WEBVTT 00:00.000 --> 00:02.240 The following is a conversation with Kyle Vogt. 00:02.240 --> 00:05.120 He's the president and the CTO of Cruise Automation, 00:05.120 --> 00:08.000 leading an effort to solve one of the biggest 00:08.000 --> 00:10.880 robotics challenges of our time, vehicle automation. 00:10.880 --> 00:13.120 He's a cofounder of two successful companies, 00:13.120 --> 00:17.040 Twitch and Cruise, that have each sold for a billion dollars. 00:17.040 --> 00:19.880 And he's a great example of the innovative spirit 00:19.880 --> 00:22.160 that flourishes in Silicon Valley. 00:22.160 --> 00:25.760 And now is facing an interesting and exciting challenge 00:25.760 --> 00:30.040 of matching that spirit with the mass production 00:30.040 --> 00:32.800 and the safety centered culture of a major automaker, 00:32.800 --> 00:34.440 like General Motors. 00:34.440 --> 00:36.520 This conversation is part of the MIT 00:36.520 --> 00:38.560 Artificial General Intelligence series 00:38.560 --> 00:41.040 and the Artificial Intelligence podcast. 00:41.040 --> 00:44.840 If you enjoy it, please subscribe on YouTube, iTunes, 00:44.840 --> 00:47.640 or simply connect with me on Twitter at Lex Friedman, 00:47.640 --> 00:49.800 spelled F R I D. 00:49.800 --> 00:53.480 And now here's my conversation with Kyle Vogt. 00:53.480 --> 00:55.600 You grew up in Kansas, right? 00:55.600 --> 00:58.080 Yeah, and I just saw that picture you had to hit know 00:58.080 --> 01:00.400 there, so I'm a little bit worried about that now. 01:00.400 --> 01:02.480 So in high school in Kansas City, 01:02.480 --> 01:07.200 you joined Shawnee Mission North High School Robotics Team. 01:07.200 --> 01:09.120 Now that wasn't your high school. 01:09.120 --> 01:09.960 That's right. 01:09.960 --> 01:13.920 That was the only high school in the area that had a teacher 01:13.920 --> 01:16.080 who was willing to sponsor our first robotics team. 01:16.080 --> 01:18.360 I was gonna troll you a little bit. 01:18.360 --> 01:20.320 Jog your mouth a little bit with that kid. 01:20.320 --> 01:22.880 I was trying to look super cool and intense. 01:22.880 --> 01:23.720 You did? 01:23.720 --> 01:25.680 Because this was BattleBots, this is serious business. 01:25.680 --> 01:28.840 So we're standing there with a welded steel frame 01:28.840 --> 01:30.240 and looking tough. 01:30.240 --> 01:31.800 So go back there. 01:31.800 --> 01:33.840 What does that drew you to robotics? 01:33.840 --> 01:36.480 Well, I think, I've been trying to figure this out 01:36.480 --> 01:37.920 for a while, but I've always liked building things 01:37.920 --> 01:38.760 with Legos. 01:38.760 --> 01:39.920 And when I was really, really young, 01:39.920 --> 01:42.360 I wanted the Legos that had motors and other things. 01:42.360 --> 01:44.840 And then, you know, Lego Mindstorms came out 01:44.840 --> 01:48.280 and for the first time you could program Lego contraptions. 01:48.280 --> 01:52.560 And I think things just sort of snowballed from that. 01:52.560 --> 01:56.800 But I remember seeing, you know, the BattleBots TV show 01:56.800 --> 01:59.320 on Comedy Central and thinking that is the coolest thing 01:59.320 --> 02:01.200 in the world, I wanna be a part of that. 02:01.200 --> 02:03.680 And not knowing a whole lot about how to build 02:03.680 --> 02:06.880 these 200 pound fighting robots. 02:06.880 --> 02:11.000 So I sort of obsessively poured over the internet forums 02:11.000 --> 02:13.440 where all the creators for BattleBots would sort of hang out 02:13.440 --> 02:16.120 and talk about, you know, document their build progress 02:16.120 --> 02:17.120 and everything. 02:17.120 --> 02:20.520 And I think I read, I must have read like, you know, 02:20.520 --> 02:24.280 tens of thousands of forum posts from basically everything 02:24.280 --> 02:26.400 that was out there on what these people were doing. 02:26.400 --> 02:28.920 And eventually, like sort of triangulated how to put 02:28.920 --> 02:33.040 some of these things together and ended up doing BattleBots, 02:33.040 --> 02:34.800 which was, you know, it was like 13 or 14, 02:34.800 --> 02:35.960 which was pretty awesome. 02:35.960 --> 02:37.680 I'm not sure if the show's still running, 02:37.680 --> 02:42.000 but so BattleBots is, there's not an artificial intelligence 02:42.000 --> 02:44.200 component, it's remotely controlled. 02:44.200 --> 02:46.720 And it's almost like a mechanical engineering challenge 02:46.720 --> 02:49.560 of building things that can be broken. 02:49.560 --> 02:50.680 They're radio controlled. 02:50.680 --> 02:53.880 So, and I think that they allowed some limited form 02:53.880 --> 02:56.600 of autonomy, but, you know, in a two minute match, 02:56.600 --> 02:58.800 you're, in the way these things ran, 02:58.800 --> 03:00.720 you're really doing yourself a disservice by trying 03:00.720 --> 03:02.360 to automate it versus just, you know, 03:02.360 --> 03:04.760 do the practical thing, which is drive it yourself. 03:04.760 --> 03:06.960 And there's an entertainment aspect, 03:06.960 --> 03:08.240 just going on YouTube. 03:08.240 --> 03:11.200 There's like some of them wield an axe, some of them, 03:11.200 --> 03:12.200 I mean, there's that fun. 03:12.200 --> 03:13.760 So what drew you to that aspect? 03:13.760 --> 03:15.400 Was it the mechanical engineering? 03:15.400 --> 03:19.400 Was it the dream to create like Frankenstein 03:19.400 --> 03:21.080 and sentient being? 03:21.080 --> 03:23.960 Or was it just like the Lego, you like tinkering stuff? 03:23.960 --> 03:26.000 I mean, that was just building something. 03:26.000 --> 03:27.960 I think the idea of, you know, 03:27.960 --> 03:30.920 this radio controlled machine that can do various things. 03:30.920 --> 03:33.800 If it has like a weapon or something was pretty interesting. 03:33.800 --> 03:36.440 I agree, it doesn't have the same appeal as, you know, 03:36.440 --> 03:38.520 autonomous robots, which I, which I, you know, 03:38.520 --> 03:40.320 sort of gravitated towards later on, 03:40.320 --> 03:42.720 but it was definitely an engineering challenge 03:42.720 --> 03:45.600 because everything you did in that competition 03:45.600 --> 03:48.480 was pushing components to their limits. 03:48.480 --> 03:52.960 So we would buy like these $40 DC motors 03:52.960 --> 03:54.840 that came out of a winch, 03:54.840 --> 03:57.280 like on the front of a pickup truck or something. 03:57.280 --> 03:59.240 And we'd power the car with those 03:59.240 --> 04:01.120 and we'd run them at like double or triple 04:01.120 --> 04:02.440 their rated voltage. 04:02.440 --> 04:04.160 So they immediately start overheating, 04:04.160 --> 04:06.920 but for that two minute match, you can get, you know, 04:06.920 --> 04:08.680 a significant increase in the power output 04:08.680 --> 04:10.560 of those motors before they burn out. 04:10.560 --> 04:12.760 And so you're doing the same thing for your battery packs, 04:12.760 --> 04:14.360 all the materials in the system. 04:14.360 --> 04:15.560 And I think there was something, 04:15.560 --> 04:17.800 something intrinsically interesting 04:17.800 --> 04:20.360 about just seeing like where things break. 04:20.360 --> 04:23.360 And did you offline see where they break? 04:23.360 --> 04:25.040 Did you take it to the testing point? 04:25.040 --> 04:26.120 Like, how did you know two minutes? 04:26.120 --> 04:29.680 Or was there a reckless, let's just go with it and see. 04:29.680 --> 04:31.320 We weren't very good at battle bots. 04:31.320 --> 04:34.200 We lost all of our matches the first round. 04:34.200 --> 04:36.240 The one I built first, 04:36.240 --> 04:38.120 both of them were these wedge shaped robots 04:38.120 --> 04:39.800 because the wedge, even though it's sort of boring 04:39.800 --> 04:41.240 to look at is extremely effective. 04:41.240 --> 04:42.600 You drive towards another robot 04:42.600 --> 04:44.720 and the front edge of it gets under them 04:44.720 --> 04:46.760 and then they sort of flip over, 04:46.760 --> 04:48.280 it's kind of like a door stopper. 04:48.280 --> 04:51.920 And the first one had a pneumatic polished stainless steel 04:51.920 --> 04:54.880 spike on the front that would shoot out about eight inches. 04:54.880 --> 04:56.240 The purpose of which is what? 04:56.240 --> 04:58.800 Pretty ineffective actually, but it looked cool. 04:58.800 --> 05:00.880 And was it to help with the lift? 05:00.880 --> 05:04.080 No, it was just to try to poke holes in the other robot. 05:04.080 --> 05:05.960 And then the second time I did it, 05:05.960 --> 05:09.560 which is the following, I think maybe 18 months later, 05:09.560 --> 05:14.400 we had a titanium axe with a hardened steel tip on it 05:14.400 --> 05:17.200 that was powered by a hydraulic cylinder, 05:17.200 --> 05:20.400 which we were activating with liquid CO2, 05:20.400 --> 05:23.880 which had its own set of problems. 05:23.880 --> 05:26.320 So great, so that's kind of on the hardware side. 05:26.320 --> 05:28.360 I mean, at a certain point, 05:28.360 --> 05:31.240 there must have been born a fascination 05:31.240 --> 05:32.440 on the software side. 05:32.440 --> 05:35.520 So what was the first piece of code you've written? 05:35.520 --> 05:38.600 If you didn't go back there, see what language was it? 05:38.600 --> 05:40.600 What was it, was it EMAX, VAM? 05:40.600 --> 05:44.640 Was it a more respectable, modern ID? 05:44.640 --> 05:45.800 Do you remember any of this? 05:45.800 --> 05:49.840 Yeah, well, I remember, I think maybe when I was in 05:49.840 --> 05:52.440 third or fourth grade, I was at elementary school, 05:52.440 --> 05:55.040 had a bunch of Apple II computers, 05:55.040 --> 05:56.680 and we'd play games on those. 05:56.680 --> 05:57.760 And I remember every once in a while, 05:57.760 --> 06:01.320 something would crash or wouldn't start up correctly, 06:01.320 --> 06:03.960 and it would dump you out to what I later learned 06:03.960 --> 06:05.800 was like sort of a command prompt. 06:05.800 --> 06:07.600 And my teacher would come over and type, 06:07.600 --> 06:09.440 I actually remember this to this day for some reason, 06:09.440 --> 06:12.160 like PR number six, or PR pound six, 06:12.160 --> 06:13.840 which is peripheral six, which is the disk drive, 06:13.840 --> 06:15.920 which would fire up the disk and load the program. 06:15.920 --> 06:17.880 And I just remember thinking, wow, she's like a hacker, 06:17.880 --> 06:20.760 like teach me these codes, these error codes, 06:20.760 --> 06:22.720 that is what I called them at the time. 06:22.720 --> 06:23.760 But she had no interest in that. 06:23.760 --> 06:26.480 So it wasn't until I think about fifth grade 06:26.480 --> 06:29.120 that I had a school where you could actually 06:29.120 --> 06:30.600 go on these Apple II's and learn to program. 06:30.600 --> 06:31.920 And so it was all in basic, you know, 06:31.920 --> 06:34.240 where every line, you know, the line numbers are all, 06:34.240 --> 06:35.640 or that every line is numbered, 06:35.640 --> 06:38.000 and you have to like leave enough space 06:38.000 --> 06:40.760 between the numbers so that if you want to tweak your code, 06:40.760 --> 06:42.600 you go back and if the first line was 10 06:42.600 --> 06:44.680 and the second line is 20, now you have to go back 06:44.680 --> 06:45.640 and insert 15. 06:45.640 --> 06:47.960 And if you need to add code in front of that, 06:47.960 --> 06:49.720 you know, 11 or 12, and you hope you don't run out 06:49.720 --> 06:51.880 of line numbers and have to redo the whole thing. 06:51.880 --> 06:53.240 And there's go to statements? 06:53.240 --> 06:56.920 Yeah, go to and is very basic, maybe hence the name, 06:56.920 --> 06:58.200 but a lot of fun. 06:58.200 --> 07:00.800 And that was like, that was, you know, 07:00.800 --> 07:02.600 that's when, you know, when you first program, 07:02.600 --> 07:03.560 you see the magic of it. 07:03.560 --> 07:06.640 It's like, just like this world opens up with, 07:06.640 --> 07:08.200 you know, endless possibilities for the things 07:08.200 --> 07:10.600 you could build or accomplish with that computer. 07:10.600 --> 07:13.400 So you got the bug then, so even starting with basic 07:13.400 --> 07:16.720 and then what, C++ throughout, what did you, 07:16.720 --> 07:18.200 was there a computer programming, 07:18.200 --> 07:19.880 computer science classes in high school? 07:19.880 --> 07:22.680 Not, not where I went, so it was self taught, 07:22.680 --> 07:24.640 but I did a lot of programming. 07:24.640 --> 07:28.560 The thing that, you know, sort of pushed me in the path 07:28.560 --> 07:30.600 of eventually working on self driving cars 07:30.600 --> 07:33.280 is actually one of these really long trips 07:33.280 --> 07:38.000 driving from my house in Kansas to, I think, Las Vegas, 07:38.000 --> 07:39.480 where we did the BattleBots competition. 07:39.480 --> 07:42.760 And I had just gotten my, I think my learners permit 07:42.760 --> 07:45.080 or early drivers permit. 07:45.080 --> 07:48.280 And so I was driving this, you know, 10 hour stretch 07:48.280 --> 07:50.600 across Western Kansas where it's just, 07:50.600 --> 07:51.800 you're going straight on a highway 07:51.800 --> 07:53.640 and it is mind numbingly boring. 07:53.640 --> 07:54.960 And I remember thinking even then 07:54.960 --> 07:58.080 with my sort of mediocre programming background 07:58.080 --> 08:00.040 that this is something that a computer can do, right? 08:00.040 --> 08:01.440 Let's take a picture of the road, 08:01.440 --> 08:02.880 let's find the yellow lane markers 08:02.880 --> 08:04.880 and, you know, steer the wheel. 08:04.880 --> 08:06.600 And, you know, later I'd come to realize 08:06.600 --> 08:09.800 this had been done, you know, since the 80s 08:09.800 --> 08:12.760 or the 70s or even earlier, but I still wanted to do it. 08:12.760 --> 08:14.840 And sort of immediately after that trip, 08:14.840 --> 08:16.280 switched from sort of BattleBots, 08:16.280 --> 08:18.640 which is more radio controlled machines 08:18.640 --> 08:21.800 to thinking about building, you know, 08:21.800 --> 08:23.600 autonomous vehicles of some scale, 08:23.600 --> 08:25.080 start off with really small electric ones 08:25.080 --> 08:28.280 and then, you know, progress to what we're doing now. 08:28.280 --> 08:30.040 So what was your view of artificial intelligence 08:30.040 --> 08:30.880 at that point? 08:30.880 --> 08:31.880 What did you think? 08:31.880 --> 08:35.040 So this is before there's been waves 08:35.040 --> 08:36.680 in artificial intelligence, right? 08:36.680 --> 08:39.480 The current wave with deep learning 08:39.480 --> 08:41.760 makes people believe that you can solve 08:41.760 --> 08:43.520 in a really rich, deep way, 08:43.520 --> 08:46.200 the computer vision perception problem. 08:46.200 --> 08:51.200 But like before the deep learning craze, 08:51.320 --> 08:52.800 you know, how do you think about 08:52.800 --> 08:55.320 how would you even go about building a thing 08:55.320 --> 08:56.920 that perceives itself in the world, 08:56.920 --> 08:59.160 localize itself in the world, moves around the world? 08:59.160 --> 09:00.360 Like when you were younger, I mean, 09:00.360 --> 09:02.120 as what was your thinking about it? 09:02.120 --> 09:03.960 Well, prior to deep neural networks 09:03.960 --> 09:05.360 or convolutional neural nets, 09:05.360 --> 09:06.520 these modern techniques we have, 09:06.520 --> 09:09.040 or at least ones that are in use today, 09:09.040 --> 09:10.280 it was all heuristic space. 09:10.280 --> 09:12.920 And so like old school image processing, 09:12.920 --> 09:15.040 and I think extracting, you know, 09:15.040 --> 09:18.000 yellow lane markers out of an image of a road 09:18.000 --> 09:21.160 is one of the problems that lends itself 09:21.160 --> 09:23.760 reasonably well to those heuristic base methods, you know, 09:23.760 --> 09:26.760 like just do a threshold on the color yellow 09:26.760 --> 09:28.520 and then try to fit some lines to that 09:28.520 --> 09:30.320 using a huff transform or something 09:30.320 --> 09:32.280 and then go from there. 09:32.280 --> 09:34.800 Traffic light detection and stop sign detection, 09:34.800 --> 09:35.920 red, yellow, green. 09:35.920 --> 09:38.160 And I think you can, you could, 09:38.160 --> 09:39.840 I mean, if you wanted to do a full, 09:39.840 --> 09:41.960 I was just trying to make something that would stay 09:41.960 --> 09:43.520 in between the lanes on a highway, 09:43.520 --> 09:44.960 but if you wanted to do the full, 09:46.920 --> 09:48.960 the full, you know, set of capabilities 09:48.960 --> 09:50.520 needed for a driverless car, 09:50.520 --> 09:53.360 I think you could, and we've done this at cruise, 09:53.360 --> 09:54.440 you know, in the very first days, 09:54.440 --> 09:56.320 you can start off with a really simple, 09:56.320 --> 09:58.000 you know, human written heuristic 09:58.000 --> 09:59.800 just to get the scaffolding in place 09:59.800 --> 10:01.720 for your system, traffic light detection, 10:01.720 --> 10:02.960 probably a really simple, you know, 10:02.960 --> 10:04.760 color thresholding on day one 10:04.760 --> 10:06.520 just to get the system up and running 10:06.520 --> 10:08.640 before you migrate to, you know, 10:08.640 --> 10:11.080 a deep learning based technique or something else. 10:11.080 --> 10:12.800 And, you know, back in, when I was doing this, 10:12.800 --> 10:15.120 my first one, it was on a Pentium 203, 10:15.120 --> 10:17.840 233 megahertz computer in it. 10:17.840 --> 10:19.920 And I think I wrote the first version in basic, 10:19.920 --> 10:21.600 which is like an interpreted language. 10:21.600 --> 10:23.760 It's extremely slow because that's the thing 10:23.760 --> 10:24.800 I knew at the time. 10:24.800 --> 10:27.840 And so there was no, no chance at all of using, 10:27.840 --> 10:30.440 there's no computational power to do 10:30.440 --> 10:33.480 any sort of reasonable deep nets like you have today. 10:33.480 --> 10:35.360 So I don't know what kids these days are doing. 10:35.360 --> 10:37.920 Are kids these days, you know, at age 13 10:37.920 --> 10:39.360 using neural networks in their garage? 10:39.360 --> 10:40.200 I mean, that would be awesome. 10:40.200 --> 10:43.040 I get emails all the time from, you know, 10:43.040 --> 10:46.160 like 11, 12 year olds saying, I'm having, you know, 10:46.160 --> 10:48.760 I'm trying to follow this TensorFlow tutorial 10:48.760 --> 10:50.800 and I'm having this problem. 10:50.800 --> 10:55.800 And the general approach in the deep learning community 10:55.800 --> 11:00.200 is of extreme optimism of, as opposed to, 11:00.200 --> 11:02.000 you mentioned like heuristics, you can, 11:02.000 --> 11:04.800 you can, you can separate the autonomous driving problem 11:04.800 --> 11:07.520 into modules and try to solve it sort of rigorously, 11:07.520 --> 11:09.040 where you can just do it end to end. 11:09.040 --> 11:11.840 And most people just kind of love the idea that, 11:11.840 --> 11:13.360 you know, us humans do it end to end, 11:13.360 --> 11:15.360 we just perceive and act. 11:15.360 --> 11:17.040 We should be able to use that, 11:17.040 --> 11:18.720 do the same kind of thing with your own nets. 11:18.720 --> 11:20.920 And that, that kind of thinking, 11:20.920 --> 11:22.840 you don't want to criticize that kind of thinking 11:22.840 --> 11:24.640 because eventually they will be right. 11:24.640 --> 11:26.360 Yeah. And so it's exciting. 11:26.360 --> 11:28.720 And especially when they're younger to explore that 11:28.720 --> 11:30.640 is a really exciting approach. 11:30.640 --> 11:35.480 But yeah, it's, it's changed the, the language, 11:35.480 --> 11:37.240 the kind of stuff you're tinkering with. 11:37.240 --> 11:40.920 It's kind of exciting to see when these teenagers grow up. 11:40.920 --> 11:43.760 Yeah, I can only imagine if you, if your starting point 11:43.760 --> 11:46.720 is, you know, Python and TensorFlow at age 13, 11:46.720 --> 11:47.800 where you end up, you know, 11:47.800 --> 11:51.040 after 10 or 15 years of that, that's, that's pretty cool. 11:51.040 --> 11:53.760 Because of GitHub, because the state tools 11:53.760 --> 11:55.440 for solving most of the major problems 11:55.440 --> 11:56.920 that are artificial intelligence 11:56.920 --> 12:00.240 are within a few lines of code for most kids. 12:00.240 --> 12:02.280 And that's incredible to think about, 12:02.280 --> 12:04.280 also on the entrepreneurial side. 12:04.280 --> 12:08.520 And, and, and at that point, was there any thought 12:08.520 --> 12:11.960 about entrepreneurship before you came to college 12:11.960 --> 12:15.160 is sort of doing your building this into a thing 12:15.160 --> 12:17.800 that impacts the world on a large scale? 12:17.800 --> 12:19.840 Yeah, I've always wanted to start a company. 12:19.840 --> 12:22.600 I think that's, you know, just a cool concept 12:22.600 --> 12:25.240 of creating something and exchanging it 12:25.240 --> 12:28.360 for value or creating value, I guess. 12:28.360 --> 12:31.120 So in high school, I was, I was trying to build like, 12:31.120 --> 12:33.600 you know, servo motor drivers, little circuit boards 12:33.600 --> 12:36.920 and sell them online or other, other things like that. 12:36.920 --> 12:40.320 And certainly knew at some point I wanted to do a startup, 12:40.320 --> 12:42.840 but it wasn't really, I'd say until college until I felt 12:42.840 --> 12:46.720 like I had the, I guess the right combination 12:46.720 --> 12:48.960 of the environment, the smart people around you 12:48.960 --> 12:52.360 and some free time and a lot of free time at MIT. 12:52.360 --> 12:55.800 So you came to MIT as an undergrad 2004. 12:55.800 --> 12:57.080 That's right. 12:57.080 --> 12:59.040 And that's when the first DARPA Grand Challenge 12:59.040 --> 12:59.880 was happening. 12:59.880 --> 13:00.720 Yeah. 13:00.720 --> 13:03.360 The timing of that is beautifully poetic. 13:03.360 --> 13:05.680 So how'd you get yourself involved in that one? 13:05.680 --> 13:07.080 Originally there wasn't a 13:07.080 --> 13:07.920 Official entry? 13:07.920 --> 13:09.520 Yeah, faculty sponsored thing. 13:09.520 --> 13:12.760 And so a bunch of undergrads, myself included, 13:12.760 --> 13:14.160 started meeting and got together 13:14.160 --> 13:17.800 and tried to, to haggle together some sponsorships. 13:17.800 --> 13:20.120 We got a vehicle donated, a bunch of sensors 13:20.120 --> 13:21.600 and tried to put something together. 13:21.600 --> 13:24.640 And so we had, our team was probably mostly freshmen 13:24.640 --> 13:26.800 and sophomores, you know, which, which was not really 13:26.800 --> 13:30.960 a fair, fair fight against maybe the, you know, postdoc 13:30.960 --> 13:32.840 and faculty led teams from other schools. 13:32.840 --> 13:35.000 But we, we got something up and running. 13:35.000 --> 13:37.400 We had our vehicle drive by wire and, you know, 13:37.400 --> 13:42.440 very, very basic control and things, but on the day 13:42.440 --> 13:46.800 of the qualifying, sort of pre qualifying round, 13:46.800 --> 13:50.840 the one and only steering motor that we had purchased, 13:50.840 --> 13:52.600 the thing that we had, you know, retrofitted to turn 13:52.600 --> 13:55.760 the steering wheel on the truck died. 13:55.760 --> 13:58.440 And so our vehicle was just dead in the water, couldn't steer. 13:58.440 --> 13:59.880 So we didn't make it very far. 13:59.880 --> 14:00.920 On the hardware side. 14:00.920 --> 14:03.000 So was there a software component? 14:03.000 --> 14:06.200 Was there, like, how did your view of autonomous vehicles 14:06.200 --> 14:08.040 in terms of artificial intelligence 14:09.520 --> 14:10.720 evolve in this moment? 14:10.720 --> 14:12.400 I mean, you know, like you said, 14:12.400 --> 14:14.080 from the 80s has been autonomous vehicles, 14:14.080 --> 14:16.720 but really that was the birth of the modern wave. 14:16.720 --> 14:20.080 The, the thing that captivated everyone's imagination 14:20.080 --> 14:21.520 that we can actually do this. 14:21.520 --> 14:26.000 So how, were you captivated in that way? 14:26.000 --> 14:27.600 So how did your view of autonomous vehicles 14:27.600 --> 14:29.000 change at that point? 14:29.000 --> 14:33.760 I'd say at that point in time, it was, it was a curiosity 14:33.760 --> 14:35.840 as in like, is this really possible? 14:35.840 --> 14:38.440 And I think that was generally the spirit 14:38.440 --> 14:43.280 and the purpose of that original DARPA Grand Challenge, 14:43.280 --> 14:45.520 which was to just get a whole bunch 14:45.520 --> 14:48.680 of really brilliant people exploring the space 14:48.680 --> 14:49.880 and pushing the limits. 14:49.880 --> 14:51.960 And, and I think like to this day, 14:51.960 --> 14:54.160 that DARPA challenge with its, you know, 14:54.160 --> 14:57.120 million dollar prize pool was probably one 14:57.120 --> 15:00.840 of the most effective, you know, uses of taxpayer money, 15:00.840 --> 15:03.320 dollar for dollar that I've seen, you know, 15:03.320 --> 15:06.720 because that, that small sort of initiative 15:06.720 --> 15:10.440 that DARPA put put out sort of, in my view, 15:10.440 --> 15:12.560 was the catalyst or the tipping point 15:12.560 --> 15:14.800 for this, this whole next wave 15:14.800 --> 15:16.120 of autonomous vehicle development. 15:16.120 --> 15:17.160 So that was pretty cool. 15:17.160 --> 15:20.240 So let me jump around a little bit on that point. 15:20.240 --> 15:23.240 They also did the urban challenge where it was in the city, 15:23.240 --> 15:25.920 but it was very artificial and there's no pedestrians 15:25.920 --> 15:27.640 and there's very little human involvement 15:27.640 --> 15:30.480 except a few professional drivers. 15:30.480 --> 15:31.640 Yeah. 15:31.640 --> 15:33.560 Do you think there's room, and then there was 15:33.560 --> 15:35.360 the robotics challenge with human robots? 15:35.360 --> 15:36.200 Right. 15:36.200 --> 15:38.720 So in your now role as looking at this, 15:38.720 --> 15:41.640 you're trying to solve one of the, you know, 15:41.640 --> 15:43.120 autonomous driving, one of the harder, 15:43.120 --> 15:45.480 more difficult places in San Francisco. 15:45.480 --> 15:47.320 Is there a role for DARPA to step in 15:47.320 --> 15:49.680 to also kind of help out, like, 15:49.680 --> 15:54.000 challenge with new ideas, specifically pedestrians 15:54.000 --> 15:55.880 and so on, all these kinds of interesting things? 15:55.880 --> 15:57.680 Well, I haven't thought about it from that perspective. 15:57.680 --> 15:59.280 Is there anything DARPA could do today 15:59.280 --> 16:00.680 to further accelerate things? 16:00.680 --> 16:04.880 And I would say my instinct is that that's maybe not 16:04.880 --> 16:07.040 the highest and best use of their resources in time 16:07.040 --> 16:10.640 because, like, kick starting and spinning up the flywheel 16:10.640 --> 16:12.720 is I think what they did in this case 16:12.720 --> 16:14.200 for very, very little money. 16:14.200 --> 16:15.800 But today this has become, 16:16.880 --> 16:19.000 this has become, like, commercially interesting 16:19.000 --> 16:20.680 to very large companies and the amount of money 16:20.680 --> 16:23.040 going into it and the amount of people, 16:23.040 --> 16:24.840 like, going through your class and learning 16:24.840 --> 16:27.200 about these things and developing these skills 16:27.200 --> 16:29.120 is just, you know, orders of magnitude 16:29.120 --> 16:30.840 more than it was back then. 16:30.840 --> 16:33.080 And so there's enough momentum and inertia 16:33.080 --> 16:36.520 and energy and investment dollars into this space right now 16:36.520 --> 16:39.960 that I don't, I don't, I think they're, 16:39.960 --> 16:42.200 I think they're, they can just say mission accomplished 16:42.200 --> 16:44.320 and move on to the next area of technology 16:44.320 --> 16:45.360 that needs help. 16:46.280 --> 16:49.120 So then stepping back to MIT, 16:49.120 --> 16:50.880 you left MIT Junior Junior year, 16:50.880 --> 16:53.080 what was that decision like? 16:53.080 --> 16:55.680 As I said, I always wanted to do a company 16:55.680 --> 16:59.080 or start a company and this opportunity landed in my lap 16:59.080 --> 17:01.960 which was a couple of guys from Yale 17:01.960 --> 17:04.240 were starting a new company and I Googled them 17:04.240 --> 17:06.720 and found that they had started a company previously 17:06.720 --> 17:10.640 and sold it actually on eBay for about a quarter million bucks 17:10.640 --> 17:12.880 which was a pretty interesting story. 17:12.880 --> 17:15.760 But so I thought to myself, these guys are, you know, 17:15.760 --> 17:19.080 rock star entrepreneurs, they've done this before, 17:19.080 --> 17:20.720 they must be driving around in Ferraris 17:20.720 --> 17:22.320 because they sold their company. 17:23.320 --> 17:26.000 And, you know, I thought I could learn a lot from them. 17:26.000 --> 17:28.320 So I teamed up with those guys and, you know, 17:28.320 --> 17:32.000 went out during, went out to California during IAP 17:32.000 --> 17:36.440 which is MIT's month off on one way ticket 17:36.440 --> 17:38.040 and basically never went back. 17:38.040 --> 17:39.280 We were having so much fun, 17:39.280 --> 17:42.040 we felt like we were building something and creating something 17:42.040 --> 17:44.440 and it was gonna be interesting that, you know, 17:44.440 --> 17:46.800 I was just all in and got completely hooked 17:46.800 --> 17:49.640 and that business was Justin TV 17:49.640 --> 17:52.400 which is originally a reality show about a guy named Justin 17:53.720 --> 17:57.120 which morphed into a live video streaming platform 17:57.120 --> 18:00.320 which then morphed into what is Twitch today. 18:00.320 --> 18:03.720 So that was quite an unexpected journey. 18:04.720 --> 18:07.000 So no regrets? 18:07.000 --> 18:07.840 No. 18:07.840 --> 18:09.120 Looking back, it was just an obvious, 18:09.120 --> 18:10.720 I mean, one way ticket. 18:10.720 --> 18:12.760 I mean, if we just pause on that for a second, 18:12.760 --> 18:17.680 there was no, how did you know these were the right guys? 18:17.680 --> 18:19.520 This is the right decision. 18:19.520 --> 18:22.640 You didn't think it was just follow the heart kind of thing? 18:22.640 --> 18:24.520 Well, I didn't know, but, you know, 18:24.520 --> 18:26.520 just trying something for a month during IAP 18:26.520 --> 18:28.240 seems pretty low risk, right? 18:28.240 --> 18:30.760 And then, you know, well, maybe I'll take a semester off. 18:30.760 --> 18:32.280 MIT's pretty flexible about that. 18:32.280 --> 18:33.840 You can always go back, right? 18:33.840 --> 18:35.680 And then after two or three cycles of that, 18:35.680 --> 18:36.960 I eventually threw in the towel. 18:36.960 --> 18:41.920 But, you know, I think it's, I guess in that case, 18:41.920 --> 18:44.880 I felt like I could always hit the undo button if I had to. 18:44.880 --> 18:45.720 Right. 18:45.720 --> 18:49.600 But nevertheless, from when you look in retrospect, 18:49.600 --> 18:51.680 I mean, it seems like a brave decision. 18:51.680 --> 18:53.200 You know, it would be difficult 18:53.200 --> 18:54.320 for a lot of people to make. 18:54.320 --> 18:55.440 It wasn't as popular. 18:55.440 --> 18:58.120 I'd say that the general, you know, 18:58.120 --> 19:01.480 flux of people out of MIT at the time was mostly 19:01.480 --> 19:04.120 into, you know, finance or consulting jobs 19:04.120 --> 19:05.720 in Boston or New York. 19:05.720 --> 19:07.840 And very few people were going to California 19:07.840 --> 19:09.080 to start companies. 19:09.080 --> 19:12.240 But today, I'd say that's probably inverted, 19:12.240 --> 19:15.400 which is just a sign of a sign of the times, I guess. 19:15.400 --> 19:16.080 Yeah. 19:16.080 --> 19:21.560 So there's a story about midnight of March 18, 2007, 19:21.560 --> 19:25.720 where TechCrunch, I guess, announced Justin TV earlier 19:25.720 --> 19:29.080 than it was supposed to a few hours. 19:29.080 --> 19:30.360 The site didn't work. 19:30.360 --> 19:32.520 I don't know if any of this is true, you can tell me. 19:32.520 --> 19:36.200 And you and one of the folks at Justin TV, 19:36.200 --> 19:39.240 Emma Shear, coded through the night. 19:39.240 --> 19:41.440 Can you take me through that experience? 19:41.440 --> 19:47.160 So let me say a few nice things that the article I read quoted 19:47.160 --> 19:49.600 Justin Khan said that you were known for bureau coding 19:49.600 --> 19:53.520 through problems and being a creative genius. 19:53.520 --> 19:59.440 So on that night, what was going through your head? 19:59.440 --> 20:01.400 Or maybe I put another way, how do you 20:01.400 --> 20:02.520 solve these problems? 20:02.520 --> 20:05.480 What's your approach to solving these kind of problems 20:05.480 --> 20:07.080 where the line between success and failure 20:07.080 --> 20:09.680 seems to be pretty thin? 20:09.680 --> 20:10.680 That's a good question. 20:10.680 --> 20:13.400 Well, first of all, that's nice of Justin to say that. 20:13.400 --> 20:16.880 I think I would have been maybe 21 years old then 20:16.880 --> 20:18.800 and not very experienced at programming. 20:18.800 --> 20:22.680 But as with everything in a startup, 20:22.680 --> 20:24.720 you're sort of racing against the clock. 20:24.720 --> 20:27.320 And so our plan was the second we 20:27.320 --> 20:32.600 had this live streaming camera backpack up and running 20:32.600 --> 20:33.600 where Justin could wear it. 20:33.600 --> 20:35.320 And no matter where he went in the city, 20:35.320 --> 20:36.400 it would be streaming live video. 20:36.400 --> 20:37.960 And this is even before the iPhones, 20:37.960 --> 20:40.880 this is like hard to do back then. 20:40.880 --> 20:41.800 We would launch. 20:41.800 --> 20:45.160 And so we thought we were there and the backpack was working. 20:45.160 --> 20:47.080 And then we sent out all the emails 20:47.080 --> 20:49.920 to launch the company and do the press thing. 20:49.920 --> 20:53.000 And then we weren't quite actually there. 20:53.000 --> 20:55.880 And then we thought, oh, well, they're 20:55.880 --> 21:00.160 not going to announce it until maybe 10 AM the next morning. 21:00.160 --> 21:01.880 And it's, I don't know, it's 5 PM now. 21:01.880 --> 21:03.640 So how many hours do we have left? 21:03.640 --> 21:08.000 What is that, like 17 hours to go? 21:08.000 --> 21:10.440 And that was going to be fine. 21:10.440 --> 21:11.440 Was the problem obvious? 21:11.440 --> 21:13.280 Did you understand what could possibly be? 21:13.280 --> 21:16.520 Like how complicated was the system at that point? 21:16.520 --> 21:18.840 It was pretty messy. 21:18.840 --> 21:22.760 So to get a live video feed that looked decent working 21:22.760 --> 21:25.680 from anywhere in San Francisco, I 21:25.680 --> 21:28.600 put together this system where we had like three or four 21:28.600 --> 21:29.880 cell phone data modems. 21:29.880 --> 21:32.200 And they were like, we take the video stream 21:32.200 --> 21:35.600 and sort of spray it across these three or four modems 21:35.600 --> 21:38.080 and then try to catch all the packets on the other side 21:38.080 --> 21:39.480 with unreliable cell phone networks. 21:39.480 --> 21:41.080 Pretty low level networking. 21:41.080 --> 21:41.720 Yeah. 21:41.720 --> 21:44.760 And putting these sort of protocols 21:44.760 --> 21:47.560 on top of all that to reassemble and reorder the packets 21:47.560 --> 21:49.720 and have time buffers and error correction 21:49.720 --> 21:50.960 and all that kind of stuff. 21:50.960 --> 21:53.960 And the night before, it was just 21:53.960 --> 21:56.280 staticky. Every once in a while, the image would go 21:56.280 --> 21:59.640 staticky and there would be this horrible like screeching 21:59.640 --> 22:02.080 audio noise because the audio was also corrupted. 22:02.080 --> 22:04.600 And this would happen like every five to 10 minutes or so. 22:04.600 --> 22:08.080 And it was a really, you know, off of putting to the viewers. 22:08.080 --> 22:08.880 Yeah. 22:08.880 --> 22:10.200 How do you tackle that problem? 22:10.200 --> 22:13.280 What was the, you're just freaking out behind a computer. 22:13.280 --> 22:16.880 There's the word, are there other folks working on this problem? 22:16.880 --> 22:18.120 Like were you behind a whiteboard? 22:18.120 --> 22:22.000 Were you doing a hair coding? 22:22.000 --> 22:23.840 Yeah, it's a little lonely because there's four of us 22:23.840 --> 22:26.880 working on the company and only two people really wrote code. 22:26.880 --> 22:29.200 And Emmett wrote the website in the chat system 22:29.200 --> 22:32.400 and I wrote the software for this video streaming device 22:32.400 --> 22:34.280 and video server. 22:34.280 --> 22:36.240 And so, you know, it was my sole responsibility 22:36.240 --> 22:37.320 to figure that out. 22:37.320 --> 22:39.440 And I think it's those, you know, 22:39.440 --> 22:42.200 setting deadlines, trying to move quickly and everything 22:42.200 --> 22:44.200 where you're in that moment of intense pressure 22:44.200 --> 22:46.960 that sometimes people do their best and most interesting work. 22:46.960 --> 22:48.800 And so even though that was a terrible moment, 22:48.800 --> 22:50.760 I look back on it fondly because that's like, you know, 22:50.760 --> 22:54.720 that's one of those character defining moments, I think. 22:54.720 --> 22:59.480 So in 2013, October, you founded Cruise Automation. 22:59.480 --> 23:00.200 Yeah. 23:00.200 --> 23:04.200 So progressing forward, another exceptionally successful 23:04.200 --> 23:09.920 company was acquired by GM in 2016 for $1 billion. 23:09.920 --> 23:14.120 But in October 2013, what was on your mind? 23:14.120 --> 23:16.360 What was the plan? 23:16.360 --> 23:19.840 How does one seriously start to tackle 23:19.840 --> 23:22.800 one of the hardest robotics, most important impact 23:22.800 --> 23:24.960 for robotics problems of our age? 23:24.960 --> 23:28.760 After going through Twitch, Twitch was, 23:28.760 --> 23:31.480 and is today pretty successful. 23:31.480 --> 23:36.880 But the work was, the result was entertainment mostly. 23:36.880 --> 23:39.840 Like the better the product was, the more we would entertain 23:39.840 --> 23:42.760 people and then, you know, make money on the ad revenues 23:42.760 --> 23:43.760 and other things. 23:43.760 --> 23:45.000 And that was a good thing. 23:45.000 --> 23:46.320 It felt good to entertain people. 23:46.320 --> 23:49.120 But I figured like, you know, what is really the point 23:49.120 --> 23:51.120 of becoming a really good engineer 23:51.120 --> 23:53.160 and developing these skills other than, you know, 23:53.160 --> 23:53.960 my own enjoyment. 23:53.960 --> 23:55.760 And I realized I wanted something that scratched 23:55.760 --> 23:57.680 more of an existential itch, like something 23:57.680 --> 23:59.440 that truly matters. 23:59.440 --> 24:03.680 And so I basically made this list of requirements 24:03.680 --> 24:06.160 for a new, if I was going to do another company. 24:06.160 --> 24:08.000 And the one thing I knew in the back of my head 24:08.000 --> 24:12.320 that Twitch took like eight years to become successful. 24:12.320 --> 24:14.880 And so whatever I do, I better be willing to commit, 24:14.880 --> 24:17.000 you know, at least 10 years to something. 24:17.000 --> 24:20.400 And when you think about things from that perspective, 24:20.400 --> 24:21.760 you certainly, I think, raise the bar 24:21.760 --> 24:23.200 on what you choose to work on. 24:23.200 --> 24:24.320 So for me, the three things where 24:24.320 --> 24:27.120 it had to be something where the technology itself 24:27.120 --> 24:28.960 determines the success of the product, 24:28.960 --> 24:31.840 like hard, really juicy technology problems, 24:31.840 --> 24:33.600 because that's what motivates me. 24:33.600 --> 24:36.280 And then it had to have a direct and positive impact 24:36.280 --> 24:37.640 on society in some way. 24:37.640 --> 24:39.200 So an example would be like, you know, 24:39.200 --> 24:41.560 health care, self driving cars because they save lives, 24:41.560 --> 24:43.600 other things where there's a clear connection to somehow 24:43.600 --> 24:45.200 improving other people's lives. 24:45.200 --> 24:47.160 And the last one is it had to be a big business 24:47.160 --> 24:50.200 because for the positive impact to matter, 24:50.200 --> 24:51.240 it's got to be a large scale. 24:51.240 --> 24:52.080 Scale, yeah. 24:52.080 --> 24:53.840 And I was thinking about that for a while 24:53.840 --> 24:55.960 and I made like a, I tried writing a Gmail clone 24:55.960 --> 24:57.640 and looked at some other ideas. 24:57.640 --> 24:59.480 And then it just sort of light bulb went off 24:59.480 --> 25:00.440 like self driving cars. 25:00.440 --> 25:02.360 Like that was the most fun I had ever had 25:02.360 --> 25:04.040 in college working on that. 25:04.040 --> 25:05.960 And like, well, what's the state of the technology 25:05.960 --> 25:08.440 has been 10 years, maybe times have changed 25:08.440 --> 25:10.800 and maybe now is the time to make this work. 25:10.800 --> 25:13.320 And I poked around and looked at the only other thing 25:13.320 --> 25:15.480 out there really at the time was the Google self driving 25:15.480 --> 25:16.680 car project. 25:16.680 --> 25:19.600 And I thought surely there's a way to, you know, 25:19.600 --> 25:21.600 have an entrepreneur mindset and sort of solve 25:21.600 --> 25:23.520 the minimum viable product here. 25:23.520 --> 25:25.200 And so I just took the plunge right then and there 25:25.200 --> 25:26.680 and said, this, this is something I know 25:26.680 --> 25:27.840 I can commit 10 years to. 25:27.840 --> 25:30.760 It's probably the greatest applied AI problem 25:30.760 --> 25:32.000 of our generation. 25:32.000 --> 25:34.240 And if it works, it's going to be both a huge business 25:34.240 --> 25:37.040 and therefore like probably the most positive impact 25:37.040 --> 25:38.280 I can possibly have on the world. 25:38.280 --> 25:40.920 So after that light bulb went off, 25:40.920 --> 25:43.000 I went all in on cruise immediately 25:43.000 --> 25:45.560 and got to work. 25:45.560 --> 25:47.360 Did you have an idea how to solve this problem? 25:47.360 --> 25:49.640 Which aspect of the problem to solve? 25:49.640 --> 25:53.720 You know, slow, like we just had Oliver from voyage here 25:53.720 --> 25:56.560 slow moving retirement communities, 25:56.560 --> 25:58.080 urban driving, highway driving. 25:58.080 --> 26:00.400 Did you have like, did you have a vision 26:00.400 --> 26:03.560 of the city of the future or, you know, 26:03.560 --> 26:06.400 the transportation is largely automated, 26:06.400 --> 26:07.240 that kind of thing. 26:07.240 --> 26:12.240 Or was it sort of more fuzzy and gray area than that? 26:12.240 --> 26:16.640 My analysis of the situation is that Google's putting a lot, 26:16.640 --> 26:19.200 had been putting a lot of money into that project. 26:19.200 --> 26:20.760 They had a lot more resources. 26:20.760 --> 26:23.720 And so, and they still hadn't cracked 26:23.720 --> 26:26.200 the fully driverless car. 26:26.200 --> 26:28.520 You know, this is 2013, I guess. 26:29.480 --> 26:33.360 So I thought, what can I do to sort of go from zero 26:33.360 --> 26:35.600 to, you know, significant scale 26:35.600 --> 26:37.280 so I can actually solve the real problem, 26:37.280 --> 26:38.640 which is the driverless cars. 26:38.640 --> 26:40.480 And I thought, here's the strategy. 26:40.480 --> 26:44.080 We'll start by doing a really simple problem 26:44.080 --> 26:45.560 or solving a really simple problem 26:45.560 --> 26:48.080 that creates value for people. 26:48.080 --> 26:50.040 So it eventually ended up deciding 26:50.040 --> 26:51.800 on automating highway driving, 26:51.800 --> 26:54.240 which is relatively more straightforward 26:54.240 --> 26:56.440 as long as there's a backup driver there. 26:56.440 --> 26:58.480 And, you know, the go to market 26:58.480 --> 27:00.240 will be able to retrofit people's cars 27:00.240 --> 27:02.240 and just sell these products directly. 27:02.240 --> 27:04.520 And the idea was, we'll take all the revenue 27:04.520 --> 27:08.320 and profits from that and use it to do the, 27:08.320 --> 27:10.920 to sort of reinvest that in research for doing 27:10.920 --> 27:12.600 fully driverless cars. 27:12.600 --> 27:13.960 And that was the plan. 27:13.960 --> 27:15.720 The only thing that really changed along the way 27:15.720 --> 27:17.360 between then and now is, 27:17.360 --> 27:19.000 we never really launched the first product. 27:19.000 --> 27:21.680 We had enough interest from investors 27:21.680 --> 27:24.120 and enough of a signal that this was something 27:24.120 --> 27:25.000 that we should be working on, 27:25.000 --> 27:28.400 that after about a year of working on the highway autopilot, 27:28.400 --> 27:31.040 we had it working, you know, at a prototype stage, 27:31.040 --> 27:33.120 but we just completely abandoned that 27:33.120 --> 27:34.960 and said, we're gonna go all in on driverless cars 27:34.960 --> 27:36.480 now is the time. 27:36.480 --> 27:38.120 Can't think of anything that's more exciting. 27:38.120 --> 27:39.720 And if it works more impactful, 27:39.720 --> 27:41.360 so we're just gonna go for it. 27:41.360 --> 27:43.440 The idea of retrofit is kind of interesting. 27:43.440 --> 27:44.280 Yeah. 27:44.280 --> 27:46.880 Being able to, it's how you achieve scale. 27:46.880 --> 27:47.880 It's a really interesting idea, 27:47.880 --> 27:51.120 is it's something that's still in the back of your mind 27:51.120 --> 27:52.800 as a possibility? 27:52.800 --> 27:53.640 Not at all. 27:53.640 --> 27:57.080 I've come full circle on that one after trying 27:57.080 --> 27:58.880 to build a retrofit product. 27:58.880 --> 28:01.240 And I'll touch on some of the complexities of that. 28:01.240 --> 28:04.240 And then also having been inside an OEM 28:04.240 --> 28:05.400 and seeing how things work 28:05.400 --> 28:08.320 and how a vehicle is developed and validated. 28:08.320 --> 28:09.360 When it comes to something 28:09.360 --> 28:11.280 that has safety critical implications, 28:11.280 --> 28:12.520 like controlling the steering 28:12.520 --> 28:15.280 and other control inputs on your car, 28:15.280 --> 28:17.720 it's pretty hard to get there with a retrofit. 28:17.720 --> 28:20.520 Or if you did, even if you did, 28:20.520 --> 28:23.280 it creates a whole bunch of new complications around 28:23.280 --> 28:25.400 liability or how did you truly validate that? 28:25.400 --> 28:27.480 Or, you know, something in the base vehicle fails 28:27.480 --> 28:29.880 and causes your system to fail, whose fault is it? 28:31.560 --> 28:34.080 Or if the car's anti lock brake systems 28:34.080 --> 28:36.680 or other things kick in or the software has been, 28:36.680 --> 28:38.240 it's different in one version of the car. 28:38.240 --> 28:40.080 You retrofit versus another and you don't know 28:40.080 --> 28:43.000 because the manufacturer has updated it behind the scenes. 28:43.000 --> 28:45.400 There's basically an infinite list of long tail issues 28:45.400 --> 28:46.240 that can get you. 28:46.240 --> 28:47.760 And if you're dealing with a safety critical product, 28:47.760 --> 28:48.960 that's not really acceptable. 28:48.960 --> 28:52.160 That's a really convincing summary of why 28:52.160 --> 28:53.160 it's really challenging. 28:53.160 --> 28:54.360 But I didn't know all that at the time. 28:54.360 --> 28:55.480 So we tried it anyway. 28:55.480 --> 28:57.160 But as a pitch also at the time, 28:57.160 --> 28:58.400 it's a really strong one. 28:58.400 --> 29:00.720 That's how you achieve scale and that's how you beat 29:00.720 --> 29:03.360 the current, the leader at the time of Google 29:03.360 --> 29:04.720 or the only one in the market. 29:04.720 --> 29:06.840 The other big problem we ran into, 29:06.840 --> 29:08.240 which is perhaps the biggest problem 29:08.240 --> 29:10.280 from a business model perspective, 29:10.280 --> 29:15.280 is we had kind of assumed that we started with an Audi S4 29:15.440 --> 29:16.880 as the vehicle we retrofitted 29:16.880 --> 29:18.760 with this highway driving capability. 29:18.760 --> 29:21.040 And we had kind of assumed that if we just knock out 29:21.040 --> 29:23.360 like three make and models of vehicle, 29:23.360 --> 29:25.880 that'll cover like 80% of the San Francisco market. 29:25.880 --> 29:27.400 Doesn't everyone there drive, I don't know, 29:27.400 --> 29:30.240 a BMW or a Honda Civic or one of these three cars? 29:30.240 --> 29:32.040 And then we surveyed our users and we found out 29:32.040 --> 29:33.480 that it's all over the place. 29:33.480 --> 29:36.680 We would, to get even a decent number of units sold, 29:36.680 --> 29:39.880 we'd have to support like 20 or 50 different models. 29:39.880 --> 29:42.200 And each one is a little butterfly that takes time 29:42.200 --> 29:44.800 and effort to maintain that retrofit integration 29:44.800 --> 29:47.120 and custom hardware and all this. 29:47.120 --> 29:49.240 So it was a tough business. 29:49.240 --> 29:54.240 So GM manufactures and sells over nine million cars a year. 29:54.280 --> 29:58.560 And what you with crews are trying to do 29:58.560 --> 30:01.160 some of the most cutting edge innovation 30:01.160 --> 30:03.000 in terms of applying AI. 30:03.000 --> 30:06.040 And so how do those, you've talked about it a little bit 30:06.040 --> 30:07.760 before, but it's also just fascinating to me, 30:07.760 --> 30:09.360 we work a lot of automakers. 30:10.560 --> 30:12.880 The difference between the gap between Detroit 30:12.880 --> 30:14.680 and Silicon Valley, let's say, 30:14.680 --> 30:17.320 just to be sort of poetic about it, I guess. 30:17.320 --> 30:18.680 How do you close that gap? 30:18.680 --> 30:21.480 How do you take GM into the future 30:21.480 --> 30:24.840 where a large part of the fleet would be autonomous perhaps? 30:24.840 --> 30:28.520 I wanna start by acknowledging that GM is made up of 30:28.520 --> 30:30.240 tens of thousands of really brilliant, 30:30.240 --> 30:32.720 motivated people who wanna be a part of the future. 30:32.720 --> 30:35.240 And so it's pretty fun to work with them. 30:35.240 --> 30:37.480 The attitude inside a car company like that 30:37.480 --> 30:41.240 is embracing this transformation and change 30:41.240 --> 30:42.360 rather than fearing it. 30:42.360 --> 30:45.440 And I think that's a testament to the leadership at GM 30:45.440 --> 30:47.680 and that's flown all the way through to everyone 30:47.680 --> 30:49.280 you talk to, even the people in the assembly plants 30:49.280 --> 30:51.200 working on these cars. 30:51.200 --> 30:52.040 So that's really great. 30:52.040 --> 30:55.160 So starting from that position makes it a lot easier. 30:55.160 --> 30:59.160 So then when the people in San Francisco 30:59.160 --> 31:01.400 but cruise interact with the people at GM, 31:01.400 --> 31:02.960 at least we have this common set of values, 31:02.960 --> 31:05.000 which is that we really want this stuff to work 31:05.000 --> 31:06.040 because we think it's important 31:06.040 --> 31:07.440 and we think it's the future. 31:08.360 --> 31:11.520 That's not to say those two cultures don't clash. 31:11.520 --> 31:12.440 They absolutely do. 31:12.440 --> 31:14.760 There's different sort of value systems. 31:14.760 --> 31:17.960 Like in a car company, the thing that gets you promoted 31:17.960 --> 31:22.600 and sort of the reward system is following the processes, 31:22.600 --> 31:26.080 delivering the program on time and on budget. 31:26.080 --> 31:30.440 So any sort of risk taking is discouraged in many ways 31:30.440 --> 31:34.000 because if a program is late 31:34.000 --> 31:36.200 or if you shut down the plant for a day, 31:36.200 --> 31:37.560 you can count the millions of dollars 31:37.560 --> 31:39.600 that burn by pretty quickly. 31:39.600 --> 31:43.800 Whereas I think most Silicon Valley companies 31:43.800 --> 31:48.280 and in cruise and the methodology we were employing, 31:48.280 --> 31:50.080 especially around the time of the acquisition, 31:50.080 --> 31:53.800 the reward structure is about trying to solve 31:53.800 --> 31:56.120 these complex problems in any way, shape or form 31:56.120 --> 31:59.640 or coming up with crazy ideas that 90% of them won't work. 31:59.640 --> 32:02.920 And so meshing that culture 32:02.920 --> 32:05.480 of sort of continuous improvement and experimentation 32:05.480 --> 32:07.400 with one where everything needs to be 32:07.400 --> 32:08.480 rigorously defined up front 32:08.480 --> 32:12.760 so that you never slip a deadline or miss a budget 32:12.760 --> 32:13.600 was a pretty big challenge 32:13.600 --> 32:16.960 and that we're over three years in now 32:16.960 --> 32:18.360 after the acquisition. 32:18.360 --> 32:20.480 And I'd say like the investment we made 32:20.480 --> 32:23.600 in figuring out how to work together successfully 32:23.600 --> 32:24.440 and who should do what 32:24.440 --> 32:26.360 and how we bridge the gaps 32:26.360 --> 32:27.680 between these very different systems 32:27.680 --> 32:29.520 and way of doing engineering work 32:29.520 --> 32:30.920 is now one of our greatest assets 32:30.920 --> 32:32.320 because I think we have this really powerful thing 32:32.320 --> 32:35.560 but for a while it was both GM and cruise 32:35.560 --> 32:37.440 were very steep on the learning curve. 32:37.440 --> 32:38.920 Yeah, so I'm sure it was very stressful. 32:38.920 --> 32:39.960 It's really important work 32:39.960 --> 32:43.680 because that's how to revolutionize the transportation. 32:43.680 --> 32:46.640 Really to revolutionize any system, 32:46.640 --> 32:48.200 you look at the healthcare system 32:48.200 --> 32:49.680 or you look at the legal system. 32:49.680 --> 32:52.040 I have people like Laura's come up to me all the time 32:52.040 --> 32:53.920 like everything they're working on 32:53.920 --> 32:55.960 can easily be automated. 32:55.960 --> 32:57.480 But then that's not a good feeling. 32:57.480 --> 32:58.320 Yeah. 32:58.320 --> 32:59.160 Well, it's not a good feeling, 32:59.160 --> 33:01.200 but also there's no way to automate 33:01.200 --> 33:06.200 because the entire infrastructure is really based 33:06.360 --> 33:08.360 is older and it moves very slowly. 33:08.360 --> 33:11.560 And so how do you close the gap between? 33:11.560 --> 33:13.880 I haven't, how can I replace? 33:13.880 --> 33:15.720 Of course, Laura's the one be replaced with an app 33:15.720 --> 33:17.920 but you could replace a lot of aspect 33:17.920 --> 33:20.160 when most of the data is still on paper. 33:20.160 --> 33:23.400 And so the same thing with automotive. 33:23.400 --> 33:26.080 I mean, it's fundamentally software. 33:26.080 --> 33:28.560 So it's basically hiring software engineers. 33:28.560 --> 33:30.320 It's thinking of software world. 33:30.320 --> 33:32.560 I mean, I'm pretty sure nobody in Silicon Valley 33:32.560 --> 33:34.640 has ever hit a deadline. 33:34.640 --> 33:36.000 So and then on GM. 33:36.000 --> 33:37.400 That's probably true, yeah. 33:37.400 --> 33:39.920 And GM side is probably the opposite. 33:39.920 --> 33:42.720 So that's that culture gap is really fascinating. 33:42.720 --> 33:45.160 So you're optimistic about the future of that. 33:45.160 --> 33:47.440 Yeah, I mean, from what I've seen, it's impressive. 33:47.440 --> 33:49.400 And I think like, especially in Silicon Valley, 33:49.400 --> 33:51.440 it's easy to write off building cars 33:51.440 --> 33:53.120 because people have been doing that 33:53.120 --> 33:54.960 for over a hundred years now in this country. 33:54.960 --> 33:57.080 And so it seems like that's a solved problem, 33:57.080 --> 33:58.840 but that doesn't mean it's an easy problem. 33:58.840 --> 34:02.280 And I think it would be easy to sort of overlook that 34:02.280 --> 34:06.080 and think that we're Silicon Valley engineers, 34:06.080 --> 34:08.960 we can solve any problem, building a car, 34:08.960 --> 34:13.200 it's been done, therefore it's not a real engineering 34:13.200 --> 34:14.600 challenge. 34:14.600 --> 34:17.480 But after having seen just the sheer scale 34:17.480 --> 34:21.360 and magnitude and industrialization that occurs 34:21.360 --> 34:23.280 inside of an automotive assembly plant, 34:23.280 --> 34:25.840 that is a lot of work that I am very glad 34:25.840 --> 34:28.200 that we don't have to reinvent 34:28.200 --> 34:29.480 to make self driving cars work. 34:29.480 --> 34:31.680 And so to have partners who have done that for a hundred 34:31.680 --> 34:32.960 years and have these great processes 34:32.960 --> 34:35.720 and this huge infrastructure and supply base 34:35.720 --> 34:38.760 that we can tap into is just remarkable 34:38.760 --> 34:43.760 because the scope and surface area of the problem 34:44.560 --> 34:47.400 of deploying fleets of self driving cars is so large 34:47.400 --> 34:50.320 that we're constantly looking for ways to do less 34:50.320 --> 34:52.920 so we can focus on the things that really matter more. 34:52.920 --> 34:55.360 And if we had to figure out how to build and assemble 34:55.360 --> 35:00.120 and test and build the cars themselves, 35:00.120 --> 35:01.640 I mean, we work closely with GM on that, 35:01.640 --> 35:03.240 but if we had to develop all that capability 35:03.240 --> 35:08.240 in house as well, that would just make the problem 35:08.320 --> 35:10.200 really intractable, I think. 35:10.200 --> 35:14.880 So yeah, just like your first entry at the MIT DARPA 35:14.880 --> 35:17.680 challenge when it was what the motor that failed 35:17.680 --> 35:19.000 and somebody that knows what they're doing 35:19.000 --> 35:20.040 with the motor did it. 35:20.040 --> 35:22.080 It would have been nice if we could focus on the software 35:22.080 --> 35:23.880 and not the hardware platform. 35:23.880 --> 35:24.800 Yeah, right. 35:24.800 --> 35:27.080 So from your perspective now, 35:28.080 --> 35:29.960 there's so many ways that autonomous vehicles 35:29.960 --> 35:34.280 can impact society in the next year, five years, 10 years. 35:34.280 --> 35:37.080 What do you think is the biggest opportunity 35:37.080 --> 35:39.360 to make money in autonomous driving, 35:40.560 --> 35:44.720 sort of make it a financially viable thing in the near term? 35:44.720 --> 35:49.120 What do you think would be the biggest impact there? 35:49.120 --> 35:52.160 Well, the things that drive the economics 35:52.160 --> 35:53.600 for fleets of self driving cars 35:53.600 --> 35:56.440 are there's sort of a handful of variables. 35:56.440 --> 36:00.400 One is the cost to build the vehicle itself. 36:00.400 --> 36:03.720 So the material cost, what's the cost of all your sensors, 36:03.720 --> 36:05.200 plus the cost of the vehicle 36:05.200 --> 36:07.560 and all the other components on it. 36:07.560 --> 36:09.520 Another one is the lifetime of the vehicle. 36:09.520 --> 36:12.480 It's very different if your vehicle drives 100,000 miles 36:12.480 --> 36:14.800 and then it falls apart versus 2 million. 36:16.720 --> 36:18.840 And then if you have a fleet, 36:18.840 --> 36:22.920 it's kind of like an airplane or an airline 36:22.920 --> 36:26.120 where once you produce the vehicle, 36:26.120 --> 36:27.880 you want it to be in operation 36:27.880 --> 36:30.760 as many hours a day as possible producing revenue. 36:30.760 --> 36:32.480 And then the other piece of that 36:32.480 --> 36:35.280 is how are you generating revenue? 36:35.280 --> 36:36.880 I think that's kind of what you're asking in. 36:36.880 --> 36:38.400 I think the obvious things today 36:38.400 --> 36:40.080 are the ride sharing business 36:40.080 --> 36:42.760 because that's pretty clear that there's demand for that. 36:42.760 --> 36:46.240 There's existing markets you can tap into and... 36:46.240 --> 36:47.960 Large urban areas, that kind of thing. 36:47.960 --> 36:48.800 Yeah, yeah. 36:48.800 --> 36:51.200 And I think that there are some real benefits 36:51.200 --> 36:54.520 to having cars without drivers 36:54.520 --> 36:56.040 compared to sort of the status quo 36:56.040 --> 36:58.520 for people who use ride share services today. 36:58.520 --> 37:01.040 You know, your privacy, consistency, 37:01.040 --> 37:02.440 hopefully significantly improve safety, 37:02.440 --> 37:05.120 all these benefits versus the current product. 37:05.120 --> 37:06.520 But it's a crowded market. 37:06.520 --> 37:08.000 And then other opportunities 37:08.000 --> 37:09.600 which you've seen a lot of activity in the last, 37:09.600 --> 37:12.560 really in the last six or 12 months is delivery, 37:12.560 --> 37:17.560 whether that's parcels and packages, food or groceries. 37:17.800 --> 37:20.320 Those are all sort of, I think, opportunities 37:20.320 --> 37:23.640 that are pretty ripe for these. 37:23.640 --> 37:26.000 Once you have this core technology, 37:26.000 --> 37:28.080 which is the fleet of autonomous vehicles, 37:28.080 --> 37:30.920 there's all sorts of different business opportunities 37:30.920 --> 37:32.080 you can build on top of that. 37:32.080 --> 37:34.520 But I think the important thing, of course, 37:34.520 --> 37:36.440 is that there's zero monetization opportunity 37:36.440 --> 37:37.520 until you actually have that fleet 37:37.520 --> 37:39.160 of very capable driverless cars 37:39.160 --> 37:41.040 that are as good or better than humans. 37:41.040 --> 37:44.120 And that's sort of where the entire industry 37:44.120 --> 37:45.920 is sort of in this holding pattern right now. 37:45.920 --> 37:47.960 Yeah, they're trying to achieve that baseline. 37:47.960 --> 37:51.520 But you said sort of not reliability consistency. 37:51.520 --> 37:52.360 It's kind of interesting. 37:52.360 --> 37:54.200 I think I heard you say somewhere, 37:54.200 --> 37:55.440 not sure if that's what you meant, 37:55.440 --> 37:58.240 but I can imagine a situation 37:58.240 --> 38:01.200 where you would get an autonomous vehicle. 38:01.200 --> 38:04.560 And when you get into an Uber or Lyft, 38:04.560 --> 38:05.960 you don't get to choose the driver 38:05.960 --> 38:07.320 in a sense that you don't get to choose 38:07.320 --> 38:09.080 the personality of the driving. 38:09.080 --> 38:12.040 Do you think there's room 38:12.040 --> 38:14.120 to define the personality of the car 38:14.120 --> 38:15.040 the way it drives you, 38:15.040 --> 38:17.600 in terms of aggressiveness, for example, 38:17.600 --> 38:21.120 in terms of sort of pushing the boundaries. 38:21.120 --> 38:22.760 One of the biggest challenges in autonomous driving 38:22.760 --> 38:27.760 is the trade off between sort of safety and assertiveness. 38:28.600 --> 38:30.920 And do you think there's any room 38:30.920 --> 38:35.920 for the human to take a role in that decision? 38:36.040 --> 38:38.080 Sort of accept some of the liability, I guess. 38:38.080 --> 38:41.000 I wouldn't say, no, I'd say within reasonable bounds, 38:41.000 --> 38:42.280 as in we're not gonna, 38:43.200 --> 38:44.360 I think it'd be higher than likely 38:44.360 --> 38:46.600 we'd expose any knob that would let you 38:46.600 --> 38:50.240 significantly increase safety risk. 38:50.240 --> 38:53.080 I think that's just not something we'd be willing to do. 38:53.080 --> 38:56.760 But I think driving style or like, 38:56.760 --> 38:59.120 are you gonna relax the comfort constraints slightly 38:59.120 --> 39:00.160 or things like that? 39:00.160 --> 39:02.400 All of those things make sense and are plausible. 39:02.400 --> 39:04.480 I see all those as nice optimizations. 39:04.480 --> 39:06.760 Once again, we get the core problem solved 39:06.760 --> 39:08.120 in these fleets out there. 39:08.120 --> 39:10.440 But the other thing we've sort of observed 39:10.440 --> 39:12.560 is that you have this intuition 39:12.560 --> 39:15.400 that if you sort of slam your foot on the gas 39:15.400 --> 39:16.680 right after the light turns green 39:16.680 --> 39:18.160 and aggressively accelerate, 39:18.160 --> 39:19.720 you're gonna get there faster. 39:19.720 --> 39:22.080 But the actual impact of doing that is pretty small. 39:22.080 --> 39:23.680 You feel like you're getting there faster, 39:23.680 --> 39:26.680 but so the same would be true for AVs. 39:26.680 --> 39:29.640 Even if they don't slam the pedal to the floor 39:29.640 --> 39:31.000 when the light turns green, 39:31.000 --> 39:32.520 they're gonna get you there within, 39:32.520 --> 39:33.600 if it's a 15 minute trip, 39:33.600 --> 39:36.400 within 30 seconds of what you would have done otherwise 39:36.400 --> 39:37.800 if you were going really aggressively. 39:37.800 --> 39:40.760 So I think there's this sort of self deception 39:40.760 --> 39:44.440 that my aggressive driving style is getting me there faster. 39:44.440 --> 39:46.640 Well, so that's, you know, some of the things I study, 39:46.640 --> 39:48.760 some of the things I'm fascinated by the psychology of that. 39:48.760 --> 39:50.640 And I don't think it matters 39:50.640 --> 39:52.240 that it doesn't get you there faster. 39:52.240 --> 39:55.520 It's the emotional release. 39:55.520 --> 39:59.080 Driving is a place, being inside our car, 39:59.080 --> 40:00.880 somebody said it's like the real world version 40:00.880 --> 40:02.920 of being a troll. 40:02.920 --> 40:04.960 So you have this protection, this mental protection, 40:04.960 --> 40:06.640 and you're able to sort of yell at the world, 40:06.640 --> 40:08.200 like release your anger, whatever it is. 40:08.200 --> 40:10.040 But so there's an element of that 40:10.040 --> 40:12.000 that I think autonomous vehicles 40:12.000 --> 40:15.400 would also have to, you know, giving an outlet to people, 40:15.400 --> 40:19.120 but it doesn't have to be through driving or honking 40:19.120 --> 40:21.200 or so on, there might be other outlets. 40:21.200 --> 40:24.040 But I think to just sort of even just put that aside, 40:24.040 --> 40:26.880 the baseline is really, you know, that's the focus, 40:26.880 --> 40:28.200 that's the thing you need to solve, 40:28.200 --> 40:31.000 and then the fun human things can be solved after. 40:31.000 --> 40:34.680 But so from the baseline of just solving autonomous driving, 40:34.680 --> 40:36.000 you're working in San Francisco, 40:36.000 --> 40:38.960 one of the more difficult cities to operate in, 40:38.960 --> 40:42.080 what is the, in your view currently, 40:42.080 --> 40:45.040 the hardest aspect of autonomous driving? 40:46.880 --> 40:49.200 Negotiating with pedestrians, 40:49.200 --> 40:51.400 is it edge cases of perception? 40:51.400 --> 40:52.760 Is it planning? 40:52.760 --> 40:54.520 Is there a mechanical engineering? 40:54.520 --> 40:57.040 Is it data, fleet stuff? 40:57.040 --> 41:01.200 What are your thoughts on the more challenging aspects there? 41:01.200 --> 41:02.240 That's a good question. 41:02.240 --> 41:03.520 I think before we go to that though, 41:03.520 --> 41:05.080 I just want to, I like what you said 41:05.080 --> 41:07.600 about the psychology aspect of this, 41:07.600 --> 41:09.680 because I think one observation I've made is, 41:09.680 --> 41:11.760 I think I read somewhere that I think it's, 41:11.760 --> 41:13.880 maybe Americans on average spend, you know, 41:13.880 --> 41:16.520 over an hour a day on social media, 41:16.520 --> 41:18.280 like staring at Facebook. 41:18.280 --> 41:20.080 And so that's just, you know, 41:20.080 --> 41:21.600 60 minutes of your life, you're not getting back. 41:21.600 --> 41:23.120 It's probably not super productive. 41:23.120 --> 41:26.200 And so that's 3,600 seconds, right? 41:26.200 --> 41:29.160 And that's, that's time, you know, 41:29.160 --> 41:30.600 it's a lot of time you're giving up. 41:30.600 --> 41:34.080 And if you compare that to people being on the road, 41:34.080 --> 41:35.360 if another vehicle, 41:35.360 --> 41:37.600 whether it's a human driver or autonomous vehicle, 41:37.600 --> 41:39.840 delays them by even three seconds, 41:39.840 --> 41:41.920 they're laying in on the horn, you know, 41:41.920 --> 41:43.280 even though that's, that's, you know, 41:43.280 --> 41:45.280 one 1,000th of the time they waste 41:45.280 --> 41:46.360 looking at Facebook every day. 41:46.360 --> 41:48.640 So there's, there's definitely some, 41:48.640 --> 41:50.040 you know, psychology aspects of this, 41:50.040 --> 41:50.880 I think that are pretty interesting. 41:50.880 --> 41:51.720 Road rage in general. 41:51.720 --> 41:52.960 And then the question, of course, 41:52.960 --> 41:54.960 is if everyone is in self driving cars, 41:54.960 --> 41:57.560 do they even notice these three second delays anymore? 41:57.560 --> 41:58.920 Because they're doing other things 41:58.920 --> 42:01.720 or reading or working or just talking to each other. 42:01.720 --> 42:03.200 So it'll be interesting to see where that goes. 42:03.200 --> 42:05.120 In a certain aspect, people, 42:05.120 --> 42:06.360 people need to be distracted 42:06.360 --> 42:07.360 by something entertaining, 42:07.360 --> 42:09.160 something useful inside the car 42:09.160 --> 42:10.960 so they don't pay attention to the external world. 42:10.960 --> 42:14.240 And then, and then they can take whatever psychology 42:14.240 --> 42:17.400 and bring it back to Twitter and then focus on that 42:17.400 --> 42:19.640 as opposed to sort of interacting, 42:20.920 --> 42:23.200 sort of putting the emotion out there into the world. 42:23.200 --> 42:24.560 So it's an interesting problem, 42:24.560 --> 42:26.960 but baseline autonomy. 42:26.960 --> 42:28.760 I guess you could say self driving cars, 42:28.760 --> 42:31.680 you know, at scale will lower the collective blood pressure 42:31.680 --> 42:33.920 of society probably by a couple of points 42:33.920 --> 42:35.760 without all that road rage and stress. 42:35.760 --> 42:37.480 So that's a good, good externality. 42:38.560 --> 42:41.760 So back to your question about the technology 42:41.760 --> 42:43.760 and the, I guess the biggest problems. 42:43.760 --> 42:45.560 And I have a hard time answering that question 42:45.560 --> 42:48.680 because, you know, we've been at this, 42:48.680 --> 42:51.440 like specifically focusing on driverless cars 42:51.440 --> 42:53.520 and all the technology needed to enable that 42:53.520 --> 42:55.160 for a little over four and a half years now. 42:55.160 --> 42:58.080 And even a year or two in, 42:58.080 --> 43:02.960 I felt like we had completed the functionality needed 43:02.960 --> 43:04.800 to get someone from point A to point B. 43:04.800 --> 43:07.280 As in, if we need to do a left turn maneuver 43:07.280 --> 43:08.960 or if we need to drive around a, you know, 43:08.960 --> 43:11.800 a double parked vehicle into oncoming traffic 43:11.800 --> 43:13.840 or navigate through construction zones, 43:13.840 --> 43:15.960 the scaffolding and the building blocks 43:15.960 --> 43:17.800 was there pretty early on. 43:17.800 --> 43:22.360 And so the challenge is not any one scenario or situation 43:22.360 --> 43:25.520 for which, you know, we fail at 100% of those. 43:25.520 --> 43:28.960 It's more, you know, we're benchmarking against a pretty good 43:28.960 --> 43:31.320 or pretty high standard, which is human driving. 43:31.320 --> 43:33.320 All things considered, humans are excellent 43:33.320 --> 43:36.240 at handling edge cases and unexpected scenarios 43:36.240 --> 43:38.400 where it's computers are the opposite. 43:38.400 --> 43:43.080 And so beating that baseline set by humans is the challenge. 43:43.080 --> 43:46.520 And so what we've been doing for quite some time now 43:46.520 --> 43:50.760 is basically it's this continuous improvement process 43:50.760 --> 43:55.000 where we find sort of the most, you know, uncomfortable 43:55.000 --> 43:59.840 or the things that could lead to a safety issue 43:59.840 --> 44:00.960 or other things, all these events. 44:00.960 --> 44:02.520 And then we sort of categorize them 44:02.520 --> 44:04.560 and rework parts of our system 44:04.560 --> 44:06.200 to make incremental improvements 44:06.200 --> 44:08.040 and do that over and over and over again. 44:08.040 --> 44:10.160 And we just see sort of the overall performance 44:10.160 --> 44:12.120 of the system, you know, 44:12.120 --> 44:13.960 actually increasing in a pretty steady clip. 44:13.960 --> 44:15.360 But there's no one thing. 44:15.360 --> 44:17.360 There's actually like thousands of little things 44:17.360 --> 44:19.880 and just like polishing functionality 44:19.880 --> 44:21.640 and making sure that it handles, you know, 44:21.640 --> 44:26.120 every version and possible permutation of a situation 44:26.120 --> 44:29.200 by either applying more deep learning systems 44:30.120 --> 44:32.960 or just by, you know, adding more test coverage 44:32.960 --> 44:35.760 or new scenarios that we develop against 44:35.760 --> 44:37.160 and just grinding on that. 44:37.160 --> 44:40.120 We're sort of in the unsexy phase of development right now 44:40.120 --> 44:41.800 which is doing the real engineering work 44:41.800 --> 44:44.120 that it takes to go from prototype to production. 44:44.120 --> 44:46.960 You're basically scaling the grinding. 44:46.960 --> 44:50.560 So sort of taking seriously the process 44:50.560 --> 44:54.040 of all those edge cases, both with human experts 44:54.040 --> 44:57.520 and machine learning methods to cover, 44:57.520 --> 44:59.320 to cover all those situations. 44:59.320 --> 45:00.760 Yeah, and the exciting thing for me is 45:00.760 --> 45:03.000 I don't think that grinding ever stops 45:03.000 --> 45:04.840 because there's a moment in time 45:04.840 --> 45:08.760 where you've crossed that threshold of human performance 45:08.760 --> 45:10.000 and become superhuman. 45:11.200 --> 45:13.560 But there's no reason, there's no first principles reason 45:13.560 --> 45:17.560 that AV capability will tap out anywhere near humans. 45:17.560 --> 45:20.280 Like there's no reason it couldn't be 20 times better 45:20.280 --> 45:22.120 whether that's, you know, just better driving 45:22.120 --> 45:24.240 or safer driving or more comfortable driving 45:24.240 --> 45:26.800 or even a thousand times better given enough time. 45:26.800 --> 45:31.480 And we intend to basically chase that, you know, forever 45:31.480 --> 45:32.840 to build the best possible product. 45:32.840 --> 45:33.960 Better and better and better 45:33.960 --> 45:36.400 and always new edge cases come up and new experiences. 45:36.400 --> 45:39.520 So, and you want to automate that process 45:39.520 --> 45:40.720 as much as possible. 45:42.680 --> 45:45.160 So what do you think in general in society 45:45.160 --> 45:48.200 when do you think we may have hundreds of thousands 45:48.200 --> 45:50.200 of fully autonomous vehicles driving around? 45:50.200 --> 45:53.560 So first of all, predictions, nobody knows the future. 45:53.560 --> 45:55.360 You're a part of the leading people 45:55.360 --> 45:56.560 trying to define that future, 45:56.560 --> 45:58.560 but even then you still don't know. 45:58.560 --> 46:02.240 But if you think about hundreds of thousands of vehicles, 46:02.240 --> 46:05.840 so a significant fraction of vehicles 46:05.840 --> 46:07.600 in major cities are autonomous. 46:07.600 --> 46:10.800 Do you think, are you with Rodney Brooks 46:10.800 --> 46:13.960 who is 2050 and beyond? 46:13.960 --> 46:17.200 Or are you more with Elon Musk 46:17.200 --> 46:20.600 who is, we should have had that two years ago? 46:20.600 --> 46:23.840 Well, I mean, I'd love to have it two years ago, 46:23.840 --> 46:26.120 but we're not there yet. 46:26.120 --> 46:28.480 So I guess the way I would think about that 46:28.480 --> 46:31.240 is let's flip that question around. 46:31.240 --> 46:34.200 So what would prevent you to reach hundreds 46:34.200 --> 46:36.320 of thousands of vehicles and... 46:36.320 --> 46:38.200 That's a good rephrasing. 46:38.200 --> 46:43.200 Yeah, so the, I'd say that it seems the consensus 46:43.200 --> 46:45.200 among the people developing self driving cars today 46:45.200 --> 46:49.200 is to sort of start with some form of an easier environment, 46:49.200 --> 46:52.200 whether it means lacking, inclement weather, 46:52.200 --> 46:55.200 or mostly sunny or whatever it is. 46:55.200 --> 46:59.200 And then add capability for more complex situations 46:59.200 --> 47:00.200 over time. 47:00.200 --> 47:05.200 And so if you're only able to deploy in areas 47:05.200 --> 47:07.200 that meet sort of your criteria 47:07.200 --> 47:09.200 or that the current don't meet, 47:09.200 --> 47:13.200 operating domain of the software you developed, 47:13.200 --> 47:16.200 that may put a cap on how many cities you could deploy in. 47:16.200 --> 47:19.200 But then as those restrictions start to fall away, 47:19.200 --> 47:22.200 like maybe you add capability to drive really well 47:22.200 --> 47:25.200 and safely and have you rain or snow, 47:25.200 --> 47:28.200 that probably opens up the market by two or three fold 47:28.200 --> 47:31.200 in terms of the cities you can expand into and so on. 47:31.200 --> 47:33.200 And so the real question is, 47:33.200 --> 47:35.200 I know today if we wanted to, 47:35.200 --> 47:39.200 we could produce that many autonomous vehicles, 47:39.200 --> 47:41.200 but we wouldn't be able to make use of all of them yet 47:41.200 --> 47:44.200 because we would sort of saturate the demand in the cities 47:44.200 --> 47:47.200 in which we would want to operate initially. 47:47.200 --> 47:49.200 So if I were to guess what the timeline is 47:49.200 --> 47:51.200 for those things falling away 47:51.200 --> 47:54.200 and reaching hundreds, thousands of vehicles. 47:54.200 --> 47:55.200 Maybe a range is better. 47:55.200 --> 47:57.200 I would say less than five years. 47:57.200 --> 47:58.200 Less than five years. 47:58.200 --> 47:59.200 Yeah. 47:59.200 --> 48:02.200 And of course you're working hard to make that happen. 48:02.200 --> 48:05.200 So you started two companies that were eventually acquired 48:05.200 --> 48:08.200 for each $4 billion. 48:08.200 --> 48:10.200 So you're a pretty good person to ask, 48:10.200 --> 48:13.200 what does it take to build a successful startup? 48:13.200 --> 48:18.200 I think there's sort of survivor bias here a little bit, 48:18.200 --> 48:20.200 but I can try to find some common threads 48:20.200 --> 48:22.200 for the things that worked for me, which is... 48:24.200 --> 48:26.200 In both of these companies, 48:26.200 --> 48:28.200 I was really passionate about the core technology. 48:28.200 --> 48:31.200 I actually lay awake at night thinking about these problems 48:31.200 --> 48:33.200 and how to solve them. 48:33.200 --> 48:35.200 And I think that's helpful because when you start a business, 48:35.200 --> 48:37.200 there are... 48:37.200 --> 48:40.200 To this day, there are these crazy ups and downs. 48:40.200 --> 48:43.200 One day you think the business is just on top of the world 48:43.200 --> 48:45.200 and unstoppable and the next day you think, 48:45.200 --> 48:47.200 okay, this is all going to end. 48:47.200 --> 48:50.200 It's just going south and it's going to be over tomorrow. 48:52.200 --> 48:55.200 And so I think having a true passion that you can fall back on 48:55.200 --> 48:57.200 and knowing that you would be doing it 48:57.200 --> 48:58.200 even if you weren't getting paid for it 48:58.200 --> 49:00.200 helps you weather those tough times. 49:00.200 --> 49:02.200 So that's one thing. 49:02.200 --> 49:05.200 I think the other one is really good people. 49:05.200 --> 49:07.200 So I've always been surrounded by really good cofounders 49:07.200 --> 49:09.200 that are logical thinkers, 49:09.200 --> 49:11.200 are always pushing their limits 49:11.200 --> 49:13.200 and have very high levels of integrity. 49:13.200 --> 49:15.200 So that's Dan Kahn in my current company 49:15.200 --> 49:17.200 and actually his brother and a couple other guys 49:17.200 --> 49:19.200 for Justin TV and Twitch. 49:19.200 --> 49:23.200 And then I think the last thing is just, 49:23.200 --> 49:26.200 I guess, persistence or perseverance. 49:26.200 --> 49:29.200 And that can apply to sticking to 49:29.200 --> 49:33.200 having conviction around the original premise of your idea 49:33.200 --> 49:36.200 and sticking around to do all the unsexy work 49:36.200 --> 49:38.200 to actually make it come to fruition, 49:38.200 --> 49:41.200 including dealing with whatever it is 49:41.200 --> 49:43.200 that you're not passionate about, 49:43.200 --> 49:47.200 whether that's finance or HR or operations or those things. 49:47.200 --> 49:49.200 As long as you are grinding away 49:49.200 --> 49:52.200 and working towards that North Star for your business, 49:52.200 --> 49:54.200 whatever it is and you don't give up 49:54.200 --> 49:56.200 and you're making progress every day, 49:56.200 --> 49:58.200 it seems like eventually you'll end up in a good place. 49:58.200 --> 50:00.200 And the only things that can slow you down 50:00.200 --> 50:01.200 are running out of money 50:01.200 --> 50:03.200 or I suppose your competitor is destroying you, 50:03.200 --> 50:06.200 but I think most of the time it's people giving up 50:06.200 --> 50:08.200 or somehow destroying things themselves 50:08.200 --> 50:10.200 rather than being beaten by their competition 50:10.200 --> 50:11.200 or running out of money. 50:11.200 --> 50:14.200 Yeah, if you never quit, eventually you'll arrive. 50:14.200 --> 50:16.200 It's a much more concise version 50:16.200 --> 50:18.200 of what I was trying to say. 50:18.200 --> 50:21.200 So you went the Y Combinator out twice. 50:21.200 --> 50:23.200 What do you think, in a quick question, 50:23.200 --> 50:25.200 do you think is the best way to raise funds 50:25.200 --> 50:27.200 in the early days? 50:27.200 --> 50:30.200 Or not just funds, but just community, 50:30.200 --> 50:32.200 develop your idea and so on. 50:32.200 --> 50:37.200 Can you do it solo or maybe with a cofounder 50:37.200 --> 50:39.200 like self funded? 50:39.200 --> 50:40.200 Do you think Y Combinator is good? 50:40.200 --> 50:41.200 Is it good to do VC route? 50:41.200 --> 50:43.200 Is there no right answer or is there, 50:43.200 --> 50:45.200 from the Y Combinator experience, 50:45.200 --> 50:47.200 something that you could take away 50:47.200 --> 50:49.200 that that was the right path to take? 50:49.200 --> 50:50.200 There's no one size fits all answer, 50:50.200 --> 50:54.200 but if your ambition I think is to see how big 50:54.200 --> 50:57.200 you can make something or rapidly expand 50:57.200 --> 50:59.200 and capture a market or solve a problem 50:59.200 --> 51:02.200 or whatever it is, then going the venture 51:02.200 --> 51:04.200 back route is probably a good approach 51:04.200 --> 51:07.200 so that capital doesn't become your primary constraint. 51:07.200 --> 51:10.200 Y Combinator, I love because it puts you 51:10.200 --> 51:13.200 in this sort of competitive environment 51:13.200 --> 51:16.200 where you're surrounded by the top, 51:16.200 --> 51:19.200 maybe 1% of other really highly motivated 51:19.200 --> 51:22.200 peers who are in the same place. 51:22.200 --> 51:26.200 In that environment I think just breeds success. 51:26.200 --> 51:28.200 If you're surrounded by really brilliant 51:28.200 --> 51:30.200 hardworking people, you're going to feel 51:30.200 --> 51:32.200 sort of compelled or inspired to try 51:32.200 --> 51:35.200 to emulate them or beat them. 51:35.200 --> 51:37.200 So even though I had done it once before 51:37.200 --> 51:41.200 and I felt like I'm pretty self motivated, 51:41.200 --> 51:43.200 I thought this is going to be a hard problem, 51:43.200 --> 51:45.200 I can use all the help I can get. 51:45.200 --> 51:46.200 So surrounding myself with other entrepreneurs 51:46.200 --> 51:48.200 is going to make me work a little bit harder 51:48.200 --> 51:51.200 or push a little harder then it's worth it. 51:51.200 --> 51:54.200 That's why I did it, for example, the second time. 51:54.200 --> 51:57.200 Let's go full soft, go existential. 51:57.200 --> 52:00.200 If you go back and do something differently in your life, 52:00.200 --> 52:06.200 starting in high school and MIT, leaving MIT, 52:06.200 --> 52:08.200 you could have gone to the PhD route, 52:08.200 --> 52:13.200 doing startup, going to see about a startup in California 52:13.200 --> 52:15.200 or maybe some aspects of fundraising. 52:15.200 --> 52:17.200 Is there something you regret, 52:17.200 --> 52:20.200 not necessarily regret, but if you go back, 52:20.200 --> 52:22.200 you could do differently? 52:22.200 --> 52:24.200 I think I've made a lot of mistakes, 52:24.200 --> 52:26.200 pretty much everything you can screw up, 52:26.200 --> 52:28.200 I think I've screwed up at least once. 52:28.200 --> 52:30.200 But I don't regret those things. 52:30.200 --> 52:32.200 I think it's hard to look back on things, 52:32.200 --> 52:34.200 even if they didn't go well and call it a regret, 52:34.200 --> 52:37.200 because hopefully it took away some new knowledge 52:37.200 --> 52:39.200 or learning from that. 52:42.200 --> 52:45.200 I would say there's a period, 52:45.200 --> 52:47.200 the closest I can come to this, 52:47.200 --> 52:49.200 there's a period in just in TV, 52:49.200 --> 52:54.200 I think after seven years where the company was going 52:54.200 --> 52:57.200 one direction, which is towards Twitch and video gaming. 52:57.200 --> 52:58.200 I'm not a video gamer. 52:58.200 --> 53:01.200 I don't really even use Twitch at all. 53:01.200 --> 53:04.200 I was still working on the core technology there, 53:04.200 --> 53:06.200 but my heart was no longer in it, 53:06.200 --> 53:08.200 because the business that we were creating 53:08.200 --> 53:10.200 was not something that I was personally passionate about. 53:10.200 --> 53:12.200 It didn't meet your bar of existential impact. 53:12.200 --> 53:16.200 Yeah, and I'd say I probably spent an extra year or two 53:16.200 --> 53:20.200 working on that, and I'd say I would have just tried 53:20.200 --> 53:22.200 to do something different sooner. 53:22.200 --> 53:26.200 Because those were two years where I felt like, 53:26.200 --> 53:29.200 from this philosophical or existential thing, 53:29.200 --> 53:31.200 I just felt that something was missing. 53:31.200 --> 53:34.200 If I could look back now and tell myself, 53:34.200 --> 53:35.200 I would have said exactly that. 53:35.200 --> 53:38.200 You're not getting any meaning out of your work personally 53:38.200 --> 53:39.200 right now. 53:39.200 --> 53:41.200 You should find a way to change that. 53:41.200 --> 53:44.200 And that's part of the pitch I used 53:44.200 --> 53:46.200 to basically everyone who joins Cruise today. 53:46.200 --> 53:48.200 It's like, hey, you've got that now by coming here. 53:48.200 --> 53:51.200 Well, maybe you needed the two years of that existential dread 53:51.200 --> 53:53.200 to develop the feeling that ultimately 53:53.200 --> 53:55.200 it was the fire that created Cruise. 53:55.200 --> 53:56.200 So you never know. 53:56.200 --> 53:57.200 You can't repair. 53:57.200 --> 53:58.200 Good theory, yeah. 53:58.200 --> 53:59.200 So last question. 53:59.200 --> 54:02.200 What does 2019 hold for Cruise? 54:02.200 --> 54:05.200 After this, I guess we're going to go and talk to your class. 54:05.200 --> 54:08.200 But one of the big things is going from prototype to production 54:08.200 --> 54:09.200 for autonomous cars. 54:09.200 --> 54:10.200 And what does that mean? 54:10.200 --> 54:11.200 What does that look like? 54:11.200 --> 54:14.200 2019 for us is the year that we try to cross over 54:14.200 --> 54:17.200 that threshold and reach superhuman level of performance 54:17.200 --> 54:20.200 to some degree with the software and have all the other 54:20.200 --> 54:23.200 of the thousands of little building blocks in place 54:23.200 --> 54:27.200 to launch our first commercial product. 54:27.200 --> 54:30.200 So that's what's in store for us. 54:30.200 --> 54:32.200 And we've got a lot of work to do. 54:32.200 --> 54:35.200 We've got a lot of brilliant people working on it. 54:35.200 --> 54:37.200 So it's all up to us now. 54:37.200 --> 54:38.200 Yeah. 54:38.200 --> 54:41.200 So Charlie Miller and Chris Vell is like the people I've 54:41.200 --> 54:42.200 crossed paths with. 54:42.200 --> 54:43.200 Oh, great, yeah. 54:43.200 --> 54:46.200 It sounds like you have an amazing team. 54:46.200 --> 54:49.200 So like I said, it's one of the most, I think, one of the most 54:49.200 --> 54:52.200 important problems in artificial intelligence of this century. 54:52.200 --> 54:53.200 It'll be one of the most defining. 54:53.200 --> 54:55.200 It's super exciting that you work on it. 54:55.200 --> 54:59.200 And the best of luck in 2019. 54:59.200 --> 55:01.200 I'm really excited to see what Cruise comes up with. 55:01.200 --> 55:02.200 Thank you. 55:02.200 --> 55:03.200 Thanks for having me today. 55:03.200 --> 55:08.200 Thank you.