah. And uh I think that's a good summary of of what I think. Um and if you were to try to to come up with a taxonomy of possible sources of mistakes that people make, um I would say that there's maybe two main sources of mistakes that people can for why people might make mistakes. Um and the first is this kind of representational issue where people just either don't really believe you or they don't uh they're not able to construct a causal model that's that would actually allow them to to solve the the to solve the statistical problem. Um there's a different way that they could be wrong is that if the problem is super complicated then then of course you I mean it's it's kind of it's a bit of a trivial point

in a way. It's like um asking people if I if I ask you you know what is the 17th digit of pi and you give me an answer that is no better than chance, there's a sense in which you're not being normative. But it's just that the computational demands that I put on you are too high. And so that that is one way that that you can make mistakes. And I mean one one other domain in which I have shown uh that people are relatively quite they're quite good at following the normative principles of probabilistic inference is um to in the social domain. So when they have to infer the welfare trade-off ratio of other people, then they have to solve a problem that is actually quite sophisticated. If you were to think of, you know, they have to make probabilistic computations over a

causal model of how other people make decisions. And yes, despite the complexity of the problem, they're actually quite good at uh inferring how much someone else values them. So this is a a series of experiments that we've done with uh with John and Lida, um, where the simplified version of this experiment is that you see someone kind of uh sharing a pot of money uh in a particular context uh and you see the decision that they have made. Uh and then there's a slightly different context where maybe the cost to them of giving you money is a bit higher. And I um and I asked. Something, yeah, yeah. It's a good picture. And so in that in that context, so you know what they have done in that previous situation.

And I ask you, what is the probability that they will give the food to you? And so you you can compute what is the like given some assumptions, what is the like the normative probability that uh a Bayesian model would give. And what we find is that first people are quite good, the people's answers are quite aligned with the answers that are given by the Bayesian model. And the second cool thing that we have found is that we also asked people well how grateful are you when the person actually because you actually at the end you you learn what they have what they have done. And so if the if the person in the experiment gives you uh gives you something. We ask people, well how grateful are you that you are given that thing?

And if they decide to take the selfish action, we ask, well how angry do you feel? And then what we found is that the best predictor of how angry or grateful the people feel is not the best predictor is not how much they get or how much they fail to get. Because we control for those things tightly. So like the the kind of the payoffs that are involved or controlled across trials um and the sort of opportunity cost is also controlled. But what we can do is we can actually compute the welfare trade-off ratio that an ideal Bayesian observer would infer the other person has on the basis of the decisions that you observe. And then this inferred welfare trade-off ratio is actually quite a good predictor of

gratitude and anger. So basically if you infer that if even if a if an uh an ideal Bayesian observer infers that yes, this person does not care at all about me, then I'm going to be quite hang angry. Even if you end up getting a lot. Yes, yeah. Yeah. But yeah, so it's kind of the the intuition is that suppose at some point they give me a lot, but because it was trivial for them to give me. Yeah. The cost of was super super low. And then they were gonna throw it in the garbage or whatever, and you just happen to be there. Yeah, whatever. Yeah, exactly. Yeah, exactly. And then on the other occasion, they kind of they could have benefited me by a little bit, uh, at again a trivial cost to themselves.

Um, and they decide, oh yeah, you know what, I don't really feel like giving it to you. I'm going to take the two cents or whatever, instead of giving you ten dollars. Uh, and here, you know, you just lost ten dollars, it's not that much, but like, ah, you did it just to save two cents. Like that, that's you just you just don't care about me at all. And so that is what elicits anger. And of course, this is not this is not my own theory, this is the work of uh of Aaron Cell. But what we were able to show in that paper is actually you can uh create a computational account of of how anger uh follows from people's from observations of of what the other person uh has done. What what does that mean?

I mean, I think people might have two questions, and this is both hard. Um we we we you know you and I have talked about this today. Like what does it mean to offer a computational um account because you straddle two communities we've talked a lot about functionalism or like a software level description Um, but then there's also like a technical methodological use of the term computationalism. Uh and then there's like computers, like a laptop, like can for people, because there's different audiences, can you explain what what does that mean, both to like a general audience, but also like evolutionary psychologists who might be used to like functionalism, functionalism. Yeah, what does that mean, computational?

Yeah, yeah. Um that's I think a very important point. And it's a bit of an unfortunate fact about the history of psychology that you have just two different meanings of computational. So the first meaning of computational is maybe the one that will be most familiar to people across the sciences. Um so it's the idea that we are going to use uh we're going to use mathematics or we're going to use computers in order to to help us do our science. And so within that meaning of computational, uh, you are engaging in in computational work every time you compute a linear regression, or you're when you compute a t-test or something like that, because you're using a computer, you're using mathematical principles in order to make inferences.

Right. It was sort of just standard, it's just sort of like rules of deduction and just sort of force like you know if you know what the inputs are just some deterministic process produces some outputs that's that's compute that that's that you know right and it's just certain ones work for whatever reason you don't have to do it on computer to you can also exactly you can you could compute your T test on on paper. Yeah exactly dinosaurs roam the earth yeah and to do that. I see the I don't remember anything from it. Yes exactly sorry yes Uh I wrote a horse to the schoolhouse and played with an advocate. Yeah, exactly. When there was it to CSB, actually we had to compute matrix multiplications by hand in the

case. Yeah, we it's probably the Greg Ashby. Okay, yeah. Uh for that one semester because he wasn't live. But yeah, I mean and we'll we were giving given no explanations for why it was. Exactly. I did the same thing, yes, exactly. Um anyway. But yeah, so that's the everyone understands computation in this sort of sort of way of using either math or for more complicated systems, uh computer code, but it's still the same principle where it's just deterministic rules, um, it's just it's just deduction in sort of a scripted way that is sort of known. Yeah. Yeah. And obviously this is a super powerful tool, uh and it's it it's helpful in in in quite a lot of ways. Um but the second meaning of computational that is more specific to I guess psychologists

and neuroscientists is uh being interested in the computations that are performed by an organism like typically the computations that are performed in someone's brain. So it's really making theories about information processing. And the interestingly, the two meanings can be dissociated. So I can have a theory of the computation that is that someone is performing, or that a theory of the computation that a given part of the brain is performing, uh, that can be expressed in in verbal terms. the like uh another version of that is that I can have a very simple computational theory. So if you think uh if we go back to the theory of of anger uh that Aaron Cell gave, the the equation is is quite simple.

It's just the I mean for those who don't know, it's an equation that determines how um someone decides whether to help based on this parameter that's called their wealth fair trade-off ratio. and then you you decide to help given that your welfare trade-off ratio is above uh so if the costs and the benefits of your action of the potential actions are related to the welfare trade-off ratio in a certain way. Um so this is a computational theory in both senses, right? Because it you can express it with an equation. Um, and this is you're also telling something about the the computation that is that is taking place in someone's brain. Right. So if Weber Trade F ratio is one, it's like, well, if you would get $10 and I would get um

nine, I would just give it to you because that's more. They're that's just the same. But then it's like a par it's just it very it changes the cost and benefits. So like for my son, I might have a a a positive one where like even if if if even if I get uh ten but he gets seven, I still might give it to him because I care about him more than I care about me. And so it's it's like a very positive of our trade-off ratio. But then if it's like a negative one and yeah or a very negative value, then um I might, you know, um be willing to incur a cost to make sure you don't get a benefit. So just to make it slightly more concrete. Yeah. Sorry, but that's just Yeah no no of course that's a yeah that's a great summary.

And so yeah and so uh uh I think and I mean a lot of evolutionary psychologists think that having those computational theories, you know, those theories of what is the computation is extremely valuable. Um and and this is valuable even when those computational theories are actually not expressed in uh in computational terms. I don't uh even if you don't have equations or computer code in order to explain the theory. Yeah. Of course, maybe the the ideal case is when you have both, right? When you have a computational theory that is expressed in equations or in computer code. But then there's another quadrant in the space of possibilities. There's a quadrant where you're using very complicated mathematics, but you're not saying

anything about information processing. Exactly. And and that is, I mean, it's often very useful, right? Sometimes you might be doing sociology and don't care at all about how the mind works. Uh it's a very um valuable thing to do. Yeah. But then the the sort of problematic part of this quadrant is that when you are a psychologist and you say, look, I have a computational theory of X, and your computational theory is indeed there's a lot of very complicated mathematics. You can implement it, you can run nice simulations. But when you look at it a little bit better, then you're not actually really saying anything about information processing in the mind. You're just modeling how the observed variables are

actually related to each other. Yeah, I think that's a really important profound point. Yeah, you and I were talked about this. We really, what does it mean to be computational is like a paper we should really do. I think a good concrete example of this would be just describing, like to make it super concrete for people, just a bicycle, like to the extent people know how bicycle works. Like you could have a verbal model of how bicycle works, like you know, you push the pedals, it turns the crank, it pulls the chain, it pulls the cassette, it turns the wheel. That's all verbal. I could have a very nice mathematical description of like how fast the tire rotates, and like sort of some other parameter, like I don't know how often the chain vibrates, that has

nothing to do with explaining how the thing works. And if you just like if I typed up the functional description and if I made a beautiful graph of the of the uh um fluff and you just did superficially like which one looks science or people would be like oh the beautiful graph you have numbers like there's a pragmatic effect that like once you know the function of something so well you you can avoid being qualitative and you can become quantitative. But I believe there's been an enormous often cynical, frankly, thing where it's like I if I make something quantitative, I can b lead people into believing that we know understand it so well that we're beyond the qualitative phase and now we're doing

quantitative science. It's the equivalent of like putting on a lab coat and being like I'm a scientist. Um and yeah, it's not a straw man to say like I have sat in on many, many, many, many, many, many, many, many, many talks that were this quadrant. Yeah. Yeah, yeah. And there's a the there's a version of that that ends up connecting back to the issue of correlation and causation, which is this kind of industry of mediation analysis and structural equation models. Uh you can have increasingly complicated mathematical models where um someone collects a lot of data um and then they put them inside the computer and then the computer crunches the numbers for hours. And then what comes back is a chart where you have uh you have little causal errors between

one variable and another. And you can have a quite a complicated graph. And again, it looks like you have produced a description, a mechanistic description of the system. Um but very often those things um or kind of increasingly sophisticated ways of convincing yourself that you have causal insights, whereas you're you're just describing a pattern of correlations. Yeah. And we shouldn't, you know, I mean, oftentimes the the computational aspect of it can be quite simple, as in the case of anger. Like it's actually when you think about the design of anger, it's shockingly simple. You know, you're trying to figure out what uh cost a person would be willing to inflict on you if i if in exchange for whatever benefit to them.

It's just a ratio, literally a ratio. You're trying to figure that out. And then your anger and gratitude should be calibrated in response to that. Like that's a pretty simple computational model. It doesn't require a huge page of equations or graphs. And so we should be open to the possibility that whatever system we're trying to explain in the mind should be as simple or as complicated as it needs to be for solving the particular adaptive problem. And we shouldn't have a prior that it necessarily needs to be very sophisticated. Or we should at least, you know, be open to the possibility that some parts of it might be actually very simple and some parts of it might be complicated depending on what

perspective you're taking or what part you're looking at. In terms of the the possible inputs to the welfare trade-off ratio calculation, there might be many different inputs and many different contextual cues that that go into that calculation. And that part might be very complicated. But what what what is spit out in the end is actually very simple uh it's just a ratio of two numbers. Yeah um I think uh I think you could even go further than that and and say that uh even if you start with the baseline assumption that's everything will be quite complicated, um there is often value in having a simplified model. Yeah where you sort of say well of course we know that the system doesn't really work

exactly like that but um it's actually the case that we have a very simple model that gives us a lot of traction. Um and I mean this is actually the if we look at the history of science, this is the way that that things have worked, right? I mean uh kepp Kepler's the Kepler's law of laws of motion, they're not they're not a complete description of kind of the way that the solar system works. We we didn't like it's not like the general theory of of gravitation uh and relativity sprung up fully formed. Um we had to start with very simple models and then we added a lot of complexity. And so yeah, probably in the sciences of the mind, there's no reason that things should work in the uh otherwise.

This was a long oh sorry, David, go ahead. I I feel like we often confuse which parts should be simple and which parts should be complicated. The deeper principles should be simple. What results from the operation of those simple principles can be mind bogglingly complicated. I mean, the best example of this is the theory of evolution itself. You know, it's extremely simple. You can lay it out in a couple sentences. Um, Darwin laid it out in a couple sentences. Uh, and yet that simple idea, that simple principle, gives rise to the most mind bogglingly complex forms of biological organization, including the brain that's producing these words, is a is a product of the of that of those simple principles.

So that to me is the best example of an extreme difference between the simplicity of the underlying causal process and the complexity of what root emerges from that causal process. And I think we should be trying to have the same kind of insights about the mind. We should be trying to hone in on what the simple principles are that are producing the complexity of mental and behavioral phenomena. And I think evolution helps you focus your lens on what those simple principles ought to be. What are the underlying goals of an evolved organism? Uh and and what what are the the parameters that we need to be looking for in order to guide behavior adaptively. Yeah, Tedek, I think I derailed I think you triggered me when you said computational.

I think this is like a long digression. You were saying you have a computational model like ten minutes ago. Do you remember what you were talking about? If not, I have we can go on to something else. I think I was just talking about the the computational model uh and the instantiation of the the model of how anger follows from a sort of a Bayesian inference about the um uh about the the person's welfare threat of ratio. Right. Because I guess in in previous work by Aaron Cell, there was this idea that yes, if you if the cost and benefit ratio is such that um this gives you information about the person's welfare total ratio, then you will calibrate your anger accordingly. And so what what we did in that paper is kind of to really flesh out uh like how this would

work in in Bayesian terms. Yeah. In Bayesian terms, just so it's not like jargon or fancy. Let me may let me see if I do a simple like roughly six year old level description that you tell me if it's accurate. So like one version of sort of like let's use the the cloud rain example, right? You could do something where you're sort of like, well, every time it rains, how often are there clouds? And then you could just sort of say, Well, that's how I'm going to figure out how likely it is it's going to rain. That's one version. Um, but a smarter version would you would say, Well, look, how how when it's raining, when are there clouds? But you should look at another category too, which is how often are there clouds and it's

not raining? And then you should combine those two. And roughly speaking, that's what being Bayesian is. It's looking at both things and not just the one. Is that a fair, super simple gloss? Yeah, yeah, yeah. If we want to go even more at a more foundational level, maybe the idea behind Bayesian thinking is that um there's a lot of uncertainty in the world, And so, I mean, with some exception, with some exceptions, we can never be certain of most things because of the uncertainty. And so the best that we can do is to assign some number to our belief that something is the case. So maybe I look out the window and it's I see clouds in the sky, and maybe maybe it's going to rain, but I'm not quite sure.

Maybe I will say that there's a 0.9 chance that it's going to rain. And so this idea is that we can quantify our belief in something by assigning numbers. And then the the good way to do this, if we are going to engage in sound reasoning, is for those numbers to be probabilities. Because when we have probabilities, then we can work with them and we can combine them in particular ways. Is that you should really uh work with probabilities in certain instead of just truth values. And then this and this kind of framework is called uh Bayesianism after Bayes theorem, because Bayes theorem tells you what is the right way to update your the number that you assign to a particular proposition based on the evidence that you have.

Yeah. It's uh we can do yeah, we could do a whole two hours on what is a probability. We should not. Um I wanted to I wanted to slightly move direct, change direction. Oh, David, do you have something? I don't want to of just like we've talked a lot about causation versus correlation, and we've sort of been like, well, look, the mind is actually pretty good at like picking out causation, But there's another sense in which causation doesn't matter at all. And I think this gets into your free will work, which is imagine you have a causal account. Like, I don't know, let's imagine you're up on a stage, it's made out of wood, and you slip on a banana and you hit your head really hard. And you're like, what caused me to hit my head?

There's a long stretch of causal sequences, including like the people who chopped down the tree that made the stage, the person that grew the banana, right? Like there, there's a long stretch of causality that all led up to you hitting your head. But we, we meaning like psychological design, focuses in on not all of that causation, but the part of the causation that has the highest effective sort of turnover, where it's sort of like, where's the biggest bang for my buck? Like, if I don't want to slip, should I like go to where they grow the trees and like convince them to not grow the trees? Should I like, you know, uh go to the banana grower and not do that? No, it's like, well, that banana shouldn't have been on the ground there.

And so, like a notion of causation is also nar it's not just all actual causation, it narrows down to like as David describes, sort of like points of intervention or things I could do something about, which is what you would expect an organism to do. And I that's coming from your work, correct? That like I think you're also making this point that sort of an interventionist view of causation, I think there's other people doing it too, is also very evolutionary in the sense of even if you grant causal views, it really is a a a shrinking down of all possible causal sequences to those things you can functionally do something about. Yeah. Uh and actually the first time I heard about this problem was uh a talk by by John

Toby at HBS. And I don't think he ever published anything of the sort, but I mean I think this is one of those things. It's a great example of of instinct blindness. This is a problem that we never think about. Basically, I mean it's um it's it's something that's has been it has taken uh quite a long time before people actually started to realize that well this is actually an important psychological problem. What is going on in people's brain uh when they do that? Because this is one of those problems that we solve so effortlessly and unconsciously that it just does not register that that we are doing it. That sounds weird. You have to like make effort to be like, well, it's the growers of the trees.

It's like weird to come up with that. It's I mean that's why when I sometimes talk about my work, people look at me funny because why is this kind of thing? Um another kind of a classical example of of this kind of situation, this kind of problem, is that suppose that you want to know what started a forest fire, uh what was the cause, and that you know that well there were a lot of dry leaves on on the ground and it was very hot on that day, there was a lot of wind, um, there was a lightning bolt that struck a tree, there was some oxygen in the air, and so on and so forth. Um it seems almost obvious that it was the lightning bolt that was the cause of the forest fire. Even though there are many things that also contributed.

So um the oxygen was in a sense also the the the cause of the fire because if we had magically removed the oxygen from the air, then the fire would not have happened. But it feels almost ludicrous to kind of ask the question why do we think it's the lightning bolt and not the oxygen? It's like, of course it's the it's not the oxygen. This is um an intuition that our brain is giving is giving to us in such a vivid way that it doesn't even feel that we had to do any computations in in the first place. And so and I think that that's why people didn't really think a lot about that problem, uh, and a lot of work that went into this. So, this is called the problem of causal selection. Sometimes people talk about the problem of causal responsibility and so on and so forth.

Um, so a lot of the work on this uh sprung up from issues in moral psychology. So people say, Well, I know sometimes we have an intuition that some people are more morally responsible for a bad outcome. Um, and when we try to analyze this concept of moral responsibility, then we find that well, causality has a lot to do with it, uh, which is itself uh its own interesting thing, right? Because uh, you know, you can imagine uh a thought experiment where I have an intent to do something bad, I want to I don't I want to harm David, And then David just gets harmed in exactly the way that I wanted, but actually, there was no connection, causal connection between my intent and his harm. People will not punish me as much because there was no causal connection between the bad

intent and the bad outcome. And so this is a bit of a tangent to say that in the process of analyzing how people uh think about moral responsibility, uh philosophers and social psychologists and experimental philosophers came to wonder well, why how is it that we attribute causal responsibility in the first place? Um, and so yeah, I think that is interesting because it's we had to recognize that there is an information processing to be solved here, but we did not show in the field did not really come to it in a very direct way. So, um, how how do we figure out when someone is morally responsible? I mean, certainly, you know, the the causal part of things is is relevant, but it do we figure it out in the same way that we figure out that the lightning bolt caused

the fire, or is it there's something different going on in the realm of moral responsibility? So uh maybe it it might be useful to talk about the the information processing logic of causal responsibility in the first place before we extend it to the the the realm of of moral responsibility. Um so so one one one sort of obvious idea for how we assign responsibility is that we engage in in counterfactual thinking. Is that we imagine a counterfactual possibility where we remove the lightning bolt and see and we see what happens to the forest fire. And then we say, oh without the lightning bolt, then there's no forest fire. And in intuitively it produces the correct result. But one problem with with that is that we are now kind of forced to judge that the

presence of oxygen in the air was also a cause of the forest fire, right? Because well, if there was no oxygen, then there was a forest fire. Um an example that people often use is so you can have this situation where um you ask me to water your plants while you're away on a trip. Okay, so I'm the one that's supposed to water you, and I kind of forget uh, and then your plants wither and die. So you can say, well, um my my my omission from watering the plants was the cause of your plants dying, which seems like a fine thing to say. And this is because if I actually had decided to water your plants, then they would have survived. So in that counterfactual uh the outcome has changed. But then you can also say, well, no, I think it's the Pope that caused your plants to die.

Because if the Pope had come to your flat and watered your plants, then they would not have died. Yeah, there's a lot of counterfactuals. If you hadn't plant if you didn't have plants, you wouldn't have killed them. I mean if there was no water, you wouldn't have killed the space of possible counterfactuals that are relevant is is just infinite, right? And so and so people need to have a way of just focusing on the counterfactuals that are actually relevant to the problem of causal attribution. Um and so a lot of what is going on is this that some counterfactual possibilities are just more relevant than others. And there is some work that has looked at what makes a counterfactual relevant for the

purpose of causal attribution. And so roughly um a counterfactual is relevant if it it was kind of high probability to to begin with. So the pope coming to your to your flat and watering your plants is just was so unlikely that this would happen that this is not a relevant possibility. And in the same way, the present like the absence of oxygen in the air is such a remote possibility that we're not going to consider it. Um, also, people don't really focus on counterfactual possibilities that deviate too much from what actually happened. So the idea is sorry, like just to make so like in an alien world where oxygen was variable but lightning was not, those aliens would think that oxygen, the oxygen caused the fire.

Yeah, yeah, yeah, absolutely. Um and then I mean where it gets maybe a little bit more spooky is that um kind of social and and moral norms also seem to have an influence on what we think what or the the relevant kind of kind of actual possibilities. So it might be that I don't know, um I I every day like uh Dave and I kind of come to water your plants. But you kind of repeatedly ask Dave not to come in. So it's kind of like he's not supposed to be here, but Dave comes in anyway. And I'm really the one that's supposed to come in and water your plants. Um so then if kind of suddenly none of us come, neither of us come and your plants die, it seems that I like I'm the person that is responsible for your plants dying because

I'm the one that was supposed to be here. That's like the counterfactual possibilities. The counterfactual possibility where I am I actually show a patcher apartment seems to be the most relevant counterfactual. And this is this is something for which I kind of I kind of wish there was a strong adaptationist answer for why this should be a functional requirement of causal attribution. But I don't I don't have super well formed ideas about why this is the case. Well, wouldn't it be that I mean what unpacking relevant it comes back to interventionism, right? So like if the point is what is what's the thing that David can do going forward, it's gonna be it's gonna be weird to say, well, Dave, I was wrong.

You should have been coming all along. Uh uh like he has to recalibrate himself and me, but just you, right? So, like, you know, he asked you to water the plants, but you didn't. He asked me to not be there, but I was there. And we didn't water the plants. The the most efficient, the the biggest bang for your buck causal interventionist thing is to recalibrate how much you are taking into account what he's saying about um the right about that. Like you you just fail to do something you heard him say. So that's like the I mean it's not exactly couldn't it couldn't it just be reframed in terms of the more likely counterfactual um being a more being more relevant because if you think about people responding to social

expectations or me doing what you request of me, like the probability that I will do what you request of me is very high, right? Um or or the the probability that I I'd forget is also pretty high, right? Um and so it may it may be that um responding to requests or social expectations is just is very relevant because um the uh the the degree to which we take them into account or not is just at baseline a very high likelihood thing to uh consider. Is that yeah, yeah. I think that that's that's one possible explanation. It's a little bit of a a deflationary explanation that all people are actually just tracking probability. Um I think I think that there is some evidence against this where people have done

experiments where they really emphasize well no like the the kind of frequency of those things is is really the same. Um, but then you could counter argue that you know to what extent the participants actually believe what you're telling them, and maybe they're still um using those social norms uh as a as a QT probability, that's for sure. Um but I I'm I'm I'm kind of inclined to believe that there is There is still an effect of kind of social expectations on the relevance of countervactuals. And one possible explanation to that is what Dave suggested about identifying good points of intervention. Maybe another possibility is that when we engage in causal attribution, we are trying to communicate things about our beliefs, about our values, about our

own model of the world. And so one way that we can communicate what we expect of others is to actually uh make causal attributions that will contain cues about what's what we expect them to do. Yeah. So that would predict, for example, like let's say there's um I don't know, we're holding up a I'm trying to think of a good example. This is a dumb example, but let's imagine like we're holding up a wall that might fall over. and or or we need to, right? Like we're we've noticed this wall we've all worked hard on it. We need to. And I I tell like 10 people, the wall's falling over, it's gonna like um cause our uh it's gonna like hit, it's gonna ru hurt us in some way, right? And then another 10 people I don't tell.

And all else equal, the wall falls over. And then you just say, like, well, who's responsible for the wall falling over? My intuition is it's the 10 people I told. And that's just a readout of like who's more likely to if I add a little bit of information or who's easier to convince um, you know, like the that's the that's more and more responsibility. So I think it comes back to the that's the interventionist account right there. That um that it's going to be easier to intervene if I've already told somebody and there's common knowledge that there's a cost and there's a problem versus I haven't even said anything yet to you. Um that's enough to tip it towards the first 10 and not the second 10.

Yeah and and it is definitely the case that the the counterfactual where those people show up is much more relevant than the counterfactual where it is the 10 people who have not heard the information that that show up. Yeah. I mean, I I personally, like this is my own bias. I don't like it relevance language because I find that obscuring. Because I think it's sort of an outcome of the interventionist calculation. Um, it doesn't like the relevance is sort of a placeholder for like this is what you think. Not that relevance is calculated and then that's what causes it to be uh the outcome. It's that it's relevant because it's the outcome calculated based on interventionist logic. That's m that's my own personal bias.

But um interestingly the sort of the the relevance of something is not there is not always a direct a direct mapping uh between the relevance of a counterfactual and whether it is selected as the cause. which was like I think one of the most interesting discoveries in this research on causal attribution. Um, and I can so I can first give a like an example of an experiment that shows this: that the the most relevant thing is not always the um the what people pick as a a causal explanation. Um and then I can maybe say a few words about my own preferred explanation for for this pattern. And so you can imagine that there's this person, call them Bob, they're trying to graduate. And they're almost past all their requirements for graduating.

They just need to pass two different exams in order to graduate. So they need to pass the history exam, which is famously very easy, and they need to pass the math exam, which is famously very hard. Everyone, almost everyone fails at the math exam, right? And so they really need to pass both of them in order to graduate. And suppose that Bob goes, he and he passes the math exam and the history exam. Um, and therefore he graduates. Uh, when you tell people that story, um, they will say, Well, I think Bob graduated because he passed the math exam. And so here we have a situation where um it seems that the most relevant counterfactual is kind of failing the math exam because it was very likely in the first place that he

would fail this very difficult exam. And also people think that yes, it was passing the math exam that was the cause. So here we have we have an alignment between the what counterfactual is relevant and causal attribution. But you can also modify the story in a very uh slight way, in such a way that actually you will get the reverse pattern. So now if you say, well, um, and this is Alice, and she also needs to graduate. Um, but Alice will graduate if she passes either the history exam or the math exam. And again, the math exam is the most difficult thing, but she only needs one of them. Um then again, Alice passes the math exam and the history exam and she graduates. Uh, when you ask participants uh why did Alice graduate, they will tell you, Well, it is

because she passed the history exam. I don't know, I don't know if you to what extent you share this intuition, but maybe it was you know that's the she only needed to pass one of these. Uh and since the history exam was the the easy one, that's kind of that's the one that actually explains why she graduated. Um and and in that case, it is still the case that if we think about what is the most relevant counterfactual, um, the one that is easiest to come to mind in a way, um, that is the counterfactual where she fails the mathematics exam. Because again, the math exam is super difficult, and so it's quite likely that she would have failed it. But despite the fact that the most relevant counterfactual is the one where she fails the math exam, people have this intuition that

it's really it was passing the history exam that caused her to graduate. And really all that we did is change one word in the story. I have intuitions about what might explain it. I I do. But m maybe Dave had a methodological comment. No, no, no. No, yeah. Um okay. Uh it might be that what we're simulating counterfactually is how likely um she would be to study for the history exam versus the math exam. Like whether she would she would really invest in trying to pass the history exam because that's gonna get her the most bang for her buck. And so um we uh so we're imagining that and that it would be really unlikely if she uh if she invested all of her effort studying for the math exam. So it's interventionless, sort of again.

Yeah, and it's again it goes back to this notion of what is the optimal point of intervention. Um yeah, I like this explanation. Um however, I don't think it can explain like all possible cases. Okay. So you can you can run the same version of this experiment where the the two events are completely out of your control. So you have no in order to you're playing some sort of casino like chance game. And in order to win the game you need to kind of draw a a ball from an urn. And you kind of you're going to draw at random. And so maybe you're going to win the game if you get a a green ball from either one of two boxes in that would be one condition. Um and in the other condition, you win if you get a successful draw from the earning

from uh from both both boxes. And so in that kind of setup, you still find the same effects. And then since the outcome of a given draw is a random one, you can't really uh give an explanation in terms of like how much someone's going to prepare. Okay. That makes sense. But isn't it still an intervention in the sense of like what's the outcome that has the biggest bang for the buck? Or like what do you need to what what goal do you need to do? If if you were like, well, how was I going to do if I can manipulate the system, what's the thing I need to do to get the outcome I want? Yes, that that's definitely one one explanation. And people have tried to formalize this insight in computational terms and it explains some

of the data but not all of it. But I mean the the most parsimonious explanation that I came up with for that is that um when people make a causal judgment they they imagine not just one um counterfactual possibility but they they simulate many of them. So they they kind of sample, simulate some number of counterfactuals. And each counterfactual is simulated with a probability that is that is proportional to kind of its what we would call its relevance. So it is simulated in proportion to how uh to its prior probability, simulated in proportion to how close it was to the the actual event and so on and so forth. And then once you have this collection of different counterfactual possibilities, you can

compute the correlation between the cause and the effect across those counterfactuals. Yeah, so if you're trying to think, oh is it the passing the history exam that caused Alice to win? Then you compute what is the correlation between passing the history exam and uh Alice graduating in the worlds that you have simulated. And so when you do that, you find that well, there is in almost all worlds, she she graduates, right? And she passes the history exam. And it kind of doesn't really matter whether she passes the math exam or not. And so the correlation between passing the history exam and graduating is quite high. And that's why you think that it was passing the history exam that was the cause of her

graduating. And now we can apply the same logic to the case of Bob. So just as a reminder, Bob needed to pass both the math exam and the history exam in order to graduate. Like he really had to have both. And so you kind of engage in the same process of simulating counterfactual possibilities. And now you see that well, there's going to be many worlds in which Bob passes the history exam, but he fails the math exam. And so he graduates. So it means that now you have a high correlation between passing the math exam and graduating. And this then explains why we think that, well, in the case of Bob, it was passing the math exam that caused him to graduate. Let me make sure. So in other words, it's the crucial difference of all these counterfactuals are run

through. It's sort of what's the crucial difference that leads to the outcome desired versus not desired. Yeah. Yes, yes. But this idea of, I think this idea of a crucial difference is a nice intuition for what is going on. And this kind of process of counterfactual simulation and computing correlation is an explicit information processing description of what is going on. Right. It's a crucial difference right across those counterfactuals. Yeah, yeah. Yeah. Interesting. Um presumably these counterfactual simulations are happening unconsciously. It it seems like it introspectively when when I was thinking about the the Bob problem or or the the math test problem, like I didn't have a vivid, you know, simulation in my head

where you know he he passed in this case, he did he didn't in this case, he graduated in this case. It's just kind of I just had an intuition. So presumably these simulations are happening unconsciously. Yeah, and I I think I assume that's yeah, it is the case for most people. I myself don't really have uh the phenomenology of running those simulations. Um and so from uh from a cognitive perspective, there are two different possible accounts of what is going on. There's is an account where we can we do run those simulations uh but it's just implicit and unconscious. Um there is maybe another account where we don't even need to to really engage in simulation. It instead we our brain has found a way, kind of kind of a shortcut that allows you

to compute what would be the result of those um of those simulations. Kind of maybe as an analogy for what this possibility might look like, you might think about you know the process, the much simpler process of computing probabilities. So I might want one way that you could compute the probability that a dice is going to come up for is to just simulate many rows, many rolls of the dice, and you count up the number of fours. But you don't necessarily have to do that. You can use the laws of probability theory to compute that. And so there might be a way that our brain has also evolved some shortcut way of computing what would be the outcome of such counterfactual simulations. Would an example of that just be that like there's a representation of high variability and

low variability, for example, like associated with like you know doing good in math versus this? And so even just that common sense knowledge about the variability um curtails some of the needing to run through the simulations. Yeah, I mean it might be that if I kind of give you like a novel scenario where you have an unfamiliar thing, then uh your brain needs to engage in simulation. Um but if I ask you about something that is much more common common knowledge, I mean in the sense that those are familiar things, then you can um you can use a sort of a pre-computed uh Yeah, or even just like if you have like imagine like a physics is a really good example. Like let's just say the physics, your folk the folk physics um system in your head, the

physics engine in your head, um just calculates based on certain experiences, how precarious something is. Um And then there's principles that then just can get extended automatically to new instances based on the extracted cues. So, like, you know, is a beam precarious, like when it's up straight up and down versus like laying on the ground. And there in that has been based on past runs of the of the physics and seeing these things fall. But now there's sort of that gets represented and attributed in new instances without running the simulation. That on its own, if you see a new instance where the the counterfactual depends on the likelihood of stability versus falling, that involves no running of the counterfactuals.

It's just that you already there's already a system there that plugs into one is more likely, one is less likely, and you don't have to run which the counterfactuals. It would be a concrete example of that. Yeah, yeah, I definitely think that something like that could happen. Uh I think yeah, I think it's it's an example, it's uh an illustration of the fact that the model itself uh is really couched at the kind of the David Mars computational level of of analysis where like okay, here is the computation that is that is that the system is trying to implement. Um when it comes to how the s the the mind is doing that, yeah, uh concretely it kind of all backs are off current. Yeah, yeah, exactly.

And the I think intuitions about what's easier and harder are really bad then too, because you know, it's yeah, it's what it feels like to me to do something is not how hard or easy it is for a computational system. Yeah. Yeah. I mean, cause what's interesting to me is that there are cases where this is conscious that I can think of pretty easily. And it seems like when I do have a phenomenology of causal reasoning, it tends to fit your model pretty well. So like the example I'm thinking of is I'm playing chess and I really want to get checkmate. I might actually consciously simulate all the different moves and counter moves. If I do this, you might do this in response. If you do if you do that, I'll do this in response.

Uh and I might simulate a bunch of different possibilities. You know, if I move this piece versus this piece, and I might look at the correlation it consciously in my head of moving this piece and checkmate further down the line. And if that correlation is high, like, oh, and all these different simulated possibilities, moving this piece helps me get to checkmate, then I'll move that piece. So it seems like uh there are some cases where we we are doing this uh in in in a full-fledged conscious way. And I'm curious if you if you have any thoughts about when we would need to do that consciously and when we wouldn't. Yeah, something that is related to to that point is that uh even though even if we assume

that the system is kind of mostly implicit, um it's not um I don't I I don't want to use the term encapsulated here in in in Dave's presence because this we I'll leave. You can definitely like uh like a verbally uh intervene on that system. Um so you know it's people have done this experiment where you know again you have the case of um Alice who needs to graduate. And this time I I don't tell you anything about the difficulty of the exams, I just tell you there's the history exam and the the the the math exam, right? And I tell you, okay, like so she she passes the math exam, she passes the history exam. And I tell you, well, just try to imagine you know what would have happened if she had failed the math exam.

Kind of try to really simulate the counterfactual. And then after you have done that, I ask you, well, what do you think was the cause of her passing? And then people will say, oh, this was the history exam. Right. And so you kind of sort of can artificially push people to to simulate one counterfactual more. And then this will have an effect on their causal attribution. So you can sort of even if the process is implicit, you can sort of um use explicit communication in order to influence what people are going to do. That's fascinating. And I think then the other part of your question is when are you going to engage in sort of more conscious simulation? I think that at least one part of the answer is that if you are in a novel

situation, maybe this is the first time that you've been playing this chess, this particular kind of chess game. I you I guess you you were never in that position before, then you need to engage in explicit simulation. Um but maybe if this is a situation that you're very familiar with, then all the computation has just been done and you can rely on schemas that have been pre computed on the basis of your past experience. Interesting. So maybe like the opening moves are just gonna be wrote, but like the end game, that's where my conscious reasoning might kick in. Yeah, yeah. And I I think I think that uh to a large extent, you know, the people who develop uh AI systems that play chess or go, they have this kind of hybrid architecture where they will

rely a lot on those kinds of uh cached pre-computed um sort of calculations of what is the value of a particular move. And then they will com combine this with more explicit simulations. Um and it's quite likely that the human brain is designed along more or less the same principles. Interesting. Um, let's talk about theory of mind. Uh I mean, talking about nebula is hard to understand, um, obtuse stuff. Um uh I I I have a very soft spot in my heart for theory of mind. Theory of mind like the ability to attribute mental states. Mental states are just like causation, as we talked about, like where in some sense they do exist, right? Like in the sense of it's a convention of like, of course, I have a desire and a belief.

And um, but there's another sense in which they're placeholders for common outcomes in the world. Um, so there it's very confusing and weird and abstract too. So um, so where does theory of mind come in? Um and is essentialism part of it, or is that a separate part? Is is uh because because in addition to causation, you've tackled theory of mind and essentialism. Um uh so that's like the trifecta of of of of of abstraction. Uh what what is the what is up with theory of mind and um and essentialism and are they related? Yeah. Well, I think that the nice thing is as we could do like a a two hour discussion on each. Exactly. Uh yeah. Get comfortable. Get comfortable. The bad news I maybe for the sake of kind of uh uh for the sake of tractability,

uh I'll focus on theory of mind uh for now. Sure. Um so theory of mind uh is yeah, I mean as you said it's quite interesting because there's that's a lot of abstraction that's we can't really see someone's desires or someone's beliefs um yet we have to attribute beliefs and desires and and so on and so forth to other people in order to even start understanding their behavior. If I were to just try to model you in physical terms Then it would be very difficult. I mean, I would predict maybe you're going to say to stay roughly at the same place over time, that there you will be some sort of spatial temporal continuity to your behavior, but that's roughly how far I'm going to go. Whereas when I attribute to you mental states, then um the richness of my explanation of

your behavior is going to be much uh much uh much more effective. And so this is something that has been of quite a lot of interest to psychologists because obviously we want to know what is the functional logic of this system that has been so important. Evolutionarily, this is something that social cognition is, especially in a species like ours, it is super important. If you cannot predict the behavior of other people and if you cannot exert some degree of influence on them, then you're not going to go very far. Right. And so a lot of a lot of research and theory of mind has looked at um our ability to attribute beliefs. And there are good reasons to to that. Because in order to really predict

um the someone's behavior you need to engage in what people call uh a meta representation. A r meta representation is just a fancy term to say a representation of a representation. Okay. Um and it's because if I represent the representations that are inside your mind, then I can I will know how you behave. Uh and uh a standard test for whether someone has meta representations is whether they are able to imagine that someone might uh represent the world in a different way than than than your own. Okay. So and then to give a some simple example, um, there is a mug on on the table on my on my office, and I assume you can see it. And so I can represent that um that maybe I'll okay I can represent that

Dave has a representation of the fact that there is a mug on the table. But to really have a handle on to really be able to prove that you have a representation, I need to show that you are able to represent the world as being different from it actually is. And so maybe you might have uh I might have a sophisticated hologram of a table, of a of a mug on the table instead of an actual mug. If that is the case, I know that there is in fact no mug on the table. But I know that you believe that there is a mug on the table. Um and so a lot of research and theory of mind has been looking at this ability to attribute false beliefs to to others. Um and in in some of my recent work, I've been trying to to see how

uh maybe uh how we can think about theory of mind in that involves uh representations that are not as uh as demanding as as that what makes that demanding uh so so in in the case of like representing that you know maybe you think that there is a a mug on the table whereas Yeah, like let's say you move the mug and I didn't know that. Like I just have a false. So like you told me you showed me the mugs on the table, it's out of frame. You are uh like a false belief, naughty character, you move the mug, I didn't know it. So now you know where the mug is, uh, but I don't, but I so I have a false belief. Yeah. Yeah. Um intuitively it's not necessarily demanding, but I mean if if you think about the the

logic of meta representation, uh it's if you really take it to the extreme that in then it means that when i represent your mind then i will construct a whole model of your own of your model of the mind right and that's obviously that's a lot to represent right so i need to represent like how i think the world is and then i need to have a another representation that is basically the the world as I think uh Dave sees it. And so if you take on to the extreme it it's really um this idea of meta representation really uh involves the commitments that I'm going to have one model of the world for each person that I'm trying to to uh to as you interact with so just to make sure i understand um the the you're arguing against a sort of an argument

which is the way that we represent other people's mental states is we create an additional model of an entire other human brain and what it currently is representing in addition to our own, and that we multiply that every time for new states of the world and for new agents. And that uh that sort of naive model is unlikely to be true. Yes. Yeah, that's uh that's a good summary. And I don't think that's that anyone is really claiming that we do that. But if you if you kind of look at the disclaim that theory of mind involves meta-representation and that you run all the way with it, then it seems to be the logical consequence. And so what is your what's your claim that that is instead of that claim, uh what is your

argument? What's going on? Yeah. Right. And so an interesting question is indeed now no what are ways that the system could actually work and not completely run out of memory uh every time you're trying to predict someone's behavior. Um and here the the idea that we have been putting forward is that you don't always need meta representation. And in fact in in most cases you you don't actually need it. So in the case where I'm not playing any tricks on you and there is actually a mug on my table and you see the mug and now you know that there is a mug on the table, then I don't need to attribute to you the necessarily the a meta representation that there is a mug on the table, that a kind of a belief that there is a mug on a table.

Instead I can just note that well you know that there is a mug on the table. And so what this does um is that I have a physical description and a physical I have my own physical model of the world. So in my physical model of the world, there is a mug on the table. And then in addition to that, I just add a very simple binary representation that oh Dave knows this fact. So now instead of having to duplicate my own model of the world in sort of the meta representation that I attribute to you. I can just uh have this one binary, one like on or off tag about whether you know this. And that the def what that on off tag means is uh all the inferences and consequences that I have for the mug, the other agent has as well.

I'm not saying that that's uh the full model because of course it's not right, because like I might see it from a different angle and you would need that additional. But but just to be clear, like what it means is that there's no um there's no difference between what you all the things that follow about the mug for you are applied to me, but they don't have to get applied to me. It's just that there's representation of everything that is represented in me about the mug also applies to this other agent. Yes, yes, exactly. So it's like one piece of information instead of duplicating a brain. Yeah. Yeah. Yeah. In the same way that, you know, we scan the the the physical environment and we tag physical objects with various properties.

We might say this, you know, coffee cup tag belongs to Dave, uh, contains hot liquid, um is, you know, this sh shade of blue or whatever. Um uh you know uh whatever. And in addition to that, another tag is Dave knows it. And we just oh or didn't Dave knows it's there. And we just have that tag to, you know, everything we know, um, Dave also knows. Yeah. Right. Yeah, yeah. That that's yeah, it's such a great description of of the proposal. Yeah. Just one more feature of the of the object. Um yeah so I mean this is this was just an idea um this this is not an idea that actually came from from us this existed in in the literature um but this had been proposed as just like one possibility for how theory of mind might work.

And what we we did actually in in our project was to try to take a more adaptationist stance towards this proposal. And we said, well, can we actually formalize what is the adaptive problem here? And we started with this fact that, well, we obviously we need to predict and explain the behavior of others, but we also have a constraint on on our information processing capacity. Like we don't have infinite computation, we don't have infinite memory and so on and so forth and so forth. And so the design of our theory of mind system needs to somehow trade off between those two different considerations. We want to have a lot of predictive power while at the same time like not run out of computation every time we interact with someone.

And this is this idea that computation is a constraint is really salient in the context of theory of mind where it's very easy to run out of memory when you're predicting the behavior of an entity that is as complex or even more complex as than than you are. And so once you you set up this optimization problem where you try to trade off between those two different things, The nice thing is that you can you can solve for it. You can ask, well, what would be the optimal cognitive strategy for solving that problem? And we did that in a in a by setting up a toy model. So we had sort of a simple agent-based model, and in that agent-based model, the the agent, there's one agent that is basically doing a series of very

simple theory of mind tasks. So there's a like a false belief task, there's a task when you need to predict what someone who has knowledge of a fact is going to do and so on and so forth. And then and this you can sort of uh then optimize the cognitive strategy that this agent is going to use. And when we did that, we find that indeed if your agent is sufficiently constrained in terms of its information processing capacities, then it will use a strategy that is very similar to the one that I just described, where it will kind of abandon any pretension of creating meta representations. Instead, it will just have its own physical description of the world, and then it will add a kind of a binary tag about whether or like this fact is known or not known by the other

person. Is one way to think about it then that um so like I don't I'm not a um uh I don't know, I'm not like left-handed. So like let's say there's a guitar a right-handed guitar and a left-handed guitar, and I reach for the right-handed guitar and then I I somebody else, I see somebody else making that same decision, but they're left-handed, they reach for the left-handed guitar. My model of predicting what they would have done is wrong when I just use what I would do. Right. So then I have to say, I have to have an entry for what's the minimal extra information I need to add to still use me as the machine for predicting behavior, but then change one variable to generate the right prediction.

Is one idea that, like, all these kinds of things of like what a false belief actually is, or what a motivation is, or a personality variable, they're all just sort of settings on what's the minimal thing I need to add to my own inference engine about what I would do next to then just predict the behavior of others. That in that sort of all these, all these representations that we think of are just sort of uh just slight tweaks to this otherwise very complicated system that we don't have to repeat and duplicate. Um so it's just that for for for Tedig and for David, I just have my own pre, it's like an equalizer, you know, and it's sort of like I have the presets for David such that when I think of David or I see David, I can pull those up and I know my own

self-simulation system. I need to add these 20 or 10,000 tweaks to all those little inference engines, but that's only little 10 10,000 bits, right? It's just small tweaks to the the big machine. Is that a fair? I mean, I realize it's sort of abstraction, but is that a fair way to characterize the picture here? I think it's it's uh what you are describing is is a more general extension of the general principle that we should like to what the extent that we can, the the system should be minimizing the computational complexity uh by using all tricks that it has available. But the the the coincide behind our model is that um even before we are trying to predict the behavior of others, We have our own model of the physical world

because it is, I mean, it's super useful. It's very useful to know what the world looks like. And the nice thing about other people is that they also live in the same world. And so by virtue of the fact that you're living in the same world, you will end up creating representations that are actually quite similar. And so this is a kind of a nice fact about the problem of social prediction is that you're predicting the behavior of agents that also live in the same the s the same world. And so it means that to a large extent you can use your own understanding of w of what the world looks like in order to predict their their behavior. It also means you can tell when an error occurs, right? Because that's one of the other problems, is like a lot of the machine learning, what the

pro the trick is to know when an error occurs, and that's a discrepancy between what they do, which you have data for, and what you would do, which you also have data for. Yes, yes. And so really like really from this point of view, yes, the the main challenge is to uh um to compute what what are those other those aspects of the world as I represent it that other people also are aware of. And this is a so I mean an interesting thing about that is that um this this model, even though it is relatively simple, it makes quite interesting predictions. Because you can look at the for a particular setting of the the parameters, you can look at the optimal cognitive strategy that an agent comes up with.

And you can see well to what extent does this actually describe the behavior of actual organisms. And here a particularly interesting case is when we look at non-human primates. Okay, because non-human primates um they are much more uh constrained in their information processing capacities as we are. And maybe they don't have to use theory of mind to such an extent that we do because um the the sort of social demands of their ecology are are not as as strong. And so you can you can sort of set up an agent that has not a lot of information processing capacity for the purpose of theory of mind. And then you can look at the strategy that this simulated agent is using. And one prediction that follows from that is that they will behave in a certain way

in tasks that are a little bit like a false belief task, except that you have changed one thing. So this task is called the accidental true belief task. So this is a task where I so suppose I I hide an apple in front of Dave. So like you're aware that I'm hiding an apple in this in this one box. And then you go away. And while you're away, I will remove the apple from that box. But then I put it back in the same box. Okay. So what it means is that you you have a belief that the apple is in that box. But the there's a sense in which a belief is kind of lucky. Like because it was just uh a chance event that I put it back in the same box. Or um and to make it, let's say you know you you you remove it from the box and put it on

the windowsill and then a bird magically flies in and picks it up and puts it back in the box. Yes. Another tree that's inner. Yeah. So we just really got lucky. A forest fire that was caused by a lightning sunway. Yeah. Um and so in that situation something surprising happened, which was that kind of the simulated agent um sort of stops making any prediction. It was like, oh, like the situation is too complicated to me for me. I'm just going to throw my hands up in the air. and basically uh i'm going to predict that you you will look that i that dave will look in any any of the boxes i'm not going to make any bet about where dave is going to look um and it turns out that this is exactly how monkeys behave if you give them

uh that that kind of task And so so it so it seems in that there's a lot of similar experiments that are done on non-human primates, uh, where they do things that look a little bit weird, um, but that are quite well predicted by this kind of version of the model where you assume that the agent does not have a lot of information processing capacity. Interesting. Well, fascinating stuff. I feel like uh I could keep picking your brain on this for hours, but we are um approaching the end of our time. And it and uh is there any topic that you wish we would have talked about that we that we should have asked you about? Anything that uh Oh yeah, um I think yeah, I think the another thing on which we could

have spent a lot of time is this concept of uh intentional action that is also related to theory of mind that it's its own uh its own thing. Well here's here's I would I'd like to ask how did you uh discover we haven't done origin stories in a while. Um how did you come across of psychological Yeah um so I've always been quite interested in the general question of how the mind works. Um but I I didn't really know it back then when I was in high school. Um and so I decided to because I was also interested in in history. Um and so I decided to to do political science as a major. Uh and that's when in the process of majoring in in political science, I I read this book by David Hume. And back to Edinburgh.

Full circle. Yeah. And so I was kind of blown away by the fact that oh this guy was asking just deep questions about about the mind and it made me realize well uh this is actually something that I find super interesting and so I ended up double majoring in in philosophy um and this was this is what led me to do uh then a master's in cognitive science in in Lyon. Uh and and the the master's was was interesting there was a lot of interesting stuff about you know behavioral neuroscience, cognitive psychology uh and so on and so forth. Um but while taking those courses, I always had a nagging feeling that there was a set of important questions that were not asked. Um and then of course like now I know that I was um internally

complaining about the fact that there were there was not much of an evolutionary perspective being being uh offered. Um and then it came time to find uh an internship, like a research internship for my master. and there was this new researcher uh at the university called Pascal Boyer. Uh so I reached out to him and like can I attend your lab meeting? Um and then during the lab meeting he's sort of like laid out the perspective for what his research over the next few few years what was going to be and I had this kind of the this click of oh this this is what I want to do. And and and uh a year later I was at UCSV working with John Anita. That's fantastic. So wait, Pascal was where? So he he was in Lyon at the time.

He had gotten a grant to stay there for a few years. Fantastic. Yeah. We have to have Pascal on. Oh, I can't believe you haven't had it. I know, right? Yes. Yeah. Well, Latik, thank you so much for being on. Um my brain is full. I've I've exhausted the computational capacity in my head. Yeah, this is great. This is our pleasure, David. It was a really fun conversation. David, what causes a good podcast? Uh well, hold on. Let me run through all the counterfactual worlds in my head. Uh computing. Copper wires. Copper wires. Yes. Oxygen in the atmosphere is definitely a a critical component to a good podcast. If there was no oxygen, then uh we wouldn't be alive. So that so um yeah uh so leave us a kind review and

talk about um all the things all the things that could have gone wrong such that you wouldn't have discovered this podcast. Yeah that's a great content for your kind review. Um, you can also visit us on Patreon and that might have a causal impact on the continuation of this podcast. Maybe we can afford some. This is very uh it feels very Well, we're just giving you ways you can intervene on the world. Yeah, exactly. you have you have free will to do any of these things. Yeah, but it there's we unintended consequences to things too. So you know, be careful what you do. That's true. That's true. Um uh I'll see you next week, David. Yep, sounds good. All right, well see everybody. Okay. See ya.