Behind The Tech with Kevin Scott - Chris Urmson: Aurora CEO - Autonomous Driving

Episode Date: January 30, 2020

Learn about the latest thinking for autonomous driving technology from Aurora CEO Chris Urmson, who led Google’s self-driving car program. This former Carnegie Mellon professor also tells us about his days with DARPA and shares predictions for 2020 and beyond. Listen to other Microsoft podcasts at aka.ms/microsoft/podcasts

People talk about Silicon Valley engineers being risk takers. I think it's actually the opposite. It's the realization that if you go and try one of these things and you're actually good at what you do, if it fails, it fails. You have a job the next day at somewhere else, right? You'll have this wealth of experience that people will value. Hi, everyone. Welcome to Behind the Tech. I'm your host, Kevin Scott, Chief Technology Officer for Microsoft. In this podcast, we're going to get behind the tech.

We'll talk with some of the people who made our modern tech world possible and understand what motivated them to create what they did. So join me to maybe learn a little bit about the history of computing and get a few behind-the-scenes insights into what's happening today. Stick around. Hello, and welcome to our first episode of Behind the Tech in 2020. I'm Christina Warren, Senior Cloud Advocate at Microsoft. And I'm Kevin Scott. All right. So, Kevin, it is 2020, which is both the new year and I guess a new decade, although people will get weird about the technicalities.

And it's always a great chance to kind of look back at what's happened over the last 10 years and reflect on new opportunities. Yeah, I mean, it is, I think, in our industry and for human beings in general, really easy to get completely used to new innovations that enter our lives. But like when you think back 10 years ago, the world looked like a very different place than it looks right now. So smartphones were just catching on. They were nowhere near as ubiquitous as they are right now. And the things that you could do on them were far, far more constrained than they are right now. I mean, for God's sake,

people were renting movies from Blockbuster in 2010. Right, right. Blockbuster was actually still a thing. And Instagram hadn't even been invented yet. Totally different world. Like, you know, now that we've hit 2020, do you have any forecasts about what the next year in tech might bring or even the next decade? Well, I think one of the themes that we spent a bunch of time chatting about last year on the podcast was artificial intelligence and machine learning. And I think we are certainly going to see the trends that had started in the prior years continue to accelerate. It's one of the reasons why I'm really interested in chatting with our guests today. So autonomous vehicles, for instance, I believe are going to make a ton of progress over the next couple of years

in particular. And I'm just sort of really looking forward to seeing some of that stuff play out. Yeah, I couldn't agree more. It's funny. I don't have a driver's license, but I've actually been on a few self-driving car panels over the years. And I think the technology behind it is so fascinating, which is why I'm really, really excited about your conversation with today's guest, Chris Urmson. And Chris is an engineer who's known for his work in pioneering self-driving car technology. Yeah. And, you know, one of the reasons that I'm especially interested in self-driving cars and am looking forward to this conversation that we're about to have with Chris is that there's so many ways that the world is going to change for the good once we are able to put this technology into the hands of lots of different companies.

So one of the things that we'll hear about Aurora is they are a company building the self-driving car technology as a platform for other companies to use to build autonomous applications. And so, you know, like one of the things that I'm sort of hopeful that will come into the world in the not-too-distant future is some technologies that may help my grandmother. So I'm lucky enough to have a grandma that's still alive. She's 89 years old and lives in a very rural place in Virginia. And she can still drive, which is awesome, but the day is coming where she's not going to be able to drive her car in the same way that she is right now. And, like, then it begs the question of how she has access to all of the things that she needs in order to help her live an independent life. So how does she get her prescription medicines? Like, how does she get her groceries and, you know, just sort of the staple things that

she needs to exist? And one of the things that I think could be really incredibly beneficial with these self-driving technologies is, like, the possibility that you'll be able to have autonomous deliveries for people like my grandmother. I think you're absolutely right. I think the potential for this stuff is really fantastic. So let's hear more about some of the potential for this technology from Chris Ermson. Chris is the co-founder and CEO of Aurora, a company that builds self-driving vehicle technology.

Before founding Aurora, he was CTO of Google's self-driving car program. Prior to that, Chris was a faculty member of the Robotics Institute at Carnegie Mellon University, where he was the technical director of the DARPA Urban and Grand Challenge teams. I'm really excited to hear what he's up to these days. Hey, Chris. Welcome to the show. Thanks for having me. So I'd love to start by learning how you got interested in technology in the first place as a kid.

Were you taking engineering classes or programming classes when you were in high school? Or you just discovered that in college? Back when I was in high school, there wasn't really computer science in high school. And so, I bought some kind of Tandy x86 clone or whatever, you know, back when I was in probably ninth or tenth grade from money from my paper route. And, you know, tried to learn to program at first where you go, you know, I don't know if you recall this, but you go to the bookstore and you'd buy, you know, this paperback book that was, you know, program whatever it was. And it was just the source code listing. And this is before CD-ROMs even. Yeah.

Which people probably don't even remember now. That's right. You know, before that, actually, we'd bought a Commodore 64. And, of course, that was exciting because it didn't have a tape drive. Right. Right? It didn't have a floppy disk drive. It didn't have a floppy, yeah.

And it had five and a quarter inch disks. That's what it had, yeah. Yep. So anyway, so we was doing that. And then there's this language, C++, which seemed to be the hot new thing. And so started actually the first program language I really learned was C++.

Wow, that's rough. Yeah, no, it was a little crazy. I mean, and I guess on some level, Wow, that's rough. Yeah, no, it was a little crazy. And I guess on some level, C++ is a challenging first language, but the good thing is after you've mastered it. It's all downhill. It's all downhill. And so did you know from all of this experience in high school that you wanted to get a computer science and engineering degree? You know, I was up in Canada, so applied to a variety of schools, got into a couple of them.

And then in my senior year, I met a girl. Turns out now she's my wife. And decided I wanted to stay at the University of Manitoba, which is right in central Canada, Manitoba, and got into the computer engineering school. Computers seemed like, you know, they had a future. Yep. And so you got your undergraduate degree and you went straight to grad school, right? That's right.

And you went to grad school at Carnegie Mellon. Yep. So how did you know you wanted to go to CMU? One day, I was in the engineering building just outside the library, and there was this poster next to the elevator that showed this robot crawling out of a volcano. And I saw that. I thought, that's really cool, right? Like, I like robots. I like space. This thought, that's really cool, right? Like, I like robots. I like space.

This seems exciting. And my girlfriend said, you know, you should really apply if you think that's cool. And I figured, you know, it's Carnegie Mellon. It's like there's no way. How am I, you know, at the University of Manitoba, how am I possibly going to get in there? And, you know, they made a mistake. Well, no, I think they did not make a mistake. You know, we'll see.

But yeah, no, it was really, I've been fortunate throughout my career to kind of look at, hey, this seems interesting. This seems cool and fun. And, you know, gone and tried it. And it's mostly seemed to work out so far. And so what was the experience like at Carnegie Mellon? Because I remember it still is to this day like just one of the most extraordinary places in the world, especially to do computer science and robotics. And like the Robotics Institute is just fantastic.

So what was that like as a grad student? For me, personally, it was eye-opening, right, that you would have people come and lecture who had written the textbooks I'd used in undergraduate. And, you know, so meeting these people, meeting people that worked at NASA, meeting people who worked at DARPA, meeting, you know, people from, you know, Microsoft or Intel at the time. And just it opened up this whole other world of possibility. And then the faculty were great, right? There wasn't really, as a graduate student, you didn't see politics. You saw people working together. We got to work on cool things. You know, one of the things I love about Carnegie Mellon is that it's very much a systems school. You know, they've got incredibly deep, strong, fundamental,

theoretical underpinnings, but it's about make it work and see it out in the real world and learn about that part of, you know, the development and engineering process and actually touching real things. And so, it was fantastic. And, you know, I got to go up to the Arctic Circle, and we had a robot up there driving around. I got to go down to the Atacama Desert with a robot and, you know, explore that. It was just, it was an incredible experience. What was the first useful robot that you worked on? Useful? I don't know that I've worked on a truly useful robot yet. We're getting closer.

No, we, so the robots we built for going to the Arctic, this was a robot called Hyperion. And we were exploring how can you make a robot think about, how can you make it so a robot could operate perpetually? So, one of the challenges, you send a robot to Mars and you put solar panels on it. Well, it can only operate when there's enough sunlight. And it can only operate when there's a communication window back to Earth. So we were looking at both how do you plan so that if you, say, launch it to the pole of a planet, that you have constant power by rotating and driving such that the solar panels always pointed at the sun. And then how do we make science discovery automated? So instead of asking, you know, should I look at this rock?

Should I look at this rock? Should I look at this rock? Have the robot go and look at a bunch of rocks and then try and figure out, hey, this one was unique and interesting in some way and send that back so that you could maximize the use of the narrow comb bandwidth that was available.

So that was cool. Didn't go anywhere other than as a research experiment, but some of the technology ended up making it happen. And what year was that? That would have been 2001, I think. So that was before the, like, the big deep neural network computer vision revolution. Yeah. Oh, very much so.

This was the, you know, you spend five years working on your PhD and you make something 20% better by coming up with a new set of feature vectors and, you know, you earn your PhD. And so your PhD was, were you, I'm guessing when you're working on robots, there's such broad systems. So like there's, there's the software and the software is very complicated. It's all the way from control loops to perception to planning, and it's just a ton of complexity there. And then there's also all of this complexity on the electromechanical side of things. Like, how do you make it light enough?

How do you give it the sort of strength and durability to do the things that you want it to do? How do you power it? How do you make it resistant to the environment? So, did you have one thing or the other that you specialized in or that you gravitated towards? So, I was definitely more on the software side and kind of the software and systems side of how do you think about the whole thing working together. You know, I'm definitely not a mechanical engineer. Now, I kind of know enough to be dangerous and really frustrating, I'm sure, for the mechanical engineering folks that I work with. But on the software side, I kind of worked at this intersection of motion planning and perception. And so say a little bit more.

What is motion planning? Yeah. So motion planning is figuring out how do you make the vehicle move through the world. And there's a bunch of different techniques you can use for it. And there's variations where you're thinking about just kinematically what the limits are. And then you think about kinodynamic motion planning where you're actually counting for the fact that there are dynamics of motion and inertias. And so the earliest robots, like this robot we took up to the Arctic, it moved at 15 centimeters a second. And so, you know, put that in context, that's like a slow person with a walker.

Yep. And it had to be that slow. Why? Well, one, we actually probably couldn't plan much faster than that. Gotcha. So it was like the speed of the processing. Yeah. It was complicated. It was truly off-road. It was also solar-powered.

So, there's a power budget limit given we were driving this thing around. And so, there's only so much power we could put through the gearboxes and the motors. Fascinating. But so, even though you were on the software side, like if you are writing planning software and like the planning software, I mean, like we just are getting to the point where deep learning systems can learn a kinematic model from scratch. But so like these kinematic models are basically they're sort of a model of the physics of the system. And so you do have to understand mechanics in order to write the software. Yeah, you really have to look at the kinematics of the vehicle and figure out, you know, in our case, this vehicle, the Hyperion robot was four-wheel drive,

and it had this cool design, made it much more fun, where the front axle was just, there was no active actuation on it. So, if you turned one wheel faster than the other, that would turn the axle. And then if you want to change the trajectory of the overall vehicle, you're actually driving all four wheels and you're kind of slipping the back wheels and steering, you know, steering the front wheels. And, you know, it's actually a really interesting, as you say, you know, you have to understand the mechanics of it. You have to understand the forces that are interacting to make it work well. Yeah. And then, I mean, just sort of the perception part of navigating the real world is also complicated. Some of the, I mean, for a long time before the, some of the more recent advances in perception

that are being driven by deep neural networks and GPUs and whatnot, like, perception was one of those, like, real intelligence, artificial intelligence gaps that we really weren't sure when we were going to be able to bridge. And so, you were doing all of this work where we were still in that universe of confusion about will we ever be able to get the perception to be as good as a human? Oh, yeah. And again, it's been fun to watch

what we can actually bring into scope and start to solve. So back then, this robot was designed to drive around in the Arctic. It turns out one of the benefits of being in the Arctic is there's nothing there that moves, right? There's just rocks. And so we were using stereo vision systems to reconstruct 3D geometry and then estimating, you know, the load-bearing surface from this and thus figuring out which were the flat bits we could drive on. You know, we were not particularly adventurous in the train we would drive over. And then, you know, the next step from that was looking at, there's this great program

that I wasn't part of at Carnegie Mellon called Crusher. And this was this six-wheeled, multi-ton thing that could move at, you know, I don't know, 30 miles per hour, but it was like a little tank. And so there, they were figuring out the load-bearing surface, but they were also figuring out, like, what's the stuff you can just drive through? Because if you're driving a tank, it turns out the grass doesn't get in your way, right? The small trees don't get in your way. And so, you know, that was the next level of figuring out what was vegetation, what

was the load-bearing surface, you know, what's the slope I can really drive on? And then a lot of the stuff that I've been associated with over the last decade has been figuring out what was vegetation, what was the low-bearing surface, you know, what's the slope I can really drive on. And then a lot of the stuff that I've been associated with over the last decade has been, you know, the actual underlying geometry of the world is pretty simple, right? You drive on a road, it's flat. But now you have all these actors moving through it, whether it's vehicles or bicyclists or motorcycles or pedestrians or, you know, ducks or wheelchairs. Right. Right. Right. And so now you have to separate the stuff that moves from the background and be able to track that and then understand that is a car or that is a pedestrian because that,

your model of how they're going to behave in the future is somewhat a function of what you think it is, right? Right. A pedestrian is very unlikely to move at 60 miles an hour. Right. You know, whereas a car, you know, that can happen. And so the way you interact with these vehicles and actors changes. Interesting.

Yeah. And so while you were at Carnegie Mellon, like, were you a graduate student or faculty? Like, the DARPA Grand Challenge, like, happens. And I remember this was just one of those shockingly cool things as a computer scientist. Like I wasn't a roboticist or like even an AI person. And I was like paying very close attention to this program and what was happening because at the time it was just sort of shocking the notion that you could have an autonomous vehicle that could like navigate a complicated environment by itself. So, you were part of that. It was really cool. So, this was, I was a graduate student, and then I worked for a company,

and then I was a faculty member. So, I kind of went through three different phases of life, I guess, there. And yeah, it was really cool. And it was a grand challenge because back in 2002, we were not sure you could actually do this. So – And so what was the this? Yeah. So the first – so there was three different events. There was two grand challenges and then the Urban Challenge. And the Grand Challenge was to get a robot, we called it a robot back then, to drive across the desert to 150 miles,

nominally from Los Angeles to Las Vegas. And it had to do it in less than 10 hours, and it had to do it on a given day. So, you know, we showed up, launched the robots in the morning, and, you know, some number of hours later, they come back, hopefully. And when it kicked off in 2000 and probably late 2002, 2003, we were, like I said, we had conversations where is this even doable?

There had been a lot of great work in this space before. There'd been No Hands Across America, which was a Carnegie Mellon program that had actually used neural networks to drive a car down the freeway, or at least steer a car down the freeway. It didn't do gas and brake. And there'd been some great work by Ernst Dickmanns in Europe and some interesting work in Japan. But the idea you'd drive off-road for 150 miles was just like it was impossible. The state of the art was some guys had a Humvee driving at, I want to say, 10 meters per second through a field with giant hay bales. And the idea was don't hit the hay bales. And they mostly didn't hit the hay bales. Mostly. Mostly, right?

And so, the idea we could go across the, you know, across the Nevada desert for 150 miles, like on command. And again, a lot of the robotic research at the time and still today, rightfully so, is you work really hard at this thing. It's such a brittle system that you get a video at working the one time, and that's your conference submission. And that's the right way it should be. In an academic setting, you're kind of trying to prove out ideas. So, the idea that we would show up one morning and, you know, the starting gun would shoot and it had to go and it had to work that time was just, it was exhilarating. And the problems, you know, it was a classic under-defined problem to begin with where, you know, we notionally were worried we'd just have to drive across tumbleweed and cactus for 150 miles and truly be off-road. Now, it kind of morphed into

drive down the trails, which made it more viable and, you know, more useful, honestly. But even that, when, you know, a trail isn't the smooth, perfect ground that we see on the road, right? There's ridges to it or rutting across it. And, yeah, it was a heck of a lot of fun. And so, when you were doing all of this, did you think that you were going to be a computer science professor for your entire career? When I, yeah, it was kind of the mission. So, you know, I, when I was an undergraduate student was, I want to go to graduate school. And then when I was a graduate student, I was pretty convinced I wanted to be a faculty member at Carnegie Mellon because

it seemed like a great place. Yeah. And ultimately got the opportunity to be that. So, it was exciting. Sort of an extraordinary thing. And so, you know, in my mind, like as someone who also thought that they were going to be a computer science professor for a while, like getting the faculty gig at Carnegie Mellon is like pretty much the top of the mountain. Yeah. So, why do something different? Yeah. So, you know, we had this chance to work on these challenges for a couple of years, and that was exciting. And then I spent a couple of years on the faculty and worked with Caterpillar.

We were automating these dump trucks that were the size of houses, right? Like, 400-ton thing moving at 45 miles an hour. Shockingly big. It's amazing. Just incredibly cool. And a great team with Caterpillar, a great team at the university. And I had been talking with Sebastian Thrun, who was at Google at the time, and we were thinking we should do something around self-driving cars. Again, we still didn't – we called them, you know, car-bow-bots or something.

I don't know what we called them, but we hadn't invented that term for him yet. And, you know, he was at Google, and I got a phone call saying, you know, effectively, like, we'd actually like to do self-driving cars at Google. And this is 2008. So Google had quietly just very recently acquired Android. And that was, you know, it was a search engine. I was like, what on earth? Why? This doesn't make any sense.

And we talked about it, and I came to realize that they, you know, Larry and Sergey thought they had this incredible engineering resource and wanted to go solve cool, interesting problems. And they were asking me to come out and lead the team. And I thought, well, you know, and initially it was a two-year gig. The idea was we'd come out, see what we could do for two years. And, you know, it's kind of like the Dread Pirate Roberts, you know, good day today will probably kill you in the morning, right? Like, well, you're going to be gone in two years. And... Yeah. And I want to like, I want to draw a line under this because I think now people probably almost take for granted that Google having Waymo and like all of the self-driving car stuff is, was just sort of an

inevitable consequence of how things were going. But I was at Google from 03 to 07. So I left before you got there, but you started close enough after my last day that, like, the company, it was a search and ads company. And, like, the idea that, I mean, you had two things. You had this company that, like, it wasn't obvious, I think, to anyone that they should be building self-driving cars. And, like, I'm not saying that in a pejorative way. It's just, you know. It's not what they did. Yeah, it's not what they did.

And the state of the technology was also nascent enough where it just wasn't obvious that commercially, like, you could make a profit on any of this stuff on any sort of reasonable time horizon. And so them having the idea to do that and you sort of saying this would be a cool enough thing to do to take a sabbatical for my tenure track job at the number one computer science program and that required a lot of courage all the way around. Yeah, no, and it was, yeah, and it was a big decision because, this was the thing that I had aspired to for the last seven years. And it wasn't obvious that it was a good idea. But it seemed like, you know, my wife and I, we just had our first son and our second. And we're thinking, well, I've got this opportunity here, but, you know, we could go try California for a couple years. And we're like, you know, this is where the crazy people with the Birkenstocks are.

So, you know, if we're going to go, we may as well go give where in the fall, the new faculty members come in and talk about their research and their aspirations and whatnot. And there was three of us. And, you know, I came in and did my, you know, little thing. And at the end of it said, and by the way, I'm – you know, now it's – because I think I was special faculty and then I was faculty, faculty. By the way, I'm leaving for two years. Thanks for having me. You know, and went over about as well as you might think. But, you know, it was, it's to their credit, they let me do it.

And to Google's credit and Larry and Sergey's, you know, they had the foresight to try this out. And it really was, can we, we went into it with no degree of confidence that it would go anywhere. Just go try it. What did your dad think of this? Or your mom? They were very supportive, right? It just seemed like a neat opportunity. And, yeah, I think if I had left to try a startup, I think that would have been a little more, you know, I can tell you actually, they were a little more skeptical of that.

But, you know, the opportunity to go there, they got it. Yeah. I know my mom was really anxious for reasons I didn't quite understand when I left Google to go do a startup. Yeah. It's, like, just very hard for her to reason about, like, how you could leave a good job like that to go try this, like, very risky thing. Yeah. No, and it is, right? And I think one of the things that people outside of Silicon Valley and haven't been here don't realize is that it's not really.

That, you know, people talk about Silicon Valley engineers being risk takers. And I think it's actually the opposite. It's the realization that if you go and try one of these things and you're actually good at what you do, if it fails, it fails. You'll have a job the next day at somewhere else, right? And you'll have this wealth of experience that people will value. And I think that is something that it's hard, you know, I'll categorize this as, you know, East Coast people, but, you know, kind of more conventional business folks don't kind of have that sense of the opportunities that are around. And maybe we've just been here during a particularly fortuitous time.

Yeah, but I do think it's a piece of advice that I give people all the time is that if you, however you do it, like whether it's like be a software engineer in Silicon Valley where you've got a bunch of career opportunities or whether it's avoiding taking on like huge amounts of debt early on in your life, if you can give yourself the opportunity to be able to have choice in what you work on and to like always choose the thing that's interesting and looks like it's going to give you a bunch of learning, like you will probably do better than you would in other circumstances. Oh, for sure. There's, you know, this, what is it?

If you do something you love, you'll never work a day in your life. Yeah. Right? Like that, it's not just that, right? It's that you get the best out of yourself. You get the best out of people when they're invested in the thing that they're doing, when they're passionate about it. Yeah.

You know, and it's, I guess it's easy for me to say, having been in a position where I've had the opportunity to do this. And I, you know, understand that not everybody has the same opportunities that I've had. But if you can find that, if you can, if you are able to seek it out, it seems like the way that you get a chance to shine and the way that you help build something awesome. So, tell me about the early days of the self-driving car program at Google. So it started, it was like you and Sebastian. Yeah, well, there was, I want to say there was a half dozen of us who started at the beginning. There was me and Dmitry Dalgov,

who's now the CTO there. He and I started on the same day. And Dirk and Mike Montemurlo, who was actually an office mate of mine at Carnegie Mellon and is one of the brilliant minds in simultaneous localization and mapping. And who else? There was, oh, Henrik.

So we were very much a black project. So we were kind of sworn to secrecy. We were kind of hanging out in the Google Maps building, but people didn't really know what we were kind of sworn to secrecy. We were kind of hanging out in the Google Maps building, but didn't really, you know, people didn't really know what we were working on because it was kind of weird. And it was how do we learn as quickly as possible? How do we show, you know, how do we convince ourselves that we could actually build something meaningful with this technology? And so, we had two goals when we started. The first one was to drive 100,000 miles on public roads, and then the other was to drive 10 100-mile really interesting routes. And the idea was kind of to get kind of

statistically interesting coverage and, like, focused ability to learn about particular places we could go. And so – And it was, like, drive safely 400,000 miles on public roads, right? I mean, like, I know that it's implicit in what you're saying. Yeah, of course. It bears saying because one of the things that I think you all did extraordinarily well is, like, you were very thoughtful and even conservative in a way with the choices that you made to make sure that safety was always the number one thing. Yeah, that was one of the things that I helped instill in the organization early on. And we had a great group of people who would have done it without me.

But how do we think about training people to operate the vehicle? How do we transfer the knowledge from the engineering team to the operations team so they understood what was going on? What processes we put in place so that we have people checking one another? How do we make sure that we trust the release software going out to them? So yeah, that was a really important part of it. And that's one of the fundamental bits of culture that I helped instill there that I'm very proud of and one of the cores of the way we operate at Aurora as well. Yeah. So, you had these two goals.

Yeah. And nascent technology. So, and this is 2008. 2009. Yeah. I started February 2009. Okay. So, yeah.

And so, we did exactly what engineers do. We optimized to the constraints. And so, for the 100,000 miles, the vast majority of that was driving on 280 between 85 and Sneath. And so, we had a fleet of Toyota Priuses. And if you lived here in 2000 and 2010, you would have seen a bunch of them going up and down on the freeway there, gathering 100,000 miles. Yeah, you guys were on my commute. My startup was in San Mateo, and I lived in San Jose. So, I was on the 280 every day. Yeah. And, you know, they got incrementally better over time, and, you know, we burned through the 100,000 miles.

It was really kind of a fun time because, you know, people didn't know what they were, right? We had people think they were weather trackers. They thought we were storm chasers. They thought this was one of my favorite. I was at a gas station, and somebody was convinced we had kind of a perpetual motion machine, which is on its face a problem because we're at the gas station. But, you know, like the laser that was spinning on the roof was clearly a wind turbine. And at the time, we had this encoder on the rear wheel. And so that, you know, their model was we were gathering from the wind from the car driving, and then we were recuperating energy through the encoder on the wheel.

And it's one of these things where we couldn't tell them what we were doing, so we kind of nod along. I knew a lot of people who knew Google's business who thought that they might have something to do with Street View. Oh, yeah. Yeah, I remember we were definitely that. We had one time we were driving back, we were up in Sacramento or Tahoe testing, driving back through that stretch of 80, just kind of where it's really flat. And it turns out there's a part of that where there's like a parallel road. And we're driving along and this motorbike passes us, cuts over to the parallel road, and then pops a wheelie. And for like three quarters of a mile, it's just doing a wheelie beside our car. And I am sure that for the next two or three years, that guy was still checking for the Street View footage

of him doing his wheelie and thought, yeah. That's awesome. Good times. So, what had to change after you accomplished the initial goal? So, the initial goals were good ones, but, like, that is not, like, how you get this initial proof point that the technology at least can work some of the time on traffic, on public roads with real people. So now, how does this turn into a product? What is the application we're going to go after? What is kind of the MVP? The first thing we thought about was basically what looks like autopilot today on Tesla,

except we set a very, very high reliability goal so that you really could kind of take your eyes off the road for periods of time. And so we spent a couple of years working towards that. And, you know, I've talked in the past about some of the experience we had at that. At the point, we got to a point where it was extremely good. And then we started to dog food it. And, you know, despite warning people about the fact that, A, it was beta, and B, you know, effectively, we knew where they lived. They worked for the company. We were going to monitor what they did. We saw them, you know, being overly confident and, you know, checking things in the back of the car or putting on their makeup, all kinds of things that you wouldn't want. So then we moved from that to how do we turn this into something that really can drive, do the whole driving task for you. And, you know, in'm particularly thinking about how do we

move people. Right. And so, at some point, you decided that you were going to go start a company, which is what you're doing right now. So, Aurora, tell us a little bit about that. Yeah. So, I had a tremendous time at Google. It was an incredible opportunity to work with great people, you know, good company and go and build something neat. And then by the time we got to about 2016, it was time for me to move on. So I resigned in the middle of 2016 and then spent a few months trying to figure out what to do next. And I met with all kinds of people and, you know, from tech companies to automotive companies

to startups. And, you know, it was fascinating because after you spend that much time at a company, you kind of get used to seeing the world through a particular lens. And being able to get outside of that lens and engage with people more as an individual and kind of deconstruct the lens, you learn a lot through that process. And what became clear to me was there was a chance to help accelerate this technology. And that given the experience I had and given kind of the state of the industry at the time, it seemed like if I could find great people to found a company with, we could really do something meaningful as an independent company and, you know, build technology that mattered and see it

in the world in a way that we could do a lot of good, you know, and ultimately build a sustainable, successful business. Yep. And so, what does Aurora do? So, Aurora's mission is to deliver the benefits of self-driving technology safely, quickly, and broadly. And what that means is we want to make a driver that will make it easier for people to get around, make it less expensive for them to get around, and easier for us to ship goods and perform logistics. And you all think about yourself as a platform company. So, anybody who wants to build an application that needs self-driving cars, like, they could use your technology to, like, help realize it.

That's exactly right. We think about the thing that we can do best in the world is build the driver. And so, let's go do that. People, you know, it's easy. You look at, say, Uber, and what you see is an app that you figure, how hard can that be? And what you miss is just how complicated that business is behind the scenes. And, you know, the operational aspects of it, the messaging driver with vehicle or passenger with vehicle. If you look at, again, if you look at a company like FedEx, right, it's incredibly complicated, the logistics that they do. And so we don't want to go build that, right? We think those people know what they're doing. They're really good at it.

Similarly, you know, it's really easy to dismiss an automotive company, right? Oh, they just, you know, they just bend metal. Well, no, it turns out that they built these incredibly complicated products that operate for the next 15 years and have an immense amount of technology under the hood. And so, you know, we don't want to do that either. We're not going to be very good at that. So, let's concentrate on this one thing, this driver, and let's make it a platform that everyone can use that will make it, for the end user, it'll make it safer for them to get around. It'll make it easier and it'll make it less expensive. And then for our partners in the automotive community and in the transportation logistics communities, you know, they'll have an opportunity to help build better businesses themselves.

Right. And the idea is like there will be lots and lots of these businesses, like not a small number, but like you will have maybe hundreds or thousands of companies that are like instantiating some sort of AI driver in a vehicle of some sort. And the vehicle could be your giant 400-ton Caterpillar thing, or it could be like a consumer-owned automobile that you're using to commute to work, or it could be a completely unmanned drone vehicle of some type. Yeah, that's exactly right. I think I don't have enough imagination, and we probably as a team don't, to figure out how all the ways, all the great ways this could be used out in the world. So, again, stick to the thing we know how to do, open it up and let it flourish. So, one of the things I'm really interested in

is your perspective on what has changed in the technology over the history of autonomous vehicles from the point where you entered, which was pretty early, you know, like, you know, sort of grand challenge, you know, DARPA problem all the way through to 2020. Yeah. It's been incredible. You know, you could start by thinking about the hardware side of this and maybe even the mundane hardware side. So, vehicles are now almost all electronic. And so, you can communicate to the brake system or the steering wheel

or you can talk to the gas pedal, right? And that wasn't the case. We literally had a motor with an arm on it that would press the brake pedal in those earliest vehicles. That would compress a hydraulic cylinder somewhere. And they all still have hydraulic cylinders, but the front end of them now, almost all of them, there's no mechanical linkage. Oh, sorry, no. No, there actually is in cars. But on top of that, there's also an electronic

control. And that's actually one of the challenges. And so you have the mechanical linkage for safety? So today, yes. And this is actually one of the areas where there's still steps that have to be made before we will see broad distribution of this technology. So this is fascinating. Like, this is some of the areas where there's still steps that have to be made before we will see broad distribution of this technology. So this is fascinating. Like, this is some of the minutiae I think we, like, sort of skate over sometimes when we're thinking about these technologies.

But, like, one of the things that the auto industry does incredibly well is, like, they understand manufacturing for safety perhaps better than any other industry, maybe other than aerospace in the world. And, like, they have thought about this problem so deeply and for so long. Oh, yeah. And that's why it's so great to work with them on these kind of problems. And let's take one very specific example. So let's talk about the brake system on a car.

So the way that works today is when you press the brake pedal, you're actually pressing, you've got a metal arm that's going in and compressing a cylinder. That's the master cylinder that's then creating pressure, hydraulic pressure in your brake system. And that brake system is actually split into two circuits that are generally diagonally linked across the vehicle. So that if one of them fails, you've still got two brakes. And it turns out that the torques applied by the brakes on the wheels don't spin the car when you apply the brakes. You don't have as much authority, but you still have brake power. And that's a mechanically, like, brilliant design.

Awesome. And then what's happened is with the introduction of more advanced features, like first electronic stability control and now adaptive cruise control, there needs to be a way for a computer to actuate some part of the brake system. And with electronic stability control, you're actually modulating whether that pressure that's being applied by the master cylinder is effectively on or off around that. And with adaptive cruise control, you're actually wanting to generate pressure itself. Now, in both of these systems, the redundancy, so if the electronics, if the

electronic system fails, you're still able to apply pressure through that master cylinder system. And so, you, you know, the force of actuating the brakes is coming from your legs. If you have a robot driving, if it's a self-driving car, then there's nothing applying that force. And so we need to design new brake systems where there's parallel systems to actuate the force. And so this is one of the parts that will become standard on vehicles as we move to automate them. Similarly around the steering shaft, right, the steering column.

So, the backup in if the power steering system breaks is your, you know, your arm muscles. But if you are a self-driving vehicle with nobody grabbing the steering wheel, then we have to come up with an alternative system for backup. And so, it's, you know, interesting things like dual-wound motors or second motors on the shaft. And so in addition to like re-engineering the cars themselves to have new forms of mechanical redundancy, like you also have to think about how you make the software redundant. Like this is one of the things that wasn't obvious at all to me when I first started learning about self-driving vehicles

is I think people have in their imagination that the software is this one monolithic machine learned thing that is doing all of the perceiving of the environment and all of the, you know, sort of driving and reacting to the environmental conditions. And like that is not the way the systems are engineered. So, you don't necessarily need to make it redundant, but you need to have a certain level of resiliency in it. And redundancy is one way that you could implement that level of resiliency. And so, you could imagine having two teams working, you know, kind of in enclosed rooms and come up with solutions. Or you can do more thoughtful things about how you

take more thoughtful approaches where you think about how the system can fail and mitigate those to a point where you've reduced the risk to the point where it's no longer unreasonable. One of the benefits we have over aviation is that we're on the ground. And so we just need to take the kinetic energy out of the system and bring it to a safe state. In aviation, you know, like you'll ultimately take the kinetic energy out, but it's going to be a very bad day. So it's a slightly different problem. And so you probably don't need the full kind of the full degree of redundancy,

but you need something where you're kind of meeting your safety targets. And so how do you think about sort of interpretability of these models that are, I mean, so like one of the things that I learned about over the past couple of years is in airplanes, you have this thing called MCAS, so the Mid-Air Collision Avoidance System. And so every modern airplane has MCAS. And, like, MCAS's job is to sort of look at all of the air traffic. And if you get too close to another plane, the planes communicate with one another. And, like, one will go up and the other will go down in order to avoid the collision from happening. And there's a document. So there's two versions of this MCAS.

So there's the old version of MCAS that is described in a document that has a bunch of pseudocode and it's like hundreds of pages long. And there is a new version of MCAS, which is, you know, sort of almost like a dynamic programming-based thing. So, it's not full machine learning, but, like, all of its decision points are, like, codified in a table. And the argument is that, like, even though it's not this human-readable pseudocode, that this new system has better provability properties that it's going to do what it's going to do. Although, like, from a human perspective, it's probably less easily decipherable. And, like, if you look at the pseudocode, like, the pseudocode is sort of a wreckage. Like, it's full of inconsistencies and, you know, no one really understands 100% of, like, what the thing says that these systems should do. And so, like, one step further along this level of sort of human inscrutability are these, like, machine learn models that we're building right now where they are very large and, like, how they – you can't easily learn how they behave by just inspecting the weights of the parameters that are sort of in the interior of the model.

So, how do you think about this? Because you're working on one of the most important safety-critical systems in software engineering. Yeah, no, it's a great question. And the way you talked about it actually is a really great way to frame it because as we started to bring machine learning into these systems, people began to really worry about this. Like, can we trust the ML system? Well, if you knew anything about the way these perception algorithms had been implemented before, it's like it was pretty darn opaque. You've got feature vectors that you're creating and parameters you're tweaking and weird case statements that happen to work. So it's the same problem because by the time these perception systems actually worked, even if you human

coded it, it was not human readable per se. And so the way we think about this is there are some strict don't violate. We call them guardrails that, you know, our vehicle shouldn't plan to be in the same space as another vehicle. Our vehicle should not plan to leave the road. You know, there's a bunch of the stuff that we can kind of codify, and they're relatively straightforward and relatively, you can write a requirement for it, you can trace that and use kind of classic engineering approaches. And you can write a piece of software where a human being can look at it and say, like, this software enforces this guardrail. Right, and you can put testing around it and, you know, have confidence that you're implementing that. And then, and so we use, we do that, and then we take the machine-learned approach and basically stick it in the guardrails.

And so, you know, we don't kid ourselves in believing that we could really interpret all of the weights in that system, but we put rails around it so that we kind of bound the operating environment. And we believe that that will get us to a point where we can have enough confidence in the system that it really works. And theoretically, I mean, one of the things that we've seen over and over again is you have these moments with

complicated systems where you start off not being able to really understand and characterize the performance of the complex system, and so you don't really trust them. And in certain of these complex systems, they eventually evolve to the point where you both understand them, and they have superior performance to, like, everything that preceded them. And in some cases, you know, the thing that preceded them was like a human brain and like the software system is better at that narrow thing than a human brain. Like no, like I think it bears saying that no piece of software is close to being better than a complete human brain and a complete set of things that human brains are good at. But at narrow things, yes. And so, like, we may get to the point,

and, like, it would be a fantastic thing for the human race to, like, have an autonomous driving system that is better than a human from a safety perspective. And that's what our aspiration is. And the good news is that I don't even know that we need to be strictly better than the best human driver at all times. Because a lot of times when people are driving, they're not paying that much attention. They're distracted one way or the other.

Their mind wanders. They're inebriated. Right? There's an awful lot of error that happens not because we're operating at our best and couldn't handle it. It's because our attention is somewhere else. And one of the things that we can do for these vehicles is we can make them superhuman. And superhuman in kind of trivial ways. Like our vehicles see all the way around the vehicle all the time.

And now that I'm used to that in the technology, when I'm trying to make a lane change on the 101, it drives me nuts because, you know, make a lane change, I'm watching traffic in front, and then I need to look and assess the state of the traffic behind me. And, you know, what happens if, you know, particularly when it's kind of jerky flow of traffic, you know, if I'm taking a half second or three quarters of a second to look behind me and the guy in front of me has hit the brakes, I'm having a bad day. So, I find that incredibly stressful. Yeah. Well, and that's the thing already about a modern automobile. Like, forget whether or not it has autonomous technology in it or is, like, on a path to autonomous technology. From a sensing perspective, like, modern automobiles are already superhuman.

Like, they can see the world, like, just because they can see behind them at the same time they can see in front of them. Like, they, you know, like, your, the cars that you all, you know, pioneered at Google have lidars in them. And so, like, they can sort of see in dimensions that human beings can't see. It's really cool. And it's one of these things where this is why I believe we should be using a combination of laser radar and camera, not just one modality. Because, you know, like the people who believe that you should just use cameras, for example, they say, you know, the argument is effectively people just have two eyes, it's like cameras, we should just use that. And the obvious response to this is, look,

people don't have wheels for legs. But it turns out when we made the car, wheel is much better than trying to make legs work. So let's go use whatever engineering hacks we can to go make this safer. And so when we can use lasers, which allow direct 3D reconstruction of the geometry of the world, when we can use radars that allow us to see through certain obscurance and allow us to see through certain obscurance and allow us to see velocity instantaneously. When we use cameras to see the state of traffic lights and get higher resolution data, we should use it all. One of the things I'm really excited about at Aurora is

we just bought this company that builds this brand new kind of LiDAR technology. It's called Frequency Modulated Continuous Wave LiDAR. And what's LIDAR for folks who don't know? LIDAR is light detection and ranging. So think of it like radar, but instead of using a radio wave, you're using a beam of light. Yep. And for this technology, so the way

kind of classic LIDAR, so the LIDARs we used to use at Google, and most of the LIDARs that are out there that you'll see on cars, the way it works is you send out a pulse of light and you wait for it. It goes out in the world, bounces off something, comes back, and you measure how long it took for that little pulse to go out and come back. Because the speed of light is constant, you can get distance from that. Now, the trick is you have to make that pulse so bright that it is brighter than everything else out there because otherwise you don't see it when it comes back. Yep. And so that means it's got to be brighter than the sun.

Which is incredibly bright. Yes. Now, the good news is you're looking over a small period of time, so you can kind of do it, and maybe you don't necessarily care because not many cars come out of the sun. Yeah. But it makes it, you know, you basically have this really challenging signal and noise problem because you're doing DC measurements. Yeah. With this FMCW technology, what's happening is instead of sending a pulse out, you're sending out a continuous wave.

And you watch that go out, hit the world and come back. And now you interfere the outbound wave with the inbound wave and look at the phase difference. Oh, interesting. And now you measure the phase difference effectively and that tells you how far away the thing is that it's reflecting off of. And what's cool is that because of this neat physical property of self-heterodyne, you get 10 to 101 amplification, which is, you know, and because you're no longer at DC,

you're filtering out stuff that isn't at this frequency. So, you don't, like, turns out the sun is roughly constant. So, it doesn't, there's no frequency to that. So, yeah, it makes it so you can be much more sensitive. And then the other really cool thing that comes along with this is that you actually get to measure velocity instantaneously. So using Doppler. So that's the, you know, when you hear a siren go by and how it changes pitch, that's the

Doppler shift in that audio signal. So with this is we're measuring vehicles at distance with each one of those measurements that come back through that pulse or through that wave. So you can tell whether you're accelerating or decelerating relative to that other object. And not just whether we are exactly what the delta in speed is. Right. And so this, the reason why Aurora, why we bought this company, why we brought this technology in-house is it's magical, right? Because if you're trying to use a normal LIDAR to measure the world out,

let's say we want to see whether there's something 200 meters out in front of us on the road, and we want to know is that a truck or is that a pedestrian or is that a bicycle or something else, we need to get enough points on that, little dots back from it, that allow us to see the shape of it, and then we need to run a points on that little dots back from it that allow us to see the shape of it

and then we need to run a classifier on it so you need to get a whole lot of points on it because if I have two dots back from something that's a line I can't tell whether that's a person or a dump truck or a Bugatti or whatever and so that's a lot of points

it turns out to be a really hard problem. In contrast, if I hit that thing with an FMCW lighter and can measure the speed of it, well, if it moves at 60 miles per hour. It's not a human. Right. It may not be a, I don't know if it's a Bugatti or a dump truck, but I know it's a vehicle. And so I can classify it much sooner so I can respond to it much more quickly. Super interesting. Yeah, it's really cool stuff.

Yeah, and like that, I mean, the safety applications of that must be interesting. So like you can pay really close attention to stationary things. Yeah, low-speed things. You understand you have this whole additional signal that we can incorporate into the classification and thus make us more robust and make us react more quickly. Fascinating. So, a couple more questions before we go. So, one is, like, looking forward, like, what's the thing that you're most excited about in the autonomous vehicle space?

Of Aurora, obviously. Yeah, obviously. But, like, at Aurora, like, what's the interesting thing that's going to happen over the next handful of years? So, over the next handful of years, you know, we're going to get to the point where we have vehicles on the road that we can trust. And that they'll be doing, they'll almost certainly be in fleet applications and they'll be driving around. And I think that's going to be really exciting. Over the next year, there's a couple of things we're really excited about. One is, we talked earlier about Aurora building a driver,

and it's this platform. Well, because we now have this special LiDAR capability, we're now confident that we can build vehicles that can drive on the freeway. So, Aurora has spent a lot of time building passenger vehicles or passenger vehicle technology. We're now going to be moving into logistics and trucking. And so, that'll be a big thing for us over this next year. We're also going to be putting an awful lot of effort into codifying our safety process and kind of convincing ourselves, getting along the road to convincing ourselves that this vehicle really is sufficiently safe that we can trust it out in the world. And those are two really big kind of chunks for us over this next

year. And then, of course, we're going to continue on the core driving technology and moving that, you know, moving that from, you know, the capabilities we have today to more and more competent driving. Yeah, it's super awesome. So, last question. I do all sorts of things that are somewhat orthogonal to software engineering and technology. It's almost like meditation. My latest thing is I have access to a machine shop, and I'm doing a bunch of CNC machining, which is incredibly fun and exactly what I need to get my brain detached long enough from the things that I normally think about where I can go think even more clearly about the things that I

should be thinking about. So what's the interesting things that you do in your copious free time? Copious free time. Yeah. So I have a startup that I helped found. So that's almost all of it. Yeah. I think the two things that I like doing. So, one is I like rock climbing. So, I do indoor rock climbing because my wife is not excited about me climbing outdoors.

Yeah. when I get the chance to do it, you know, it's this blend of, you know, the physical exertion of climbing. You know, I'm not a small guy. And so, you know, there's a lot more work for me than some people. But then the really interesting part is when you climb it at the higher grades, it's a puzzle. And how do you load this rock so that you, you know, so you can actually hold it rather than just slip off, right? How do you contort your body up the wall? It's just I enjoy that. It's a workout and it's thoughtful. And then the other is we play games with my kids. And we're into these fantasy flight games, which there's this one called Imperial Assault where there's a team that plays the heroes, and then I get to play the Empire.

And so you have these Star Wars battles, which are a lot of fun. Sounds like a ton of fun. Yeah, that's super awesome. I mean, on the rock climbing thing, you must follow Alex Honnold a bit. So he's the climber who's like famous for free soloing the – what was it? Dom Wall for El Cap. I find him absolutely fascinating, just the mental preparation that goes into – I mean like there's the physical aspect, which I can't even conceive of physically how he does what he does. But the mental preparation that he goes through to be able to do what he does is like maybe even more fascinating than the physical side of it. Yeah, I thought it was really incredible.

Actually, one of the most impressive things to me about that movie was the aborted attempt. Yeah. That I, you know, the self-awareness and confidence and self-confidence, you know, particularly with the big film crew there and everything, and this has been the thing you've been working to, to be up on the face at some point and just say, you know what? I'm not feeling it. And, you know, and abort, right? Like, I think that was the most impressive thing to me in that whole movie, that he was able to do that. Yeah, I have a keen enough sense of vertigo that I haven't

actually watched the movie, but I have watched a bunch of interviews with him. It just sounds extraordinary. Yeah. Yeah. I think people that do that kind of thing, there's just something special about them, right? In one way or another. Yeah. I love this incredible diversity in human beings where there's so many people who are so obsessed with being really great at their particular thing. Yeah. And I just love it. I love it all. Like, it makes me so happy to, like, see this guy. And it's, like, not just him.

Like, there's so many of these rock climbers who are just extraordinary what they do and like i'm amazed at concert pianists and i'm amazed at roboticists i mean it's just yeah like what a great thing it is to be a human being it is right there's i don't know if you've watched this the youtube video people are people are awesome uh and it's it's that it's like all these crazy things, you know, people juggling while on a unicycle or doing, you know, and it's just like, yeah, like the diversity that is humanity and the fact that people, there's so many things out there and people can pick them up. And yeah, I wonder about my lack of imagination, right? That, you know, I see some of these things and I think, you know, you see it after the fact and you think, that's incredible.

And then I think, I could, it would never have occurred to me to do that. And, you know, it's a bit of a shame some days. Well, I'm sure that there are plenty of people who look at you and have the same reaction. I think building autonomous robots and trying to help self-driving cars come into the world seems like an inscrutable, almost impossible thing to some folks. But thank goodness we have people who are captivated by the idea of trying to do it and have the determination to try to make it happen. Well, thanks very much. So with that, thank you for being on the show today. Oh, my pleasure. Thanks.

Awesome. All right. So that was Kevin Scott chatting with Chris Urmson. And, you know, Kevin, as we were discussing before the interview, one of the things I think that is so interesting about Aurora and about Chris's journey is the fact that, you know, he's not, Aurora's not just building, you know, self-driving hardware, which a lot of different companies and startups are doing, but they're really looking at, you know, building an entire platform around self-driving technologies.

Yeah. And I think that's a really, I mean, one of the things that Chris said in the interview about this platform approach that really rings true to me is, like, he doesn't believe that they have sufficient imagination inside of Aurora to conceive of all of the many things that you may want to do with an AI driver. And, like, I totally agree with that. So, I think it's really presumptuous to think that one company or even one industry is going to have all of the best ideas about what to do with a really transformational technology like autonomous driving. And it makes me really happy that there's a company like Aurora out there that is building things in a platform way where hundreds or thousands or tens of thousands of companies will be able to use their technology to realize the vision that they have for what you could do with an autonomous driving agent. Yeah, no, I think you're right.

The potential there is so great. And I think that as interesting as it is and important as it is for lots of different companies to be something that people can build off of, you know, a platform, which has historically been what has made other companies, whether they're, you know, technology-based or otherwise, really successful. So, I agree with you. I think the potential there is tremendous. Yeah, the other thing, too, that's interesting that I've been thinking a lot about, and I write a little bit about this in the book that I've got that's coming out in April, is the whole thing about platform companies is that you measure your success in terms of the successes of people who are using their technology to accomplish their goals. And so, if you think about AI and its potentially negative disruptive impacts on the

world, you're much less likely to have like weird sort of re-concentrations of economics with a platform company than you are with someone who's trying to build a set of technologies that completely disrupt an entire industry. So the platform gives a whole bunch of people opportunities to build new economically viable businesses that do useful things for large numbers of people, that it allows you to have more, you know, sort of inclusive and diverse points of view and perspectives participating in the development of products. And like, I think net-net, that is a good thing for the world to like have these like really complicated platform components that are available to a huge number of people so they can go, you know, sort of express their creativity and where the economics are lined up in a way where the platform, the builders of the complicated technology only get rewarded when the people that they're empowering are successful themselves. No, totally. And I mean, at a certain point, you could even think about it like, you know, disruption will come, but it will come from those people who are building on top of things.

It's not necessarily going to be directly, you know, tied to the platform itself, which I think is probably a good way of doing things and making things more available to lots of different types of people. Yep. Awesome. Well, that's the lots of different types of people. Yep. Awesome. Well, that's the end of our show for today. And as always, we would love to hear from you. You can reach out to us anytime at BehindTheTech at Microsoft.com. Tell us what's on your mind. Tell us your New Year's resolution.

And of course, be sure to tell all your friends, your colleagues, you know, your self-driving cars, if that's what you've got, you know, your assistants, your voice assistants, your Uber drivers, whatever. Be sure to tell them all about our show. And thank you for listening. All right. See you next time.