Erik Brynjolfsson is an economist at Stanford. Please support this podcast by checking out our sponsors:
- Vincero: https://vincerowatches.com/lex to get up to 25% off + free shipping
- Four Sigmatic: https://foursigmatic.com/lex and use code LexPod to get up to 60% off
- ExpressVPN: https://expressvpn.com/lexpod and use code LexPod to get 3 months free
- Cash App: https://cash.app/ and use code LexPodcast to get $10

EPISODE LINKS:
Erik's Twitter: https://twitter.com/erikbryn
Erik's Website: https://www.brynjolfsson.com/
The Second Machine Age (book): https://amzn.to/33f1Pk2
Machine, Platform, Crowd (book): https://amzn.to/3miJZ76

PODCAST INFO:
Podcast website: https://lexfridman.com/podcast
Apple Podcasts: https://apple.co/2lwqZIr
Spotify: https://spoti.fi/2nEwCF8
RSS: https://lexfridman.com/feed/podcast/
Full episodes playlist: https://www.youtube.com/playlist?list=PLrAXtmErZgOdP_8GztsuKi9nrraNbKKp4
Clips playlist: https://www.youtube.com/playlist?list=PLrAXtmErZgOeciFP3CBCIEElOJeitOr41

OUTLINE:
0:00 - Introduction
2:56 - Exponential growth
7:24 - Elon Musk exponential thinking
9:41 - Moore's law is a series of revolutions
15:03 - GPT-3
16:42 - Autonomous vehicles
23:43 - Electricity
28:12 - Productivity
33:19 - Why is Twitter and Facebook free?
43:36 - Dismantling the nature of truth
46:56 - Nutpicking and Cancel Culture
53:11 - How will AI change our world
59:12 - Existential threats
1:01:05 - AI and the nature of work
1:07:11 - Thoughts on Andrew Yang and UBI
1:13:03 - Economics of innovation
1:19:09 - Effect of COVID on the economy
1:28:22 - MIT and Stanford
1:32:56 - Book recommendations
1:36:01 - Meaning of life

CONNECT:
- Subscribe to this YouTube channel
- Twitter: https://twitter.com/lexfridman
- LinkedIn: https://www.linkedin.com/in/lexfridman
- Facebook: https://www.facebook.com/LexFridmanPage
- Instagram: https://www.instagram.com/lexfridman
- Medium: https://medium.com/@lexfridman
- Support on Patreon: https://www.patreon.com/lexfridman

The following is a conversation with Eric Brinjalsen.

He's an economics professor at Stanford and the director of Stanford's Digital Economy Lab.

Previously, he was a long, long time 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.

4 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 1 of 2 camps, the fear mongering camp or the technological utopianism camp.

As always, the future will land us somewhere in between.

I prefer to wear 2 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 mention Andrew Yang, the presidential candidate who has been 1 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, 234 hours or more, and in person.

Even more so in the times of COVID, which requires getting tested nonstop, 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 Fat Stars on Apple Podcasts, follow on Spotify, support on Patreon, connect with me on Twitter at Lex Friedman.

And now, here's my conversation with Eric Brynjolfsson.

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?

Yeah, that quote, I remember posting that.

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 2 or 3 days and 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 I did my last in-person conference that week, I was flying back from Las Vegas, and I was the only person on the plane wearing a mask.

And the flight attendant came over to me, she looked very concerned, she kind of put her hands on my shoulder, she was touching me all over, which I wasn't thrilled about.

And she goes, you know, do you have some kind of anxiety disorder, are you okay?

And I was like, no, you know, it's because of COVID.

But I was worried because I knew I could see, or I suspected, I guess, that that doubling would continue.

And pretty soon we had thousands of times more cases.

Most of the time when I use that quote, I try to, 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 walk for 10 minutes, you get about 10 times as far away as if you walk for 1 minute.

You know, that's the way our physical world works.

But if something doubles for 10 times as long, you don't get 10 times as much, you get a thousand times as much.

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 that can be good in the case of things creating wonders, but it can be dangerous in the case of viruses and other things.

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 in which aspects of our society does exponential growth apply?

Well, you can learn the math, but The truth is our brains, I think, tend to 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, 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.

It's hard to feel, it's hard to retrospect and really acknowledge how much has changed in just a couple of years or 5 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.

It feels pretty calm and normal and gradual.

A lot of stuff, you know, I think there are parts of the world, most of the world is not exponential.

The way humans learn, 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 because some of them are exponentially more dangerous, and our intuitions and our human skills and our institutions that just don't change very fast at all.

And that mismatch, I think, is at the root of a lot of the problems in our society, the growing inequality and other dysfunctions in our political and economic systems.

So 1 guy that talks about exponential functions a lot is Elon Musk.

He seems to internalize this kind of way of exponential thinking.

He calls it first principles thinking, sort of the kind

going to the basics, asking the question, like what were the assumptions of the past, how can we throw them out the window?

How can we do this 10X much more efficiently and constantly practice in that process?

And also using that kind of thinking to estimate sort of when, you know, create deadlines and estimate when you'll be able to deliver on some of these technologies.

Now, it often gets him in trouble because he overestimates, like he doesn't meet the initial estimates of the deadlines, but he seems to deliver late, but deliver.

Like, what are your thoughts about this whole thing?

I think going to first principles, I talked about 2 ways of getting more of a grip on the exponential function.

And 1 of them just comes from first principles.

If you understand the math of it, you can see what's going to happen.

And even if it seems counterintuitive that a couple of dozen of COVID cases can become thousands or tens or hundreds of thousands of them in a month, it makes sense once you just do the math.

And first, I think he also benefits from hanging out in Silicon Valley.

And he's experienced it in a lot of different applications.

So it's not as much of a shock to him anymore.

But that's something we can all learn from.

In my own life, I remember 1 of my first experiences really seeing it was when I was a grad student and my advisor asked me to plot the growth of computer power in the US economy in different industries.

And there are all these exponentially growing curves.

In each industry, it was just taking off.

And 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 if it goes anything like that, we're gonna have orders of magnitude more computer power than we did at that time.

they often talk about it as just, so the exponential function is actually a stack of S-curves.

So basically it's, you milk or whatever, take the most advantage of a particular little revolution and then you search for another revolution and it's basically revolutions stacked on top of revolutions.

Do you have any intuition about how the heck humans keep finding ways to revolutionize things?

Well first let me just unpack that first point that I talked about exponential curves but no exponential curve continues forever.

It's been said that if anything can't go on forever, eventually it will stop.

And it's very profound, but It seems that a lot of people don't appreciate that half of it as well either.

And that's why all exponential functions eventually turn into some kind of S curve or stop in some other way, maybe catastrophically.

I mean, it went up and then at some point, it starts saturating the pool of people to be infected.

There's a standard epidemiological model that's based on that.

And it's beginning to happen with Moore's law or different generations of computer power.

The remarkable thing, as you allude, 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 1 and do that generation after generation with new materials, new processes, and just extend it further and further.

I don't think anyone has a really good theory about why we've been so successful in doing that.

It's great that we have been, and I hope it continues for some time.

But 1 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 2 to 3 years if there's 1 component of it That's not keeping up then the economic incentives become really large to improve that 1 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.

Do you think some version of the Moore's Law will continue?

I mean, 1 version that has become more important is something I call Cume's Law, which is named after John Cume, who I should mention was also my college roommate, but he identified the fact that energy consumption has been declining by a factor of 2.

And for most of us, that's more important.

The new iPhones came out today as we're recording this.

I'm not sure when you're going to make it available.

And for most of us, having the iPhone be twice as fast, it's nice.

But having 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 is probably more important for many applications than just the raw speed.

Other dimensions of Moore's law are in AI and machine learning.

Those tend to be very parallelizable functions, especially deep neural nets.

And so instead of having 1 chip, you can have multiple chips, or you can have 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 more important and we're seeing progress in them.

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.

It's fascinating to think, like, can I actually continue?

Us figuring out more and more tricks on how to train networks faster and faster.

As you look at image recognition, as you mentioned, I think it's a function of at least 3 things that are coming together.

1, we just talked about faster chips, not just Moore's law, but GPUs, TPUs, and other technologies.

I mean, we are awash in digital data today in a way we weren't 20 years ago.

Photography, I'm old enough to remember, it used to be chemical, and now everything is digital.

I took probably 50 digital photos yesterday.

I wouldn't have done that if it was chemical.

And we have the internet of things and all sorts of other types of data.

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, as they mentioned at OpenAI, there have been significant improvements in the techniques.

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 techniques of software and algorithms, and soon you're getting into millionfold improvements.

You know, somebody brought this up, this idea with GPT-3 that it's, so it's trained in a self-supervised way on basically internet data.

And that's 1 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.

I know, we make it to the point where it's consumed basically all of human knowledge or all digitized human knowledge.

But the interesting thing with bottlenecks is people often use bottlenecks as a way to argue against exponential growth.

They say, well, there's no way you can overcome this bottleneck.

ways to overcome whatever bottlenecks the critics come up with, which is fascinating.

I don't know how you overcome the data bottleneck, but probably more efficient training algorithms.

Yeah, well, you already mentioned 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 in China, but all around us.

In some applications, you can simulate the data, video games, some of the self-driving car systems or simulating driving.

And of course, that has some risks and weaknesses.

But you can also, if you want to exhaust all the different ways you could beat a video game, you could just simulate all the options.

Can we take a step in that direction of autonomous vehicles?

I'm actually talking to the CTO of Waymo tomorrow.

And obviously, I'm talking to Elon again in a couple weeks.

What's your thoughts on autonomous vehicles?

Like where do we stand as a problem that has the potential of revolutionizing the world.

Well, I'm really excited about that, but it's become much clearer that the original way that I thought about it, most people thought about it, like will we have a self-driving car or not, is way too simple.

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 noticed that you're right next door to Toyota Research Institute.

We're going to, we're supposed to talk, it's kind of hilarious.

So there's kind of the, I think it's a good counterpart to say what Elon is doing, and hopefully they can be frank in what they think about each other, because I've heard both of them talk about it.

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 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, driving through the snow 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 is going to do at these intersections that aren't really right angles and aren't very well described.

That's a much harder problem and requires understanding human motivations.

So there's a continuum there of some places where the cars will work very well and others where it could probably take decades.

So you mentioned 2 companies that actually have cars on the road.

There's the Waymo approach that it's more like we're not going to release anything until it's perfect and we're going to be very strict about the streets that we travel on, but it better be perfect.

Well, I'm smart enough to be humble and not try to get between.

I know there's very bright people on both sides of the argument.

I've talked to them and they make convincing arguments to me about how careful they need to be and the social acceptance.

Some people thought that when the first few people died from self-driving cars, that would shut down the industry.

Of course, there's still a concern that there could be setbacks if we do this wrong.

Your listeners may be familiar with the different levels of self-driving, level 12345I think Andrew Yang has convinced me that this idea of really focusing on level 4, 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 can 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 kind of another way of progressing to make it more feasible over time.

I mean, that's kind of like the Waymo approach, which is they just now released, I think just like a day or 2 ago, a public, like anyone from the public in the Phoenix, Arizona, to you can get a ride in a Waymo car with no person, No driver.

Oh, they've taken away the safety driver?

For a while now, there's been no safety driver.

Because, I mean, I've been following that 1 in particular, but I thought it was kind of funny about a year ago when they had the safety driver, and then they added a second safety driver because the first safety driver would fall asleep.

I'm not sure they're going in the right direction with that.

No, they've way more particular done a really good job of that.

They actually have a very interesting infrastructure of remote like observation.

So they're not controlling the vehicles remotely, but they're able to, it's like a customer service.

I bet they can probably remotely control it as well, but that's officially not the function that they use now.

can see that being really, because I think the thing that's proven harder than maybe some of the early people expected was there's a long tail of weird exceptions.

of the cases, but then there's something that's just never been seen before in the training data and Humans, you know more or less can work around that although let me be clear and note There are about 30,000 human fatalities just in the United States and maybe a million worldwide So they're far from perfect, But I think people have higher expectations of machines.

They wouldn't tolerate that level of death and damage from a machine.

And so we have to do a lot better at dealing with those edge cases.

And also the tricky thing that, if I have a criticism for the Waymo folks, there's such a huge focus on safety where people don't talk enough about creating products that people, that customers love, that human beings love using.

You know, It's very easy to create a thing that's safe at the extremes, but then nobody wants to get into it.

Yeah, well, back to Elon, I think part of his genius was with the electric cars.

Before he came along, electric cars were all kind of underpowered, really light, and they were sort of wimpy cars that weren't fun.

And the first thing he did was he made a Roadster that went 0 to 60 faster than just about any other car and went the other end.

And I think that was a really wise marketing move as well as a rise technology move.

It's difficult to figure out what the right marketing move is for AI systems.

That's always been I think it requires guts and risk taking, which is what Elon practices.

I mean, to the chagrin of perhaps investors or whatever.

It also requires rethinking what you're doing.

I think way too many people are unimaginative, intellectually lazy.

And when they take AI, they basically say, what are we doing now?

How can we make a machine do the same thing?

Maybe we'll save some costs, we'll have less labor.

And yeah, it's not necessarily the worst thing in the world to do, but it's really not leading to a quantum change in the way you do things.

When Jeff Bezos said, hey, we're going to use the internet to change how bookstores work, and we're going to use technology.

He didn't go and say, okay, let's put a robot cashier where the human cashier is and leave everything else alone.

That would have been a very lame way to automate a bookstore.

He went from soup to nuts, said, Let's just rethink it.

And that's been the story of these general purpose technologies all through history.

In my books, I write about electricity and how for 30 years, there was almost no productivity gain from the electrification of factories a century ago.

Now, it's not because electricity is a wimpy, useless technology.

It's because at first, they really didn't rethink the factories.

It was only after they reinvented them, and we describe how in the book, then you suddenly got a doubling and tripling of productivity 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 more creativity than most people have.

Yeah, well, sure, I'll tell you what happened.

Before electricity, there were basically steam engines or sometimes water wheels.

See all Lex Fridman transcripts on Youtube

Erik Brynjolfsson: Economics of AI, Social Networks, and Technology | Lex Fridman Podcast #141