Transcription
Um, thank you so much. Thank you to the mayor and thank you to all of Seoul for having me here. It is such a pleasure to be here. It does feel when you land in Korea like you are in the future. Um, and there's something to be said for that.
I'm going to talk to you today less about the technology and more about us because I've become pretty obsessed over the last 5 years with the relationship between human and machine. And in particular now the relationship between society and machine.
Now before I talk about societal thresholds, we should talk about what we're doing. What is actually the pursuit of AI? Right now we are building machines that possess human intellectual equivalence and soon superiority. That is something that we are getting quite good at. Right now, in the last 5 years, we've observed a near parabolic improvement in what the machines are capable of doing. Going from product recommendations to novel science support.
But somehow even more important than how good the machines are getting, in my opinion, is this. How inexpensive the machines are getting. Somehow, more importantly and lost I think on most people, is the rapid decline in inference cost. Some of this of course was sparked by Deep Seek's R1 in January this year, but what we are seeing right now is a precipitous decline in the cost to actually run the models. To use the technology.
And I care about this for a lot of reasons. One, I care about this because anytime in history we've observed a critical resource decline in cost this much, there has been a very good economic outcome. Water, foodstuffs, electricity, the internet, these all followed the path from scarce to abundant and good things happen. And the other reason I care about this, it's the best evidence we have that AI is approaching utility or commodity. Outside of the frontier models, people will have access to the best technology at exceptionally low cost.
And that I think means this. I think that the world will soon have access to unmetered intelligence. This is the biggest idea that I could sell you today and I have only 15 minutes. So instead of talking to you about all that it is, let me tell you what it is not. Unmetered intelligence is not universal brilliance. This does not mean that everyone in the world will be smart. It means that everyone in the world will have access to brilliance and what we do with it will be up to us. In the same way that just because you can read doesn't mean you do, in the same way because you have the internet does not mean you do research, unmetered intelligence does not mean that everyone will behave brilliantly.
But as I start looking at unmetered intelligence, what are the implications? The big question that I keep coming up to is in a world of unmetered intelligence, what is the limit of AI adoption? What is the limit of automation? Where do we actually decide that the human experience is sufficient? And that's led me to explore the theory of societal thresholds. A technological threshold asks the question simply, what can a machine do? That's all it asks. And a societal threshold asks the question, what do we want a machine to do? And I care about this a lot now for a number of reasons, not least of which is companies are going to build things that no one wants and we'll talk about that in a second. And also companies are going to build things that we should not want. There are a bunch of places that we could use this technology that we should not. There are a bunch of things that we should do to protect our humanity.
But in order to explore this better, I've gone back in time recently to a to study the history of societal thresholds. Cuz what I really want to know is how and why does a societal threshold move? How and why does a does a society decide that a technology is something that we can adopt? And there are a bunch of historic examples. Nuclear power, CRISPR, GMOs, tons of interesting technologies that we could use but we don't. But one historical one is really fun. The elevator.
In 1854 Otis Elevator invents the modern people mover and starts selling a bunch of these elevators. By 1855 hundreds of tall buildings have a purchased a thing that safely moves people up and down. The problem, by 1856 a bunch of these tall building owners have called up Otis, no one wants to ride in the elevator. Why not? Well, they're afraid they're going to die. So Otis has a problem. He's met the technological threshold. He has built a machine that can safely move people up and down a building, but he hasn't met the societal threshold. He hasn't convinced people that they should use it. So he goes back to the drawing board and comes up with three ways he thinks he can convince people to use his elevator. The first, he puts a mirror in the back of every elevator. So when you get in, you'll be a little distracted by your own image. The second, he puts music in every elevator. So when you get in, you'll be a little entertained. And the third, he puts an operator in the elevator. He puts a physical person in every elevator to push a button. Giving you the sense of human agency. And it works. It moves the societal threshold. And then we all use elevators. And now we can look back and sort of laugh.
That introduces a really important question, which is what are our elevators? What are the things that we should obviously be using this technology for that we are not? And it is wonderful to have Martin here because the obvious answer to our AI societal threshold is the autonomous vehicle. Now so much has been said about the autonomous vehicle and I want to be careful not to be too redundant. But what is remarkable today is this. 1.3 million people will die on the roadways this year. 1.3 million. The autonomous vehicle polls at 25% approval globally. 25% of the population wants an autonomous vehicle on their road. Why? Why? Why do we stare in the face of catastrophe and say no to the most obvious solution?
Well, we went and studied this. And what I'm going to present to you I think are the three reasons that a societal threshold moves as told by the autonomous vehicle. The first reason is really simple. We love control. Humans love control. When I first started doing this study, I had this idea that there was a freedom element. Some of this was, this is a true story. A woman came up to me after a talk. I was espousing the importance of Waymo and Tesla and the progress they'd been making in 2019. And a woman came up to me afterwards, she put her foot she put her finger in my chest and she said, you shouldn't talk about autonomous vehicles like that. It's very un-American. Cars are freedom machines and if they take away our cars, what's next? I said, okay, my bad. So I had this idea that it was control. All right, I had this idea that it was freedom. It's not freedom.
And we figured this out by studying two populations. We went and studied a bunch of really wealthy people. Really wealthy people it turns out like to fly planes and race cars. And they don't race those cars because they're going anywhere. It turns out they just ride around a track. They come right back to where they are. And they don't fly the planes because there aren't pilots to do it. They like these things because of the control. And then we went to Disneyland. Anybody here been to Disneyland recently? Anaheim? Thank you very much. Disneyland has added a very famous ride called Autopia. Autopia, built in 1954, one of the one of the six original rides at Disneyland. It is a it is a ride on which children age 3 through 8 drive around a cement circuit 2 miles an hour on gas-powered cars. It is ugly, it smells bad, it looks bad. Disney hates it because it has no Disney representation and yet every time they go to change it, every parent says, don't get rid of this ride. Why? Because since 1954, its inception, it has been the most popular ride at Disneyland. More than Space Mountain, more than the Matterhorn. They don't talk about it, but it's the ride that every 3 through 8-year-old wants to do. Why? Because everyone knows that when you put your foot on the pedal and you turn a wheel, there is a visceral thrill. You don't need to be a genius or even an adult to appreciate the power of controlling a machine. And what we now face in a world where we are asking machines to take over things on our behalf, even those things that are killing us, it is going to be very hard for a lot of people to surrender that control.
The second reason, technological thresholds are moving way too quickly. Martin knows this problem exceptionally well. When I started doing this research in 2019, autonomous vehicles were not that good. When I started talking about this problem in 2015, it was a pie in the sky and they were decidedly dangerous. By 2019, they were good but not great. Human drivers were still better in a lot of conditions. Then something happened. By the time we finished the study, autonomous vehicles had gotten exceptionally good. Guess how much that moved the polls. Guess how much that changed people's opinions on autonomous vehicles. Martin knows this answer, zero. And in fact, it appears that people are less interested in autonomous vehicles today than they were 4 years ago. The answer to this is pretty clear. Most people have no idea how good these machines are. Outside of Pittsburgh and Phoenix, San Francisco, Austin, a few places around the world, most people have never seen an autonomous vehicle, fewer people yet have been in one. And they have no idea this technology has actually progressed to a place where we want it.
And I understand this problem exceptionally well because I spoke at a medical conference 2 months ago espousing the progress of imaging in radiology pathology. And when I was done, a radiologist in a crowd of 2,000 people stood up, raised his hand, and he said, "You shouldn't talk about it that way. Radiology pathology is still not very good with computer vision." And he cited a study from December 2024. And I had to say, "Sir, I'm sorry to embarrass you, but that study is old. Radiology pathology as of last month is in fact now a solved problem for every scan except the cardiology chest scan." If a radiologist does not know that his industry has been turned over by technology, what chance does the average person have? That is a huge problem as we build these applications helping people go through the journey so that the societal threshold can catch up to the technological progress.
But the last reason is the most important. It's the one that I cannot stop staring at, and it's the one that I think we will talk about basically forever. Humans have an exceptional tolerance for machine failure, and we have none for human failure. We are Sorry. Humans have an exceptional tolerance Thank you, Martin. Have an exceptional tolerance for human failure, and we have none Exceptional tolerance for for human failure and none for machines. This is the problem. This is why we can watch a Tesla crash on autopilot and declare the whole program a failed state. We can watch a Waymo swerve off the road and declare everything to be shut down. But when a drunk driver causes a huge car pileup, we go, "Oh, that's just the cost of doing business. That's just how it goes." This is a huge problem. This is a weird problem, and I don't have a bunch of solutions, but I have two proposals.
The first, humans believe that mechanical failure is systemic failure. When we see a Tesla crash, our assumption is all Teslas will crash. In fact, the opposite is true. We can solve mech- mechanical problems systemically. Humans do not possess this quality. We fail every day in the exact same way over and over and over collectively and individually. Getting older is just getting a little bit wiser such that you're failing a little bit less. That's all it is.
The second reason is even more fascinating in my opinion. We have no empathy for machines. We are very willing to watch each other try and fail. We are very willing to say, "That's okay. You did your best." We make movies about a lot of Olympic failures. But when your washing machine breaks, you're like, "All right, time for a new one. Get this piece of crap out of here." This is an interesting phenomenon, and I actually think one that we should embrace. The precision that we expect from machines, but not humans, is a fascinating space between humans and machines, and one that has actually protected us for a long time. It's why this building will never collapse. It's why planes don't fall out of the sky. But it is going to set a very high boundary for our adoption of machines, especially in cases where they can save our lives.
Thank you so much.