Can we model cognition? And what is a model anyway? with Paul Kelly

Mechanistic and Dynamical Approaches with Paul Kelly and Andrea Hiott.
This episode delves into the philosophical and practical distinctions between mechanistic and dynamical models, highlighting their significance in cognitive science and climate science. Paul and Andrea also unpack the 3M principle, surrogative reasoning, and equilibrium models, emphasizing the role of models in scientific inquiry and their broader implications for understanding complex systems. Perfect for enthusiasts of cognitive science, philosophy, and anyone fascinated by the dynamic intersection of these fields.

Find more about Paul and his work at: https://pauljkelly.weebly.com/

Love & Philosophy Beyond Dichotomy Substack: https://lovephilosophy.substack.com/

#model #cognition #love #philosophy

For those who want to delve deeper: https://communityphilosophy.substack....

Note: These conversations take place approximate 5 to 8 weeks before they post

Join the Substack: https://lovephilosophy.substack.com/
Go deeper with Community Philosophy: https://communityphilosophy.substack.com

Join this channel to get access to perks, or just give a dollar to support:
   / @waymaking23  

#love #philosophy #mind #andreahiott

Transcript

How to Model Cognition with Paul Kelly: Beyond mechanistic and dynamical?

[00:00:00] Hello, everyone. This is a conversation, a P H D conversation. With Paul Kelly. Let me just say from the start that. I'm not a pan cyclist or an, enactivist or any other IST. Or ism. Even though those words come up here and it might sound like I'm one or the other. I'm not, although maybe that means I'm both or everything. 

I mean, to get beyond dichotomy. Is to talk about what we are in different ways. So. Maybe I'm both. Maybe I'm neither, maybe both of those at the same time in some interstitial place, in any case. The point is I find all of them helpful in different ways. Depending on what I'm looking at or trying to understand about the world. And Paula and I get into this in terms of modeling cognition and what that means. And he's gone really deep in his scholarship and research and understanding of these topics. 

He's really clear about, about them. He really almost sees the world differently. Um, at [00:01:00] least when people talk about the word model, And he tries to explain that and. Does explain that very well here. Uh, these two different groups, again, depending where you come from, but. Roughly, you can sort of divide it into a mechanistic and a dynamical approach to modeling cognition. So we talk about what these are. And I'm not an insider. 

This is a little bit of insider baseball, so to speak insider modeling, but, uh, I'm not an insider to modeling, at least not this kind. So you don't have to be an insider to learn about modeling cognition. You don't even have to follow all this. Don't have to follow baseball, so to speak. Uh, but we talk about that. 

What is, what is a model? What does it mean? Why do we need them? Why does it get confused? For mapping territory. You know, if you listen to this show, I'm a little bit obsessed with trying to figure out. How we confuse our maps and our territory and how both are important and how to distinguish [00:02:00] between them. So we touched on that issue here in terms of modeling cognition and what it means to be mechanistic or dynamic. 

Why do we think we have to choose between those? Is there another way to look at this? Uh, that's part of the research I'm doing too. I learned about surrogate reasoning, quite an interesting term. Which Paul talks about. Quite a bit here. And even more interesting. Once we've gone through these terms, we get to this example of climate science and how. In that avenue, we might already be finding a different way to think about modeling. 

That's not either or mechanistic or dynamic, dynamic. Uh, Paul brings this up. It's really illuminating. Uh, very thankful to him for that. And for that example, And for this conversation, it's a long one. I'm just going to leave it as it is. And I hope you're having a good day out there wherever you are. And thank you so much for being here with us for a little while. All right. 

Be well, bye. 

Andrea Hiott: Hey, Paul. It's good to see you. [00:03:00] Thanks for coming on to talk today. 

Paul Kelly: Yeah, it's good to see you too. 

Andrea Hiott: Where are you? You're in Wisconsin. 

Paul Kelly: I am in Madison, Wisconsin, 

Andrea Hiott: um, so we're going to talk about a lot of big subjects today, and, um, I just want to say from the beginning that I am definitely not caught up on all the papers and references, but that's the point of having these conversations is to just open up new doors.

And also, I'm going to be asking you a lot of big questions, which doesn't mean that you necessarily have to have all the answers to them because there's a huge, huge amount of literature behind this idea of, um, modeling cognition, I guess. Is that a fair way? Do you think of yourself as studying that or thinking about that?

How to model cognition? 

Paul Kelly: Yeah, I think that that's a good question to start and I imagine the pressing questions that leap to mind when you phrase the question that way are What does it mean to model something and what [00:04:00] is cognition? And these are two very big questions and very controversial questions, but the idea that we could model cognition in a particularly fruitful and perhaps even explanatory way has become a really hot topic in the last couple of decades.

Andrea Hiott: Yeah, definitely. Um, and yeah, it's sort of, it seems like we're assuming we know what cognition is if we think we could model it in a way, but then in another way. The whole point of trying to model it is to find some understanding of it. So do you think about that much, or do you just assume, you know, what cognition is?

Or do you feel like you're looking for what cognition is, or do you just not think about it too much? 

Paul Kelly: Well, I'm, I'm interested in what cognition is, and nobody seems to have a clear sense of what cognition is, from what I can tell. Nobody also has a clear sense of what it means to model something. So there's a sense in which these two concepts using them together the ways in which the concepts can move in terms of our uncertainty is [00:05:00] in both directions.

Um, yeah, what cognition is, whether or not there could be even a strict notion of cognition that's defined in a particular way. The way that philosophers often think about providing a definition of something is by advancing conditions that are both necessary and sufficient for the thing to be present.

And there's lots of examples to try to illustrate this. One of the most common historical ones is the term bachelor, right? If you try to analyze the concept bachelor and provide conditions that are both necessary, that are required for something to be a bachelor. And sufficient in the sense that if those conditions were present, that would be enough to guarantee that the person was a bachelor.

You get something like something is a bachelor if and only if it's an unmarried man, right? 

Speaker 3: Right. 

Paul Kelly: Those conditions are individually necessary and jointly sufficient for being a bachelor. Whether or not we could do something like that to sort of capture [00:06:00] the essence of what it means for something to be cognitive in terms of a state or a process is much less clear.

And people have tried to propose definitions over the years and. It's contested and maybe there's a sense in which that's a problem because there's this fundamental notion of cognition, which we're deploying that we don't have a clear definition of, 

Speaker 3: or 

Paul Kelly: maybe practically speaking, we don't need a strict definition.

In the absence of necessary and sufficient conditions, we can still do good inquiry and investigate something about the mind, even if this concept itself isn't well defined in a strict sense. 

Andrea Hiott: Yeah, exactly. There can be a way to, um, think of it in terms of regularities or things in common. Um, but even with that bachelor example that you started with, I feel like we already assume a certain, um, set of references when we start to even ask if something is necessary and sufficient, but we won't get into all that today because I, I want to talk about modeling [00:07:00] cognition, but, um, before we really dig into what a model is, how did you start get into thinking about all of this?

Like what brought you to philosophy and to thinking about models of cognition? 

Paul Kelly: Yeah, I mean, I, I got into philosophy by taking which I imagine is the case for a lot of people in undergraduate class in philosophy. And prior to that, um, I'm not even sure I really knew of it as, as a discipline. And it was surprising to me.

That there was this field where people were asking these very abstract, general questions. And that there were these people, philosophy professors, who somehow made a living doing this. Just sitting around and scratching their head and talking to each other about these big, general ideas. Um, and it really intrigued me.

Um, and the more I got interested in philosophy and the more classes I took, the more I started to become interested in questions about science, um, and there's a subdiscipline in philosophy, the philosophy of science, which I became attracted to. [00:08:00] Um, in the philosophy of science, there's lots of different questions that people ask.

What science is how science proceeds, if there's a scientific method. Um, but also the achievements of science, the successes of science how science can help us understand the world or can give us truth in some meaningful sense. And a lot of philosophy of science over the last couple centuries has been focused on trying to make sense of the notion of a scientific theory.

Scientific theories seem really important. They seem to be something that we need. develop and construct and use to try to navigate the world and maybe understand or explain things in the world. Um, and philosophers of science have spent a lot of time thinking about theories. But in the last couple decades, people have realized that in actual scientific practice, Sometimes theories are referred to, but there's another term, model, which practicing scientists use constantly and oftentimes without much reflection about what they're [00:09:00] actually talking about.

Um, the way that the term model is used by someone who's an economist or someone who is a cognitive scientist or someone who is modeling you know, the spread of COVID and the spread of disease, They're all using this term model but there's no general consensus among the scientists anyway about what this means or how it supposedly represents the kinds of things in the world that scientists take them to be representations of.

So, all of these questions that philosophers of science have been spending centuries talking about in terms of theory, about how theories are confirmed and how they represent and how they help us understand, can now be reframed in language of models. And so we have additional philosophical work to do to try to figure out what these things are and how they enable us to have a sort of scientific understanding or picture of the world.

Andrea Hiott: Right. So we want to understand what's going on in the world. I mean, just in a really general [00:10:00] way, we're trying to figure things out. We start to become aware of ourselves and that something is going on and we start to try to understand it. And people have come up with all sorts of theories since there were, I mean, we can go back, you know, And back and back about what's really happening.

What is life? What is mind? What is all of this kind of stuff? And there's a lot of theories about it, as you said, but when did we start making models in the sense of what I think you think of as like, not, not mental representations, but external externalizing? And, and trying to understand cognition through models that were not mental models, not something that, I mean, something that we can share externally as an object or as a diagram or as something like that.

I mean, did that start, do you, do you have any idea about the history of that? Have you studied that much? Thought about, you know, When we started to do that? Probably. I I know drawing stuff in the sand. Yeah. [00:11:00] 

Paul Kelly: Yeah. I mean, I think it's a great question. Um, and it really in many ways cuts to the core of what we were talking about just a couple minutes ago, which is that we're not even really sure what models are.

Depending on how we answer that question will obviously impact our assessment of the historical question of when we started this practice of modeling. Um, I guess that's what I'm trying 

Andrea Hiott: to get into a little bit too as you answer this question is trying to understand because you know, we're coming at it, you've read all this papers about what modeling is, people discuss it and discuss it and discuss it, but if we're just going to start in a really general place, you know, where did that start and then why did it become controversial about what a model is.

You know, is it just language that we're just fighting over definitions or I mean, how do you see that kind of process as you're answering that other question about where did it even get even started? 

Paul Kelly: Yeah, I, I don't think it's just arguing about language and how we use these terms. I think there is fundamentally a difference in the kind of, to use a word [00:12:00] you just used, um, the notion of mental representation.

Which has been actively discussed in philosophy and the sciences, psychology and so on for a very long time. And this idea of a model based representation. 

Andrea Hiott: You see those as different things. 

Paul Kelly: Yeah, I see them as different. And I also see them as. Both different in terms of what they're referring to, but also they might require very different philosophical analyses and proposals to try to make sense of what's going on.

So the notion of mental representation one obvious kind of mental representation, is the notion of a belief. Right? Um, and there's been a lot of philosophical work trying to make sense of what it means to believe something whether or not particular sort of physical arrangements of things in the natural world, maybe a brain state or a brain process, could be about something.

The common term that's been used to talk about this notion of mental representation, especially in terms of belief and so on, is the idea that such states have intentionality. [00:13:00] Intentionality is a fancy philosopher's way of talking about something being about something else. Uh, sometimes There's so much 

Andrea Hiott: controversy about even that word too, and what it really means.

Are we going to go with Bretano Are we going to go I mean,

But and, and it's right what you're saying, I guess what I'm trying to get at before we get into all that is mental representations already require something like language, something like this, what I was talking about, this external space where we've communicated about something to even start thinking about what a mental representation is.

It doesn't mean that we didn't have what those words are referring to before, but there's some way in which all of this language and thought and philosophy, trying to understand all of this stuff, um, I'm trying to get at sort of, do you think that there was a way in which Okay, it was something like a model was required before we could even notice that we have something like a mental representation.

Paul Kelly: Um, I don't think so. So, I, I think historically speaking, we can think about it this way. [00:14:00] So there are various sort of simplistic, if, if the scientific picture of the world that we have, where there's this evolutionary process and that life develops and becomes increasingly complex and the kinds of mental abilities or dare I say representational abilities of these various organisms get more and more complicated.

There was a period of time where physical states, if mental representations are in fact you know, physical arrangements of these natural properties, they emerged in the world, right? There was a period of time where there were no mental representations, and then there was a period of time where mental representations came into the world.

Andrea Hiott: I don't know if I agree with that. Okay. 

Paul Kelly: Yeah, depending on your particular view about mentality, right? You might reject that if you're like a panpsychist or something. But it seems to me that that's a plausible view. And then later on, we get this complicated epistemic practice of modeling. And perhaps it depends on having certain kinds of mental representations available to the organism.

But this [00:15:00] practice of using a phrase that you used previously, externalizing our reasoning, right? And pointing at a particular thing which we take to be a representation of some other thing, that's in many ways the hallmark of what modeling is. And so over the last couple decades, my sense is that the convergence of people writing about this topic, has sort of developed on the idea that what defines modeling as a kind of representation, in contrast to, say, mental representations, is that it engages in what's come to be known as surrogative reasoning.

And surrogative reasoning is where you take something to be a surrogate or a stand in for some other thing. And then you reason about the surrogate. Maybe you manipulate the surrogate, maybe you see how the surrogate behaves and so on in various interventions. And then, based on your observations of the surrogate, you make inferences about the target, the actual phenomena that you're trying to explain.

And so that kind of process requires [00:16:00] a much more elaborate kind of mental set of faculties, right, where you can imagine that this thing, the surrogate, is like this other thing in certain respects. The surrogate itself could be a math equation, it could be a diagram, it could even be a physical object.

Imagine a model train set, right, which is taken to be a representation of, say, a larger train. Um, the crucial thing is that mental representation which we might describe in terms of intentionality and try to naturalize in the way that various philosophers have tried to do. in many ways refers to states that just sort of emerge out of the physical interaction of, I don't know, our brains and the environment that we're in.

But then there's this complicated epistemic practice of modeling and modelers, which comes significantly later, it would seem. 

Andrea Hiott: So I guess what do you think when you're using the term mental and mental representation, of course, we know there's this large debate about whether or not we can think of representations as being in the brain or if they're in the body.

So I [00:17:00] guess we should clarify a little bit. Like when you're thinking about. Mental and representations. Are you really talking about states of the brain or something like that? Or are you talking about, um, only in a more philosophical sense of the way those terms are used to think about propositional states and beliefs and so on?

Are you like really sticking to the literature or are you thinking there's something in the body and in the brain that's also a representation and that being like the distinction between this model, for example? Yeah. 

Paul Kelly: Yeah, I mean, it's, it's a great question. Um, it, in many ways it connects up with the earlier question about what cognition it is.

There's extensive debate. And what a model is. 

Andrea Hiott: Cause I think of Carl Friston, who says we don't have models. We are models, you know, so it can be used in so many weird, different ways that I wonder how you see it. Yeah. 

Paul Kelly: Yeah. Um, I, I think that. In addition to trying to define what cognition is there's another closely nearby related question, which would probably be informed depending on how we define it, which are the [00:18:00] boundaries of cognition, where it is, right?

If we define cognition in a very particular way, maybe we think that it's in the brain. Or maybe if we define cognition in a slightly different way, it seems to apply to processes and states that extend beyond. the brain into aspects of the body or even more radically into aspects external to our physical body.

And you have the extended notion of cognition, right? Um, which particular view do I adopt? I'm not sure. In many ways, I think it's an empirical question. It's a question on which of these kinds of theoretical frameworks are going to be the most successful and fruitful in trying to investigate the plurality of different kinds of phenomena and questions that we are interested in as, as reasoners.

Fundamentally, what is a belief or a state like that? I think the philosophy does say something significant here. It seems to be that a belief [00:19:00] Amongst other things, perhaps, is at least a particular attitude towards the truth value of a proposition. Right? Where a proposition is a thing that we utter or that is capable of being true or false.

So there are lots of things that we utter, like questions or commands and so on, that don't have truth values. If I ask you, what time is it? That particular utterance isn't true, it isn't false. It's neither. It doesn't have that property. But some things that we say can have truth values the sun is larger than the moon, right?

That is in fact true. That's a proposition. And I could have a belief about it. And I think that philosophy can give us some sense of what beliefs are, or maybe other states like desires and hopes and fears and so on. And then there's this joint enterprise with the actual empirical inquiry of trying to figure out how that's possible.

And there can be a give and take on both ends. Maybe the philosophers have a sort of overly simplistic and sort of [00:20:00] a empirical conception of what these various concepts are that could be informed by the science. But it's also true that the scientists would probably benefit from being familiar with the way these terms have been developed and analyzed in the philosophical tradition as well.

And so trying to have this be a joint enterprise of both philosophers and scientists trying to hone in on the best. An optimal understanding of these different concepts is, I think, clearly the way to go. 

Andrea Hiott: Yeah. And even just for everyday life and people, it can help to learn how to think critically, just in a way it, I don't know if you would say it's modeling it, but to start to model your own thought process in a way you don't have to be doing it in a necessarily analytic philosophy or, or logic way, but it can help, right.

To learn those ways to think about belief and so forth too, even if you're not. Trying to be a philosopher or a scientist, there's a way in which that critical thinking about it, drawing attention to it, um, [00:21:00] can be helpful. But let's get to how you got into this idea of modeling because you're really looking at models of cognition, right?

And different forms, like I'll let you, I mean, I think you're focusing more on dynamical models of cognition. Is that right? 

Paul Kelly: Yeah, I'm interested in dynamical models of cognition, but also just the bigger question of models trying to explain any kind of phenomenon. So even the idea that models could explain non cognitive phenomena like just physical phenomena.

There are lots of different kinds of models that are used in psychology and in other types of cognitive sciences and neuroscience and so on, but there's also other fields out there. There's, there's physics, there's economics and so on, and they all use this term model. Is there a unified analysis that we can give?

Of what models are, and is there a related unified analysis of how these things predict, describe, or explain [00:22:00] the things that these scientists are taking them to do? What is it about that, 

Andrea Hiott: what is it about that that's so interesting to you? What do you think? What's, what's behind that? What, what are we trying to do with all these models?

Because you're right. It's, I mean, I don't know, I haven't done a keyword search, but that is a word that everyone uses in all disciplines. And it's very important to the way we come to believe in things or think of things as facts and so on. So in your own interest in philosophy of science and thinking about this, do you have, is there something that really motivates you about that?

Or is it just curiosity? 

Paul Kelly: I think it's two main things. Um, the first is that As a philosopher, alarm bells go off in my head when someone starts appealing to a concept that it's not clear that they have a precise notion of what it is. Um, sometimes that's fine. As we mentioned earlier, it could very well be that we have various concepts we appeal to that can do good work and that can be genuinely scientific, but don't have clear boundaries.

So the notion of [00:23:00] species, for example, in biology is a common one. The boundaries between species isn't well defined, but we still think that scientists aren't misspeaking when they talk about species existing and reason with the concept. However, I sat in on a psychology lab for the last semester, and every single week at the meetings, the psychologists would give presentations, and they would use the term model.

They would say, oh, I'm using this model, or I consulted this model, and so on. And, I think it was like the first meeting, I raised my hand and asked, well, what does that mean? And I realized I was the outsider, I was the philosopher stepping in. Philosopher has 

Andrea Hiott: arrived. Yes. 

Paul Kelly: Yeah, they started proposing on, off the cuff, because they clearly hadn't thought about this in any detailed way.

potential definitions of what they meant by model. But each one of them either didn't represent the way that they were using the term previously, or was easily susceptible to counterexamples of how this term is used. So practicing scientists are using this concept without a clear sense of what it means.

[00:24:00] And I think that given that, I think philosophers can do good work to help the scientists, working with them, obviously, to try to hone in on a particular conception that captures their ordinary usage but also is, gives us an illuminating account of how this epistemic object works in the right way.

Speaker 3: Yeah. So 

Paul Kelly: that's the first reason. The second reason I think that talking about models is valuable is because there are big questions in the philosophy of science that relate to big questions in philosophy writ large, right? As you said, we want to understand the world 

Speaker 3: and 

Paul Kelly: we want to understand our place in it and what we are fundamentally.

And insofar as science can help us, answer those questions, or at least make traction on trying to answer those questions, this notion of models seems really important. People try to model internal states that we have, external things in our environment, um, and trying to think through this complicated epistemic practice [00:25:00] that scientists engage in and the philosophical implications of it, as well as the assumptions can give us a richer sense of our understanding of the world and how we come to it and how we can do it better.

Andrea Hiott: Yeah, I definitely agree with that. And I guess that's why I was pushing about this external representation thing and this trying to understand, think of a model as almost a shared space where people can come, um, to clarify and observe the kinds of things that might have annoyed you about that psychology class.

That it wasn't that those people are wrong necessarily, but they've come from a particular place of reference and trajectory. And if the term model has just been supposed, assumed, never really thought about, no, attention just not turned to it. Then they don't think about how they're using it, and they just assume, and in a kind of way we use cognition too, that it's clear what is meant by it.

But, um, so I guess I really agree with you about the role that the [00:26:00] philosopher can play, that you could play in, in, in clarifying that. But I do want to push you a little bit um, what do you think, is there some kind of danger in not understanding what a model is? Because for me Something I think about a lot, and the reason I'm bringing this up, is how we confuse the model with the process itself.

And, as we get into your work, I would, I mean, that's something I feel like happens a lot, I don't know, maybe you don't think it happens a lot, but where we, this kind of messiness that you were bringing up, can become a kind of, um, that we're assuming we understand. really important things about the world that were actually just perhaps only one representation of how it might have been, but we end up, you know, in wars over these kinds of things or, or, or so on.

So I don't know that map territory, um, confusion, does that relate at all to any of what you were just saying? Or, 

Paul Kelly: yeah, I think that that's, [00:27:00] that's well put. Um, I think The dangers, insofar as we don't reflect on what models are there are a variety of them. And in many ways, we could sort of distinguish the different dangers in terms of the different kinds of models that are out there.

I mean, earlier you alluded to the idea that sometimes the term model is used in a way to capture sort of model behavior or some model that's taken to be sort of illustrative or an exemplar of some kind of practice. That's a way that people use the term. Um, and that's a sort of normative conception of what the model is tracking, right?

That we ought to engage in a particular practice or structure our actions in a particular way. Like the model 

Andrea Hiott: behavior or role model, that kind of use of it. Is that what you mean? 

Paul Kelly: Right, exactly. And so that's a usage of the term model. It's not obviously exactly the same as the sort of scientific notion, right?

But they're clearly related because as you said, these are different representations. There are things that we use to try [00:28:00] to understand and try to triangulate and guide our various actions in a particular way. And they perform important sort of epistemic roles in our reasoning. And if we're not reflective about why we have the particular kinds of idealized roles and models in mind we could be led astray.

Um, Relatedly, in the scientific case, Oftentimes, I am also frustrated by similar kinds of worries expressed where it seems like various people perhaps conflating their representation, the model, for the thing they're trying to represent, be it, say, the world. 

Speaker 3: Yeah. 

Paul Kelly: Or the brain. Or the brain. So. So to take a very sort of, a clear example, I think, of this being the case, um, there is a certain kind of model, a certain kind of surrogate that we could use to try to describe some phenomena, which is sometimes described as a mechanistic model.

And the defining [00:29:00] hallmark of a mechanistic model is that it posits various kind of structural relations between various components. There are spatial and temporal relations between them. And the interaction of those various things are taken to produce. a particular phenomena. Now, this is an extremely successful kind of model.

Speaker 3: Oh, yeah. And 

Paul Kelly: it's built off a particular kind of analogy, historically speaking, between, say, the clock and how the clock works and why the hands move that they do, as they do. That's the phenomena, and those are the mechanistic details underneath. 

Speaker 3: Yep. 

Paul Kelly: Now, there's lots of different kinds of phenomena that are extremely well understood in terms of mechanistic models, but oftentimes people take the success of these kinds of models as reason to think that the phenomena themselves or maybe all phenomena that we could possibly encounter are always mechanistic or can always be explained solely in terms of a mechanistic model.

And that, as you pointed out, [00:30:00] confuses perhaps One critique someone might offer, facts about the model for facts about the world and this hasty generalization that all phenomena or all constituents of the world are best understood in terms of mechanistic descriptions and explanations might not be correct.

Yeah, that's, 

Andrea Hiott: that's very well put, yeah. 

Paul Kelly: There are lots of different possibilities of the different kinds of models that we can deploy, and there's also lots of different kinds of epistemic aims that we might have in mind. For people who don't, I mean, 

Andrea Hiott: epistemic aims, you just mean, I mean, epistemic is like a theory of knowledge and so on.

So you're, we're back to this idea of trying of, of theorizing or coming up with some kind of a knowledge trajectory, right? 

Paul Kelly: Yeah. So different kinds of epistemic aims is just a, a phrase to capture that there are different purposes. that we could use a model for. And some of them are specifically to try to gain something related to knowledge.

So maybe we just want to accurately describe some phenomena. Maybe we want to try to predict [00:31:00] the phenomena. 

Speaker 3: Yeah. 

Paul Kelly: Maybe we want to try to explain the phenomena. Or maybe we want to try to understand the phenomena. And philosophers have tried to spend a decent amount of time trying to make sense of what those things are.

And there's all these new questions, which are how models might facilitate these various kinds of cognitive achievements. Um, so the, the mechanistic approach, um, might be a good one for various kinds of phenomena, but a danger involved with this kind of reasoning which depends on a sort of.

potentially overly simplistic account of the types of models or the kinds of things we can use models for is certain assumptions about the models directly contributing to our metaphysical picture of what the world actually looks like. And so I think you're right that oftentimes people conflate the map for the territory, right?

They conflate the model, the representation for the thing that it's supposed to be a model of. But there's also an interesting sort of feedback loop, if you think about it, [00:32:00] between these two things. If we look at the history of science, people would develop a particular theory or a particular model of the way the world is.

And insofar as it was successful, it would inform their metaphysical picture of the world, which would then contribute to later developments of the epistemic enterprise of trying to figure out the way the world is. And that in and of itself, I don't think is elicit. I don't think that that's epistemically vicious.

I think there should be this feedback loop between our sense of the way the world is and the way we reason about it, but we need to do it very carefully because underlying assumptions might make us presuppose facts about the world that aren't warranted by our reasoning. 

Andrea Hiott: Yeah, I think that's very well put.

And you can't separate it. I mean, I think of something like, um, how the world changed with the, um, the theories of Einstein or something with quantum physics or it's once you, once there is some kind of new proposition about how the world works that can be somehow understood on a [00:33:00] public level, it does change whatever we mean by cognition and consciousness in a very, in a very big way.

And that's, you can't really separate all of that. But as you were talking, I was thinking about this idea of mechanism that maybe we can dig into it a little more. Um, I was remembering this 1998, or I tried to look back at it, just before we came on paper by I can't even remember the authors, but I know Carver is the last author and it's the one, That's always referred to about mechanism these days.

Um, I I remember my master's test to get into school for neuroscience. One of the papers that you could write about was the one what is it? It's like mechanisms, thinking about mechanisms, thinking about 

Paul Kelly: mechanisms. Yeah. 

Andrea Hiott: And, um, I don't know how you feel about this and this is, this is a sloppy way to put it, but in that paper, one of the mechanisms that they describe is the neural polarity it's, it's definitely depolarization of neurons in some kind of way that is the, is the model. But this idea [00:34:00] of mechanism, that first sentence um, I did write it down. Okay. In many fields of science, what is taken to be a scientific explanation requires providing a description of a mechanism. There's already a lot of the words that you brought up in there. There's explanation, there's description, there's mechanism.

And, um, In that, in that paper though, I think is the definition where it describes mechanism in the way you just unpacked as like activities and entities, right? That are structured in a certain way. Um, and we have these space time I'm sure you can describe this much better than me, but, um, maybe you can talk about the importance of, of that paper.

And is that where you think Is that where we're really looking at for this definition of mechanism now? Because of course in that paper they say mechanism has been used for years and years and years and so on and so forth, but um, I feel like at least in the little world I'm in of neuroscience it's always gone back to that paper.

So when you say mechanistic, what are you kind of like, What's your references? Is that, is that one of the papers, or is there something I'm [00:35:00] missing? 

Paul Kelly: No, that's, that's exactly right. So that paper I think is one of the most cited recent philosophy papers of the last several decades. 

Speaker 3: Yeah. 

Paul Kelly: So it's, it's this paper thinking about mechanisms.

Um, it's this co authored paper With amongst other people Carl Craver and Lindley Darden, um, and Peter Mockamer. Um, and it's, it's come to be known as an extremely important touchstone in the discussion about mechanisms and also discussions about models. In many ways, the conversation about models, I think, is wasn't quite as developed at the time.

And so a lot of these discussions were focused on what it means for a mechanistic model to explain. And the conversation has sort of taken a step back now and said, well, what is a model to begin with? Mechanistic models are one type amongst others. Um, your question of, well, what is a mechanism and what are mechanistic models is a really good one.

And even among [00:36:00] people who, are sometimes described as members of the new mechanists. People who think that an extremely fruitful way of thinking about explanation description, prediction, and scientific practice across the board, especially within cognitive science, Is in terms of mechanistic perspectives disagree with each other about what mechanisms fundamentally are, and also the relationship between models and the normative constraints that a commitment to mechanism would require.

Um, so that paper in many ways sort of kicked off this entire cottage industry of people talking about mechanisms and a lot of detail. I actually had a conversation with Lindley Darden, one of the, one of the authors who said that they had worked on this paper for two or three years, trying to revise the precise wording very, very carefully before it got published.

And even so There are people who are in the mechanistic camp who think that there are confusions and problems with it. Um, One person who has done a lot of work on this [00:37:00] is Stuart Glennon. Stuart Glennon is a self identified member of the New Mechanists. He wrote a book a few years ago trying to provide a very accessible overview of the sort of mechanistic approach.

And he's become well known for being an advocate of a view that's sometimes called minimal mechanism, which is trying to distill down just to the bare essentials, what we mean by a mechanism. And then a mechanistic model is understood to be an attempt to represent that, that particular thing. Um, so whether or not there are precise.

Necessary and sufficient conditions for being a mechanism is controversial, but the general idea, I take it, is that in actual scientific practice, especially within certain fields like cognitive science, people talk in terms of mechanisms. People use this phrase, and moreover, their assessments, their judgments that a particular phenomena has been adequately explained often depends on their sense that they've [00:38:00] identified a mechanism that produces it or underlies that phenomena.

Andrea Hiott: So something like neuronal depolarization, which I think, if I'm wrong, I'm very sorry to the authors, is what they use as a big example of an, of a mechanism in that paper. So it's become that we try to identify certain mechanisms, um, but those are defined as entities and activities, right? They're not, I'm trying to lead towards trying to get a little bit of a definition of what a mechanistic model is so that we can then look at what a dynamic model is in contrast, if it's even possible.

Paul Kelly: Yeah. So the, the, the general convergence, my sense is that what a mechanism is fundamentally, and this comes from Stuart Glennon's like minimal characterization, but the bare minimum that needs to be the case for something to be called a mechanistic, you know, representation is that it specifically refers to entities.

and relations, the interaction of which is taken to be what produces, [00:39:00] underlies, or there's a few other kinds of verbs that specifically are trying to capture certain kinds of relations. the phenomena. And the general idea is that when a model is mechanistic, it tries to represent the phenomena, but then it also has this other thing that's represented, which is taken to be what underlies, right, the phenomena.

And to connect with our previous conversations, this term wasn't popularized or wasn't, you know, widely used, but really what's going on is that what's, what's underlying this in terms of the model is the surrogate. It's the diagram, right? The diagram that the scientist puts forward. is supposed to try to make posits about the various kinds of parts and relations that are really there in the world.

that produce or underlie, right, the phenomena that you're trying to investigate as a scientist. 

Andrea Hiott: Okay, so let me just dumb it down for myself a little bit. So you have something like a process um, the phenomena, I think is what you're also calling [00:40:00] it, or what it's also often called in the literature.

Um, that's the actual ongoing, let's say, neuronal depolarization that we've named depolar, not neuronal depolarization, but it's just going, going, going. Okay. But we're saying, okay, that's a mechanism. Um, and then we're going to model it in the way that you just described with this kind of surrogate or external thing.

So the mechanism isn't the model. The mechanism is sort of the process, but I feel like that all gets very confusing. Um, I feel like often just already, even with trying to make, take something out of the body and call it a mechanism is already confusing. A model, I don't know. I mean, that's very messy, but, um, I really, I do want to get into it a little bit because I think it also connects a bit to the more recent I guess you could say trend towards dynamical systems, complexity, and [00:41:00] those kinds of models, which already start in that place that I just described of understanding we're always working with processes and that there is no kind of system that you take out of, its other possible connections to other systems.

Um, but for me, this overlaps a lot with the mechanistic and the, and the dynamic. So I don't know in your, in your research, how you've, or how people distinguish those. Um, I did read a little bit of this Gervais paper that you talk about. Um, in your own work some, and I think in that paper, the distinction is sort of between, um, like a discrete, like with the mechanistic model, you can have these kinds of discrete series of, um, operations or something, whereas in the dynamic model, it's more like.

Describing the changes itself. Um, but I still don't really understand this at all. Maybe you do, maybe you can help me. 

Paul Kelly: Yeah. I mean, those are, those are great questions. Um, the first part about well, what's the difference [00:42:00] between the mechanism and the mechanistic model. So let's, let's pick an example so we can get traction on it that we seem to have a more intuitive grasp of it.

So take. My watch, right? Let's take the paradigmatic phenomenon 

Andrea Hiott: loves to think about the watch. It's a good one. 

Paul Kelly: Everyone likes the watchmaker, right? This is the clockwork universe. So here's a phenomena. We could describe the movement of the minute hand and the hour hand on the surface of the watch face.

That's just the description of the phenomena. And let's say I want to investigate and I want to simply describe that in a compact and efficient way. I could develop various math equations that talk about the behavior of these things and how they relate. That one moves around the whole clock face, one that's every hour, and the other one every 12 hours, and so on.

Right? Now let's ask another question. Why? Why do the hands move the way that they do? And to try to answer that, We wouldn't just have a description that's [00:43:00] accurate of the phenomenon, right? We would also want to dig a little deeper and create a representation. Maybe I write it on a piece of paper, or maybe it's in my computer or something of say, the movement of my watch, all the various gears and springs and so on inside of it.

Now, what we don't want to confuse ourselves with is the difference between the diagram, the surrogate. and the actual physical mechanism underlying my particular token watch. One is a representation that we develop and construct and intervene upon to justify or facilitate inferences that we make about this thing.

So we should always keep in mind the difference between the model, the surrogate, in this case, the diagram, let's say, of the various components and the actual components themselves. Right? Those are very different. And keeping that in mind people, I think, of a particular kind of psychological persuasion or [00:44:00] inclination are inclined to view things in a mechanistic way.

And are inclined to prioritize the consideration of mechanistic models and hypotheses over other kinds of models and hypotheses that we might have considered. And so you alluded to an entire other class of models that are explicitly not mechanistic, which are called dynamical models. Now the kinds of questions that motivate investigating some phenomena mechanistically in many ways are assumptions about, hey, let's think about it as if it were a watch, right?

Let's think about processes inside the brain as if there's essentially a movement, these various springs and gears and levers. Inside the brain, which would explain, you know, the depolarization of neurons. Or something more complicated like long term memory or something like that. Um, but that's not the only way that we can think about it, right?

The selection of the mechanistic model was a selection of a [00:45:00] variety of different options. We could have tried to represent the various processes inside of this watch, not in terms of positing particular components defined by their spatio temporal location, but perhaps simply in terms of. functions of these various components without making any claims about physically speaking, how they're instantiated.

Or maybe we talk about the various landscape of different trajectories that the various objects inside of it seem to move or that the object as a whole seems to move. Maybe we flesh out the various possible states that it could be in. And then given that abstract sort of landscape, we can see that it sort of converges and ebbs and flows in particular ways.

Now that kind of model is not a mechanistic model. It doesn't make specific posits of structural components and relations, and it doesn't decompose in the right kind of way where we can zoom in on one particular thing and talk about it in detail. It's this kind of complicated smooth terrain. [00:46:00] And there's been a lot of debate in the last couple decades from people who are sympathetic to mechanistic models and people who come from dynamical systems theory.

Who developed these state spaces and so on about how to make sense of these models and what epistemic work they can do. Mechanists, for example, often we'll say that dynamical models might be helpful in describing a phenomena and maybe they can help inspire future avenues of research, but they can't explain anything to really explain something you need to get down to the nuts and bolts, you need to talk about the gears and the levers.

And so there's this wary, I think, from people who are sympathetic to the dynamical systems approach. That this is presupposing certain normative constraints. on the kinds of things that can do explanatory work, given antecedent assumptions you have about the nature of the phenomena itself. Maybe there are certain underlying philosophical assumptions about everything being like a clock that maybe if we reflect on we can realize might be incorrect.[00:47:00] 

Andrea Hiott: Right. I think that's what was, is so attractive about thinking of mechanisms, about the paper and about also starting to think of something like the body or the brain in a mechanistic way. I think there's an excitement to think about it like as if it is a computer or as if it is a watch. Um, and there's actually, that can be very beneficial.

But I think what bothers me is that. That is already to me, um, what I was trying to get out at the beginning, an external representation. It's not actually the process itself, but it's, it's taking the territory, it's taking that territory, mapping it, and then thinking the map is the territory. It, it, it seems to me quite often, um, even in something like neuroscience, where of course it's very helpful to, Say this is a mechanism, something in the body, like with the nervous system, neuronal depolarization or whatever, but you haven't actually [00:48:00] extracted that in any way, and it's not your watch.

Um, I mean, the watch is a great example. These are easy examples, because they're actually already externalized. They aren't, Dynamic processes to begin with. I mean, something like time and all of this, we're in a way modeling that already just to create by creating the watch. So we're like meta meta representing when we use the watch as our example, or the computer as our example.

Um, which was what I was trying to get out with that mental representations thing, like to know you have a mental representation. In the way that, philosophically, you need something like language, which for me is also a model or a representation. So, I know this is getting messy, but, for me, um, and I don't think I'm completely alone, I mean, even in Gervais, he's saying dynamical and mechanistic can be very similar, but I feel like what the mechanistic literature is helping us understand and describe is actually not mechanistic, which is a [00:49:00] very weird way to say it.

Um, and that actually it's, It's very helpful. It's one way to understand a process and modeling it dynamically is too. But in both cases, you have some kind of process that actually isn't itself mechanistic. Does that make any sense at all? Or can you kind of like show me where maybe others have addressed that?

Paul Kelly: Yeah. I mean, I think that, I think that that's right, that one potential upshot someone might take from this conversation about mechanisms is that the actual processes in the world. Are perhaps much more complicated than a sort of. what we might think of as naive mechanistic kind of approach is able to capture.

I think also in the mechanistic literature though, to sort of flesh out some conceptual distinctions that might be helpful, very often mechanists will make claims and use the term mechanism, but will actually be referring to different theses. And [00:50:00] it's not entirely obvious that that's what's going on. So, Arnon Levy has a paper from a few years ago called Three Kinds of Mechanism.

And he distinguishes three ways that one could be committed to certain mechanistic commitments. The first one he describes, if I remember correctly, is sort of like a metaphysical mechanist. Right, you think that the world and the things in the world are composed of mechanisms. That's a particular thesis you could have.

And then, you might hold different views about the best way of investigating those things. But, that's a particular metaphysical thesis someone could accept. One could also be what, what we might call an explanatory mechanist, right? They think that the best kinds of explanations that one can provide are mechanistic explanations.

Maybe they're not the only kinds of explanations, but they're the best kind, and we should actively seek them out, something like that. Or you could be what he refers to as a strategic mechanism, or a methodological mechanist, which is to say, as a practicing scientist, I have lots of different hypotheses I could dedicate [00:51:00] my time to, but I'm going to prioritize the mechanistic hypotheses because I think that strategically speaking, that's.

the most effective way of perhaps obtaining explanations or tracking metaphysical relations and so on. So these are different kinds of commitments that we would all term as mechanistic, and the relationships between them and the justificatory relationships between them might be complicated, right?

Someone might say, well, the reason why I should prioritize mechanistic hypotheses over others is because those are the explanations that I want, Why do I want those explanations? Because I think they track the metaphysics in the world, right? And so on. Um, but each one of these can come apart, right?

Someone could be committed to the metaphysical picture, but not the explanatory picture, or not the strategic picture. Imagine someone who is pursuing dynamical systems theory which just, to specify that in a little bit more detail, is a specific approach that doesn't view things fundamentally as, say, like a watch but more in [00:52:00] terms of a dynamical system, a system that changes over time, and trying to flesh out the possible states that the system can be in and also trying to map particular kinds of trajectories that similar systems tend to move through that state space.

Maybe you want to approach the world like that, but also think metaphysically there's just mechanisms out in the world. That's a coherent I don't even know 

Andrea Hiott: what that means, just mechanisms out in the world. I mean, what the, what does that mean? 

Paul Kelly: So Carl Craver, for example, who's one of these, you know, extremely strong advocates.

He's the, he's 

Speaker 3: the With this word. When 

Paul Kelly: people talk about the new mechanist, Craver is always the person that people discuss. Craver is of the opinion that the world, fundamentally is composed of mechanisms. The phrase that he sometimes uses is that ontically the world is composed of mechanisms.

And there are mechanisms that produce phenomena over extended periods of time. And there are also mechanisms that constitutively underlie [00:53:00] phenomena at a particular time slice. So all kinds of things in the world are parts of mechanisms. And he takes this metaphysical commitment to in many ways inform his particular views about the normative scope of what kinds of explanations scientists should be pursuing or should accept as genuine explanations.

Um, this is in many ways a resurgence of this sort of clockwork picture of the world, right, where the world literally is this extremely elaborate set of causal interactions between component parts that, if we were to reference properly, would explain, you know, the phenomena. 

Andrea Hiott: It just seems to assume that those parts are static.

And I think we know pretty well with our science that there's no static part. And, um, I mean, first of all, I think this literature has brought so much to us. So I'm not trying to say it's like wrong or bad or something like this, but I do think there's a way in which just thinking of everything as a mechanism is a little.[00:54:00] 

It doesn't make a lot of sense to me anymore because, which is again what I'm trying to get at is because the mechanism is our way of representing this ongoing process to try to understand it. So if we just say everything is a mechanism, okay, it just means that we can model everything. But all those things are still changing.

There's no one static. Um, mechanism and something like an ontic explanation in the way that he would use it once you have add explanation to it. You've mechanized it and you've modeled it in a way. Um, I know those aren't the same words, but to me, you've, you've used some kind of representative system or you've made a representation of it.

It's not just the ontic anymore. It's some kind of little snapshot of it, of it, and all that can be very, very, very good. Um, it just, I don't, yeah, it's, it's a wording I don't understand, but, but, I could go on about that forever, but what I want to get at is your work a bit too, and so, we have these two camps.

I mean, I'm generalizing in a very big way, but [00:55:00] after all this time in literature, it's been what half a, I mean, 50 years or something, um, of thinking about mechanism and, and, and, 1998, I think was that paper Thinking of Mechanisms and, and there's been these kind of two camps can we generalize it and say, there are these mechanists and then now there's kind of people who think about dynamical models.

And those things seem, as you said, they, they don't go together or do they, I don't know, but a lot of the criticism of the dynamic model is mechanist, right? So could you unpack that a little bit? Like why. If I'm a mechanist, would I not want to also accept that something could be, that dynamic models could be explanatory or descriptive or yeah, maybe we can also go into what those words mean a little bit, but maybe you can just help me understand what are those two, what's the tension between those two as they've evolved up, you know, to this point.

Paul Kelly: Yeah, it's, it's a great question. 

 There's been this sort 

of emergence of a variety of different schools of thought, [00:56:00] but two of them that, over the last two decades or so, have disagreed about these assessments that scientists should make concerning how we should assess models are what are called the mechanists. which are philosophers and cognitive scientists who think that approaching this conversation with an awareness of how philosophical of science, philosophy of science debates about mechanistic explanation have sort of played out, can inform this discussion.

And then people who are dynamical systems theorists, people who say, let's have the kinds of models that we develop be ones of state spaces and trajectories through terrains and so on. Now, Mechanists are of the opinion and they're explicit about this in their writing David M. Kaplan explicitly in several of his writings, 

um,

describe a phenomena 

in some compact, efficient way. It might be super helpful to start by trying to represent the possible states, [00:57:00] That, say, if the phenomenon is some system, for example, and the way it behaves, just the states of the system, and maybe getting a sense of the various trajectories that move through it, and various areas of the trajectory, the landscape that the system tends to converge to.

That's very interesting, in terms of describing the phenomena. And in many ways, it might be really illuminating. If we just have a bunch of data points, and we don't see any sort of pattern between them, representing them in terms of a landscape might be very fruitful. Mechanists agree with that. very much.

Maybe, moreover such a representation might be predictive. Maybe if we have this terrain the various possible states the system could be in and the trajectories, and we say, what would happen if the system started in this particular state? Where would it likely go? Let's make a prediction. That model might facilitate those predictions.

But what the mechanists want to say is that there's something special about saying a model is an explanation, or that it [00:58:00] explains a phenomena. And that is different than simply describing the phenomena or predicting the phenomena. To say that you explain the phenomena on the, say, mechanistic view, you need to have good reasons to think that you're tracking the causal structure of the world that produces the phenomena.

So let me give an example that hopefully will intuitively track these distinctions. Imagine, you know, it's a long time ago, and we're, we're cavemen, to use the, you know, thing that you alluded to, and we look up and we see the moon, right? And each night, the phase, the moon looks slightly different.

It has different phases. And then moreover, we look at the ocean, and the ocean seems to come in and out at a particular rate, a particular pattern, there's tides, right? Now we could describe, say, the movement of the tides in a particular way, right? Maybe some math equation. But then, in addition to that descriptive account, when we realize the moon's [00:59:00] phases seem to correspond or correlate in a very interesting way to the tides, we could predict the tides.

and how they would move on the basis of the phases of the moon. That's a very interesting model. It's a correlation though. It's fundamentally a correlation. To say that you have a model that explains the tides, right, you would have to supplement that predictive model with some kind of causal oomph. It would actually have to represent something like gravity.

Right? Some sort of causal mechanism by which this particular process unfolds. So, developing an accurate description of the phenomenon is the first step in inquiry. And having a predictive model is extremely good, insofar as prediction is the main aim that you want. But explanation is something that's special.

And what the mechanists want to say is that when we start talking about explanatory models, only one kind of model is going to fit the bill. And those are mechanistic models, because the world itself is composed of mechanisms. Dynamical systems models [01:00:00] might be really helpful in describing, might be really helpful in predicting when it comes, when the rubber meets the road, when we want to try to explain what's going on.

The only game in town are mechanistic models and representations. That's what the mechanists want to say. And as you alluded to People who are proponents of dynamical systems theory often object to this characterization, and they say, well, there are various kinds of metaphysical assumptions that the mechanists might be making to try to advocate for this constraint on the kinds of models that we would deem legitimately as explanatory, um, and so maybe this is unjustified.

But then there's this lingering question, even if that's true, when is a dynamical model more than just a predictive model? When is it actually genuinely explanatory? What, if anything, makes it the case that appealing to some dynamical feature of the terrain of the state spaces provides an explanation, as opposed to merely a more accurate prediction?

So, [01:01:00] even if the mechanists are mistaken in terms of overtly restricting the class of explanatory models, there's still important, interesting philosophical work to be done in explicating or unpacking what it means for a model, regardless of the type, to explain anything. as opposed to merely predict or describe.

Andrea Hiott: Yeah, that's well, but I, I think, um, I wonder about, um, something like, I mean, that's a great explanation you gave with that. We need something like explanatory, like gravity, for example, you said but even we could there are always ways we can zoom in and out of different scales and start to understand that something like gravity is a little bit of a, a prediction or it's, It's a probability, um, in a particular landscape because, you know, you go to the moon and it's going to be different.

Or, um, if something changes with certain conditions of regularities on the earth, it's going to be different. So [01:02:00] I feel like those, a lot of the mechanistic models are assuming things don't change. And I feel like I mean, this is very general and I feel like the dynamic systems is assuming everything's always changing and you can always see it from places.

And I don't see that those need, I see that as the, as a primary conflict, um, not that both aren't actually great ways of understanding the world. But I don't know. What do you think about that? 

Paul Kelly: Yeah, I'm not so sure. I think that there are different kinds of explanations we can give, and a common distinction that's made is between providing a causal explanation of something, and causation is a kind of thing that occurs over time.

So a particular event happens, Which then causes some other event to occur, right? One billiard ball hits another billiard ball at t1, which causes the second billiard ball to move away at t2, something like that. It's extended over time. That's also a 

Andrea Hiott: very linear way of, of looking at [01:03:00] things too. I mean, causation 2 we could debate forever, but yes, you can always kind of, in a linear way, say this happened and then this happened.

Um, but go ahead. 

Paul Kelly: But the, the, the main point is just that, Causation extends over time, right? That's a kind of thing that extends over time. But that's not the only kind of explanation we can give. Interestingly, over the last, I don't know, century or more, philosophers of science have recognized there's another kind of explanation that's sometimes called constitutive explanation.

And this is where you explain something at a particular time. So, why does the vase have the particular structure that it does at T1? No reference to T2 or anything else, just to T1. And maybe you zoom in and you focus on the arrangement of the various atoms and so on and things, the pieces of the you know, the vase.

That's what gives us its structure. That's what explains its structure at T1. But it's not. a temporally extended claim about causation. [01:04:00] So one might read the mechanists to say, oh, the mechanists are just trying to promote an account of constitutive explanation. They're focused on the components and the relations at a particular time.

Whereas maybe dynamical systems theory talks about temporally extended processes. But I'm not sure that's entirely quite right. The mechanists think that their models are aren't necessarily limited to representing processes or states of affairs just at a particular time slice. They think they also can capture the behavior phenomena over time.

So take a mechanistic model of the movement of the watch. It's true that we can sort of pause our playback of the watch at a particular time and just zoom in on particular gears and levers. But we can also, depending on how elaborate the model is or the surrogate, play it forward and actually watch how the various components interact over time.

So, it's not seemingly essential to a mechanistic model or mechanistic approach to investigating a phenomena that you [01:05:00] only look at a particular time slice. They often talk about phenomena occurring over time. Dynamical systems theorists do that too. And maybe they have better resources to try to accommodate this sort of, you know, the temporally extended kind of notion of the phenomena being extended over time and space and so on.

But I'm not quite sure it's right to say that the mechanists don't have any resources to talk about things extending over time and that they're inappropriately prioritizing sort of uh, synchronous or constitutive kind of approaches to things. Does that make sense at all? 

Andrea Hiott: Yeah, it makes sense. I guess I wouldn't say that, um, that they're necessarily trying to do that.

I'm just, I'm trying to understand really why you would have to reject The dynamical model approach, if you're a mechanist, exact I mean, maybe we can get into, um, the 3m, for example, maybe that would help illuminate this a little bit like, um, and [01:06:00] towards, because you've found a resolution, right?

A little bit between, between these. So I guess there's two things we need to talk about if you still have time. Yeah, it's just maybe a little bit more about, um, explanatory, like why that's so important. And I guess we could bring in the 3M a little bit too. Or I don't know, how, how would you like to, to move a little bit towards your, What you've, what you've done to show.

Paul Kelly: Yeah. So the main kind of normative principle that has emerged among the mechanists that if correct, definitely precludes dynamical models from being categorized as explanatory. is what's called 3M. And so the 3M principle is called 3M because it stands for the Model to Mechanism Mapping Requirement.

Model to Mechanism Mapping. So there are three M's there. And the [01:07:00] precise wording of the 3M principle has changed over the years, where mechanists proposed it, and then they tried to tweak it, and then they revised it, and so on and so forth. What's interesting about the 3M principle the way that it's proposed, at least initially there's a footnote.

And the footnote says, We take ourselves to say that this restriction on what models can be explanatory specifically applies to cognitive science but we see no reason why it shouldn't be extended elsewhere. So it seems like the authors have this idea in mind that maybe this is a restriction on Models being explanatory in science across the board unless we're given reasons to think otherwise, maybe we should just accept this as precluding other kinds of models regardless of the discipline that they're being developed in.

The particular details of the 3M principle and the precise wordings we don't need to get into, and moreover the wordings themselves have changed over the years. But the general idea is that it establishes a [01:08:00] necessary condition for a model to be explanatory. A model is explanatory only if it satisfies certain constraints.

Andrea Hiott: That's what I meant by static. That's the static that I meant. I didn't mean in the world so much as in the way that we need always this, but go ahead. 

Paul Kelly: Yeah, I think I see. So, so by static you mean that the, the proposal they're making is making a very static or restrictive demarcation, right? Between types of models, right?

Speaker 3: Yeah. 

Paul Kelly: Yeah. And that's, that's right. If the 3M principle is correct, right? Whenever we have a model and we have the sense that maybe it's explanatory, we should consult this principle. And if the model fails, the principle, it's not explanatory. Our initial sense of it being explanatory wasn't right. So it's an interesting discussion, right?

Trying to figure out what normative principles. we should adopt that [01:09:00] constrain our assessments of what kinds of models we categorize as explanatory as opposed to predictive and descriptive and so on. Right. I mean, like for me, 

Andrea Hiott: 3M is really helpful, but it could be one way to do it. It doesn't have to be the way.

And I, you know, if you look at from a dynamical systems approach, maybe if you're changing the references and the context in which you're doing this, um, process, you would need to change that a little bit. But, in any case, yeah. 

Paul Kelly: Yeah, so you might think that the 3M principle is a good way of unpacking necessary conditions for a model to be explanatory in a mechanistic sense.

Right? And interestingly, I mentioned that among new mechanists, there's disagreement about various things. Stuart Glennon, who is clearly a mechanist, in his book, calls it, his view, 3M star. And 3M star is essentially the same, but he specifies this is just for mechanistic models. Whereas the original formulation is not qualified at all.

Right? It applies to [01:10:00] all models across the board. 

Speaker 5: Okay. 

Paul Kelly: Um, it could also well be, that you alluded to, that if that's right, maybe we should be what we might call explanatory pluralists. That maybe our judgments about what kinds of models are explanatory depend on the kind of model it is, or maybe the kind of discipline that we're coming from.

Maybe the kinds of models that can do genuine explanatory work in cognitive science are different than ones that can do explanatory work in economics or sociology or virology or something like that. Maybe that wouldn't be too surprising, but what the Maybe you can get 

Andrea Hiott: to the same place with different models, like in category theory, that kind of shows us, right.

If you, which is something closer that you can sort of zoom out and you can see, you can get the same answer with two different models and it's a valid. That's 

Paul Kelly: definitely true. Yeah. So, so insofar as we're thinking about this kind of inquiry based on particular questions that we ask, and then we consult these surrogates or these models to try to answer those questions, we could get [01:11:00] the same kinds of answers.

even though the models that we consulted were very different, right? That's, that's definitely a live possibility. Um, thinking about our inquiry in terms of asking questions gets very complicated. There are particular views that think that questions are really fundamental to this process of inquiry.

Mechanists like Karl Craver are in many ways, I don't want to say dismissive but think that people have overemphasized the importance of questions in terms of our scientific inquiry. And he's an advocate of, as we mentioned earlier, the ontic conception of explanation. He thinks one usage that is of the term explanation, which is totally plausible, and maybe is the important one, uses the term explanation to refer to the things in the world, not our epistemic practices or representations about them that we would ask questions concerning or something.

But with that aside, the 3M principle is an attempt, right, to make a proposal about how [01:12:00] we should constrain our judgments about what models are explanatory. And it, in many ways, that's very fruitful, right? It's, it's gesturing towards a conversation that's worth having. But even among mechanists. Some of them think it's too restrictive.

Some of them think it should only apply to our evaluations of mechanistic models and not models across the board, right? Stuart Glennon with his 3M star, 

For both. 

Andrea Hiott: So if you have the watch again and you have the 3M, you're basically mapping the parts of the watch as you described it to the model itself, right?

Or, or how would you, like, how would you illustrate that with the watch? 

Paul Kelly: Yeah. So, so the 3M principle. A lot of people have responded to it. And I think a lot of the responses often misunderstand what the principle is saying. People will say things like, oh, the 3M principle requires that a model not only be mechanistic, but be perfect.

That it perfectly maps every single component to some posit within the model. But that's not what the principle says. It's just a [01:13:00] necessary condition that there needs to be at least one mapping. that is successful. And that seems like a really low bar, right? What the 3M principle says fundamentally is that here's a necessary condition on a model being explanatory.

There needs to be at least one posit, I don't know, a gear that is actually in the mechanism that underlies the phenomena. If your model doesn't even successfully pick out one gear that's in the mechanism, there's no way that it could be explanatory, even partially. Because what it means to explain is to track that thing, the mechanism underneath the phenomenon.

So it's incorrect to say that the three M principle, if we accept it requires that a model be perfect, because if that's the requirement, no model is ever going to be explanatory, right? No models. Perfect. Even the most robust, clear mechanistic model of say the movement of my watch will have some imperfections won't accurately capture every single atom right in the particular right place.

So [01:14:00] that's. That's too much. There's an entire separate literature that the mechanists have developed trying to make sense of what do we mean when we say that a mechanistic model is complete. And so, that gets really complicated. But the 3M principle is not that. The 3M principle is just saying, listen, to say that a model is explanatory, there needs to be at least one mapping.

But even that low bar Can you give an 

Andrea Hiott: example of one that is? That, that fulfills it? Because, I mean, the watch for me, you would have to map the watch to time or something to really make, for it to matter. I mean, if you're going to create a model of a watch, I don't know, I mean, can you give me an example of where that's actually valid?

That it that works. It was not necessarily, but there's at least this mapping as you just described of one, one to one, some simple. Yes. 

Paul Kelly: So taking the watch, um, mechanistic models and models in general are usually thought of as being composed of two main parts. The first is a representation of the phenomena, [01:15:00] and if that's accurate, that's great.

And if that's all you have, you have just a descriptive model. 

Andrea Hiott: So are we talking about drawing it on a piece of paper, a watch? Is that gonna work? I mean, what's 

Paul Kelly: So in the case of the watch, imagine I'm just accurately describing the movements of the hands on the face, and that's all I'm doing. I'm just saying one moves around the whole thing every hour, and one moves around it every 12 hours.

Right? So just, just 

Andrea Hiott: what you say can be the model. Just your 

Paul Kelly: It, it, so far as the model, so the model is just a representation of something and if all the model purports to do is just describe, then that's all it does, right? And so I could have a math equation which describes the movements of the watch hands and I'll be like, that's my model and that's the phenomena.

And it captures it in a really compact, nice, elegant way in the math equation. The equation by itself doesn't seem to explain anything, right? It describes accurately the phenomena that I'm interested in and that's great. That's a good place to start. However, [01:16:00] in addition to just describing a phenomena, there's this other part that we can also have to make it, say, a mechanistic model, which is a surrogate, a diagram with a bunch of various components and relations and so on that we think bring about this phenomena, right?

And if that, what the 3M principle says is that that's how we're going to, that's the thing we're going to zoom in on to evaluate if this mapping occurs. Right? There's the thing in the world, the movement of the watch, there's the surrogate, the diagram that we're putting forward, and there's the phenomena itself, right, and our representation of it.

Speaker 5: What's the phenomena itself if it's not the watch? What's the phenomena? Time? 

Paul Kelly: So the phenomena is just going to be the moving of the 

Speaker 5: hand. 

Paul Kelly: The movement of the hands, right? Just the movement of the hands, right? Why, now we can ask, why did the hands move the way that they do? And to answer that, what the 3M principle says is that our model, to be genuinely explanatory, Needs to have at least one of the posits.

So [01:17:00] one of the gears that it says actually in the mechanism, if my model posits a bunch of years where there's just no mapping whatsoever, it just totally misrepresents what's going on underneath here. I mean, let's say it thinks there's a computer inside of it, or let's see. It says there's a bunch of sawdust just underneath here.

There's, there's no mapping. And so it's not explanatory. And so what's interesting about the 3M principle is that it imposes this constraint on the kinds of models that can satisfy it. Dynamical models, which just don't make posits of particular components like gears and levers and springs, are going to fail even that low bar.

They're not even going to have a single thing mapping over. To the underlying mechanism, right? 

Speaker 3: Yeah. 

Paul Kelly: In the case of the watch, it's very clear. There is a mechanism, right? I can crack it open. I can look at it, right? 

Andrea Hiott: Yeah, because it's already itself created as a mechanism. So, I mean, 

Paul Kelly: yeah, 

Andrea Hiott: it's I think you referenced Greg Calendar, who's actually one of my, he's my philosophy of science advisor, [01:18:00] but, um, and, and some of his work where they talk about, um, You know, you could have a salt shaker or something.

Is that the one you where you've decided that this represents, um, Paris on a, on the table as you're discussing, like how far Paris is from Berlin or something. Um, and that can represent it. So there's this kind of context that's also important, but I mean, I don't want, I don't mean to get. out of the, out of the discussion.

But I, instead of the watch, maybe we can think about the planets. And if you make, cause we all know this as a kid, right? The planets, I mean, then we have something that's not the watch because the watch to me is already a mechanism that's externally replicated, but the planets aren't, but then we create a replica of it.

Like with styrofoam, I think is one example you use of the planets in styrofoam. And so then the model is, The mechanism is the solar system, right? The model is the styrofoam. Is that right? And what is the, where, how do you, what's 3M? Can you explain [01:19:00] 3M with that example? 

Paul Kelly: Yeah, sure. I think so. Maybe, maybe this will be helpful to get a little traction.

So, so Glennon, I think very helpfully proposes what he calls the the model's first approach to understanding what mechanisms are. And he wants to say that mechanisms are individuated. by the phenomena that they produce. So in the case of the solar system, there's a whole host of different phenomena we could, we could zoom in on, right?

So I might want to have the phenomena be the rotation of Jupiter, or I might want to say it's the orbit of Venus, etc, etc, right? Now, those are particular phenomena. And in fact, they're particular propositions, we could write them down, right? And they're in fact true. assuming they are true and our representation of them is right, yeah?

And then we could ask, once we have this proposition in mind, what explains it? And the mechanism on the mechanistic view is what's going to be the causal processes and also constitutive relations that lurk underneath that [01:20:00] make it the case that the particular proposition that we're investigating, the phenomena, is true.

Um, the solar system is an interesting one because it's one that we're familiar with. And it's also one that is governed seemingly by gravity, which we alluded to earlier. And even though mechanists sometimes talk about gravity as a mechanism, Glennon calls it a fundamental mechanism. It's in many ways a limiting case.

Because, remember back to how we described what a mechanism is. It's a bunch of component parts that interact to produce something. But gravity doesn't have parts. It's it's this fundamental force, insofar as it exists in the way that we think about it, causes things to move in particular ways. And so, Or it's an interaction 

Andrea Hiott: between a lot of other parts, which we, 

Paul Kelly: yeah, which we call it.

And so what, what Glennon wants to say is that there are. limiting aspects where the mechanistic approach [01:21:00] in some ways sort of breaks down and gravity is one limiting case where there is just no deeper level in which there are parts that seemingly bring it about unless maybe there's some complicated quantum interaction that produces gravity or something.

I just think 

Andrea Hiott: once you're talking about anything that hasn't already been produced like a watch or some kind of thing that's already a machine that's produced Um, then you get into this problem, but I, I will save that. I will bracket that because I want to come back to, okay, so we have 3M and basically if I'm a mechanist, I would say a dynamical model doesn't satisfy that.

Right. It can't possibly, or could it? 

Paul Kelly: That's right. It can't, it just fundamentally constitutionally, the way they're set up and the kinds of posits that dynamical systems. make just, just are never going to have this mapping. But here's an interesting point. It's not just dynamical models that get excluded by this.

Because of the specific requirement that the posits be not just successfully mapped, but mapped to structural, [01:22:00] physical, spatio temporally defined things, There are other kinds of models that also don't satisfy the mopping the mapping that are very common in cognitive science. So-called functional models, for example.

Speaker 3: Mm-Hmm. 

Paul Kelly: that describe something. So, solely in terms of inputs and outputs of particular components, let's say, but without making additional claims about the physical properties of the components themselves. A lot of people, I think, have the intuition that that kind of model is at least capable of providing some explanatory power.

But 3M would say no. They would say that's a good start, but you need to do more. You need to tell us about the actual gears. Tell us about the physical properties. If you don't do that, the mapping's not going to obtain, you're going to fail this normative constraint of 3M and the model won't be explanatory.

So this relates to, you mentioned the Gervais paper. I think he talks about the idea of upgrading. a model. And so, insofar as you have a dynamical model, sometimes what a mechanist might say is, that's great, that's a first [01:23:00] step. Now put mechanistic detail into it. Convert it into a mechanistic model so it satisfies our principle, and then it'll be explanatory.

Now, that's definitely one way of doing it, right? You could have a dynamical systems model, and also add in a bunch of claims that might satisfy this mapping. But when you do that, what you do is you change the kind of model. You make it a mechanistic model to satisfy this requirement. which we all agreed from the get go mechanistic models can satisfy the requirement.

The question is if a model that doesn't have this mechanistic structural detail is still plausibly construed as explanatory. That's the, the main disagreement. 

Andrea Hiott: So there's this judgment, I guess, by mechanists, some mechanists, or all mechanists, I don't know, that dynamic, dynamical models are non explanatory, right?

But you seem to just be hinting that, that there are models, um, That, that could be mistaken, actually, and that we could maybe show that in pretty, an easy [01:24:00] way, right, that they could be maybe explanatory. Would you agree with that? I mean, first of all, do all mechanists think that dynamical models are non explanatory?

Paul Kelly: All mechanists that I'm aware of, however, I recently learned at a conference that I was at in Florida, that in a recent talk Bill Bechtel, who is a Historically a strong advocate of these various mechanistic approaches was asked explicitly about contemporary dynamical systems theory and if they could be explanatory.

And he said, maybe. It's possible. He's the only one that I'm aware of. And everybody at the conference was very surprised. And people tried to qualify and said, well, it was just at a conference. He didn't put it in writing yet, but there may be some, some progress on this issue. However, the main core new mechanists like Kaplan and Craver, especially I think, no these, they're in writing, in multiple articles, proposing the 3M principle, but also explicitly [01:25:00] engaging with certain dynamical systems theory models and just saying, listen, these things just can't do it.

They can't be explanatory. 

Andrea Hiott: Okay. Well, this is just, I mean, on a really general level, if you're listening to this, this just seems crazy because, um, we can use dynamical models to help us understand things. I mean, that's just. That happens, right? So is that not a form of explanation if it's helping us understand something?

Does it really have to be something like gravity, which by the way, as we've already discussed, isn't, I don't think fully explanatory either in meeting all these conditions, but that would be a whole, that would be two hour discussion. But just, okay. So you've written about this too a bit if we just understand something in a certain way, if, if the model is helping us understand something, can we, can that be a form of explanation, or do we just have to throw, that also doesn't work for the mechanism, or, or for the literature?

Paul Kelly: I think bringing in the notion of understanding is probably a way of, illuminating this conversation in a [01:26:00] really rich way. So this debate has been phrased in the language of explanation. The 3M principle is proposed to say, listen, here's a constraint on the kinds of models that can explain. And there've been a couple people Gervais, you mentioned also my, my advisor, Larry Shapiro, have said, listen, maybe we should just sidestep this issue and say, these models can clearly help us understand.

And that's good enough, right? That's a good kind of. epistemic or cognitive achievement, and we can sort of move on. But interestingly, there's this other related literature on the nature of understanding as a mental state and the different constraints on it. And one popular view among people who write about understanding is that what it means to understand, at least when that use that term is used in the scientific context, scientific understanding, is to grasp or recognize in explanation.

Andrea Hiott: Yeah, so if that particular view, 

Paul Kelly: yeah, if that view is correct, then conceding that dynamical models can provide understanding [01:27:00] implicitly commits you to the idea that they're at least somewhat explanatory. You're grasping something that's explanatory. And as you point out, I think intuitively there are facts that can be represented in a dynamical model, which intuitively really seem explanatory.

So take, for example, um, imagine a piece of paper. Right? And imagine I drop a piece of a drop of water on the paper and it moves in a particular pattern as the, the paper moves. Right? So, why does it behave that way? Well, you could try to give a mechanistic explanation as to why it behaves that way, given the structural nature of the paper and the physical properties of the water and so on.

But imagine the paper is folded and the water like dips into the crevice. Well, it seems like appealing to the shape of the paper, the terrain, the landscape, explains why the water pools where it does. Maybe not completely, maybe it's not an exhaustive explanation, maybe it's [01:28:00] compatible with a whole host of other kinds of explanations we could give, but by itself it seems to be doing some explanatory work.

And moreover, when you learn that the paper is curved, It seems like you understand why the water behaves the way that it does, because you're grasping an explanation. Or so these proponents of this view would say. So, I think that examples like that seem very intuitive. Our intuitive judgment is that it's appropriate to deem such things as explanatory.

But it runs into tension with these normative principles that mechanists are advocating for. 

Andrea Hiott: I mean, I also just think on a really, I mean, to get out of philosophy, out of this world of philosophy, which is we have to be very precise about all of this, but just thinking of dynamical systems in the world or something like complex, complexity theory or fractal, I mean, if you really look at the math of it, you can see that it's actually had explanatory.

The math has explained [01:29:00] things that have helped us create new mechanisms electronically. I mean, you can look at actual kind of devices, you know, that, that this is at least facilitated understanding enough to create some new mechanisms in the world. I'll just leave that to the side right now. But I mean, if you just even look at Stephen Wolfram and that whole kind of thing and the technology he's created out of it, I mean, there's, Obviously some understanding going on, but I'll leave that to the side right now because I think there's also another way in which you've maybe shown that this is mistaken to think of, um, dynamical models as non explanatory.

Um, I think there's even two more ways, and one has to do with this motif, right? Druskola, or I'll let you explain it, but what's another way we could maybe think about how Um, this might be wrong to judge dynamical models as non explanatory. 

Paul Kelly: Yeah. So in the recent sort of explosion of interest in machine learning systems and artificial neural networks, everybody talks about AI and chat [01:30:00] GPT and so on.

And the way that these systems work is very unclear. Interestingly, what's fascinating about them is that they are able to seemingly perform various kinds of tasks. Various inputs generate various outputs, but the way in which they're doing so is not clear. 

Andrea Hiott: Right. 

Paul Kelly: Related to our previous conversation or at least the conversation earlier, people refer to these things as models.

But there's also this other conversation trying to model what the heck is going on inside of this model. 

Andrea Hiott: Yeah. It's very meta. It's 

Paul Kelly: very unclear. And so one way that we try to understand these things, um, is through dynamical systems theory, where we try to make sense of the various state spaces that these systems are in, either physically or conceptually and then try to map the movements through them.

And what's really interesting is that there's been a recent development where if you take one of these systems And you train it to perform a particular kind of task. And so it gets quite good at it. And then you, at [01:31:00] a later time, deploy it to try to perform a related kind of task. It'll be much better at that related task than just an untrained network.

And intuitively that's not terribly surprising, but from a mechanistic perspective, it is surprising, right? Because the mechanistic details aren't known to us and also don't seem absolutely essential. to explaining why the particular thing is better than an untrained network. 

Speaker 3: And 

Paul Kelly: so there's been this development of this new idea of a dynamical motif.

And the idea is that when you train the network to perform the original kind of task, it constructs a certain kind of conceptual kind of state space. where things are represented prototypically or as exemplars or something that in many ways constitute its capacity. Why it's able to be effective at this input and output task.

And then when you deploy it on this new kind of task, it exploits or repurposes those kinds of dynamical properties. It exploits the various attractors or the [01:32:00] prototypes or so on that it's representing. 

Speaker 3: But 

Paul Kelly: notice what we've done. If we can see that as an explanation of its capacity, we're appealing to dynamical properties.

We're not appealing to nuts and bolts. We're not appealing to springs and levers and gears. And so such an explanation would once again fail 3M, but that seems implausible. 

Andrea Hiott: You don't have these discrete parts, or what I was getting at by the static way of of, of mapping. Um, and yet you still have what could be understood as, as a mechanism if, if it weren't so, um, if it weren't so, determined to be stuck to something like that mapping.

I don't know. But yeah, that's, that's, yeah. Yeah. 

Paul Kelly: So, so one important point perhaps to note is that it is possible that there's a mechanistic explanation for why the systems behave they do. But it would be so complicated. The kinds of, what it would constitute a component is entirely unclear. The kinds of interactions that are inside an artificial network.

It [01:33:00] just seems like a waste of 

Andrea Hiott: energy. I guess that's what I'm trying to get at is that. It's not that these are wrong. One is wrong and the other is right. Like I don't, it's a little hard for me to understand why we have to say it's wrong. It's just the dynamical approach might be better in this case for understanding, um, than a mechanistic.

Um, but if you really wanted to dig into some certain part of that, maybe a mechanistic mapping would actually be helpful too. But yeah, but that's not the only, I mean, there's also other, you know, Other areas of society and science, which are using models that I think in a way that that term would, you would agree with it is a model, um, are explanatory, but that would still be rejected by this notion of mechanism.

Isn't that right? Something like, I think you talk about equilibrium models or something like that. Could you? maybe illuminate that too a little bit, like that other possibility. 

Paul Kelly: Yeah, so in addition to there being a seeming tension between the [01:34:00] acceptance of 3M and these intuitive judgments we have, about what kinds of models are explanatory, like the, the bent paper example with the water, or actual scientific practice in the case of this posit of a dynamical motif explaining the behavior of these artificial neural networks.

There are also, if we extend 3M to areas beyond cognitive science, there are other kinds of models that also just get excluded from the category of providing an explanation. Um, one particular example which I think helps illustrate this is an extremely successful kind of model in economics and biology called an equilibrium model.

And an equilibrium model can be used to try to represent the dynamics or change of, say, a predator and prey relationship over time in a population, why they ebb and flow as they do, or in economics between supply and demand moving as they do relative to each other. There's a fascinating relationship between these various kinds of variables, and they seem to converge on [01:35:00] what are called equilibrium points.

the systems tend to sort of hone in on particular kinds of areas of the state space over and over again. And these equilibrium models posit equilibrium points, but these kinds of posits seem very similar, if not pretty much the same, as the kinds of posits that you get in dynamical systems theory about attractors and repellers and state spaces and so on.

And so I am suggesting that we not throw the baby out with the bathwater. That if We accept the 3M principle, and we take it to apply to all of science, it precludes these models, which are deemed explanatory. But why shouldn't we, as cognitive scientists, or as people investigating cognition or some mental process, try to employ all the various kinds of models, and why should we think that that kind of enterprise in cognitive science Is unique in that only mechanistic models are the, the permitted ones.

It seems like special pleading on the part of the [01:36:00] mechanists to say no. It may be the case that in economics and in biology you can appeal to equilibria points, but not here. Not here. It's not explanatory. Why? It seems like that's an arbitrary demarcation that needs a justification. 

Andrea Hiott: Yeah, I would agree.

Especially if we go zoom out to the why are we really doing this and what does it really, why does it really matter? And um. Yeah, it would seem that we would want something that we would want to be able to get to answers in many different ways. Um, not only one way or, and it seems like we already do that, right?

I mean, I think you say this, it's the things you've just described are very consistent with what we're doing in science and already, I mean, they're there. So if they don't fit so well with this very strict mechanistic. model, it doesn't seem we even have the choice whether we reject them or [01:37:00] not. It just seems we need to maybe open up, open up the discussion about, um, what that there could be more than one kind of model.

Or do you see it like that? Do you see this is a real choice? Do we have to choose between something like a dynamical model and a traditionally mechanistic? Do you really think that's a choice that has to be made?

Paul Kelly: Choice arises when it comes to models in different ways. So imagine you're a practicing scientist and you only have so much time and energy on your hand. What kind of model do you want to develop to investigate some phenomena? Maybe you want to develop a mechanistic model or maybe you want to develop a dynamical model.

Those would require different kinds of pursuits. Sure, but that's 

Andrea Hiott: not an objective choice. I mean, you're not saying one is wrong and the other is right. You're saying this one fits better here, which is wonderful. Very practical. Yeah. 

Paul Kelly: I also think that when it comes to this other issue of do we need to choose a particular kind of model as being genuinely explanatory, Perhaps 3M principle is setting up a [01:38:00] false dichotomy.

I think maybe it's helpful to import in a recent kind of development in another field that talks about models constantly, which is climate science. So the philosophy of climate science, which has become very important and is very significant, has converged generally on this idea that no particular climate model.

is by itself just true. Rather, you can't simply just evaluate models, climate models, or in many ways models across the board, just in terms of their truth value. Insofar as the model refers to the surrogate, like a diagram or an equation, it doesn't even really make sense to say that it is true. The surrogate, what is true of the surrogate is that it has a particular kind of relationship to the target in certain degrees and respects.

And so the general terminology from Wendy Parker, amongst others, is that we should evaluate models as being adequate for purpose. This adequacy for purpose account of model [01:39:00] evaluation. And so an individual model, insofar as our purpose is to simply describe the phenomena, a particular kind of model might be well suited for that purpose.

But if we want to predict the phenomena, maybe a different kind of model is what we want. Or if we want to explain the phenomena, maybe a different one. Or maybe, to use a term that's come up a couple times, if we want to understand the phenomena, which is a different kind of constraint on the kinds of mental abilities we have it requires a different kind of model.

In climate science in particular, the kinds of models that are being developed, in terms of software, simulations, and so on, are so complicated. No individual agent like you or me could represent what's going on in there. Even so, we might think that that model is tracking an explanation, but we might need to dumb it down and simplify it so that we can understand it.

Understanding requires a certain mental state being present in a particular agent relative to, say, some model or some phenomena or [01:40:00] surrogate. So, I think the conversation, in many ways, about explanation presupposed a bunch of things about the purposes of what models are for. That, what are models for? Oh, to explain.

And how should we explain them? Oh, mechanistically. But once you take a step back and realize there are different kinds of models and different kinds of ways you can use them for different ends, The conversation becomes much richer. 

Andrea Hiott: Absolutely, yeah. And I would say it's even urgent. I mean, I'm glad you brought up the climate science because there's this either or thing that we were talking about at the beginning, how these theories and models actually affect climate change.

And I think that's, that's everyday life in a, in a very real way too, once you zoom out and having to choose between something like a mechanistic model and a dynamic model might seem like we don't need to worry about it too much. But I do think, you know, it is connected a lot, right? Because we are using this in our science and it is informing just overall how we see the world.

And if we always think we have to choose between [01:41:00] these rather than what you just described, which model is best for the purpose. That's a very different orientation with which we're just coming to a problem. But I can also understand how something like the mechanistic viewpoint, how it's given us so much.

And if you're really, I mean, if you're in academia and you're really invested in that, and you've built your career on that, then there, you probably feel like that needs to continue in a certain way. So there's a lot of complicated stuff in there, but somebody like Wendy Parker, who I don't know, and I'll definitely look it up now, but would, would, would she, or, or, you know, Believe you have to choose between something like a mechanistic and a dynamic model.

I mean, I don't think 

Paul Kelly: so. I think, I think what, what Parker is going to say is that take climate science as a clear example. That's the primary thing that she works on. 

Andrea Hiott: Right. 

Paul Kelly: When we develop these kinds of models, we need to be clear about what phenomena we're talking about and also what particular thing we're trying to achieve.

Speaker 3: If 

Paul Kelly: we're only [01:42:00] interested in trying to have the model represent, um, just give us accurate predictions about sea level rise in a hundred years, That's the only thing that we want the model to give us a prediction of. 

Speaker 3: Yeah. 

Paul Kelly: It could be that the particular things we need to build into the model to get that accurate prediction might involve complicated causal stuff involving the greenhouse effect and so on, things like that.

But maybe not. It could be that there are other variables, proxies, which are caused by similar kinds of processes, which would generate accurate predictions. And if all we care about are the predictions, that might be good enough. If we want to try to explain why the sea level is rising right now, or why it will rise in the particular way that it will in 100 years, we might need to employ a different kind of model.

So in many ways, I think what Parker and people who are advocates of this adequacy for purpose view are going to say is, We need to be very specific about the kinds of questions we're asking, the kinds of aims that we want to achieve, and then also, given that, the kinds of models that [01:43:00] we select. 

Speaker 3: And the kinds 

Paul Kelly: of inferences that we make on the basis of our model should be very explicit and very carefully worded.

And we shouldn't presuppose that one particular kind of model or one particular kind of aim is the end all be all. In many ways, and I think this is an insight not just from Parker, but from. the modeling literature generally, it just doesn't make any sense to say that some model is just true to court, right?

Without any qualification, without any context. 

Speaker 3: Yeah. 

Paul Kelly: Models can be, say true things, they can imply true things about certain systems in terms of them being like them in particular degrees and respects, but the object itself, say a math equation, without any qualification, or the diagram, without any qualification isn't true.

It doesn't make any sense to say that it's true. It's not the kind of thing that has that property. 

Andrea Hiott: But again, it's back to that top territory map confusion, which I really think is at the root of a lot of misunderstanding that reverberates into very urgent problems, especially [01:44:00] thinking about something like climate change, where we think there's going to be only one answer.

We have this model of There's one model that we have to choose, like we have to choose the right model and we have to, there's not a kind of plurality or a way of, of understanding that there's many different positions, that there's many different ways to get to an answer, that there's many different trajectories.

I mean, that's what I'm trying to do in my work too, is find out a way, how can we think about positions from different scales and how, what I hear you saying when you, or with Wendy Parker too, I'm definitely going to read that as. You have to understand what landscape you're working in, what's the position from which you're measuring, what's the goal, what are you trying to figure out?

I mean, all of that matters for the model you're going to use, but also the answer that you get. And you don't just stop there. I mean, there should, that's what science is. We're trying to understand from many different positions in the most rigorous And I guess I could be a little more biased towards something like [01:45:00] dynamical systems because the math or, you know, like I mentioned category theory or something, or it's at least more in that orientation where you can start to understand mathematically or whatever, on whatever level, how to hold many different models together.

as plausible, um, without needing to choose them, but rather choose them according to what landscape you're in and where you want to go and, and, and so forth. So I don't know, does that, to bring it back to, to your work, I mean, is that connected at all to this unificationist account? Is that completely something different?

Would that also jive with this theoretical hypothesis approach that you're coming at where you need both? The model, the map and the context or the theory. Um, I mean, are you, does that, does that resonate with, with where you're going to with your research too? 

Paul Kelly: Yeah. I mean, I think it does.

I think a important topic, [01:46:00] which I think people are just starting to really dedicate a lot of effort to try to figure out is there's this common scientific practice referred to as integrating models. where you have different models, perhaps of the same phenomena, and then you try to integrate them into some general model.

Another possibility is that you have multiple models of the same phenomena, which make slightly different kinds of claims when you interpret them. Wendy Parker has focused on that, because when it comes to climate models, There's not just one, right? There's a variety of them. And then there's this practical question of how to either integrate them or take the judgments or the verdicts from these models, the inferences we're making on the basis of them, and make, square them with one another.

Insofar as one model makes a prediction that's in tension with another, which one should we give more credence to? Is there some sort of normative discussion to be had about weighing them off one another? Um, that can be a really complicated one. And it's also [01:47:00] exacerbated by, as you mentioned, the importance of the discussion, right?

These are significant, they're significant practical upshots, depending on whether we think climate will change in a particular way, in a particular time period, and communicating that to the public is really difficult. With that said, though, Sometimes, at first blush, it might appear as if two models are in tension with one another, but if you dig deeper, they're not.

And so, if we think about these models in terms of the way that I'm advocating for, and I think it's consistent with the modeling literature in general, is that these things, in terms of their metaphysics, are just surrogates, right? We take the diagram or the equation to be a stand in for something, and then, here's the crucial point, we, as modelers, interpret.

those surrogates as indicating certain propositions being true. We engage in an inference, and we say because the equation moves this way, the phenomena will behave this way, and so on. That kind of inference is often very not, it's not explicit [01:48:00] in the science. People just take themselves to sort of read off from the model, some fact about the phenomena, but it's actually an interpretation.

It's actually an inference from this surrogate. And once you realize that, and you make explicit what the surrogate has built into it, it could be that the propositions you get out the other end aren't incompatible. Maybe they're making different kinds of claims. about the phenomena and can be integrated in a particular way.

Andrea Hiott: Yeah, and there's often you can find that there are patterns overlapping, but once you have a way of being very clear and specific about what regularities were studied, what went into it, um, you know, in, in the way that you were describing with Parker, or really, you know, What's the question? What's the landscape?

What's, what are we actually dealing with? Instead of everyone just assuming, as we often do, that we're all coming from the same space. And then, of course, a model looks, or the findings of, from another model might look very much in contrast with ours because we're taking our landscape as the landscape from which that model was done, [01:49:00] which is almost always not the case.

So it's more, I think that's where the becomes helpful in something like a lot of the math that's associated with something like complexity and dynamical systems because you do start to get ways of being able to, in the way that we started this discussion with, um, put something, represent something externally, model it in a sense that can be understood better, that you can start, you know, you can start to understand that there are many different approaches getting to the same spot.

That is a very hard thing, I think, for all of us to understand as scientists and philosophers, but also. everyday life, right? But it also feels very crucial, in a sense, if we're going to deal with some of these urgent challenges that we face right now. So, yeah, I don't know. Um, I mean, we've been talking now for a couple hours, but I'd like to hear, you know, what your thoughts are on that.

And also just this, where are you going? What's, um, you've shown that, [01:50:00] I think with your work, you're, you're showing that this criticism of the non explanatory Of the dynamical model by the mechanic perspective has can be challenged clearly. Um, but then what's the step after, like, how does it all fit together?

I don't know if it has to do with this unification account or, or what, but I wonder like how you're seeing that in your work right now. I feel like it's probably connected to this Parker and everything we're talking about, but maybe we can clarify it a little bit. 

Paul Kelly: Yeah, I mean, I think it is, it is at least if if my current thoughts about it are on track for anything, they are all connected, that this question of, of what models are, the different kinds of aims that we can deploy them for and also this unificationist account that you alluded to and also these conversations about the relationship between explanation and understanding.

I think that they're all sort of intertwined in a really complicated way. Okay. And I think that these different kinds of literatures, like the literature and the [01:51:00] philosophy of climate science, and modeling there, and this literature about mechanistic explanation, and the normative discussion about what principles we should accept about that, can really fruitfully, you know, be put into conversation with each other.

Um, I think maybe a way of thinking about it from the way that I'm sort of coming at it is, In actual scientific practice, there's this phenomena that's been identified, referred to as model transfer. And this is where one particular discipline I don't know, say it's physics, you're developing the ideal gas law, for example, realizes there are particular patterns that accurately describe the behavior of some kind of system.

And then there are other people working in very different kinds of disciplines, say, I don't know, population genetics or something. And they look at the ideal gas law and they look at the populations and they go, wait a minute. It looks like this could helpfully describe, I'm using that word specifically, describe the behavior of this various population and the dynamics.

So the [01:52:00] model, in many ways, that's a mis, incorrect way of speaking, the surrogate, right, the equation, is now being deployed to have a new target. It's representing something different. Now, when that's done, that's really interesting in terms of, scientific description, right? Maybe this new model transferring over can describe the behavior of this population.

But it seems like there's another deeper question, which is, can it explain? If it explains in the case of the ideal gas law and the behavior of these various gases, does it also explain when it's deployed over here? What needs to be the case for it to be not just a mere description of the population, but actually accurately tracking some explanation and maybe providing some understanding as well.

And it seems to me the resources from dynamical systems theory can perhaps give us a sense of, to use a fancy philosopher's phrase, in virtue of what 

Speaker 3: these 

Paul Kelly: models are explanatory. 

Speaker 3: Yeah. 

Paul Kelly: And they're explanatory in virtue of [01:53:00] or because they're tracking various dynamical states and processes that are occurring within, you know, the underlying structure of the phenomena.

The dynamics of the ideal of the gas. and the dynamics of the population might have similar kinds of state spaces, might have similar attraction points and similar trajectories. And if that's true, then identifying that through this active model transfer isn't just merely descriptive. It could be explanatory and might provide genuine understanding.

So I think that when that's done, the kind of explanation that you obtain we have to think about. There are different accounts of what an explanation is fundamentally. The mechanists are in many ways committed to this ontic conception of explanation. There are other accounts though. There's the pragmatic account of explanation, which views explanations as answers to why questions.

There's also other accounts, like the unificationist account from [01:54:00] Kitcher amongst others. And the sort of hallmark of the unificationist account is that what it means to explain, at least often, is that you're coming to recognize. that the particular instance is an instance of some more general pattern or structure.

And the way that Kitcher frames it is in terms of argument forms. You come to recognize that a similar argument or a similar principle can be deployed within different contexts to do explanatory work. And perhaps I'm thinking that could be a way of making sense of this idea of model transfer and dynamical systems.

Because what you're doing when you transfer the ideal gas law into a biological context is you're recognizing and unifying, in some ways, this pattern, 

Speaker 3: right? Yeah. 

Paul Kelly: And the hope is that recognizing this pattern isn't just purely superficial. It's not just descriptively the case. Dynamical systems theory might give us reason to think that it's actually representing not just the [01:55:00] superficial behavior, but actual aspects of the processes underneath that are explanatory if they exist.

Andrea Hiott: Yeah, I, I think that's a very well said, and I I'm really on board with that because I mean, in my work and actually with a lot of people I've talked to, um, as part of the research, we end up coming back to trying to figure out, trying to focus more on patterns instead of parts. I think though we needed to focus on parts.

Even in the kind of this literature in the nineties and early two thousands of mechanism, it was important to focus on the parts, observe them, notice they're there. That's not none of that needs to be said. That's not bad. There's no, you know, we learned a lot from that, but now we're trying to understand more patterns and we have more tools and technologies like with dynamical systems and maths to start to look at patterns in a different way.

And And make it a little more rigorous in terms of, um, how we can understand across disciplines, [01:56:00] as you were saying what those patterns can show us about how to be healthier, how to address certain urgencies, those kinds of things that really do matter. And that, you know, we are actually. Trying to understand with all this.

Um, but also just on an everyday sense, when you were talking, I'm thinking that's how it works when you just connect with other people or when you, um, have a kind of a good day or something, what you're really doing is, you know, connecting certain patterns. So I think there's a lot of different scales of why this matters and, and, and why it's important.

Um, and I'm glad you're, you're working on it in the way that you are. What's what's the main challenge or obstacle just to end that you feel like you have to deal with? I mean, we've talked a lot about the mechanistic and the dynamic, um, butting heads or something. Is that, does, is that really fair to say that's the main issue that you're trying to reconcile?

Because I, I looked briefly at your prospectus for your PhD and you've actually, it seems a lot broader than [01:57:00] that. So I guess like just to end maybe. Yeah, what's, what's, what's like the bigger state space or, or landscape that you're looking at now and trying to what's driving you, what's driving the research right now?

Paul Kelly: Yeah, I think, I think you're right. I mean, the questions I'm interested in, I think, could be expressed in a much broader way. And the, the debate or the disagreement between people about mechanistic models or dynamical models and if they're explanatory, and in many ways it's just a, a case study. It's just a particular instance that illustrates the complicated kinds of philosophical assumptions that are at play in these kinds of assessments.

The very general kinds of questions that I'm interested in are questions about, you know, models and what they are and how they represent anything. How is that possible? And to connect to the beginning of our conversation, how is this notion of representation the same or different or related to if it is [01:58:00] other notions of representation, right?

You've alluded to the idea of maps throughout this discussion. 

There's also there's paintings. There's also beliefs and desires and so on. There are arguments. There's all these different things that we think represent other things. Is model based representation its own kind of thing, and does it require some sort of unique kind of analysis, perhaps different from these others?

Um, and then relatedly, I want to know what it means for a model to explain something. There's lots of different conceptions of what explanation is. There's the ontic explanation, the pragmatic one, the unificationist, and so on. And fundamentally, I think there's a philosophical kind of urge to try to understand the world.

And we think that science is in many ways a good or maybe even the best way of going about trying to obtain an understanding of the world. Models seem fundamental to that. And so trying to figure out what's going on is, is really [01:59:00] essential. Um, there's a whole host of different assumptions that are perhaps at play in this.

You alluded to the idea, which I think is really well put between patterns and parts. I hadn't thought about that before that. Maybe really what's lurking in the bottom here in terms of the disagreement that we've been focusing on, is that the mechanists think that parts can do explanatory work. 

Speaker 3: Yeah.

Paul Kelly: And the patterns that are picked out by these other approaches aren't doing the kind of work that we want it to. Maybe patterns can do explanatory work, and maybe, this is a very speculative philosophical view, maybe parts in many ways Fundamentally are just patterns. 

Andrea Hiott: Yeah, I think, I think so. I mean, you know, that goes back to physics, right.

With the particle and wave and you can study them both ways and both ways are good, but yeah. Choosing between them is not possible, I guess. So I just have to ask. 

Paul Kelly: Go ahead. I was just, so the big general question that you asked is what am I sort of struggling with now [02:00:00] or what am I sort of looking forward to in terms of this?

Um, I hope that I can sort of. move this discussion forward in a particular way. And also, hopefully, I think it's very abstract right now, trying to make it more practical. I think contemporary discussions about the importance of AI and normative questions about these machine learning models and the various assumptions that are built into them and how they're deployed.

There's good philosophical work to be done there because people are taking these things up and deploying them for particular purposes that have real consequential practical upshots in terms of climate science, but also in terms of using these kinds of things. to try to assess people's rates of recidivism, and then determining sentences for their prison sentences on the basis of the output of a model, without any real reflection about what these models are, the various assumptions involved with them, or if they [02:01:00] should be performing this kind of role that they're taking them to do.

So, I hope to sort of move this conversation, not just in the abstract about how we should investigate cognition and so on, But also have a sort of robust philosophically informed discussion about how models are playing a significant role in our everyday life and will continue to do so going forward.

Andrea Hiott: Wonderful. I think that's very important because it is one of those things that gets accepted and used even in courts and so on without actually being well understood and part of zooming in and really understanding that is. What we need to do to clarify these problems, so I wish you a lot of luck with that.

I have three just really quick questions. Um, what's the difference between a surrogate and a model? I mean, you don't have to go into a big explanation, but you seem to not use them interchangeably. But I have trouble not seeing them as I mean, a surrogate is standing in for something, right? Which isn't Like a model could be applied to many different things, or I don't know what's [02:02:00] maybe you can just clarify that.

Cause it's, it's bothering me a little bit. 

Paul Kelly: Yes. It's a, it's a great question. Um, the term model is the thing that is just sort of ubiquitous that people use across the board. And as we said earlier, there's often not really general consensus about what it's supposed to be referring to. The literature on what models are, it's generally converged on this idea that the defining features that models are used with surrogative reasoning.

And so surrogates are the things that stand in. Are the surrogates the model, or is the model something more than that? That's controversial. People who program models or develop models will often refer to the equation or the software or the object as the model. But there's a sense in which, if we're speaking precisely, and you're literally just referring to the object, Then maybe you should say the surrogate.

What additional things need to be present for it to be a model? Well, there's different proposals. You alluded to this idea that maybe there needs to be a theoretical [02:03:00] hypothesis. in addition. So take, for example, a set of styrofoam balls that are arranged in the arrangement of the planets in our solar system just by itself.

Is that a model? Well, it's just a set of physical objects that have certain kinds of relations to the planets in the solar system. That relation obtaining by itself doesn't make it a representation, right? Because the spatial relationship between the various balls. and the planets also obtains between the planets and the balls.

But we don't say that the planets are a model of the styrofoam balls. So that relationship is symmetric. The representational relationship is asymmetric. The thing is a representation of something else. 

So 

what does that additional work to make it about something? Ronald Geary has proposed a few years ago, um, that it's a theoretical hypothesis.

So it's, it's a proposition that someone asserts. They say [02:04:00] that set of styrofoam balls is like those planets in this particular way. And once you can join the theoretical hypothesis with the actual reference, the surrogate, then it becomes a model. So a model has two components. It's the thing, the surrogate.

But also this suggestion, this proposal, 

Speaker 3: this 

Paul Kelly: hypothesis that it has this relation that obtains between it and its target. So 

Andrea Hiott: you can't pretend those are detached, I guess. Um, I mean, it, it also goes back to this Parker thing too, where you, you need to, or what we were talking about, where, you know, what's the, what's the purpose, what's the landscape.

All of that matters actually for even something like how you're going to, you know, Notice the relationship between the styrofoam balls and the planets. So it's weird that we would just dismiss all of that and not think of it, you know, it, it is not detached. So 

Paul Kelly: it's not detached at all. It's a, it's a, on this view, the theoretical hypothesis is a constituent.

It's a part of what makes the thing a model. And to see that, recognize the fact that the styrofoam balls. [02:05:00] If I attach a different theoretical hypothesis to it, would just be about something different. I can say this represents the hydrogen atom. 

Andrea Hiott: Absolutely. 

Paul Kelly: It's the same surrogate, but the target is different because of the hypothesis being different.

Andrea Hiott: Right. Which is why you need that, that space, that landscape, that embeddedness as part of it. And also I think that connects ontic and epistemic kind of, we can go into it a lot here, but that too, you would need, I mean, it's a, it's a bridge toward understanding that differently, but here's the second because we already talked a long time.

Um, do you think language is something like a model? I mean, how do you, I know that's a huge question, but just, I can't help but wonder, you know, cause we, we, you, when we were talking about the watch, you were saying like, Oh, I could just tell you this. And that's a kind of a model. So is language itself a model?

Or how do you see that? 

Paul Kelly: Depends what you mean by language. I think if what you mean language, I guess. Like an alphabet. A language? Yeah, 

Andrea Hiott: yeah. Like English or Spanish or [02:06:00] Because you've externalized, right? It's in a way, it's what we've been talking about, where you've created this thing that's symbolic, that then becomes a kind of, um, connecting space to explore and 

Paul Kelly: I think in general, I'm going to say no.

And here's, here's the reason why. If the defining feature of model representations is surrogative reasoning being deployed on them, where you investigate the surrogate and the structure of the surrogate or properties about the surrogate are used to infer. properties about the referent? That's not the case with language.

Take the word atom, A T O M, right? And writing it on a piece of paper. And now imagine zooming in and analyzing the letter A and the letter T, right? And trying to manipulate the letter T in a particular way. It's not the same kind of thing. I'm not using it as a stand in for the thing that I take it to be a representation of.

So, [02:07:00] linguistic reference. Linguistic representation, or when you combine various linguistic components into, say, a proposition. Like a sentence, propositional representation, which could be beliefs or desires or something. That's just a very different kind of representation than the external model. That is say a physical object or an equation that I'm using as a stand in that I investigate itself.

So the kinds of reasoning that's used when we deploy models seems fundamentally different in many ways from the kind of reasoning that we deploy when we use language. With that said, language is often used to express what are called arguments. chains of reasoning that have a particular relationship, premises to conclusions.

And a lot of people think that maybe when we reason with models, even though it's not explicit, we're actually deploying and appreciating certain arguments. So maybe there is a kind of intimate relationship between language and model based reasoning, but at least at a superficial characterization, they [02:08:00] seem very different.

Andrea Hiott: Yeah, I guess the language seems representational. I mean, I think we could agree that language is representational. And that's probably more what I was trying to say, I guess, about that common space that we, it's represented. And so it's like a sort of third thing that we can then agree upon. But then maybe something like an argument could be understood as a model, which, I mean, I don't 

Paul Kelly: Perhaps.

So if the defining feature is that When we have a model at play with a surrogate, we zoom in and investigate the particular dynamics of the surrogate. With a sentence, it's not clear that we do that. Sentences are clearly representational words are clearly representational, but if they're representational in the same way that models are, it's not clear.

Arguments are an interesting case, because we do zoom in and focus on the structure in a particular way, and we maybe manipulate the argument, move the pieces around. 

Mm hmm. 

So it's closer to the kind of reasoning that takes place with the model, but it's not clear that it's at least a [02:09:00] paradigmatic instance.

Perhaps it's a limiting case. 

Andrea Hiott: Very interesting. And the last thing that came to mind that I just want to ask about is something like a book or a movie or something like this, where, um, again, you, we have this kind of, I mean, I'm thinking of like scientific papers and we're, we've been talking a lot about how, when we create models and we create theories and so on, it's, it's changing and it's important, but so too, when we create something like a movie or a TV series or a book and I just, I, I'm not sure how to, if those connect at all, or if you've thought about that in any way, Where is the representational or the modeling, um, going on?

It seems like something like a dynamical model would be able to look at something like that in a different way than a mechanistic model, which is what made me think of it. But yeah, it's chaotic what I just said, but I'm just throwing it out there. Have you, have you thought about, have you thought about anything like that?

Paul Kelly: Yeah, it's a, it's a really interesting question. So there's lots of different kinds of representations. [02:10:00] Um, and trying to evaluate in virtue of what they are representations is really complicated. It's sort of, I think, a really interesting point to note that the modeling literature and the recent development of it.

Often, initially emerged out of conversations that came from aesthetics, 

Speaker 3: because 

Paul Kelly: in aesthetics people were saying, well, what does it mean for something to be a representation, right? What does it mean for the Mona Lisa to be about Mona Lisa? Oh yeah, definitely. And so there's this really interesting conversation to be had about aesthetic objects that we construct.

So imagine the book that you have in mind is a piece of fiction. Right? Or imagine the movie is, I don't know, a sequence, I don't want to, I don't know particular views about the metaphysics of movies and what they fundamentally are, if it's the, the film and the projector, or if it's the images projected on the screen or something.

But imagine a world where you have those physical objects, but there are no more minds anymore. Something cataclysmic happens and there are no more minds, but there are these physical objects in the world. Are they still [02:11:00] about anything? It's not clear, right? They're definitely various objects. They definitely might have certain kinds of relations that obtain between them and things that did or didn't exist at a particular time.

But for something to be about something else and to be a representation, there needs to be, it would seem, an extremely important sort of feedback loop between the thing that is recognizing it. As a representation or the thing that's asking questions about it that makes it a representation. Mm-Hmm. . 

Speaker 3: And 

Paul Kelly: so in this debate about models, there's a disagreement between what are called inferential and representational list, where the representation list, think the representations are out there and then we recognize them and we reason with them.

Where the inferential says, well, what makes something a representation is that there's beings like you or me. Who are asking certain kinds of questions. And that process makes it a representation. So if we imagine a world with no minds on the inferentialist view, there are no representations, there are no models.

You could [02:12:00] have scribblings on pieces of paper, or you could have various diagrams and equations written down. But if no agents are there to recognize it as such and reason with them, they're not about anything. 

Speaker 3: Um, 

Paul Kelly: so it's a complicated question and it cuts to the core of, you know, What it means for anything to be about anything else.

Andrea Hiott: It's fascinating. And I actually think, um, it could be addressed in a similar way to how we dealt with this dichotomy of mechanistic and, and dynamic and ended up getting to patterns. I think you could get. Look at something like the, that dichotomy you just laid out too, in terms of patterns and, and, and the Susan Parker view is what I'm calling it now, even though I've never read her.

But, um, in, in asking these questions of actually what are we measuring and what are we thinking about? And, and then you always have some agent or observer because. That's the space that you're in. So yeah, maybe, um, the pattern patterns instead of parts makes, makes more sense there too, but that would have to be a whole [02:13:00] other discussion for now.

I would just say, thank you very much for this. And you've given me a lot to think about and, um, it's very rich what you're doing and I wish you a lot of luck with it. I think it's important. 

Paul Kelly: Oh, well, thank you very much. Thank you for having me on. I enjoyed, enjoyed the conversation and had me think about things in ways that I hadn't previously.

So I appreciate it. 

Andrea Hiott: Well, good. I'm glad. All right. Well, have a nice day there in Wisconsin. It's night here. So I got to go have dinner. 

Paul Kelly: Have a good night. 

Andrea Hiott: All right. Bye.