George Hotz: Comma.ai, OpenPilot, and Autonomous Vehicles | Lex Fridman Podcast #31

1. 0:00 – 15:00

   The following is a…

   1. LFLex Fridman0:00

      The following is a conversation with George Hotz. He's the founder of Comma.ai, a machine learning based vehicle automation company. He is most certainly an outspoken personality in the field of AI and technology in general. He first gained recognition for being the first person to carrier unlock an iPhone and since then, he's done quite a few interesting things at the intersection of hardware and software. This is the Artificial Intelligence podcast. If you enjoy it, subscribe on YouTube, give it five stars on iTunes, support it on Patreon, or simply connect with me on Twitter @lexfridman, spelled F-R-I-D-M-A-N. And I'd like to give a special thank you to Jennifer from Canada for her support of the podcast on Patreon. Merci beaucoup, Jennifer. She's, uh, been a friend and an engineering colleague for many years since I was in grad school. Your support means a lot and inspires me to keep this series going. And now here's my conversation with George Hotz. Do you think we're living in a simulation?

   2. GHGeorge Hotz1:06

      Y- yes, but it may be unfalsifiable.

   3. LFLex Fridman1:10

      What do you mean by unfalsifiable?

   4. GHGeorge Hotz1:12

      So if the simulation is designed in such a way that they did like a formal proof to show that no information can get in and out, and if their hardware is designed to, for the, anything in the simulation to always keep the hardware in spec, it may be impossible to prove whether we're in a simulation or not.

   5. LFLex Fridman1:32

      So they've designed it such that it's a closed system, you can't get outside the system?

   6. GHGeorge Hotz1:37

      Well, maybe it's one of three worlds. We're either in a simulation which can be exploited, we're in a simulation which not only can't be exploited but like the same thing's true about VMs. Um, a really well-designed VM, you can't even detect if you're in a VM or not.

   7. LFLex Fridman1:50

      (laughs) That's brilliant. So we're, uh, it's, yeah, so the simulation's running on a, on a virtual machine?

   8. GHGeorge Hotz1:56

      Yeah. But now i- in reality all VMs have ways to detect.

   9. LFLex Fridman1:59

      That's the point. I mean, is it, uh, y- you've done quite a bit of hacking yourself, uh, and so you should know that, uh, really any complicated system will have ways in and out.

   10. GHGeorge Hotz2:11

       So this isn't necessarily true going forward. I spent my time away from Comma, I learned, uh, Coq.

   11. LFLex Fridman2:19

       Mm-hmm.

   12. GHGeorge Hotz2:19

       It's a dependently typed, like, uh, it's a language we're writing math proofs in.

   13. LFLex Fridman2:23

       Mm-hmm.

   14. GHGeorge Hotz2:24

       And if you write code that compiles in a language like that, it is correct by definition. The, the types check its correctness.

   15. LFLex Fridman2:32

       Mm-hmm.

   16. GHGeorge Hotz2:33

       So it's possible that the simulation is written in a language like this, in which case, you know...

   17. LFLex Fridman2:39

       Yeah, but that, that can't be sufficiently expressive a language like that.

   18. GHGeorge Hotz2:43

       Oh, it can.

   19. LFLex Fridman2:44

       It can be?

   20. GHGeorge Hotz2:45

       Oh, yeah.

   21. LFLex Fridman2:46

       Okay. Well so you, uh, hm, all right, so-

   22. GHGeorge Hotz2:48

       The simulation doesn't have to be Turing-complete if it has a scheduled end date. Looks like it does actually with Entropy.

   23. LFLex Fridman2:54

       I, uh, I don't think that, uh, a simulation as, that results in, uh, something as complicated as the universe wo- would have a form of proof th- of correctness, right? Uh, it's possible of course, uh...

   24. GHGeorge Hotz3:10

       We have no idea how good their tooling is, and we have-

   25. LFLex Fridman3:13

       True.

   26. GHGeorge Hotz3:13

       ... no idea how complicated the universe computer really is. It may be quite simple.

   27. LFLex Fridman3:18

       It's just very large, right?

   28. GHGeorge Hotz3:19

       It's very, it's definitely very large.

   29. LFLex Fridman3:22

       But the fundamental rules might be super simple.

   30. GHGeorge Hotz3:24

       Yeah. Conway's game of life kind of stuff.
2. 15:00 – 30:00

   I'm not even sure…

   1. LFLex Fridman15:00

      at the distribution of, of smart people, smart people are generally good. And then this other person, I was talking to Sean Carroll, the physicist, and he was saying, "No, good and bad people are evenly distributed amongst everybody." My sense was good hackers are in general good people and they don't want to mess with the world. What's your sense?

   2. GHGeorge Hotz15:21

      I'm not even sure about that. Like, I have a nice life. Crime wouldn't get me anything. But if you're good and you have these skills, you probably have a nice life too, right? Like-

   3. LFLex Fridman15:38

      Right, you can use to further things. But is there an ethical... Is there some... Is there a little voice in your head that says, uh, "Well, yeah, if you could hack something to where you could hurt people and you could earn a lot of money doing it, though. Not hurt physically perhaps, but disrupt their life in some kind of way." I- isn't there a little voice that says-

   4. GHGeorge Hotz16:02

      Um, well, two things. One, I don't really care about money.

   5. LFLex Fridman16:06

      Right.

   6. GHGeorge Hotz16:06

      So like the money wouldn't be an incentive. The thrill might be an incentive.

   7. LFLex Fridman16:10

      The thrill.

   8. GHGeorge Hotz16:10

      But when I was 19, I read Crime and Punishment.

   9. LFLex Fridman16:13

      Right. Good.

   10. GHGeorge Hotz16:14

       That was another, that was another great one that talked me out of ever really doing crime.

   11. LFLex Fridman16:18

       (laughs)

   12. GHGeorge Hotz16:19

       Um, 'cause it's like, "That's gonna be me. I'd get away with it," but it would just run through my head. Even if I got away with it, you know? And then you do crime for long enough, you'll never get away with it.

   13. LFLex Fridman16:28

       That's right, in the end. That's a good reason to be good.

   14. GHGeorge Hotz16:32

       I wouldn't say I'm good.

   15. LFLex Fridman16:33

       (laughs)

   16. GHGeorge Hotz16:33

       I would just say I'm not bad.

   17. LFLex Fridman16:35

       You're a talented programmer and a, a hacker in a good positive sense of the word, word. You've, uh, played around, found vulnerabilities in various systems. What have you learned broadly about the design of systems and so on from that, from that whole process?

   18. GHGeorge Hotz16:53

       You learn to not take things for what people say they are.

   19. LFLex Fridman17:02

       Mm-hmm.

   20. GHGeorge Hotz17:02

       But you look at things for what they actually are.

   21. LFLex Fridman17:06

       Hmm. Yeah.

   22. GHGeorge Hotz17:08

       I understand that's what you tell me it is, but what does it do?

   23. LFLex Fridman17:11

       Right. And you have nice visualization tools to really know what it's really doing.

   24. GHGeorge Hotz17:16

       Oh, I wish... I'm a better programmer now than I was in 2014. I said, "Cura, that was the first tool that I wrote that was usable." I wouldn't say the code was great. I still wouldn't say my code is great. It's better.

   25. LFLex Fridman17:28

       So how was your evolution as a programmer except practice? You went... You started with C, at which point did you pick up Python? 'Cause you're pretty big in Python now.

   26. GHGeorge Hotz17:37

       Now, yeah. In, uh, in college. Uh, I went to Carnegie Mellon when I was 22.

   27. LFLex Fridman17:42

       Nice.

   28. GHGeorge Hotz17:42

       Um, I went back, I'm like, "All right, I'm gonna take all your hardest CS courses and we'll see how I do," right? Like, did I miss anything by not having a, a real, uh, undergraduate education?

   29. LFLex Fridman17:50

       Yeah.

   30. GHGeorge Hotz17:51

       Took, uh, operating systems, compilers, AI, and their like, uh, freshman reader math course. Um, and som-
3. 30:00 – 45:00

   Well, first time I…

   1. GHGeorge Hotz30:00

      I can gather a data set and train this net in, in, in weeks." And I did.

   2. LFLex Fridman30:04

      Well, first time I tried the implementation of MobileEye in a Tesla, I was really surprised how good it is.

   3. GHGeorge Hotz30:10

      Mm-hmm.

   4. LFLex Fridman30:11

      Uh, it's quite incredibly good 'cause I thought it's... Just 'cause I've done a lot of computer vision, I thought it'd be a lot harder to create a system that that stable. Uh, so that, I was personally surprised. It's, you know-... uh, I have to admit it 'cause I was kind of skeptical before trying it 'cause I thought, uh, uh, it would go in and out a lot more, it would get disengaged a lot more, and it's pretty robust. Uh, so what's... how, how, how hard was the problem when you t- t- uh, when you tackled it?

   5. GHGeorge Hotz30:42

      So I think AP1 was great, like, uh, Elon talked about disengagements on the 405, down in LA where like the lane marks were kind of faded, um, and the Mobileye system would drop out.

   6. LFLex Fridman30:52

      Mm-hmm.

   7. GHGeorge Hotz30:53

      Uh, like, I had something up and working that I would say was, like, the same quality-

   8. LFLex Fridman30:59

      Mm-hmm.

   9. GHGeorge Hotz31:00

      ... in three months.

   10. LFLex Fridman31:02

       Same quality? But how do you know? You, you, you say stuff like that-

   11. GHGeorge Hotz31:06

       Yeah.

   12. LFLex Fridman31:06

       ... confidently, but you can't... uh, and I love it, but, uh-

   13. GHGeorge Hotz31:09

       Well-

   14. LFLex Fridman31:09

       ... the question is, you can't... you, you're kind of going by feel 'cause you tested it out.

   15. GHGeorge Hotz31:13

       You're going by feel, absolutely.

   16. LFLex Fridman31:15

       Yeah.

   17. GHGeorge Hotz31:15

       Absolutely. Like, like I would take... I had, I borrowed my friend's Tesla.

   18. LFLex Fridman31:18

       Yeah.

   19. GHGeorge Hotz31:18

       I would take AP1 out for a drive.

   20. LFLex Fridman31:20

       Yep.

   21. GHGeorge Hotz31:20

       And then I would take my system out for a drive.

   22. LFLex Fridman31:22

       And it seems reasonably like, uh, the same. (sighs) So the 405, how hard is it to create something that could actually be a product that's deployed? I mean, uh, I've, I've read an article where Elon dis-, uh, responded, said something about you saying that, um, to build autopilot is, uh, is more complicated than a single-

   23. GHGeorge Hotz31:48

       Mm-hmm.

   24. LFLex Fridman31:48

       ... George Hodgs-level job. How hard is that job to create something that would work across the, globally?

   25. GHGeorge Hotz31:58

       Um, I don't think global is the challenge, but Elon followed that up by saying, "It's gonna take two years and a company of 10 people."

   26. LFLex Fridman32:04

       Yeah.

   27. GHGeorge Hotz32:04

       And here I am, four years later with a company of 12 people, and I think we still have another two to go.

   28. LFLex Fridman32:10

       (laughs) Two years? So yeah, so what do you think, um, what do you think about, uh, how Tesla's progressing with autopilot V2, V3?

   29. GHGeorge Hotz32:19

       I think we've kept pace with them pretty well. I think Navigator and autopilot is terrible. We had some demo features internally of the same stuff, and we would test it, and I'm like, "I'm not shipping this even as, like, open source software to people."

   30. LFLex Fridman32:35

       Why do you think it's terrible?
4. 45:00 – 1:00:00

   Right. …

   1. GHGeorge Hotz45:01

   2. LFLex Fridman45:01

      Right.

   3. GHGeorge Hotz45:02

      But overall, mostly, yeah.

   4. LFLex Fridman45:03

      That's so cool that you know all this stuff. That's, uh... I don't, uh, often talk to people that... 'cause it's such a rare car, unfortunately, currently.

   5. GHGeorge Hotz45:10

      We, we, we bought one-

   6. LFLex Fridman45:11

      Yeah.

   7. GHGeorge Hotz45:11

      ... explicitly for this. We, we-

   8. LFLex Fridman45:12

      That's awesome.

   9. GHGeorge Hotz45:12

      ... lost like, like 25K in the deprecation, but I feel it was worth it.

   10. LFLex Fridman45:16

       I was very pleasantly surprised that, uh, GM system was so innovative, uh, and really, uh, that, that it wasn't advertised much, wasn't talked about much.

   11. GHGeorge Hotz45:27

       Yeah.

   12. LFLex Fridman45:28

       Um, and I was nervous that it would die. (laughs) That it would disappear.

   13. GHGeorge Hotz45:31

       Well, I-

   14. LFLex Fridman45:32

       And that-

   15. GHGeorge Hotz45:32

       They, they put it on the wrong car. They should've put it on the Bolt and not some weird Cadillac that nobody bought. Um-

   16. LFLex Fridman45:36

       I think that's going to be into... uh, they're saying, at least, it's going to be into their entire fleet. So what do you think about... and if... as long as we're on the driver monitoring, uh, what do you think about Elon Musk's claim that driver monitoring is not needed?

   17. GHGeorge Hotz45:52

       Normally, I love his claims. That one is stupid. That one is stupid, and, you know, he's not going to have his level 5 fleet by the end of the year. Hopefully, he's like, "Okay, I was wrong. I'm going to add driver monitoring." Because when these systems get to the point that they're only messing up once every thousand miles, you absolutely need driver monitoring.

   18. LFLex Fridman46:14

       So let me play dev- 'cause I agree with you, but let me play devil's advocate.

   19. GHGeorge Hotz46:17

       Sure.

   20. LFLex Fridman46:17

       One possibility is that without driver monitoring, people are able to monitor, uh, s- uh, self-regulate, monitor themselves. You know, that... so your idea is obviously-

   21. GHGeorge Hotz46:30

       You're seeing all the people sleeping in, in Teslas? Uh...

   22. LFLex Fridman46:33

       Uh, yeah. Well, I'm a little skeptical of all the people sleeping in Teslas because, um, I, I've, I've stopped paying attention to that kind of stuff because I want to see real data. There's too much glorified... it's doesn't-

   23. GHGeorge Hotz46:47

       Yeah.

   24. LFLex Fridman46:47

       ... feel scientific to me. So I want to know, you know, what... how many people are really sleeping in Teslas versus sleeping... I've... I was driving here sleep-deprived in a car with no automation.

   25. GHGeorge Hotz46:59

       Yeah.

   26. LFLex Fridman46:59

       I was falling asleep.

   27. GHGeorge Hotz47:01

       I agree that it's hype-y. It's just like... you know what? If you want to put driver monitoring, I, I, okay, I rented a... my last autopilot experience was I rented a Model 3 in March-

   28. LFLex Fridman47:10

       Mm-hmm.

   29. GHGeorge Hotz47:11

       ... and drove it around. The wheel thing is annoying. And the reason the wheel thing is annoying, we use the wheel thing as well, but we don't disengage on wheel. For Tesla, you have to touch the wheel just enough-

   30. LFLex Fridman47:21

       Yeah.
5. 1:00:00 – 1:15:00

   Yeah. …

   1. LFLex Fridman1:00:00

      and perception is similar to what you're describing, which is really turning into a... not some kind of modular thing, but really do... formulate it as a learning problem-

   2. GHGeorge Hotz1:00:11

      Yeah.

   3. LFLex Fridman1:00:11

      ... and solve the learning problem with scale. So how many years... uh, point one is how many years would it take to solve this problem or- or just how hard is this freaking problem?

   4. GHGeorge Hotz1:00:21

      Well, the cool thing is I think there's a lot of value that we can deliver along the way.

   5. LFLex Fridman1:00:30

      Mm-hmm.

   6. GHGeorge Hotz1:00:30

      I think that you can build lane keeping assist, actually, plus adaptive cruise control, plus, okay, looking at Waze, extends to, like, all of driving.

   7. LFLex Fridman1:00:46

      Yeah, most of driving, right?

   8. GHGeorge Hotz1:00:47

      Right? Oh, your adaptive cruise control treats red lights like cars. Okay.

   9. LFLex Fridman1:00:51

      So let's jump around, which you mentioned that you didn't like, uh, Navigator and Autopilot.

   10. GHGeorge Hotz1:00:55

       Yeah.

   11. LFLex Fridman1:00:55

       What advice... how would you make it better? Do you think as a feature that if it's done really well, it's a good feature?

   12. GHGeorge Hotz1:01:02

       I think that it's too reliant on like hand-coded hacks for like h- how does Navigate on Autopilot do a lane change? It actually does the same lane change every time, and it feels mechanical. Humans do different lane changes.

   13. LFLex Fridman1:01:15

       Mm-hmm.

   14. GHGeorge Hotz1:01:16

       Humans sometimes will do a slow one, sometimes do a fast one. Navigate on Autopilot, at least every time I used it, it did the identical lane change.

   15. LFLex Fridman1:01:23

       How do you learn... I mean, this is a fundamental thing, actually.

   16. GHGeorge Hotz1:01:26

       Yeah.

   17. LFLex Fridman1:01:26

       Is, uh, the braking and then accelerating something that's still, uh... Tesla probably does it better than most cars, but it still doesn't do a great job of creating a comfortable, natural experience. And Navigate on Autopilot is just lane changes and an extension of that. So how do you learn to do a natural lane change?

   18. GHGeorge Hotz1:01:49

       So we have it, and I can talk about how it works. So I feel that we have the solution for lateral. Uh, we don't yet have the solution for longitudinal. There's a few reasons longitudinal is harder than lateral. The lane change component, the way that we train on it very simply is like, our model has an input for whether it's doing a lane change or not.

   19. LFLex Fridman1:02:13

       Mm-hmm.

   20. GHGeorge Hotz1:02:14

       And then when we train the end-to-end model, we, w- hand label all the lane changes 'cause you have to.

   21. LFLex Fridman1:02:19

       Mm-hmm.

   22. GHGeorge Hotz1:02:19

       I've, I struggled a long time about not wanting to do that, but I think you have to 'cause y- or-

   23. LFLex Fridman1:02:24

       For the training data.

   24. GHGeorge Hotz1:02:25

       For the training data, right? Uh, we actually, we have an automatic ground truther which automatically labels all the lane changes.

   25. LFLex Fridman1:02:30

       Is that possible?

   26. GHGeorge Hotz1:02:31

       To automatically label lane changes?

   27. LFLex Fridman1:02:32

       Yeah.

   28. GHGeorge Hotz1:02:33

       Yeah, detect the lane. I c- when it crosses it, right? And I don't know if they get that, that high percent accuracy, but it's like 95, good enough.

   29. LFLex Fridman1:02:38

       Okay.

   30. GHGeorge Hotz1:02:39

       Now, I set the bit when it's doing the lane change in the end-to-end learning. And then I set it to zero when it's not doing a lane change. So now if I wanted to do a lane change at test time, I just put the bit to a one and it'll do a lane change.
6. 1:15:00 – 1:16:34

   Oh, don't trust me.…

   1. LFLex Fridman1:15:01

      believe you, but I have to take it with a grain of salt because, I mean, you, you are an inspiration because you basically don't care about a lot of things that other companies care about. You don't try to bullshit, in, in a sense, like make up stuff, so d- drive up valuation. You're really very real and you're trying to solve the problem. I admire that a lot. What I don't necessarily fully, can't trust you on-

   2. GHGeorge Hotz1:15:26

      Oh, don't trust me. That's true.

   3. LFLex Fridman1:15:26

      ... with all due respect, is how good it is, right? I can only ... But I also know how bad others are. And so (laughs) that-

   4. GHGeorge Hotz1:15:33

      I'll say two, I'll say two things about d- trust but verify, right?

   5. LFLex Fridman1:15:36

      Yeah.

   6. GHGeorge Hotz1:15:36

      I'll say two things about that. One is try, uh, get in a 2020 Corolla-

   7. LFLex Fridman1:15:42

      Yeah.

   8. GHGeorge Hotz1:15:42

      ... and try OpenPilot 0.6 when, when it comes out next month. Um, I think already, you'll look at this and you'll be like-

   9. LFLex Fridman1:15:49

      Damn.

   10. GHGeorge Hotz1:15:49

       ... "This is already really good." And then I could be doing that all with hand labelers and all with, with, like, like the same approach that, like, MobileEye uses.

   11. LFLex Fridman1:15:57

       Mm-hmm.

   12. GHGeorge Hotz1:15:58

       When we release a model that no longer has the lanes in it, that only outputs a path-

   13. LFLex Fridman1:16:03

       Mm-hmm.

   14. GHGeorge Hotz1:16:05

       ... then think about how we did that machine learning and then right away, when you see, and that's going to be in OpenPilot, that's gonna be in OpenPilot before 1.0, when you see that model, you'll know that everything I'm saying is true, because how else did I get that model?

   15. LFLex Fridman1:16:16

       Good. This-

   16. GHGeorge Hotz1:16:17

       You'll know that what I, what I'm saying is true about the simulator, right?

   17. LFLex Fridman1:16:19

       Yeah, yeah. That's super exciting.

   18. GHGeorge Hotz1:16:20

       Yeah.

   19. LFLex Fridman1:16:20

       That's super exciting. And, um ...

   20. GHGeorge Hotz1:16:22

       But, like, you know, I listened to your talk with Kyle, and Kyle was originally building, uh, the, the aftermarket system, and he gave up on it because of technical challenges.

Episode duration: 1:59:36