You observe a phenomenon, and come up with an explanation for it. That’s true for scientists, but also for literally every person. (Why won’t my car start? I bet it’s out of gas.) But there are literally an infinite number of possible explanations for every phenomenon we observe. How do we invent ones we think are promising, and then decide between them once invented? Simon DeDeo (in collaboration with Zachary Wojtowicz) has proposed a way to connect explanatory values (“simplicity,” “fitting the data,” etc) to specific mathematical expressions in Bayesian reasoning. We talk about what makes explanations good, and how they can get out of control, leading to conspiracy theories or general crackpottery, from QAnon to flat earthers.
Support Mindscape on Patreon.
Simon DeDeo received his Ph.D. in astrophysics from Princeton University. He is currently an Assistant Professor in Social and Decision Sciences at Carnegie Mellon University, and External Professor at the Santa Fe Institute.
- Web site
- Carnegie Mellon web page
- “From Probability to Consilience: How Explanatory Values Implement Bayesian Reasoning,” Wojtowicz and DeDeo
- Axiom of Chance blog
- Google Scholar publications
[accordion clicktoclose=”true”][accordion-item tag=”p” state=closed title=”Click to Show Episode Transcript”]Click above to close.
0:00:00.1 Sean Carroll: Hello, everyone. Welcome to the Mindscape Podcast. I’m your host, Sean Carroll. Obviously, on a podcast like mine, we’re often going to be talking to scientists of various sorts, theoretical scientists in particular, some of my favorites, not all of them. But a lot of times, we’re talking about the ideas for how we might explain some scientific phenomenon, coming up with a new theory, whether it’s dark matter or evolutionary biology or whatever it’s going to be.
0:00:26.4 SC: You might remember, in fact, that I had Lee Smolin on the podcast very recently. And despite the fact that we work in very similar areas and we’re personally very friendly, we have different ideas about how to go about building the better next generation scientific theories. Why is that? How can two scientists who are both more or less trained in the same way come up with very different preferences when it comes to building new explanations?
0:00:51.7 SC: That’s what we’re on about today, on today’s podcast, with Simon DeDeo. Simon actually started as a theoretical cosmologist, much like myself, but he switched into some combination of statistics and data-driven social science and cognitive science. So it’s a wonderfully difficult specialty to pin down. But he’s both at Carnegie Mellon University and also the Santa Fe Institute. So we overlap a lot in our intellectual interests. And what Simon talks about in the paper that we’re going to be discussing today is how explanations work. And honestly, explanations in this sense is kind of a synonym for a theoretical viewpoint or formalism to answer some kind of question. So you have some phenomenon, whether it’s, “My car broke down,” or, “There is dark matter in the universe,” and you want to explain it.
0:01:39.0 SC: Now, what happens is you can invent an explanation, and different people will prefer different kinds of explanations for different reasons. So what Simon and his collaborators have done is to break down the different parts of Bayesian analysis that go into making a good explanation and sort of quantify different preferences or different values you may have for choosing your personal preferred kind of explanation. Explanations have different kinds of good aspects. An explanation can be simple, it can be powerful, it can be close to things we already understand, it can explain many things at once. These are all good things, but sometimes they fight against each other. Sometimes an explanation can be simple, but not powerful. It’s simple for only one phenomenon, or it’s both simple and powerful, but utterly different than anything we know.
0:02:28.6 SC: So how to balance these is kind of a human subjective thing. And we talk about both how scientists actually do this when you have legitimate scientific disagreements: The many-worlds interpretation of quantum mechanics versus the Copenhagen interpretation or pilot-wave theories or something like that. What are the different values that the practitioners have that allow them to prefer in an intellectually respectable way these different explanations in a situation where we don’t know which one is right yet? And what’s fascinating about Simon’s analysis is it goes beyond the case where everything’s working. I can say that I like many-worlds as my favorite theory, but I have absolute respect for people who don’t if they have principled reasons for preferring some other things.
0:03:16.0 SC: Sometimes people pick wrong explanations because they’re failing at balancing these different kinds of values. So there’s sort of a continuum between high-level scientific discourse about unknown theories and complete crazy talk, conspiracy theories. Why do people believe QAnon or that school shootings are false flag operations or something like that? Well, you can understand that in terms of them putting all of their eggs in one explanatory basket. All of their values are concentrated on comprehensiveness rather than simplicity or something like that. So it’s not just that there are sensible people and crackpots. There’s a continuum of ways that we can try to explain the world, and you can try to analyze the similarities and differences between conspiracy theorists and the world’s best theoretical physicists. It gives you a lot of food for thought about how we go about explaining things in our everyday lives. So with that, let’s go.
[music]
0:04:26.7 SC: Simon DeDeo, welcome to the Mindscape Podcast.
0:04:28.3 Simon DeDeo: Hi, Sean.
0:04:29.2 SC: So we’ve had a lot of people, at least a few people, talking about misinformation, disinformation, conspiracy theories that people fall in love with, things like that. And I think that what we’ll be talking about today, you’ll correct me if I’m not getting this exactly right, but rather than focusing on the huge wrongness of believing in a conspiracy theory or a crackpot scheme, we’re thinking about explanations more broadly and saying that there are people who believe in conspiracy theories or crackpot schemes, but their kind of belief, the way they justify it, etcetera, is of a piece with perfectly respectable scientific beliefs. We should sort of… It’s a matter of balancing things here and there rather than a matter of a completely different way of thinking. Is that fair?
0:05:12.6 SD: Yeah. We’ve been working on these questions for a while, and one of the inspirations for me at least was a couple of years ago now, David Deutsch wrote a book called The Beginning of Infinity, and it’s a lovely book. It’s nuts, but it’s lovely. And Deutsch really focuses on something that I had thought about, but not as clearly for many years, which was this idea that you explain things, and that explanation is not simply something we do to feel better about the world, but it’s actually this enormously generative process. We had this problem, there was this puzzle for me when I first started working outside of physics because people in biology want to predict things or at least the outside world wants to predict things. Biologists themselves don’t usually. When we work on social behaviour, people want to say, “Can you predict the next war,” or whatever. And so Deutsch really crystallized for me the way in which that’s not actually what this game is about. And not only is it not what this game is about, it’s like that’s okay.
0:06:29.6 SD: So the fact that we can’t predict things, especially not about the future, is not a sign that we’re doing something wrong, but a sign that we’re doing something different. So Deutsch, that was a sort of way in. And then that kind of percolated in my head for many years. When I got to CMU, I started working with a graduate student, Zac Wojtowicz, and Zac I played around with a lot of different ideas, and we started to think about, could we build a theory of explanation making. Actually, there was another impetus for this, which is a logician at CMU, Devi Borg, and so Devi and I talked for a while about… There’s this joke, right? Why do mathematicians try to prove things they already know are true? Right, this is the worst.
0:07:20.5 SD: Science is about producing experiments with uncertain outcomes, so what’s the goal of a mathematician. No-one’s shocked Fermat’s last theorem has no solution, right. And so Devi and talked about this, and just to dig a little into this, in science people have, let’s say, tried to even formalize what… The next experiment you should do is, right, so this is called optimal information design or it goes many different terms. So you can say, should we build the next Hubble Space Telescope? Or should we build the LHC? Which one is going to give us more information and so you pick the one that will give you more information.
0:08:02.5 SD: The one thing you shouldn’t do is do an experiment where you are guaranteed to know the outcome. So then we talk, Devi and I talk, and you say, why does a mathematician care about proving something they already know is true? And Devi’s answer is something like P versus NP, these really challenging problems. One of the intuitions is that the solution is going to discover unexpected connections between very different branches of mathematics, so it’ll give us a new view onto things we thought we understood because of the way in which these different chunks have to be brought together.
0:08:40.0 SC: This is very relevant to my interests, as they say on the internet, because of course in quantum mechanics, we have a situation where generations of physicists have been brainwashed into thinking that we don’t need to understand what’s happening. All we need to do is make the predictions for the outcomes of observations. And the counter-example I try to say is, look, what if you had a black box into which you could ask literally any outcome of a specific physical situation. Does that count as solving all of physics? Anything you want to know, but you have to re-ask it, it doesn’t tell you what the theory is, it just tells you when you collide these two billiard balls, they scatter off at this particular angle. I don’t think that counts as a good theory of physics, we want the theory for reasons beyond that.
0:09:24.7 SD: Right. Right, right, right, right. So if we’re building a theory of how to, exactly how to shake the box, we have a goal in shaking the box in a certain way, and then the question, yeah, what’s the goal? How do we know when we’ve reached that goal? How do we know if we’re on the right track? And so someone like Deutsch articulates this fact that if what we care about is explanation, then Zac and I dig in and we say, alright, let’s try to build… And this is a joke, right, let’s try to build a unified theory of explanation-making, and this then, this then rolls into a year of enormous fun trying to make sense of the ways in which… We have a whole bunch of different experiments that we do, for example, lab experiments on people in a psych lab where we say… And this is not my work.
0:10:19.9 SD: Tania Lombrozo’s one of the big figures in this, so you might tell somebody, okay, look, there’s this alien, these aliens, and they fall victim to these kinds of diseases, and this disease gives these symptoms and that disease give that symptom, and this alien comes into your consulting room and he has spots and a cough, and diagnose this alien. And you can look at the different choices people make. In a sense a diagnosis by proxy is an explanation for the symptoms that the person brings in, or the alien brings in. And so by looking at the choices they make and attributing, let’s say, a pair of diseases or a single disease to the alien, you learn a lot about the kind of preferences we have, the ways in which we select explanations one from another.
0:11:08.0 SC: And let me, just to be super duper clear, we’re using the word explanation here almost in the sense of a theory or a model. It’s not like someone says, well, explain black holes to me, and we think we know what black holes are, but there’s this pedagogical attempt to explain them in a comprehensive way. You’re using the word explanation to really mean knowledge of what is going on behind the scenes that could help us and then we can sort of choose between competing explanations that can’t all be right.
0:11:35.9 SD: Right, so this is great, right? So our model is a Bayesian model, meaning we ask questions about the things we have to hand, say the facts in the world, and we ask the ways in which they match the models we have of the world, and so it’s a very specific kind of model, it’s a generative model, which says it gives you degrees of belief in the things that could happen. So in that sense we’ve really boiled it down to a paradigm that in one sense doesn’t really match any kind of science we do, like condensed matter physics is not a generative model in the sense that it specifies the probability of different things happening.
0:12:16.4 SC: Sorry, what is the definition of a generative model?
0:12:19.5 SD: Right, so a generative model tells you not just the things that might happen or the things that could happen, but also gives you degrees of likelihood for the different things that could happen. It’s a way… You can think of it as a mini simulation that you have controls over, so you can see inside the simulator, and that simulator, you push a button and it projects a possible world onto the screen. When you ask, Sean, what is an explanation more broadly, this gets into a big question that let’s say we can split into two parts.
0:13:01.4 SD: An explanation has to do with empirical facts, it’s a way of accounting for stuff that has happened or stuff you know. But it’s also an account of how that stuff could have happened, and generally, it’s an account of other things that might happen, other things you could look for, other things that could happen in the future or are happening somewhere else. So there’s two pieces when we think about an explanation. One piece, when they evaluate one, one piece is, you asked me to explain this, right, can you? Meaning, can you tell me a story in some way in which the thing that happened comes true.
0:13:42.7 SC: Yep, sure.
0:13:43.7 SD: Right, so that’s kind of this empirical piece, right, and then there’s another piece that’s sitting there, which is this much troubling piece is, how do I feel about the larger structure of that explanation, including the things it might do for things that could happen in the future, let’s say, things that I haven’t observed yet, also things like what does the explanation look like, how succinct is it, how concise is it, versus how elaborate it is. So when we refer to things like Occam’s razor, what we’re talking about is that second chunk of the ways in which we might value an explanation.
0:14:20.1 SC: And in some, this is going to be very vague and then you’ll fix it, but one of the nice things about your way of talking about explanations, is it is Bayesian, like you said, and we talk about Bayesian stuff on the podcast all the time, so the folks have heard the term, but all we really need is the idea that there’s a prior, there’s some pre-existing probability or credence that the model is right, and then you update because information comes in. And what you do in some sense is sort of decompose those two parts, the prior part and the likelihood part, into different values that an explanation can have. Is that fair?
0:14:58.0 SD: Right, yeah. So we put explanations into the super collider, we spin them really fast, and it turns out that what we thought were two things are actually, let’s say, at least four things.
0:15:12.2 SC: Tell us what the four things are.
0:15:16.4 SD: Sorry?
0:15:16.4 SC: Tell us those four things.
0:15:17.3 SD: I will tell you… The human curiosity is infinite, right? And of course, what does it mean to diagnose or just to discover these sub-pieces of a mathematical equation, what we’re really doing is saying, this piece picks up on a psychological value, so in fact, our minds are sensitive, let’s say, not just to the prior and the posterior, the things you knew beforehand, and how well you do on the next bit, but also at these little pieces, right. So, for example, in the case of the empirical side of the evaluation of explanations, we split into two pieces, one is descriptiveness, we call it. Descriptiveness is just how likely is the stuff that happened, taken individually, given your explanation. So I’ve got a whole bunch of things that happened I want you to explain.
0:16:14.1 SD: You give me an explanation. If I pick, let’s say, any piece of what I’ve seen at random, how well is the explanation doing…
0:16:21.8 SC: As opposed to… Sorry, let me back up, because we’re trying to convert into words something that is mathematically very clear if we could see the equation in front of us, which we can’t. So in Bayes’ theorem, there is this idea of the likelihood, which is to say the probability that you would be getting that evidence if this particular explanation were the right one. And I think what I hear you saying is, yes, that’s a thing, but we’re going to divide that up, we’re going to divide this probability of getting the evidence, given the explanation, into first the probability of getting this particular piece of data… Is that right?
0:16:56.7 SD: Mm-hmm. That’s right, yeah. So I have a story about, let’s say… Gosh, now I’m on the spot, let’s think of a good one. I have a story about why my food tastes this way, okay, let’s say, and it’s hot and it’s salty, so I have a certain explanation, and let’s say that it’s hot and it’s salty because I put it in the open, right. Now, that explanation does really well to explain why it’s hot, it doesn’t do so well to explain why it’s salty. I might also say it’s hot and salty because I put it in the oven and then I salted it.
0:17:42.6 SD: So now, okay, what I’ve done is I’ve explained two things in the world. In some sense, I’ve explained them somewhat separately. I’ve accounted for this fact with this part of the explanation, this other fact with this other part of the explanation, and in that case, sort of trivially, the only value, the only empirical value there is what we call descriptiveness, because it does just as well explaining one part of the data as the other part, as in fact both parts together. So there’s a different piece, though, here that’s coming in or that comes in, which is, and maybe this is a better way to say it, Sean, some explanations link things together.
0:18:24.9 SD: Some explanations not only say what’s happening, but also say, this thing happening is in some way correlated with this other thing happening, right. So I can put food in the oven or I can salt the food, but these two things in that explanation aren’t really dependent upon each other. So my explanation hasn’t linked together these two things. There are explanations that do link things together, and I know you’re more the foodie than I am, don’t you like salt crust and cook something? There’s some way that when you cook something in salt, it’s better. I don’t know, help me, Sean, right. Let’s imagine there is.
0:19:10.5 SC: You’ve picked a tough example. I think that the best explanation I can think of is, you put salt in it and then you put it in the oven. Sorry, sorry about that.
0:19:17.3 SD: You put it in the oven, right. Maybe you’re cooking it like pure, literally sodium, right? Maybe you’re cooking it Pittsburgh style where the salting and the heating is something that has to go together. In that case, the reason it’s hot and it’s salty is it being cooked Pittsburgh style.
0:19:38.1 SC: I think that we would agree… So good, so what we’re doing here is taking terms and an equation, Bayes’ theorem, and we’re relating them to human values, we like it when… If you just said on the street, do you like explanations that cover many things at once and link them together, people would say yes, and now you’re able to identify where that value is expressed in Bayes’ theorem. But what’s interesting to me here is that this idea that the explanation can link the data together is in what you’re calling the empirical part, it’s not in the prior, it’s not just part of saying, well, it’s a simple cohesive theory of everything, that’s ordinarily what I would think of as living in the prior credence before we collect any evidence at all.
0:20:27.3 SD: Right, right. So this is great. I should say, Sean, I have a better example, which I think will work much better.
0:20:34.4 SC: Give us the good example.
0:20:35.5 SD: We could rewind, or should we just plow onward?
0:20:37.4 SC: Plow onward, because we can see gears turning. That’s good for people.
0:20:42.1 SD: Okay, good, good, good, good. We can… Thinking a lot. So your intuition is not crazy, right? There is something about this feature we call co-explanation, which is how it links things that we’ve observed together, and that seems to be linked to just how the theory would do in general, that’s something we call, in fact, we have a name for it, we call it unification. The critical part here, and this is, it’s in some sense a trivial distinction, is that co-explanation only deals with the facts you have to hand, not necessarily the facts that you could get in the future. So it’s simply… Somebody presents you a scenario or a situation with a bunch of features in it, he may also present you with different kinds of explanations that make the observed things dependent upon each other or not. So sometimes it’s this very sort of blinkered view of the world. It’s only one part.
0:21:50.4 SC: In other words, let’s imagine that there were 50 different ways I could characterize the food I’m eating, and I had a theory that only had two inputs that explained all 50 of them, but if the two that I had actually measured were saltiness and heat, and those two needed both the two inputs to explain them, then at that level, I haven’t really co-explained anything, even though the theory in itself was quite simple and unified for these particular data, it was not doing that particular value.
0:22:23.9 SD: Right, exactly. So your theory would be, let’s say, high in unification. It’s making a lot of promises, it’s tempting you, right, its says, let’s go out there and see, and at some point we’re going to come up to the dark side of this. QAnon certainly has this feature, it’s like, you know, there’s a lot of things that might be linked and I wonder what you’ll find if you look for yourself, right? So there are these kind of promises that theories make, and the promises, of course, are deeply satisfying when they pay off. And so I think about When I teach, that co-explanatory moment is a lovely moment. You’ve thought this, like I’ve been telling you this, now check this out, and you pull… You pull away the curtain and lo and behold, people see this thing in a new light and they understand the old thing and the new thing as being somehow linked together.
0:23:18.0 SC: I remember telling my friends in undergraduate school, you know, what I learned is that when you do a Lorentz transformation in special relativity, like you move to a different reference frame, it’s just like doing a rotation in space, except it’s in spacetime and you could see them go like, o, why didn’t they tell us that, that’s awesome. It’s unifying these two different things.
0:23:42.7 SD: Yeah, so that’s funny, I’m trying to think of a better example. So let me give you one so we can maybe use it as we go forward. Instead of food, let’s say, being hot versus salty, let’s imagine you have an undergraduate and you see they’ve taken a class in, let’s say, French, and a class in neuroscience. And so, one explanation is they’re interested in French, they’re interested in neuroscience, right. So these two facts are not dependent upon each other, you can basically like French independently of neuroscience, but another explanation would be this is a linguistics major. And a-ha, okay, right, there is now this common explanation that says you’re a linguistics major, the fact that you took neuroscience is connected to the fact that you took French. The discovery of majors enables you to make sense of what all the undergraduates were doing in that school, these kind of hidden common causes driving the classes they’re showing up in.
0:24:46.9 SC: So I think I am getting it. The virtue of the value that we’re pinpointing here is not just the simplicity of the underlying explanation, but the fact that it relates these two particular effects that we have seen. It’s not just that it explains both, it actually relates them to each other in a palpable way.
0:25:04.0 SD: One doesn’t move if the other doesn’t move, right, that sort of thing, right? You have a sore throat and a runny nose, right. One answer could be you have allergies and you were screaming. Another explanation could be you have one of these things, you have a virus. There’s one thing I want to… It’s funny, Sean, ’cause you did the same thing I did when we first started working on this, which is you said, oh, a co-explanatory theory is also a simple one, but it’s not, right. So you can have some enormous cascade of coincid… Of like, I’ll tell you why the bus was late and your watch was broken. It’s because there’s a malevolent conspiracy that is chasing you through the world, arranging things so that the guy who fixed your watch timed it to break and stopped the bus. And yes, that does make these facts in the world dependent upon each other. It’s a deeply non-simple theory.
0:26:06.0 SC: Or even if you just had 100 different kinds of data, and your theory was that data 1 is correlated in this way with data 2 always. And data 2 is always correlated with data 3 in this way, and it just goes on for 100 different correlations independently, then you’ve co-explained everything, but you don’t have a simple theory.
0:26:26.0 SD: You’ve co-explained the pairs. So if you think about it, now just imagine filling out that pyramid. So these two things are co-explained by this cause and these two things are co-explained by this cause. Those hidden things there, those two causes, maybe they’re co-explained. So now you get this big branching tree where everything goes back to the godhead or something. But your instincts aren’t wrong, right, because… Your first thought was, “Wow, but that’s simple. It’s kind of cool that these two things are dependent upon each other.” And I think now it’s kind of a nice situation here where we see that the things we like may lead us astray.
0:27:05.6 SD: If somebody leads you to perceive these two things are correlated, we kind of like it. We think, “Oh, that’s elegant, that’s simple.” Simplicity is a value. But let’s hold on a second, because maybe there are some other pieces here that are in play. And this of course, this was our odyssey of discovery for us. It’s realising that a lot of things that we think of as values are… Or a lot of things we think of as good things are actually combinations of values. And they’re combinations, let’s say, of well-weighted values. So this is… You want to cook a good explanation, don’t put too much fish sauce in. A little goes a long way. So it was difficult for us at first to kind of break apart our own values. I won’t speak for Zac. For me it was default to the break apart our own values to start to see how things could go right or could go wrong along each of these axes.
0:28:05.2 SC: Well, that’s right. So these are values and not only just value-neutral values, but they’re good values, we would like to be able to co-explain, we would like to be able to be descriptive. But they’re going to compete against each other. That’s going to be the trick. So we have on the table descriptiveness, co-explanation. What are the other values?
0:28:24.1 SD: Excellent, alright, so let’s sweep over to the… Let’s call it the theoretical side, theoretical values, or the values of the…
0:28:30.2 SC: Those were the… You were doing the empirical side and now we’re doing the theoretical side.
0:28:35.6 SD: Exactly, right. So descriptiveness, co-explanation, it’s all about, “Show me the money. Show me the data. Let’s see what the explanation is doing for the stuff that’s causing us all these problems.” So then let’s go over to the theoretical side. We split this into two parts. One part, we called domain-dependent value. I’ll tell you… I promise, I will tell you what that is in a second. The other part we call, then you’ll be happy, Sean, we call this simplicity.
0:29:08.0 SC: That’s good.
0:29:10.4 SD: So let’s dig into the easy one, which is domain-dependent value. Domain-dependent value is your intuition’s just about the stuff the explanation is about. So the example of a car mechanic, you bring your car in and it’s like, “Oh, yes, it’s a Volkswagen. Those years, it’s the insulation that goes bad.” So these kinds of experiences, their intuition, it’s tested knowledge. It’s all of these things that go into your sense of how the world works in this domain. So most of our domain-dependent values, they’re invisible until you change fields and you’ll get really upset because you present a biologist with a theory that explains all the data, and it’s really beautiful, and they are like, “That’s just not how life works. It just doesn’t work that way.” And you’re like, “Tell me.” And they’re like, “You have to actually do this for 10 years, 20 years.”
0:30:07.8 SD: That’s a real part. It’s a real part that’s sitting there. It’s obviously a bit harder to study in the lab because either everybody has the same domain-dependent values, we all have the same folk physics, folk biology, sort of stories about the way the world works. And then of course, in the things we know very well, we have very specific values, but it’s tough to get enough car mechanics in the lab at the same time to understand how these things play out.
0:30:35.9 SC: How does this relate to the idea that if we have a new idea or some specific explanation, it should try to cohere and be compatible with things that we already think are true elsewhere? Is that the same idea or a related idea?
0:30:49.3 SD: I think we really wanted to keep this kind of simple. We wanted to say this is just… How likely do you think explanations of this pattern are? You may have some deep theory, but let’s go back to the car mechanic. There’s 50 different models of cars. You work on cars long enough, you learn that the Volkswagen has this insulation problem and the Subaru tends to have a piston problem. This is not fitting into some grand theory about car manufacturing and the industry. It might, if you were a certain kind of car mechanic, but by and large, you’re working off of, let’s say… It doesn’t have to be stimulus response, but just things you’ve noticed and remembered. Maybe there’s some pattern recognition involved, but we really wanted to kind of keep this as almost an un-theorized chunk. These are the things that make the biologists say, “Life doesn’t work that way.” And then you say, “Why?” And they’re like, “You know, get out of my office.”
0:31:51.2 SC: Let me try another example. So you and I both used to be cosmologists. We’ve moved on in our lives, but something that cosmologists often do is to say, “Well, maybe there’s some scalar field or some modification of gravity, some new fundamental physics that stretches out that affects things in galaxies and clusters that we’ve never noticed here in the lab.” And you tell this to particle physicists or to quantum field theorists, and they’re like, “No, effective field theories don’t work that way. You don’t have weird things showing up at long distances in the infrared that wouldn’t show up at short distances. We’re pretty confident in the long distances.” Is that an example?
0:32:30.1 SD: That’s a good example. I think… Let’s push even further away from… You know, particle physicists, let’s say, have stories about how effective field theories work, but… Okay, you want a physicist, I’ll give you a physicist, so Scott Aaronson has a wonderful blog post about why I won’t read your proof that P equals NP or P not equal MP. It’s kind of a long list. And none of these… I actually have to remember, but most of these reasons do not involve his theories about the way P versus NP work. One of his great examples is, if you’re using mathematics that I just don’t think is powerful enough, like category theory, no, category theory is not going to do this. No, it isn’t because Scott has tried or had some great theory about why category theory won’t work, it’s just like, it’s great, but it’s just sandbox stuff compared to what you would need.
0:33:34.0 SC: Well, Scott, of course, former Mindscape guest. A more recent Mindscape guest, Julia Galef, she and I were talking about the… How one deals with crackpot theories in physics. And the point being, so I have a way of dealing with crackpot theories in physics as a physicist, but she says she’s not a physicist, so how should she approach grandiose claims from people she doesn’t know, and because she’s not a physicist, she has to rely on signals from the person who has the theory, like do they recognize the problems that their theory has, are they willing to say they might be wrong, or are they just like, no, I’m oppressed by this system and I’m a genius that no one else has ever been before, and so that’s a different kind of domain-specific knowledge, it’s almost like the psychological domain, like what are the features that you’re likely to have in a crackpot versus a respectable scientist?
0:34:27.1 SD: Exactly, I think that’s even… That’s a nice example, ’cause it really gets at the un-theorized version here. And there’s nothing wrong with a GeoCities’ web page in the 1990s, like John Baez had a GeoCities’ webpage, and he talked about the fundamental theories of physics. So when you turned on that page and it was… John probably didn’t have this, right, we had the animated GIF of the flames and the under construction sign, and it says how to understand Feynman diagrams, you’re like, great, here we go. The same, literally the same website in 2020, it’s like, this is… Don’t, right, stay away. And I think there’s… Again, I’m not sure I want to put Zac on the spot here, but my sense is that these domain-dependent priors, a lot of this is much more what we might call tacit knowledge, the stuff you know, but you can’t say.
0:35:25.1 SD: So Julia, when she says, okay, I have to watch the person. She’s not crazy. There’s good reasons to do that, but when Scott says category theory is just not powerful enough, you’re like, well, can you tell me why?
0:35:41.2 SC: Not really, yeah.
0:35:41.6 SD: And he’s like, it’s my feel. Like I’ve been juggling these things for years. So this is great, it’s funny, Sean, Zac and I didn’t spend too much time on this, this part, partly ’cause we had a word limit, but the domain-dependent priors are there and they matter, right? The other piece, this is where all the excitement starts showing up, right, because this is where we start judging what you might call the aesthetics of a theory, so not how it does on the data to hand, not whether it fits your gut and your gut meaning your experiences of empirical life, but the ways in which the theory look. If you dig deep enough into this, things get very strange, and sort of beautifully strange.
0:36:33.6 SD: We talked about many different values that fall under, let’s say, the simplicity value, succinctness, like literally, how short is the explanation. Can you say it in five words or less? You know, my dad used to say, if someone can’t tell you what their job is in a sentence, they don’t have a real job, right? That’s may be a little unfair, but succinctness is something that might tell you whether the explanation is a good one, right. Why were you late? Oh, okay. So probably this is not… You believe this explanation just because it takes a long time, so unfair? Maybe.
0:37:15.3 SD: Concision, a slightly more advanced version that talks about, okay, maybe you switched languages here, is it concise in this language, does a Lorentz transform look particularly elegant in a hyperbolic [0:37:27.0] ____ universe. We can talk about unification, for example, which is the way it links things together, we can count the number of hidden causes, we can say, your example of there’s 100 things and they’re each pairwise linked together, and then we link all the pairs together, we know, okay, fine, it’s log number of things, causes roughly.
0:37:52.1 SD: So just counting causes, Occam’s razor, we debate this a little bit, but one simple thing is to count parameters. How many parameters are in your theory, if you have a mathematical theory. When I was a grad student, I gave a talk on my latest dumb scalar energy theory, and who was it? Some famous physicists said how many parameters does your theory have, and I said, whatever, three, and he just walked out ’cause that was just too many parameters for him, right. It was, it didn’t matter what I could explain, it didn’t matter how everything linked everything together, it’s just that’s too many parameters, I gotta walk my dog.
0:38:33.5 SC: So you’re like, I’m leaving, I’m going to social science.
0:38:37.5 SD: Right, well, and so… There are some interesting… We’ll talk a little bit about the social science stuff because there’s some interesting stuff that happens there with respect to simplicity. You can keep digging, though, so take the value of succinctness. One way is to say not just succinctness in English, but like succinctness in… And here we go, right, succinctness in the computer program that if you ran it would print the explanation out in the language you can read. This has claim to being in some sense the right definition of simplicity, or let’s say concision or succinctness, because it makes it sort of language-independent.
0:39:24.0 SD: We know roughly speaking that whatever language you write in it won’t change that value very much. So you write your code to generate the model in Python versus Lisp versus C. It’s maybe a constant offset. So this idea, and it’s called Kolmogorov complexity, also called Chaitin complexity, it has many different names for it, because partly it was invented during the Cold War. We had many people in the West and they had Kolmogorov, and so he gets his name on literally everything. Let’s call it Kolmogorov complexity. So this is, in one sense, is the ultimate value, if we could perceive this value, we would know the true simplicity of an explanation, right?
0:40:12.5 SD: Now, should we value that true simplicity? Maybe, right, let’s put that question aside and just say, whether it’s a virtue or not, let’s just say, how could we come to know that value, and it turns out it’s logically impossible that you could know this value. We can say what it is, but unless you’re like Roger Penrose, and you think that humans in some sense transcend the Turing world, if we can’t be efficiently or inefficiently simulated by a computer, unless you think that, we have no contact with this value…
0:40:51.9 SC: It is uncomputable.
0:40:53.9 SD: It’s value, it’s uncomputable, right, and I go on about this at length, uncomputable means uncomputable and you can’t compute it anyway, so you can’t approximate this value, because any approximation you do will have unknown error and then you say, fine, I will compute the error, which of course is uncomputable, right.
0:41:13.2 SC: And so actually this is a very good opportunity for me to make sure I understand this, ’cause Scott and I wrote a paper, Scott Aaronson, and we need to mention the fact that Kolmogorov complexity is uncomputable and I didn’t understand it, and he finally taught it to me, so let me see if I remember. I mean, ’cause the obvious counter-argument is, given any language, I will just write every computer program from the shortest one to the longest one. I will keep writing longer and longer computer programs until I output the output I want to get.
0:41:45.1 SC: And the reason why that fails is because of the halting problem, because you will eventually hit computer programs that never terminate and you don’t know whether it will terminate or not, so if you… In your enumeration of every computer program, if you don’t actually by luck output what you’re looking for, you will never be able to get to what you’re looking for.
0:42:06.0 SD: Right, exactly. That’s one way, it’s all of the… Your way is blocked at every turn, it’s like the Dungeons and Dragons game, and the DM is logic, and the DM, you can’t get out of the room. So every door you try, there is some problem with it, right, so your example, is, oh, if you just enumerate every computer program, then eventually you’ll hit one that will never stop, this is the halting problem, right, so okay, the halting problem blocks you that way. People have different creative solutions for how to solve this, every creative solution runs into a kind of Gödel-like problem. Every Gödel-like problem secretly is a problem of self-reference.
0:42:51.3 SC: And so we’re saying that simplicity is something that we hold in great value, but can’t really quantify it?
0:42:58.5 SD: It’s like, it’s the… Don’t cancel me, Sean. It’s the atheist God, it’s the negative theology of the ultimate theory. We will never know, we actually might have it already, but we can never know that we have it.
0:43:11.1 SC: Got it, okay.
0:43:13.5 SD: So, and just to drive the intuition here, I love… You can transform this into the halting problem, another thing you can say is… And this is what’s called Berry’s paradox, right, you have some way to name all the numbers, okay, find the shortest number that you can’t… Or the smallest number you can’t name in less than 50,000 words. I have just named that number in less than 50,000 words, so there’s some paradox here in talking about how difficult it is to name things, in part because when you name things, you can also talk about names. So that’s the kind of self-referential aspect here.
0:43:51.2 SD: If we were somehow able to ban all self-reference from our theories, we could actually compute the simplicity, but I’m not sure we’d want to do that, because most good theories in some sense can refer back to themselves. And I guess in some very simple physics theories that may not be the case if they’re sitting purely as a set of, let’s say, discrete update equations, it may be possible to think about the shortest way to specify them, but if you could say, okay, look, what’s the context relying on. No, because those are uncomputable as well. It’s really hard, right.
0:44:31.2 SC: It’s very hard.
0:44:32.3 SD: It’s really hard to have theories that are boring enough such that you can know with absolute confidence their simplicity. So I guess the only one I can think of is like a Markov model. Markov models, we actually do know, we can sort of compute how simple they are.
0:44:46.5 SC: Okay, but at the end of the day, where we are here is you have enumerated, denumerated? Four values, so to remind everyone, the descriptiveness, the co-explanation, the domain-specific knowledge, was it domain-specific?
0:45:02.6 SD: Priors, domain-specific priors.
0:45:04.3 SC: And the simplicity. So these are the things that we look for in an explanation, and when I read your list, it reminded me there was a famous or semi-famous list that Thomas Kuhn came up with. I don’t know if you’re familiar with this, but when Kuhn wrote The Structure of Scientific Revolutions and said, well, there are paradigm shifts and you can’t even judge one paradigm from within another one, he was accused by his detractors of being a relativist, of saying no scientific knowledge or progress is possible, etcetera.
0:45:30.9 SC: And he didn’t think of himself that way, so he wrote a follow-up piece where he said, well, no, I’m just saying kinda there’s no algorithm for doing it, but there are values that we have, and he listed seven values, and I forget what they were. But one of them was fruitfulness, like if the theory would not only explain what it’s explaining, but it has the promise of explaining other things. So my question is, does your list of four values purport to be it? Is it comprehensive and exhaustive? Are these the values that we have when it comes to explanation, or are they some of them? How should we think about that?
0:46:07.6 SD: Right, right. So I would say Kuhn’s fruitfulness is probably our unification, unification being the co-explanation you’d get if you observed lots of other stuff and the theory turned out to be true. So it’s a piece there, we give it a name. In the end, it would be nice if we had a normative theory of explanation, meaning we know which ones… We know when we’ve got it right. Really what we have here is a psychological theory. We’re interested in the axes along which a theory can get better or worse, that we perceive. So it’s a little bit… The extreme version would be like, okay, it can be salty, it can be sweet. How do we as explainers, human explainers, look at the world, and look at explanations?
0:47:04.2 SD: Then you can say okay, well, maybe we’re not so bad at it, because as David Deutsch says, we’re amazing, infinitely capable creatures. Maybe we’re onto something in having these values, but we can also over and undervalue them, so there’s a nice line here. Values can be both virtues and vices. We can value the wrong things, or we can value them too much, or too little. So Kuhn’s list of seven, I wouldn’t be surprised if many of them aligned with the pieces we have. He may have found others that don’t align that way. A couple of things could be going on. One is he could be wrong, right?
0:47:48.1 SC: Yup. [chuckle] Always possible.
0:47:51.0 SD: Another is that he could have a psychological value that for him is very real, but that he has learned. So another piece here that we have is that we can… Not only not only do we have, let’s say, baked-in values… So when you study children, you discover… Psychologists who study this, discover that children like co-explanation. They like sweet things and they like co-explanations.
[laughter]
0:48:15.7 SD: So some of these are kind of baked-in, but others are sort of trainable. So it’s probably the case that you and I, as people who like physics, had a heavy weight on certain values, let’s say the unification value, the simplicity value, but it was exaggerated over time because all of our charismatic teachers gave us candy when we valued simplicity more. Conversely, anthropologists value simplicity less, and it’s in part because they know the world is not that simple when it comes to people. So I remember there was a simulation, I saw people doing a simulation of the civilization developing in the American South West, pre-contact, and they had this model where there were people on a landscape and choosing where to walk and choosing where to settle and build houses, and I’m sitting there, and I’m sort of boiling with upsetness as a physicist.
0:49:17.1 SD: It was just that I’m new to the Santa Fe Institute where this was happening at the time. I was like, “This is terrible. There’s two kinds of houses. How many parameters does this theory have?” And then this great anthropologist raised his hand, one of the big figures, and he’s like, “But where are the turkeys? Why don’t you have the wild turkeys in the simulation ’cause that’s important, right?” And he was right. That matters. That’s an important part of explaining what’s going on, where the turkeys are, but with somebody with a different set of values would say, “This is just getting too extreme. We need a different… The fact that we’re adding these things into the explanation is making it worse and not better.”
0:49:58.4 SC: Well, I guess this is one of the things I wanted to ask because you have these different values, and as you just highlighted very clearly, in the real world, they compete, sometimes, anyway. Obviously, if there’s one theory that is both simpler and more co-explanatory and more descriptive, it will win, but sometimes a theory is less simple but more descriptive, etcetera, and then you have to balance, and that’s harder. So I guess two questions. One is: Does the right way to balance these values pop out of Bayes’ theorem or something like that? Have you mathematically proven the right way? Or, question number two, can we empirically figure it out? Can you go back in the history of science and say, “Well, this person was valuing simplicity and this person was valuing their domain-specific priors, and look who won,” and sort of tote up a score card?
0:50:55.8 SD: Yes. Well, let’s say partly. We have a story about the proper weighting between descriptiveness and co-explanation. There’s a proper ratio in which you should value these. Ratio and units and the correct units is one.
0:51:14.5 SC: That’s a very simple explanation. [chuckle]
0:51:15.9 SD: So you should value these two things equally in a certain way. But the real challenge here is when it comes to the theoretical values. What we don’t have is a normative, let’s say, or the optimal or the ideal way to talk about the theoretical values. There are people who will tell you that, for example, simplicity has to be a value. And you say, “Well, why does simplicity have to be a value?” And you talk to them long enough, and it turns out it’s because the universe is simple. “Okay. How do you know it’s simple?” “Well, it turns out that’s the simplest explanation for why… ” So on that side, we don’t have an answer, but you ask, of course, the right question is: Can we just go see how people do?
0:52:09.6 SD: Can we see how well people who valued this kind of explanation over that one, how they’ve done in historical time? So we’re actually… We have a project on this, it’s really fun. We have all the data from the proceedings of the Royal Society, the Royal Society of London. So this is the first scientific institution. It’s formed in 1666, or at least the Journal starts then.
0:52:35.5 SC: 1660.
0:52:36.1 SD: It’s… Sorry?
0:52:38.5 SC: The Society started in 1660. I know that ’cause I was literally reading about it yesterday. [chuckle]
0:52:42.7 SD: Excellent. Okay. The Journal starts a little bit later, maybe… I can’t remember now. It’s like 1663. This is why I drive historians crazy. It starts on the order of 10 to the 3 years after the birth of Christ. So…
0:52:58.1 SC: Cosmologists.
0:53:00.1 SD: We have this data on essentially how scientists are putting ideas together over time. And how do we track the ideas? We do some magical pattern recognition on the text, we look at the patterns we find. We as scientists say, okay, these patterns are making sense. So we can detect the magnetism idea, we can detect the electricity idea, we can detect the magnetic substances topic, which is a different one. It’s like, magnetism, what’s up with that? And it’s sort of sad, ’cause our data only goes to the late 1800s, 1887. So we know they’re not going to figure out why iron becomes magnetic until the 20th century. It’s just not going to happen for you.
0:53:48.3 SD: And so you sort of feel bad, you want to say, No. Other things they do figure out, though, of course. And famously, what they figure out is that there is this global conspiracy between the electric and the magnetic fields, which we call the electromagnetic field. These forms, and it’s tricky to call them co-explanation, because these are not… These topics or these ideas are not just about the observed things, but also discussions of the ideas themselves, but we can track how these ideas link together over time.
0:54:24.8 SD: The first thing you find is that this value, at least in science, kind of appears out of nowhere. So the first hundred years of science is people putting ideas together in somewhat arbitrary ways. Now, it could be they… It could be they know that ideas should be linked together, and no one’s agreed on how to link them together, so they may have the value, or perhaps more likely, I would say, they haven’t yet learned that what you should be doing is finding the ideas that tend to link together and working on those. So preferring, and in fact, we can see this happening, preferring to work on ideas that reliably connect to particular other ideas. So this is kind of wild. We can see people moving towards, and it starts, this preference starts around early 1800s. Actually, one of the first people, one of the… The issue in which it starts is the issue in which they published Ben Franklin’s kite experiment. Can we date it that precisely? No. But that’s what the Bayesian model says, right. Great Pennsylvania physicist.
0:55:36.2 SD: But beginning around the early 1800s, we have this era, which lasts maybe about 100 years, maybe 50 years, we have this era where people start connecting ideas together. So people know, oh, idea A has something to do with idea B, and not C. Whereas earlier, you look 50 years before, 100 years before, they’re like, Earthquakes. Cows? Maybe cows? And you’re like, No. It’s not… That’s not going to happen. But at some point, they start putting these ideas together. Let’s say, the overall unification level in science starts to rise. We can also see people, as I said, people choosing to work in… On areas that are linking together. So we see this emergence of a value, or at least the value begins to articulate itself in the record.
0:56:34.6 SC: Can I… With respect to that, can I… This might be relevant to a very long-standing puzzle that I’ve wondered about, which was… ‘Cause my first trade book, From Eternity To Here, was on entropy, and I read a lot about Boltzmann and Maxwell and their discussions, etcetera. And one of the big objections to Boltzmann was, he was deriving the second law of thermodynamics, that entropy increases, as a probabilistic statement, but it wasn’t absolute. And people thought it should be absolute. They thought it was a law. They thought it was a law separate from Newton’s laws of classical mechanics. And that always puzzled me, like, how were you allowed to think that? It was the same stuff. How could a gas obey Newton’s laws and the second law if they weren’t compatible with each other? And are you telling me that maybe that possible incompatibility just wasn’t something that would throw itself in their faces and make them bothered by it at that point in scientific history?
0:57:36.7 SD: That’s great. It’s funny, Sean, one of the papers in our data, it’s a bit ironic, is Baez’s original paper. So the Reverend Baez publishes his article, and it’s an amazing article, it’s actually… He was… He died, and somebody said, Baez told me the following. It’s an amazing article, because actually everything’s in there. So the idea that models make probabilistic predictions, the idea that you can go both directions, that you can go from model to data, data to model. It’s all sitting in there. But somehow, this just didn’t catch fire for people. The idea that knowledge could be probabilistic, that this was a good representation of what we know or how to know things just didn’t take off. Baez’s that… Those ideas are sitting there and don’t get linked, which is sort of interesting. So in one sense, I can tell you, you’re right, no one connected these together.
0:58:35.7 SD: In that case… How do you punt it? You might punt that to the domain priors, right? Newton is about precision. Newtonian laws are about determinism, the solar system will go forever running like a clock, so somehow the idea that there would be some relationship between Newton’s, that style and probabilistic reasoning, it’s a little bit like the Scott Aaronson idea that that’s just not powerful enough. But in this case, it’s like, no, that’s just not the right thing. We’ve been doing this for 200 years, but they were just wrong. So I think… That’s a lovely example, Sean.
0:59:15.1 SD: It’s easy for us to focus on the successes. The obvious one is we can see them put electricity and magnetism together, which is a beautiful thing. We also see them put electricity and electrochemistry together, which is an interesting one, because electricity is more of a theory, electrochemistry is the experiment. So we can start to see people… You see what Joe Moka [0:59:39.2] ____ calls this virtuous cycle between theory and experiment, better theories let you build better devices to get you better theories. So we see some of these loops. The one that I really love, actually, I should say, it’s not just about physics, so we also see people connect demography and agriculture, birth and death, life spans, child mortality. People start to realize this, it probably has something to do with how people eat, right. They connect agriculture to geology, whether your crops grow probably has something to do with the underlying rocks. We see them connect metabolism to actually to agriculture, another one, metabolism ’cause it’s like the cow eats stuff and it processes it and it poops it out.
1:00:31.8 SD: And we should probably, if we understand that, that will probably help us explain phenomena that we’ve noticed in agriculture and so on and so forth down the line. So it isn’t just the… It isn’t just what we might call the exact sciences that are linking together, but also these might… We see paleontology connect to botany as they start realizing that they can learn about the deep history of plant life, it’s not just about dinosaurs. This connection stuff is kind of all over the place, and I’ll tell you the story and then we can chat a little bit more, but the one I really love is… Okay, so they connect electricity and magnetism, right. But as we may remember, light is a form of electricity and magnetism. Hertz’ paper is in our final issue before it goes under copyright or whatever, right. So at the end, they’ve put this together, they’ve linked electricity, magnetism and light, they realize this is all… These things all go together, but we can see them make these connections, we can see them start to connect magnetism and light, like 40, 50 years earlier, which is wild, right. Now, what are they doing? They’re like, there’s this magnetic substance and I shine light through it, and something’s weird happening, look, guys, help me out here.
1:01:58.4 SD: What is it about? Magnetism seems to be doing something to light, like what… So you can see them start to make those connections, and so one of the things that it brings up for us is, can you sort of look into the future, so can you see where the next advances are going to be? Now, that would be… In some sense, you can’t really do that, it’s impossible to predict an unpredictable future, and science is in some sense unpredictable, but it maybe gives us a sense at least that we’re on the right track here as to what’s making for good explanations for them.
1:02:34.8 SC: And I do want to give you enough of a chance to talk about what makes for terribly bad explanations as well, because this is a very fun part to what you’ve done. I mean, the different values that you’ve pinpointed, like we said, they compete against each other and it can be in some sense not completely algorithmic, how you weigh them against each other, so there’s a failure mode where you over-value a particular kind of thing and undervalue the others. So why don’t you tell the Timothy McVeigh story? I think that’s a great example. I really like that one.
1:03:05.5 SD: Yeah, okay, I’ll tell you two stories. Timothy McVeigh is a really interesting one. So this is the conspiracy surrounding the Oklahoma City bombing, this sort of the worst terrorist attack on American soil by that point, 1995, white supremacist attack. What happened? Well, Timothy McVeigh and some co-conspirators built a fertilizer bomb, put it into a U-Haul truck, drove it to the federal building in Oklahoma City, set a bunch of timers, left it there, it blew the building to pieces. Many, many people died. The devastation is quite shocking, actually reading about it in retrospect, we sort of forget how insane it was. Just one example just of the sort of horror of this is many people died just because of the broken glass that flew out from the explosion, so the building also collapsed, so this was crazy.
1:04:04.0 SD: The other thing about this attack that’s relevant here is that no one knew who did it. This was actually an extraordinarily meticulously done thing, at least at first. They couldn’t… Who drove the truck? Where did the bomb come from? They had no leads. It was not… There were not mistakes made, let’s say, at least early on, in this investigation, so what happens? Well, McVeigh sets this bomb off and he walks away, it’s on a timer, he’s far enough away, five minutes later the thing blows up. McVeigh gets in a car and leaves. So he’s in his car, the car he’s driving doesn’t have license plates on it. He gets on the highway and he starts speeding, he gets pulled over by a cop for not having license plates and for speeding, and the cop notices he’s got a firearm, and it turns out it’s like an unregistered firearm, so he gets thrown in jail.
1:05:10.5 SD: And it’s only when he’s been in jail for three days that they figure out it’s him. So there’s this great puzzle. How could somebody so able to pull off this plan, this very elaborate plan, be so stupid? Even I know this. Don’t drive a car without license plates. And if you do, don’t get on the highway and speed. Don’t blow past a cop. So how do you put this together? The explanation, of course, you and I have about the Oklahoma City bombing and, spoiler, it’s true, was that he was essentially an insane, deeply evil person who drew a few people in to help him blow up this building, spurred by… Let’s just call it white supremacist rhetoric. That’s our explanation.
1:05:58.9 SD: The problem is, that it’s low in what we might call descriptiveness. That explanation postulates that Timothy McVeigh was not a moron, at least in certain relative future planning abilities, and yet he acts like a moron. If you’re not a moron, doing what Timothy McVeigh did after the bombing is a low probability event. So now our explanation is suffering on a certain value. It’s suffering on this empirical value of descriptiveness. It’s like, say, to go back to our earlier example, why is this student taking French and neuroscience together? Ah, because they’re a romance language major. Okay, well, that actually makes the neuroscience less likely, ’cause they got other requirements. But you’re, “No, no. Come on, there’s other reason.”
1:06:48.6 SD: Some people look at this and they’re told this story, and they say, “No, I can explain that.” It wasn’t McVeigh. McVeigh, yes, he’s a moron. McVeigh was a patsy. He didn’t pull this off. Actually, alright, let’s go, right, actually, it’s a government conspiracy and the Bureau of Tobacco and Firearms is involved and… So they have an explanation which can actually account for these facts.
1:07:11.7 SC: They’re better at fitting the data.
1:07:15.6 SD: They’re a better fit to the data, and so you know, of course, what’s coming, what’s going wrong for them. I can over-fit to the data. And so, of course, the theory they need not only is strange, not only does it violate our domain-dependent priors… And the joke is, our domain-dependent prior is that the government’s not that good at doing anything, so how could it do this. But also it’s violating these simplicity priors because once you assume there’s a conspiracy, okay, who else is in on it? Why did the cops not see it if the cops were in on it? And the cop who wasn’t got shot in a mysterious way. And the newspaper guy is… He disappeared. So of course, this explanation ramifies outward. That’s the case where we over-value an anomalous point or we over-value dealing with the anomalous point. That’s a classic case from our point of view of going extreme on that first value of descriptiveness, the desire to make sense of every last detail.
1:08:12.5 SC: And this is probably not true, but I can’t help but hypothesize that this might have something to do with one’s fondness for detective shows and novels, because when you’re in a mystery fiction, there are no coincidences. You’ll hear the detective say, “I don’t believe in coincidences.” And every fact turns out to be really, really important later on. And I think that’s partly why Alex Rosenberg, who was another former guest on the show, he likes to say that we make a cognitive mistake by over-valuing stories. We tell ourselves a story that makes everything fit. Sometimes things just happen. And it’s hard to weigh those two values of admitting that sometimes things just happen with the satisfaction you have of fitting it all together into a matrix.
1:09:00.3 SD: I think that’s right. I can’t remember if they let us keep this, but Zac and I talk about, or wanted to talk about, perhaps, yeah, there’s this enormous pleasure in a detective novel when all the facts are connected together. It’s this co-explanatory moment at the end and there’s always the scene, Poirot brings everybody into the room. There’s a great [1:09:22.7] ____ so… Umberto Eco in his book, Name of the Rose, and this is not a spoiler, but Name of the Rose is a great… This is mild spoiler. Name of the Rose is a great book because in fact, there is no co-explanatory moment. It’s a detective story without co-explanation. And of course, I think Eco, actually, he sort of talks about this a little bit in maybe an introduction essay he wrote about it. So that’s a great example of how he breaks the convention.
1:09:50.5 SD: But you’re right. Alex says… I think he’s got a piece of the puzzle here when he talks about stories, because stories often contain this co-explanation, that’s what makes them appealing, but there’s also more. So I would say, for example, if we’re right about how these empirical values work, you would expect conspiracy theories to involve not just stories, but anomalous facts. So this meme, jet fuel can’t melt steel beams, which also turns out not to be relevant, but it’s a fact. That fact comes in, as well as dramatic accounts and stories and narratives as well. So there’s this empirical side that I think we can’t neglect, these little pieces that people find attractive, or at least seem to be part of the appeal.
1:10:45.9 SC: And presumably things like QAnon or flat Earth beliefs are of the same spirit where you’re explaining a bunch of things. I guess flat Earth is not quite the same thing, it’s not a conspiracy theory. But QAnon is the classic, it’s the epitome of explaining everything by having a million moving parts in your theory.
1:11:04.2 SD: Right, so let me… So we talked about one way to go wrong, which is this descriptiveness one. Let me talk about the other one, which is the co-explanation going wrong, because this is a different way that things can break. So that little fragment of the McVeigh Oklahoma City bombing conspiracy theory is clearly a case where people need to fix an anomalous piece. Jet fuel can’t melt steel beams fixes an anomalous piece. Co-explanation is a different appeal, and I think this is partly working in QAnon. I think I can tell this story, because I told it once before, but I won’t… This will be the last time I tell this story ’cause it’s such a good story, and I don’t want to over-play it. But I’ll give you an example of co-explanation that I fell victim to. Here we go.
1:11:56.0 SC: We all have.
1:11:58.1 SD: We all have. So this was when I first moved to Santa Fe, and I was in the cafe, and Santa Fe is full of very interesting people, some of whom are crazy. One of the jokes is the people in Santa Fe, they were so disorganized that their car broke down on the way to San Francisco, right. So that’s your group. And so I’m in the cafe, and my first week there, and I meet this guy, and he sort of buttonholes me and he starts telling me his conspiracy theory, and I’m sort of, don’t make enemies, it’s your first week in town, so I’m listening and he’s, the conspiracy he’s explaining to me, and I think this guy is long out of this, right, but at the time, he was telling me this conspiracy theory known as the Sovereign Citizens Movement.
1:12:45.7 SD: So the Sovereign Citizens Movement is a conspiracy theory that is so elaborate that I could… If I remembered, I could tell you the whole thing, but just to give you a fragment of it. It involves the idea that British common law in some way meant that the United States government in the 1800s could not borrow money on the credit of the citizens, like somehow this couldn’t work, right. So what they did was for every citizen, they created a fictitious identity called your US name, so everybody’s carrying around a fictitious identity called the US name, and it’s the US name identity that for example, has to obey the law, right.
1:13:33.8 SD: The US name identity is the one that has to pay taxes, you yourself, actually, you only, it turns out, are answerable to like the [1:13:41.2] ____ of your tax.
1:13:43.7 SC: I’ve never heard this one.
1:13:43.8 SD: So this is sort of funny, but also obviously the guy keeps going and eventually it turns into… Obviously, there’s anti-Semitic features of all this, right. One of the nice things about this movement is people know it very well, because one of the things it tells you to do is write a letter to the IRS telling them that you’re not going to pay taxes. So it turns out they keep those letters, right, so this is, it’s a movement that’s easy for the government to track. So he’s explaining this to me, and I know at some point I am waiting for it to get dark. But he says, look, I’ll tell you something, Simon, I’ll tell you about your US name. The government has to deal with you in your US name capacity all the time. When it does, when you get a letter or something like that, your name will be printed in all capital letters. So, take your wallet out, Sean, do you have your wallet?
1:14:34.2 SC: I don’t have it with me right now, but I’ll trust you.
1:14:35.1 SD: Okay, take your wallet, okay. So if your listeners do this, you can test, you can test your your co-explanation problem here. It’s like, your US name from the government is all in capital letters, now take out your wallet, look at your driver’s license and my name’s in all capital letters.
1:14:49.7 SC: My goodness. What about the passport?
1:14:51.6 SD: And I was like, for this moment, Sean, I was like… Just for a moment, I was like, there’s something… Oh, my God, there’s something to this.
1:14:57.6 SC: No other way of accounting for this.
1:15:00.7 SD: And so this is the classic… Yeah, it’s a classic co-explanation moment, because what he’s done, of course, is link all of these facts he’s given me about British common law, or whatever, not that I believe this, right? But it’s like he’s told me all about Jefferson, [1:15:12.9] ____ and then he’s given me this story that links with this totally unexpected fact that my driver’s license has my name in all caps. So this is now all connected, and this is kind of a lovely feeling, like the world very briefly got sort of brighter and the colors got a little brighter. And then you sort of shake it off and you’re like, no, this is crazy. But maybe I think the message, or the larger message, Sean, is that these are values, we have them. People who let’s say fall victim to these conspiracy theories, and even if people… Even if someone doesn’t go crazy, there are some very negative features of believing a conspiracy theory, one of them actually being that we love to explain things to each other, it’s a human… It’s one of the things we do all the time.
1:16:02.9 SD: If your explanatory values go wrong, you can’t enjoy this with other people, right. So you and I sit around and it’s like, I don’t know, Sean, we would never have this conversation. It was like, what about the Bulls? The Chicago Bulls are doing really well this year, and he’d be like, what’s your theory? Well, let me explain it this way, that way, we can have a lot of fun. But then the guy with the weird values, it’s like, “It’s the Jews.” And you’re like, “Hey, don’t say that. That’s insane.”
1:16:26.9 SD: But also it’s like, That’s not how we… That’s not a satisfying explanation, even if it wasn’t creepy. So you lose this ability to spend time with others, and people mention this, so some studies of QAnon people, this is part of it is they sort of get exiled from their friend groups, not even because they’re being weird or sexist, racist or anything, but because they’re just no longer fun to do this fundamental human thing with. You can enjoy a sunset with them, but you can’t explain why Trump won the election or why Aunt Sally’s so upset this week. So there are some real downsides, but I think our message here is that those stumbles, those falls that people have, they’re not some alien ridiculous axis that’s completely orthogonal to you and I.
1:17:22.9 SC: Yeah, exactly.
1:17:24.9 SD: That their reasoning… Their reasoning has gone wrong, but they are still reasoning, they’re not babbling, and the question people always ask us is, okay, you’ve explained these things, can you predict what to do about them or what intervention will work? And so the answer is no, we don’t do prediction, but the explanation, I think, does potentially… It gives us part of the puzzle, because if we see the membership of anything like QAnon as in part a disorder of explanation-making, well, how do we fix that? How do we get people’s values back into balance?
1:18:05.4 SD: One piece, and this is work that we’ve been doing, Chloe Perry and I have been doing, Chloe is now at the University of Michigan, is the idea that you have weird explanatory values. You can’t hang out with your friends, you go on the internet, and what you do on the internet is not just talk about jet fuel, can’t make steel beams, but you also share and reinforce these sort of malfunctioning values as well, so you’re surrounded by people who don’t just believe crazy things, but also have the wrong meta principle for adopting them. It suggests that part of this is disconnecting people from an epistemic value system as well as their particular beliefs, right.
1:18:47.9 SD: So it isn’t just saying, “It’s ridiculous that you think Timothy McVeigh was part of this conspiracy,” that people can be smart and idiots the same day, but also making sense of the way that they’re connecting things together on a larger scale.
1:19:05.3 SC: But part of what I thought you were saying is that the actual set of epistemic values that the conspiracy theorists have is the same set of values that we very level-headed natural scientists have, they’re just weighting them different, and that sounds like a harder thing to… Or maybe it’s an easier thing, like maybe since we rely on the same values, we can sort of speak to those values and bring people back. I don’t know, is there an optimistic message here?
1:19:38.2 SD: So I gave a talk, actually, Zac and I together gave a talk to the philosophers at the University of Pittsburgh, so Pitt philosophy is like the best philosophy in the universe, neo-Hegelian pragmatism sometimes. But one of the people in that seminar made this lovely remark, he’s like, “You have an Aristotelian theory of epistemic values,” meaning so in Aristotle it’s all about this balance. Don’t be foolhardy, but don’t be a coward. Where do you find yourself on that continuum? Weakness versus strength. You need to find some, like the intelligence here, the wisdom here is finding the correct location on that line, not getting yourself killed, but not running from a fight you may need to fight.
1:20:28.8 SD: We have an Aristotelian theory of these epistemic values, they become virtues, or you become, let’s say, epistemically virtuous when you’re at the right point here.
1:20:38.8 SC: The golden mean.
1:20:40.9 SD: Now, does Aristotle tell us how to treat anger? Maybe, right, people who are too foolhardy, people who don’t take enough, risk, maybe. But it tells us maybe one way in which these values are operating.
1:20:54.9 SC: No, that’s actually really good. I’m becoming more Aristotelian in that sense, myself, in the sense… I mean, not only is it that you balance things and you look for a harmonious middle point, but that there is no algorithm for doing it, right, that there is some kind of human choice. Wisdom is the word you used, phronasis or something like that, I’m sure that the Greeks would say. But… Okay, good, so maybe let’s… To wrap this up then, put it to work. Let’s do some worked example here. How would we think about, let’s say, I have lots of worked examples I can pick from, but let’s say panpsychism versus physicalism about consciousness.
1:21:33.4 SC: So here are two explanations for consciousness. One is that the world is just protons and neutrons and electrons obeying the core theory, and there is some higher level emergent description, and then we call that consciousness. There’s another explanation in which, no, no, no, there’s that stuff and it obeys those laws, but there’s also intrinsically mental aspects of the physical stuff which are not captured by the standard laws of particle physics, etcetera, and those extra intrinsically mental aspects are needed to explain conscious experiences.
1:22:07.0 SC: So here’s two explanations, and you have a bunch of values and I can kind of… I can imagine how I would think that the physics one is simpler, maybe, but what would you say to someone who wanted to evaluate these using your values?
1:22:20.6 SD: I love this example so much, ’cause it’s wonderful, so let’s deal with the property dualism one, right, the panpsychism one, that says, look, there’s just this other property we all have, which is consciousnessiness. This would be, I would say, a highly descriptive theory. There is these things, and then there’s these other things, and physics does these things, and we just have to stick in this other story here that they’re not connected, there’s no co-explanation. This conscious property is in no way correlated with that physical property in any deep way. The properties live in distinct worlds of matter and spirit, so it’s winning on descriptiveness, right? It’s not winning so much on co-explanation.
1:23:20.5 SD: But I want to flip this around now, ’cause I know what you like, I know what I like, which is this emergent story. There’s this stuff on the bottom here and it’s bubbling up to produce these emergent phenomena, but that’s actually… It’s not that simple, because how many layers between quark and belief are there, right? How many… Okay, oh, gosh, I guess you have to do the molecules and the molecules… Okay, so now you gotta get the substances and… There’s a lot of space between the standard model and conscious experience under that other one, so in that case, that almost, you can make the pitch, like that’s a real conspiracy theory.
1:24:04.8 SC: Well, I think this is… I make this point in my recent quantum mechanics book about people who believe in, let’s say, hidden variables or pilot wave Bohmian theories versus people who believe in many-worlds. Not only are they different theories and they can argue about which one is right, and we don’t yet know, but also they both claim they have the simpler theory, but the sense of simplicity is different, and I think that this is one of the reasons why I like your set of values, because I can see them coming apart. The Everettians will say, look at my equations. There’s only one, there’s one equation, that’s all, and everything, you just work on that equation and massage it and think about it really hard, and you get everything. That’s simplicity for me.
1:24:47.5 SC: And the Bohmians say, well, no, look at the world, and I can easily locate it in my theory, there’s particles, like there they are, in my theory, I don’t need all these layers of explanation, so that’s what I call simple, but maybe that’s actually more descriptiveness. I’m not sure.
1:25:03.8 SD: I think that’s actually… I think that’s a great way to distinguish the many worlds from these more pilot wave things, right. You have to grind through a lot of mathematics to get from the many-worlds to the classical world, right. Now, in one sense, this is not simple, like David Wallace’s book is 500 pages, but if you think about this as, let’s say, a program that you run to derive these things, the program actually might be quite short, in the sense that if you… You just have to be really smart, meaning you have to be able to run that “program,” but the program might be short.
1:25:44.7 SC: Yeah, so let me… There’s the shortness of the program, but there’s also how long it’s got to run, right?
1:25:50.2 SD: Right, yeah, the logical depth, they call it, how long the program runs, all of these… These are wonderful. I think one of the pieces… You know, Sean, we could talk about this forever, right, but one of the pieces in the non-many-worlds stories, like the Bohmian story, I think a lot of this is domain-dependent prior, right. Because I think one of the responses to the many-worlds, it’s like, that’s just insane, right? Like, come on. And I think that’s an appeal to, let’s say, common sense in the same way, and I keep going back to this, Aaronson, Scott Aaronson is an example of reasons why I won’t read your proof, and again, Scott, this is a joke blog post, right? But Scott’s like, it would be ridiculous if that’s how you prove P not equal to NP. Like, come on, right.
1:26:36.6 SC: Yeah, I’m just not going to take it seriously.
1:26:39.2 SD: So I think that’s a domain-dependent story there, so I think that’s another piece. It may also suggest why people have been so resistant to the many-worlds theory over the years, is that we can convince each other that something’s beautiful, it turns out, we can say, okay, this is more beautiful than that, and we learn these values. It may be much harder to push around the tacit knowledge, it may be much harder to push around these things that we’ve learned that we don’t even know we know, like we can’t quite articulate.
1:27:06.6 SC: I think that’s true, but I also think that even when people share the tacit knowledge, there’s a way they balance the different values is… My favorite thing that I got out of what you’ve been saying here, I’ll give you one last example that you can run with or not, as you choose. Lee Smolin was very recently on the podcast, and he said something, not on the podcast, but years ago, that always really, really struck me, ’cause I couldn’t… It was clearly true, and I couldn’t explain it, and maybe you’re helping me explain it. He said, “Isn’t it weird how people who believe in the Everett interpretation of quantum mechanics also always believe that computers will some day be conscious?” And it’s not because Everett implies that computers will be conscious, but the kind of person who’s like, yes, just give me the simple rules and I’ll derive everything, is likely to buy both Everett and buy that consciousness is substrate-independent.
1:28:04.2 SD: That… This is a lab experiment one could do on MTurk, if you could get enough physicists or the IRB to actually talk to one. I think this, I don’t know, Sean, again, we could go on forever, but one of the things that I think we’ve missed or it’s a big opportunity, is to look at these more, let’s call them exalted forms of reasoning. We tend to look at minute level snap judgments that people make, this tells us a lot, but we’re missing the culture of explanation-making and that’s, I think… I mean, I love this example, I think it would probably hold up, we should run it at the next APS, I’ll have to ask people.
1:28:44.6 SC: Do you know about the philpeople.org, the website for philosophers, so David Chalmers and David Bourget, I guess. If you’re a philosopher, you can have a profile site on philpeople, and one of the fun things they do is they ask you your opinions about all sorts of hot button philosophical issues, so you could totally cross-correlate those. The data are there.
1:29:10.5 SD: Oh, my God, that’s amazing. Yeah, no, the data is there. I mean, that… I will do that this afternoon, Sean, because that’s a great example. I’m sure there’s 50 questions, and philosophy tends to have this, it’s like, do you believe in analytic identity or do you believe… Or who knows, they have a vocabulary there that enables pretty simple binary answers, so…
1:29:32.8 SC: Well, yeah, all the questions are multiple choice, so it’s not like… There’s no essays, so it’s like consciousness, physicalism, panpsychism, whatever it is, dualism, and you know, quantum mechanics, Everett, da da dah. Yeah. Well, that’s good, I always like to…
1:29:47.0 SD: I’m sold. I love it, I love it, Sean, this is…
1:29:50.4 SC: I like to end the podcast on an optimistic note, and there is no more optimistic note than giving the podcast guest work to do as a research program, so I’m glad we were able to do that. Simon DeDeo, thanks so much for being on the Mindscape Podcast.
1:30:01.6 SD: Thank you, Sean.
[music][/accordion-item][/accordion]
Us good Bayesians have known how to explain explanatory values for decades.
While most academics are bayesians, I’m glad to the rest of the world coming along!
Pingback: Sean Carroll's Mindscape Podcast: Simon DeDeo on How Explanations Work and Why They Sometimes Fail | 3 Quarks Daily