Are Human Drivers Finally Obsolete?

23:53A legendary roboticist named Sebastian Thrun. Sebastian Thrun, he was at the first Grand Challenge. He didn't bring a team. He wasn't participating. DARPA wanted to show off some other projects they'd been funding, including one of his robots. So he brings the robot, and so he's there. And he watches this disaster, and he thinks, I can do better than this. I looked at the very first iteration of this grand challenge where I didn't participate. I was a spectator. This, of course, is Sebastian Thrun. He grew up in West Germany, moved to the U.S., taught at Carnegie Mellon before moving to Stanford.

24:29Watching that day, he saw this fundamental error he believed all the entrance had made. I saw that all the teams treated this like a hardware problem. They looked at this and said, we have to build bigger wheels and bigger chassis and so on. And I looked at this and said, well, wait a minute. The challenge really is to build a self-driving car that can drive to the desert. I can get a rental car that can do it just fine, provided there's a person inside. And the challenge is really to take the person out of the driver's seat and replace it by a computer. That is not a problem of bigger tires. That's actually really a software problem.

25:09Sebastian Thrun had a dual background, robotics and artificial intelligence, which probably explains his focus here on the robot driver's mind. He was thinking about something else, too. The military wanted this tech to replace a relatively small number of drivers in its war zones. But Sebastian was already imagining something bigger. What would happen to traffic deaths worldwide if one day everyone had access to a driverless car? I had experiences of losing people in my life to traffic accidents, and I felt we lost over the million people in the world to traffic accidents. Wouldn't it be amazing if DARPA invented something that would save a million lives a year?

25:47In October of 2005, 43 teams have brought their vehicles to compete in a unique event. A race driven not by testosterone, but computer code. Chapter 3. Machine Learning.

26:08The race course is a circular maze that zigzags for 132 miles. 18 months later, for the second Grand Challenge, DARPA doubled the bounty,$2 million. This footage is from a PBS documentary called The Great Robot Race, narrated, to my mild joy, by John Lithgow. Familiar faces have returned. Chris Urmson, back with the Carnegie Mellon team, this time with two vehicles, Highlander and Sandstorm. Anthony Lewandowski back with his motorcycle, which still doesn't work. He's knocked out in the qualifiers. And now there's also Stanford's entrant. Compared to Sandstorm, the bulked-up Hummer, the car looks measly.

26:45A blue SUV donated by Volkswagen. A baby-faced Thrun smiles next to his soccer mom-looking vehicle. The vehicle's name is Stanley. So Stanley is nothing else but Stanford. But it also gives the vehicle a personality. We think of the vehicle more and more as an intelligent decision maker. And Thrun really brought more artificial intelligence, which at the time, we're talking 2005, was still rather primitive, especially compared to what we have today. But he could use it to teach his vehicle how to recognize the road and how to do it much faster. faster. They found a dirt road out near Stanford, and they drive it down a dirt road and have the car's cameras record what they were seeing.

27:34The robot Stanley was able to train itself as it went. And the way it worked is, its eyes looked way ahead, and it could see stuff way at distance. When it drives over the stuff, you could tell if it was a good place to drive or not, because it could measure how slippery or how bumpy the road was. And then you could then retroactively train and say, this green stuff over there, it's something good to drive on, a.k.a. grass. And this brownish stuff, a.k.a. mud, is not so good to drive. And so it was able to detect patterns and generalize from what it had learned? Yeah, absolutely. And it did this like 30 times a second.

28:12I mean, just like a person. The race kicks off with Stanley sandwiched between Carnegie Mellon's two behemoths.

28:19Timothy B. Lee:Highlander leads the pack, followed by Stanley and Sandstorm. What happens in the second race? The second race is as successful as the first race is disastrous. Nearly every entrant in the second race would go further than Sandstorm had in the first. Multiple vehicles would finish the course. The real question was who would do it fastest. And so at what point was it clear to you that you were going to win? Well, once we passed the front-running team, we kind of saw the vehicle descend into what was the hardest part of the race course, a very, very treachery mountain pass. And we saw at a distance a dust cloud.

29:03We saw a helicopter. We saw little features that made us believe, wow, there's something happening that's magical. And this dust cloud then all of a sudden turned bluish because the car was blue and came closer. And then it came first to the finish line. It was unbelievably magical. At the end of the dock, over some criminally corny piano music, Sebastian Thrun gives his post-race interview. He's dressed a lot like a race car driver, watching you could forget he wasn't in the car. It was just amazing to see this community of people. That community succeeded today. Behind me, there are three robots that made it all the way through the desert.

29:38And all three of them did the unthinkable. It's such a fantastic success for this community. I think we all win. A made-for-TV kumbaya moment, still years before the race to build driverless cars, would enter its cutthroat phase.

29:58What would happen next is that a small band of lunatics would take driverless cars out of the desert, start secretly driving them on public roads in the state of California. They would do this at the behest of a man who had been observing from the stands that day, disguised in a hat and sunglasses who watched the challenge while his mind spun.

30:22That's after a short break.

30:29I'm Stephen Dubner, and you are listening to a special episode of the podcast Search Engine here on Freakonomics Radio. We will be right back.

30:44Tyler Reddick and Chumba Casino, a winning combination. Another checkered flag for Tyler Reddick in 2311 Racing, this time at Coda in the Chumba Casino Toyota Camry. That's three wins this season. Talk about a lucky streak. Chumba Casino, proud partner of Tyler Reddick and 2311 Racing. Play for free at ChumbaCasino.com. Let's Chumba! No purchase necessary. VGW Group Voidware prohibited by law. CT &C's 21 Plus, sponsored by Chumba Casino. Hi, I'm Willa Paskin, the host of Decoder Ring, Slate's podcast about cracking cultural mysteries. On Decoder Ring, we dive down rabbit holes and obsessively explore questions hiding in plain sight.

31:24Like, why is protein in everything these days? And are British teeth really as bad as Americans like to think they are? And why, in the late 1990s, did people suddenly become re-obsessed with swing music? All that and more on Decoder Ring. Listen to new episodes every two weeks where you listen to podcasts.

31:46Hey there, it's Stephen Dubner. Today we are running an episode from the Search Engine podcast with host PJ Vogt. Chapter four, something actually useful for the world. The race in the desert had been designed as a spectacle, something flashy to draw out America's smartest roboticists. But it had drawn another person who'd come for his own reasons. Google's Larry Page arrived at the DARPA Grand Challenge in a baseball hat and sunglasses, a disguise He found Sebastian Thrun and buttonholed him asking him a million highly specific questions about things like the wavelength his LiDAR system used But this meeting in the desert this was not actually their first introduction Well the first time I met Larry was a bit earlier he had built a small little robot that acted as a telepresence for meetings and he was trying to drive it around the Google offices instead of himself going to a meeting with a robot.

32:41And he sent me a message and said, I'm going to show you the robot I've built. And I, in a spirit of like craziness, I sent a message back saying, Larry, I'm so glad that Google lets you use 20 % of the time to do something useful for the world. I couldn't. I either expected a rapid response or never hear from him again. It turns out I was lucky. He responded immediately. I took his robot, I fixed it the next 24 hours, and he was very happy. Larry Page, it turned out, had actually been interested in autonomous vehicles since at least grad school. That's what he'd wanted to do his thesis on before being guided by some wise PhD advisor toward search engines instead.

33:27Now, as a spectator at DARPA's second grand challenge, he could see real-world evidence that autonomous vehicles might actually be a thing. At first, Larry Page hires Sebastian Thrun, along with fellow DARPA contestant Anthony Lewandowski, just to build what will become Google Street View. They'll actually modify the system that Stanley the Car's roof-mounted cameras had used to begin photographing American streets. But before long, Larry Page returns to Sebastian with his dream of a driverless car.

33:59Timothy B. Lee:And so how soon after arriving at Google does Project Chauffeur begin? Like Larry Page says to you, I have a mission. Like, how does this happen? This is an embarrassing moment for me. It's about two years later, 2009, where I sit in my cubicle and Larry Page comes by and says, Sebastian, I think you should build a self-driving car that can drive anywhere in the world. And my immediate reaction was, no, taking the technology we built for this empty desert and put it in the middle of Market Street in San Francisco is going to kill somebody. and Larry would come back the next day with the same idea and I would give him the same answer and both of us got increasingly more frustrated like, God damn it, it can't be done and eventually he came and said, look Sebastian, okay, I get it you can't do it I want to explain to Eric Schmidt, the CEO at the time and Sergey Brin, my co-founder, why it can't be done can you give me the technical reason why it can't be done and that's the moment of incredible pain because I go home and I can't think of a technical reason why not It was this kind of moment where I felt, look, I'm the world expert on self-driving cars, and I'm the person who denies that it can be done.

35:11That taught me an incredibly important lesson about experts. For the rest of my life, I decided experts are usually experts of the past, not the future. And if you ask an expert about innovation, something crazy new, they're the least likely person to say, yes, it can be done.

35:29Timothy B. Lee:So this is where the Google self-driving car project begins in 2009. It's led by Sebastian, joined by others from the DARPA challenges. The methodical Chris Urmson was running most things day to day. Anthony Lewandowski, the flashy motorcycle guy, would work on hardware. Dmitry Dolgov, another DARPA veteran, would be responsible for planning and optimization. It was a secret project. They'd report directly to Larry Page, a small enough team that there'd be no bureaucracy, few emails, fewer meetings. Just 11 engineers who writer Alex Davies says represented some of the best young talent in the country.

36:03And so Google builds this very quiet team and it says to them, build us a self-driving car. And because that goal is super nebulous, they give them two challenges. They say, safely log 100 ,000 miles on public roads, but they also give them a challenge called the Larry 1K. So Larry and Sergey and I sat together and the two of them carved out a thousand total miles of road surface in California. They open up Google Maps and they just click around and they look for 10 separate 100-mile routes that are really tricky. Absolutely everything, like the Bay Bridge and Lake Ta 'o and Highway 1 to Los Angeles.

36:52and Market Street and even Crooked Lombard Street. And they say to the team, you have to drive each of these 100-mile routes without one human takeover of the system, without one failure of the car. To get off to a running start, the team licenses the code from Stanford's DARPA Urban Challenge vehicle. Anthony Lewandowski goes to a local Toyota dealership and buys eight Priuses, takes them back to Google, and retrofits them to accept a computer as a driver. He hooks that computer driver electronically into the brakes, the gas, the steering. These Priuses get a radar system behind the bumper. Cameras, a LiDAR system spinning 360 degrees on top.

37:33LiDAR, like radar, but it shoots lasers instead of radio waves. At first, the team gives each Prius a cool name, like Knight Rider. But I think we quickly realized that we're not going to be able to name all these vehicles as we scale up our fleet, And so we just started to number them like, you know, Prius 27.

37:51Timothy B. Lee:This is Don Burnett. He'd been a researcher working on autonomous submarines. He lost a friend in a car accident, separately got in a bad accident himself, and decided he wanted to do work on self-driving cars. That's how he eventually ended up on the team in its early days. I was on the motion planning and behavior decision-making team, and my responsibility was to work on the nudging behavior. Nudging, when a big truck passes a human driver on the right, the driver will nudge a little to the left. For us, it's an instinct. Don's job was to teach a computer to nudge. You're trying to encode the behavior that you would use as a driver under kind of partially good perception.

38:31And it's a really tricky problem.

38:33Timothy B. Lee:A team of academic roboticists, some of whom had had friends die in cars, spending Google's money to see if they could make driving safer. It was a weird era.

38:47There's this big concert venue near Google's offices called the Shoreline Amphitheater. In 2009, you could have seen Sheryl Crow there, the killers, Fish.

38:57Timothy B. Lee:But the most interesting show that year was one almost nobody knew about. In the venue parking lot, on days when there was no concert, no tour buses around to see them, the Google team would run its first test runs of their driverless cars, essentially hiding in plain sight. A Prius driving itself around the amphitheater parking lot with an attentive safety driver sitting behind the wheel, just in case. The team was making sure the basics functioned, that the sensors could really recognize another car, that the computer in the car was abiding by their orders. These were the baby steps. They'd happen in this parking lot and at an empty airplane runway that was close to their offices.

39:36Spring 2009, the team tries actual real road driving for the first time. Chris Ermsson takes one of the Priuses out on the central expressway. Speed limit, 45 miles per hour. There are humans driving here. And immediately, outside the confines of the empty parking lot and empty airplane runway, here's what's clear. They had a real problem. The car was swerving wildly. It was weaving around like a drunken sailor. And we realized that the scale of the runway was such that you didn't notice the one or two foot kind of oscillation it had in lateral control. And you put it on Central Expressway and suddenly, you know, yep, turns out actually that's a problem.

40:21One more problem to fix.

40:27listening to this story it's funny because i can imagine it giving me a totally different feeling than it does a tech company with nobody's permission was testing driverless cars on public roads in california i don't know why that strikes me as being about invention instead of just hubris and impunity maybe it's because i know that google would be one of the few tech companies whose driverless cars would not cause any fatal accidents in testing. And that the team would just take more safety precautions than the other companies who'd rush in later to catch up with them once this was an arms race.

41:03The way these cars were designed, the safety driver sat behind the steering wheel, ready to take over. In the other seat was their partner, watching the monitor displaying a graphical interface designed by Dmitry Dolgov. The people watching the screen would call out problems ahead, some discrepancy between what the sensors were seeing and what was actually in the road. This is what teaching a car to drive actually looked like. Two-person team is manning the cars, logging errors, going back to the office to troubleshoot, and then updating the code.

41:33Timothy B. Lee:I asked Don Burnett about this era. And while you're doing this, and then you leave work and you get in your car that you drive as a human, did you find yourself thinking more carefully, like, how do I know what I know when I'm driving? Like, you're trying to teach a machine by day. Did it affect how you thought about human driving by night? Almost obnoxiously so to any passengers in the car with me. I was obsessed with one big question, which is, why do humans drive the way they drive? And it turns out there were no good answers. And I still think they're not great answers. And instead of actually answering that question, we've just turned to machine learning to infer the deep truths behind why humans do what they do.

42:16And so there's some basic principles that you can understand. Like we try to minimize lateral acceleration, meaning you don't want to be thrown to the outside of your car when you're making a turn. So you're going to slow down. But how much do you slow down, right? And it turns out that's contextual. Don gave me an example. So you're trying to figure out the right speed and angle for the car on one of those tight, curvy on-ramps onto the highway. You want it to feel comfortable for a passenger. Don says you can work out the math. the lateral acceleration is two meters per second squared. But the surprising thing is, that number only applies on the on-ramp.

42:54If I put you at a cul-de-sac in a neighborhood and you were going to do a U-turn at the end of the cul-de-sac, even though the speed is significantly slower, if you did two meters per second squared of lateral acceleration around a cul-de-sac, you would tell your driver they were crazy. It would feel incredibly uncomfortable. Like, incredibly uncomfortable. You would feel like you were in Mario Kart. Yes, it would feel Mario Kart. And remember, this is a force. So the physical feeling on your body is exactly the same. But the contextual awareness of the situation, of speeding up to get on the highway versus making a U-turn in a residential street, tricks your brain into feeling opposite about the situation.

43:42And so it turns out the limit for a cul-de-sac is around 0.75. It's almost three times less than you would be willing to tolerate as you accelerate onto a highway. And so there were things like that where you couldn't just say humans have specific physical restrictions, right, from a forces perspective. The context matters. And when the context matters, now all of a sudden anything is game. So things like that is where I spent my time as a researcher trying to figure out, okay, how are we going to make this comfortable for passengers? All these little problems to solve. But there was one gift, which was that the team at this point had an overarching goal uniting them.

44:25The DARPA challenge had told them, drive across this patch of desert. The Larry 1K challenge told them, drive these 10 routes without human intervention.

44:34Timothy B. Lee:The specificity of the mission meant they never had to squabble about why they were there. By 2010, just a year in, the team was really on a roll. They start knocking out routes. Each one of the routes was unique and distinct and different and had its own challenges. Down Route 1, Silicon Valley to Carmel. The bridges run. We had to go across all of the bridges in the Bay Area, starting in Mountain View, finishing crossing the Golden Gate Bridge. It's Chris Hermsen in the car. It's Anthony Lewandowski in the car. I was in the car with Dimitri, Chris, and Anthony. It was the four of us in the Prius.

45:12They were figuring out the technology much faster than they thought they could. The Larry 1K was set up like a video game, meaning they'd get to try the route over and over until they could complete it without a single human takeover. Then they'd move on to the next one. It was really a proof-of-concept exercise. Can you even make this happen once? When they fail a route, they know what the car can't handle, So they go back and say, you have to be better at doing X, Y, Z. And then we got back to the office. We regrouped. We went back out, I think, at like 11 p.m. And by 1 a.m., we had completed the route.

45:51Timothy B. Lee:They buy a bottle of Corbell champagne. They all write their names on it. Corbell,$13.99 a bottle. The champagne they have at Trader Joe's. They had one for every route they completed. And one by one, they pick off the Larry 1K routes. And they think this is going to take them about two years when they start out. And they do it in a little bit more than a year. Nearly twice as fast as they had expected. By fall of 2010, they're done. Here's Chris Urmson. And I think we had a big party up at Sebastian's house in Los Altos Hills. So, you know, it was pretty spectacular, right? They throw each other in the pool.

46:29They celebrate. And then they're not entirely sure what to do next. It was kind of, okay, and now what? The team had pulled off a kind of miracle in a year. A driverless car with human supervision, with lots of human coding, but still a driverless car successfully navigating some very tricky roads in California. They'd done this safely. They'd done it quickly. And now things would begin to wobble. Competition would arrive. The team itself would begin to schism. and one member, a person who believed the team was moving too slowly, would actually take matters into his own hands in a particularly extreme way.

47:12After the break, mutiny. Hey there, Stephen Dubner. That again is PJ Vogt, and you are listening to an episode of the Search Engine podcast here on Freakonomics Radio. We will be right back.

47:34You know what's worse than a long wait? Flipping through old magazines in a waiting room. Instead, I power up Chumba Casino. Slots, blackjack, solitaire, roulette, bingo, and more. Boredom doesn't stand a chance. Take a few virtual laps with me. Let's Chumba. Play now at ChumbaCasino.com. No purchase necessary. VGW Group Voidware prohibited by law. CT &Cs 21 Plus, sponsored by Chumba Casino. Hi, I'm Willa Paskin, the host of Decoder Ring, Slate's podcast about cracking cultural mysteries. On Decoder Ring, we dive down rabbit holes and obsessively explore questions hiding in plain sight. Like, why is protein in everything these days?

48:15And are British teeth really as bad as Americans like to think they are? And why, in the late 1990s, did people suddenly become re-obsessed with swing music? All that and more on Decoder Ring. Listen to new episodes every two weeks where you listen to podcasts.

48:33Timothy B. Lee:Welcome back to the show.

48:38Timothy B. Lee:As early as 2010, Google's driverless car project had developed some very impressive self-driving technology. But what they were struggling to decide was this. What was the actual product they were developing here? Here's Sebastian Thrun. We had a lot of debates inside Google what the right business model was. At some point, we actually had a big debate. We should just buy Tesla. and Tesla was worth$2 billion at the time. I remember this. Maybe we should have in hindsight. But joking aside here, there was a debate whether this is more of an assistive technology or a disruptive replacement technology.

49:18Timothy B. Lee:Basically, should they follow the route that Tesla ultimately would? Design self-driving as a feature in your car, something that could take over sometimes but still need human monitoring? Or was it better to wait until the car could fully drive itself? Theron would eventually come around to this version of self-driving. Specifically, he'd come around to the idea of self-driving robo-taxis. A taxi service type system is way more capital efficient than ownership. An owned car is being used about 4 % of the time and it's parked 96 % of the time. Imagine a city without parked cars, where every car is being utilized, call it 50 % of the time.

49:54which means we have like only 10 % number of cars needed that we need today when we all own cars. That's going to happen. There's no absolute question.

50:04Timothy B. Lee:What Sebastian is describing here so matter-of-factly is a fairly radical reimagination of American cities. The idea that robo-taxis would be so cheap and widely available that most people just wouldn't own cars, that we could put something else, anything else, in the places where we put most of our parking lots and parking spaces, That is a far-fetched idea, just given how much of American identity is tied into personal car ownership. A far-fetched idea, and for it to begin to happen, Google would have to bring a product to market. But the years passed, and they didn't. And some people who were there felt stuck.

50:43Timothy B. Lee:Don Burnett says he believes life at Google got dangerously cushy. The food was great, the money was too. These former academics making much more than they'd ever expected. There was a lack of urgency on the team to actually make something viable. We had a funding supply that effectively felt infinite. And maybe it was, maybe it wasn't. But it certainly felt infinite. And when you have infinite funding, you're not forced to make hard decisions. You're not forced to focus. You're not forced to look at the opportunity, the market, the customer, and be the best. It was more like, hey, let's take our time.

51:23Let's make sure we do it right. which is on its face a good principle. But at the end of the day, I think the lack of urgency wasn't for everyone. And within the team, you get Team Chris and Team Anthony. And they start butting heads all the time.

51:41Timothy B. Lee:Chris and Anthony, meaning Chris Urmson, official head of the project, versus Anthony Lewandowski, who I still think of as the motorcycle guy. The main difference in their approach is how quickly they want to move. Anthony is very okay with risk, we'll say. He gets one of these cars and he's driving it back and he lives in Berkeley, works in Palo Alto. He's just using this car like on the Bay Bridge every day, probably outside the bounds of what the team actually wanted. And he's not like necessarily logging data. He's just enjoying his self-driving car and taking it all over the place. Chris comes from an academic background.

52:21He's that Canadian, very nice, very careful, very risk-averse. When I asked Chris Ermsson about all this, his memory was slightly different. In his memory, Team Anthony was pretty much just Anthony. And Anthony, he said, was a move fast and break things kind of guy.

52:41Timothy B. Lee:Move fast and break things, a motto famously coined by Mark Zuckerberg. It defines a way of developing technology which once might have felt cute and revolutionary, but which today, at least to me, feels pretty irresponsible. Chris didn't think that philosophy was an option for their team. Even if their cars were statistically safer than human drivers, he knew that the first news story about a self-driving car in a fatal accident was going to be a huge deal. Anecdote was going to demolish data if they weren't extremely careful. By all accounts, Anthony Lewandowski felt differently. But he actually wasn't the only one.

53:19Here's Don Burnett. There were some people on the team, very famously, including myself, that started to get the itch kind of towards the three to four year mark. The itch of like, okay, where is this going? Who is it for? How are they going to use it? Where are they going to use it? And I felt like the leadership didn't have great answers to that. There was no commercial race, right? We had no competition and there was no market for the product. But competition would soon arrive in the form of Uber.

53:52This was the oh shit moment for me. Uber announced their self-driving program. And I remember, like it was yesterday, waking up, reading the news, going to my desk in the morning and thinking, oh crap, these guys are going to eat our lunch.

54:09Timothy B. Lee:In 2013, then CEO of Uber, Travis Kalanick, had gotten a ride in one of Google's prototype driverless cars. Sitting in a taxi without a human driver, he'd understood that this could mean the end of his company. And so Uber had plunged headlong into the driverless car race. The company hired nearly half of Carnegie Mellon's top robotics lab. And not long after, we also know through court records and emails, that Uber also began communicating with Anthony Lewandowski, who in 2016 would leave Google, quitting just before he could be fired for recruiting team members away, including Don Burnett. Anthony would then start his own autonomous vehicle company.

54:51Uber would soon buy that company for almost$700 million, even though the company had no product and was only months old, which raised a mystery. Why would Uber pay so much for a company whose only assets seem to be its people? This is where Google goes into its computer security logs and realizes that not long before he left, Anthony Lewandowski downloaded something like 14 ,000 technical files onto his computer and moved them onto an external disk.

55:23Timothy B. Lee:Obviously, you can't do that. I mean, I'm assuming obviously you can't do that. No, you definitely cannot do that.

55:33And this is the kind of thing that maybe if he had stayed there, this is the kind of thing Anthony would have done. And he would have been like, oh, it's just so I could have access to it somewhere else. And he probably would have gotten away with it. But when you then go and work for Uber and start running their direct competitor self-driving car program, that's when you get in trouble. And that's when what's technically called Waymo at this point, Google's program, sues Uber and puts Anthony at the center of an enormous legal battle between these tech giants. Secrets and subterfuge in Silicon Valley, a former Google engineer has been charged with stealing files from Alphabet's self-driving car project and taking them to Uber.

56:21Specifically, it involves a former lead engineer of Google's self-driving car unit, Anthony Lewandowski. Now, he's accused of using his personal laptop and downloading more than 14 ,000... In 2016, Google had just spun its driverless car unit into a new entity, Waymo. Waymo sued Uber. Uber had to settle to the tune of$245 million. And in a separate criminal trial, Anthony Lewandowski pled guilty to stealing trade secrets. Afterwards, Uber continues their driverless car program without him, continuing to pursue its move-fast, break-things strategy, which in 2018 leads to the death of a woman named Elaine Hertzberg.

57:02Uber is hitting the brakes on its self-driving cars after one of them hit and killed a woman in Arizona. The vehicle was in autonomous mode, but it did have a safety driver on board. But a police report later indicating the safety driver was streaming TV shows on her phone for three hours that night, including at the time of the crash.

57:23Timothy B. Lee:The way this story was reported, nearly everyone blamed the safety driver. She was on her phone. She was streaming an episode of The Voice. Tempe investigators saying had Vasquez been paying attention to the road, she could have stopped the car 42 feet before impact. The NTSB slamming Uber... There was some important additional context, which was that Uber's robot driver was also just much worse than Waymo's. A statistic I found jaw-dropping. At this point, Waymo's safety drivers were having to take over from the car once every 5 ,600 miles. Uber's safety drivers that year had to intervene more than once every 13 miles.

58:02Timothy B. Lee:Despite that, five months before the crash, over employee objections, Uber had cut its safety crews. Instead of two humans, they just used one. One safety driver overseeing a robot driver that was arguably not ready to be on public roads. In the last moments of Lane Hertzberg's life, the robot spent an indefensible 5.6 seconds, trying and failing, to guess the shape in the road that was a human body pushing a bike. Over those 5.6 seconds, the robot kept reclassifying her. Was she an unknown object? A vehicle? A bicycle? During that time spent wondering, the car did not slow down. Soon after Elaine Hertzberg's death, Uber halted its testing program.

58:46Uber has temporarily suspended its driverless fleet nationwide as the NTSB, police, Uber, and the National Highway Traffic Safety Administration investigate.

58:56Timothy B. Lee:We reached out to Uber for comment. a spokesperson said that the fatal collision was indeed a tragedy, which had a significant impact on Uber and the entire industry. There would be other competitors who would shut down after similar accidents. There would also be Tesla, which by 2020 was publicly marketing a product the company called full self-driving, but which absolutely was not. Meanwhile, Waymo had slowly continued to develop its tech. Their robo-taxis would be ready for riders by 2020. The team had gotten an unexpected boost from a technology that was, at the time, very little understood.

59:33In 2026, when most people talk about artificial intelligence, the conversation defaults to products like ChatGPT and Claude. But artificial intelligence has been a core part of driverless cars going back two decades. In the 2010s, neural net advances meant that you could now begin to feed a computer system large amounts of data. and watch as its perception, prediction, and decision-making abilities improved. Here's Sebastian Thrun. That technology of massive data training was with us from the get-go, but has become more and more and more and more important. The surprise for all of us has been that size matters.

1:00:11When you put a million documents into an AI, it's fine. A hundred million is fine. But when you put a hundred billion documents into an AI, it is unbelievably smart. and that I think shocked everybody, myself included.

1:00:26Timothy B. Lee:The Google Brain team, the deep learning people, started working with the driverless car team to use training data to help the computer driver learn things like how to better predict when another car was about to suddenly switch lanes, how to more reliably spot pedestrians. Over the years, as the car drove more miles, as the team gathered more data, plugged that data into their AI systems and tweaked those systems, the engineers say the robot driver kept improving. As they tested the car in new weather conditions, they discovered problems that required hardware fixes. For instance, in Phoenix, Waymo had to design miniature wipers for their car's LiDAR sensors to deal with the dust storms and heavy rains.

1:01:02Timothy B. Lee:In 2020, Waymo finally debuts to the public in Arizona. In the years after, it'll roll out to 10 more American cities. A funny consequence of Waymo's long development cycle is that the public's attitude towards Silicon Valley has just really changed in that time. There's more suspicion towards Google than there was back in 2009 when the project first started. And so now, many people look at the Waymo driver with a raised eyebrow, with a question immediately on their lips. Chapter 5. Are you a good driver? All right, autonomous vehicles can now get you around Atlanta. The future of driving through Austin is here, except it comes without a driver.

1:01:43The 911's live-hailing app is now taking passengers in Miami. A fleet of white electric Jaguars covered in 40 different sensors, cameras, radar, LIDAR. It's an expensive car, as much as$150 ,000 by some estimates. In the news stories, you see the inside, where the human driver would normally sit. There's an empty seat. You're not allowed in. With a steering wheel in front of it, vestigial, it turns itself. Cars without drivers are here. Yeah, it sounds like something out of the Jetsons, but get ready because you may look over at the car next to you and see it rolling down the street.

1:02:18Timothy B. Lee:The TV newscasters always use the same gee whiz tone. They can never resist a Jetsons reference. In every city, the influencers hop in to record testimonials for their daily serving of clout. So in today's video, I'm about to take my first ever driverless car. It's with an app called Waymo. Waymo is basically driverless car Uber, where it's like ride service. You call it, go wherever you need it to go, but there's no driver. You guys, this is creepy. It's like I'm being driven around by a ghost person. It's a little terrifying. It is definitely... Robo taxis poll hilariously badly. According to J.D.

1:02:53Power, a data analytics firm, among people who've not ridden in one, consumer confidence is at 20%. But among people who have taken a ride, the number shoots up to 76%. It's a thing I didn't capture in this story. But when I sat in one a couple years ago, I just found it persuasive as an experience. You know what? I'm not as nervous as I thought I was going to be. This is actually quite relaxing. Nice gradual turn. Felt very safe. You know, it was kind of freaky at first, but now it's pretty chill. It's a smooth ride, though. It wasn't driving fast.

1:03:25Timothy B. Lee:It wasn't jerking. It's driving like you always hope your Uber driver would. So I guess that's one of the big... Chris Ermsson, the methodical team leader, had left Google years ago. But he told me about his experience as a civilian consumer trying a Waymo out in the world. My universal experience has been, and you can tell me if this was your experience, the first couple of minutes in the vehicle, it's, huh, that's crazy. There's nobody behind the wheel. Ooh, swimming with sharks. And then a few minutes in, it's like, okay, you know, is this just going to drive? Is that all it does? And then, you know, 10 minutes in, people are looking at their phones.

1:04:04Timothy B. Lee:People tend to feel safe in these cars, but are they, actually? So we know that the Waymo driver has now driven over 200 million real-world miles, and they've released safety data so far for the first 127 million miles. Waymo's fairly transparent. They released their crash and safety data unredacted to the public. By contrast, Tesla redacts the details of its crashes. The company says they are confidential business information. In Waymo's case, I've looked at the data. I've looked at how the company interprets it, how skeptical, independent researchers interpret it. I wanted to walk through it with an autonomous vehicle reporter I trust.

1:04:44Timothy B. Lee:His name is Timothy B. Lee, author of the newsletter Understanding AI. I asked him how much our picture of the Waymo safety data has been evolving. So it's been pretty consistent the last couple years. They are scaling up, and so all the numbers get bigger, like the total number of miles get bigger, The number of crashes is getting bigger, but the light crashes per mile have not changed a ton. Waymo says, and I think this is correct, that it's roughly 80 % safer in terms of crashes that are severe enough to trigger an airbag, crashes severe enough to cause an injury, and also crashes involving vulnerable road users like pedestrians or bicyclists.

1:05:23Timothy B. Lee:So 80 % fewer airbag crashes than human drivers, and actually 90 % fewer crashes that cause a serious injury. Some independent experts have small quibbles with the methodology, but broadly they find Waymo's data credible. Timothy pointed out there's one very important thing we don't know, the fatal crash comparison. For every 100 million miles humans drive, we cause a little over one fatal crash. The Waymo driver has driven 200 million miles without causing a fatal crash. But statistically speaking, that could still be a fluke. Some academics have suggested we'd need about 300 million miles to have statistical confidence.

1:06:03Timothy B. Lee:In the hundreds of millions of miles the Waymo driver has traveled, it was involved in two fatal crashes, which it did not appear to cause. Here are the details of those crashes. In one, a speeding human driver rear-ended a line of vehicles at a stoplight. There's an empty Waymo in the line of struck cars. In another crash, a Waymo was yielding for a pedestrian. It was rear-ended by a motorcycle. The motorcycle driver was then struck by a second car. That's everything. When Timothy B. Lee looks at the entire safety picture, the results we have so far from this big experiment Waymo is conducting on American roads, what he sees is mainly promising.

1:06:43So far, it's been better than human drivers. And so far, I think the case for allowing them to continue the experiment is very strong. Which doesn't mean we shouldn't scrutinize this Waymo experiment as it continues. I find myself paying a lot of attention to Waymo crashes, which isn't hard. They make headlines. The most harrowing one recently was this January. A child near an elementary school in Santa Monica is struck by a Waymo. A child ran across the street from behind a double parked car and a Waymo hit the kid. Santa Monica police saved a child. A 10-year-old girl was not hurt.

1:07:16Timothy B. Lee:The company issued a statement. Waymo said its driver had braked hard, reducing speed from 17 to under 6 miles per hour, a faster reaction they claimed than a human driver would have been capable of. What happened next at the accident scene actually answers a question I'd had. What does a Waymo do after a car crash, since there's no human driver to help? Waymo employs what they call human fleet response agents, human beings who can't remotely drive the cars, but who the car can ask questions to if it gets confused. In Santa Monica, the Waymo called one of those humans. The human called 911, and this is the strangest part of Waymo's statement.

1:07:53Timothy B. Lee:Apparently, the car then waited at the scene of the accident until the police dismissed it. That's what we know so far, but there's two federal agencies investigating this crash, and so we'll have a full report in the future. One problem that's not really captured in the safety data that I've seen is what I'd call troubling edge cases. You see them in videos on social media. A Waymo gets stuck at a dead stoplight or blocks an emergency vehicle. Or, an example Timothy gave, Waymos were driving past stopped school buses in Austin. I think it's reasonable to say this is like a clear-cut rule that the vehicle should follow this rule.

1:08:27These educators are still very rare. And so if it's a one in 10 million thing, I think it's not that big a deal as long as they are making progress, which for most of these I think they are.

1:08:35Timothy B. Lee:Timothy pointed to one area where Waymo's not been as transparent as he'd like. Those human response agents, some of which are based here, some in the Philippines, there's questions about what specifically they do and about how this will all work as Waymo scales up. We asked Waymo for comment on everything you heard in this episode, especially the recent safety incidents. A spokesperson said that the data to date indicates that the Waymo driver is already making roads safer in the places where they operate and says that Waymo continues to work with policymakers and regulators to improve its technology.

1:09:07Timothy B. Lee:That's the safety picture so far, which to me, after many months of looking at this and talking to experts, looks pretty good. As Waymo continues its rollout, other companies are quickly falling behind. Amazon's new driverless taxi is launching in Las Vegas this summer and it's expected to arrive in LA next year. There's other robo-taxi companies like Amazon Zoox. Uber is back in the mix, not making technology, but partnering with these robo-taxi companies. WeRide recently struck a partnership with Uber to bring its AVs to Abu Dhabi, another sign that... And many of those early Waymo engineers are now CEOs of autonomous companies themselves.

1:09:44Timothy B. Lee:Dmitry Dolgov is actually co-CEO at Waymo. But other team members run driverless trucking companies. Got Don Burnett, founder and CEO of Kodiak AI. Don, thank you so much for joining us. It's good to see you again. Don Burnett is head of Kodiak AI, which has its technology deployed in driverless trucks in the Permian Basin. Please welcome CEO of Aurora, Chris Urmson. A big round of applause. Chris Urmson now heads Aurora, which currently has semi-trucks on Texas highways. And my personal favorite plot development, which just emerged this week. I just broke on the information that Uber founder Travis Kalanick is starting a new self-driving car company with financial backing from Uber and in partnership with Anthony Lewandowski.

1:10:28Now, for those who have been...

1:10:29Timothy B. Lee:They say there's no second acts in American lives. Somehow, both of these men seem to be on their fourth. The big picture, though, is that everywhere in America today that you see a driver, taxi, truck, food delivery, there are several companies working on the robot version. Trying their best to make driver, as a job, start to go the way of the knocker-upper, of the lamplighter. Those knocker-uppers, by the way, they disappeared quietly. The lamplighters did not. Writer Carl Benedict Frey tells the story of the Lamplighters Union, how their strikes plunged New York City briefly into darkness, to the delight of lovers and thieves.

1:11:08In Verviers, Belgium, the Lamplighters' strikes turned violent, ending in an attack on the local police headquarters the army was brought in.

1:11:17Timothy B. Lee:The Lamplighters lost their fight, in part just because they were so outnumbered. But the drivers today, fighting to save their livelihoods, are a significantly bigger force. Please stand up. Everybody that's rideshare, union members, or someone who drives a vehicle, stand up. 4.8 million Americans drive for a living. It's one of the most common jobs we have. And these workers do not plan to surrender to the California tech companies. They're doing this because they stand to make an unfathomable amount of money if they eliminate driving jobs for working class people. I understand it is business.

1:11:56It is capitalism. But not in my city at the expense of our jobs.

1:12:02Timothy B. Lee:These drivers are represented by unions, backed by politicians. And in cities across America, blue cities, they're organizing. So far, they're winning. Humans drive this city, not machines. Labor drives this city. Keep the workers in the workforce. If it works in another city, great. Have fun. Not here. Not Boston. Thank you.

1:12:27Timothy B. Lee:Next week, the fight to save a job. To save the human driver. Don't miss this one.

1:12:48Many thanks to PJ Vogt and the entire search engine team for this story. You will hear part two right here on Freakonomics Radio very soon. Until then, take care of yourself. And if you can, someone else too. Freakonomics Radio is produced by Renbud Radio. You can find our entire archive on any podcast app. It's also at Freakonomics.com, where we publish transcripts and show notes. For Search Engine, this episode was produced by Emily Maltaire. The show was created by PJ Vogt and Shruti Pinamineni. Garrett Graham is their senior producer. Leah Reese Dennis is their executive producer. Fact-checking was done by Mary Mathis.

1:13:24And sound design and original composition by Armin Bizarian. Their production intern is Piper Dumont. For Freakonomics Radio, this episode was produced by Dalvin Abouaji and edited by Ellen Frankman. The Freakonomics Radio network staff also includes Augusta Chapman, Eleanor Osborne, Elsa Hernandez, Gabriel Roth, Elaria Montenacourt, Jasmine Klinger, Jeremy Johnston, Tao Jacobs, and Zach Lipinski. Our theme song is Mr. Fortune by the Hitchhikers, and our composer is Luis Guerra. As always, thanks for listening.

1:13:57Is it possible that you were really stoned on painkillers in that first Waymo ride? I mean, I wasn't stoned on painkillers, and I don't think I was stoned at all. I think I really had a sense of normal technological awe.

1:14:16The Freakonomics Radio Network, the hidden side of everything. Healthcare can feel complicated. That's why Optum uses technology to connect the people and processes that make healthcare easier, more affordable, and more effective. We're making it clearer for you to know exactly what your benefits cover. And to help you better manage your health, we're coordinating care between your doctors and your technology. We believe better, simpler healthcare is always possible. That's healthy optimism. That's Optum. Visit Optum.com to learn more.

1:15:12What are you waiting for? Play Chumbo Casino and enter for your chance to win. Let's Chumbo! Only available in permitted states. No purchase necessary. See terms and conditions for details. VGW Group. Void or prohibited by law. 21 Plus. Sponsored by Chumbo Casino.