Lex Fridman PodcastErik Brynjolfsson: Economics of AI, Social Networks, and Technology | Lex Fridman Podcast #141
EVERY SPOKEN WORD
150 min read · 30,325 words- 0:00 – 2:56
Introduction
- LFLex Fridman
The following is a conversation with Erik Brynjolfsson. He's an economics professor at Stanford and the director of Stanford's Digital Economy Lab. Previously, he was a long, longtime professor at MIT, where he did groundbreaking work on the economics of information. He's the author of many books, including The Second Machine Age, and Machine Platform Crowd, co-authored with Andrew McAfee. Quick mention of each sponsor, followed by some thoughts related to the episode. Vincero Watches, the maker of classy, well-performing watches. Four Sigmatic, the maker of delicious mushroom coffee. ExpressVPN, the VPN I've used for many years to protect my privacy on the internet. And Cash App, the app I use to send money to friends. Please check out these sponsors in the description to get a discount and to support this podcast. As a side note, let me say that the impact of artificial intelligence and automation on our economy and our world is something worth thinking deeply about. Like with many topics that are linked to predicting the future evolution of technology, it is often too easy to fall into one of two camps, the fearmongering camp or the technological utopianism camp. As always, the future will land us somewhere in between. I prefer to wear two hats in these discussions, and alternate between them often. The hat of a pragmatic engineer, and the hat of a futurist. This is probably a good time to, uh, mention Andrew Yang, the presidential candidate who has been one of the high profile thinkers on this topic, and I'm sure I will speak with him on this podcast eventually. A conversation with Andrew has been on the table many times. Our schedules just haven't aligned, especially because I have a strongly held-to preference for long form, two, three, four hours or more, and in person. I work hard to not compromise on this. Trust me, it's not easy. Even more so in the times of COVID, which requires getting tested non-stop, staying isolated, and doing a lot of costly and uncomfortable things that minimize risk for the guest. The reason I do this is because, to me, something is lost in remote conversation. That something, that magic, I think is worth the effort, even if it ultimately leads to a failed conversation. This is how I approach life: treasuring the possibility of a rare moment of magic. I'm willing to go to the ends of the world for just such a moment. If you enjoy this thing, subscribe on YouTube, review it with five stars on Apple Podcasts. Follow on Spotify, support on Patreon, connect with me on Twitter @lexfridman. And now, here's my conversation with Erik Brynjolfsson.
- 2:56 – 7:24
Exponential growth
- LFLex Fridman
You posted a quote on Twitter by Albert Bartlett saying that, "The greatest shortcoming of the human race is our inability to understand the exponential function." Why would you say the exponential growth is important to understand?
- EBErik Brynjolfsson
Yeah, that quote, I remember posting that. Uh, it's actually a reprise of something Andy McAfee and I said in The Second Machine Age. But I posted it in early March, when COVID was really just beginning to take off, and I was really scared. There were actually only a couple dozen cases, maybe less at that time, but they were doubling every, like, two or three days. And, you know, (laughs) I could see, "Oh my God, this is gonna be a catastrophe, and it's gonna happen soon." But nobody was taking it very seriously, or not a lot of people were taking it very seriously. In fact, I remember, uh, did my last, um, c- in-person conference that week. I was flying, uh, back from Las Vegas, and, uh, I was the only person on the plane wearing a mask.
- LFLex Fridman
Yeah.
- EBErik Brynjolfsson
And the flight attendant came over to me. She looked very concerned. She kind of put her hands on my shoulders. She was touching me all over, which I wasn't thrilled about, and she goes, "Oh, you know, are, do you have some kind of anxiety disorder? Are you okay?"
- LFLex Fridman
Yeah.
- EBErik Brynjolfsson
And I was like, "No, you know, it's 'cause of COVID." And she's like-
- LFLex Fridman
So this is early March?
- EBErik Brynjolfsson
Early March. But, um, you know, I was worried because I knew, I could see, or I suspected I guess, that, that, that doubling would continue and it did, and, and pretty soon we had thousands of times more cases. Most of the time when I use that quote, I try to th- you know, it, it's motivated by more optimistic things like Moore's Law and the wonders of having more computer power. But in either case, it can be very counterintuitive. I mean, if you, if you walk for 10 minutes, you get about 10 times as far away as if you walk for one minute. (laughs) You know, that's the way our physical world works. That's the way our brains are wired, uh, um, but if something doubles for 10 times as long, uh, you don't get 10 times as much. You get 1,000 times as much, and after 20 it's a billion, after 30 it's a, a tr- no sorry, after 20 it's a million, after 30 it's a billion, and pretty soon after that, it just gets to these numbers that you can barely grasp. Our world is becoming more and more exponential, mainly because of digital technologies, so more and more often, our intuitions are out of whack and, uh, and that can be good, in the case of things creating wonders, but it can be dangerous in the case of viruses and other things.
- LFLex Fridman
Do you think it generally applies... Like is there spaces where it does apply and where it doesn't? How are we supposed to build an intuition about w- in which aspects of our society does exponential growth apply?
- EBErik Brynjolfsson
Well, you know, you can learn the math, but the truth is our brains, I think, tend to be, learn more from experiences. So we just start seeing it more and more often. So hanging around Silicon Valley, hanging around AI and computer researchers, I see this kind of exponential growth a lot more frequently, and I'm getting used to it, but I still make mistakes. I still underestimate some of the progress in, just talking to someone about GPT-3 and how rapidly natural language has improved, but I think that as the world becomes more exponential, w- we'll all start experiencing it more frequently.... the danger is that we may make some mistakes in the meantime using our old kind of caveman intuitions about how the world works.
- LFLex Fridman
Well, the weird thing is that it always kinda looks linear in the moment. Like, the every, you know, it's hard to feel, it's hard to, uh, retros- like, introspect and really acknowledge how much has changed in just a couple of years or five years or 10 years with the internet. If we just look at advancements of AI or even just social media, all the various technologies that go under the digital umbrella.
- EBErik Brynjolfsson
Yeah.
- LFLex Fridman
It feels pretty calm and normal and gradual and-
- EBErik Brynjolfsson
Well, a lot of stuff, you know, I think there are parts of the world, most of the world that is not exponential, you know. Um, the way humans learn or the way organizations change, the way our whole institutions adapt and evolve, those don't improve at exponential paces. And that leads to a mismatch oftentimes between these exponentially improving technologies, or let's say changing technologies, 'cause some of them are exponentially more dangerous, and our intuitions and our human skills and our institutions that, that just don't change very fast at all. Um, and that mismatch, I think, is at the root of a lot of the problems in our society, the, the growing, you know, inequality and, uh, other, other, uh, dysfunctions in our political and economic systems.
- 7:24 – 9:41
Elon Musk exponential thinking
- EBErik Brynjolfsson
- LFLex Fridman
So, one guy that talks about exponential functions a lot is Elon Musk.
- EBErik Brynjolfsson
Mm-hmm.
- LFLex Fridman
Uh, he seems to internalize this kind of way of exponential thinking.
- EBErik Brynjolfsson
Mm-hmm.
- LFLex Fridman
Uh, he calls it first principles thinking, sort of the kinda-
- EBErik Brynjolfsson
Yeah.
- LFLex Fridman
... going to the basics, asking the question, like, what were the assumptions of the past? How can, how can we throw them out the window? How can we do this 10X much more efficiently and constantly practicing that process? And also using that kind of thinking to estimate sort of, uh, when, you know, create deadlines and estimate when you'll be able to deliver on some of these technologies. Now, it, it often gets him in trouble because he overestimates, like he, uh, he doesn't meet the initial estimates of the deadlines.
- EBErik Brynjolfsson
Mm-hmm.
- LFLex Fridman
But he seems to deliver late but deliver.
- EBErik Brynjolfsson
Right. Well-
- LFLex Fridman
And which is kind of an interesting Like, what are your thoughts about this whole Elon thing?
- EBErik Brynjolfsson
Well, no, I, I, I think, uh, we can all learn from Elon. I think going to first principles, I, I talked about two ways of, of getting more of a grip on the exponential function. And one of them just comes from first principles, you know. If you understand the math of it, you can see what's gonna happen, and even if it seems counterintuitive that a couple of dozen of COVID cases could become, uh, thousands or tens or hundreds of thousands of them in a, in a month, um, it makes sense once you just do the math. Um, and I think Elon tries to do that a lot. I, you know, in fairness, I think he also benefits from hanging out in Silicon Valley.
- LFLex Fridman
Mm-hmm.
- EBErik Brynjolfsson
And he's experienced it in a lot of different applications. So, you know, it's not as much of a shock to him anymore. But that's, uh, that's something we can all learn from. Uh, uh, in my own life, I remember one of my first experiences, um, really seeing it was when I was a, a grad student and my, my advisor asked me to plot the growth of computer power in the US economy-
- LFLex Fridman
Mm-hmm.
- EBErik Brynjolfsson
... uh, in different industries, and there were all these, you know, exponentially growing curves (laughs) . And I was like, "Holy shit, look at this." Each, in each industry, it was just taking off. And, you know, you didn't have to be a rocket scientist to extend that and say, wow, this means that... This was in the late '80s and early '90s, that, you know, if it goes anything like that, we're gonna have orders of magnitude more computer power than we did at that time. And, of course, we do.
- 9:41 – 15:03
Moore's law is a series of revolutions
- LFLex Fridman
So, you know, when people look at Moore's law, they often talk about it as just... So the exponential function is actually, um, a stack of S-curves.
- EBErik Brynjolfsson
Mm-hmm.
- LFLex Fridman
So basically, it's you milk or whatever, take, uh, the most advantage of a particular little revolution and then you search for another revolution, and it's basically-
- EBErik Brynjolfsson
Yes.
- LFLex Fridman
... revolutions stacked on top of revolutions.
- EBErik Brynjolfsson
Mm-hmm.
- LFLex Fridman
Do you have any intuition about how the heck humans keep finding ways to revolutionize things?
- EBErik Brynjolfsson
Well, first let me just unpack that first point that I talked about exponential curves, but no exponential curve continues forever.
- LFLex Fridman
Mm-hmm.
- EBErik Brynjolfsson
Um, it's been said that if anything can't go on forever, eventually, it will stop. (laughs) And, and-
- LFLex Fridman
That's very profound. (laughs)
- EBErik Brynjolfsson
It's very profound. But it's, it seems that, that a lot of people don't appreciate that half of it as, as well either.
- LFLex Fridman
Yeah.
- EBErik Brynjolfsson
And that's why all exponential functions eventually turn into some kind of S-curve or, or, or stop in some other way, maybe catastrophically. And that had happened with COVID as well. I mean, it was, it went up and then it sort of, you know, at some point, it starts saturating the, the pool of people to be infected. Um, there's a standard epidemiological model that, that's based on that. Um, and it's beginning to happen with Moore's law or different generations of computer power. It happens with all exponential curves. The re- remarkable thing, as you alluded in the second part of your question, is that we've been able to come up with a new S-curve on top of the previous one and do that generation after generation with new materials, new processes, and just extend it further and further. Um, I don't think anyone has a really good theory about why we've been success- successful in doing that. Um, it's great that we have been, and, uh, I hope it, it continues for some time. But it's, uh, you know, one beginning of a theory is that there's huge incentives when other parts of the system are going on that clock speed of doubling every two to three years. If there's one component of it that's not keeping up, then the economic incentives become really large to improve that one part. It becomes a bottleneck. And anyone who can do improvements in that part can reap huge returns, so that the resources automatically get focused on whatever part of the system isn't keeping up.
- LFLex Fridman
Do you think some version of the Moore's law will continue?
- EBErik Brynjolfsson
Some version, yes. It a- it is. I mean, one version that has, um, become more important is something I call Koumi's law, which is, uh, named after John Koumi, who I should mention was also my college roommate. But he identified the fact that energy consumption has been declining by a factor of two, and for most of us, that's more important, you know? The new iPhones came out today as we're-
- LFLex Fridman
Mm-hmm.
- EBErik Brynjolfsson
... recording this. I'm not sure when you're gonna, uh, make it available.
- LFLex Fridman
Very soon after this, yeah.
- EBErik Brynjolfsson
Um, and for most of us, you know, having the iPhone be twice as fast, you know, it's nice, but having it, the battery life longer, that would be much more valuable. And the fact that a lot of the progress in chips now is reducing energy consumption, um, is probably more important for many applications than just the raw speed. Other dimensions of Moore's law are, um, in AI and machine learning. Um, those tend to be very parallelizable functions, um, especially, uh, deep neural nets. And, uh, so instead of having one chip, you can have multiple chips, or you can have a, a, a GPU, a graphic processing unit, that goes faster, and now special chips designed for machine learning like tensor processing units. Each time you switch, there's another 10X or 100X improvement above and beyond Moore's law. So, I think that the raw silicon isn't improving as much as it used to, but these other dimensions are becoming important, more important, and we're seeing progress in them.
- LFLex Fridman
I don't know if you've seen the work by OpenAI where they show the exponential improvement of the training of neural networks just literally in the techniques used. So that, that's-
- EBErik Brynjolfsson
Right.
- LFLex Fridman
... almost like the algorithm, the... It's, it's fascinating to think, like, can they actually continue as, as figuring out more and more tricks on how to train networks faster and faster?
- EBErik Brynjolfsson
Well, the progress has been staggering. You know, if you look at image recognition, as you mentioned, I, I think it's a function of at least three things that are coming together. One, we just talked about faster chips, not just Moore's law, but GPUs, TPUs, and other technologies. The second is just a lot more data. I mean, we are awash in digital data today in a way we weren't 20 years ago. Uh, photography, I'm old enough to remember, it used to be chemical, and now everything is digital. I took, uh, you know, probably 50 digital photos yesterday. Um, I wouldn't have done that if it was chemical. And, and we have, um, the internet of things and all sorts of other types of data. Our... When we walk around with our phone, it's just broadcasting a huge amount of digital data that can be used as training sets. And then last but not least, um, as they mentioned at OpenNI, op- as they mentioned at OpenAI, there have been significant improvements in the techniques. You know, the core idea of deep neural nets has been around for a few decades, but the advances in making it work more efficiently have also improved a couple of orders of magnitude or more. So, you multiply together, you know, a hundredfold improvement in computer power, a hundredfold or more improvement in data, hundredfold improvement in, uh, in techniques of software and algorithms, and soon you're getting into million-fold improvements.
- 15:03 – 16:42
GPT-3
- EBErik Brynjolfsson
- LFLex Fridman
Y- you know, somebody brought this up, this idea with GPT-3 that it's, uh... So it's trained in a self-supervised way on basically internet data.
- EBErik Brynjolfsson
Mm-hmm.
- LFLex Fridman
And that's one of the... I've seen arguments made, and they seem to be pretty convincing, that the bottleneck there is going to be how much data there is on the internet, which is a fascinating idea that it literally will just run out of human-generated data to train on. It's, uh...
- EBErik Brynjolfsson
Right. I know. We may get to the point where it's consumed basically all of human knowledge-
- LFLex Fridman
Yeah.
- EBErik Brynjolfsson
... or all digitized human knowledge. Yeah.
- LFLex Fridman
And that would be the bottleneck. I mean... But the, the, the interesting thing with bottlenecks is y- people often use bottlenecks as a way to argue against exponential growth. They say, "Well, there's no way you can overcome this bottleneck." But we seem to somehow keep coming up in new ways to, like, overcome whatever bottlenecks the, uh, the critics come up with, which is fascinating. I don't know how you overcome the data bottleneck-
- EBErik Brynjolfsson
Mm-hmm.
- LFLex Fridman
... but probably more efficient training algorithms.
- EBErik Brynjolfsson
Yeah. Well, you already mentioned that, that, that these training algorithms are getting much better at using smaller amounts of data. We also are just capturing a lot more data than we used to, especially-
- LFLex Fridman
True.
- EBErik Brynjolfsson
... in China. (laughs)
- LFLex Fridman
Yeah.
- EBErik Brynjolfsson
But, but all around us. So those are both important. You know, in, in some applications you can simulate the data. You know, video games, um, some of the, the self-driving car systems are, you know, simulating driving and, uh, of course, that has some risks and weaknesses, but you can also, in- in, if you want to, you know, exhaust all the different ways you could beat a video game, you could just simulate all the o- all the options.
- 16:42 – 23:43
Autonomous vehicles
- EBErik Brynjolfsson
- LFLex Fridman
Can we take a step in that direction of autonomous vehicles?
- EBErik Brynjolfsson
Mm-hmm.
- LFLex Fridman
I'm actually talking to the CTO of Waymo tomorrow.
- EBErik Brynjolfsson
Mm-hmm.
- LFLex Fridman
And obviously, I'm talking to Elon again in a couple weeks. What's your thoughts on autonomous vehicles? Like, where do we stand-
- EBErik Brynjolfsson
Mm-hmm.
- LFLex Fridman
... w- what, like-
- EBErik Brynjolfsson
Well-
- LFLex Fridman
... as, as a, as a problem that has the potential of revolutionizing the world?
- EBErik Brynjolfsson
Yeah. Well, you know, I'm really excited about that, but, uh, it's become much clearer that the original way that I thought about it and most people thought about it, like, you know, will we have a self-driving car or not, is way too simple. The, the better way to think about it is that there's a whole continuum of how much driving and assisting the car can do. I, I noticed that you're right next here, to your, next door to Toyota Research Institute.
- LFLex Fridman
That's a to- that's a total accident. I love the TRI folks, but yeah.
- EBErik Brynjolfsson
Have you talked to Gill Pratt?
- LFLex Fridman
Uh, yeah, we're going to... We were supposed to, uh, talk. It's ki- it's kinda hilarious.
- EBErik Brynjolfsson
So there's kind of the op- I think it's a good counterpart to, to what Elon is doing, um, and hopefully they can be frank in how, what they think about each other because I've heard both of them talk about it. Um, but they're much more, you know, this is an assistive, a guardian angel that watches over you as opposed to try to do everything. I think there are some things like driving on a highway, you know, from LA to Phoenix where it's mostly good weather, straight roads, that's...... close to a solved problem, let's face it. In other situations, you know, driving through the snow in, in Boston, where the roads are kind of crazy, and most importantly, you have to make a lot of judgments about what the other driver's gonna do at these intersections that aren't really right angles and aren't very well-described. It's more like game theory. (laughs)
- LFLex Fridman
Yeah.
- EBErik Brynjolfsson
Um, that's a much harder problem and requires understanding human motivations and, um ... uh, so there's a continuum there of some places where the cars will, uh, work very well, and others where it could probably take decades.
- LFLex Fridman
What do you think about the Waymo? So you mentioned two companies that are, actually have cars on the road.
- EBErik Brynjolfsson
Mm-hmm.
- LFLex Fridman
There's the Waymo approach that it's more like, "We're not going to release anything until it's perfect, and we're gonna v- be very strict-
- EBErik Brynjolfsson
Yeah.
- LFLex Fridman
... about the s- the streets that we travel on-
- EBErik Brynjolfsson
Mm-hmm.
- LFLex Fridman
... but it better be perfect."
- EBErik Brynjolfsson
Yeah. Well, I'm smart enough to be humble (laughs) and not try to get between ... I, I, I know there's very bright people on both sides of the argument.
- LFLex Fridman
Yeah.
- EBErik Brynjolfsson
I've talked to them, and they make convincing arguments to me about how careful they need to be and the social acceptance. Um, some people thought that when, uh, the first few people died from self-driving cars, that would shut down the industry, but i- it was more of a blip, actually. And, you know, so that was interesting. Um, of course, there's still a concern that if, if, uh, if there could be setbacks if we, we do this wrong. You know, your listeners may be familiar with the different levels of self-driving, you know, level one, two, three, four, five. I think, uh, Andrew Ng has convinced me that this idea of really focusing on level four, where you only go in areas that are well mapped rather than just going out in the wild, is the way things are gonna evolve. But you could just keep expanding those areas where you've mapped things really well, where you really understand them, and eventually they all become kind of interconnected. And that could be a, a, a, kind of another way of progressing, um, to make it more, uh, feasible over time.
- LFLex Fridman
I mean, that's kinda like the Waymo approach, which is they, uh, they just now released, I think just, like, a day or two ago, a public, like, anyone from the public in the, um, in the Phoenix, Arizona-
- EBErik Brynjolfsson
Mm-hmm.
- LFLex Fridman
... uh, to, uh, you know, you can get a ride in a Waymo car with no person, no driver, and then-
- 23:43 – 28:12
Electricity
- EBErik Brynjolfsson
In, in, in my books, I write about, like, electricity and how for 30 years, there was almost no productivity gain from the electrification of factories a century ago.
- LFLex Fridman
Mm-hmm.
- EBErik Brynjolfsson
And that's not because electricity is a wimpy, useless technology. We all know how awesome electricity is. It's 'cause at first, they really didn't rethink the factories. It was only after they reinvented them, and we describe how in the book. Um, then you suddenly got a doubling and tripling of productivity and growth. But it's the combination of the technology with the new business models, new business organization. That just takes a long time, and it takes, um, more creativity than most people have.
- LFLex Fridman
Can you maybe linger on electricity? 'Cause that's a fun one. Like-
- EBErik Brynjolfsson
Yeah. Well, well, sure. I'll tell you what, what happened. B- before electricity, there were ...... basically steam engines or sometimes water wheels. And to power the machinery, you had to have pulleys and crankshafts. And you really can't make them too long because they'll, they'll break with torsion. So all the equipment was kind of clustered around this one giant steam engine. You can't make small steam engines either because of thermodynamics. So if you have one giant steam engine, all the equipment clustered around it, multi-story. They'd have it vertical to minimize the distance as well as horizontal. And then when they did electricity, they took out the steam engine. They got the biggest electric motor they could buy from General Electric or someone like that, and nothing much else changed. (laughs)
- LFLex Fridman
Yeah.
- EBErik Brynjolfsson
It took until a generation of managers retired or died 30 years later that people started thinking, "Wait, we don't have to do it that way." You can make electric motors, you know, big, small, medium. You can put one with each piece of equipment. There's this big debate if you read the management literature between what they call group drive versus unit drive, where every machine would have its own motor. Well, once they did that, once they went to unit drive, those guys won the debate, um, then you started having a new kind of factory which is sometimes s- spread out over acres, single story, and each piece of equipment had its own motor. And most importantly, they weren't laid out based on who needed the most power, they were laid out based on what is the workflow of materials. (laughs) You know, assembly line, let's have it go from this machine to that machine to that machine. Once they rethought the factory that way, huge increases in productivity. It was just staggering. People like Paul David have documented this in their research papers, and, uh, you know, I, I think that there's a... that is a lesson you see over and over. It happened when the steam engine changed manual production. It's happened with the computerization, you know, people like Michael Hammer said, "Don't automate, obliterate." Um, in each case the big gains only came once smart entrepreneurs and managers basically reinvented their industries. I mean, one other interesting point about all that is that during that reinvention period, you often actually not only don't see productivity growth, you can actually see a slipping back, measured productivity actually falls. I just wrote a paper with Chad Syverson and Daniel Rock called The Productivity J Curve, which basically shows that in a lot of these cases you have a downward dip before it goes up, and that downward dip is when everyone's trying to like reinvent things. And you could say that they're creating knowledge and intangible assets, but that doesn't show up on anyone's balance sheet. It doesn't show up in GDP. So it's as if they're doing nothing. Like take self-driving cars, we were just talking about it. There have been hundreds of billions of dollars spent developing self-driving cars, and basically no chauffeur has lost his job. No taxi driver. I guess I gotta check out the ones that-
- LFLex Fridman
It's a big J curve. (laughs)
- EBErik Brynjolfsson
Yeah. So there's a bunch of spending and no real-
- LFLex Fridman
Yeah.
- EBErik Brynjolfsson
... consumer benefit. Now, they're doing that in the belief, I think the justified belief, that they will get the upward part of the J curve and there will be some big returns, but in the short run, you're not seeing it. That's happening with a lot of other AI technologies, just as it happened with earlier general purpose technologies, and it's one of the reasons we're having relatively low productivity growth lately. Uh, you know-
- LFLex Fridman
Mm-hmm.
- EBErik Brynjolfsson
... as an economist, one of the things that disappoints me is that as eye-popping as these technologies are, you and I are both excited about some of the things they can do, the economic productivity statistics are kind of dismal. We actually, believe it or not, have had lower productivity growth in the past about 15 years than we did in the previous 15 years, in the '90s and early 2000s. And so, that's not what you would have expected if, if these technologies were that much better, but I think we're in a, in kind of a, a long J curve there. Personally, I'm optimistic we'll start seeing the upward tick, um, maybe, maybe as soon as next year. But, um, uh, the past decade has been a bit disappointing if you thought there was a one-to-one relationship between cool technology and higher productivity.
- 28:12 – 33:19
Productivity
- EBErik Brynjolfsson
- LFLex Fridman
Well, what would you place your biggest hope for productivity increases on? Because you kind of said at a high level AI, but, eh, if I were to think about what has been so revolutionary in the last 10 years, I would... or 15 years, and thinking about the internet, I would say things like, um, hopefully I'm not saying anything ridiculous, but e- everything from Wikipedia to Twitter.
- EBErik Brynjolfsson
Mm-hmm.
- LFLex Fridman
So like these kind of websites, not so much AI, but like I would expect to see some kind of big productivity increases from just the connectivity between people and, uh, the access to inf- more information.
- EBErik Brynjolfsson
Yeah. Well, so that's another area I've done quite a bit of research (laughs) on actually, is these free goods like Wikipedia, Facebook, Twitter, Zoom. We're actually doing this in person, but-
- LFLex Fridman
Yeah.
- EBErik Brynjolfsson
... almost everything else I do these days is, uh-
- LFLex Fridman
With Zoom.
- EBErik Brynjolfsson
... is online. Um, the interesting thing about all those is most of them have a price of zero, you know? What do you pay for Wikipedia? Maybe like a little bit for the electrons to come to your house? (laughs)
- LFLex Fridman
(laughs) Yeah.
- EBErik Brynjolfsson
Basically zero, right?
- LFLex Fridman
Yeah.
- EBErik Brynjolfsson
Um...
- LFLex Fridman
I ta- take a small pause and say I donate to Wikipedia often. You should too, 'cause it makes a-
- EBErik Brynjolfsson
Good for you. Yeah. So, but what does that do, mean for GDP? GDP is based on the price and quantity of all the goods things bought and sold. If something has zero price, you know how much it contributes to GDP? To a first approximation, (laughs) zero.
- LFLex Fridman
Mm-hmm.
- EBErik Brynjolfsson
So these digital goods that we're getting more and more of, we're spending more and more hours a day consuming stuff off of screens, little screens, big screens. Um, that doesn't get priced into GDP. It's like they don't exist. Um, that doesn't mean they don't create value. I get a lot of value from watching cat videos and reading Wikipedia articles and listening to podcasts, even if I don't pay for them. Um, so we've got a mismatch there. Now in fairness, economists, since Simon Kuznets invented GDP and productivity, all those statistics back in the 1930s, he recognized, he in fact said, "This is not a measure of well-being. This is not a measure of welfare. It's a measure of production."... but almost everybody has kind of forgotten hi- that he said that, and they just use this. Like, how well off are we? Well, what was GDP last year? It was 2.3% growth or whatever. Um, that is how much physical production, but it's not the value we're getting. We need a new set of statistics, and I'm working with some colleagues, Avi Collis and others, um, to develop something we call GDP-B. GDP-B measures the benefits you get, not the cost. If you get benefit from Zoom or Wikipedia or Facebook, then that gets counted in GDP-B even if you pay zero for it. So, you know, back to your original point, um, I think there is a lot of gain over the past decade in these digital goods that doesn't sh- that doesn't show up in GDP, doesn't show up in productivity. By the way, productivity is just defined as GDP divided by hours worked, so if you mismeasure GDP, (laughs) you mismeasure productivity by the exact same amount. Um, that's something we need to fix. I'm working with the statistical agencies to come up with a new set of metrics, and, uh, you know, over the coming years I think we'll see... We're not gonna do away with GDP, it's very useful, but we'll see a parallel set of accounts that measure the benefits.
- LFLex Fridman
How, how difficult is it to get that B in the GDP-B?
- EBErik Brynjolfsson
It's, it's pretty hard. I mean, the thing, one of the reasons it hasn't been done before is that, you know, you can measure at the cash register what people pay for stuff, but how do you measure what they would've paid, like what the value is? That's a lot harder, you know. How much is Wikipedia worth to you? That's what we have to answer. And to do that, what we do is, um, we can use online experiments. We do massive online choice experiments. We ask hundreds of thousands, now millions of people to do lots of sort of A/B tests. How much would I have to pay you to give up Wikipedia for a month?
- LFLex Fridman
Mm-hmm.
- EBErik Brynjolfsson
How much would I have to pay you to-
- LFLex Fridman
Brilliant.
- EBErik Brynjolfsson
... stop using your phone? And in some cases, it's hypothetical. In other cases, we actually enforce it, which is kind of expensive. Like, we, we pay somebody $30 to stop using Facebook and we see if they'll do it, and-
- LFLex Fridman
Yeah.
- EBErik Brynjolfsson
... and some people will give it up for $10. Some people won't give it up even if you give them $100.
- LFLex Fridman
That's awesome.
- EBErik Brynjolfsson
And then you get a whole demand curve. You get to see what all the different prices are and what, how much value different people get. And not surprisingly, different people have different values. We find that women tend to value Facebook more than men.
- LFLex Fridman
Hm.
- EBErik Brynjolfsson
Old people tend to value it a little bit more than young people. That was interesting. I think young people maybe know about other networks that I don't know the name of that are better than Facebook. Um, and, uh, and so you get to see these, like, these patterns, but, you know, e- every person's individual. And then if you add up all those numbers, you start getting a, uh, an estimate of the value.
- LFLex Fridman
Okay. First of all, that's brilliant. Is this a work that, uh, will soon eventually be published?
- 33:19 – 43:36
Why is Twitter and Facebook free?
- EBErik Brynjolfsson
is already out.
- LFLex Fridman
You know, it's kind of a fascinating mystery that Twitter, Facebook, like, all these social networks are free, and it seems like almost none of them, except for YouTube, have experimented with removing ads for money.
- EBErik Brynjolfsson
Mm-hmm.
- LFLex Fridman
Can you, like... Do you understand that from a both economics and a, a product perspective?
- EBErik Brynjolfsson
Yeah, it's something that... You know, so I teach a course on digital business models, or u- used to at MIT. At, at Stanford. I'm not quite sure. I'm not teaching until next spring. I'm still thinking what my course is gonna be. Um, but there are a lot of different business models, and when you have something that has zero marginal cost, there's a lot of forces, especially if there's any kind of competition, that push prices down to zero. But you can have ad-supported systems. You can bundle things together. Um, you can have volunteer... You mentioned Wikipedia. There's donations. And, uh, I think economists underestimate the power of volunteerism and, and donations, um, you know, National Public Radio. Actually, how do you... D- this podcast, how is this, uh... What's the revenue model?
- LFLex Fridman
There's sponsors at the beginning and then-
- EBErik Brynjolfsson
Okay.
- LFLex Fridman
... and people... The funny thing is, I tell people they can... It's very e- I tell them to s- timestamp. So if you want to skip the sponsors, you, you're free. Uh, but the... It's funny that a bunch of people... So I read-
- EBErik Brynjolfsson
Yeah.
- LFLex Fridman
... the, the advertisement and then a bunch of people enjoy reading it and it's-
- EBErik Brynjolfsson
Well, they may learn something from it, and also, from the advertiser's perspective, those are people who are actually interested, you know?
- LFLex Fridman
Exactly.
- EBErik Brynjolfsson
Like, I mean, the example I sometimes give is, like, I, I bought a car recently, and all of a sudden all the car ads were, like, interesting to me. (laughs)
- LFLex Fridman
(laughs)
- EBErik Brynjolfsson
I was paying attention.
- LFLex Fridman
Exactly. Exactly.
- EBErik Brynjolfsson
And then, like, now that I have the car, like, I sort of zone out on them. Okay, but that's fine. The car companies, they don't really want to be advertising to me if I'm not gonna buy their product.
- LFLex Fridman
Yeah.
- EBErik Brynjolfsson
Um, so there are a lot of these different revenue models and, you know, i- i- it... It's a little complicated, but the economic theory has to do with what the shape of the demand curve is, when it's better to monetize it with charging people versus when you're better off doing advertising. Um, in short, on a... When the, when the demand curve is relatively flat and wide, um, like generic news and things like that, then you tend to do better with, um, advertising. If it's, uh, a good that's only useful to a small number of people, but they're willing to pay a lot, they have a very high, uh, value for it, then u- advertising isn't gonna work as well and you're better off charging for it. Both of them have some inefficiencies. And then when you get into targeting and you get into these other revenue models, it gets more complicated. But there's some economic theory on it. I also think, to be frank, there's just a lot of experimentation that's needed because, um, sometimes things are a little counterintuitive, especially when you get into what are called two-sided networks or platform effects, where, um, you may grow the market on one side and harvest the revenue on the other side. You know, Facebook tries to get more and more users, and then they harvest the revenue from advertising. Um, so that's another way of, of kind of thinking about it.
- LFLex Fridman
Is it strange to you that they haven't experimented?
- EBErik Brynjolfsson
Well, they are experimenting, so I- I- I... You know, they are doing some experiments about what the will- willingness is of-
- LFLex Fridman
I see.
- EBErik Brynjolfsson
... for people to pay.
- LFLex Fridman
Yeah.
- EBErik Brynjolfsson
Um, I, I, I think that when they do the math, it, it's gonna work out that, that they still are better off with an advertising-driven model.... but, um-
- LFLex Fridman
What about a mix? Like this is what YouTube is, right?
- EBErik Brynjolfsson
Yeah. Yeah.
- LFLex Fridman
It's, uh, you, uh, you allow the person to decide, the customer to decide-
- EBErik Brynjolfsson
Yeah.
- LFLex Fridman
... exactly which model they prefer.
- 43:36 – 46:56
Dismantling the nature of truth
- LFLex Fridman
I mean, implicit in what you're saying now is a hopeful message that with platforms, we can take a step towards, uh, greater and greater popularity of truth. But the more cynical view is that what the last few years have revealed is that there's a lot of money to be made in dismantling the, even the idea of truth.
- EBErik Brynjolfsson
Mm-hmm.
- LFLex Fridman
That nothing is true.
- EBErik Brynjolfsson
Mm-hmm.
- LFLex Fridman
And ev- as a thought experiment, I've been, you know, thinking about if it's possible that our future will have, like, the idea of truth as something we won't even have.
- EBErik Brynjolfsson
Mm-hmm.
- LFLex Fridman
Do you think it's possible, like, in the future that everything is on the table in terms of truth, and we're just swimming in this kind of digital economy-
- EBErik Brynjolfsson
Mm-hmm.
- LFLex Fridman
... where ideas are just little toys that are not at all connected to reality?
- EBErik Brynjolfsson
Yeah. I, I think that's definitely possible. I'm not a technological determinist, so I don't think that's inevitable. I don't think it's inevitable that it doesn't happen. I mean, the thing that I've come away with every time I do these studies and, and I emphasize it in my books and elsewhere is that technology doesn't shape our destiny. We shape our destiny. So, just by us having this conversation, I hope that your audience is gonna take it upon themselves as they design their products and they think about it and they use products, as they manage companies, how can they make conscious decisions to favor truth over falsehoods, favor the better f- kinds of societies and not abdicate and say, "Well, we just build the tools." I think there was a, a saying that, uh, that, was it the German scientists when they were working on the, uh, the missiles, um, in, in late World War II? You know, they said, "Well, our job is to make the missiles go up. Um, where they come down, that's someone else's department."
- LFLex Fridman
(laughs)
- EBErik Brynjolfsson
Um, and, you know, that's obviously a, a, not the r- I think it's obvious that's not the right attitude that technologists should have, that engineers should have. They should be very conscious about what the implications are. And if we think carefully about it, we can avoid the kind of world that you just described where, where truth is all relative. There are going to be people who benefit from a world of where people don't check facts and where truth is relative and popularity or, or fame or money, um, is orthogonal to truth. But one of the reasons I suspect that we've had so much progress over the past few hundred years is the invention of the scientific method, which is a really powerful tool or meta-tool for finding truth and favoring things that are true versus things that are false. If they don't pass the scientific method, they're less likely to be true. And, uh, that has this, the societies and the people and the organizations that embrace that have done a lot better, um, than the ones who haven't. And so, I'm hoping that people keep that in mind and continue to try to embrace not just the truth, but methods that lead to the truth.
- LFLex Fridman
So, maybe on a more personal question-
- EBErik Brynjolfsson
Mm-hmm.
- LFLex Fridman
... if one were to try to build a competitor to Twitter-
- EBErik Brynjolfsson
Mm-hmm.
- LFLex Fridman
... what would you advise? Is there, um, a m- I mean, m- I mean, the, the bigger, the, the meta question, is that the right way to improve systems?
- EBErik Brynjolfsson
Yeah.
- 46:56 – 53:11
Nutpicking and Cancel Culture
- EBErik Brynjolfsson
No. I, I, I think that the underlying premise behind Twitter and all these networks is amazing, that we can communicate with each other and, and I use it a lot. There's a subpart of Twitter called Econ Twitter where, you know, we economists, uh, tweet to each other and, and talk about new papers. Something came out in the NBER, the National Bureau of Economic Research, and we share about it. People critique it. I think it's been a godsend 'cause it's really sped up the scientific process, if you can call economic scientific. Um-
- LFLex Fridman
Does it get divisive in that little-
- EBErik Brynjolfsson
It, sometimes, yeah, sure. Sometimes it does. It can also be done in nasty ways and, you know, there's the bad parts. But the good parts are great because you just speed up that clock speed of learning about things, you know. Instead of, like in the old, old days and you're waiting to read it in a journal or the not so old days when you'd see it posted, um, on a, on a website and you'd read it. Now, on Twitter, like, people will distill it down and there's a real art to, to getting to the essence of things.
- LFLex Fridman
Mm-hmm.
- EBErik Brynjolfsson
So that's been great. Um, but, um, it certainly, we all know that (laughs) Twitter can be a cesspool of misinformation. And, uh, like I just said, unfortunately, misinformation tends to spread faster on Twitter than truth. And there are a lot of people who are very vulnerable to it. I'm sure I've been fooled at times. There are, uh, agents, whether from Russia or from political groups or, um, others that explicitly create efforts at misinformation and efforts at getting people to hate each other, or even more important lately I've discovered is, um, is nutpicking. You know the idea of nutpicking?
- LFLex Fridman
No. What's that?
- EBErik Brynjolfsson
It's a good term. Um...
- LFLex Fridman
(laughs)
- EBErik Brynjolfsson
Nutpicking is when you find, like, an extreme nutcase on the other side, and then you amplify them and make it seem like that's typical of the other side.
- LFLex Fridman
Yes.
- EBErik Brynjolfsson
So you're not literally lying. You're taking some...... idiot, you know, ranting on the subway.
- LFLex Fridman
Yeah.
- EBErik Brynjolfsson
Or just, you know, whether they're in the KKK or Antifa or whatever, they just, and you, normally, nobody would pay attention to this guy. Like 12 people would see him, it'd be the end. Instead, with video or whatever, you, you get mill- tens of millions of people say it and, and I've seen this. You know, I look at it and like I get angry. I'm like, "I can't believe that person did such things, it's so terrible."
- LFLex Fridman
(laughs) Yeah.
- EBErik Brynjolfsson
"Let me tell all my friends about this terrible person."
- LFLex Fridman
Yeah, exactly. (laughs)
- EBErik Brynjolfsson
(laughs)
- LFLex Fridman
Yeah.
- EBErik Brynjolfsson
And, uh, and it's, uh, it's a great way to generate division. I, I, I talk to a, a friend who studied Russian misinformation campaigns, and they're very clever about literally being on both sides of some of these debates. They would have some people pretend to be part of BLM, some people pretend to be white nationalists, and they would be throwing epithets at each other, saying crazy things at each other, and they're literally playing both sides of it. But their goal wasn't for one or the other to win, it was for everybody to get, be hating and distrusting everyone else. So these tools can definitely be used for that, and they are being used for that. It's been super destructive for our democracy and our society, and the people who run these platforms, I think, have a social responsibility, a moral and ethical personal responsibility, to do a better job, and to shut that stuff down, as well. I don't know if you can shut it down, but to, to design them in a way that, that, you know, as I said earlier, favors truth over, over falsehoods, and favors positive types of, um, uh, communication versus destructive ones.
- LFLex Fridman
And just like you said, it's also on us, which I, I try to be all about love and compassion, empathy on Twitter. I mean, one of the things-
- EBErik Brynjolfsson
Yeah.
- LFLex Fridman
... nitpicking is a fascinating term. One of the things that people do that's, I think, even more dangerous is nitpicking applied to individual statements of good people. So basically-
- EBErik Brynjolfsson
Yeah.
- LFLex Fridman
... worst-case analysis in, uh, computer science is, uh, taking, sometimes out of context, but sometimes in context, uh, a statement, one statement-
- EBErik Brynjolfsson
Right.
- LFLex Fridman
... by a person, uh, like I've been, because I've been reading The Rise and Fall of the Third Reich. Uh-huh.
- EBErik Brynjolfsson
I've often talk about, uh, Hitler on this podcast with folks.
- LFLex Fridman
Uh-huh.
- EBErik Brynjolfsson
It is so easy-
- LFLex Fridman
That's really dangerous. (laughs)
- 53:11 – 59:12
How will AI change our world
- EBErik Brynjolfsson
- LFLex Fridman
So you've written quite a bit about how artificial intelligence might change our world.
- EBErik Brynjolfsson
Mm-hmm.
- LFLex Fridman
How do you think... If, if we look forward, again, it's impossible to predict the future, but if we look at trends in, from the past, and we try to predict what's gonna happen in the rest of the 21st century, how do you think AI will change our world?
- EBErik Brynjolfsson
(laughs) That's a big question.
- LFLex Fridman
(laughs)
- EBErik Brynjolfsson
I, you know-
- LFLex Fridman
Yeah.
- EBErik Brynjolfsson
... I'm mostly a techno-optimist. I'm not at the extreme, you know, the singularity is near end of the spectrum. But I, I do think that we are likely in for some significantly improved living standards, some really important progress. Even just the technologies that are already kind of, like, in the can that haven't diffused. You know, when I talked earlier about the J-curve, it can take 10, 20, 30 years for an existing technology to have the kind of profound effects. And when I look at whether it's, you know, vision systems, voice recognition, problem-solving systems, even if nothing new got invented, we would have a few decades of progress. So I, I'm excited about that, and I think that's gonna lead to us being wealthier, healthier. I mean, the healthcare is probably one of the applications that I'm most excited about. Um, so that's good news. I don't think we're gonna have the end of work any time soon. Um, there's just too many things that machines still can't do. Um, when I look around the world and, and think of whether it's, it's childcare or healthcare, cleaning the environment, um, interacting with people, scientific work, artistic creativity,These are things that, for now, machines aren't able to do nearly as well as humans. Even just something as mundane as, you know, folding laundry or whatever. And s- many of these, I think, are gonna be years or decades before machines catch up. You know, I may be surprised on some of them, but, but overall, I think there's plenty of work for humans to do. There's plenty of problems in society that need the human touch. So we'll have to repurpose. We'll have to, as machines are able to do some tasks, people are gonna have to re-skill and move into other areas, um, and that's probably what's gonna be going on for the next, you know, 10, 20, 30 years or, or more. A kind of big restructuring of society. We'll get wealthier and people will have to do new skills. Now, if you turn the dial further, I don't know, 50 or 100 years into the future, then, you know, maybe all bets are off. Then it's possible that, that machines will be able to do most of what people do. You know, say, one or two hundred years, I think it's even likely. And at that point, then we're more in the sort of abundance economy. Then we're in a world where there's really little for h- that humans can do economically, uh, better than machines, other than be human. And, uh, you know, well, that will take a transition as well, kind of more of a transition of how we get meaning in life and, and what our values are. But, but shame on us if we screw that up. I mean, that should be like great, great news.
- LFLex Fridman
Yeah.
- EBErik Brynjolfsson
And it kinda saddens me that some people see that as like a big problem. You know, I think that would be, should be wonderful if, if people have all the health and, and material things that they need and, and can focus on loving each other and discussing philosophy and playing and, and doing all the other things that don't require work.
- LFLex Fridman
Do you think you'll be surprised to see what the 20... Like, if we were to travel in time 100 years into the future, do you think you'll be able to... Like, if I gave you a month to, like, talk to people. No, like let's say a week. Would you be a- would you be able to understand what the hell is going on?
- EBErik Brynjolfsson
You mean if I was there for a week?
- LFLex Fridman
Yeah, if you were there for a week.
- EBErik Brynjolfsson
Uh, 100 years in the future?
- LFLex Fridman
Yeah. So like, so I'll give you one thought experiment is like, isn't it possible that we're all living in virtual reality by then? Like-
- EBErik Brynjolfsson
Yeah. No, I think that's very possible. You know, I've played around with some of those VR headsets and they're not great, but I mean, the average person spends many waking hours staring at screens right now. You know, they're kinda low res compared to what they could be in 30 or 50 years. Um, but certainly games and why not, um, any other interactions could be done with VR, and that would be a pretty different world and we'd all, you know, in some ways be as rich as we wanted. You know, we could have castles-
- LFLex Fridman
Right.
- EBErik Brynjolfsson
... and we could be traveling anywhere we want. Um, and it could obviously be multisensory. So that would be, that would be possible and, you know-
- LFLex Fridman
With Neuralink as well.
- EBErik Brynjolfsson
... of course, there's people, you know. Uh, y- uh, you've had Elon Musk on and others, you know, there are people, Nick Bostrom, you know, makes the, the simulation argument that maybe we're already there. (laughs)
- LFLex Fridman
(laughs) We're already there. So but, but in general, or do you not even think about i- in this kind of way, you're self-critically thinking how good are you as an economist at predicting what the future looks like? Y- do you ever-
- EBErik Brynjolfsson
Well, it starts getting... I mean, I feel reasonably comfortable next, you know, 5, 10, 20 years in terms of, um, that path. When you start getting truly superhuman artificial intelligence, um, kind of by definition, (laughs) I'd be able to think of a lot of things that I couldn't have thought of and create a world that I couldn't even imagine. And, uh, so I, I'm not sure I can, I can predict what that world is going to be like. One thing that A- AI researchers, AI safety researchers worry about is what's called the alignment problem. When an AI is that powerful, then, um, they can do all sorts of things and you really hope that their values are aligned with our values. And it's even tricky defining what our values are. I mean, first off, we all have different values.
- LFLex Fridman
Mm-hmm.
- EBErik Brynjolfsson
And secondly, maybe if we were smarter, we would have better values, like, you know, I like to think that we have better values than they did in 1860, um, and, or in, you know, the year 200 BC on a lot of dimensions, things that we consider barbaric today, and it may be that if I thought about it more deeply, I would also be morally evolved. Maybe I'd be a vegetarian or, or do other things that, that, uh, right now, um, whether my future self would consider kind of immoral. So, um, that's a tricky problem, getting the AI to do what we want, assuming it's even a friendly
- 59:12 – 1:01:05
Existential threats
- EBErik Brynjolfsson
AI. I mean, I should probably mention, there's a non-trivial other branch where we destroy ourselves, right? I mean, there's a lot of exponentially improving technologies that could be ferociously destructive, um, whether it's in nanotechnology or biotech and weaponized viruses, AI, and other things that-
- LFLex Fridman
Nuclear weapons.
- EBErik Brynjolfsson
Nuclear weapons, of course.
- LFLex Fridman
The old school technology.
- EBErik Brynjolfsson
Yeah. Good old, good old nuclear weapons that could, uh, could be devastating or even existential, um, and new things yet to be invented. So that's a branch that, you know, I, I think is, is pretty significant and there are those who think that one of the reasons we haven't been contacted by other civilizations, right? Is that, is that once you get to a certain level of complexity and technology, there's just too many ways to go wrong. There's a lot of ways to blow yourself up and people, uh, or I should say species end up falling into one of those traps. The great filter.
- LFLex Fridman
The great filter. I mean, there's an optimistic view of that. If there is literally no intelligent life out there in the universe or at least in our galaxy, that means that we've passed at least one of the great filters or some of the great filters-
- EBErik Brynjolfsson
Right.
- LFLex Fridman
... that we survived, but-
- EBErik Brynjolfsson
Yeah. No, I th- I think, I think it's Robin Hanson who has a good way of, maybe others, they have a good way of thinking about this, that if there are no other intelligent c- creatures out there and that we've been able to detect, one possibility is that there's a filter ahead of us, and when you get a little more advanced, maybe in 100 or 1,000 or 10,000 years-... things just get destroyed-
- LFLex Fridman
Oh, boy.
- EBErik Brynjolfsson
... for some reason.
- LFLex Fridman
Yeah.
- EBErik Brynjolfsson
The other one is the great filter's behind us. That, that would be good, is that most, um, planets don't even evolve life, or if they don't evolve life, they don't evolve intelligent life. Maybe we've gotten past that, and so now maybe we're on the good side of the, of the great filter.
- 1:01:05 – 1:07:11
AI and the nature of work
- EBErik Brynjolfsson
- LFLex Fridman
So (laughs) uh, if we sort of rewind back and look at the, the thing where we could say something a little bit more comfortably at five years and 10 years out.
- EBErik Brynjolfsson
Mm-hmm.
- LFLex Fridman
You've, uh, written about jobs.
- EBErik Brynjolfsson
Mm-hmm.
- LFLex Fridman
And, uh, the impact on sort of our economy and the jobs, i- in terms of artificial intelligence that might, it might have. It's a fascinating question, what kind of jobs are safe? What kind of jobs are not?
- EBErik Brynjolfsson
(laughs)
- LFLex Fridman
Can you maybe speak to your intuition about how we should think about AI changing the landscape of work?
- EBErik Brynjolfsson
Sure. Absolutely. Well, this is a really important question, because I think we're very far from artificial general intelligence, which is AI that can just do the full breadth of what humans can do. But we do have human-level or super-human level narrow intelligence, narrow artificial intelligence. Um, and, you know, obviously my calculator can do math a lot better than I can (laughs) , and there's a lot of other things machines can do better than I can. So which is which? We actually set out to address that question, um, with Tom Mitchell. I wrote a, uh, paper called What Can Machine Learning Do? that was in Science. And it, we went and interviewed a whole bunch of AI experts and kind of synthesized what they thought machine learning was good at and wasn't good at, and, uh, we came up with what we called a rubric, uh, basically a set of questions you can ask about any task that will tell you whether it's likely to score high or low on, uh, suitability for machine learning. And then we've applied that to a bunch of tasks in the economy. Um, in fact, there's a data set of all the tasks in the US economy, believe it or not. It's called O*NET. Um, the US government put it together, part of the Bureau of Labor Statistics, and they divide the economy into about 970 occupations, like, you know-
- LFLex Fridman
Mm-hmm.
- EBErik Brynjolfsson
... bus driver, economist, primary school teacher, radiologist. And then for each one of them, it, they describe which tasks need to be done.
- LFLex Fridman
Mm.
- EBErik Brynjolfsson
Like for radiologists, there are 27 distinct tasks. So we went through all those tasks to see whether or not a machine could do them. And what we found, interestingly, was-
- LFLex Fridman
Brilliant study, by the way.
- EBErik Brynjolfsson
Thank you.
- LFLex Fridman
That's so awesome. (laughs)
- EBErik Brynjolfsson
Yeah. Thank you. Um, so what we found was that there was no occupation in our data set where machine learning just ran the table and did everything, and there was almost no occupation where machine learning didn't have like a significant ability to do things. Like take radiology. A lot of people, I hear it said, you know, "It's the end of radiology," and one of the 27 tasks is read medical images. Really important one, like it's kind of a core job (laughs) , and machines have basically gotten as good or better than radiologists. There was just an article in, uh, Nature last week, but, you know, they've been publishing them for the past few years, um, showing that, uh, um, machine learning can do as well as humans on many kinds of diagnostic imaging tasks. Um, but other things radiologists do, you know, they sometimes administer conscious sedation. Uh, they sometimes do physical exams. They have to synthesize the results and explain to, to the other, uh, uh, doctors or to the patients. In all those categories, machine learning isn't really up to snuff yet. So that job, we're gonna see some, a lot of restructuring. Um, parts of the job they'll hand over to machines, others humans will do more of. And that's been more or less the pattern in all of them. So, you know, to oversimplify a bit, we're gonna see a lot of restructuring, uh, reorganization of work, and it's real, gonna be a great time, it is a great time for smart entrepreneurs and managers to, to do that reinvention of work. Not gonna see mass unemployment. To get more specifically to your question, the kinds of tasks that machines tend to be good at are a lot of routine problem solving, mapping, uh, inputs X into outputs Y. If you have a lot of data on the Xs and the Ys, the inputs and the outputs, you can do that kind of mapping and find the relationships. They tend to not be very good at, for even now, fine motor control and dexterity, um, emotional intelligence and, and human interactions, um, and thinking outside the box, creative work. If you give it a well-structured task, machines can be very good at it, but even asking the right questions, that's hard. There's a quote that Andrew McAfee and I use in our book, Second Machine Age. Um, apparently, uh, Pablo Picasso was shown an early computer, and he came away kind of unimpressed. He goes, "Well, I don't see all the fuss. All that does is answer questions."
- LFLex Fridman
(laughs)
- EBErik Brynjolfsson
And, you know-
- LFLex Fridman
Yeah.
- EBErik Brynjolfsson
... to him, the interesting thing was asking the questions.
- LFLex Fridman
Yeah. Try to replace me, GPT-3.
- EBErik Brynjolfsson
(laughs)
- LFLex Fridman
I dare you, although some people think I'm a robot. You have this cool plot that shows, um... (laughs) I just remember where economists land. Uh, where-
- EBErik Brynjolfsson
Mm.
- LFLex Fridman
... I think the X axis is the income.
- EBErik Brynjolfsson
Yes.
- LFLex Fridman
And then the Y axis is, I guess, aggregating the information of how replaceable the job is, or I think there's an index.
- EBErik Brynjolfsson
I think it's the Suitability for Machine Learning index. Exactly.
- LFLex Fridman
Yeah. Yeah.
- 1:07:11 – 1:13:03
Thoughts on Andrew Yang and UBI
- EBErik Brynjolfsson
- LFLex Fridman
So on th- on this topic of, uh, the effect of AI o- on our, on our landscape of work, s- one of the people that have been speaking about it in the public domain, public discourse is the presidential candidate, Andrew Yang.
- EBErik Brynjolfsson
Yeah.
- LFLex Fridman
Uh, what are your thoughts about Andrew? What are your thoughts about UBI that, uh, universal basic income that he made one of the core ide- by the way, he has like hundreds of ideas about like everything. It's-
- EBErik Brynjolfsson
He does.
- LFLex Fridman
... kinda int- it's kinda interesting.
- EBErik Brynjolfsson
Yeah.
- LFLex Fridman
But what are your thoughts about him?
- EBErik Brynjolfsson
Well-
- LFLex Fridman
And what are your thoughts about UBI?
- EBErik Brynjolfsson
Let me answer, you know, the, the, the, this, um, question about his broader c- you know, approach first. I mean, I just love that.
- LFLex Fridman
Yeah.
- EBErik Brynjolfsson
He's really thoughtful, analytical. I agree with his values. So that's awesome. And, and, and he read my book, and, and-
- LFLex Fridman
(laughs)
- EBErik Brynjolfsson
... mentions it sometimes, so it makes me even more exciting. Um, and, uh, the, the thing that he really made the centerpiece of his campaign was UBI. And I was originally kind of a fan of it, and then as I studied it more, I became less of a fan, although beginning to come back a little bit. So let me tell you a little bit of my evolution. You know, as an economist, we have, uh, looking at the, the problem of people not having enough income and the simplest thing is, "Well, why don't we write them a check?" (laughs)
- LFLex Fridman
Right.
- EBErik Brynjolfsson
Problem solved. But then I talked to my sociologist friends and people be- and, and they really convinced me that just writing a check doesn't really get at the core values. You know, Voltaire once said that, uh, work solves three great ills, boredom, vice, and need. And, uh, you know, you can deal with the need thing by writing a check, but people need a sense of meaning, meaning, they need something to do, and, um, when, uh, you know, say, steel workers or coal miners, um, lost their jobs and were just given checks, alcoholism, depression, divorce, all those social indicators, drug use all went way up. People just weren't happy just sitting around collecting a check. Um, maybe it's part of the way they were raised, maybe it's something innate in people that they need to feel wanted and needed. So it's not as simple as just writing people a check. You need to also give them a way to have a sense of purpose. And that was important to me. And the second thing is that, as I s- mentioned earlier, you know, we are far from the end of work. Uh, you know, I don't buy the idea that there's just like not enough work to be done. I see s- like our cities need to be cleaned up.
- LFLex Fridman
Yeah.
- EBErik Brynjolfsson
And r- robots can't do most of that. You know, we need to have better childcare. We need better healthcare. We need to ca- take care of people who are mentally ill or older. We need to repair our roads. There's so much work that require at least partly, maybe entirely a human component. So rather than like write all these people off, let's find a way to repurpose them and keep them engaged. Um, now that said, I do... would like to see more buying power, um, from people who are sort of at the bottom end of the spectrum. The economy has been designed and evolved in a way that's, I think, very unfair to a lot of hardworking people. I see super hardworking people who aren't really seeing their wages grow over the past 20, 30 years, while some other people who have been, you know, super smart and/or super lucky have, uh, have had, you know, have made billions (laughs) or hundreds of billions. And, uh, I don't think they need those hundreds of billions to have the right incentives to invent things. I think if you talk to almost any of them, as I have, you know, they don't think that they need an extra $10 billion to, to do what they're doing. Most of them probably would, would love to do it for only a billion (laughs) -
- LFLex Fridman
(laughs)
- EBErik Brynjolfsson
... or maybe for nothing.
- LFLex Fridman
For, for nothing, many of them, yeah.
- EBErik Brynjolfsson
I mean, you know, uh, uh, uh, an interesting point to make is, is, you know, like do we think that Bill Gates would have founded Microsoft if tax rates were 70%?
- LFLex Fridman
Mm-hmm.
- EBErik Brynjolfsson
Well, we know he would have because there were tax rates of 70% when he founded it, (laughs) you know?
- LFLex Fridman
Yeah.
- EBErik Brynjolfsson
So, um, I don't think that's as big a deterrent, and we could provide more buying power to people. My own favorite tool is, uh, the earned income tax credit, which is basically a way of supplementing income of people who have jobs and giving employers an incentive to hire even more people. The minimum wage can discourage employment, but the earned income tax credit encourages employment by supplementing, uh, people's wages, you know? If, if, uh, the employer can only afford to pay him $10 for a task, um, the rest of us pick in- kick in another $5 or $10 and bring their wages up to $15 or $20 total, and then they have more buying power. Then entrepreneurs are thinking, "How can we cater to them? How can we make products for them?" And it becomes, uh, a self-, uh, reinforcing system where people are better off. Andrew Yang and I had a good discussion where he, uh, suggested instead of a, uh, universal basic income, he would suggested... or instead of an unconditional basic income, how about a conditional basic income where the condition is you learn some new skills, we need to reskill our workforce, so let's make it easier for people to, uh, uh, find ways to get those skills and get rewarded for doing them? Th- that's kind of a neat idea as well.
- LFLex Fridman
That's really interesting. So, I mean, one of the questions... one of the dreams of UBI is that you provide some little safety net while you retrain, or-
- EBErik Brynjolfsson
Right.
- LFLex Fridman
... while you learn a new skill.
- 1:13:03 – 1:19:09
Economics of innovation
- LFLex Fridman
So I'm a total... How do I put it nicely about myself? I'm totally clueless about the economy. It's not totally true, but pretty good approximation. Uh, if you were to try to fix our tax, tax system... And, uh, or maybe from another side, if there is fundamental problems in taxation or some fundamental problems about our, our economy, what would you try to fix? What would you try to speak to?
- EBErik Brynjolfsson
You know, I definitely think our whole tax system, our political and economic system has gotten more and more screwed up over the past 20, 30 years. I don't think it's that hard to make headway in improving it. I don't think we n- need to totally reinvent stuff. A lot of it is what, what I've heard elsewhere with Andy and others, called Economics 101. (laughs) You know, there are just some basic principles that have worked really well in the 20th century that we sort of forgot, you know, in terms of investing in education, investing in infrastructure, welcoming immigrants, having a tax system that, um, was more progressive and fair. At one point, tax rates were on, on top incomes, were, were significantly higher, and they've come down a lot to the point where, in many cases, they're lower now than they are for, for poorer people. Um, so, and we could do things like an earned income tax credit. To get a little more wonky, I'd like to see more Pigouvian taxes. What that means-
- LFLex Fridman
What's that?
- EBErik Brynjolfsson
... is you tax, uh, things that are bad instead of things that are good. So right now we tax labor, we tax capital, and which is unfortunate because one of the basic principles of economics, if you tax something, you tend to get less of it. (laughs) So, you know, right now there's still work to be done a- and still capital to be invested in. But instead, we should be taxing things like pollution and congestion. Um, and if we did that, we would have less pollution. So a carbon tax is a, you know, almost every economist would say it's a no-brainer, whether they're, um, Republican or Democrat. Greg Mankiw, who was head of George Bush's Council of Economic Advisers or, or, uh, Dick Schmalze, who is the, uh, another Republican economist, agree and, of course, uh, a lot of, uh, Democratic, uh, economists agree as well. If we taxed carbon, we could raise, uh, hundreds of billions of dollars. We could take that money and redistribute it through an earned income tax credit or other things so that overall, um, our tax system would become more progressive. We could tax congestion. One of the things that kills me as an economist is every time I sit in a traffic jam, I know that it's completely unnecessary.
- LFLex Fridman
(laughs)
- EBErik Brynjolfsson
It's, this is complete waste of time.
- LFLex Fridman
You can just visualize the cost and productivity that this is creating. (laughs)
- EBErik Brynjolfsson
All the pr- Exactly, because they are taking cost from me and all the people around me. And if they charged a congestion tax, they would take that same amount of money and people would, it would streamline the roads. Like, when you're in Singapore, the traffic just flows 'cause they have a congestion tax. They listen to economists. They invited me and others to go talk to them. Um, and then w- I'd still be paying, I'd be paying a congestion tax instead of paying in my time, but that money would now be available for healthcare, be available for infrastructure, or be available just to give to people so they could buy food or whatever. So it's just, it saddens me when, you know, when you sit, when you're sitting in a traffic jam, it's like taxing me and then taking that money and dumping it in the ocean, just like destroying it. So there are a lot of things like that that, uh, economists... And I'm not, I'm not like doing anything radical here. Most, you know, good economists would, would, would, I probably agree with me point by point on these things. And we could do those things and our whole economy would become much more efficient, it'd become fair. Invest in R&D and research, which is s- close to a free lunch is what we have. My, uh, erstwhile MIT colleague, Bob Solow, got the Nobel Prize, not yesterday, but, but about 30 years ago, for, um, describing that most improvements in living standards from, come from tech progress. And Paul Romer later got a Nobel Prize for noting that investments in R&D and human capital can speed the rate of tech progress. So if we do that, then we'll be healthier and wealthier.
Episode duration: 1:39:49
Install uListen for AI-powered chat & search across the full episode — Get Full Transcript
Transcript of episode NOReE-3EBhI
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome