So, let's say you're in high school. Your parents come home one day and they say, 'We've decided to get you a private tutor.' And boy, do we have the perfect guy for the job. He knows everything. He has literally memorized every fact and finding in every discipline. He can explain every concept to you with whatever metaphor or analogy you want. You're confused about bio. You need some Harry Potter metaphor to remember the difference between the sympathetic and parasympathetic nervous system. He is so good at that. And best of all, he's cheap. He's going to live in the basement. He never has to eat or sleep. He's available 24/7. How good does that sound? Sounds pretty good to you. And so one day this tutor comes by the house and you ask him, 'Is it true? Can you really memorize every fact? Teach me every single subject with Harry Potter analogies?' Of course, the tutor says, 'And there's something else I can do for you as well. In fact, I've done it for most of my past clients. If you want me to just take the biology test for you, I can do it. If you want to just skip the geography lesson and have me just write the geography essay for you, I can do that, too. Boring book. You don't want to read The Grapes of Wrath? I get it. It's long. I'll read the book for you and write the book report.'
He's the best teacher ever. And he's the worst tutor in the world. A tutor so terrible, in fact, that the more you rely on him, the dumber you get.
This isn't a magical tutor. It's not a make-believe thing. In fact, millions of students now have access to it, and we're learning the results. A team of researchers from California recently studied the effect of artificial intelligence, chatbots, on math scores on a popular learning website since 2015. These students could have used AI to unpack math concepts, come up with mnemonics for the quadratic equation. They could have used AI to become better at math. But what actually happened? On problems that students could copy into their chatbots, performance improved since 2022. But when students were given tests away from their computers, their scores plummeted. In fact, their scores after the invention of AI were worse than they had been before the invention of AI. The tutor who could teach them anything with any metaphor taught them to be dumber.
There's several lessons here. I think one lesson is the difference between the capability of a technology and its use. Just because we can use Wikipedia or Claude to become experts on quantum mechanics doesn't mean the primary use case of Wikipedia or Claude is to increase the number of particle physicists. The second lesson comes from the media theorist Marshall McLuhan who smartly said that every technology is both an extension and an amputation. A machine that extends human intelligence is also a machine that cuts off the need for humans to be intelligent. Like many technologies, AI forces us to ask a hard question. When you're in possession of a tool that can reduce struggle and friction, you have to ask yourself, what is the value of friction? When is the struggle the point? And when is the solution the point?
Ethan Mollick, I think, is probably the single best explainer of what AI can do for people and the difference between what it can do in theory and how it's actually used in practice. He's the author of a new book, Coexistence. And today we talk about the paradoxes of AI. AI makes experts more powerful while automating the junior work that creates experts. It makes our work smarter and faster by automating the process that makes us smart and fast. And according to Ethan, every challenge currently presented by AI will get harder and harder as its capabilities improve exponentially. I'm Derek Thompson. This is Plain English.
This episode of Plain English is presented by Comcast Business. Did you know Comcast Business detected 79.3 billion cyber threats last year? That's why they're introducing Security Edge Preferred. It's big business cyber security designed for small business. Get started with Comcast Business Internet for $50 a month for one year on a qualifying internet package. And find out how to get one month of Security Edge Preferred on them. Plus, ask about a $500 prepaid card. Limited time offer. Restrictions apply. Requires 300 megabits per second internet, Security Edge Preferred, and wireless connect. Call today. Ethan Mollick, welcome to the show.
So your book opens with forecasters trying to predict the future of artificial intelligence and math, and they gave AI a 2.3% chance of winning math Olympiad gold by 2025. Last year, superforecasters gave a Millennium Prize problem a 1.7% chance of being solved. Now OpenAI has said it has already solved one. I mean, we're in October. Maybe it'll solve another one in the next 24 hours. Just starting here at this forecasting failure to anticipate how powerful AI would be even in just a few months. What does that failure teach us about the capabilities of AI?
I think it teaches us a sort of lesson that I think feels like people have heard this since COVID, right? Which is humans don't think in exponentials very well, no matter how hard we try. And I think that it's tough, even the world's best forecasters, and they are really good, right? Like it's a straight line on a logarithmic graph. Like the ability level has not actually varied very much over the last four years, right? It's actually ticked up a little bit. And I think it's really hard for us to grasp what that means, right? It's better on a bunch of dimensions we've never had to measure before. It's not horsepower, it's not speed, it's intelligence, for whatever that is. And I think it's really hard to adjust what that means. I think that's a lot of failure I see constantly is an assumption that the future looks like the past, and that's not looking true anymore.
It seems to me that forecasters have actually been wrong in two directions. I think one group has underrated AI's growing capabilities in, for example, solving complicated math problems. But another group, I think, has overrated the real-world consequences of AI's capabilities. If you look at the unemployment rate, it looks basically normal. If you look at productivity growth, it's up a little bit, but there's a variety of reasons for that. And it's hard to see AI in the productivity numbers. If you look at corporate returns, they look pretty great. But if you kind of just like put your hand over the chip companies and the memory companies, a lot of the economy is saying we're not necessarily witnessing some enormous ROI from all this AI that's flowing through our bureaucratic structures. And so it raises this question that you and I have gone back and forth a bit online. And so I'd love to get your thesis on the record here. If AI's capabilities are growing so fast, growing exponentially, why are the real-world consequences of AI growth moving along so slowly?
So this is kind of where I'll argue, like, I'm surprised how fast it is, right? I am a professor at a business school, an economic sociologist, and the lesson you learn, and I say entrepreneurship, but the lesson you learn is like entrepreneurs should never win, right? Like, you know, just to say you have back, they should never win because all of the resources, all the money, all the customers are all with large companies. Why do they win? Because big companies can't change. So the fact that we're actually seeing in surveys high levels of ROI already, even if that hasn't been translated into, you know, 30 million is not enough to move the needle for a lot of these large companies, but they are seeing positive gains over a very short period of time that some of them are starting to reconfigure how they work. Like to me, that's fast, not slow. So I'm not surprised that we don't see this happening, especially because the current phase of AI we're in is single-player. It is about individuals getting performance gains. And you could see that, right? I can vibe code an entire game overnight. I can write academic papers faster than ever before. But when I start writing just systems, multiplayer, many people working together, there are still huge gaps. How does that even work? How do we work together in this? So you'll see, like in coding, a coder will be a hundred times more productive, but they're still in a two-week-long sprint that requires input from people every two weeks in different kinds of ways. You're not going to have a change. In fact, what you're going to have is huge amounts of token spend, and you're not going to see an impact. I don't think that that is a long-term persistent problem, but I think that it is a very obvious short-term problem and something I talk about a lot in the book as well.
There's two concepts that you've used that I think helped to make your argument especially sticky. You've argued or quoted other people arguing that jobs are like O-rings and that companies are like garbage cans. Why is the economy, in your view, made up of garbage cans and O-rings?
Right. So we have two tragedies here in different ways, right? So the O-ring is named after the Space Shuttle Challenger, where 16,000 parts or whatever worked great that day in 1986, but the O-ring failed and caused a catastrophe. And part of that is, you know, I talk about the jagged frontier a lot in this book and last book was a term we coined in our papers. And it's the idea that AI is good at some stuff you wouldn't expect, bad at some stuff you wouldn't expect. It only takes one or two places for AI to be bad at a job for you to still need a human to sub in for that because you can't have everything work great and then one thing fail at most jobs. It's not additive or little bits add and subtract from each other. So the O-ring means that jobs are not easily kind of replaced one for another with AI because if AI can't do human tasks exactly, you still have value accruing to human work, and you see that happening even in coding where like AI now writes code, but you still need human engineers to manage it. Now maybe that will change in the next couple of months, and then we'll see another part of the O-ring get stuck. So jobs are evolving very rapidly, but there's still these sticking points, and especially as you go to organizations. And then the garbage can model is one of the more depressing models of organizational theory, which is that basically companies are just bundles of problems and solutions all looking for each other, piled on top of each other. So there is no plan, or at least the plan is only a small part of how organizations work, and organizations are instead full of like contingencies and complicated aspects and evolutionary dead ends, and there's loops of work that no one ever knows even exists anymore. And you know, organizations just do things, and as a result, that makes it hard both for to improve them immediately. It also means that the job of redesigning organizations to work with AI is more complicated than just, well, we will supplement this task because that task is on paper somewhere but doesn't represent the actual garbage can of the company itself. So you can see why we call this a garbage can metaphor, right? Everything is tossed together like waste in a garbage can and mixed together randomly, and sometimes it works out and sometimes it doesn't.
Like you, I try to write about and cover artificial intelligence for a wide range of readers and listeners, some of whom are working at or even executives at the frontier labs, and some of whom absolutely despise the frontier labs, hate the acronym AI, hate everything about this subject, and wish we could just get into a time machine and go back to October 2022. And I think there are a lot of reasons for this. I mean, you see it even in the survey data where if you ask people who use AI, do you think it's helpful? By a four-to-one margin, they say yes. And you ask in the same survey, do you think AI is bad for society? And by a four-to-one margin, they say yes. So there's all sorts of interesting discrepancies here, but one that I would love you to dilate on a bit is that we're all using the word AI to describe what are often completely different products. When the frontier labs say our AI solved Navier-Stokes or some other math problem, they're talking about an internal model that won't be made public maybe for months. Then you have the most advanced public models that people spend a lot of money for, that corporations or individuals subscribe to. But only about 3% of individuals in the country actually personally subscribe to any AI. Many of them, if they're using it on a monthly basis, are using some model that's been public for years, which means that it was internally available even more years. So how do you think about this, I think, very difficult public communication challenge of talking about a product that is, at least in my last recounting, three different things: the most powerful internal AI models, which are much better than the most powerful public AI models, which are much better than the models that most people are using on a week-by-week basis. How do you address this problem?
And you're probably short a few, by the way. There's also the AI model that lives in people's heads from science fiction, right? Which determines a lot of what they think the future is, is Terminator, right? Or whatever they watch. And there's the AI, and then there's still inside companies a big issue is AI still means machine learning on your own data, where data was the new oil, and all the other techniques around artificial intelligence. So there's a lot of mess here. I would say that part of this is that it is confusing, right? Like we have a general-purpose technology, so there's many things happening at once, and the uses are different everywhere as well. I would say the biggest lack of clarity, though, is not feeling that exponential curve, right? And it seems very corny and very Silicon Valley to be like, feel the exponential pill, AGI pill, but like the truth is that models are getting better very, very quickly, and your old views are kind of out of date. I was actually, I think I'll remember the number correctly, but I was just trying to figure out how good the best model my last book was when it came out, which was two years ago, by the way. And both when I started writing it and when I finished writing it, and for several months afterwards, the best model in the world was GPT-4. It was the best model for about a year. And GPT-4 is now the 181st best model in the world right now. And there are 19 models that can run on your phone that are better than GPT-4. And that was like what everyone was freaking out about, you know, a few months ago. Now the average best model changes 70 days or less, right? So we're in this very fast acceleration period. And I do think everyone's lagging. I think that you can even use the smartest model in the world right now available to you, and you're not trying to solve Navier-Stokes. So you don't even know what it can do capably. So there's this huge mismatch, the overhang, right, between AI capability and where people's perception of AI is. And people are making all sorts of predictions based on all kinds of missing assumptions. Kind of nobody knows anything.
Reading your book, I felt like in many ways it's a book about tensions, about paradoxes that we live with when we work with AI, when we read about AI. And I want to go through at least four paradoxes that I drew from your new book. The first paradox I'm thinking of as the expertise paradox. And that is the idea that AI makes experts more powerful while often automating the kind of junior work that actually creates expertise. Maybe unpack a little bit about what you see here and how you think we're going to grow the next generation of experts if we're automating the process by which they are made.
Yeah. And the great thing is, since I wrote that, is like we actually have more evidence that both things are happening, right? I mean, maybe not great, but a true thing. So on one hand, there's a nice paper out of Anthropic suggesting that the more expertise you have, the better returns you get from using AI models. So you get better answers and you get more answers, right? And that makes sense because as AI has gotten better, it produces a huge amount of material, and if you're an expert, you can glance at that material right away and know whether it's good or bad or how to change it, just like you could if someone handed you a report. You're not going to read the whole thing, you flip through it. Or, you know, if you're an expert on, you know, podcasts, you know how this interview is going right now and are making judgment calls in your head that you couldn't describe to anyone else, right? So expertise has real value in working with complex systems. At the same time, you know, the way we teach expertise is apprenticeship. Like we don't have a way of teaching expertise fully in schools. So what we do is you send someone to be an intern at a podcasting studio, and they do the grunt work of doing the pre-interviews with me or whatever else, and you give them feedback, hopefully very kind feedback, maybe some people yell at them, and they learn the ropes that way. And you didn't want to do that interview. So they get to do it, and they get feedback as a result, and they have to prove to you that they're good. And it's been this really nice cycle that works well under capitalism and under state-planned societies, everything else, which is junior people want to strive to get ahead. They have, we need someone to do the work for senior people who don't want to do it, and you learn as you go. And all of that broke down with AI because junior people are much worse than AI, which means that junior people would be better off, you know, having AI do the work, and they don't have the expertise to assess whether it's good or bad. So they'll just turn in AI assignments. And the middle managers who used to be assigning work to junior people and giving them feedback, AI does a better job than that. It has no human emotions. So I'd rather just assign it to AI. So nobody learns anything anymore. The learning was incidental to the process, even if it was the point in the long term, right? So we were very lucky that we had a system where expertise creation was incidental to a process that worked really well in any kind of, you know, for 4,000 years, essentially. And so we have to rethink how we rebuild expertise. And that means we have to decouple efficiency from learning, which is not an easy thing to do in companies.
Not an easy thing to do at universities either. We had David Deming on this show a few months ago, and we talked about what I think is a really knotty problem of how do you teach artificial intelligence in an educational community that should be all about making thinking difficult so that people struggle at college, so they grow the muscles they can use when they graduate. But on the one hand, it seems if you teach with too much AI at the Harvards and Whartons of the world and every other, you know, the Virginia Techs and University of Michigan, then you are automating the process by which people become more intelligent, and that means they're going to end up dumber than they otherwise were. On the other hand, if you say, well, then we obviously can't teach with AI in college because it's allowing students to end-run the whole learning process, then you graduate thousands of people into a labor market where they might be expected to use AI, and you have forbidden their contact with this general-purpose technology. How do you think about the right way of balancing the virtue of learning at college and the virtue of colleges to serve as a pre-professional institution so that people graduate smart but also with some kind of facility with a technology they'll likely use in their first job?
Okay, so this will be my most controversial point and something academics don't say, but I don't think this is that hard a problem. I think transitioning to the state we need to be in is a harder problem. But we actually have a lot of control over people in universities and K-12 education. And we had to do things a very particular way. No one asked for AI to come and wreck the way we were doing things. But it has, and we actually have the evidence already, right? So if you use AI as a learning shortcut, which most people will naturally do, either because you want to cheat or just because you're asking for advice and the AI gives you the answer and you think you've learned something you haven't. We're very bad at knowing when we're learning. Then you destroy your learning opportunities, right? If, on the other hand, AI is used as a tutor where it's not giving you the answers but it's pushing you, we now have enough studies that show positive effects in meta-analysis RCTs at Harvard, at Stanford, and in Kenya, other places where we actually know AI can be a very powerful learning tool. So the answer isn't that hard, right? We need to use AI as a learning tool outside of class and inside of class do what we've known for the last 30 years of pedagogy works best, which is active learning experiences where in class it's discussions, cases. We have to test people in class without laptops nearby. We have to have in-class discussions. Like it's not the end of the, it's not that hard a process actually. Getting there will be hard, but like there is a better form of learning that awaits us in the end if we get there, all based on how we know how to do pedagogy today, right? And I think then the other piece of this puzzle is what does it mean to teach AI? I'm much more skeptical about there is an AI to teach or that people have to learn to use AI in college. Like I don't know what that skill is. I don't know what learn to use AI as a skill is. I know when CHROs of companies are saying, oh, the kids today, they're AI native. They're not AI native. They're just Claude. Like they're just pasting stuff into Claude. They're not learning to use AI. If returns to AI use are returns to expertise, then you're not an expert in AI. You're an expert in a topic area where you can use AI to do it.
Does the last thing you just said suggest that professors should be much more aggressive about limiting, even prohibiting any kind of AI use because there is no benefit to becoming fluent with the technology between 18 and 22? At college, your job is just to become good at learning, and when you graduate, you'll just learn and be filled with the AI that exists then when, you know, Bain or whoever else hires you.
So I think it's a little more complicated than that. I think that every class doesn't need to be an AI class, right? I absolutely think we should, like this is an emergency situation. We should be teaching AI in schools, but I feel like there's, you know, I'm a professor. The way professorship works is everyone kind of does their own thing in most cases. And so you have a lot of people, everybody's integrating AI into their class in all kinds of ways, some of which are hurting learning, some of which are wrong about AI, some of which are, you know, have trouble keeping up. So we should be thinking about how we teach AI skills. Like it helps to know a little bit how LLMs work. Not a lot, but a little bit. It helps to know about where current models stand. But I don't think every class needs to be an AI class. And I don't think we know what that even looks like. And by the way, there's lots of room for AI pedagogical tools. In my classes, I have AI. People have to teach an AI that I've trained to be a kind of a dumb student. So you have to teach it. That's a great way of learning is teaching. You build a case study with it. You run simulations where the AI is a game master. There's lots of interesting educational opportunities out here to use AI for helping learning that we will have to keep uncovering and working on. But AI as a subject, it might infuse everything you do, but you wouldn't want to take every class during the Industrial Revolution and make it a steam engine class. Like it's not necessary, right? And so I think in the same kind of way, like there is value in this, but there is a little bit of like everyone jumping on board of like we need AI literacy, AI certification. Well, I can tell you we've been studying at Wharton and all of the sort of prompting techniques I used to teach people, none of them work anymore. Like if you learned how to do, if you learned AI last year, right, in your recent college grad and you had a formal class on this, none of that works anymore. Think step by step is not a useful thing to tell an AI model. Personas no longer change AI performance levels. Like almost all the stuff you've learned now is now useless. So I think we need higher-order thinking. I mean, AI looks a lot like project management right now. So maybe everyone has to learn a project management class.
I think very related to what we've just been discussing, which is the expertise paradox, is what I'm calling the skill paradox. AI can make people much more productive. But as you've said in several studies, AI seems to make people better at certain tasks. But when you take away the AI, their scores plummet because it's revealed that they've been using the machine as a crutch rather than a tutor. When you're working with LLMs, how do you know? How do you feel when you're on one side of the line using AI as a tool versus another side of the line using AI to automate your thinking?
It is a really hard process, right? Like, so I do it by establishing bright lines for myself, right? Like I write every draft myself, and it's tempting sometimes because I'm like, I don't really want to do, like, you know, this is another post, another article. Like, but I have to. Like I write the drafts myself, and I absolutely turn to AI afterwards. And always the most depressing thing is like there'll be two lines of AI writing in like a piece of work, and everyone's like, that was a great line. I'm like, we wrote everything else. But I think that there is this kind of sense of like, you know, like this is the hardest challenge I have because we will absolutely deskill ourselves every chance we get. But part of the message of the book, I think, is be thoughtful about AI use. And it's tough to say that, right? Like, you know, humans don't like effort, right? We've had to teach ourselves that lifting weights is good when it obviously is something our body is designed to avoid. Like we don't want to do this, right? We don't want diet. And in the same way, learning requires desirable difficulties. It requires difficulty. We don't want to do that. And I think choosing what skills we are willing to give up, right? You know, and what skills we aren't is going to be our big challenge here. And I think building incentive systems to keep the right skills is going to be hard, especially companies, which aren't built for that kind of thing, right? Like it's one thing, again, university is we can control this. I can give you a test and make that test your grade, and you don't have AI on that test, and I can reveal whether or not you've learned things, and I can help set you up in a world where all the AI tutors are available to you, and the class experiences are amazing, and TAs are available, and like, and it ultimately has to come down to you. In a company, though, that's where things get really weird, right? Because not only do I not want to do that work that the AI is doing, you know, and deskill myself, but also the company doesn't want me to do that work, and they want, because it's more efficient for me not to do it. So strategic skill retention is going to be a very big deal in the future.
Going back to the metaphor of gym, it's just really interesting to me as I reflect on how frequently in the last two years the gym metaphor has felt more appropriate for the thing I'm obsessed with. So initially, you know, the gym was necessary because we had to be hunter-gatherers. We had to walk for 12 hours a day. And then we invented a specialized labor economy that meant that we didn't have to walk that much. And so if we wanted to remain fit, we had to go to this special thing and simulate some ancient processes in order to get back that fitness. That's gym 1.0. Then when I wrote The Antisocial Century about the decline of socialization, I was really interested in people who said, you know, we've thought of friendship and community and neighborliness as something that naturally comes out of us because of the need to, you know, reach out to people for support and meet friends and keep friendships. But now that the smartphone is here and it's easy to keep in touch with thousands or hundreds of people as sort of third-party allies on Twitter or Instagram, and you can sit on your couch all day and watch Netflix and go on TikTok, we need to think about hangouts as a means of social fitness, like get out of your house and practice being around people. And now what I'm hearing from you is there's a third gym that we need, that, you know, AI, because it's so good at doing our homework and our work and our jobs for us, we need to remember that sometimes the friction is the point. The struggle is the point. And we need to think of aspects of intellectual endeavor as a kind of gym, a special place or time to go to to work on skills that might feel hard in the moment. And we know we don't have to do it, but we think it might be good for us in the long run. It's just funny as I'm reflecting on the idea that this concept of the gym keeps coming up in all these segments as we find new ways of making life comfortable and technologically rich. I want to move on to the third paradox here that I wrote down, which is again related. It's what I'm thinking of as the knowledge paradox. In the last few months, OpenAI has released hundreds of math findings, proofs, and these are proofs that mathematicians themselves sometimes find quite alien when they work through the logic. I think one that you referenced on Twitter said a proof read like it was written on psychedelics. And I'm interested in how you think about this tension between AI creating knowledge without understanding. That is to say, it's creating the proof before humans understand the thing necessary to get to that proof. How do you think about this question of whether it matters that we're creating knowledge without human understanding?
This is part of our challenge, right? Like we didn't have to unpack different kinds of intelligences really because there was one kind of intelligence, which was human. So we talk about arrays and different types of things, but like we all kind of work the same way. And AIs are remarkably unalien, right? They're not the alien computer that we were expecting from like Star Trek. It's all logic. It would blow up if you said the word love to it, does not compute, explodes. Like they talk like emotional people for a variety of reasons, real and unreal. And you know, but the alienness starts to come up when we talk about intellectual work, right? And in this kind of, and also we haven't had to examine our own intellectual work. So one of the interesting things about mathematicians, right, is to the outside world, proofs were the big deal, right? That was the goal, is like these are the Millennium Problems. If we solve them, we're in great shape. And the mathematicians have been protesting and probably were saying before this, but are protesting publicly, not like this, right? The proofs weren't actually the point. Each proof would spin off new ideas for us that we never discovered before, and the process of solving it would bring the math community together around new problems and spaces that were interesting, and that's all being ruined. And in some ways, they're absolutely right, right? Like this, we're unraveling math as we know it, and we have to reconstruct this field. I think that's going to be happening in lots of spaces too, about reconstructing what we actually meant, unbundling this idea of, you know, intellectual engagement. I mean, you can now generate a pretty darn good podcast from my book or our tweets or anything you want, but it's not our podcast. How do we defend the idea that the human touch in writing or in proof, that the way we do knowledge and acquire it is valuable? I think it's going to be a big challenge coming up. So I'm absolutely of the belief that there is something real here, you know, and then, but then you have the second-order problem, which is, okay, well, do we need the humans to do math at all anymore? And you know, right now we do, but six months ago it was very obvious we needed humans, and six months from now it may not be, right? The proofs get generated, and then does AI then, does math become a game the way chess has, right? Which is like human achievement matters, but it's sort of in a bounded space. I think Bill Gates was talking about leaving reserves of problems for humans to solve, which feels a little bit unnerving to me as an outcome. But I think that we do have this kind of idea of trying to figure out what is the job that we are trying to do. And when we say not like that AI, what do we mean by that? And I think that's going to be important. I think there's a lot of not like that AI that's going to be important for us to think through.
As I was reading your book, I had a conversation with the mathematician Steven Strogatz about this phenomenon you're circling, which is mathematicians being indignant, offended, even quite moved to tears by how many math problems were solved by OpenAI, seemingly solved by OpenAI in a way that seemed to nuke their fields of research. And on the one hand, you know, you would expect a mathematician to say or to argue forcefully that not like this OpenAI, this is terrible. Having AI solve math problems without human understanding is a crime against the field. But he also shared this story of a character from his last book, MIT computer scientist Regina Barzilay, who survived breast cancer in her early 40s and was really stunned in the process to learn how much doctors didn't seem to know about breast cancer diagnosis. And so she led a team that trained AI on hundreds of thousands of mammograms, and they developed a technology that predicts breast cancer cases many years in advance better than any team of human doctors. And Strogatz asked Barzilay, 'Do you understand what your AI is doing?' And she said, 'No.' And he said, 'Doesn't it worry you that you don't understand your own invention?' And she said, 'Why would we limit a life-saving model to what humans can understand?' And on the one hand, I am very sympathetic to the tears of mathematicians. And on the other hand, I'm very sympathetic to wanting to not die of the cancers that killed my parents. I want to live. And if living requires a life-saving model that humans don't understand the inner mechanics of, I don't give a... The years of life are worth more than the qualia of those scientists understanding. And so I would love you to reflect just a little bit more on this tension between the scientific value of human understanding and the real-world value of just science and scientific progress for the sake of scientific progress, if it is indeed saving and extending lives.
