Skip to content
Lex Fridman PodcastLex Fridman Podcast

George Hotz: Hacking the Simulation & Learning to Drive with Neural Nets | Lex Fridman Podcast #132

George Hotz (geohot) is a programmer, hacker, and the founder of Comma.ai. Please support this podcast by checking out our sponsors: - Four Sigmatic: https://foursigmatic.com/lex and use code LexPod to get up to 40% & free shipping - Decoding Digital: https://appdirect.com/decoding-digital - ExpressVPN: https://expressvpn.com/lexpod and use code LexPod to get 3 months free EPISODE LINKS: Comma.ai's Twitter: https://twitter.com/comma_ai Comma.ai's Website: https://comma.ai/ George's Instagram: https://www.instagram.com/georgehotz George's Twitch: https://www.twitch.tv/georgehotz George's Twitter: https://twitter.com/realgeorgehotz Comma.ai YouTube (unofficial): https://www.youtube.com/channel/UCwgKmJM4ZJQRJ-U5NjvR2dg PODCAST INFO: Podcast website: https://lexfridman.com/podcast Apple Podcasts: https://apple.co/2lwqZIr Spotify: https://spoti.fi/2nEwCF8 RSS: https://lexfridman.com/feed/podcast/ Full episodes playlist: https://www.youtube.com/playlist?list=PLrAXtmErZgOdP_8GztsuKi9nrraNbKKp4 Clips playlist: https://www.youtube.com/playlist?list=PLrAXtmErZgOeciFP3CBCIEElOJeitOr41 OUTLINE: 0:00 - Introduction 2:31 - Will human civilization destroy itself? 5:18 - Where are the aliens? 10:05 - Tic Tac UFO and Bob Lazar 12:33 - Conspiracy theories 14:36 - The programming language of life 18:57 - The games that humans play 27:27 - Memory leaks in the simulation 29:58 - Theories of everything 31:43 - Ethereum startup story 39:30 - Cryptocurrency 48:57 - Self-help advice 52:37 - Comma.ai 54:30 - Comma two 1:03:19 - Tesla vs Comma.ai 1:12:22 - Driver monitoring 1:26:03 - Communicating uncertainty 1:27:51 - Tesla Dojo 1:34:19 - Tesla Autopilot big rewrite 1:40:37 - How to install the Comma Two 1:45:13 - Openpilot is Android & Autopilot is iOS 1:54:28 - Waymo 2:05:41 - Autonomous driving and society 2:07:53 - Moving 2:10:58 - Advice to Startups 2:24:00 - Programming setup 2:27:01 - Ideas that changed my life 2:35:06 - GPT-3 2:38:26 - AGI 2:42:29 - Programming languages that everyone should learn 2:49:02 - How to learn anything 2:51:34 - Book recommendations 2:59:57 - Love 3:01:46 - Psychedelics 3:04:07 - Crazy CONNECT: - Subscribe to this YouTube channel - Twitter: https://twitter.com/lexfridman - LinkedIn: https://www.linkedin.com/in/lexfridman - Facebook: https://www.facebook.com/LexFridmanPage - Instagram: https://www.instagram.com/lexfridman - Medium: https://medium.com/@lexfridman - Support on Patreon: https://www.patreon.com/lexfridman

Lex FridmanhostGeorge Hotzguest
Oct 22, 20203h 8mWatch on YouTube ↗

EVERY SPOKEN WORD

  1. 0:002:31

    Introduction

    1. LF

      The following is a conversation with George Hotz, AKA Geohot. His second time on the podcast. He's the founder of Comma AI, an autonomous and semi-autonomous vehicle technology company that seeks to be to Tesla autopilot what Android is to the iOS. They sell the Comma 2 device for $1,000 that, when installed in many of their supported cars, can keep the vehicle centered in the lane even when there are no lane markings. It includes driver sensing that ensures that the driver's eyes are on the road. As you may know, I'm a big fan of driver sensing. I do believe Tesla autopilot and others should definitely include it in their sensor suite. Also, I'm a fan of Android and a big fan of George for many reasons, including his nonlinear, out-of-the-box brilliance and the fact that he's a superstar programmer of a very different style than myself. Styles make fights, and styles make conversations. So I really enjoyed this chat, and I'm sure we'll talk many more times on this podcast. Quick mention of a sponsor, followed by some thoughts related to the episode. First is Four Sigmatic, the maker of delicious mushroom coffee. Second is Decoding Digital, a podcast on tech and entrepreneurship that I listen to and enjoy. And finally, ExpressVPN, the VPN I've used for many years to protect my privacy on the internet. Please check out the sponsors in the description to get a discount and to support this podcast. As a side note, let me say that my work at MIT on autonomous and semi-autonomous vehicles led me to study the human side of autonomy enough to understand that it's a beautifully complicated and interesting problem space, much richer than what can be studied in the lab. In that sense, the data that Comma AI, Tesla autopilot, and perhaps others like Cadillac super crews are collecting gives us a chance to understand how we can design safe, semi-autonomous vehicles for real human beings in real-world conditions. I think this requires bold innovation and a serious exploration of the first principles of the driving task itself. If you enjoyed this thing, subscribe on YouTube, review it with five stars on Apple Podcast, follow on Spotify, support on Patreon, or connect with me on Twitter @lexfridman. And now, here's my conversation with George Hotz.

  2. 2:315:18

    Will human civilization destroy itself?

    1. LF

      So last time I started talking about the simulation. This time, let me ask you, do you think there's intelligent life out there in the universe?

    2. GH

      I've always maintained my answer to the Fermi paradox. I think there has been intelligent life elsewhere in the universe.

    3. LF

      So intelligent civilizations existed, but they've blown themselves up. So your general intuition is that intelligent civilizations quickly... Like, there's that parameter in, in the Drake equation. Your sense is they don't last very long.

    4. GH

      Yeah.

    5. LF

      How are we doing on that? Like, ha- have we lasted pretty, pretty good?

    6. GH

      Oh, no.

    7. LF

      Or we do?

    8. GH

      Oh, y- yeah. I mean, not quite yet.

    9. LF

      (laughs)

    10. GH

      Well, �... Or could I tell you as Yudkowsky, the IQ required to destroy the world falls by one point every year.

    11. LF

      Okay, so technology democratizes the destruction of the world.

    12. GH

      When can a meme destroy the world?

    13. LF

      (laughs) It kinda is already, right?

    14. GH

      Somewhat. In a way. And I think, I don't think we've seen anywhere near the worst of it yet. World's gonna get weird.

    15. LF

      Well, maybe a meme can save the world. Have you thought about that?

    16. GH

      Yeah.

    17. LF

      The meme lord Elon Musk fighting on the side of good versus the, uh, the meme lord of the darkness, which is, uh... I'm not saying anything bad about Donald Trump-

    18. GH

      (laughs)

    19. LF

      ... but he is the, the lord of the meme on the dark side. He's a Darth Vader of memes.

    20. GH

      I think in every fairy tale, they always end it with, "And they lived happily ever after." And I'm like, "Please tell me more about this happily ever after. I've heard 50% of marriages end in divorce. Uh, why doesn't your marriage end up there?" You can't just say, "Happily ever after." So-

    21. LF

      (laughs)

    22. GH

      ... i- it's... The thing about destruction is it's over after the destruction. Um, we have to do everything right in order to avoid it, and, uh, one thing wrong... I mean, actually, this is what I really like about cryptography. Uh, cryptography, it seems like we live in a world where the defense wins. Um, versus, like, nuclear weapons, the opposite is true. It is much easier to build a warhead that splits into 100 little warheads than to build something that can, you know, take out 100 little warheads. Uh, the offense has the advantage there. Um, so maybe our future is in crypto, but-

    23. LF

      (laughs)

    24. GH

      Uh...

    25. LF

      So cryptography, right. The Goliath is the, the defense, and then all the different hackers are the, uh, are the Davids, and that equation is flipped for nuclear war, because there's so many... Like, one nuclear weapon destroys everything, essentially.

    26. GH

      Yeah, and it is much easier to, uh, attack with a nuclear weapon than it is to like... The technology required to intercept and destroy a rocket is much more complicated than the technology required to just, you know, orbital trajectory send a rocket to somebody.

    27. LF

      So, okay,

  3. 5:1810:05

    Where are the aliens?

    1. LF

      your intuition that the, there were intelligent civilizations out there, but it's very possible that they're no longer there.

    2. GH

      Um-

    3. LF

      That's kind of a sad picture.

    4. GH

      They enter some steady state. They all wirehead themselves.

    5. LF

      What's wirehead?

    6. GH

      Um...

    7. LF

      (laughs)

    8. GH

      Stimulate, stimulate their pleasure centers-

    9. LF

      (laughs)

    10. GH

      ... uh, and just, you know, live forever in this kind of stasis.

    11. LF

      Oh.

    12. GH

      Th- they become... Well, I mean, I think the reason I believe this is because where are they? If there's some reason they stopped expanding, because otherwise they would have taken over the universe. The universe isn't that big, or at least, you know, let's just talk about the galaxy, right? 70,000 light years across. Uh, took that number from Star Trek Voyager, I don't know how true it is.

    13. LF

      (laughs)

    14. GH

      But, um, (laughs) uh, yeah, that's not big, right? 70,000 light years is nothing.

    15. LF

      ... for some possible technology that you can imagine that can leverage, like, wormholes or something like that?

    16. GH

      Oh, you don't even need wormholes. Just a von Neumann probe is enough. A von Neumann probe and a million years of sub-light travel, and you'd have taken over the whole universe. That clearly, uh, didn't happen, so something stopped it.

    17. LF

      So you mean if you... Right, for- for like a few million years, if you sent out probes that travel close. What's sub-light? You mean close to the speed of light?

    18. GH

      Let's say 0.1c.

    19. LF

      And it just spreads. Interesting. Actually, that's an interesting calculation. Huh.

    20. GH

      So-

    21. LF

      But what makes you think that we'd be able to, uh, communicate with them? Like, uh, yeah. What- what's, uh, why do you think we would be able to- to comprehend intelligent lives that are out there? Like, even if they were among us, kind of thing. Like, or even just flying around.

    22. GH

      Well, I mean, that's possible. It- it's possible that there is some sort of prime directive. Uh, that'd be a really cool universe to live in, um, and there's some reason they're not making themselves visible to us.

    23. LF

      Right.

    24. GH

      But it makes sense that they would use the same... Well, at least the same entropy.

    25. LF

      Well, you're implying the same laws of physics. I don't know what you mean by entropy in this case.

    26. GH

      Oh, uh, yeah.

    27. LF

      So-

    28. GH

      I mean, if entropy is the scarce resource in the universe?

    29. LF

      (laughs) So what do you think about, like, Steven Wolfram and everything is a computation, and then what if they are traveling through this world of computations? So if you think of the universe as just information processing, then, uh, what you're referring to with- with entropy, and- and then these- these pockets of interesting complex computations swimming around, how do we know they're not already here? How do... How do we know that this, like all the different amazing things that are full of mystery on Earth are just like little footprints of intelligence from light years away?

    30. GH

      Maybe. I mean, I tend to think that as civilizations expand, they use more and more energy, uh, and you can never overcome the problem of waste heat. So where is their waste heat?

  4. 10:0512:33

    Tic Tac UFO and Bob Lazar

    1. GH

    2. LF

      Yeah. Okay. What do you make about all the... I recently (laughs) came across this, uh, having talked to, uh, David Fravor. I don't know if you caught what the- the videos that Pentagon (laughs) released and, uh, the New York Times reporting of the UFO sightings. So I kind of "looked into it," and there's actually been like hundreds of thousands of UFO sightings, right? And a lot of it, you can explain away in different kinds of ways. So one is it could be interesting physical phenomena. Two, it could be people wanting to believe, and therefore they conjure up a lot of different things that just... You know, when you see different kinds of lights, some basic physics phenomena, and then you just conjure up ideas of possible out there mysterious worlds. But, you know, it's also possible, like you have a case of, um, David Fravor, who is a Navy pilot who's, you know, a leg- as legit as it gets in terms of humans who are able to perceive things in the environment and make conclusions whether those things are a threat or not. And he and several other pilots saw a thing, I don't know if you followed this, but they saw a thing that they've since then called tic toc that moved in all kinds of weird ways. They don't know what it is. It could be technology developed by- by the United States and they're just not aware of it, and the surface level from the Navy, right? It could be different kind of lighting technology or drone technology, all that kind of stuff. It could be the Russians and the Chinese, all that kind of stuff. And of course, their mind, our mind, can also venture into the possibility that it's from another world. Have- have you looked into this at all? What do you think about it?

    3. GH

      I think... I think all the news is a PSYOP.

    4. LF

      (laughs)

    5. GH

      Uh, I think that the most plausible-

    6. LF

      Nothing is real. (laughs)

    7. GH

      Yeah, I listened to the, uh... I think it was Bob Lazar-

    8. LF

      Yeah.

    9. GH

      ... um, on- on Joe Rogan, and like...... I believe everything this guy is saying, and then I think that it's probably just some, like, MKUltra kind of thing, you know? (laughs)

    10. LF

      Uh, what do you mean?

    11. GH

      Like, they- they- they, uh, you know, they made some weird thing and they called it an alien spaceship, you know? Maybe it was just to, like, stimulate young physicists minds, tell them it's alien technology and we'll see what they come up with, right?

  5. 12:3314:36

    Conspiracy theories

    1. LF

      Do you find any conspiracy theories compelling? Like have you pulled at the string of the- of the rich complex world of conspiracy theories that's out there?

    2. GH

      I think that, uh, I've heard a conspiracy theory that conspiracy theories were invented by the CIA in the 60s to discredit true things.

    3. LF

      Yeah.

    4. GH

      Um, so, you know, you can go to ridiculous conspiracy theories like flat earth and Pizzagate and, uh, y- you know, these things are almost to hide, like, conspiracy theories that like, you know, remember when the Chinese, like, locked up the doctors who discovered coronavirus? Like, I tell people this and I'm like, "No, no, no, that's not a conspiracy theory. That actually happened. Do you remember the time that the money used to be backed by gold and now it's backed by nothing? This is not a conspiracy theory. This actually happened."

    5. LF

      Well, that's one of my worries today with the idea of fake news, is that when nothing is real, then, like, you dilute the possibility of anything being true by conjuring up all kinds of conspiracy theories.

    6. GH

      Mm-hmm.

    7. LF

      And then you don't know what to believe, and then, uh, the- like the idea of truth, of objectivity has lost completely. Everybody has their own truth.

    8. GH

      So you used to control information by censoring it. And then the internet happened and governments were like, "Oh shit, we can't censor things anymore. I know what we'll do." (laughs) You know? It's the old story of, uh, the story of, like, tying a flag where the leprechaun tells you his gold is buried, and you tie one flag and you make the leprechaun swear to not remove the flag, and you come back to the field later with a shovel and there's flags everywhere.

    9. LF

      (laughs)

    10. GH

      (laughs)

    11. LF

      That's one way to maintain privacy, right? Is, like, in order to protect the contents of this conversation, for example, we could just generate, like, millions of deep fake conversations where you and I talk and say random things.

    12. GH

      Yeah.

    13. LF

      So this is just one of them and nobody knows which one was the real one. This co- this could be fake right now.

    14. GH

      Classic steganography technique.

  6. 14:3618:57

    The programming language of life

    1. GH

    2. LF

      (laughs) Okay, another absurd question about intelligent life, 'cause, uh, you know, you're- you're an incredible programmer, outside of everything else we'll talk about, just as a programmer. Uh, do you think intelligent beings out there, the civilizations that are out there, had computers and programming? Did they, do we naturally have to develop something where we engineer machines and are able to encode both knowledge into those machines and instructions that process that knowledge, process that information to- to make decisions and actions and so on? And would those programming languages, if you think they exist, be at all similar to anything we've developed?

    3. GH

      So I don't see that much of a difference between, quote unquote, natural languages and programming languages.

    4. LF

      Hmm.

    5. GH

      Um...

    6. LF

      Yeah.

    7. GH

      I think there's so many similarities. So when asked the question, uh, what do alien languages look like? I imagine they're not all that dissimilar from ours, and I think translating in and out of them, uh, wouldn't be that crazy.

    8. LF

      Well, it's difficult to compile, like, DNA to Python and then to C. There- I mean, there is a little bit of a gap in- in the kind of languages we use for- for, uh, Turing machines and the kind of languages nature seems to use, a little bit. Maybe that's just, we just haven't cr- we haven't understood the kind of language that nature uses well yet.

    9. GH

      DNA is a CAD model. It's not quite a programming language. It has no sort of, uh, serial, uh, execution. It's not quite a- a l- yeah, it's- it's a CAD model. Um, so I think in that sense, we actually completely understand it. The problem is, um, you know, uh, well, simulating on these CAD models, I- I played with it a bit this year, is- is super, uh, comp- computationally intensive if you want to go down to, like, the molecular level, um, where you need to go to see a lot of these phenomenon, like protein folding. Um, so yeah, it's not that it's- it's not, it's not that we don't understand it, it just requires a whole lot of compute to kind of compile it.

    10. LF

      For human minds, it's inefficient both for the pro- for the data representation and for the programming.

    11. GH

      Yeah. It runs well on raw nature.

    12. LF

      Yeah.

    13. GH

      It runs well on raw nature, and when we try to build, uh, emulators or simulators for that, uh, well, they're mad slow. And I've tried it. (laughs)

    14. LF

      (laughs) It- it runs in, uh, yeah, you- you've commented elsewhere, I don't remember where, that, uh, one of the problems is simulating nature is tough, and if you want to sort of, uh, deploy a prototype, I- I forgot how you- you put it but it made me laugh, but animals or humans would need to be involved-

    15. GH

      Mm-hmm.

    16. LF

      ... in- in order to- in order to try to run some prototype code on, um, like if we're talking about COVID and viruses and so on.

    17. GH

      Yeah.

    18. LF

      If- if you were trying to engineer some kind of defense mechanisms, like a vaccine, uh, against COVID or all that kind of stuff, that doing any kind of experimentation like you can with, like, autonomous vehicles would be very technically cost- e- technically and ethically costly.

    19. GH

      Well, I'm not sure about that. I think you can do tons of, uh, crazy biology in- in test tubes. I think my bigger complaint is more, oh, the tools are so bad.

    20. LF

      Like literally? You mean like- like-

    21. GH

      I'm not-

    22. LF

      ... libraries and-

    23. GH

      Oh, no, I'm not pipetting shit. Like, you're handing me a pi- I gotta... No. No.

    24. LF

      (laughs)

    25. GH

      No. No. There has to be some.

    26. LF

      ... like, automating stuff and-

    27. GH

      Yeah.

    28. LF

      ... and like the ... Yeah, but human biology is messy. Like it seems-

    29. GH

      But like look at those Theranos videos. They were a joke. It's like, it's like a little gantry. It's like a little XY gantry high school science project with the pipette. I'm like, "Really?"

    30. LF

      Gotta be something better. Well-

  7. 18:5727:27

    The games that humans play

    1. LF

      L- let's get all the crazy out of the way. Uh, so a bunch of people asked me, since we talked about the simulation last time, uh, we talked about hacking the simulation. Do you have any updates, any insights about how we might be able to go about hacking simulation if we indeed do live in a simulation?

    2. GH

      I think a lot of people misinterpreted the point of that South by talk. Um, the point of the South by talk was not literally to hack the simulation. Uh, I think that this ... We ... This is, this is an idea is literally just, I think, theoretical physics. I think that's the whole, you know, the whole goal, right? You want your grand unified theory, but then, okay, build the grand unified theory, search for exploits, right? I think we're nowhere near actually there yet. My hope, uh, with that was just more to like, like, "Are you people kidding me with the things you spend time thinking about? Do you understand, like, kinda how small you are? You are, you are bytes in God's computer. Really?" And the things that people get worked up about, and, you know?

    3. LF

      So basically, it, it was more a message of, uh, we should humble ourselves, that w- we get to, uh ... Like what-

    4. GH

      Well-

    5. LF

      What are we humans in this byte code?

    6. GH

      Yeah.

    7. LF

      I mean-

    8. GH

      And not just, not just humble ourselves, but like, like I'm not trying to, like, make people feel guilty or anything like that. I'm trying to say, like, literally look at what you are spending time on, right?

    9. LF

      What are you referring to? Are you referring to the Kardashians? What are we talking about?

    10. GH

      Um ...

    11. LF

      Twitter?

    12. GH

      I'm referring to ... No, the Kardashians, see everyone knows that's kind of fun. I'm, I'm referring more to, like, the economy.

    13. LF

      Words?

    14. GH

      You know, this idea that we gotta up our stock price.

    15. LF

      (laughs)

    16. GH

      Like ... Or, or what is, what is the goal function of humanity?

    17. LF

      You don't like the game of capitalism? Like, you don't like the games we've constructed for ourselves as humans?

    18. GH

      I'm a big fan of, of, of capitalism. I don't think that's really the game we're playing right now. I think we're playing a, a different game where the rules are rigged. Um ... (laughs)

    19. LF

      Okay, which games are interesting to you that we humans have constructed and which aren't? Which are productive and which are not?

    20. GH

      Actually, maybe that's the real point of the, of the talk. It's like, stop playing these fake human games. There's a real game here. We can play the real game. The real game is, you know, nature wrote the rules. This is a real game. There still is a game to play.

    21. LF

      But if you look at ... Sorry to interrupt. But if ... I don't know if you've seen the Instagram account Nature Is Metal. The game that nature seems to be playing is a lot, a lot more cruel than we humans want to put up with, or at least we see it as cruel.

    22. GH

      Of course.

    23. LF

      It's like the bigger thing eats the smaller thing and, uh, does it to impress another big thing so it can mate with that thing, and that's it. That (laughs) that seems to be the entirety of it.

    24. GH

      Well-

    25. LF

      There's no art. There's no music. There's no Comma AI. There's no Comma One, no Comma Two, no George Hotz with his brilliant talks at South by Southwest.

    26. GH

      See, I disagree though. I disagree that this is what nature is. I think nature just provided basically a, uh, open world MMORPG. And, um ... You know, here it's open world. I mean, if that's the game you wanna play, you can play that game.

    27. LF

      Well that ... Isn't that g- isn't that beautiful? I don't know if you played Diablo. They used to have a, I think, cow level where it's (laughs) ... Uh, so everybody would go just ... They figured out this, like, the best way to gain, like, experience points is to just slaughter cows over and over and over.

    28. GH

      (laughs)

    29. LF

      (laughs) And, uh, so they figured out this little sub-game within the bigger game that this is the most efficient way to get experience points, and everybody somehow agreed that getting experience points in an RPG context where you always wanna be getting more stuff, more skills, more levels, keep advancing, that seems to be good. So might as well spend, sacrifice actual enjoyment of playing a game, exploring a world-

    30. GH

      Mm-hmm.

  8. 27:2729:58

    Memory leaks in the simulation

    1. GH

    2. LF

      So that talk in your thinking here is not literally referring to a, a simulation in that our, our universe is a kind of computer program running in a computer. That's more of a thought experiment.

    3. GH

      Um-

    4. LF

      Do you also think of the potential of the sort of, uh, Bostrom, Elon Musk and others that talk about an actual program that simulates our universe?

    5. GH

      Oh, I don't doubt that we're in a simulation. I just think that it's not quite that important. I mean, I'm interested only in simulation theory as far as like it gives me power over nature. Uh, if it's totally unfalsifiable, then who cares?

    6. LF

      I mean, what do you think that experiment would look like? Like somebody, uh, on Twitter asked, "Ask George what signs we would look for to know whether or not we're in the simulation," which is exactly what you're asking, is like the, the step that precedes the step of knowing how to get more power from this knowledge is to get an indication that there's some power to be gained. So get an indication that there, you can, uh, discover and exploit cracks in the simulation, or it doesn't have to be, you know, in the physics of the universe.

    7. GH

      Yeah. Show me, I mean, like a memory leak could be cool.

    8. LF

      Huh.

    9. GH

      Some scrying technology, you know.

    10. LF

      What, what kind of technology?

    11. GH

      Scrying.

    12. LF

      What's that?

    13. GH

      Oh, that's a weird, uh... Scrying is the, is the, uh, paranormal ability to, uh, like, like remote viewing, like being able to see somewhere where you're not. Um, so, you know, I don't think you can do it by chanting in a room, but, um, if we could find... It's a memory leak, basically. Uh-

    14. LF

      (laughs) It's a memory leak. Yeah. You're able to access parts you're not supposed to.

    15. GH

      Yeah, yeah, yeah.

    16. LF

      And thereby discover a shortcut.

    17. GH

      Yeah. Maybe, I mean, memory leak means the other thing as well, but I mean like, yeah, like an ability to read arbitrary memory.

    18. LF

      Yeah.

    19. GH

      Right. Um, and that one's not that horrifying, right? The, the write ones start to be horrifying.

    20. LF

      Read, right. So it's the, the reading is not the problem.

    21. GH

      Yeah. It's like Heartbleed for the universe. (sighs)

    22. LF

      Oh boy. The writing is a big, big problem. It's a big problem, 'cause the moment you can write anything, even if it's just random noise, that's terrifying.

    23. GH

      I mean, even without, even without that, like even some of the, you know, the nanotech stuff that's coming, I think is...

    24. LF

      I don't know if you're paying attention,

  9. 29:5831:43

    Theories of everything

    1. LF

      but actually Eric Weinstein came out with a theory of everything. I mean, not came out, he's been working on a theory of everything in the physics world called geometric unity. And then for me, from computer science person like you, Steven Wolfram's theory of everything of like hypergraphs is super interesting and beautiful, but that not from a physics perspective, but from a computational perspective. I don't know. Have you paid attention to any of that?

    2. GH

      So again, like what would make me pay attention and like why, like, I hate string theory is...... okay, make a testable prediction, right? I, I'm only interested in, I'm not interested in theories for their intrinsic beauty. I'm interested in theories that give me power over the universe. So if these theories do, I'm very interested. Um-

    3. LF

      Can I just say how beautiful that is? Because a lot of physicists say, "I'm interested in experimental validation." And they skip out the part wh- where they say, "To give me more power in the universe." I just love the, um-

    4. GH

      No, I want, I want, I want-

    5. LF

      ... the clarity of that.

    6. GH

      I want 100 gigahertz processors. I want transistors that are smaller than atoms. I want, like, power.

    7. LF

      (laughs) That's, uh, that's true and that's where people from aliens to this kind of technology where people are worried that governments, like, who owns that power? Is it a George Hotz? Is it thousands of distributed hackers across the world? Is it governments? You know, is it, uh, Mark Zuckerberg? There's a lot of (laughs) people that, uh, I don't know if anyone trusts any one individual with power, so they're always worried.

    8. GH

      It's the beauty of blockchains.

    9. LF

      That's the beauty of blockchains, which we'll talk about.

  10. 31:4339:30

    Ethereum startup story

    1. LF

      On Twitter, somebody pointed me to a story, uh, a bunch of people pointed me to a story a few months ago where you went into a restaurant in New York, now you can correct me if any of this is wrong, and ran into a bunch of folks from a company in the crypto company who are trying to scale up Ethereum. And they had a technical deadline related to a Solidity to OVM compiler. So these are all Ethereum technologies. So you stepped in, they recognized you, uh, pulled you aside, explained their problem, and you stepped in and helped them solve the problem, uh, thereby creating l- legend status story.

    2. GH

      (laughs)

    3. LF

      So, uh, can you, uh, tell me the story with a little more detail? It seems kind of incredible. This, did this happen?

    4. GH

      Yeah, yeah. It's a true story. It's a true story. I mean, they, they wrote a very flattering, uh, account of it. Um, they, so Optimism is the spin, the company's called Optimism. It's a spinoff of Plasma. Uh, they're trying to build L2 solutions on Ethereum. So right now, uh, every Ethereum node has to run every transaction on the Ethereum network. Um, and this kind of doesn't scale, right? Because if you have n computers, well, you know, if that becomes 2n computers, you actually still get the same amount of compute, right? This is, this is like, like all of one scaling, um, because they all have to run it. Okay, fine. You get more blockchain security, but like, the blockchain's already so secure. Can we trade some of that off for speed? Uh, so that's kind of what these L2 solutions are. They built this thing which kind of, um, kind of sandbox, uh, for Ethereum contracts so they can run it in this L2 world and it can't do certain things in L world, in L1.

    5. LF

      Can I ask you for some definitions? What's L2?

    6. GH

      Oh, L2 is layer two. So L1 is like the base Ethereum chain, and then layer two is like a computational layer that runs, um, elsewhere, but still is kind of secured by layer one.

    7. LF

      And I'm sure a lot of people know, but Ethereum is a cryptocurrency, probably one of the most popular cryptocurrencies second to Bitcoin, and a, a lot of interesting technological innovations there. Maybe you can also slip in whenever you talk about this any things that are exciting to you in the Ethereum space, and why Ethereum?

    8. GH

      Well, I mean, Bitcoin, uh, is not turing-complete. Ethereum is not technically turing-complete with a gas limit, but close enough.

    9. LF

      With a gas limit. What's the gas limit? Resources?

    10. GH

      Uh, yeah, I mean, no computer's actually turing-complete.

    11. LF

      Right.

    12. GH

      Right.

    13. LF

      (laughs)

    14. GH

      They're gonna find that ram, you know, give me some, I can actually solve the whole problem.

    15. LF

      What is g- what's the word gas limit? You just have so many brilliant word, I'm not even gonna ask.

    16. GH

      Well, that's what that's, no, no, that's not my word. That's Ethereum's word.

    17. LF

      Gas limit.

    18. GH

      Ethereum, you have to spend gas per instruction. So like different op codes use different amounts of gas, and you buy gas with Ether to prevent people from basically DDoSing the network.

    19. LF

      So, uh, Bitcoin is proof of work, and then what's Ethereum?

    20. GH

      It's also proof of work. Uh, they're working on some proof of stake Ethereum 2.0 stuff. But right now it's, it's proof of work. It uses a different hash function from Bitcoin that's more ASIC resistant because you need ram.

    21. LF

      So we're all talking about Ethereum 1.0?

    22. GH

      Yeah.

    23. LF

      So what, uh, what were they trying to do to scale this whole process?

    24. GH

      So they were like, well, if we could run contracts elsewhere, um, and then only save the results of that computation, uh, you know, well, we don't actually have to do the compute on the chain. We can do the compute off chain and just post what the results are. Now the problem with that is, well, somebody could lie about what the results are. So you need a resolution mechanism. And the resolution mechanism can be really expensive, uh, because, you know, you just have to make sure that like the person who is saying, "Look, I swear that this is the real computation. I'm staking $10,000 on that fact." And if you prove it wrong, yeah, it might cost you $3,000 in gas fees to prove wrong, but you'll get the $10,000 bounty. So you can secure using those kind of systems. Um, so it's effectively a sandbox which runs contracts, uh, and like, just like any kind of normal sandbox, you have to like replace syscalls with, um, you know, calls into the hypervisor.

    25. LF

      Uh, sandbox, syscalls, hypervisor. What do these things mean?

    26. GH

      Uh, so like-

    27. LF

      As long as it's interesting to talk about.

    28. GH

      Yeah, I mean you can take like the, the Chrome sandbox is maybe the one to think about, right? So the Chrome process that's doing a rendering, uh, can't for example, read a file from the file system.

    29. LF

      Yeah.

    30. GH

      It has, if it tries to make an open syscall in Linux, the open syscall, you can't make it open syscall. No, no, no. Uh, you have to request from the kind of, uh, hypervisor process or like, uh...... what it's called in Chrome, but, um, the, kind of, "Hey, could you open this file for me?" And then it does all these checks and then it passes the file handle back in if it's approved.

  11. 39:3048:57

    Cryptocurrency

    1. GH

    2. LF

      What are your thoughts on crypto in general? So what, what parts technically or philosophically do you find especially beautiful maybe?

    3. GH

      Oh, I'm, I'm extremely bullish on crypto long term. Not any specific crypto project, but this idea of... well, two ideas. One, um, the Nakamoto consensus algorithm is, I think, one of the greatest innovations of the 21st century. This idea that people can reach consensus, you can reach a group consensus using a relatively straightforward algorithm, um, is wild. And, like, uh, you know, Satoshi Nakamoto, people always ask me who I look up to, it's like, "Whoever that is."

    4. LF

      Who do you think it is?

    5. GH

      I mean, I-

    6. LF

      Is it Elon Musk? Is it, is it you?

    7. GH

      (laughs) It is definitely not me, and I do not think it's Elon Musk. But yeah, this idea of, uh, groups reaching consensus in a decentralized yet formulaic way is one extremely powerful idea from crypto. Um, maybe the second idea is this idea of smart contracts. When you write a contract, uh, between two parties, any contract, um, this contract, if there are disputes, it's interpreted by lawyers. Lawyers are just really shitty overpaid interpreters. Imagine you had-

    8. LF

      Yeah.

    9. GH

      Let's talk about them in terms of a, in terms of like, let's compare a lawyer to Python, right? So lawyer... Well, okay.

    10. LF

      That's brilliant.

    11. GH

      So, yeah.

    12. LF

      Oh, I never thought of it that way. It's hilarious. (laughs)

    13. GH

      So, so Python, I'm paying, I'm paying, um, you know, even 10 cents an hour, I'll use the nice Azure machine. I can run Python for 10 cents an hour.

    14. LF

      Yeah.

    15. GH

      Um, lawyers cost $1000 an hour. So Python is, is, is, um, 10,000X, uh, better on that axis. Um, lawyers don't always return the same answer. Um, Python almost always does. Uh...

    16. LF

      Cost.

    17. GH

      Yeah, I mean, just, just, just cost, reliability. Everything about Python is so much better than lawyers. Um, so if you can make smart contracts, this whole concept of code is law, I, I love and I would love to live in a world where everybody accepted that fact.

    18. LF

      So, so may- maybe, uh, you can talk about what smart contracts are?

    19. GH

      So let's say, um, let's say, you know, we have a, uh, even something as simple as a safety deposit box, right? A safety deposit box that holds a million dollars. I have a contract with the bank that says two out of these three parties, uh, must, uh, be present to open the safety deposit box and get the money out. So that's a contract for the bank, and it's only as good as the bank and the lawyers, right? Let's say, you know, somebody dies, and now, oh, we're gonna go through a big legal dispute about whether, "Oh, well, was it in the will? Was it not in the will?" What, what... Like, it's just so messy and the cost to determine truth is so expensive.... versus a smart contract, which just uses cryptography to check if two out of three keys are present, well, I can look at that, and I can have certainty in the answer that it's going to return. And that's what all businesses want is certainty. You know, they say businesses don't care. Viacom, YouTube, YouTube is like, "Look, we don't care which way this lawsuit goes. Just please tell us, so we can have certainty."

    20. LF

      Yeah, I wonder how many agreements in this world, 'cause we're, we're talking about financial transactions only in this case, correct? The smart, the smart contracts.

    21. GH

      Oh, you can go to, you can go to anything. You can go, you can put a prenup in the Ethereum blockchain.

    22. LF

      (laughs) Mar-

    23. GH

      Hmm?

    24. LF

      A married smart contract-

    25. GH

      Sorry, divorce lawyers, sorry.

    26. LF

      ... or a relationship.

    27. GH

      You're gonna be replaced by Python. (laughs)

    28. LF

      (laughs) Uh, okay, so that's, uh ... So tha- that's, that's another beautiful idea. Do you think there's something that's appealing to you about any one specific implementation? So if you look 10, 20, 50 years down the line, do you see any, like, Bitcoin, Ethereum, any of the other hundreds of cryptocurrencies winning out? Is there ... Like, what's your intuition about the space? Are you just sitting back and watching the chaos and who cares what emerges?

    29. GH

      Oh, I don't, I don't speculate. I don't really care. I- I don't really care which one of these projects wins. Uh, I'm kind of in the Bitcoin is a meme coin camp. I mean, why does Bitcoin have value? It's technically kind of, you know ...

    30. LF

      Look, yeah.

  12. 48:5752:37

    Self-help advice

    1. LF

      it does seem that y- you're really knowledgeable about a lot of different topics. So the natural question, a bunch of people ask this, which is, uh, how do you keep learning new things? Do you have, like, practical advice if you were to introspect, like taking notes, allocate time-

    2. GH

      God.

    3. LF

      ... or you just mess around and just allow your curiosity to drive?

    4. GH

      I'll, I'll write these people a self-help book and I'll charge $67 for it.

    5. LF

      (laughs)

    6. GH

      I, I will, I will write-

    7. LF

      What's-

    8. GH

      ... I will write-

    9. LF

      ... what's chapter one?

    10. GH

      I will write on the cover of the self-help book, "All this advice is completely meaningless-"

    11. LF

      Yeah.

    12. GH

      "... but you're gonna be a sucker and buy this book anyway."

    13. LF

      Yeah.

    14. GH

      And the one lesson that I hope they take away from the book is that I can't give you a meaningful answer to that.

    15. LF

      That's interesting. Let me translate that-

    16. GH

      (laughs)

    17. LF

      ... is you haven't really thought about what it is you do, systematically. Because you could reduce it, and there's some people, I mean, I've met brilliant people that, um, I mean, this is really clear with athletes. Some are just, you know, the best in the world at something and they, they have zero interest in writing a se- like, a self-help book abo- or how to master this game. And then there are some athletes who become great coaches and they love the analysis, perhaps the over-analysis. And you right now, at least at your age, which is an interesting, you're in the middle of the battle, you're like the warriors that have zero interest in writing books. Uh, so you're in the middle of the battle, so you, you have... yeah.

    18. GH

      This is, this is a fair point. I do think I have a certain aversion to, um, this kind of deliberate, intentional way of living life.

    19. LF

      You eventually... the hilarity of this, especially since this is recorded, (laughs) it will have revealed beautifully the absurdity when you finally do publish this book. I guarantee you, you will.

    20. GH

      But it's not-

    21. LF

      The story of Comma AI could be, maybe it'll be a biography written about you. That'll be, that'll be better, I guess.

    22. GH

      Well, and you might be able to learn some cute lessons if you're starting a company like Comma AI from that book. But if you're asking generic questions like, "How do I be good at things?" Dude, I don't know. (laughs)

    23. LF

      Well, learn- I mean, the interesting-

    24. GH

      Do them a lot.

    25. LF

      (laughs) Do them a lot. But the interesting thing here is learning things outside of your current trajectory, which is what it feels like from an outsider's perspective. I mean, the, uh, you know, th- that, I don't know if there's a advice on that, but it, it is an interesting curiosity when you become really busy, you're running a company. (laughs)

    26. GH

      (laughs) Part-time. (laughs)

    27. LF

      Yeah. But like, there's a natural inclination and trend, like just the, the, the momentum of life carries you into a particular direction of wanting to focus and this kind of dispersion that curiosity can lead to gets harder and harder with time 'cause you're, you get really good at certain things and it sucks trying things that you're not good at, like trying to figure them out. And you do this with your live streams is you're on the fly figuring stuff out. You don't mind looking dumb.

    28. GH

      No.

    29. LF

      Like (laughs) you just figure it out, figure it out pretty quickly.

    30. GH

      Sometimes I try things and I don't figure them out quickly.

  13. 52:3754:30

    Comma.ai

    1. GH

    2. LF

      All right, let's talk about Comma AI. What's the mission of the company? Let's, like, look at the biggest picture.

    3. GH

      Oh, I have an exact statement. Solve self-driving cars while delivering shippable intermediaries.

    4. LF

      So long-term vision is have fully autonomous vehicles and make sure you're making money along the way.

    5. GH

      I think it doesn't really speak to money, but I can talk, I can talk about what solve self-driving cars means. Solve self-driving cars, of course, means, um, you're not building a new car, you're building a person replacement, uh, that person can sit in the driver's seat and drive you anywhere a person can drive with a human or better level of safety, speed, quality, comfort.

    6. LF

      And what's the second part of that?

    7. GH

      Delivering shippable intermediaries, um, is, well, it's a way to fund the company, that's true, but it's also a way to keep us honest. Uh, if you don't have that-

    8. LF

      Yeah.

    9. GH

      ... it is very easy with this technology to think you're making progress when you're not. I've heard it best described on Hacker News as you can set any arbitrary milestone, meet that milestone, and still be infinitely far away from solving self-driving cars. (laughs)

    10. LF

      So it's hard to have, like, real deadlines when you're like Cruise or a Waymo when, uh, you don't have, uh, revenue. Is that... I mean, is revenue essentially-

    11. GH

      Rev-

    12. LF

      ... th- the thing we're talking about here?

    13. GH

      Revenue is, is... capitalism is based around consent. Capitalism, the way that you get revenue as a co- uh, real capitalism, Comma's in the real capitalism camp, there's definitely scams out there, but real capitalism is based around consent.

    14. LF

      Yeah.

    15. GH

      It's based around this idea that, like, if we're getting revenue, it's because we're providing at least that much value to another person. When someone buys a $1,000 Comma 2 from us, we're providing them at least $1,000 of value or they wouldn't buy it.

  14. 54:301:03:19

    Comma two

    1. GH

    2. LF

      Brilliant. So can you give a whirlwind overview of the products that Comma AI provides, like, uh, throughout its history and today?

    3. GH

      I mean, yeah. The past ones aren't really that interesting. It's kind of just been refinement of the same idea. Uh, the real only product we sell today is the Comma 2.

    4. LF

      Which is a piece of hardware with cameras-

    5. GH

      Yeah. Um, so the Comma Two, I mean, you can think about it kind of like a person. Uh, you know, and future hardware will probably be even more and more person-like. Um, so it has, uh, you know, eyes, ears, a mouth, a brain, uh, and a way to interface with the car.

    6. LF

      Does it have consciousness? Just kidding. That was a trick question because... (laughs)

    7. GH

      I don't have consciousness either.

    8. LF

      Okay. (laughs)

    9. GH

      Uh, me and the Comma Two are the same.

    10. LF

      You're the same?

    11. GH

      I have a little more compute than it. It only has, like, the same compute as a-

    12. LF

      Oh, interesting.

    13. GH

      ... bee. Uh, you know.

    14. LF

      You're more efficient energy-wise for the compute you're doing.

    15. GH

      Far more efficient energy-wise.

    16. LF

      Huh.

    17. GH

      20 petaflops, 20 watts. Crazy.

    18. LF

      But you lack consciousness.

    19. GH

      Sure.

    20. LF

      Do you fear death? You do. You want immortality.

    21. GH

      Of course, I fear death.

    22. LF

      Does Comma AI fear death? I don't think so.

    23. GH

      Of course, it does. It very much fears. Well, it fears negative loss. Oh, yeah.

    24. LF

      (laughs) Okay. So Comma-

    25. GH

      Yeah.

    26. LF

      So Comma Two, when did that, that come out? That was a year ago? No. Two...

    27. GH

      Uh, early this year.

    28. LF

      Wow. Time... It feels like-

    29. GH

      Yeah.

    30. LF

      2020 feels like, uh, it's taken 10 years to get to the end of it.

  15. 1:03:191:12:22

    Tesla vs Comma.ai

    1. GH

      large.

    2. LF

      So th- that's an interesting distinction to draw. So if you look at the way Tesla's approaching the problem and the way you're approaching the problem, which is very different than the rest of the self-driving car world, so let's put them aside, is you're treating most... The driving task is a machine learning problem, and the way Tesla's approaching it is with a multitask learning where you break it, the task of driving into hundreds of different tasks, and you have this multi-headed neural network-

    3. GH

      Mm-hmm.

    4. LF

      ... that's very good at performing each task, and there, there's presumably something on top that's stitching stuff together in order to, uh, make controlled decisions, policy decisions about how you move the car. But what that allows you, there's a brilliance to this because it allows you to, um, master each task, like lane detection, uh, stop sign detection, uh, traffic light detection, uh, drivable area segmentation, uh, you know, vehicle, bicycle, pedestrian detection, uh, there's some localization tasks in there, also predicting, uh, like, yeah, predicting how the, the entities in the scene are gonna move. Like, everything is basically a machine learning task, where there is a classification, segmentation, prediction, and it's nice because you can have this entire engine, data engine that's mining for edge cases for each one of these tasks, and you can have people like engineers that are basically masters of that task, like become the best person in the world at, uh, as you talk about the cone guy for, uh, for, for Waymo.

    5. GH

      Yeah, yeah, the good old cone guy.

    6. LF

      The, the, becoming the best person in the world at, uh, cone, (laughs) at, uh, at cone detection. Uh, I, so there's, that's a compelling notion from a supervised learning perspective, aut- automating much of the process of edge case discovery and retraining a neural network for each of the individual reception tasks, and then you're looking at the machine learning in a more holistic way, uh, basically doing end-to-end learning on the driving task, sup- supervised, trained on the data of the actual driving of people that use Komma.ai, like actual human drivers doing manual control plus the moments of disengagement that, uh, maybe with some labeling could indicate the failure of the system. So you have the, you have a huge amount of data for positive control of the vehicle, like, successful control of the vehicle, both maintaining the lane as, as I think you're also working on longitudinal control of the vehicle, and then failure cases where the vehicle does something wrong that it n- needs disengagement. So, like, what, why do you think you're right and Tesla is wrong on this? A- and then (laughs) do you think, do you think you'll come around to the Tesla way or do you think Tesla will come around to, to your way?

    7. GH

      If you were to start a chess engine company, would you hire a bishop guy?

    8. LF

      See, y- we have, uh, this is Monday morning quarterbacking, is, uh, yes, probably.

    9. GH

      (laughs)

    10. LF

      So-

    11. GH

      Oh, oh, our rook guy. Oh, we stole the rook guy from that company. Oh, we're gonna have real good rooks.

    12. LF

      Well, there's not many pieces, right? You can, uh... Y- uh, there's not many guys and gals to hire. You just have a few that work on the bishop, a few that work on the rook.

    13. GH

      (laughs) But is that not ludicrous today to think about in, in a world of AlphaZero?

    14. LF

      But AlphaZero is a chess game. So the, the, the fundamental question is...... how hard is driving compared to chess? Because-

    15. GH

      Longer.

    16. LF

      So long term, end-to-end will be the right solution. The question is, how many years away is that?

    17. GH

      End to end's gonna be the only solution for level five.

    18. LF

      For the only way we get there.

    19. GH

      Of course. And of course, Tesla's gonna come around to my way. And if you're a rook guy out there, I'm sorry.

    20. LF

      (laughs) The cone guy.

    21. GH

      Cone guy.

    22. LF

      I don't know.

    23. GH

      We're gonna specialize each task. We're gonna really understand rook placement. Yeah.

    24. LF

      I understand the intuition you have. I mean, that, that, uh, is very compelling notion that we can learn the task end-to-end, like the same compelling notion you might have for natural language conversation. I'm not sure, 'cause one thing you sneaked in there is the assertion that it's impossible to get to level five without this kind of approach. I don't know if that's obvious. That-

    25. GH

      I don't know if that's obvious either. I, I don't actually, uh, mean that. I think that it is much easier-

    26. LF

      Much easier.

    27. GH

      ... to get to level five with an end-to-end approach. I think that the other approach is doable, but the magnitude of the engineering challenge may exceed what humanity's capable of.

    28. LF

      So, but what do you think of the Tesla data engine approach, which to me is an active learning task is, is kind of fascinating, is, so is breaking it down into these multiple tasks and mining their data constantly for, like, edge cases, for these different tasks.

    29. GH

      Yeah, but the tasks themselves are not being learned. This is feature engineering.

    30. LF

      Yeah, I mean, it's, it's, um, it's a higher obstruction level of feature engineering for the different tasks. It's task engineering in a sense.

  16. 1:12:221:26:03

    Driver monitoring

    1. LF

      Now, on that little transition, I mean, how do you make the transition work? Is, uh, is this where driver sensing comes in? Like, how do you make the... 'Cause you said 100 miles. Like, is, is there some sort of human factor psychology thing where people start to overtrust the system, all those kinds of effects? Once it gets better and better and better and better, they get lazier and lazier and lazier. Uh, is that... Like, how do you get that transition right?

    2. GH

      First off, our monitoring is already adaptive. O- our monitoring has already seen adaptive.

    3. LF

      Driver monitoring, is this the camera that's looking at the driver? You have an infrared camera in the...

    4. GH

      ... our policy for how we enforce the driver monitoring is scene-adaptive.

    5. LF

      What's that mean?

    6. GH

      Well, for example, in one of the extreme cases, um, if you, uh, if the car is not moving, we do not actively enforce driver monitoring. Right?

    7. LF

      Right.

    8. GH

      Um, if you are going through a, uh, like a 45 mile an hour road with lights, um, and stop signs and potentially pedestrians, we enforce a very tight driver monitoring policy. If you are alone on a perfectly straight highway, um, and this is, it's all machine learning, none of that is hand-coded. Actually, the stop is hand-coded, but...

    9. LF

      So there's some kind of machine learning estimation of risk?

    10. GH

      Yes.

    11. LF

      Yeah. I mean, I've always been a huge fan of that. That, that's, uh-

    12. GH

      Because-

    13. LF

      Uh, it's difficult to do every step into that direction is a worthwhile step to take. It might be difficult to do really what like us humans are able to est- estimate risk pretty damn well, whatever the hell that is. That feels like one of the nice features of us humans, uh, 'cause like we humans are really good drivers when we're really like-

    14. GH

      Mm-hmm.

    15. LF

      ... tuned in. And we're good at estimating risk like when are we supposed to be tuned in.

    16. GH

      Yeah. And, you know, people are like, "Oh, well, you know, why would you ever make the driver monitoring policy less aggressive? Why would you always not keep it at its most-"

    17. LF

      Yeah.

    18. GH

      "... aggressive?" Because then people are just gonna get fatigued from it.

    19. LF

      Yeah. When they get annoyed, you want them-

    20. GH

      Yeah.

    21. LF

      You, you want, you want the experience to be pleasant and-

    22. GH

      Well, obviously, I want the experience to be pleasant, but even just from a straight-up safety perspective.

    23. LF

      Hmm.

    24. GH

      If you alert people when they look around and they're like, "Why is this thing alerting me? There's nothing I could possibly hit right now," people will just learn to tune it out. People will just learn to, to tune it out, to put weights on the steering wheel, to do whatever to overcome it. And remember that you're always part of this adaptive system. So, all I can really say about, you know, how this scale is going forward is, yeah, something we have to monitor for. Ooh, we don't know. This is a great psychology experiment at scale, like we'll see.

    25. LF

      Yeah, it's fascinating.

    26. GH

      Track it. And making sure you have a good understanding of attention is a very key part of that psychology problem.

    27. LF

      Yeah. I think, I mean, you and I probably have a different, come to it differently, but to me, it's an, it's a fascinating psychology problem to explore something much deeper than just driving. It's, uh, it's such a nice way to explore human attention and human behavior, uh, which is why, again, we've probably both criticized Mr. Elon Musk on the-

    28. GH

      Yeah.

    29. LF

      ... on this one topic from different avenues, uh, so both offline and online, I had little chats with Elon and ... Like, I love human beings as a, as, um, as a computer vision problem, as an AI problem, it's fascinating. He wasn't so much interested in that problem. Is like in order to solve driving, the whole point is you want to remove the human from the picture.

    30. GH

      Mm-hmm.

Episode duration: 3:08:46

Install uListen for AI-powered chat & search across the full episode — Get Full Transcript

Transcript of episode _L3gNaAVjQ4

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.

Add to Chrome