Plain English with Derek Thompson - The Self-Driving Revolution Is Real—and It Could Be Spectacular

Episode Date: November 15, 2024

What would a world of self-driven cars look like? How would it change shopping, transportation, and life, more broadly? A decade ago, many people were asking these questions, as it looked like a boom in autonomous vehicles was imminent. But in the last few years, other technologies—crypto, the metaverse, AI—have stolen the spotlight. Meanwhile, self-driving cars have quietly become a huge deal in the U.S. Waymo One, a commercial ride-hailing service that spun off from Google, has been rolled out in San Francisco, Phoenix, Los Angeles, and Austin. Every week, Waymo makes 150,000 autonomous rides. Tesla is also competing to build a robo-taxi service and to develop self-driving capabilities. There are two reasons why I’ve always been fascinated by self-driving cars: The first is safety. There are roughly 40,000 vehicular deaths in America every year and 6 million accidents. It’s appropriate to be concerned about the safety of computer-driven vehicles. But what about the safety of human-driven vehicles? A technology with the potential to save thousands of lives and prevent millions of accidents is a huge deal. Second, the automobile was arguably the most important technology of the 20th century. The invention of the internal combustion engine transformed agriculture, personal transportation, and supply chains. It made the suburbs possible. It changed the spatial geometry of the city. It expanded demand for fossil fuels and created some of the most valuable companies in the world. The reinvention of last century’s most important technology is a massive, massive story. And the truth is, I’m not sure that today’s news media—a category in which I include myself—has done an adequate job representing just how game-changing self-driving technology at scale could be. Today’s guest is Timothy Lee, author of the Substack publication Understanding AI. Today I asked him to help me understand self-driving cars—their technology, their industry, their possibility, and their implications. If you have questions, observations, or ideas for future episodes, email us at [e-mail]. Host: Derek Thompson Guest: Timothy Lee Producer: Devon Baroldi Learn more about your ad choices. Visit podcastchoices.com/adchoices

Did you know that scientific studies have found most people lie once every 10 minutes? In my new podcast, Truthless, I'm talking to people about the lies, they tell, from faking illnesses in high-pressure moments to making up stories on national TV. From Spotify and the Ringer Podcast Network, I'm Brian Phillips. Listen to Truthless on Spotify or wherever you get your podcasts. Today, the self-driving revolution and how it could change the future of transportation. There's a famous graph in tech history that's called the Gartner hype cycle. It looks like a wave, up and down and up, where the X-axis represents time, and the Y-axis

represents hype. So a new technology debuts, and, of course, expectations go through the roof, the peak of excitement. But almost no invention arrives fully formed. It needs a series of micro-inventions to make it cheaper, safer, more effective, more scalable. And because it's incredibly difficult to scale a new technology, inevitable disappointments tend to plunge this new technology into a pit of despair. But over time, people continue to work and tinker and the tech gets better in the absence of any circus of hype, until it's eventually, suddenly, everywhere. The Gardner hype cycle is a cliche in tech, but that's because, like most things that become cliches, it's true enough. And nowhere is it more true than in the brief history of self-driving cars.

In 2005, DARPA, the Defense Advanced Research Projects Agency, which was founded during the Cold War under Eisenhower, created a grand challenge for vehicles to drive themselves. It built a 100,000, mile route through the Mojave Desert, and the winner of that 2005 competition was a team from Stanford led by a professor named Sebastian Thrun. In 2009, Google's co-founders hired Thrun to start a secret research project, to build an autonomous vehicle, a self-driving car. For several years, Thrun worked at X, the company's innovation lab, building the software and hardware necessary for a fleet of cars that could orient themselves in real-world conditions. After all, driving through a desert is nice, but there are no people in the deserts, no crosswalks, no pedestrians, no drunk drivers.

Google's first autonomous fleet consisted of modified Toyota Priuses, to which they attached cameras, radar, a technology called LiDar, which shoots lasers at a little spinny gadgets from which the car creates a 3D map of its surroundings in real time. In 2016, Google named this self-driving car company, Waymo. It was around this time in the mid-2010s that I started picking up on this big surge and hype around self-driving cars. I attended events and conferences where techno-optimists and auto executives confidently predicted that we would have fully autonomous vehicles sluicing through the streets by the end of the decade. By 2020, we were invited to imagine a world in which self-driving, had completely remade the city, and by extension, the rest of the country. But 2020 came and went, and self-driving cars were practically nowhere to be seen,

from the peak of excitement, down into the trough of disappointment. For self-driving technology, it seemed, the last mile problem was painfully literal. Meanwhile, autonomy wasn't changing the roads, but it was being subtly built into the driving, cars that we did buy and own. If in the last few years you've bought a mildly fancy car, it will have some self-driving capabilities or self-parking capabilities

that just didn't exist several years ago. So rather than a big revolution in self-driving, cars without drivers, it seemed like we were stuck with a smaller revolution in driving. Cars with drivers, but got a little bit better every year in taking the wheel from you.

But while many people's attention has flittered to other sparkling points on the tech horizon, including generative AI programs like chat GPT, self-driving cars have quietly become a reality. Waymo One, a ride hailing service like Uber, has rolled out in San Francisco, Phoenix, Los Angeles, and Austin. The company is looking to go deeper into trucking as well. As the Gartner hype cycle curve predicted,

the peak of excitement and the trough of disappointment, have been followed by the slow and steady march of technological progress. There are two reasons why I've always been fascinated by self-driving cars. The first is safety. There are roughly 40,000 vehicular deaths in America every year and 6 million accidents. It's appropriate to be concerned about the safety of robot-driven cars, but what about the safety of human-driven cars? We're not that great at driving.

And a technology that was better at driving, less distracted, more reactive, could save thousands of lives and prevent millions of accidents. That is a huge deal. Second, the automobile was arguably the most important technology of the 20th century. The invention of the internal combustion engine transformed agriculture, transportation, supply chains. It made the suburbs possible. It changed the spatial geometry of the city. It expanded demand for fossil fuels and created some of the most. valuable companies in the world. The reinvention of last century's most important technology

is a massive story. And the truth is, I'm not sure that today's news media, a category in which I include myself, has done an adequate job representing just how game-changing self-driving technology at scale could be. Today's guest is Timothy Lee, the author of the Substack Understanding AI. Today, I ask him to help me understand self-driving cars, their technology, their industry, and their implications. I'm Derek Thompson. This is plain English. Tim Lee, welcome with the show. Derek is great to be on.

I'd love you to make me smarter first about how self-driving cars work. I remember from my first ride in AWAMO maybe six or seven years ago out in the Bay Area, they've got a bunch of cameras mounted around this vehicle. It's like it's a normal car with a bunch of cameras in the top and on the sides, including the famous LIDAR machine, which creates a 3D image of the road. Maybe it's best to ask the question this way. How do self-driving cars perceive the world? And how do they use that perception to orient themselves through the world? Sure.

So on the first question, how do they perceive the world? They've got a bunch of sensors, but the three most important are cameras, which you mentioned, which are just digital cameras, just like you have your phone. On radar, which is the same are in some conventional cars and on boats and stuff, which those are most useful for telling velocities. So if you're like on the freeway and you want to see if the car had to be with decelerating, like radars are very useful for that. And the third one, which is most unique to self-driving cars is the LIDAR. The Waymo cars, for example, have these big spinning things at the top that have lasers. They fire laser beams in every direction several times a second. And they create a 3D point cloud.

And so this gives them a very precise map of all the physical objects in the world. And then they'll combine the data from the ladder with the data from the cameras, because the cameras will kind of give a visual of what is this? And so they know there's an object there. The camera helps them figure out what's there. And then they end up with this 3D model of how the world works. And so the software that does that first step of like kind of building the model that's usually called the perception part of the self-diving software.

And then it does a lot of kind of internal simulation where it maps out. Here are some possible ways the future could unfold. And then based on those possible scenarios, it figures out, like, what's the safest thing to do in the next second, the next two seconds, the next three seconds, it plans out a trajectory. And this ends up working pretty well. And how similar is the technology being used to make sense of the world? How similar is that to the transformer technology that powers large language models like chat GPT? So this has been something that's been changing in the industry recently. I think we'll talk about later about how this industry has evolved.

But self-driving cars have been around for about 15 years. The transformer is only about six or seven years old. And so early on, they had different techniques. But in the last three or four years, the leading companies have always started using the transformer in its self-driving software. And the architecture is pretty similar. It's the same basic approach. I don't think we need to get into details of how it works. But the transformer is very good at taking large amounts of data and extracting patterns from it.

And what that is meant is that the more recent generations of self-dribes, driving car software, do much more learning from experience. If you gather lots and lots of data of real world human driving or real world self-driving car driving with feedback of whether it did it right or not, it can then extract patterns from that and say, okay, in this particular situation, I should be this far from the curb, I should give the driver ahead of me this much space, because part of what's difficult about driving is there's the rules you see in the handbook, but then there's also lots of kind of subtle rules of thumb and dynamic kind of situations that are really hard to write,

explicit software rules. And so neural networks, in particular, transformers are really good at just looking at how have people done this in the past and figuring out implicitly, like, what's the role that the human is following here and then doing a similar action that makes the car drive as it seems like it's a human driver. Yeah, it's kind of funny to think about an analogy between chat, ChbT, getting smart about the world by reading Reddit, by reading human impressions of what is true and untrue, and a car making sense of the world, rather than reading Reddit, it's reading the road, it's reading human interactions, it's reading the fact that, you know, sometimes that pedestrian looking the other way will sometimes walk into the street.

And so you have to slow down as you're reaching the intersection because pedestrians, you know, don't have perfect LIDAR awareness of what's around them and they'll walk in the other street. And so they're essentially reading the world and processing the world in a way that is, um, is similar to in some ways the way the Chachibati makes sense of the corpus of the internet. Yeah, one of the places of the Transformers has been most useful is, the earlier generations of self-driving cars really had trouble at intersections with a lot of different vehicles. Because to know what you should do, you have to predict what the other cars are going to do. But then what they are going to do is going to depend on what you do and what the other cars do.

So you have this like computational explosion of a number of possible permutations you have to think of. And it's a little ways that it's a little bit like the conversation the cars are having, where each car sort of makes a move and then the other ones respond. And so in the same way that LLM predicts the next word, what the transformer does in the prediction part of this software is it kind of predicts the next move. It says, okay, each car gets a move, and we build a network that kind of learn the strategy that each car plays, and that kind of plays this game out, you know, inside a data center, inside a computer. And they found that that work much better than a computer

programmer trying to explicitly write out, okay, the rule that each car should follow is X, because it's very complicated, and you have this kind of computational explosion if you try to do it very explicitly. There is a really beautiful and profound idea here that we're not going to go into deeply that so much of what we think of is knowledge and life, is synthesis of information plus next moment prediction. Like this is what so much perception is, this is what conversation is, I don't know the words that are about to come out of my mouth

as I speak to you, but my brain is doing some kind of transformer business that's allowing me to process the world and sort of speak next tokens, predict the near future. One of the reasons I find the I revolution so interesting is that I think philosophically these similarities between how a technology speaks English

and how a technology drives on a road. It doesn't seem like these are the same process, but there are philosophical and computational ways in which they are the exact same process, and I find that incredibly rich. We're not going to do a whole philosophy 410 seminar as much as I would actually love to kind of do that. Let's talk about Waymo.

In my open, I offered a brief history of self-driving cars up to the moment. Why don't you plunge us into the moment with what's going on with Waymo, the industry leader in self-driving? What's the state of play? How many rides a week is, Waymo giving and how fast is this service going? Yeah, so Waymo is active in about four cities.

They have significant installations in Phoenix and San Francisco. They're scaling up Los Angeles quite quickly. They just opened the, remove their wait list in Los Angeles. So I think hundreds of thousands of customers have access now. Austin, they're just starting out in. And they just recently announced they're doing 150,000 rides per week, which is up from 10,000, about 18 months ago. So, you know, 15 acts over a year and a half.

I don't know if they'll continue with that, right? But you play that out two or three years, and it's going to be in a lot of cities. And it has the potential to be in a lot of cities and be used by a lot of people quite soon. And just to get people a bit of context, 150,000 rides a week is Waymo. Uber, I found, is about 130 million rides a week.

Does that sound directionally accurate, essentially, that Waymo does one-oneth, the number of rides as Uber? Yeah, I have not looked up that specific statistic recently, but yes, that sounds about right. I mean, it's still a small thing. I mean, it's only in four cities. There's dozens of cities in the U.S.

And in the cities where Waymo is, they're still in restricted areas. So like in San Francisco, it's just in the city of San Francisco and a little bit down the peninsula, but, you know, Oakland or East Bay or any of those, you know, Marin, there's lots of parts of San Francisco Bay Area they're not in yet.

And so we almost still has a ton of room to grow. But it's also on the flip side, it's more than a token kind of scale. I mean, two or three years ago, they had like hundreds of rides a week. And so they're in the kind of steep part of the growth curve right now. And if you figure if they're doing 10x growth every year or two, that means that they could be Uber size in like three to five years.

So I don't know if that growth will continue, but they're growing at a rate where they could get to be a serious player in this market in like a single-digit number of years. So the big mystery to me about self-driving cars, as I took my eye off the ball over the last few years, is in 2017, 2018, you had all these auto executives saying, we're going to get self-driving all over the road in three years, five years, in the early 2020s,

you're not going to be able to believe how many cars are self-driving. And 2020 came and went, and self-driving cars were basically nothing. They were pittance, like, you know, maybe like 17 total rides in a week or something. But as you said, just in the last two years, while everyone's been looking at large language models and chat to BT and talking about, crypto and then not talking at crypto and talking about the metaverse, then not talking about the

metaverse. Meanwhile, inch by inch, mile by mile, you had self-driving cars really start to take off, especially in these four cities. What made the difference? I think there was any one thing. So there's this classic model car, the Gartner hype cycle that applies to a lot of technology like that. You think about the dot-com boom. You had a big boom in the 1990s. Early 2000, people said, oh, that internet thing's kind of over. And then late 2000s, early 2010, it's, it's it came around back and in fact turned out in fact it was a big deal. I think you saw the same thing with self-driving cars in 2015, 2016, 2017. There was a ton of hype around this technology. I think Ford said they were going to have a car with no steering wheel by 2021. I think Lyft said that by 2021,

half their rides would be driverless. And so a ton of money came in, a ton of people invested. And it just turned out the problem was harder than people thought. And in particular, what people call edge cases, you know, weird situations that you can build a car on private test. or do simulations in the show, you know, this works really well. You can do demos and say, look, it can drive this route really well. But it turns out the world that's really complicated. And human drivers, people say humans are bad drivers, but in some ways they're really good. I mean, you know, human driven vehicles only get in a fatal crash, but once every 100 million miles.

And so if you have a self-driving technology that crashes once every 10,000 miles, in some sense, that's going to seem pretty impressive. You know, you can do a demo and it can go 20 miles and not make a mistake. You just say, that's amazing. But that might still be far short of what you need to actually be. safely you came a driver, which is a standard that most of these companies are holding themselves to. And so in 2018, Google kind of tried to, you know, Waymo kind of tried to launch a driverless taxi service. I think they weren't confident enough to the technology. They kept safety drivers.

They did a very small scale. And it seemed like kind of nothing was happening. They had this 40 square mile area of the Phoenix metro area where they had like a token number of people driving around and lots and lots of safety drivers. And anyway, and it kind of wasn't clear from the outside what was happening. But then in 2020, they actually launched the driverless service. It was still very small, but it grew and grew and grew. And I think they were just kind of plowing through that list of educations. They just had to have the car make of a stake, go back to the engineers. They had the car made this mistake.

How can we fix it? And they just had to iterate over and over again until they got to the point where it was safe enough. They felt confident to take the driver out. And the other thing they've done is they've very gradually increased the difficulty. So they started in the suburbs of Phoenix, which is sunny, never snows, wide straight streets, well-painted lines, low density. And that was a good testing ground, but a terrible market.

And so over time, it's kind of the harder, the better markets also tend to be more lucrative. I mean, the more lucrative markets tend to be more difficult technically because you think about airports, you think about downtowns. You think about, you know, streets with a lot of bars late at night. These are with a lot of chaos, a lot of unexpected things happening. And so, yeah, first they did suburban Phoenix, then they did the residential parts of San Francisco. than they did downtown San Francisco, then they try to do airports. And so they've kind of gradually worked their way up the kind of difficulty level. The one place Wemmo doesn't do yet is freeways.

They started testing drivelessly on freeways, but just because of the high speeds, the cost of a mistake is much, much higher. And the strategy these cars have, if they do come to a situation they understand, is they'll slam on the brakes and stop. It's very safe to do that if you're going 10, 15, 20 miles an hour. If you're going 70 miles an hour,

that's much harder to reliably do safely. And so anyway, that's the one kind of big, I think, situation that the way more cars don't deal with yet. But they are now doing driverless testing. I expect maybe next year that they will begin to do freeways. And so they're getting very close to the point where they can service basically all the situations we find. So the Great League Forward over the last six years has been this steady ability to solve these edge cases, to be smarter about driving in the world of humans. And it's telling that the grand challenge of DARPA, which would have to be a steady ability to solve,

of kick-started this era of self-driving cars. That was the Mahavi Desert, right? So there are no people there. There are no bars there. There's certainly no snow in the Mojave Desert. So there's ways in which a DARPA design course in Mojave Desert is the easiest of all possible easy modes. And then something like a Vermont college town on a Saturday night when it's snowing is like the hardest of all possible places because you've got drunk kids spilling out in the street and there's white falling from the sky. It makes me. think, as I'm putting this together, you mentioned the four cities that they've expanded in, Phoenix, San Francisco, Los Angeles, is it Austin? Austin, yep. The annual snowfall, and that's

Atlanta, okay, five, Atlanta. The annual snowfall in those five cities is probably about 0.1 centimeters or something. Do, why, maybe there's an obvious answer here, why is Waymo still afraid of the snow? I think that's just they want to do things incrementally. So snow is the category they haven't really solved. I don't think it's going to be a huge difficulty, but I think it's just like they can go anywhere they want, and snow is just one extra kind of difficulty factor. And so given they can have a profitable business serving a bunch of southern and southwestern cities, I think they're just starting with those. So yeah, I absolutely expect we'll see Texas, Florida, you know, probably Oregon, Washington State, Nevada,

the states like that before we get to Chicago or New York City or Boston. But I don't think that's the bottleneck for them right now. And I think two to three years from now, when they're ready to kind of go national, I think they will, I mean, they've been testing in snow. I think they'll have snow figured out by then. It sounds to me like we could be at the beginning of a real explosion in the number of cities where Waymo is operating. Is there, if my curiosity here is, why isn't the technology moving faster? If it is at the point where it's solved edge cases, why aren't there Waymo's everywhere? Why isn't there? this the most obvious technical revolution in the world, is the bottleneck to growth right now

on the manufacturing side? They just can't build enough of these computers and computer outfitted cars fast enough, or is there another implementation reason why Waymo is as successful as it is in four, almost five cities, but it's not in 50 cities? I think it's a function of the business model they've chosen. So they're building a taxi service, and a taxi service is not like a smartphone app where you can just push a button and push it out to 100 million users. You need a lot of different things to make a taxi service. So they're owning their own fleet, which is different from what Uber did, because obviously people don't have self-driving cars, so you can't have them like bring their own self-driving cars. So they have to build and own their own self-driving cars.

They have to have local operations, both to interact with local officials, but also they need people to clean and repair the cars. And then they also, um, um, um, um, um, um, um, um, um, um, um, um, um, um, um, um, um, And then they also need the technology to work in all these different places. And so all of those things interact where, like, one of the bottlenecks absolutely is manufacturing. So anytime you manufacture something in a factory, the larger the scale you can manufacture at, the cheaper the cost is. And so right now when those vehicles are quite expensive, the rumor is $100,000, $150,000 per vehicle. And a big reason for that is that they take off-the-shelf vehicles and they basically tear them down and rebuild them with the sensors and the computer inside. it would be much cheaper to redo an assembly line

where you're putting the sensors in the compute in the car as it's going down the assembly line. But a car company is not going to talk to you until you're putting orders in in the order of tens of thousands of vehicles. If they only have five operating cities and they're only operating a portion of those cities, they'd get 50,000 vehicles.

They wouldn't know what to do with them. And so there's this chicken and egg problem where they need to have operations in a bunch of cities, they need to have staff in a bunch of cities. They need to have the technology. They can handle a bunch of different situations. They need to have a big fleet of cars. And all of those things kind of support all the other.

And so I think the 10x growth we've seen in the last year or so is about as fast as you can expect the technology like this to work because there's a lot of just tedious, you know, real estate deals and hiring and stuff you need to do to be ready to have 10 times as many cars. But then you need to have that infrastructure in place in order to order the 10 times of many cars and use the bigger orders to get the lower prices, which then makes it economically viable. And so I think they're just like climbing that that hill in a way that anyway. I mean, I think even if you look at Uber or Lyft, they didn't go from zero to $100. cities overnight. They probably went a little faster than Waymo has, but I think probably took like three to five years for them to get in kind of all the major cities. And so I think, I think WIMO will be a little slower than that, but it'll be roughly to say time span, it'll be like, you know, probably three to seven years to get to the point where it's in most cities, at least

across the, you know, the south and the west. Before we move on to industry analysis and looking at Waymo in the context of some potential competitors, I definitely want to check the box of safety. You mentioned that humans make an error, or maybe it's a fatal error, once every 100 million miles. How safe are Waymo vehicles compared to humans right now? So we can't say anything definitive about fatalities yet, because in total, Waymo's vehicles have done about 25 million total miles. And so if they were as safe as humans, you would expect zero fatalities. There's been zero fatalities. So they're not less safe than a human, but we can't say they're necessarily more safe.

But if you look at less serious vehicles, less serious crashes, they've been looking pretty good. So one category that Waymo has been collecting statistics about is crashes serious enough to trigger an airbag injury. And I believe their statistic, as they estimate, it's 84% less likely. So it's about six times more likely to have an airbag crash with a Waymo vehicle versus a human vehicle. And this is, they tried to pull data for a similar area.

So like this is San Francisco and Phoenix drivers in the areas where they operate. for crashes involving injuries. I forget the exact number, but I think it's about one-third as often. So it's about roughly three times of the safe. And another thing that I think, another sign that is that these have a pretty good safety record, if you look at the kinds of crashes they get into,

the overwhelming most common crash is when mobile looks in front, another vehicle rear ends them from behind. There's no example of describing all the other way where the Waymo vehicle rear and somebody else. There's several examples where a car ran a red light and crashed into a Waymo. I don't think there's any examples of a Waymo running a red light and crashing into somebody else. And so there's crazy drivers out there.

Even a perfect driver is going to get in crashes sometimes. And if you just look at the 20 or so serious crashes that Waymo's have gotten into, 70, 80, 90 percent of those are number 120 is much fewer than you'd expect from human drivers in the same areas, but then a large majority of those are pretty clearly the fault of the other driver. And there's a few where you could kind of quibble and say, maybe Waymo could have handled this a little more gracefully. But there's no cases of really briefest ones where Waymo did something like clearly illegal or

unsafe. Like there's a couple cases. There was one case where Waymo was coming in one direction. A bike was coming from the other direction was behind a garbage truck. And then kind of the last minute turned left in front of the Waymo. Waymo slammed on the break. You could say, well, maybe it should have anticipated that or been able to break faster or whatever. But that's like, I think of a human being, if that would at least have some specific view it,

like maybe they did the best they could. The worst waymo prices are that kind of thing where you could. Maybe they should have handled it better, but there clearly seemed to be some fault on both sides. To broaden this out from Waymo for a second, I want to talk about this industry at large. At the highest level, I think it's pretty interesting that two of the most important evolutions of automaking this century are the rise of electric vehicles and the rise of self-driving vehicles. And notably, the big automakers did not lead either revolution. This is not a place where GM or Ford led in electric or in self-driving. Tesla is the leader in electric, and Waymo is the leader in self-driving.

And I do want to just pause here to point out that I do think that's very strange and interesting. The CEO of the most sophisticated rocket company in the world is leading the electric car revolution in America. And the founders of the dominant search engine put together the team that became the top self-driving car company in America. It does seem like a really interesting case of the innovator's dilemma, of upstarts coming in where the incumbents are not. In your mind, how did Waymo do it? Which is to say, what makes Waymo the dominant player in this space? What makes the technology best? I think they just started the earliest and a Timble, a really great team,

and then they had bottomless financial resources. I actually got this wrong. Earlier in my career, around 2019, when it seemed like Waymo wasn't launching, and there were all these startups that had plans to do simpler things. Like, there were companies, for example, that were going to do delivery robots. that drive on the street, but maybe have a top speed of like 20 miles an hour. And because there's no people in the vehicles that's maybe safer. And so I thought companies like that,

there was another company that was going to do a taxi service on a retirement community where the maximum speed was 25 miles an hour. And so I thought that's like an easier problem. So I thought it was kind of disruptive innovation story where these like startups with like easier problems, kind of scrappy would lap Waymo. And then all those companies went out of business. And most of them went out of business. Some are still around but are not, you know, in the position of Waymo is in.

And what I think I missed was that there's actually not like a, there's not like an easy subset of the problem. Because even if you pick an area where it's less likely that somebody will run out in front of your car or, you know, whatever is crazy thing is going to happen. Everything that could happen, every crazy thing that could happen somewhere could happen anywhere. And so you really just need to kind of hit all the, you need the general solution. And there's no alternative to that than just like grinding through it and putting, millions of miles on the road and experiencing all the problems with safety driver, which is expensive, and then having your engineers fix it and trying it again. And so I think it's just a matter of Google has bottomless resources. Larry and Sergey have control over Google, and so they can spend

those bottomless resources as long as they want, and they have just kept spending and spending and spending. They spend, I don't think they've disclosed the exact amount, but it's like billions of dollars every year for like 15 years on this. And nobody else has the kind of financial resources. I mean, Until recently, Google's biggest competitor was Cruise, which is a startup that GM acquired about five years ago. And they were putting, I think, similar, but not quite as much financial resources into Cruz. Cruise seemed a little behind, but seemed to be, like, rushing really hard to catch up. And they were not quite as carefully as Waymo. They had a couple of serious mistakes.

There was a woman that ended up getting dragged under a cruise car, and they had a big setback where they had to pause testing for, did it pause their service for about six months. and it's not clear the GM has the resources to do several more years of billions of dollars every year. And so, yeah, I think it's just, it helped a lot that Google, you know, has the reputation to attract top-tier engineers. But yeah, then they just had to, like, put in the work and spend the money to do this, like, very difficult. So in some ways, it's kind of the opposite of an innovator's telebra. The innovative televised story is you start at the low end, you do something quick and dirty and then kind of work your way up. And this was like the opposite of that is, like, you cannot have, like, I think everybody understands, like, if you kill somebody,

it could be catastrophic for your company. And so they have to be really careful. They have to, like, no expense spare to make sure they say technology safe, and it turned out like Google's the only company with the resources and the willingness to see that through. Well, another company with deep resources and visionary leadership is Tesla, which is a potential competitor here. And the Maxim never bet against Elon Musk seems to be a lesson

that the universe continues to press upon us in so many ways. Let's talk about where to be. Tesla is in the self-driving race. This is a company that, on the one hand, is impressive in its own way, in electric vehicles. Elon Musk is a very talented manager, but also has a bit of a PT Barnum-esque bluster when it comes to making promises about his technology. And correct me if I'm wrong, but it seems like he's been promising the imminence of purely self-driving Tesla vehicles for quite a while, even though that has not become reality, but the cars have made progress.

And I don't own a Tesla, but I see videos of people who do own Teslas taking their hands off the wheel and filming themselves while their Tesla seems to relatively capably drive itself around whatever city that they're in. So situate us in the state of play of Tesla. Where does that company stand

in comparison to AMO when it comes to the quality of their self-driving technology? So I think the main thing to understand about Tesla is they have a different business model. So in those business models, they build a taxi service, which means they have to go city by city, and they can spend a lot of money per vehicle because, you know, taxis are a pretty high margin business, especially if you don't have to pay a driver. Whereas Tesla's business model is they are shipping the hardware in every vehicle,

and then they try to upsell you on the software. And so one consequence of this is they can't use expensive sensors like LIDAR. And so I think that's been a disadvantage. Elon has said, though, he thinks LiDar is a crutch, and actually this is going to be an advantage for us. I don't believe that. But it's certainly true that LiDars are expensive. And so if you are, especially if you are pre-shipping the LiDar in every car,

even for customers, they're not paying for it. You know, these things cost thousands of dollars. So that wasn't viable. And so Tesla basically had to try to do this without the expensive sensors. And then because they are selling the vehicle rather than doing it as a taxi service, this, like, geography by geography approach wouldn't work because nobody's going to buy a self-driving car that only works in Phoenix, right? I mean, maybe some people in Phoenix, but even people in Phoenix are going to want to go to other cities or whatever.

And so they have to make it work everywhere before it's really a viable product. And so their approach instead has been to initially sell a driver assistance service where it doesn't drive everywhere, but it can drive on freeways, and maybe now it can drive on city roads, but you still have to pay attention to it. And they've made impressive progress, I think. I mean, certainly FSD, the full self-driving software is much better than it was one, two or three years ago. But I just think they are several years behind where Waymo is. We were talking earlier, you have to go thousands of miles between mistakes to kind of equal human drivers.

There are lots of videos of a Tesla going 10 or 20 or 30 miles without making a mistake. And that's better than it would have been a couple years ago. But Waymo's vehicles were going hundreds of miles between mistakes in like 2017. And so I think they're just, they're making progress. I think eventually they will have a driverless service that maybe is competitive with Waymo. But I think it's like three to five years if they kind of continue with their current trajectory and follow the trajectory we see with Waymo,

I think they're a few years away from being at the point where Waymo is in terms of the technical abilities of their software. It's really interesting to disentangle the Waymo versus Tesla comparison as being two different levels of comparison. One level is the technology, a reliance on LIDAR technology

versus reliance on other technologies. And the other comparison, as you said, is at the level of business model, a robotaxy service for Waymo versus an add-on for personally owned cars. It strikes me that if you knew nothing about this space except the business model comparison, you'd say, well, of course Tesla's going to win.

The vast majority of car ownership or car use in this country is individuals driving alone in cars that they own, right? Compared to the car ownership market, taxi market is absolutely tiny. So Waymo must know this. That piece of information is not exactly a tightly kept secret. We've talked about the use case of self-driving cars as driverless taxi rides, but private ownership is bigger. How close are folks to being able to buy Waymo cars that they can park in their garage or do whatever else they want with? So I think this goes back to what I said a minute ago, which is that nobody's going to buy a car that only works in Phoenix and San Francisco.

And so what I envision happening is I think Waymo will expand and expand and expand. You'll get to the point where they have service in the top 20 metro areas or something. And then they'll go to OAMs and say, we'll sell you the Waymo package where your customer can pay an extra $10,000 or something. Or maybe they'll be a monthly fee. And then you'll be able to turn on self-driving abilities in the cities where we have covered. It'll be a coverage map like a cell phone plan, and that coverage map will grow over time. Right now, I think their coverage map is just not big enough. They would make sense for them to do that, and they're very focused on throwing that taxi service.

But in the long run, I think that's absolutely something to do. And it's definitely true that, like, one reason they need to go pretty quickly is they do want to get there before the technology is a commodity, and either Tesla or any other OEM can just license the technology from somebody else. They're built it themselves. And so I definitely think they're going to- And what does OEM stand for? I'm sorry, that's a car manufacturer, the original equipment manufacturers. That's GM and Ford and Toyota. Sorry, that, yeah, a little bit of jargon there. The reason I think that Waymo's approach makes sense is that even if you sell a vehicle to the customer, if you want to be fully driverless, if you want to be able to, like, get out at the restaurant and the car going off and park itself, or you want to rent it out as a taxi, sometimes, anything like that, or even if you just want to fall asleep in the back seat, there are always going to be some situations where the car gets confused, or maybe the car just gets a flat tire, and you're going to need a certain amount of infrastructure to support.

it to rescue the cars if it gets stuck, or even just, like, interact with local officials are not going to be thrilled about having cars driving around unless they can, like, have something they can call if the cars make a problem. And so there's going to be a certain number of, a certain amount of infrastructure you need for driverless cars. That's geography-specific. And it would be really hard to build that infrastructure everywhere all at once. And so I think the incremental rollout where they do one city at a time is, like, almost

necessary. Even if you were, even if you were kind of Tesla's business model, there's a lot of, this is actually one reason I'm fairly pessimistic when the LMA says she's going to have a taxi service. You know, if that driverless taxi gets a flat tire and there's not like a Tesla store nearby, like what are they going to do? And there's things they can do, but as far as they can tell, they haven't started building that kind of infrastructure. And they're going to have to do that if they want, you know, part of their business model to be a taxi service. And, yeah, they don't seem to be doing that yet. Have you thought about the near future of self-driving cars

retracing the 2010s history of Uber and Lyft, where I feel like my experience of Uber and Lyft in the 2010s followed several stages. Stage one was, oh, cool, this thing is kind of magic that I can press a button on my phone and someone just picks me up, like I'm a rich person with a limo service. Stage two was asking the question

when I was traveling between cities, does this city have Uber? Oh, does this city have Lyft? Oh, that's cool, great. I'll press the Lyft or I'll press the lifter or the Uber button. Stage three, as you wrote a lot about, was intra-municipal fights between Uber and Lyft and various cities about whether or not they were obeying certain laws or whether Travis Kalanek was just running roughshod over existing regulations. And then stage four was, it's just normal. It's a part of our invisible infrastructure.

It's air, it's water, it's, you know, electrical power lines. It was like Uber being in cities is not interesting at all to anyone. It's just a part of modern life. Do you see Waymo? I'm just picking up as you're talking. Do you see Waymo is potentially following that script of right now we're in stage one. It's a really neat novelty and people are taking videos and posting them on social media. Step two is it's going to roll out city by city and people traveling on the country are going to ask themselves or their friends, is Waymo available here?

and then stage three, we're going to have these kind of showdowns as cities are trying to make sure that Waymo isn't or a Tesla robot taxi isn't just a bunch of autonomous robots rummaging about the city and breaking down with no one to fix them. And there'll be a little bit of a fight between some cities and the companies to fix that problem and a host of other problems that I can't even imagine right now. Is that a timeline that you find plausible? Yeah, some of that will happen. So you're already seeing, in last year, San Francisco had a lot of friction with Waymo and Cruz over things like how they dealt with first responders. It would get kind of confused in, you know, crash scenes or fire scenes. And city governments got really mad about that. They ended up kind of being protected because most of the regulations at the state level and the news administration is more favorable to this than city officials.

But anyway, so that's something that will happen. It has happened. I think it'll be a little bit different because I actually think that you'll see a lot of overlap with the. traditional ride-hailing industry, Uber just signed a deal with Waymo to be the exclusive kind of interface with customers in Atlanta and Austin. So in those cities, if you want a Waymo, the way you'll do it is you'll pull out your Uber app and you'll opt in. I'd like to have Waymo cars when it's available and they'll give you one. And then Uber's actually going to do the cleaning and charging of the vehicles. And so that's an experiment. I don't know if that'll take

off everywhere, but it could actually be like literally the way. like Waymo becomes a supplier to Uber and nothing will change about the customer experience except sometimes the car doesn't have a driver in it. The big evolution I think we'll see is I think that with Uber and Lyft, you had a period where it was really, really cheap because they were burning money to build this market and because interest rates are low. And then we had this regression where prices went up and actually I think the market shrank or at least stopped growing.

I think it's going to be the opposite with Waymo or with Waymo and maybe Cruz and other companies if they get there. Because right now, I think the novelty of it means that they can charge a little bit more. The novelty, and also that I think some people just like to not have a driver. You know, if you're a woman, don't want to be sexually harassed or just a shy person, don't want to make small talk. And so I think they're able to charge a premium right now, a small premium. I think that will continue as long as there's a small share of the market.

Then as they get to be a significant share, it'll kind of equalize. And then I think the cost goes down and down and down because there's no driver. And like fleet management should be more efficient in the long run. And so I expect in the long run, it should cost about half of what Uber or left vehicle costs to day. And so, yeah, I think you'll see a more kind of normal trajectory where it'll become just kind of another alternative Uber and Lyft initially. And then as it gets cheaper, like I think the market for it will grow a lot because people will start to notice, oh, this is like so cheap that maybe it doesn't make sense for me to own a car. And the other thing

I think will happen is that with Uber and Lyft, they have to work hard to economize on the driver's time because, you know, human time is expensive. Drivers don't want to wait around. I think with Uber and Lyft, like with Waymo, it'll be much. it's less important to do that. And so you'll be able to do things like have the car come, you know, five minutes before you need to leave and just have it waiting for you. There'll be, you know, more, that would seem as more socially acceptable

because it doesn't even get visas to anybody. It doesn't cost the company as much. And I think there'll be other conveniences like that where the kind of details of the experience will be better in some ways than Uber and Left and it'll be cheaper. And so I think it'll ultimately be like a much nicer customer experience than Uber and Lyft. But I think it'll take five or 10 years for it to really mature to the point where that's where those benefits become clear.

That point is really interesting, and it perfectly on-ramps to the next question I was going to ask you, which is, when this scales, how is the world going to change? And I have loved the seriousness with which you've described the present, but I would like to just, like, dream a little bit for the next five or ten minutes. So just, you know, you can feel free to sort of put on your speculative hat. I remember conversations I was having with people several years ago about, like, the utopian implications of self-driving cars. You know, one of them was, for example, that the city or the county, the county, the county, the county, owns or cooperates this huge municipal fleet that moves everybody around.

Um, the other, uh, is that, you know, I buy a Waymo and it drives me to work. And, you know, maybe that means I live even further away from where I work because I'm just sitting in a room and that feels kind of nice. I can, I can nap in my room with wheels. I can, you know, watch TV. Okay. So far, so normal, more or less. That's not really in the world of science fiction.

But one thing I wonder is I think about what really changes when there's no driver needed behind the wheel. You know, what if I own a Waymo and it drives itself to work? Okay, now I'm at work. My car is with me at work. But rather than park at work, I can say, you know, car, go just drive around and be a kind of Airbnb with wheels. Just be my Uber while I'm working.

go to go self-drive, pick people up, make me money while I'm in the office. At the end of the day, at 4 p.m., I'm programming you to, like, go pick up groceries before you swing by and pick me up. And, you know, you go to your Waymo loading dock at Whole Foods and somebody puts the food in the car and then it swings back around and gets me at the office. I'm not so much trying to make a specific prediction here as I am trying to designate, like, you know, several points along a spectrum of possibility. One is Waymo as a self-driving taxi, but that's a very simplistic application of this technology. It's not very imaginative. More imaginative is the idea that Waymo or any other self-driving car is like an autonomous mobile agent that I can use as an Uber driver in my family, that I can use as a butler to get me stuff.

What aspect, that's my own sort of dreamy vision of the possible futures that are open. to us if this technology scales, I would love you, I would love to hear what your dreams of this technology are. What's the self-driving future that you find most optimistically compelling? Yeah, so I think what you said is like almost right, but I think it's important not to assume

that you're still going to have one vehicle for everything. Because right now when you buy a vehicle, you want five seats because sometimes your kids are arriving. You want something with a range to do a road trip. But in a world where vehicles drive themselves, they can be much more specialized. So, for example, to go in to get groceries, you won't say your car to go get groceries.

The grocery store will have a fleet or, you know, DoorDash will have a fleet or somebody have a fleet of customized, of like dedicated grocery delivery robots. And so delivery service will just get much faster and cheaper. And so it'll be very normal if you need, you know, if you need a stick of butter, you pull out your app and you order butter

and 15 minutes later, a little robot comes up, maybe a sidewalk robot because you don't need a very big robot for this, that drops you off butter. I think this will have profound implications from retail. I mean, I think some retail stores will still exist, but there will also be a lot of, kind of like dark stores that just are, like, small warehouses that have commonly used items and have little robots that come and get the thing and take it to your house.

So that's one, like, big difference that I think is very likely to happen is you'll have the literally becomes much cheaper and faster. And so then, you know, traditional retail changes quite a lot from that. Just to jump on that, because I'd love you to keep unspooling some of these dreams you have. and admittedly these are in the realm of speculation. But one thing you made me think of is right now, when I want something from CVS, and the nearest CVS is about a seven-minute drive from my house,

well, there's only one way for me to do this. I get into my car and I drive to the CVS. But in a world in which CVS is a distributed entity across 20 different vans, that self-driving vans that exist in the Chapel Hill area, then when I want something from CVS, I don't go to CVS. The CVS van that has the stuff that I want,

you know, I need toothpaste, and I need like some paper towels, and I want some floss. It drives to me, and I walk out of my house, and I go up to the CVS van, and I pick up the stuff that I want from the CVS van. And so retail is totally transformed

from you drive to the mall to the mall drives to you. Why would you ever drive to the mall? That doesn't make sense. And that's a funny inversion. That's totally right, is unlocked by self-referral self-driving cars, again, not just being cars minus driver, but rather best being seen as

like autonomous retail agents in a way. Yeah, and so there's two models. One is the whole store kind of comes to you. The other, the store becomes a warehouse where the robot goes and gets an individual item and brings it to you. There's a startup called Robo Mart that is doing the model you talked about. I believe they're about to start testing ice cream deliveries in Baltimore, because ice cream is an example of a product that is really hard to do with a traditional like DoorDash model because it would like melt by the time they get there. But if you have a van with a freezing, in it, the van can like come to you and drop off and then you can like open the freezer up and take the ice cream you want and they've got a, you know, like a camera that detects what you

take. Anyway, so that's a model. I think there's a lot of uncertainty about, yeah, we'll come to you or will just a robot with an item come to you? Also, will it be sidewalk robots or road robots or flag robots? So a lot of questions about that. But yeah, I would expect big changes to retail over the next, you know, 20 years or so. One of the most interesting points that you made in this great interview that you did with Ben Thompson and Shetakery is the implicatory of scaled self-driving vehicles for public safety and penalties for bad driving. Today, if I, you know, get into a fender bender or I get a DUI or something, then there's a penalty, but my license isn't taken away forever. I mean, that would be quite dramatic,

considering that I have to drive myself in order to live. But in a world where people don't actually have to drive themselves in order to live. The penalties actually could be incredibly dramatic. Just unpacked that case, because I thought that was such an interesting implication. Yeah, absolutely. I mean, and it wouldn't actually do that dramatic. You could still own a car. It would just be there's some setting that's changed in your car, you know, probably remotely

via the, you know, via way more, whoever, that says we're just going to turn off the human drive the car feature and you have to use the self-driving to detectology all the time. And so, yeah, I think people talk a lot about, like, once self-driving cars are safe, are we going to make it illegal for people to drive? Yeah, I think it's not so much. I think eventually maybe we'll have that debate, but that's very far away. What I think will happen in the short term is, yeah, we'll become much less, much less tolerant of people that engage in risky behaviors.

And, you know, if you get a DUI, one of the penalties will be you just lose your driving period privileges, maybe for a year or 10 years or maybe forever. And I think people will be much more willing to impose those penalties as the self-driving becomes a more affordable and accessible option. And because everything this world eventually has to pass through the cultural filter, I also suspect that there's going to be some kind of, well, I'll make this in a question rather than a prediction. Do you think it's plausible that there could be some kind of cultural or even political filter to driving versus self-driving? An idea that, for example, if WAMO rolls out in cities where progressives tend to be overpopulated and it rolls out slower, in rural areas where Republicans tend to live, could there be a kind of cultural divide whereby self-driving is woke, so to speak, and driving your car is human. It's humanist. It's old-fashioned. It's traditional. It's what America's all about. The person with their corporeal hands on the

wheel on the open road, that's real living, not this, you know, some alchrum of living that the, you know, crazy progressives in cities. One, is it too fanciful to think that as everything in the world seems to eventually pass through this political cultural filter that self-driving cars could also? Yeah, absolutely. I think that's very likely to happen. I mean, I think that the technology will be compelling enough that the pro-technology side is going to win that, and eventually, like, everybody will be using self-driving vehicles. But this is one reason I don't think we're going to have a ban on anybody driving. And part of the reason is just that it takes 20 years for a vehicle fleet to roll over. So,

even if every new vehicle is self-driving tomorrow is going to take 20 years before everybody has a self-driving vehicle, so it'll take a long time before you can even have that debate. But then, yeah, I think there'll be a period of 10 or 20 years where some people think, oh, we should restrict human driving a lot more and other people think it's not. And, you know, it's hard to predict, like, how these things play out, but it would not be surprised to me if the kind of more Republican rural viewpoint is the, you know, we should protect our, especially protect our cars, whereas the urban liberal viewpoint is we should restrict it more, especially because I do think that self-dram taxi services are going to roll out first where they are most lucrative, which is going to be in cities, the same places where Uberrude Lyft provide service.

Eventually, I think they'll map the whole country and you'll be able to get service anywhere, but it might be five or ten years later that the most rural places. And in the same way, the cell phone networks, you know, they service the big cities first and then kind of gradually go out to the rural areas. And so there might be a period where, like, every big city has a driverless car service, but a lot of rural areas don't have it yet. and obviously then that's going to make city dwellers more comfortable to technology than people in rural areas. One last question, self-driving cars on the labor implications. There's been several analyses of BLS labor data, suggesting that the occupation of the occupation of driving across taxis and Uber and limousine services and trucking is one of the

most common occupation categories among American men. in a world where self-driving technology is good enough that we are replacing or edging in on human-driven taxisies and human-driven limousines. And then eventually, Waymo becomes very competitive in interstate trucking. And as they become competitive there, very lucrative business, a lot of other companies are going to try to compete in trucking as well. This is a highly unionized area. So there could be fights here about, you know, what exactly is. allowed, when do you think we're going to actually hit the moment where there are political economy discussions about the implications of self-driving technology on the labor force?

I think those conversations are happening already. I mean, California recently, the governor California recently vetoed bill that would have essentially banned driverless trucks that was supported by the, I think the teams, union in in California. And so I think that, like, truck driver unions are very attuned to this issue and are doing what they can to slow it down. The difficulty I think they have is that that one of the most appealing routes for driverless trucks is the, like, Arizona to Texas to Georgia kind of route down the southern United States. Those are all Republican states that are not going to have any sympathy for unions. And so, and, you know, government news and veto this bill.

So far, California is open. But I think that, um, You know, if unions manage to get blue states to restrict this technology, then we're going to have it develop quickly in the south and the west. And it'll become very mature. And that'll benefit the state's economies and probably have some safety benefits. And maybe it'll take five or ten years for the other states to come around. But I think it'll be hard for them really to stop it. One thing I think is worth thinking about here is people always, I think, overestimate how quickly these kind of transitions will. happen. And often what happens is technologies don't just replace the job, but they become

compliments. And so one thing I think you might see happen here is that the long-haul trucking, which is the kind of most tedious and most dangerous, and also the places where you'd have the most benefit from self-driving because you can drive 24 hours, whereas human drivers have to sleep, I think those will probably be automated first. Whereas there's a kind of last mile of trucking where once you get to a metro area, the driving is more difficult, but then also once you get to the distribution center, often there's paperwork you need to do or you want to help unload, or you have kind of smaller trucks that go around and, like, go from distribution center to retail stores. And often the truck driver, like, also puts the, the richest at the shelf

or brings it into the store. And so I think you might have a shift where truck driving becomes increasingly something that's like a local thing, which means it becomes somewhat of the more attractive job because you can, you know, sleep with your family every night. Long-haul trucking becomes something that's mostly automated. And there'll be a transitional period of 10 or 20 years, I think, where that's the case. And then the long run, we'll probably will reorganize stores. So maybe there's these ghost warehouses

where there aren't actually retail stores. And so we don't need to have people stocking their shelves. But yeah, I think it'll be a pretty gradual thing where certainly I wouldn't recommend a 25-year-old, like going to trucking as a career right now. But if you're a 50-year-old trucker, you probably have 10 or 15 more years of employment before you'll really see big downward pressure on wages.

So I think it'll be more gradual that some people are worried about. implementation always takes longer than people think it always takes longer than people think i think i think think you wrote the canonical article about this when it came to atms right for vox you wrote the article about how since atMs were introduced by banks what was it 1960s when you just tell the story about atms and the and banker employment yeah so the so the big growth of atoms happened in the 80s and 90s and there's an economist maiden jane basson who um who looked at the data on this and uh employment of tellers actually grew a little bit in the during that period from like 19th

1990 to 2010, where you would expect the biggest impact. And a big reason for that is that as the ATM made branches more efficient, it became cheaper to have a branch. And so banks opened a bunch more branches, each of which had fewer tellers, but the total number of tellers went up a little bit. It's a little hard to tell if this is entirely causal because there's also a lot of deregulation.

Like in 1980, it was kind of illegal to have branches. So it's a little hard to disentangle. But certainly is not the case that, like, we invented the ATM and then all the tellers got put out of business. I looked at this recently, the number of tellers has actually started to decline a little, but I think because a lot of people don't go to banks at all anymore. It's like you just pay with like apps and stuff.

So eventually I think these jobs do get replaced. But if you had been in 1990 and be like, oh, these ATMs are going to put me out a job, you definitely could finish a 20 or 30 year career as a bank teller before the really big employment effects will hit. And I think that's more through than people expect for professions like tracking. I think will definitely be eliminated in the long run, but the long run might be 10 or 20 years. And I think the lesson to pour from banking into driving is that new technologies that often seem like substitutes rarely act like a one-to-one substitute.

Instead, they dynamically change the industry in which they're introduced, and those dynamic changes often create the possibility of new services that can be rendered, and new services tend to not be able to be done by technology. They have to be done by people, and so technology often, not always, but often has the effect of raising employment rather than reducing it. Very last question for you, because this is a huge topic of debate right now in the AI space, is whether or not we're seeing really clear evidence that AI progress has hit a wall. For people who have been not following this space as closely, when ChatGBT came out in November 2022, that was technically Model 3.5 of OpenAI's GPT technology.

And in March of 2023, Chad GPT came out with GPT4. And a lot of people looked at this and said, oh, my God, this technology is moving so quickly. We went from chat GBT to GPT4, which is clearly better along a bunch of different margins. In five months, this is going to take over the world in just a matter of years. And the truth is, it has been an extraordinary explosion in productivity and efficiency. But we're seeing some evidence that generative AI performance is hitting a kind of ceiling, a kind of asymptote, even as we're throwing more and more compute power at the problem. How would you describe this sort of scaling debate that's happening right now in AI?

And do you come down firmly on one side of the other of whether or not we've hit a wall in AI progress? Yeah, so the classic story here is that there's three ingredients you need for an AI system. You need a good algorithm. You need a lot of data, and you need more computing power. And the idea is you scale all those things up. You make the model bigger. You have more data and you have more computing power. The models get better and better and better. And that is what you saw from GPD1, which in 2018 through GPP3, which is 2020, and then GTP4, 2023.

You saw a pretty steady improvement. And there hasn't been a GPD5. We're coming up on two years. Sam O'Long recently said there's not going to be GPD5 this year. So it'll be at least, you know, a two-year delay. And I don't think we know for sure. The people inside the labs, both publicly and from what I've heard privately, still believe in scaling laws. They think they're going to make bigger models and they're going to be amazing.

I'm pretty skeptical of that. And I think what I think is happening is that I think it is absolutely true that if you have more data and more compute, you get better performance. But I think the reason for that is that to a large extent that more data gives you more diversity of topics. So the way that they scaled up is they pulled down data from all across. the internet. There's a lot of different data on a lot of different topics created by a lot of different people for a lot of different purposes. And so any like topic you think to ask the model about, there's a good chance there was something in his training data that was very similar what you asked. And it does some pretty simple pattern matching with where it, you know, it says,

in my training data, I saw a question that was kind of like this. And so therefore I know kind of not quite Mab-Lib style, but it learns pretty simple patterns. And so the more data you have, the more of those patterns it has. And so the smarter it seems to be. And I think if you had 100 times more data from 100 times more topics, it would be 100, it would be more capable. The problem is that that trick of like pulling down the whole internet is something you can only do once because there's nothing 100 times bigger than the internet that you can train for.

And people talk about all we can use, you know, YouTube comments or people's chat logs or like, whatever. But like those are not, I think people, I think the quality of the data matters, the diversity of the data matters. And I don't think YouTube comments are like a high quality source of data. And so, and I think you're seeing. of this with the labs, they're spending a lot of money hiring scientists to basically generate new training data on their technical topics and getting higher and higher level of performance

on these benchmarks of like specific, like scientific and mathematical things. And that's cool. That's progress. But the thing I worry about is like most of the, a lot of the economically valuable things we want is like you put it in a new company and it has some kind of new cutting edge thing where the details about it or not on the internet because it's proprietary or it's like brand new research. And these models don't necessarily seem good at learning new things. They're good at pattern matching from things that are already in the training data. But the really valuable thing is kind of an inference time learning where you give it a new

thing in the current context and say, what should I do next or do this thing? And they really seem to get confused if you give them kind of complicated problems as opposed to kind of canned kind of math school problems. And so, yeah, I think my guess is they're, they're hitting a wall at the work at ultimately indeed different architectures to make a significant jump, but I'm glad people are doing the experiment. I might be totally wrong. Craigman, if I'm wrong, but it seems like you're saying that the bottleneck here is the data, that we essentially have one planet's internet. And it'd be great if we had 10 planets

internets that we could feed into these models, but we don't. There are not 10 planets that have advanced intelligence. And so there's only one planet's internet to fold into the system, and you can only fold that into the system once. With the current algorithm. With the current algorithm, right. It also seems to me like a separate debate that's being had is whether or not there's something about the transformer approach to AI that might have hit its ceiling.

Are these basically two ways of saying the same thing? That essentially the transformer approach was very good at taking the internet of planet Earth and turning it into GPT4-level artificial intelligence. But in order to create GPT-6-level, GPT-7-level artificial intelligence, we've already imbibed the internet. We need to do new stuff to that internet algorithmically. And that means we need a new kind of architecture to get us to beyond PhD-level expertise in all these different subjects. Is that kind of what you're saying?

Yeah. What I'm saying is that the transformer uses data very inefficiently. So you have to show it 100,000 examples of something before it understands it. And it only understands it in a fairly superficial way. It's like, you know, there's some students in college. who, like, when the professor lectures, they write down exactly what the professor says and learn these, like, very simple formula.

It's really this kind of problem. We plug this variable into this. And you've got other students who kind of sit at the back of the room and think very conceptually, and they, like, understand what the professor's saying. And then they might both do the same on, like, the test in that class. But then when they go to the next class, the student who really got the conceptual understanding

is going to have a much easier time that the student who's trying to, like, do it all by grinding and look at a million examples. What we need is that second kind of student. We need somebody who really generalizes. I mean, generalization is not a binary. It's not either, but generalize there or doesn't. There's different degrees of generalization.

It's a ladder. And I think we're, you know, the transformer was a step above the, on that ladder up from previous machine learning techniques. But I think we need a new step where you can show a 10 examples instead of 10,000, and it understands it. And it generalizes not to just to those examples, but to related examples. And I think my view is like human beings are still much better at that than Transformers. And I don't think making the Transformers is bigger as going to,

significantly changed that. So interesting. The substack is understanding AI, and it's really fantastic. Tim Lee, thank you so much. Thank you, Derek. Thank you for listening. Today's episode was produced by Devin Beraldi.

Our schedule for plain English for the next few weeks will be one episode a week on Fridays. We'll see you next week.