38. Sendhil Mullainathan Explains How to Generate an Idea a Minute (Part 2)

0:00At Amica Insurance, we know trust isn't built in a day. It's earned over time by showing up in the moments you plan for and the moments you can't. The setbacks, the wins, and the big life changes. For us, trust is what insurance is all about. And it's why customers stay with Amica three times longer. Visit amica.com and get a quote today for auto, home, and life insurance. Amica. Empathy is our best policy.

0:30At Edward Jones, we believe rich is more than caring about the latest and greatest. It's also taking care of what gives your life meaning. That's why your dedicated financial advisor meets you where you are with personalized financial strategies that help protect what matters so you can preserve your progress while creating a path forward. The key to being rich is knowing what counts. Let's find your rich together. Edward Jones, member SIPC.

1:04In this week's episode, I continue my conversation with my University of Chicago colleague, Sendhil Malinatzen, an economist, data scientist, and MacArthur Genius Award winner, who has remarkable talent both for generating ideas and explaining those ideas in a way that anyone can understand. Welcome to People I Mostly Admire with Steve Levitt. Sendel has so many interests that in the first half of our conversation, that's last week's podcast episode, I didn't even manage to get to the topics he's thought about most. So if you missed last week's episode, no worries, the order of the two episodes doesn't matter.

1:42But in this week's episode, I pick up in the middle of our conversation with me telling Sendel about my first reaction to hearing that he had written a book entitled Scarcity.

1:56One of the things I found so unbelievable about that book is that I saw the title and it's called Scarcity. And I thought to myself, this is so arrogant because economics is the study of scarcity. And for the last 200 years, we've all been studying scarcity. So who does Sandal think he is that he's going to have something new to say? And I read the book and I'm like, wow, Sandal had something new to say. So first tell us what you had new to say. And then I think bring it home with the experiments. So economists study the physical fact of scarcity. You know, everything is scarce. You buy something, you're not buying something else.

2:33There's a constraint. What we are studying here is the psychology of scarcity, the feeling of having too little. And the hypothesis is that when you have too little of something, that tends to capture your attention. Your mind automatically goes towards it. When you're very busy, your mind goes towards the things taking up your time and the deadlines, the things that are due for the poor who are scarce in money, their mind automatically goes towards, oh my God, will I be able to make rent? And for me, the most satisfying thing about the book, it's the sheer amount of feedback I've gotten from people who have experienced poverty or who are experiencing poverty saying, this captures my experience of being poor.

3:24And I've had periods of poverty, and that's what you feel. And your mind just keeps going to this thing. So I'll give you the example. This is a study not from the book. This is with Anud Shah. And what we did was we reran the same word study that we just did around the word sleep. But we ran it with a new set of words. Sindel's referencing a study we talked about in part one of this interview. If you tell people a series of words like bed, rest, awake, nap, dream, wake, doze, snore, slumber, blanket, snooze, tired. Things that revolve around sleep, but you don't actually say the word sleep. And you ask people to recall the words.

4:1450 % of people will think you said the word sleep. You can induce a false memory. Sandal is now talking about rerunning that same study, but instead of using words that relate to sleep. They're words like cash, pay, loan, dollar, gas, grocery. For rich people, if you give them that list, there's no word that they remember having heard. That wasn't there. They remember some of these words, but they don't make up a word. But for poor people, when you give them the same list, a huge fraction of them remember having heard the word money. because for the poor, you hear grocery, they think money. You hear gas, they think money.

4:52And it's like many roads lead back to money. The consequence of this is that if your mind was a processor, like a computing processor, a fraction of it beyond your control is constantly churning on these concerns about money, which obviously makes it much harder to think about anything else. I love these studies. Can you tell the one about when you bring people into the lab and you tell them not to eat beforehand? Oh, yeah, yeah. This is great. People are brought into the lab. Everyone's hungry when they show up, but some of them are given some food and now they're less hungry. And they're all given the same word search test.

5:30Imagine a grid of letters and you have to look for a word. And we say, okay, first word, cookie. And so you search for cookie. Second word, tent. And what you're really testing is after you've seen the word cookie, how long does it take for you to find the word tent? If you're not hungry, whatever, it's just another word. Just go right along. If you're hungry, it takes much longer to find the word tent because your mind is still on that cookie. Not even that real cookie, but those letters, C-O-O-K-I-E. All it takes is a little bit of a prime for the hungry people, and suddenly their mind is off thinking about food.

6:06And I think that's an experience I'm sure everybody's had, that it's very hard to think about something else when you've just been reminded of a cookie. And in these studies, the effects are huge. I mean, you talk about effects that when you prime people to think about what their problem is, we're talking about like a difference of 10 IQ points between otherwise identical people, some of whom are primed to be thinking about the problem that fills their mind. Yeah, exactly. The hypothesis is because it's taking up our bandwidth, as we call it, the effect on your effective IQ can be quite large.

6:42There's a study by Claire Ducanois where there are these sort of math exams that have word questions. And some of the word questions involve monetary things. Bob has$10. Fred has$5. And she basically shows that only for poor students, after you get a monetary-themed question, you do much worse on the question that comes after. So much so that just having one in 10 more monetary questions in an exam reduces the performance gap by 6%. Wow, that's crazy. I should tell you one of their studies that we're just finishing. Because what's been fun for me in the last 10 years is seeing people take these ideas and do new different studies.

7:26So this is a study by Suanna Oh and Supricor and Frank Schobach. So what they do is these are workers in India. They get paid every two weeks. Now, this is lean times. So you know what it's like. You haven't gotten paid yet, so you're pretty far away from your payday. So you're like, you're pretty tight on money. And what you find is if you just give them some of their paycheck, like four days early, their actual productivity goes up. They're actually just more productive because they can focus more at work. And the way that they're focusing more at work is pretty cool because what these people are doing is they're making leaf plates.

7:59They're taking lots of little leaves and stitching them together to make a plate. And when you look at the workers who have just gotten a little cash infusion and don't feel as strained, you can look at the plates they make and they just have far fewer errors. Whereas the ones who did not get the financial infusion, you can see places where they made a mistake, had to take the stitches out and do it again and again. It's striking. You can see it in the way people work. And I started to notice how much of my own work comes from is my mind fully there. That's just a big part. If I have anxiety or something that my mind keeps going to, that's the biggest detriment to all my work.

8:36This makes total sense. But it doesn't really give a lot of clues of getting out of poverty other than giving people a bunch of money. Are there other implications? I think there's definitely other implications. Economists are very sensitive to taxes. Like we try not to tax when we avoid it. But we're very insensitive to cognitive taxes. If I said to you, okay, you want to get financial aid? Go ahead, fill out this 50-page form. Yeah, that's a time tax. It's whatever. It's two hours. In fact, our theories say that's a good thing because, hey, if you really need it, you can fill it out. But you forget, just like those kids in the math exam, asking a poor person to fill out a 50-page form all about their finances is incredibly cognitively taxing.

9:23It's like asking you to think about the thing that's stressing you out the most. And so we impose these cognitive burdens on the poor without really realizing that we're imposing these cognitive burdens. And when you start looking at that, you realize how many of our programs are cognitively silly. So like TANF, this is welfare. It's a welfare program. The way you find out that you're about to hit your five-year max is you get a letter in the mail with two, three months left. How does this make any sense that we're asking you to keep track? It's absurd, especially once you realize these type of results, that it's cognitively very challenging to think about these things.

10:02So I think that angle of rethinking how we design all the programs that poor people encounter through this lens, I think of as a promising angle.

10:19So you embraced modern data science tools earlier and more fully than almost any other economist I know. And from other prior conversations we had, I know you see huge breakthroughs on the horizon at the intersection of economics and big data and artificial intelligence. So what's your vision in this era? I think what's most exciting here is we just have a fundamentally new way of doing science. Let me use the example of like an EKG. I don't know if you ever had an EKG. It's basically they put 12 leads. These are like electrical leads on your chest. It reads the electrical activity of your heart and you get these little waveforms.

11:00Now we've been doing this for a hundred and something years and we've got an expert at looking at those EKGs and oh, oh, this little glimmer is a problem. But if you ask a simple question, where do you put the leads? It's, oh, these are the places. That's the way we've always been doing it. In fact, if you trace it back, it goes back to the first person who put the leads on the EKG. There isn't even many leads on the back. And the heart exists on the back as much as on the front. I mean, there's no constraint on how many leads you can put on. So the whole thing is absurd. The second thing that's weird is you look at these waveforms and everyone is taught to read these waveforms.

11:37It's like the Rosetta Stone. It's like its own little language. But why are we looking at these waveforms? Now you can take the waveforms you're getting and actually just simulate out the electrical activity of the heart itself. In fact, it's totally plausible that within 20 years, we will say, hey, between the EKG and maybe some CT scans of your heart, we've just built a digital version of your heart. And now we can do on a digital version of the heart. And I think that's the combination of algorithms plus data science. Rather than thinking within the data frame that you have, think past the data frame.

12:15Wow, you could totally reconstruct reality in a very different way. And what I find so interesting about that is before I talked to you, I was limited in my vision of what more data and better models could do to what everybody talks about, which is we have these wiggles from EKG. and a machine could be 17 % better than the human at seeing what the wiggles meant, which is fine and it's great and it's wonderful. Like I'm all in favor of that. But it was so interesting to hear you transform it into, hey, let's build a three-dimensional model of the human heart. And then once you've got that in your mind, you say, well, I definitely want to get a model of my heart when it's healthy.

12:54And I want to have that on file because then when I start to get sick, it's probably super informative to the doctors about, oh, wait, now I can see how the heart has changed. That tells me exactly where maybe I want to operate. And then again, it just feeds back. It ripples through the whole way you think about health and about medicine. I'll give you another example. You and I have run these big experiments. You run these experiments and you collect some variables. Oh, what happened to income? What happened to whatever? And it's somewhat a little unsatisfying, I always find. Because, yeah, you saw things on income, but you're like, I don't know, maybe these people's lives change in other ways.

13:30You spent a lot of energy running a big experiment, and now you only get to see a few things. So here's a different way you could run an experiment. You collect the things that you're meant to collect, etc. But you also just have people talking to a microphone about what's happening in their lives. Now, we don't think of people just talking as data. But now, these algorithms, given that they can process language, that is every bit as data as someone's income or their test score. And so I think there's almost another possibility there from language to radically change how we even think of data collection.

14:04And I think that, to me, if you look in all research, quantitative research is in one place and qualitative research is considered something altogether different. But I think algorithms are going to put these much closer in conjunction than we think. I think in 10, 15 years, simply getting people to talk about what's happening to them might end up being on par, if not higher than simply these quantitative metrics that we have. That's fascinating. Are you at all worried by what machine learning and artificial intelligence algorithm-driven future might hold? I haven't really paid much attention to this debate about the end of work and living machines that are going to make humans redundant.

14:46Do you have thoughts about any of that? Yeah, I think this stuff is so overstated. There's two groups of people who are complicit even though they hate each other. And one group wants to say these algorithms are going to ruin our lives and it's going to get rid of all jobs. Other people are going to say, oh, these algorithms are amazing. They're going to do great things. They both agree the algorithms were super powerful, just not whether they're good or bad. The reality is this is just a tool. It's not some magical, super powerful thing. And one reason it's not going to get rid of jobs, I think, is sure, it'll automate some jobs, but it creates so many more amazing opportunities.

15:25So I'll give you one concrete example. Right now, a big share of health care costs involve people who are very sick, but who need some kind of monitoring a little bit, just enough that we have to keep them in the hospital. And so that's very expensive. They're taking up hospital beds for that reason. Between the digital technologies we have and these algorithms, you can send these people home to be cared for by people they love or by nurses that come to their home who don't need to be nearly as skilled because when something goes really wrong, the hospital system can be alerted and so on. So what you've now done is you've decomposed one part of the job, the rare event, into as something the algorithm handles, and you've created a new job to handle the everyday event.

16:16Caretaking is something a lot of people can do. The sort of highly skilled doctor, nurse who needs to be at the hospital is something only a few people can do. So you've actually created a much better package of jobs now. And if anything, this is bad for the very highly paid, but they'll always find something else to do. So I think that it's a bit of sort of drama. Here's the one thing I find unnerving. I read a book and I can't remember anything about it 15 minutes later. So let's just say that machines get the ability to read books in a few seconds and to keep all of that material stored. That machine becomes potentially better at just about anything I could do than I am.

16:59And then what happens? I just wonder whether you get in the situation where when machines are just so good, then the ownership of the machines becomes everything and everybody else is marginalized. Yeah. So that's a great way to put it, Steve. I think all of those things rely on a misleading model of artificial intelligence. And I think it starts with the name artificial intelligence. There's a scale of intelligence, ants are somewhere here, baboons are somewhere here, we're somewhere above that. Right now, algorithms are between us and baboons, or maybe slightly above us, but as computing gets more powerful, they'll get to stratospheres above us that we can't imagine.

17:41So let's call this the unidimensional scale of intelligence. So this is what leads people to say, wow, if an algorithm can play chess so well, imagine all the other things it can do. So that's like the unidimensional scale. And the unidimensional scale is just such a bad way of thinking of machine intelligence because it is astonishingly good at certain activities and astonishingly bad at other activities. So one of the things that I do, if enough people do this, Apple will fix this by hand. But I do this illustration in this AI class I teach. I take out my phone and I say, Siri, don't tell me the score to the Sixers game.

18:21And Siri will dutifully tell me the score to the Sixers game. Like it's astonishing. It was like, oh, Sixers game, I know what you want. I just used the word not. And you just have to remember, this is like the world's best engineers are working on this problem. This is not some goofy product built in a summer intern's lab. These algorithms are very dumb in some respects and very smart in other respects. And so I think one of the things that we're all trying to do is to get a better understanding of what machine intelligence really looks like, What tasks will they excel at and what tasks will they do badly at?

18:56They are definitely not a substitute for human intelligence.

19:05You're listening to People I Mostly Admire with Steve Levitt and his conversation with economist Sendhil Malinatan. After this break, they'll return to talk about Sendhil's knack for great ideas.

19:24People I Mostly Admire is sponsored by Mint Mobile. For kids, summer means freedom. And as adults, summer still means freedom thanks to Mint Mobile's summer sale. Right now, all of Mint's plans are$15 per month, even unlimited. Free yourself from big wireless and switch to Mint for high-speed data and coverage on the T-Mobile 5G network. All plans come with high-speed data and unlimited talk and text delivered on the nation's largest 5G network. You can even bring your current phone and your number. To get your new wireless plan for just 15 bucks a month, go to mintmobile.com slash admire. That's mintmobile.com slash admire.

20:01Cut your wireless bill to 15 bucks a month at mintmobile.com slash admire. That's it. There's no catch. Upfront payment of$45 for three months,$90 for six months, or$180 for 12 months. Plan required. $15 per month equivalent. Taxes and fees extra. Initial plan term only. Greater than 50 gigabytes may slow when network is busy. Includes up to 20 gigabytes hotspot. Capable device required. Availability, speed, and coverage varies. See mintmobile.com.

20:34With the Discover Cashback card, it's payback time when you earn cash back on everyday purchases. Activate and earn 5 % cash back at different categories each quarter on up to$1 ,500 in purchases. That's 5 % cash back at different places each quarter, like grocery stores, on gas, and at restaurants. It pays to discover. Terms apply. See discover.com slash 5 for details. They say that every day your business is late to AI, you fall two days behind. But how do you keep up? Fortunately, there's NetSuite Next, and it starts with a name you know. You probably know NetSuite, the AI-powered business management suite that connects financials, inventory, commerce, HR, and CRM into a single source of truth.

21:21Now, it's time to try NetSuite next. Take a huge leap with custom insights, AI agents that problem-solve alongside you, handle routine work, and answer questions like talking with a colleague. NetSuite is built for different industries to support how your business works. Whether your company earns millions or hundreds of millions, it's time for NetSuite Next, where your business meets AI. For the first time ever, you can try NetSuite Next for free. If your revenues are at least in the seven figures, go to netsuite.ai. Built for every industry, ready for every boardroom. netsuite.ai. Hi, Morgan.

22:03Hi, Steve. All right. So last week I invited you on the show and I thought it went pretty well. You helped me with the Q &A session. So let's do it again. Yeah, I think it went well, too. So this week you're going to answer a question from a listener named Brett. He's from Philadelphia. He wrote in to ask about the number of listener emails we get, specifically if we noticed an uptick after you admitted that you read every email, which you started doing a few weeks ago with our episode with economist Demby Samoyo. So I'm glad you picked that question, Morgan, because you know I love to crunch data.

22:36So there are three different regimes we've had within people I mostly admire. First, we had an email address, but we only mentioned it at the end of the episode. Right. Then I started mentioning our email address along the way, asking people to write. And then we went to the third phase where I told people that I read every email. Well, in that first phase, not surprisingly, we weren't getting very many emails. We're getting about half an email a day. Okay. In the second phase, we were getting about two emails a day. Now, after I said we were reading every email, that number went up to a still highly manageable five emails per day.

23:10So 10 times as many as we had before, but still not exactly a huge sacrifice on my part to read all the emails. Not a huge sacrifice, but I've definitely felt the uptick. However, if working at Freakonomics has taught me anything, correlation does not mean causation. So how do we know that you saying you read every email caused this increase? Oh, A plus answer. Morgan, you're finally thinking like an economist. That's great. Because in order to assess causality, you have to think that the only thing that changed was me saying that I read all the emails. But of course, lots of things changed. I've been asking readers for much more interesting feedback along the way.

23:50So actually, I don't think me saying I answered the emails had a very big impact at all. Still, let's be realistic. the number of people who write is not very big. I did another calculation, and it turns out that if you were to be a regular listener, you would be expected to write me about once every 10 years, which seemed about right. That seems about how often I write to people who I listen to. I just don't have that much to say. Yeah, that sounds right. That's a little discouraging. But the one prompt that you gave to our audience that got people really excited was at the end of our episode with David Epstein.