Speaker 1Hello, everyone. Welcome to the Mindscape Podcast. I'm your host, Sean Carroll. In physics, you know, we have this distinction between fundamental physics and all the rest of physics. Fundamental physics is the stuff deep down, the quantum field theory, the general relativity, or perhaps even string theory or whatever, whereas all the rest is the emergent levels of the world, whether it's condensed matter physics or geophysics or biophysics or what have you. The analogous distinction in economics is between microeconomics and macroeconomics. When you think about the inflation rate or unemployment or the manufacturing sector or finance, you're thinking about macroeconomics. Microeconomics is the foundational part of the field that asks, in very general terms, how people go through the day making economic decisions. It's really about how people decide. More generally, to do one thing rather than another, simply indexed by spending money or something like that. So, of course, this question of how people go about making decisions is a tough one. So there's a big debate about the foundations of economics, just as there's big debates about the foundations of physics, for that matter. For a long time, the simplest thing to do was to simply imagine that people were perfectly rational, more or less. No one ever thought that, right? No one ever actually thought, going back to Adam Smith, that people were perfectly rational all the time. But there's a simple assumption you can start making. Imagine that people have preferences and there's some probabilities and they have a utility function, and then you can derive consequences for that for economic behavior. But in recent decades, or at least, let's say, during my lifetime, so not so recent decades, people have begun to take more seriously the idea that maybe we should describe the decisions making process that real economic agents go through in somewhat more realistic terms. Today's guest, Colin Kammerer, is an economics professor at Caltech, and he's been one of the leaders in this field for a long time. Sometimes the field is called behavioral economics, so economics that takes into account how people actually behave, amazingly enough. Colin has been a pioneer, the inventor, basically, of behavioral game theory. So thinking about game theory, which is another place where you could simplify things by imagining that people are perfectly rational, but instead, checking out how people actually behave and what their limitations are, imposed by the fact that some people make mistakes, our energy and stick-to-itiveness is not infinite when it comes to playing these games, etc. In more recent decades, Colin has also pioneered neuroeconomics. After all, if you want to ask how people really behave in these economic situations, maybe you want to know what's going on in their brains when they're making these decisions. So he is literally doing neuroscientific experiments, seeing what circuits are being activated, what modules are relevant when people are making these hard economic decisions. And it's, on the one hand, a field where they've discovered a lot of really cool stuff. On the other hand, it seems still a long way to go in terms of future discoveries to be made. So that makes it perfect kind of mindscape material, and it's a field where people are able to do a lot of things. So, let's go. Colin Kammerer, welcome to the Mindscape Podcast. Glad to be here. You know, I think I wanted to start by mentioning that, you know, a lot of people have this idea of economists, classical economists anyway, as people who think that everyone is perfectly rational and they just maximize their utility. And we're going to move beyond that. But even before we move beyond that in your work, you know, how accurate is or was that stereotype?
Speaker 2I think it's, I think of it as a limiting case, like you might think about in physical science or natural, in the sense that there may be people with extraordinary experience, like a wine expert, who tastes a lot of wine and is immune to herd behavior, you know, who can, the core of utility maximization is being able to rank every single thing. It's called completeness of preferences. So, I think there are people who can do that. But it isn't, it, and maybe even there's probably a little bit like everything, a little bit of genetic predisposition to be highly rational in this sense of having a utility number for all these goods and bundles of goods too. So, you make trade-offs between safety and smoke alarm in your house. And so, I think there's a lot of people who can do that. So, trade-offs, besides utility maximization, trade-off is really the trickiest part because it assumes that people can, you know, judge two things that may not be very sensually in some ways commensurable. So, I think it's a limiting case that occurs.
Speaker 1But do you think that economists like to pretend that everyone is like that?
Speaker 2I think there's an important distinction between the first few chapters of the book and what's taught and what's in the journal. And what's actually believed. So, I think of the model of utility maximization as a benchmark of rationality as sort of a starting point. Like maybe Newtonian mechanics where, you know, objects aren't moving too fast. And so, approximations work okay. And it's, in a way, it's a model of learning or it could be a model of the end point of learning, as I was suggesting with wine experts. But it's not an end point for everyone and for all people at all different ages with all different experiences and so forth. And it leaves a lot out.
Speaker 1Yeah. So, probably a lot of people took an economics class and got through the first few chapters of the book and they think that that's all that economics is.
Speaker 2Right. Right. And it is true that in analyzing lots of reward problems, that's sort of the default way of thinking. Yeah. You know, so, for example, I was reading a paper by Chad Jones about how do you price, how much should we be spending on existential AI risk to limit it? And so, very typical economist study, which is to think of a value of a life and this probability. And there's two things we care about. We like to have good AI, but we don't want to get wiped off the face of the earth planet, you know, and you make this. So, and that's a pretty good starting point because like, how else would you begin?
Speaker 1I guess so. I don't know. I mean, it's you've dived right into an area where I kind of don't have strong opinions other than I'm not convinced by it. I mean, I don't know. I don't know. I don't know. I don't know. I don't know. I don't know. I don't know. I don't know. I don't know. I don't know. I don't know. I don't know. I don't know. I think that that's, I'm kind of at sea as to what to do in those situations. Do you have a better
Speaker 2view? No, I think, I think that, I mean, Jones is very, very cognizant about that. And I think that the best analysis really are look at sensitivity analysis. In fact, in his paper, he has a, you know, like ablation, they call it an AI. You know, if you change this assumption one at a time or in bundles, you know, how much does it move around? And it may be that the right analysis right now moves around so much that it is almost worthless. I don't know. But anyway, that's one of the most exotic examples I could think of that was top of mind of where an economic economics mindset gets you started.
Speaker 1It's a good one. So by economics mindset, do we, do we, is that almost identical with kind of using game theory to model how people would go through these strategic interactions?
Speaker 2Yes. So except for there's a little bit of a historical diagram. So game theory is really a branch of applied math originally. So John von Norman worked on it when Morgan Stern was an economist and their famous book that introduced game theory or made a gigantic leap forward and a lot of the yearly work and it continues to be done by people in applied math. But it's also, you can think of it as a candidate for how people actually behave and not just people, but genes, organisms, nation states, et cetera. It's, I think it's in a way, it's closest thing we have to a unifying theory of social science, which is used in lots of different places because of the levels, you know, can be quite different. And so game theory kind of got co-opted or adopted or it became a project in economics. So now when most microeconomics is taught, you learn about supply and demand, equilibration in markets, collusion, monopoly, and all of those things assume there's lots and lots of firms. So no one firm has to think too much about what they're going to do or the same on consumer side, it's called price taking. But in game theory, suppose you have three firms, then all of a sudden it makes sense for each of them to think about what the other's thinking, what the other's thinking about each other and so forth. So game theory, most of the important advances have been done in economics or by combinations of mathematicians who get interested in fundamental questions, which have obvious application to economics, like Bob Allman, for example, is a PhD in math, but he won the prize in economics for proving some important theorems about foundational questions.
Speaker 1I think I didn't quite realize what you just said, but it makes perfect sense. If I can rephrase it, you know, there was important work done in economics by Smith and Ricardo and Marx and whatever long before game theory came on the scene. And if I take what you just said, it sounds like what game theory adds to that is the ability to model the other agents acting with you.
Speaker 2Correct, correct. And then it actually becomes quite rich because, for example, if you play a one-shot game like the so-called person's dilemma we might expect a lot of defection yeah um you know people being not being cooperative like um you know kids in class are all using ai because they don't want to lose out um uh but if it's a repeated game and people know the history and know that there's more periods coming you know like two companies in a small town right then um then the results really change in fact it's so well known it's called a folk theorem that you can get cooperation which you you wouldn't normally get mathematically in a one situation if you if you play repeatedly and then so in parallel lots more interesting mathematical questions arise like can anything be a kind of equilibrium is there ways to pick out subsets and certain kinds and so it really was a i don't know how much mathematicians sitting in math departments contributed along the way but it's certainly the foundational questions we're going to be talking about in a little bit often people who are quite mathematically sophisticated and some like allman really did come from pure math grounds where my colleague omer tamuz here at caltech is in math and economics
Speaker 1both and i'm going to apologize ahead of time because you've done so many interesting things we're going to breeze through them in a high level way and we're happy to pause and dig in at any moment but in some sense your starting point years ago was you know can we do better than thinking that people are are purely rational utility maximizing game theorists can we actually think about how people actually think is that
Speaker 2fair correct yes the more or less the mission statement um that i use which is very close to one from dick thaler who won a nova prize for pioneering so much of this work and as a co-author of mine is we're trying to incorporate natural limits on computation willpower and self-interest to improve economic predictions and make better policy recommendations and and dick dick richard thaler um wouldn't use the word natural like i kind of inserted that because i i want people to think of an invitation to think about biological underpinnings and mechanisms and neuroscience as well as uh essentially psychology and sociology and other adjacent disciplines
Speaker 1is there a way at the risk of oversimplifying is there a way of giving a pithy answer to the question how rational are people i mean how much does the model fail of making them as perfect reasoners yeah
Speaker 2i was at the university of chicago 100th anniversary of their phd program of which i'm a graduate and somebody asked that and i think i said 75 percent rational
Speaker 1but i love that answer uh well
Speaker 2i wasn't i was trying to be not too evasive you know and i didn't want to say it depends which is the correct answer actually yeah of course and i again i think they're interesting boundary cases so there are um there are cases where you have kind of extreme rationality like the wine experts are probably kind of if they rank two wines they're going to make the same ranking every time it doesn't matter where they are so they're they're impervious to what we call context or situations and there are other cases where you can move people around quite a bit um i'll give a quick example that might be a fun one so my collaborator eric johnson who's at columbia did a study where he asked people how how likely do you think you'll live until 75 and so people give a number like um you know 90 how how likely do you think it is that you'll be dead by 75 they don't say 10 right so the two numbers you know unless there's some weird thing happening right at your 75th birthday right the two numbers should add up to 100 that's
Speaker 1pretty basic yes
Speaker 2and it's an example of something which eric and many people have studied which is a sort of a kind of like an acquiescence bias like have you asked about living long i think i'm probably gonna live long if i ask about dying i think a little bit more about dying and so you get attention maybe attention memory you know some inner process is sort of perturbing the numbers it's not like we have a heart unless you're an actuary right or an account it's not like we have a hard number in mind uh which we actually believe and then can retrieve
Speaker 1and we're going to get into neuroscience at some point but it sounds like this is you know rich for psychology as well i mean do you find like traditional psychology very
Speaker 2useful in your work oh yeah yeah yeah in fact what i just described as an example is that for example it's called a framing effect yeah which is a term that tversky and kahneman um coined to denote when when the same thing we know is consequentially the same is is just slightly the language of it is just slightly changed and it can have surprisingly large consequences in some cases and they produced a whole string of interesting framing effects if you call a salad dressing 10 fat you sell a lot fewer than if it's 90 fat free um so there are many you know examples of these framing effects and obviously advertisers and right um political consultants and everybody are in this business too and another
Speaker 1aspect or maybe maybe it's related i mean you're the expert you tell me but people are how good are people at sort of thinking many levels through the game like i'm going to do this they're going to do that therefore i should do this or do they kind of get tired
Speaker 2of all that um yeah so most of the data we have is from lab experiments where we can design a game to kind of carefully titrate that um but there's some field data too let me walk through the lab experiments because the the basic one is very simple it's called a two-thirds of the average game or we call it keynesian beauty contest because keynes had an idea about the stock market that was similar the idea is everyone picks a number from zero to 100 simultaneously without talking you write down your number we're going to compute the average and whoever's closest to the average times two-thirds is going to win a fixed price all right so what people often start out thinking well if other people are just i don't know what they're doing so let's say the average of everyone else will be 50 i'll pick 33 right that's we call that one step of thinking right but other people think you know what other people are going to pick 33 i'm going to pick 22 that's two steps of thinking right and what we typically see in a distribution in fact there's many many replications of this experiment in the lab and also in a number of newspapers have a contest where you kind of write in and they give some big prize and so the new york times for example we have data from 50 000 people doing this and it looks a lot like students all over the world the original data in germany we have data from singapore etc and typically what you see is more 33s than 22s and not that many two-thirds of 22s which is around 15 so it kind of peters out
Speaker 1yeah and no ones no ones
Speaker 2unless uh you do it with an audience of game theorists and even then you don't get that many ones and zeros you'll get fives and tens so you know highly educated but there's two pieces to this too which i'm showing which is it's not just that they're highly educated about game theory they're in a group of other people who they know are highly educated by game theory yeah right and so it's it's the we call common knowledge or there's a shared knowledge of background um and interestingly there's some experiments with chess players for example you might think that chess players can do more strategic thinking but they look very much like regular folks oh it doesn't seem to be depth of thinking seems to be somewhat domain specific which is true of a lot of psychological things there's very limited we call transfer of learning you know that's not a very general skill at least so far as we can tell and i can
Speaker 1imagine explaining that result by saying people only think it through one or two levels yes there's another explanation that people think that other people are irrational so they're just gonna foresee that and can you disentangle those yeah
Speaker 2yes yeah yeah you can so there's a few other games that disentangle them and you can also manipulate who they're playing so nobody's done this exact study but i'm pretty confident what would happen if you did if you had a group of very elite game theorists and you said okay we're gonna run this on prolific or do a contest on an app and you're gonna play those people a lot of the game theorists who would choose zeros or ones among themselves will choose 22 or 17 some number in the new york times i think it's going to be a lot of people who are going to be playing those i picked 19 and the the winner was at the time was 20. ah pretty good
Speaker 1all right you're well trained at human irrationality that's something to be an expert at um and this this goes into what i think is called the cognitive hierarchy model is that correct correct tell us about that yeah yeah so
Speaker 2there's sort of two flavors sometimes it's called level k the cognitive hierarchy model just says there's some percentage of level zeros who are either choosing randomly or they may be doing something like something but the main thing is they're not thinking very strategically so in rosemary nagel's data she was the first person to do this in her thesis which i have in my office she actually asked people afterwards why did you pick the numbers you picked and so one person picked 18 and they were close to winning much like the year of times was 20 and the person said i picked 18 because i'm 18 years old right so that's not three levels of strategic thinking that's not right and we only found out because rosemary and i are not thinking about that right now so that's what we call level zero like picking a lucky number sometimes it's the it's the middle um in some games it might be um there's another interesting game called 11 20 games sorry the game theorists are not great at marketing um and in 11 20 you pick an integer from 11 to 20 uh and what's neat about it is um you earn whatever number you picked so there's a tendency to pick 20 yeah right so that's what we call level zero like picking a lucky number but if you undercut the other person by one number you get a big bonus i see of 20. so if i think you're going to pick 20 i will pick 19 and if i think you're going to pick 19 i'll pick 18 but it doesn't go all the way down to 11 because if we both pick 11 nobody gets a bonus so if i think you're going to pick 11 i'm going to pick 20 if i can't get the bonus i'll just get the highest you get what's called a mixed equilibrium kind of like a cycle and there the focal number we or the salient number we think that level zero is pick 20 so it's as if they have a model of the game but they neglect the fact that they could get more depending what the other person does they act like they're kind of playing by themselves so it's as if they they have a model but they truncate out yeah oh your payoff may depend on other people well i don't know what they're going to do i'll just forget about that um and uh anyway so that game is another way to try to pin down you know number of levels and usually it's something like two or three um the only exception which is a little bit interesting is and i'm in a little debate with this with george loewenstein who's a behavioral economist at carnegie is language so if you think about the iterated structure of language uh depending what you call a level right you can get things like um i know that you know that your wife didn't feed the cat and therefore you need to text your daughter blah blah you know if you write it down you could write it as games with a lot of levels in it but i think the reason people might seem to be doing a lot of reasoning is because of just learning and familiarity or what's like cognitive science is called schemas you know like we have sort of scripts for the way things work and if it doesn't work that way like a tree you know you have a branch plan b and plan c so so it's possible that um that you can get higher levels but um i think it's pretty rare and maybe a specialized skill or it may be in certain domains where it's very high leveled strategic thinking works
Speaker 1and you said the chess players don't have any special benefits elsewhere is there is there some idea that other than domains where we're special experts in each person is a certain level thinker like you know is it universal among different kind of problems that they would
Speaker 2face um uh partly but not quite yeah yeah so and we do know here's a few things related to that which is we know that when you play a couple of these simple games if you do it the way a psychometrician does it which is not just play one but five or six different games like you play the 11 20 game and then you play the two-thirds of the average game and then there are several others that have a similar structure and you kind of infer what level do they seem to be using you you those are correlated with working memory like how many digits can people remember with education so we have some data comparing caltech students with pasadena city college which is a mile away so controlling for geography caltech students as you well know are extremely good analytically and they're also very homogeneous everyone is there's no poets right and yeah um drama majors um uh so they're really good and they typically the median math act is 800 um we basically choose on verbal scores yeah because there's so many they're all good at standard stem um uh and they're they were they were about in that these are all data they were about one half level extra thinking compared to the pcc which was a little less than we thought so there's some yeah you know so there's some differences in in groups and things like
Speaker 1that have you i mean this is going to be a provocative question have you compared caltech to princeton and stanford and chicago students um
Speaker 2no but but i'd have two answers to that question the provocative question and i'm not going to conjecture what the truth would be but you could guess what i would guess uh-huh and maybe your readers can't um i was on sabbatical last year thank you very much provost at harvard in the fall at mit in the spring good and i think the difference is the harvard students are not necessarily selected on iq they're not i was a
Speaker 1harvard grad student i i i found that out
Speaker 2you know exactly exactly they are they do have a lot of um a lot of in 10 years i may i may be in the top 30 under 30 in wired math and i'm not going to conjecture what the truth would be but they're going to
Speaker 1rule the world that's yeah the
Speaker 2youngest congress person in my state ever um uh and you would think a little strategic thinking wouldn't hurt in those cases but it's it's not what they're necessarily going to gravitate toward um being good at um so
Speaker 1you're gonna you're gonna put the caltech students up there yes yes i'll tell you a
Speaker 2related thing since you mentioned that and i do have data on this or somebody else has data so there's something called the cognitive reflection test which is basically the cognitive reflection test which is basically the cognitive brain teasers and uh one of them is the bat and the ball problem which is more famous i'm going to use this a trickier cool one that's called the algae pond problem the idea is there's a little bit of algae in a pond the amount doubles every day it covers the pond in 48 days when on what day is the pond half covered okay and the answer is not 24 if the algae was growing linearly yeah 24 would be the correct answer so it's the right answer to a question that's almost the right question right right it's what we call level what what danny goneman who coined this term called system one thinking it's like a snap judgment um and the right answer is 47 because if it doubles every day it has to be half of a full one day before 48 which is 47 and so shane frederick who concocted this he's a professor at the university of yale um collected a bunch of data from lots of schools including harvard and uh not princeton i can't remember everything that's on the list um and also he went to the boston common uh and just collected answers there yeah and the boston common is a big um park in the middle of in the middle of boston which is historically very important i went to a political rally there it's really super cool um and the harvard students were a little bit below boston common so on the cognitive reflection test there's three questions they're all a little different but they have the same structure with brain teaser structure right like there's an obvious wrong answer and then a subtler right answer and uh so the scale goes from zero to three the mit students i think were 2.28 or something like that we did some at claremont they were around maybe 1.6 and harvard was a little bit below boston common but as the boston common students were a little bit below boston common i asked shane about this and he said oh no no but but the boston common people the the day the day we went was an orchestra or something and they may have been above average bostonians it was the boston common goers are not a representative sample of bostonians like okay okay you did not
Speaker 1want to accept the result okay anyway anyway so the the pro the program then is okay people are not fully rational but you don't want to just say they're not fully rational you sort of want to dig into what they are doing right i mean there is a process going on in people's brains that maybe even is optimizing something but it's other than being a perfect reasoner correct correct exactly
Speaker 2and and from the get-go in behavioral economics i started in the 80s the people we always confronted with what's your theory what's a better theory and that's exactly the right scientific question so there is a kind of fringe of people early in the field of behavioral economics who were i call them hecklers i was like well my cousin you know um brought the wrong insurance plan therefore we've disproved everyone being rational all the time well like great um what's the what's the replacement theory right and the replacement theory has been essentially a series of sort of boutique theories revolving around different kinds of computational challenges so the cognitive hierarchy is one theory in game theory which has gotten quite a bit of traction including in linguistics which is a cool area and um a field called adversarial control which is basically you know two fighter jets are trying to patrol and scare off another enemy fighter jet like how does that what what are the trajectories they take and so people have applied level k type of modeling to that um so that's been pretty fruitful uh and oh let me briefly mention the field data so uh we also have a data with experiment with field data not an experiment it's a actual thing that happens so one of the implications of this level k cognitive hierarchy type thinking is is that people will not fully infer uh information from somebody else's actions for example in from 2000 to 2009 we have a sample of 600 big screen wide release movies in the u.s and other countries and in those days it was rare but occasionally but if if you see a movie poster or an ad for the movies it used to be the old days you'll remember some of your readers who are young won't um it used to be that friday reviews sometimes on tv and the newspaper were a big deal it kind of guided what people and so the studios would send movies screeners to the critics maybe on monday or tuesday they said send us your review even if it's bad we may take a blurb out and we'll say you know stunning a magnificent cinematic achievement or interesting unusual if it says if it says one word that's bad yes right? Like, like, nothing you've ever seen before. Right. Um, anyway, and so it turns out about 10% of movies were cold open, which means they were not shown to critics. And so if, and it turns out the cold open movies are the worst ones because critics later review them when they do come out and they get much lower ratings on what's called Metacritic, which is a nice aggregation of, of critic reviews. Um, and so we wrote, we took an aspect of this model that says some percentage of the movie goers think that if there's no review, that doesn't mean anything. That means it's an average movie, but they're wrong because, because the less average movies. So the fact that there's no review is bad news. In this case, no news is bad news, but psychologically it's a little tricky. It's kind of like, um, Sherlock Holmes, the dog that didn't bark, right? Like if I didn't see a review, eh, you know, maybe it got lost in the mail or they didn't have room in the paper, blah, blah. But like, if you step back and think about it a little bit. And so the prediction we made was that these cold open movies, given their later critic rating would do unusually well because you're kind of pulling in people who didn't realize that no review is a bad sign. And you can see that in the data, there's about a 15% increase in box office gross from this kind of sucker.
Speaker 1Because I did figure out, I did once at a project, we never published it, um, with Mark Wise, my colleague at Caltech, where we, um, ranked, how much money Hollywood movies make versus their, uh, budget. Oh yeah, yeah. And, uh, what we found was that Hollywood was actually extraordinarily rational as an industry. So two things, one is the money you make per dollar is the same, whether you make a hundred million dollar movie or a hundred, a hundred million dollar movie, $1 million movies. But the dispersion is also the same because if you're in the $1 million movies, there's a lot more dispersion, but there's also a lot more movies. So the average is the same. And, but they know how much money they're going to make from a hundred million dollar movie most of the time.
Speaker 2Yeah. Yeah. Yeah. Yeah. Typically the, often a couple of days or a week before opening, you'll hear it open and we expect it to open this much. And so it's some combination of analytics and then they do screenings and they have people fill out these cards and, um, and usually they're, they're fairly accurate that way.
Speaker 1So they know that what, which movies are going to do badly and they just cold open those. They don't want, they don't want. Exactly. Okay. Yeah.
Speaker 2Yeah. And the, the paper's fun. Cause there's some, um, a famous one is Pluto, uh, Mar Pluto Nash by Eddie Murphy. And Eddie Murphy was a big star at that time. And so this is where this dude is often get in trouble. It was kind of like, well, you made us a lot of money, you know, maybe you'll lose us some money. That's like, all is forgiven. Um, anyway, another one was called killers with Catherine Hagel and, uh, Ashton Kutcher, which, um, that one,
Speaker 1I don't know. Yeah. You know, else because, right.
Speaker 2Probably for the best.
Speaker 1Yeah. Yes. 21, I'm going to correct. So you, um, so you mentioned Kahneman and, uh, Tversky and system one and system two, how does your picture here, uh, fit into that? Is it, is it like a, uh, compatible or is it a modification of that?
Speaker 2Yeah, I think, I think, um, and by the way, I'm a little bit of a system of system two skeptic in the following sense. So, um, uh, when Kahneman, wrote it, his Nobel lecture about this, and there's lots of other people have studied it. He made a list of traits that system one has associative, fast, effortless, and system two is parallel serial, um, slow, deliberative, you know, extrinsic. Like I can tell you what I'm thinking. The snap judgment case, maybe I can't tell you. Um, and that's not, that's kind of pre-science, right? We don't want, we don't want a neural system to be defined by a list of words. We want to see it in the brain. Like the brain evidence is very good that there's something called the default mode network, which is what you're thinking when you're not doing anything active. It's the salience network, right? There's a number of things like language areas we know quite a lot about. And I remember asking Danny once, um, early, maybe five years after system one and system two was sort of in the air, have you presented this to neuroscientists? And has anyone found anything that looks like system one, system two dissociation? And he said, he had a very distinctive accent. He said, I, I, I did. I, I thought I was giving them a wonderful gift that they would all want to go study. Um, and part of it was neuroscientists don't believe in that much reductionism. Yeah. You know, there's an ongoing debate. You may know some bits of it about modularity from photo or, and you know, it's the cost me to send to me once say, we think there's 78 different modules that are adapted and modules to answer challenges. And, you know, there's lots of schemes for deciding like how many distinct parts of the brain there are. And so I think the, the neuroscientists felt like two is a too small of a number.
Speaker 1I see. Okay. And so it never got much traction.
Speaker 2I mean, I wish people would work at it because if it is two, that would be wild. And it may be, there'll be, for example, that system one is available in infants and system two shows up in puberty or when you're 25,
Speaker 1you know, some people never. Yeah. Yeah. Yeah.
Speaker 2You took the words out of my mouth. Um, so, but that's
Speaker 1a perfect segue because, um, you are an economist who does neuroscience and, uh, I mean, let, let's just, you know, ask the sort of, um, uh, career path question here. Was that an obvious thing to start doing as a professor of economics?
Speaker 2Uh, no, no. Um, yeah. So the pattern is, um, I like novelty and I'm, and I'm a bit of a dilettante, although, um, maybe, uh, I want to be more than a dilettante in the sense that, um, one of my proudest facts about my career is that I've published in a lot of different journals sometimes on my own, but also with typically with a tutor, a senior collaborator, who is my sense. I, that's the way to do it. Yeah. What I need to know. Um, and, um, so the first thing I worked on after graduate school that was a little bit ahead of its time was experimental economics. So the idea of doing an experiment came very late to economics. In fact, up to 1985 and maybe still Samuelson and Nordhaus's famous book that a lot of people taught from said, you know, it's sadly like astronomy and, um, uh, you know, geology, we can't do experiments. Like we can't make artificial earthquakes and create new planets, uh, in economics. And even at that point, there were hundreds of experiments already published. So I don't know what they were thinking. Um, it was like a hallucination of the human form of hallucination or denial. Um, anyway, the idea came quite late. And so I got interested in that because Charlie Plott, my colleague at Caltech, who's a major figure, really one of the main founders in this area, taught a course at university of Chicago. And I learned about experimental economics. No one was doing it. I liked doing things that no one else was doing. Yeah. I don't really like competition. Yeah. One way to do something no one else does or to combine, synthesize things, right. That are unusually unusual mixtures. Um, and then I get interested in behavioral economics. Um, and that was also controversial at the beginning and it took a lot of, you know, there were constant like battling with, um, referees. It was, it was kind of healthy because if you have, um, scientific enemies can be very good for you. They read your papers very carefully. If you make a mistake, they will tell you, you know, you're wrong. You're wrong. You're wrong. They'll tell you about it. Uh, if they think there's a, they probably are willing to have a conversation that says what evidence would convince you, you know, like, well, people do these games for $10. If you did it for $200, I'd be convinced. So if then you go do it for $200 or you do what we did and many people, you go to Vietnam, which is highly literate and poor at the time, 2010, and you do it there for the equivalent of 10 us dollars, which for them might be two days wage. Right. So as Thaler likes to say, the easiest way to win an argument is to run another experiment or run another regression. So we were always in, you know, grit mode. Like we're going to prove to you, prove to you. Okay. Here comes neuroscience. So in 1997, uh, George Lowenstein and I were together at the Institute for events, uh, sorry, not the Institute. That's Princeton. At Stanford, there's something called the Center for Advanced Study in Behavioral Sciences, CASPS, C-A-S-P-S that began in the fifties. And it's, it's an amazing institution. It's a great place to go. Um, ironically, they are always begging me, do you know anyone in economics or neuroscience who wants to come? Because most of the neuroscientists have big ongoing labs and they feel like they can't take a year away. Um, even though I think their trainees probably wouldn't mind to be honest, neuroscientists out there. Um, and there's, and there's flights from San Francisco very frequently. Um, anyway, it was at this CASPS and George Lowenstein and I were there and also Drajan Prelek who's a brilliant, um, psychologist who contributed to behavioral economics and his wife is a neuroscientist. And so George had gotten interested in neuroscience a little bit because he was good friends with John Cohen, a major cognitive neuroscientist who also kind of joined us in neuroeconomics with some of his projects. And, um, so between Danica Prelek, John Cohen and other people we knew, George and I organized a meeting in Pittsburgh that was basically, you know, our friends are going to meet your friends. And so there was a few economists, none of them were doing science they were just like i'd like to hear hear what it is and how do you study things and what are the questions economics like utility maximization what would how would you try to study that thing and it was quite an amazing amazing conference the except most of the economists were dead silent so i think they were listening number one okay um and they didn't know what to ask right um economics is a little bit of a proud tradition with some famous exceptions of like it's it's considered low status to say i don't really understand this oh that's too bad or how would this i'm interested in common knowledge in games how would this help me study that you know occasionally people will ask but anyway so they were just quiet um and by the way in that group was peter diamond al roth paul roemer uh and i may have left somebody at all three of whom later won nobel prizes now did we have a causal effect i'm just saying it's a small sample you know we don't have much statistical power and actually no none of them um pretty much did anything actually paul roemer got quite interested in some aspect of it he wrote a small paper on emotion and so um he's the one who kind of came closest but then he got distracted by lots of other things um that were more important for him to do frankly so they were kind of like sympathizers yeah you know tell us um and then start a conference at princeton which i missed uh and angus deaton was there later nobel laureate and angus again has had very far from inner scientist he's interested in demand systems you know how do you fit together how much products people like and whether these substitute products and and he had done some beautiful work on that but he basically said you know i came to this conference because we don't i don't think we really understand the source of demand like i have all these data but why do people like vanilla not ice cream it doesn't change over time does it depend on who is around them that's a great question can you guys help us understand that um and and you know if you if you taught a course on demand theory there's not a lot of neuroeconomics papers i would put on there right now but we would we would have a research agenda for example and we could say something about a few things that have been studied like self-control like when do people buy tempting goods rather than healthy goods um we have some data suggesting his dorsolateral prefrontal cortex interacting with another region but it depends on how much your brain is sort of encoding or representing health value okay so so there's some bits of cognitive neuro that speak to things in economics but it's just a very thin web of tissue it's not a nice dense um plan anyway so so the course of that i moved to caltech in 94. so this conference in 97 was a little bit in and i started to ask around to about neuroscience people i knew one was john allman who turns out as one of the world's experts on cross-species comparison of primates and an incredibly wise person so he was kind of my tutor and um uh and then we got his brain scanner um in 2003 and we had a kind of a volume discount so there was one scanner to study monkeys which was vertical because the monkeys don't like to lie down and they're squirming so you have to lift them up into it like a kind of hairdresser you know and there's another one for small animals which have tiny brains you need a big seven tesla and then they had we thought well when i was going to be a human scanner but as you may well appreciate from knowing caltech in biology now called bb biology and biomedical engineering very few people are interested in humans yeah because we're too complicated a biological too messy yeah of course yeah too messy right if you want to study a nicotine addiction like that you're not going to be able to study a nicotine addiction like that if you want to study um sleep and development you you want a an organism that's transparent when they're born so you study a zebrafish larvae with david proder right and so humans are like man you there's too much going on so we had this kind of field of dream scanner that no one was using and the staff people were really terrific and and a couple of us actually peter bosarts was one steve quartz was another said i have some ideas about how we could what we could do with the scanner and so people started to work on that and and some of it was a very caltech formula as you might recognize so peter started out as a finance professor and he got interested in risk and reward and he was like well what is risk and reward in the brain like fundamental thing in economics and he was just he was good at it and he like me was very adventurous like i don't care if i can't publish these papers i have 10 that's what 10 years for steve quartz was a wet philosopher what that means is he had a piece in philosophy he did a postdoc at the salk institute of la jolla where terry sanyofsky was and francis crick and um steve was there at the same time as reed montague and peter diane and so montague and diane and wolfram schultz wrote a very famous paper in 1997 about dopamine and that dopamine neurons were firing in midbrain as it as a reward signal reward learning signal anyway so steve was around this kind of computational neuroscientific and so he was still teaching philosophy like he would teach a horse and like will like free will you know um but he was he was interested and he'd like to dabble so he and he and peter wrote some papers i wrote some papers with them and we kind of got a start um uh and and caltech by the way was quite supportive partly they had this free scanner they just wanted to get something out of it right so for a few years it was free to use the scanner it's normally quite expensive basically subsidizing you know subsidizing a startup um and they were pretty charmed by it so so i was able to do it and i had good collaborators around
Speaker 1yeah i mean that sounds great the one thing i wanted to note when you were talking about organizing the interdisciplinary conference uh there's a piece recently by paul bloom who's a former mindscape guest um who it was very he was trashing interdisciplinary conferences not because the idea is bad but because you would invite like the best people from area one and the best people from area two but he said those aren't who you should invite you should invite the people from area one who care about area two and those might not be the ones you've already identified as the best
Speaker 2yes yeah yeah no i totally agree with him in the sense that it's very easy to punt on a mediocre interdisciplinary conference that's a waste of time that nobody would want to come back for part two exactly what he's putting his finger on which is you need people who are good at what they do like really good willing to learn language and communicate across languages like the way rational is used seven different ways and so that takes some you have to be in kind of student mode like you have to toggle from expert to student and then these are people for whom this could really be a priority like if this meeting is fun i want to go every year or a collaboration goes out of it or we write a joint grant that you know at first that for the those lenders that really like these kinds of mixtures like john templeton and sometimes darpa or whomever
Speaker 1i like the idea of student mode because it's depressing how many professors don't want to get into student mode like they figured they've they've paid that price already they don't want to do that anymore but it's kind of a shame but yeah okay so um this is a very broad question but we can narrow in what do we learn by looking at the brain is it something like oh when you're making a decision this part lights up or can we actually start identifying some of the things that those modules you mentioned
Speaker 2earlier uh yeah so like i so first of all let me knock down a couple of kind of myths or things that people often say good so one is that very few things are highly localized to one part of the brain right so so we do think about certain regions like amygdala or dlpfc i mentioned but typically we're looking for kind of a circuit you know um and so that's number one uh there's some exceptions like language is largely left lateral localized um uh for so there's left side of the brain left side of the brain yes exactly exactly and um and so um so there's some cases like that but even then the brain has a fair amount of plasticity so if you have damage in an area or with animals we can sometimes lesion the animals um uh people that people and animals can often like the brain will reorganize itself to have some functionality in a different part of the brain yeah there's a very famous uh paper over this on ferrets optic nerves and an auditory i believe where you kind of rewire um and you know there's even you know the people have hemispherectomies where they remove entire hemisphere they can do a lot of things so the brain has a certain resilience um it you know depends when it happens if it happens early in life it's better and so on and so on um uh but so one distinction i think is important is um oh so let me give you one concrete example that i think is builds on what we were talking about so when it comes to strategic thinking the terms that i prefer is mentalizing another term people use is theory of mind which is the idea that i have a it's not my favorite term but you know um that's why i like mentalizing theory of mind is the idea that that i have some kind of theory or personal representation of what's in your mind yeah okay all right and there's a lot of philosophical nitpicking about which is interesting actually about how it's works but that doesn't help guide us too much on the neuroscientific discovery uh it's been it's been a lively debate and live debates are usually good um but there is pretty good evidence on the mentalizing circuit so part of it is what's called right temporal parietal junction. So there's four main lobes of the brain, temporal, parietal is motor, occipital cortex in the back, and then frontal. And so temporal parietal junction is sort of a region that where the temporal and the parietal lobes tend to meet. There's part of the men's leg, the circuit is dorsomedial prefrontal cortex, dorsal is on the top, medial in the middle. So that's, you know, an area like a big quarter, you know, on the top of your forehead, an inch above your eyebrows. There's some regions in temporal lobe and posterior cingulate. And so if we do what's called a localizer, like if I show people a bunch of pictures, like my colleague Ralph Adolph does, and I said, this person's move, what is this person doing with their arm? You would say reaching. Right? Right. And I could give a similar picture and say, why is this person moving their arm that way? So the how question is not about the contents of the person's mind. The why question is about why are they doing it? Or you could show a person who's crying and said, what is this person's face doing? It's crying. Then you ask, why are they crying? Okay. And so if you run this why, how localizer, and you look, what are the areas of the brain that are more active in why than how, and just to help your listeners do it. The reason we do both is because whenever you put something on the screen and you have people press a button, tons of stuff is going on.
Speaker 1Everything goes crazy.
Speaker 2Yeah. Right. And so we want to, we want to have two things that are very, very similar in terms of what buttons you press, what you see on the screen, the luminosity, the emotions. So you might, so they literally would use like the same pictures that permit either a how, why judgment. And then you subtract the how activity from the why and what's left over we think is mentalizing. Okay. And yeah, and there's a couple, and there's a couple of studies, but not very many, including by us. Um, there's one nice study on the twos of the average game and people who do two steps of thinking in two thirds of the average, um, have more activity in mentalizing circuitry. Oh, okay. Right. And I wish there was, for some reason, including my laziness there, I wish there was a lot more studies on that. There's a few, and they all kind of point to the side of it, more strategic thinking as evidenced by the two. these kind of extra levels, you know, like 22s or 18 or 19 in the 11-20 game. Those extra steps of thinking are recruiting this mentalizing.
Speaker 1And how well can we, I mean, so sorry, let me back up a little bit. Given that you're able to do at least some connection between that way of thinking and something happening in the brain, what kind of story can we tell about why we think that way? Why we're not better at multi-level thinking? Does it come down to energy usage or just accidents of evolution or satisficing? I'm not quite sure.
Speaker 2Energy usage is always a good candidate, right? You know, because the brain runs on 40 watts and what evolution is generally doing is thrifting, you know, it's trying to sort of offload, to splice in approximations, offload things. There's a couple of papers. I don't know if the fossil record or anything, I should ask John Ullman, right? Actually, he's often at Santa Fe. You should ask him. I'll ask him, yeah. Whether there was some kind of rate limit, you know, that kept human brain from developing more steps of thinking. Probably it's something like there's a bit of a kind of mental arms race where if I could do more thinking, I can figure out if you're fooling me and I can fool you. But unless you're a very social, species in which that's highly advantageous, you know, you just wouldn't get enough evolutionary pressure to generate more. It's a very good question. If I had five PhD students writing theses, we would- One of them would be. All you in three years.
Speaker 1Okay, good. Well, you know, there might be people out there listening who can apply to grad school at Caltech and do that. One of the obvious ways to me that people fall a little short of being able to do that is they fall short of being able to do that. Perfectly rational is when we're trying to compare what we want right now in the moment to what might benefit us down the road, right? You know, planning for the future. Is that something that you can see in a distinction in the brain or is it-
Speaker 2Yes. Yeah, that's a great topic. So in psychology, it's often called delayed discounting. How much do you discount or underway things that you get with a delay? We often call it time discounting. You know, and interestingly, animal learning, biology, psych, and economics have cohered on a fairly standard set of definitions and tools for measuring it. And a very typical one is called smaller sooner, longer later. So the idea is you could have $10 now, or in a month, I'll send you $15. Yep. And people say, well, I want 10. So then we'll usually calibrate that. You can do this in a Bayesian optimal way. So it's- It's like, well, how about 10 or 18? Like, I still want 10, 10 or 20, right? So you can titrate this and then figure out what's an implicit interest rate. So we call it a discount factor, but it's exactly like an interest rate. And there's quite good evidence about how that works in the brain. And again, dorsolateral prefrontal cortex is quite important. So there's something called perspective thinking, which is basically when people are thinking about the future. And the perspective thinking system, which is fairly well understood. One of the early people who worked on it, by the way, was Demis Hassabis, who's now a DeepMind Nobel laureate. So a number of people in his circle were interested in this. And so we have a study related to this, which is the food study I alluded to briefly before, which is people see pictures of food and there are things like apples, yogurts, Snickers bars. And they rate them on health and taste. And each individual rates them. We don't take some rate because some people love tofu, some don't. And then we ask them to choose between goods. And what happens is when they're choosing the tempting one, there's more activity. Sorry, when they're resisting the tempting one, right? So they see tofu and Snickers. And when they choose tofu, and these are people, by the way, who recruiters said they're trying to diet. So they're people who we know have everyday self-control struggles. We didn't super carefully titrate what that means. And this is pre-GLP-1. So this is going to screw up all these studies. Actually, I don't know if it's going to screw up the studies. It becomes a really interesting thing to study separately. Interesting to study, yeah. Right, exactly. Anyway, and so you ask people about these healthy foods and then more tempting foods that they like better. And when they're able to resist temptation, there's activity in DLPFC. So we think of that. I think of it as it's also involved in working memory and lots of other major attacks of what's called executive function. So I think of that as a place that reminds you now about future consequences like health. And there's another really beautiful study on this topic. So one thing people have studied is called EFT, episodic future tagging. And the original study is so interesting. And it's really interesting because they took a calendar and they asked subjects, what are some things that are happening in the next couple of months? Like, oh, I have my PhD defense. I'm going to Santa Fe. I'm on Sean's podcast. And they had the rate, I'm going to go to the dentist. Are they good things or bad things? So how vivid is it when you think about it? And is it good or bad? And then what they did was they did the SSLL experiment. You could have $10 now or $15 in a month. But that's the control. And they did the SSLL experiment. And they did the control treatment. In the episodic treatment, they said, you can have $10 now or $15 when you're going to be on Sean Carroll's podcast in a month. So it's the time is fixed. But again, there's a kind of framing difference where you've injected this vividness, the salience. And then people are more patient. So the idea is that the episodic imaginability through perspective thinking kind of forward mentally. And it changes the discount rate.
Speaker 1Is it that or is it, I mean, in part, when you say something like, I'll give you $10 today or $20 in the future, part of me just doesn't believe you, right? Like, you know, maybe you're faking or maybe you'll forget or you'll get poor. Whereas if we tie it to an event, then your reliability goes up a little bit.
Speaker 2It could be. It could be. And actually, this thing about credibility of future rewards, a lot of people have tackled. And my colleague here, Charlie Springer, who's a really terrific experimental economist, the way he did it was to give people post-dated checks when they left. So if you pick $10 today, you get a check 24th of September. And if you pick one month, it's October 24th. It's a little sus because sometimes banks will cash post-dated checks and they would have you. But anyway, they were really cognizant about this. And another study by Paul Glimcher, who's a terrific neuroeconomist, what they did was they had people scan three or four sessions. And so by the time they did the second session, they had gotten their money from the first session. Okay. That was a way to build up credibility. Okay, good. Yeah, yeah, yeah. But this is kind of like the best way to win an argument is to run another regression or run another control experiment. So people have been very careful about this credibility thing. But it may be the episodicness somehow adds some vividness or something. I mean, we need to ask the brain more.
Speaker 1It does sound like this might tie into – I know you're doing recent work on burstiness in activity. Like if we're doing something, we'll keep doing it for a while, and then it'll just go away. Yeah, yeah. So what have you found out there?
Speaker 2Okay, yeah. So that's a great segue. Thanks, Sean. So burstiness is – well, let me tell you a story. So we have some gym data, and we have some data on about 1,500 days of gym attendance. These are people who join the gym, and then we have four years of data or so, sometimes less, sometimes more. And it's a lot of people. It's a lot of people, like 25,000 people. And so we were interested in whether you can predict when they go to the gym. Like do they form a habit of going regularly? And so my RA, Hong Ho, made a graph. I said, for giving a talk purposes, make a graph of one person's calendar year of when they went to the gym every day. And so I looked at this graph, and I thought, put some non-homogeneity.
Speaker 1Yeah. Just what I would think, of course. As one would. As one would.
Speaker 2Yeah, yeah. And that's an extremely clumsy term for burstiness. Yeah. Okay. And so on the graph, what happens is the guy goes one day, the next day, next day, next day. So he's going one day apart three times. Then there's a seven-day gap. Then he goes the next day, one-day gap. Then he goes 13 days later. And so there's a lot of short gaps and a lot of long gaps. And it turns out that the kind of default model for thinking about these kind of stochastic processes is often a Poisson arrival process. And what Poisson non-homogeneity means is if the Poisson parameter, Poisson parameter readers is a Poisson process is a gorgeous process invented by Mr. Poisson, which is French for fish. And it has a lovely, lovely structure. And it's characterized by one parameter. So there's a parameter tau, which is the mean of the distribution. And it's. So it's very easy to work with it. Like Gaussian, you have two parameters and, you know, lots of other beautiful functions have two or three, but, but, you know, for some purposes, one is very handy. And we also use it in the cognitive hierarchy, by the way. So it looks like the, the distribution of levels in these games, like two thirds of the average happens to have a Poisson shape. It's two different. It's the same Poisson, just completely different. It's just a good example of I'm like a platonic, you know, sometimes the same ideas show up powerfully in science. It's in many different places. I hope so. Like God, some of those makes our life easier, you know, rather than harder. Anyway. So, so back to the calendar. So, so we found this beautiful book by a guy called car side that was about 200 pages on burstiness. And so burstiness is another name for these lots of smalls and, and lots of longs. And the burstiness has a property that if you, if you regress the percentage of one day gaps. Against one day, the percentage of two against two days against two, right? You regress the percentage of days of length tau against tau, but take the logs, log percentage against log tau. And I did this on my computer after I got these data. I was so excited. It was 10 years ago. Claude couldn't help me. I basically counted by hand the size of the gaps and I made a plot, you know, I was really excited. It's like, I'm discovering, I'm discovering, uh, here comes. It's dopamine flowing. Um, um, and I ran a regression and it was a R squared of 0.97 and in social science, you only get that when you're regressing by mistake. One thing against the same thing or something that's really persistent, like GDP, you know, quarterly GDP is last year's GDP times one plus some small perturbation, you know?
Speaker 1Um, so sorry, just to, just to interrupt, but you know, for the, for the listeners who are, are, you know, might. You might have missed the forest there for the trees. So you're saying that there's essentially a perfect kind of power law relation here in the, in the frequency of the gaps between these events.
Speaker 2Correct. Correct. In the log frequency and the log length. Exactly. So if you, if you, I mean, like I said, we just never, we never see this, you know? No. Um, and, um, uh, so my view, especially after reading this car side book, which is just a gift was, um, um, and again, this is not that common in economics, but was, if it's that lawful empirically, there must be a simple explanation. Yeah. And it turns out there's not just one simple explanation. There's a few.
Speaker 1That's the problem with the simple relationship.
Speaker 2Yes, exactly. Exactly. Exactly. So it's like, um, so we have four, these four data sets and there's a couple of beautiful models we found in the literature that actually go back to 1954. So let me give you a quick tutorial to your, our beloved readers, uh, listeners on queuing theory. So queuing theory is a branch of applied math developed, like think of like a lot of boom and operative research around World War II or thereabouts, uh, which says if people arrive for some kind of service, these are the words that people still use. And the arrival rate is stochastic, but it's governed by some law like Poisson arrival. And then the service is stochastic governed by some kind of law. Um, what is, what are the statistical properties of how long people wait to get served? What's the mean? What's the median? If the arrival rate is faster than the service rate, it goes to infinity. Um, there's a lot of interesting applied questions. Like if you're at Vons and there's more than three people in line, there's four. What they do is they'll open up another server. And so you can study if you have a certain, you know, a server open up rule, like how does that work? Most of the queuing is simplified. If you use first in first out. So somebody arrives and nobody gets to cut in line, but it turns out starting in the fifties. And then later, um, uh, Barabasi was a famous, um, person working on this who people may have heard of. He's written a popular book and done a lot of interesting things. And, um, it turns out if you tweak queuing theory to what's called priority queuing, you can get a power law. Okay. And so priority queuing means it's not first in first out. So let's say there's a line. And with some probability P you don't serve that person who's next in line, you pick somebody else. Okay. And there's three or four very beautiful forms of this. And the different forms of priority generate power laws with different exact numerical slopes. Like Barabasi's form, the Cobham form 1954 is minus one half of minus one and a half minus three halves. Barabasi is minus one. Uh, there's a couple of others that have slightly different. I mean, the, the, this, the, the, this is, and this is all published in like statistical physics journals. It's just beautiful. Like I was just, I kind of wanted to like, maybe I should have stuck with physics after I majored in physics for one term and then gave up because it was too hard. It was too hard and there was no people in it. But, um, anyway, it was extremely beautiful work. You have to make these nice approximations and it's sort of knowing what to take out, like in some ways. Right. Um, anyway, so there was all this literature on what to do. Meanwhile, we have a. Uh, a priority type model, but we hadn't thought of it as priority queuing at all. It's called neural autopilot and we developed this as the name suggests to create a simple mathematical econ style model of habit that we might be able to see in the brain. In fact, it was inspired by things that people have been writing about the brain. And the idea is, is that the brain toggles between modes. When you're in optimization mode, you go to a store, you look at all the things in the store, you think about the prices. People don't really do this again. I mean, this is where we started out with the children exhibition, but they, they do some approximation or they, if they go to the store at the same time, like the wine expert, they may actually know, you know, what all the different pickles are on the Armenian store, which there's about a hundred, um, in, in LA. Um, so that that's, that's regular shopping, but it's very expensive. You have to look at all this stuff. You have to compare the prices. You have to make trade-offs, right? Do I buy the big jar, even though I'm not gonna eat that many pickles, um, in habit mode, you do something much simpler. What you do is you, you retrieve a memory. What did I pick last time? And then you pick up one number, which the brings, which is the reward reliability, which is basically, was it pretty much the same as last time? So it's not the reward prediction error, which is a workhorse idea from, well, from Schultz, Diane in theoretical neuroscience, it's the absolute reward prediction error. It's like the, the sign is it been, been pretty good. If that's high, you don't buy the same thing. Again, because if, if it's high, it's like, well, I don't know, I don't know if I'm getting something that's as good as last time. Right. If it's low, you just buy the same thing. So think how simple it is. You, what did I pick what it was? It's reward reliability. In fact, those might be closely coupled together and associative memory. So it's a very, it's, um, and so the idea is in habit mode, in habit mode, you just think of these two numbers, you're in and out, you can multitask, right? So you're pushing your shopping cart, you're on the phone, you're grabbing. Um, uh, Italian wedding soup is your favorite soup, your Italian wedding soup, blah, blah, blah, blah. Right. you can have a conversation. So multitasking is a hallmark of habit, right? Because you have this freed up executive function. That's the idea. Anyway, so we thought, well, let's take this neural autopilot model and see if it can fit some of these power laws. So for example, maybe when people are going to the gym, sometimes they're in habit mode. And if they go to the gym one day, they're just going to go again and again and again. And other times they break that habit and maybe they're in a habit for a different activity like running or binge watching or binge podcast listening or something, right? But they're not in the gym habit. And so we were able to show statistically, it's horribly inelegant compared to the statistical physics, I'm ashamed to admit, but that's the limits of my talent. And we're able to show that the neural autopilot model will also generate a lot of streaks of ones and then some long streaks and it generates a slope of the power laws the slopes of the water. And so we're able to show that the neural autopilot model will also generate a lot of streaks of water. And so we're able to show that the neural autopilot model will also powers we saw in our data were mostly like minus two okay um which happens to be a little bit higher than most of the beautiful mathematic ones that generates flatter steeper slopes um and so this this was uh um it was again i i think this is what interdisciplinary is good for it's like if i seeing that picture and knowing what person non-imaginary was, which I learned from Peter Bossart, because it's an idea used in finance to think about insider trading, right? It's like, you know, you know, the trades don't always arrive in this put some way you may get a burst of trading, and then a lull and then a burst of trading. And that's either because information comes in lumps, you know, or, or trading is self generating, like if I see a bunch of trades, the price moves, and that gets my attention. So anyway, so a bunch of people had been developing this in finance, Peter Bossart's infects me, it plants the seed in my hippocampus. I see this picture, I think burstiness, I go look for a book, I find Karsai, thank you, Karsai. And then I go and talk at my MIT Harvard sabbatical, I talked to Xavier Gabex, who is an extremely brilliant macro finance behavioral economist. And he wrote a paper on power laws, and derived why cities satisfy city sizes satisfy a power law. Which is a very old application from the 50s. From Herb Simon. And I said to Xavier, isn't this, this topic is still interesting, right? Because nobody in economics seems to be working on it. Like he wrote two great papers and then just stopped. He said, Yeah, I couldn't get. I said, when this is so lawful, there must be a beautiful, simple process. Yeah, yeah, yeah, yeah. So we so we're the two champions of power law study in economics. I mean, his work was much more serious. And I just was kind of dabbling on the empirical side. Anyway, that's what that's about. I mean, the
Speaker 1I'm diving into this myself, because I'm writing a couple of books that involve these things. Yeah, the power laws are, on the one hand, they're everywhere. On the other hand, they're surprisingly controversial, because people are like, is it really not quite a power law? Or is it a law of normal? And what is your mechanism? But I think for sure, it starts you down a path that makes you learn things. I'm all in favor of people noticing power laws. Here and there. Yeah. Yeah. Yeah.
Speaker 2Just to sorry to echo your point. We wrote a second paper. So, Gebex has the only paper I know of that actually test uses statistical inference to test whether it's really a power law or something else. There's a few ways you could do it. You can compare power law with alternative specifications like log normal. But his test is, I mean, these are really good econometricians, right? So, they're really good econometricians. So, what they're thinking of is not just, wow, that picture looks pretty good, 0.98 R squared. But could we reject power lawdness without a specific alternative? So, they have this really elegant test. It's very delicate. I never published it. And so, we're the first people to actually test that. And indeed, they're not quite power laws. Let's just whisper. You just have to back up and listen again. Click the minus 15 seconds button. What it is is that log normal fits a little better.
Speaker 1Okay. But that is, despite the fact that it comes across as a little bit persnickety statistician talk, it might have important implications for the generative mechanism, right?
Speaker 2Absolutely. Absolutely. And this is really nerdy, but let me just finish on that. And if you graph them, like power law and log normal, what the log normal differs is that when the lengths are very short, like one or two-day gaps in the gym, one or two, three-day gaps in the gym, there's a little bit fewer of those short lengths than would be on the power law straight line. And then the tail of extremely high lengths, like 30 days they haven't gone to the gym, just kind of falls apart. It's what's called exponential cutoff. And it just starts to not look like a straight line because there's not much data down there. Not much data, right. Yeah, of course. It just fans out. But the log normality might, I mean, your point is, I'm so glad I sent you this paper last night and you read it or thought about it, or at least it sparked the thought. Because in psychology, this is what I call refractory period, which is basically recharging, right? So after you eat and you're full, you don't eat again. Sex, exercise, sleep, all these things, you don't expect to see a lot of really small things close to each other, right? So for example, I don't think this is true at the gym, but it might be, for example, after you go to the gym one day, your muscles are a little sore and you can't go for two more days. So there's a deep biological reason why you can't have short lags, right? Or we did have data, for example, on people checking their phones, you know, and you just don't, you can't check it minute after minute after minute unless you're- I hope not, yeah. Yeah, I mean, it's bad enough as it is, right? But I think your point is really well taken, which is the hard work of comparing these things that look very close together may tell us something about a deeper mechanism. That's not in our habit idea and it's not in priority killing, but maybe there's a way to kind of splice that in or to find a domain in which you think the refractory period may, you know, erase the, may bend the curve in some interesting way.
Speaker 1And so the burstiness, of our habits, is that something you can also tie into the neuroscience? You can sort of say, oh, it's this circuit in the brain that is responsible?
Speaker 2We have not done that yet. I am not great on theoretical neuroscience. One of my skills I immodestly think is there's a lot of, I know a lot about a little, a little bit about a lot of things, but there's something I know nothing about, like consciousness, international taxation and spike trains, but I know who knows. Yeah. Right. So, and, and also the, the Neural Autopilot model, as we kind of wrote it down, we've used it to explain some things like consumer choices, the gym, posting on social media, hasn't been like deeply tested in the way I would like from, in neuroscience. There are a lot of people who have studied habit going back quite a few years. And we, I think the idea that, that habit, and more deliberative choice are fundamentally dissociated as well-established. For example, in rodents, goal-directed, what we call goal-directed or deliberative or optimizing activity is in dorsal medial striatum, habit in dorsolateral striatum. And as well as this, again, that's part of a circuit. It's not that habits are only there, right? That that circuits connects to motor circuits and things. And there's some beautiful work by John O'Doherty, my colleague, and Bernard Berlin in Australia, where they actually, they lesion different parts of the brain. And you can either get rid of habits that rats have by lesioning the habit area, or you can get rid of goal-directed actions and make them more habitual by in the other area. So the, you know, you can kind of turn on and off, at least in these, in mice, you know, at least in rodents doing very simple, you know, pressing buttons or swimming. They have these things where they swim through water and they climb onto a, a ramp, a platform. And so they're, they're extremely simple tasks. They're a little bit like things humans do, like your ritual for making coffee in the morning, or you get in the car and you either reach for the seatbelt first, or do you turn the car on first? You know, there are sort of chained motor sequences that are, that are similar, but a lot of how we're thinking about habit is much more, much higher scale.
Speaker 1I remember, you know, I walk to the office every day. I have, my house is within walking, distance. And one day I decided to walk a different route and halfway through, I realized I had no idea where I was. Like, it's not that complicated, but.
Speaker 2Exactly. It's like muscle memory, right? Procedural memory, muscle memory. Didn't know what I was doing. Yeah. Yeah. Yeah. In fact, this is a, I'm so glad you brought that up. So one, one bit of evidence for habits is what are called action slips, right? So for example, if you drive to work a certain day and then it turns out the Olympics are in town, you know, most people forget they'd start driving to work and all of a sudden they're stuck. So that would, that's called an action slip. And my favorite example of this, I have on my slide deck, which is I was flying to Philadelphia to give a talk about habits and the flight attendant said, what would you like and give you a drink? And she gave me a snack, like a little cookie. What would you like to drink? I said, I'd like a, a Sprite Zero with no ice, Sprite Zero, please no ice. So she pops down ice and then I put my napkin on top of the cookie and she said, um, by the way, did you get your cookie? So it was a double action slip, right? All within four seconds, you know, I'm getting my ice and she has to, she forgot about the cookie. Autopilot. Yeah. Right. Cause she's on autopilot.
Speaker 1I mean, I guess this is, maybe this is a good roundup final question here. Um, even though it's completely unfair for you, are there at least hypothetically ways we could go into the brain and change some of these things? Are there therapies, drugs, whatever? Can we become better at K level thinking or less temporal discounting or whatever it would be? Is that a frontier?
Speaker 2Yes. Oh yeah. Yeah. No, absolutely. I mean, it's a, um, I mean, if you think that's an unfair question, you should see my colleagues from the past 40 years.
Speaker 1I can think of more unfair ones, but I'm trying to be nice.
Speaker 2Yeah, that's not an unfair question. That's your idea of unfair question? Come on. Um, yeah. So I'll give you a couple of examples. We, we worked on one study, but it wasn't really a knockout study, so I didn't want to mention it. Um, and it wasn't like a hard conclusive result I would bet, bet the farm on. Um, so there's something called optogenetics. And one of my brilliant colleagues, I don't know if you've heard of optogenetics, but one of my brilliant colleagues here, Viviana Gradanar, who works on that, and she's a student of Carl Dyseroth, who kind of invented it and will, is going to win a Nobel prize in medicine any day now. Um, and the idea is you can inject a virus into the brain, um, certain kinds of viruses that are light sensitive. And so in one experiment, there are cells in the brain that are encoding reward prediction errors, right? This is essentially the cells that, uh, they're in rats that, um, they're in rats that are in rats that are in rats that are in rats that are in rats. And so what happens is you train the rat to press a lever and it either gets a drink reward or not. And so you train it to press and every time it gets a drink reward, you can see this sort of reward prediction error. Like, oh, that was good. Oh, that was good. That was good. And so it learns that if you press the lever, you get something good. There's another lever that you press, and sometimes it gets salty water, which it hates. So it learns, I'm not going to press that. And what you do is you optogenetically upregulate the reward prediction in your neurons. I call this neural gaslighting. Let's see what you think. I like that. You upregulate. So when it presses the lever and gets salty water, the brain neurons say, that was great. because these these you've hijacked you've convinced the brain that salty water is really good because you know where these war perch and aeronauts are yeah it doesn't get overridden by the sensory system so you can trick the rat's brain into um thinking it likes something that it really doesn't and you can tell it doesn't because when you look at its face and so on it drinks the salty water it's like brain what were like what were you doing telling me to drink that salty water you're gaslighting me right i think i invented the term rural gaslighting which is a good one um and here's another one that's um that's a little less practical and interesting but it just tells you something cool so my academic granddaughter lucia ju is a professor in beijing and she has a paper that exploits um the fact that there are certain times when you're sleeping in which you can whisper you can say things to people or play a sound and people remember them and will affect behavior okay i can't remember the the acronym for this sorry listeners anyway and so what she does is but it has to be and you have to really know a lot about sleep and you have to do it in a sleep lab or with some equipment that knows when there's this sweet spot because otherwise they'll wake up or if they're in deep sleep they won't encode it and remember again then many of your readers will know but when you're sleeping the brain's not completely comatose right right there's neural replay dreaming obviously um uh if you know dreaming is like medicine for your brain there's kind of defragging and interesting things happen right so the you know the brain is doing something it's just not you're not conscious anyway so what she does is she has people value different snack foods similar to the study we talked about snickers um potato chips yogurt and then um let's say they value food they value food for three dollars and she whispers the name of the food when they're sleeping snickers snickers snickers snickers and then they wake up and she says you know did you hear anything when you're sleeping but no i was sleeping but they value the foods that you that were that were treated that were whispered about and of course you control so you whisper about all the different foods and compare when you weren't whispering blah blah blah they value them about 10 more so i don't think this is a marketing technique well who knows but it is an anecdotal answer to your question right once we know something about circuitry are there techniques to upregulate or train um and uh just to wind up and also as you probably know there's a lot of um work on uh deep brain stimulation and parkinson's so um and there's a lot of interest in what's called transcranial magnetic stimulation or tms which is you know low voltage current outside of the brain like but close to treat depression and things like that Um, so lots of therapeutics are working along these lines. They just haven't been brought to economic decisions. I think there's a good chance if we worked on it, we could get, we could get people to be more patient, but, but it's one thing to do it in a short run of one experiment for one hour. And it's another thing to do it maybe at a sweet spot, like when children are three or four, um, and preferably in a, you don't want to have to come in a lab and be in a scanner or be whispering at night or any of those things. They're just proofs of concept, right? Make people more virtuous? Yes, my colleagues did that. Okay. So Alex Smith at Virginia Tech and Cheryl Ball used transcranial direct current stimulation, TDCS, to TPJ, the area we talked about for mentalizing. And they basically had people do these tasks where you have some money, you can keep some or give some to a charity. And then they vary kind of the trade-offs. Like the charity gets very little, you get a lot, you get a very little, charity gets a lot. And they test kind of the rationality of how consistent they are and also whether they give more to the charity. And when you upregulate right to temporal partner junction, they're a little bit more generous and a little bit more rational. They're more consistent.
Speaker 1You know what? I'm going to call it there because that was an optimistic place. Yeah. You scared me a little bit 10 minutes ago, but now you're making me think that there's an optimistic path forward in addition to all the wonderful science and economics and everything. So, Colin Kammerer, thanks very much for being on the Mindscape podcast. That was fantastic.
Speaker 2My pleasure, Sean. Thanks. you
PodcastTranscript AI