Transcript
00:00:00 The following is a conversation with George Hotz,
00:00:02 AKA Geohot, his second time on the podcast.
00:00:06 He’s the founder of Comma AI,
00:00:09 an autonomous and semi autonomous vehicle technology company
00:00:12 that seeks to be to Tesla Autopilot
00:00:15 what Android is to the iOS.
00:00:18 They sell the Comma 2 device for $1,000
00:00:22 that when installed in many of their supported cars
00:00:25 can keep the vehicle centered in the lane
00:00:27 even when there are no lane markings.
00:00:30 It includes driver sensing
00:00:32 that ensures that the driver’s eyes are on the road.
00:00:35 As you may know, I’m a big fan of driver sensing.
00:00:38 I do believe Tesla Autopilot and others
00:00:40 should definitely include it in their sensor suite.
00:00:43 Also, I’m a fan of Android and a big fan of George
00:00:47 for many reasons,
00:00:48 including his nonlinear out of the box brilliance
00:00:51 and the fact that he’s a superstar programmer
00:00:55 of a very different style than myself.
00:00:57 Styles make fights and styles make conversations.
00:01:01 So I really enjoyed this chat
00:01:02 and I’m sure we’ll talk many more times on this podcast.
00:01:06 Quick mention of a sponsor
00:01:07 followed by some thoughts related to the episode.
00:01:10 First is Four Sigmatic,
00:01:12 the maker of delicious mushroom coffee.
00:01:15 Second is The Coding Digital,
00:01:17 a podcast on tech and entrepreneurship
00:01:19 that I listen to and enjoy.
00:01:22 And finally, ExpressVPN,
00:01:24 the VPN I’ve used for many years to protect my privacy
00:01:27 on the internet.
00:01:29 Please check out the sponsors in the description
00:01:31 to get a discount and to support this podcast.
00:01:34 As a side note, let me say that my work at MIT
00:01:38 on autonomous and semi autonomous vehicles
00:01:40 led me to study the human side of autonomy
00:01:43 enough to understand that it’s a beautifully complicated
00:01:46 and interesting problem space,
00:01:48 much richer than what can be studied in the lab.
00:01:51 In that sense, the data that Comma AI, Tesla Autopilot
00:01:55 and perhaps others like Cadillac Super Cruiser collecting
00:01:58 gives us a chance to understand
00:02:00 how we can design safe semi autonomous vehicles
00:02:03 for real human beings in real world conditions.
00:02:07 I think this requires bold innovation
00:02:09 and a serious exploration of the first principles
00:02:13 of the driving task itself.
00:02:15 If you enjoyed this thing, subscribe on YouTube,
00:02:17 review it with five stars and up a podcast,
00:02:20 follow on Spotify, support on Patreon
00:02:22 or connect with me on Twitter at Lex Friedman.
00:02:26 And now here’s my conversation with George Hotz.
00:02:31 So last time we started talking about the simulation,
00:02:34 this time let me ask you,
00:02:35 do you think there’s intelligent life out there
00:02:37 in the universe?
00:02:38 I’ve always maintained my answer to the Fermi paradox.
00:02:41 I think there has been intelligent life
00:02:44 elsewhere in the universe.
00:02:45 So intelligent civilizations existed
00:02:47 but they’ve blown themselves up.
00:02:49 So your general intuition is that
00:02:50 intelligent civilizations quickly,
00:02:54 like there’s that parameter in the Drake equation.
00:02:57 Your sense is they don’t last very long.
00:02:59 Yeah.
00:03:00 How are we doing on that?
00:03:01 Like, have we lasted pretty good?
00:03:03 Oh no.
00:03:04 Are we do?
00:03:05 Oh yeah.
00:03:06 I mean, not quite yet.
00:03:09 Well, it was good to tell you,
00:03:10 as you’d ask the IQ required to destroy the world
00:03:13 falls by one point every year.
00:03:15 Okay.
00:03:16 Technology democratizes the destruction of the world.
00:03:21 When can a meme destroy the world?
00:03:23 It kind of is already, right?
00:03:27 Somewhat.
00:03:28 I don’t think we’ve seen anywhere near the worst of it yet.
00:03:32 Well, it’s going to get weird.
00:03:34 Well, maybe a meme can save the world.
00:03:36 You thought about that?
00:03:37 The meme Lord Elon Musk fighting on the side of good
00:03:40 versus the meme Lord of the darkness,
00:03:44 which is not saying anything bad about Donald Trump,
00:03:48 but he is the Lord of the meme on the dark side.
00:03:51 He’s a Darth Vader of memes.
00:03:53 I think in every fairy tale they always end it with,
00:03:58 and they lived happily ever after.
00:03:59 And I’m like, please tell me more
00:04:00 about this happily ever after.
00:04:02 I’ve heard 50% of marriages end in divorce.
00:04:05 Why doesn’t your marriage end up there?
00:04:07 You can’t just say happily ever after.
00:04:09 So it’s the thing about destruction
00:04:12 is it’s over after the destruction.
00:04:14 We have to do everything right in order to avoid it.
00:04:18 And one thing wrong,
00:04:20 I mean, actually this is what I really like
00:04:21 about cryptography.
00:04:22 Cryptography, it seems like we live in a world
00:04:24 where the defense wins versus like nuclear weapons.
00:04:29 The opposite is true.
00:04:30 It is much easier to build a warhead
00:04:32 that splits into a hundred little warheads
00:04:34 than to build something that can, you know,
00:04:36 take out a hundred little warheads.
00:04:38 The offense has the advantage there.
00:04:41 So maybe our future is in crypto, but.
00:04:44 So cryptography, right.
00:04:45 The Goliath is the defense.
00:04:49 And then all the different hackers are the Davids.
00:04:54 And that equation is flipped for nuclear war.
00:04:57 Cause there’s so many,
00:04:58 like one nuclear weapon destroys everything essentially.
00:05:01 Yeah, and it is much easier to attack with a nuclear weapon
00:05:06 than it is to like the technology required to intercept
00:05:09 and destroy a rocket is much more complicated
00:05:12 than the technology required to just, you know,
00:05:13 orbital trajectory, send a rocket to somebody.
00:05:17 So, okay.
00:05:18 Your intuition that there were intelligent civilizations
00:05:21 out there, but it’s very possible
00:05:24 that they’re no longer there.
00:05:26 That’s kind of a sad picture.
00:05:27 They enter some steady state.
00:05:29 They all wirehead themselves.
00:05:31 What’s wirehead?
00:05:33 Stimulate, stimulate their pleasure centers
00:05:35 and just, you know, live forever in this kind of stasis.
00:05:39 They become, well, I mean,
00:05:42 I think the reason I believe this is because where are they?
00:05:46 If there’s some reason they stopped expanding,
00:05:50 cause otherwise they would have taken over the universe.
00:05:52 The universe isn’t that big.
00:05:53 Or at least, you know,
00:05:54 let’s just talk about the galaxy, right?
00:05:56 That’s 70,000 light years across.
00:05:58 I took that number from Star Trek Voyager.
00:05:59 I don’t know how true it is, but yeah, that’s not big.
00:06:04 Right? 70,000 light years is nothing.
00:06:07 For some possible technology that you can imagine
00:06:10 that can leverage like wormholes or something like that.
00:06:12 Or you don’t even need wormholes.
00:06:13 Just a von Neumann probe is enough.
00:06:15 A von Neumann probe and a million years of sublight travel
00:06:18 and you’d have taken over the whole universe.
00:06:20 That clearly didn’t happen.
00:06:22 So something stopped it.
00:06:24 So you mean if you, right,
00:06:25 for like a few million years,
00:06:27 if you sent out probes that travel close,
00:06:29 what’s sublight?
00:06:30 You mean close to the speed of light?
00:06:32 Let’s say 0.1 C.
00:06:33 And it just spreads.
00:06:34 Interesting.
00:06:35 Actually, that’s an interesting calculation, huh?
00:06:38 So what makes you think that we’d be able
00:06:40 to communicate with them?
00:06:42 Like, yeah, what’s,
00:06:45 why do you think we would be able to be able
00:06:47 to comprehend intelligent lives that are out there?
00:06:51 Like even if they were among us kind of thing,
00:06:54 like, or even just flying around?
00:06:57 Well, I mean, that’s possible.
00:07:01 It’s possible that there is some sort of prime directive.
00:07:04 That’d be a really cool universe to live in.
00:07:07 And there’s some reason
00:07:08 they’re not making themselves visible to us.
00:07:10 But it makes sense that they would use the same,
00:07:15 well, at least the same entropy.
00:07:16 Well, you’re implying the same laws of physics.
00:07:18 I don’t know what you mean by entropy in this case.
00:07:20 Oh, yeah.
00:07:21 I mean, if entropy is the scarce resource in the universe.
00:07:25 So what do you think about like Stephen Wolfram
00:07:26 and everything is a computation?
00:07:28 And then what if they are traveling through
00:07:31 this world of computation?
00:07:32 So if you think of the universe
00:07:34 as just information processing,
00:07:36 then what you’re referring to with entropy
00:07:40 and then these pockets of interesting complex computation
00:07:44 swimming around, how do we know they’re not already here?
00:07:47 How do we know that this,
00:07:51 like all the different amazing things
00:07:53 that are full of mystery on earth
00:07:55 are just like little footprints of intelligence
00:07:58 from light years away?
00:08:01 Maybe.
00:08:02 I mean, I tend to think that as civilizations expand,
00:08:05 they use more and more energy
00:08:07 and you can never overcome the problem of waste heat.
00:08:10 So where is there waste heat?
00:08:11 So we’d be able to, with our crude methods,
00:08:13 be able to see like, there’s a whole lot of energy here.
00:08:18 But it could be something we’re not,
00:08:20 I mean, we don’t understand dark energy, right?
00:08:22 Dark matter.
00:08:23 It could be just stuff we don’t understand at all.
00:08:26 Or they can have a fundamentally different physics,
00:08:29 you know, like that we just don’t even comprehend.
00:08:32 Well, I think, okay,
00:08:33 I mean, it depends how far out you wanna go.
00:08:35 I don’t think physics is very different
00:08:36 on the other side of the galaxy.
00:08:39 I would suspect that they have,
00:08:41 I mean, if they’re in our universe,
00:08:43 they have the same physics.
00:08:45 Well, yeah, that’s the assumption we have,
00:08:47 but there could be like super trippy things
00:08:50 like our cognition only gets to a slice,
00:08:57 and all the possible instruments that we can design
00:08:59 only get to a particular slice of the universe.
00:09:01 And there’s something much like weirder.
00:09:04 Maybe we can try a thought experiment.
00:09:06 Would people from the past
00:09:10 be able to detect the remnants of our,
00:09:14 or would we be able to detect our modern civilization?
00:09:16 I think the answer is obviously yes.
00:09:18 You mean past from a hundred years ago?
00:09:20 Well, let’s even go back further.
00:09:22 Let’s go to a million years ago, right?
00:09:24 The humans who were lying around in the desert
00:09:26 probably didn’t even have,
00:09:27 maybe they just barely had fire.
00:09:31 They would understand if a 747 flew overhead.
00:09:35 Oh, in this vicinity, but not if a 747 flew on Mars.
00:09:43 Like, cause they wouldn’t be able to see far,
00:09:45 cause we’re not actually communicating that well
00:09:47 with the rest of the universe.
00:09:48 We’re doing okay.
00:09:50 Just sending out random like fifties tracks of music.
00:09:54 True.
00:09:55 And yeah, I mean, they’d have to, you know,
00:09:57 we’ve only been broadcasting radio waves for 150 years.
00:10:02 And well, there’s your light cone.
00:10:04 So.
00:10:05 Yeah. Okay.
00:10:06 What do you make about all the,
00:10:08 I recently came across this having talked to David Fravor.
00:10:14 I don’t know if you caught what the videos
00:10:16 of the Pentagon released
00:10:18 and the New York Times reporting of the UFO sightings.
00:10:23 So I kind of looked into it, quote unquote.
00:10:26 And there’s actually been like hundreds
00:10:30 of thousands of UFO sightings, right?
00:10:33 And a lot of it you can explain away
00:10:35 in different kinds of ways.
00:10:37 So one is it could be interesting physical phenomena.
00:10:40 Two, it could be people wanting to believe
00:10:44 and therefore they conjure up a lot of different things
00:10:46 that just, you know, when you see different kinds of lights,
00:10:48 some basic physics phenomena,
00:10:50 and then you just conjure up ideas
00:10:53 of possible out there mysterious worlds.
00:10:56 But, you know, it’s also possible,
00:10:58 like you have a case of David Fravor,
00:11:02 who is a Navy pilot, who’s, you know,
00:11:06 as legit as it gets in terms of humans
00:11:08 who are able to perceive things in the environment
00:11:13 and make conclusions,
00:11:15 whether those things are a threat or not.
00:11:17 And he and several other pilots saw a thing,
00:11:22 I don’t know if you followed this,
00:11:23 but they saw a thing that they’ve since then called TikTok
00:11:26 that moved in all kinds of weird ways.
00:11:29 They don’t know what it is.
00:11:30 It could be technology developed by the United States
00:11:36 and they’re just not aware of it
00:11:38 and the surface level from the Navy, right?
00:11:40 It could be different kind of lighting technology
00:11:42 or drone technology, all that kind of stuff.
00:11:45 It could be the Russians and the Chinese,
00:11:46 all that kind of stuff.
00:11:48 And of course their mind, our mind,
00:11:51 can also venture into the possibility
00:11:54 that it’s from another world.
00:11:56 Have you looked into this at all?
00:11:58 What do you think about it?
00:11:59 I think all the news is a psyop.
00:12:01 I think that the most plausible.
00:12:05 Nothing is real.
00:12:06 Yeah, I listened to the, I think it was Bob Lazar
00:12:10 on Joe Rogan.
00:12:12 And like, I believe everything this guy is saying.
00:12:15 And then I think that it’s probably just some like MKUltra
00:12:18 kind of thing, you know?
00:12:20 What do you mean?
00:12:21 Like they, you know, they made some weird thing
00:12:24 and they called it an alien spaceship.
00:12:26 You know, maybe it was just to like
00:12:27 stimulate young physicists minds.
00:12:29 We’ll tell them it’s alien technology
00:12:31 and we’ll see what they come up with, right?
00:12:33 Do you find any conspiracy theories compelling?
00:12:36 Like have you pulled at the string
00:12:38 of the rich complex world of conspiracy theories
00:12:42 that’s out there?
00:12:43 I think that I’ve heard a conspiracy theory
00:12:46 that conspiracy theories were invented by the CIA
00:12:48 in the 60s to discredit true things.
00:12:52 Yeah.
00:12:53 So, you know, you can go to ridiculous conspiracy theories
00:12:58 like Flat Earth and Pizza Gate.
00:13:01 And, you know, these things are almost to hide
00:13:05 like conspiracy theories that like,
00:13:08 you know, remember when the Chinese like locked up
00:13:09 the doctors who discovered coronavirus?
00:13:11 Like I tell people this and I’m like,
00:13:12 no, no, no, that’s not a conspiracy theory.
00:13:14 That actually happened.
00:13:15 Do you remember the time that the money used to be backed
00:13:18 by gold and now it’s backed by nothing?
00:13:20 This is not a conspiracy theory.
00:13:21 This actually happened.
00:13:23 Well, that’s one of my worries today
00:13:26 with the idea of fake news is that when nothing is real,
00:13:32 then like you dilute the possibility of anything being true
00:13:37 by conjuring up all kinds of conspiracy theories.
00:13:41 And then you don’t know what to believe.
00:13:42 And then like the idea of truth of objectivity
00:13:46 is lost completely.
00:13:47 Everybody has their own truth.
00:13:50 So you used to control information by censoring it.
00:13:53 And then the internet happened and governments were like,
00:13:55 oh shit, we can’t censor things anymore.
00:13:58 I know what we’ll do.
00:14:00 You know, it’s the old story of the story of like
00:14:04 tying a flag with a leprechaun tells you his gold is buried
00:14:07 and you tie one flag and you make the leprechaun swear
00:14:09 to not remove the flag.
00:14:10 And you come back to the field later with a shovel
00:14:11 and there’s flags everywhere.
00:14:14 That’s one way to maintain privacy, right?
00:14:16 It’s like in order to protect the contents
00:14:20 of this conversation, for example,
00:14:21 we could just generate like millions of deep,
00:14:25 fake conversations where you and I talk
00:14:27 and say random things.
00:14:29 So this is just one of them
00:14:30 and nobody knows which one was the real one.
00:14:32 This could be fake right now.
00:14:34 Classic steganography technique.
00:14:37 Okay, another absurd question about intelligent life.
00:14:39 Cause you know, you’re an incredible programmer
00:14:43 outside of everything else we’ll talk about
00:14:45 just as a programmer.
00:14:49 Do you think intelligent beings out there,
00:14:52 the civilizations that were out there,
00:14:54 had computers and programming?
00:14:58 Did they, do we naturally have to develop something
00:15:01 where we engineer machines and are able to encode
00:15:05 both knowledge into those machines
00:15:08 and instructions that process that knowledge,
00:15:11 process that information to make decisions
00:15:14 and actions and so on?
00:15:15 And would those programming languages,
00:15:18 if you think they exist, be at all similar
00:15:21 to anything we’ve developed?
00:15:24 So I don’t see that much of a difference
00:15:26 between quote unquote natural languages
00:15:29 and programming languages.
00:15:34 Yeah.
00:15:35 I think there’s so many similarities.
00:15:36 So when asked the question,
00:15:39 what do alien languages look like?
00:15:42 I imagine they’re not all that dissimilar from ours.
00:15:46 And I think translating in and out of them
00:15:51 wouldn’t be that crazy.
00:15:52 Well, it’s difficult to compile like DNA to Python
00:15:57 and then to C.
00:15:59 There’s a little bit of a gap in the kind of languages
00:16:02 we use for touring machines
00:16:06 and the kind of languages nature seems to use a little bit.
00:16:10 Maybe that’s just, we just haven’t understood
00:16:13 the kind of language that nature uses well yet.
00:16:16 DNA is a CAD model.
00:16:19 It’s not quite a programming language.
00:16:21 It has no sort of a serial execution.
00:16:25 It’s not quite a, yeah, it’s a CAD model.
00:16:29 So I think in that sense,
00:16:30 we actually completely understand it.
00:16:32 The problem is, well, simulating on these CAD models,
00:16:37 I played with it a bit this year,
00:16:38 is super computationally intensive.
00:16:41 If you wanna go down to like the molecular level
00:16:43 where you need to go to see a lot of these phenomenon
00:16:45 like protein folding.
00:16:48 So yeah, it’s not that we don’t understand it.
00:16:52 It just requires a whole lot of compute to kind of compile it.
00:16:55 For our human minds, it’s inefficient,
00:16:56 both for the data representation and for the programming.
00:17:00 Yeah, it runs well on raw nature.
00:17:02 It runs well on raw nature.
00:17:03 And when we try to build emulators or simulators for that,
00:17:07 well, they’re mad slow, but I’ve tried it.
00:17:10 It runs in that, yeah, you’ve commented elsewhere,
00:17:14 I don’t remember where,
00:17:15 that one of the problems is simulating nature is tough.
00:17:20 And if you want to sort of deploy a prototype,
00:17:25 I forgot how you put it, but it made me laugh,
00:17:28 but animals or humans would need to be involved
00:17:31 in order to try to run some prototype code on,
00:17:38 like if we’re talking about COVID and viruses and so on,
00:17:41 if you were trying to engineer
00:17:42 some kind of defense mechanisms,
00:17:45 like a vaccine against COVID and all that kind of stuff
00:17:49 that doing any kind of experimentation,
00:17:52 like you can with like autonomous vehicles
00:17:53 would be very technically and ethically costly.
00:17:59 I’m not sure about that.
00:18:00 I think you can do tons of crazy biology and test tubes.
00:18:05 I think my bigger complaint is more,
00:18:08 oh, the tools are so bad.
00:18:11 Like literally, you mean like libraries and?
00:18:14 I don’t know, I’m not pipetting shit.
00:18:16 Like you’re handing me a, I got a, no, no, no,
00:18:20 there has to be some.
00:18:22 Like automating stuff.
00:18:24 And like the, yeah, but human biology is messy.
00:18:28 Like it seems.
00:18:29 But like, look at those Toronto’s videos.
00:18:31 They were a joke.
00:18:32 It’s like a little gantry.
00:18:33 It’s like little X, Y gantry,
00:18:34 high school science project with the pipette.
00:18:36 I’m like, really?
00:18:38 Gotta be something better.
00:18:39 You can’t build like nice microfluidics
00:18:41 and I can program the computation to bio interface.
00:18:45 I mean, this is gonna happen.
00:18:47 But like right now, if you are asking me
00:18:50 to pipette 50 milliliters of solution, I’m out.
00:18:54 This is so crude.
00:18:55 Yeah.
00:18:56 Okay, let’s get all the crazy out of the way.
00:18:59 So a bunch of people asked me,
00:19:02 since we talked about the simulation last time,
00:19:05 we talked about hacking the simulation.
00:19:06 Do you have any updates, any insights
00:19:09 about how we might be able to go about hacking simulation
00:19:13 if we indeed do live in a simulation?
00:19:17 I think a lot of people misinterpreted
00:19:19 the point of that South by talk.
00:19:22 The point of the South by talk
00:19:23 was not literally to hack the simulation.
00:19:26 I think that this is an idea is literally just,
00:19:33 I think theoretical physics.
00:19:34 I think that’s the whole goal, right?
00:19:39 You want your grand unified theory, but then, okay,
00:19:42 build a grand unified theory search for exploits, right?
00:19:45 I think we’re nowhere near actually there yet.
00:19:47 My hope with that was just more to like,
00:19:51 are you people kidding me
00:19:52 with the things you spend time thinking about?
00:19:54 Do you understand like kind of how small you are?
00:19:58 You are bytes and God’s computer, really?
00:20:02 And the things that people get worked up about, you know?
00:20:06 So basically, it was more a message
00:20:10 of we should humble ourselves.
00:20:12 That we get to, like what are we humans in this byte code?
00:20:19 Yeah, and not just humble ourselves,
00:20:22 but like I’m not trying to like make people guilty
00:20:24 or anything like that.
00:20:25 I’m trying to say like, literally,
00:20:27 look at what you are spending time on, right?
00:20:30 What are you referring to?
00:20:31 You’re referring to the Kardashians?
00:20:32 What are we talking about?
00:20:34 Twitter?
00:20:34 No, the Kardashians, everyone knows that’s kind of fun.
00:20:38 I’m referring more to like the economy, you know?
00:20:42 This idea that we gotta up our stock price.
00:20:50 Or what is the goal function of humanity?
00:20:55 You don’t like the game of capitalism?
00:20:57 Like you don’t like the games we’ve constructed
00:20:59 for ourselves as humans?
00:21:00 I’m a big fan of capitalism.
00:21:02 I don’t think that’s really the game we’re playing right now.
00:21:05 I think we’re playing a different game
00:21:07 where the rules are rigged.
00:21:10 Okay, which games are interesting to you
00:21:12 that we humans have constructed and which aren’t?
00:21:14 Which are productive and which are not?
00:21:18 Actually, maybe that’s the real point of the talk.
00:21:21 It’s like, stop playing these fake human games.
00:21:25 There’s a real game here.
00:21:26 We can play the real game.
00:21:28 The real game is, you know, nature wrote the rules.
00:21:31 This is a real game.
00:21:32 There still is a game to play.
00:21:35 But if you look at, sorry to interrupt,
00:21:36 I don’t know if you’ve seen the Instagram account,
00:21:38 nature is metal.
00:21:40 The game that nature seems to be playing
00:21:42 is a lot more cruel than we humans want to put up with.
00:21:47 Or at least we see it as cruel.
00:21:49 It’s like the bigger thing eats the smaller thing
00:21:53 and does it to impress another big thing
00:21:58 so it can mate with that thing.
00:22:00 And that’s it.
00:22:01 That seems to be the entirety of it.
00:22:04 Well, there’s no art, there’s no music,
00:22:07 there’s no comma AI, there’s no comma one,
00:22:10 no comma two, no George Hots with his brilliant talks
00:22:14 at South by Southwest.
00:22:17 I disagree, though.
00:22:17 I disagree that this is what nature is.
00:22:19 I think nature just provided basically a open world MMORPG.
00:22:26 And, you know, here it’s open world.
00:22:29 I mean, if that’s the game you want to play,
00:22:31 you can play that game.
00:22:32 But isn’t that beautiful?
00:22:33 I don’t know if you played Diablo.
00:22:35 They used to have, I think, cow level where it’s…
00:22:39 So everybody will go just, they figured out this,
00:22:44 like the best way to gain like experience points
00:22:48 is to just slaughter cows over and over and over.
00:22:52 And so they figured out this little sub game
00:22:55 within the bigger game that this is the most efficient way
00:22:58 to get experience points.
00:22:59 And everybody somehow agreed
00:23:01 that getting experience points in RPG context
00:23:04 where you always want to be getting more stuff,
00:23:06 more skills, more levels, keep advancing.
00:23:09 That seems to be good.
00:23:10 So might as well sacrifice actual enjoyment
00:23:14 of playing a game, exploring a world,
00:23:17 and spending like hundreds of hours of your time
00:23:21 at cow level.
00:23:22 I mean, the number of hours I spent in cow level,
00:23:26 I’m not like the most impressive person
00:23:28 because people have spent probably thousands of hours there,
00:23:30 but it’s ridiculous.
00:23:31 So that’s a little absurd game that brought me joy
00:23:35 in some weird dopamine drug kind of way.
00:23:37 So you don’t like those games.
00:23:40 You don’t think that’s us humans feeling the nature.
00:23:46 I think so.
00:23:47 And that was the point of the talk.
00:23:49 Yeah.
00:23:50 So how do we hack it then?
00:23:51 Well, I want to live forever.
00:23:52 And I want to live forever.
00:23:55 And this is the goal.
00:23:56 Well, that’s a game against nature.
00:23:59 Yeah, immortality is the good objective function to you?
00:24:03 I mean, start there and then you can do whatever else
00:24:05 you want because you got a long time.
00:24:07 What if immortality makes the game just totally not fun?
00:24:10 I mean, like, why do you assume immortality
00:24:13 is somehow a good objective function?
00:24:18 It’s not immortality that I want.
00:24:19 A true immortality where I could not die,
00:24:22 I would prefer what we have right now.
00:24:25 But I want to choose my own death, of course.
00:24:27 I don’t want nature to decide when I die,
00:24:29 I’m going to win.
00:24:30 I’m going to be you.
00:24:33 And then at some point, if you choose commit suicide,
00:24:36 like how long do you think you’d live?
00:24:41 Until I get bored.
00:24:43 See, I don’t think people like brilliant people like you
00:24:48 that really ponder living a long time
00:24:52 are really considering how meaningless life becomes.
00:24:58 Well, I want to know everything and then I’m ready to die.
00:25:03 As long as there’s…
00:25:04 Yeah, but why do you want,
00:25:05 isn’t it possible that you want to know everything
00:25:06 because it’s finite?
00:25:09 Like the reason you want to know quote unquote everything
00:25:12 is because you don’t have enough time to know everything.
00:25:16 And once you have unlimited time,
00:25:18 then you realize like, why do anything?
00:25:22 Like why learn anything?
00:25:25 I want to know everything and then I’m ready to die.
00:25:27 So you have, yeah.
00:25:28 It’s not a, like, it’s a terminal value.
00:25:30 It’s not in service of anything else.
00:25:34 I’m conscious of the possibility, this is not a certainty,
00:25:37 but the possibility of that engine of curiosity
00:25:41 that you’re speaking to is actually
00:25:47 a symptom of the finiteness of life.
00:25:49 Like without that finiteness, your curiosity would vanish.
00:25:55 Like a morning fog.
00:25:57 All right, cool.
00:25:57 Bukowski talked about love like that.
00:25:59 Then let me solve immortality
00:26:01 and let me change the thing in my brain
00:26:02 that reminds me of the fact that I’m immortal,
00:26:04 tells me that life is finite shit.
00:26:06 Maybe I’ll have it tell me that life ends next week.
00:26:09 Right?
00:26:10 I’m okay with some self manipulation like that.
00:26:12 I’m okay with deceiving myself.
00:26:14 Oh, Rika, changing the code.
00:26:17 Yeah, well, if that’s the problem, right?
00:26:18 If the problem is that I will no longer have that,
00:26:20 that curiosity, I’d like to have backup copies of myself,
00:26:24 which I check in with occasionally
00:26:27 to make sure they’re okay with the trajectory
00:26:29 and they can kind of override it.
00:26:31 Maybe a nice, like, I think of like those wave nets,
00:26:33 those like logarithmic go back to the copies.
00:26:35 Yeah, but sometimes it’s not reversible.
00:26:36 Like I’ve done this with video games.
00:26:39 Once you figure out the cheat code
00:26:41 or like you look up how to cheat old school,
00:26:43 like single player, it ruins the game for you.
00:26:46 Absolutely.
00:26:47 It ruins that feeling.
00:26:48 But again, that just means our brain manipulation
00:26:51 technology is not good enough yet.
00:26:53 Remove that cheat code from your brain.
00:26:54 Here you go.
00:26:55 So it’s also possible that if we figure out immortality,
00:27:00 that all of us will kill ourselves
00:27:03 before we advance far enough
00:27:06 to be able to revert the change.
00:27:08 I’m not killing myself till I know everything, so.
00:27:11 That’s what you say now, because your life is finite.
00:27:15 You know, I think yes, self modifying systems gets,
00:27:19 comes up with all these hairy complexities
00:27:21 and can I promise that I’ll do it perfectly?
00:27:23 No, but I think I can put good safety structures in place.
00:27:27 So that talk and your thinking here
00:27:29 is not literally referring to a simulation
00:27:36 and that our universe is a kind of computer program
00:27:40 running on a computer.
00:27:42 That’s more of a thought experiment.
00:27:45 Do you also think of the potential of the sort of Bostrom,
00:27:51 Elon Musk and others that talk about an actual program
00:27:57 that simulates our universe?
00:27:59 Oh, I don’t doubt that we’re in a simulation.
00:28:01 I just think that it’s not quite that important.
00:28:05 I mean, I’m interested only in simulation theory
00:28:06 as far as like it gives me power over nature.
00:28:09 If it’s totally unfalsifiable, then who cares?
00:28:13 I mean, what do you think that experiment would look like?
00:28:15 Like somebody on Twitter asks,
00:28:17 asks George what signs we would look for
00:28:20 to know whether or not we’re in the simulation,
00:28:22 which is exactly what you’re asking is like,
00:28:25 the step that precedes the step of knowing
00:28:29 how to get more power from this knowledge
00:28:32 is to get an indication that there’s some power to be gained.
00:28:35 So get an indication that there,
00:28:37 you can discover and exploit cracks in the simulation
00:28:42 or it doesn’t have to be in the physics of the universe.
00:28:45 Yeah.
00:28:46 Show me, I mean, like a memory leak could be cool.
00:28:51 Like some scrying technology, you know?
00:28:54 What kind of technology?
00:28:55 Scrying?
00:28:56 What’s that?
00:28:57 Oh, that’s a weird,
00:28:58 scrying is the paranormal ability to like remote viewing,
00:29:03 like being able to see somewhere where you’re not.
00:29:08 So, you know, I don’t think you can do it
00:29:10 by chanting in a room,
00:29:11 but if we could find, it’s a memory leak, basically.
00:29:16 It’s a memory leak.
00:29:17 Yeah, you’re able to access parts you’re not supposed to.
00:29:19 Yeah, yeah, yeah.
00:29:20 And thereby discover a shortcut.
00:29:22 Yeah, maybe memory leak means the other thing as well,
00:29:24 but I mean like, yeah,
00:29:25 like an ability to read arbitrary memory, right?
00:29:28 And that one’s not that horrifying, right?
00:29:29 The right ones start to be horrifying.
00:29:31 Read, right.
00:29:32 It’s the reading is not the problem.
00:29:34 Yeah, it’s like Heartfleet for the universe.
00:29:37 Oh boy, the writing is a big, big problem.
00:29:40 It’s a big problem.
00:29:43 It’s the moment you can write anything,
00:29:44 even if it’s just random noise.
00:29:47 That’s terrifying.
00:29:49 I mean, even without that,
00:29:51 like even some of the, you know,
00:29:52 the nanotech stuff that’s coming, I think is.
00:29:57 I don’t know if you’re paying attention,
00:29:58 but actually Eric Weistand came out
00:30:00 with the theory of everything.
00:30:02 I mean, that came out.
00:30:03 He’s been working on a theory of everything
00:30:05 in the physics world called geometric unity.
00:30:08 And then for me, from computer science person like you,
00:30:11 Steven Wolfram’s theory of everything,
00:30:14 of like hypergraphs is super interesting and beautiful,
00:30:17 but not from a physics perspective,
00:30:19 but from a computational perspective.
00:30:20 I don’t know, have you paid attention to any of that?
00:30:23 So again, like what would make me pay attention
00:30:26 and like why like I hate string theory is,
00:30:29 okay, make a testable prediction, right?
00:30:31 I’m only interested in,
00:30:33 I’m not interested in theories for their intrinsic beauty.
00:30:36 I’m interested in theories
00:30:37 that give me power over the universe.
00:30:39 So if these theories do, I’m very interested.
00:30:43 Can I just say how beautiful that is?
00:30:45 Because a lot of physicists say,
00:30:47 I’m interested in experimental validation
00:30:49 and they skip out the part where they say
00:30:52 to give me more power in the universe.
00:30:55 I just love the.
00:30:57 No, I want. The clarity of that.
00:30:59 I want 100 gigahertz processors.
00:31:02 I want transistors that are smaller than atoms.
00:31:04 I want like power.
00:31:08 That’s true.
00:31:10 And that’s where people from aliens
00:31:12 to this kind of technology where people are worried
00:31:14 that governments, like who owns that power?
00:31:19 Is it George Harts?
00:31:20 Is it thousands of distributed hackers across the world?
00:31:25 Is it governments?
00:31:26 Is it Mark Zuckerberg?
00:31:28 There’s a lot of people that,
00:31:32 I don’t know if anyone trusts any one individual with power.
00:31:35 So they’re always worried.
00:31:37 It’s the beauty of blockchains.
00:31:39 That’s the beauty of blockchains, which we’ll talk about.
00:31:43 On Twitter, somebody pointed me to a story,
00:31:46 a bunch of people pointed me to a story a few months ago
00:31:49 where you went into a restaurant in New York.
00:31:51 And you can correct me if any of this is wrong.
00:31:53 And ran into a bunch of folks from a company
00:31:56 in a crypto company who are trying to scale up Ethereum.
00:32:01 And they had a technical deadline
00:32:03 related to solidity to OVM compiler.
00:32:07 So these are all Ethereum technologies.
00:32:09 So you stepped in, they recognized you,
00:32:14 pulled you aside, explained their problem.
00:32:16 And you stepped in and helped them solve the problem,
00:32:19 thereby creating legend status story.
00:32:23 So can you tell me the story in a little more detail?
00:32:28 It seems kind of incredible.
00:32:31 Did this happen?
00:32:32 Yeah, yeah, it’s a true story, it’s a true story.
00:32:34 I mean, they wrote a very flattering account of it.
00:32:39 So Optimism is the company called Optimism,
00:32:43 spin off of Plasma.
00:32:45 They’re trying to build L2 solutions on Ethereum.
00:32:47 So right now, every Ethereum node
00:32:52 has to run every transaction on the Ethereum network.
00:32:56 And this kind of doesn’t scale, right?
00:32:58 Because if you have N computers,
00:33:00 well, if that becomes two N computers,
00:33:02 you actually still get the same amount of compute, right?
00:33:05 This is like all of one scaling
00:33:09 because they all have to run it.
00:33:10 Okay, fine, you get more blockchain security,
00:33:12 but like, blockchain is already so secure.
00:33:15 Can we trade some of that off for speed?
00:33:17 So that’s kind of what these L2 solutions are.
00:33:20 They built this thing, which kind of,
00:33:23 kind of sandbox for Ethereum contracts.
00:33:26 So they can run it in this L2 world
00:33:28 and it can’t do certain things in L world, in L1.
00:33:30 Can I ask you for some definitions?
00:33:32 What’s L2?
00:33:33 Oh, L2 is layer two.
00:33:34 So L1 is like the base Ethereum chain.
00:33:37 And then layer two is like a computational layer
00:33:40 that runs elsewhere,
00:33:44 but still is kind of secured by layer one.
00:33:47 And I’m sure a lot of people know,
00:33:49 but Ethereum is a cryptocurrency,
00:33:51 probably one of the most popular cryptocurrency
00:33:53 second to Bitcoin.
00:33:55 And a lot of interesting technological innovation there.
00:33:58 Maybe you could also slip in whenever you talk about this
00:34:03 and things that are exciting to you in the Ethereum space.
00:34:06 And why Ethereum?
00:34:07 Well, I mean, Bitcoin is not Turing complete.
00:34:12 Ethereum is not technically Turing complete
00:34:13 with the gas limit, but close enough.
00:34:16 With the gas limit?
00:34:16 What’s the gas limit, resources?
00:34:19 Yeah, I mean, no computer is actually Turing complete.
00:34:21 Right.
00:34:23 You’re gonna find out RAM, you know?
00:34:24 I can actually solve the whole thing.
00:34:25 What’s the word gas limit?
00:34:26 You just have so many brilliant words.
00:34:28 I’m not even gonna ask.
00:34:29 That’s not my word, that’s Ethereum’s word.
00:34:32 Gas limit.
00:34:33 Ethereum, you have to spend gas per instruction.
00:34:35 So like different op codes use different amounts of gas
00:34:37 and you buy gas with ether to prevent people
00:34:40 from basically DDoSing the network.
00:34:42 So Bitcoin is proof of work.
00:34:45 And then what’s Ethereum?
00:34:47 It’s also proof of work.
00:34:48 They’re working on some proof of stake,
00:34:49 Ethereum 2.0 stuff.
00:34:51 But right now it’s proof of work.
00:34:52 It uses a different hash function from Bitcoin.
00:34:54 That’s more ASIC resistance, because you need RAM.
00:34:57 So we’re all talking about Ethereum 1.0.
00:34:59 So what were they trying to do to scale this whole process?
00:35:03 So they were like, well, if we could run contracts elsewhere
00:35:07 and then only save the results of that computation,
00:35:13 well, we don’t actually have to do the compute on the chain.
00:35:14 We can do the compute off chain
00:35:15 and just post what the results are.
00:35:17 Now, the problem with that is,
00:35:18 well, somebody could lie about what the results are.
00:35:21 So you need a resolution mechanism.
00:35:23 And the resolution mechanism can be really expensive
00:35:26 because you just have to make sure
00:35:29 that the person who is saying,
00:35:31 look, I swear that this is the real computation.
00:35:33 I’m staking $10,000 on that fact.
00:35:36 And if you prove it wrong,
00:35:39 yeah, it might cost you $3,000 in gas fees to prove wrong,
00:35:42 but you’ll get the $10,000 bounty.
00:35:44 So you can secure using those kinds of systems.
00:35:47 So it’s effectively a sandbox, which runs contracts.
00:35:52 And like, it’s like any kind of normal sandbox,
00:35:55 you have to like replace syscalls
00:35:57 with calls into the hypervisor.
00:36:02 Sandbox, syscalls, hypervisor.
00:36:05 What do these things mean?
00:36:06 As long as it’s interesting to talk about.
00:36:09 Yeah, I mean, you can take like the Chrome sandbox
00:36:11 is maybe the one to think about, right?
00:36:12 So the Chrome process that’s doing a rendering,
00:36:16 can’t, for example, read a file from the file system.
00:36:18 It has, if it tries to make an open syscall in Linux,
00:36:21 the open syscall, you can’t make it open syscall,
00:36:23 no, no, no.
00:36:24 You have to request from the kind of hypervisor process
00:36:29 or like, I don’t know what it’s called in Chrome,
00:36:31 but the, hey, could you open this file for me?
00:36:36 And then it does all these checks
00:36:37 and then it passes the file handle back in
00:36:39 if it’s approved.
00:36:41 So that’s, yeah.
00:36:42 So what’s the, in the context of Ethereum,
00:36:45 what are the boundaries of the sandbox
00:36:47 that we’re talking about?
00:36:48 Well, like one of the calls that you,
00:36:50 actually reading and writing any state
00:36:53 to the Ethereum contract,
00:36:55 or to the Ethereum blockchain.
00:36:58 Writing state is one of those calls
00:37:01 that you’re going to have to sandbox in layer two,
00:37:04 because if you let layer two just arbitrarily write
00:37:08 to the Ethereum blockchain.
00:37:09 So layer two is really sitting on top of layer one.
00:37:15 So you’re going to have a lot of different kinds of ideas
00:37:17 that you can play with.
00:37:18 And they’re all, they’re not fundamentally changing
00:37:21 the source code level of Ethereum.
00:37:25 Well, you have to replace a bunch of calls
00:37:28 with calls into the hypervisor.
00:37:31 So instead of doing the syscall directly,
00:37:33 you replace it with a call to the hypervisor.
00:37:37 So originally they were doing this
00:37:39 by first running the, so Solidity is the language
00:37:43 that most Ethereum contracts are written in.
00:37:45 It compiles to a bytecode.
00:37:47 And then they wrote this thing they called the transpiler.
00:37:50 And the transpiler took the bytecode
00:37:52 and it transpiled it into OVM safe bytecode.
00:37:56 Basically bytecode that didn’t make any
00:37:57 of those restricted syscalls
00:37:58 and added the calls to the hypervisor.
00:38:01 This transpiler was a 3000 line mess.
00:38:05 And it’s hard to do.
00:38:07 It’s hard to do if you’re trying to do it like that,
00:38:09 because you have to kind of like deconstruct the bytecode,
00:38:12 change things about it, and then reconstruct it.
00:38:15 And I mean, as soon as I hear this, I’m like,
00:38:17 well, why don’t you just change the compiler, right?
00:38:20 Why not the first place you build the bytecode,
00:38:22 just do it in the compiler.
00:38:25 So yeah, I asked them how much they wanted it.
00:38:29 Of course, measured in dollars and I’m like, well, okay.
00:38:33 And yeah.
00:38:34 And you wrote the compiler.
00:38:35 Yeah, I modified, I wrote a 300 line diff to the compiler.
00:38:39 It’s open source, you can look at it.
00:38:40 Yeah, it’s, yeah, I looked at the code last night.
00:38:43 It’s, yeah, exactly.
00:38:46 Cute is a good word for it.
00:38:49 And it’s C++.
00:38:52 C++, yeah.
00:38:54 So when asked how you were able to do it,
00:38:57 you said, you just gotta think and then do it right.
00:39:03 So can you break that apart a little bit?
00:39:04 What’s your process of one, thinking and two, doing it right?
00:39:09 You know, the people that I was working for
00:39:12 were amused that I said that.
00:39:13 It doesn’t really mean anything.
00:39:14 Okay.
00:39:16 I mean, is there some deep, profound insights
00:39:19 to draw from like how you problem solve from that?
00:39:23 This is always what I say.
00:39:24 I’m like, do you wanna be a good programmer?
00:39:26 Do it for 20 years.
00:39:27 Yeah, there’s no shortcuts.
00:39:29 No.
00:39:31 What are your thoughts on crypto in general?
00:39:33 So what parts technically or philosophically
00:39:38 do you find especially beautiful maybe?
00:39:40 Oh, I’m extremely bullish on crypto longterm.
00:39:42 Not any specific crypto project, but this idea of,
00:39:48 well, two ideas.
00:39:50 One, the Nakamoto Consensus Algorithm
00:39:54 is I think one of the greatest innovations
00:39:57 of the 21st century.
00:39:58 This idea that people can reach consensus.
00:40:01 You can reach a group consensus.
00:40:03 Using a relatively straightforward algorithm is wild.
00:40:08 And like, you know, Satoshi Nakamoto,
00:40:14 people always ask me who I look up to.
00:40:15 It’s like, whoever that is.
00:40:17 Who do you think it is?
00:40:19 I mean, Elon Musk?
00:40:21 Is it you?
00:40:22 It is definitely not me.
00:40:24 And I do not think it’s Elon Musk.
00:40:26 But yeah, this idea of groups reaching consensus
00:40:31 in a decentralized yet formulaic way
00:40:34 is one extremely powerful idea from crypto.
00:40:40 Maybe the second idea is this idea of smart contracts.
00:40:45 When you write a contract between two parties,
00:40:49 any contract, this contract, if there are disputes,
00:40:53 it’s interpreted by lawyers.
00:40:56 Lawyers are just really shitty overpaid interpreters.
00:41:00 Imagine you had, let’s talk about them in terms of a,
00:41:02 in terms of like, let’s compare a lawyer to Python, right?
00:41:05 So lawyer, well, okay.
00:41:07 That’s really, I never thought of it that way.
00:41:10 It’s hilarious.
00:41:11 So Python, I’m paying even 10 cents an hour.
00:41:15 I’ll use the nice Azure machine.
00:41:17 I can run Python for 10 cents an hour.
00:41:19 Lawyers cost $1,000 an hour.
00:41:21 So Python is 10,000 X better on that axis.
00:41:25 Lawyers don’t always return the same answer.
00:41:31 Python almost always does.
00:41:36 Cost.
00:41:37 Yeah, I mean, just cost, reliability,
00:41:40 everything about Python is so much better than lawyers.
00:41:43 So if you can make smart contracts,
00:41:46 this whole concept of code is law.
00:41:50 I love, and I would love to live in a world
00:41:53 where everybody accepted that fact.
00:41:55 So maybe you can talk about what smart contracts are.
00:42:01 So let’s say, let’s say, you know,
00:42:05 we have a, even something as simple
00:42:08 as a safety deposit box, right?
00:42:11 Safety deposit box that holds a million dollars.
00:42:14 I have a contract with the bank that says
00:42:17 two out of these three parties must be present
00:42:22 to open the safety deposit box and get the money out.
00:42:25 So that’s a contract for the bank,
00:42:26 and it’s only as good as the bank and the lawyers, right?
00:42:29 Let’s say, you know, somebody dies and now,
00:42:32 oh, we’re gonna go through a big legal dispute
00:42:34 about whether, oh, well, was it in the will,
00:42:36 was it not in the will?
00:42:37 What, like, it’s just so messy,
00:42:39 and the cost to determine truth is so expensive.
00:42:44 Versus a smart contract, which just uses cryptography
00:42:47 to check if two out of three keys are present.
00:42:50 Well, I can look at that, and I can have certainty
00:42:53 in the answer that it’s going to return.
00:42:55 And that’s what, all businesses want is certainty.
00:42:58 You know, they say businesses don’t care.
00:42:59 Viacom, YouTube, YouTube’s like,
00:43:02 look, we don’t care which way this lawsuit goes.
00:43:04 Just please tell us so we can have certainty.
00:43:07 Yeah, I wonder how many agreements in this,
00:43:09 because we’re talking about financial transactions only
00:43:12 in this case, correct, the smart contracts?
00:43:15 Oh, you can go to anything.
00:43:17 You can put a prenup in the Ethereum blockchain.
00:43:21 A married smart contract?
00:43:23 Sorry, divorce lawyer, sorry.
00:43:24 You’re going to be replaced by Python.
00:43:29 Okay, so that’s another beautiful idea.
00:43:34 Do you think there’s something that’s appealing to you
00:43:37 about any one specific implementation?
00:43:40 So if you look 10, 20, 50 years down the line,
00:43:45 do you see any, like, Bitcoin, Ethereum,
00:43:48 any of the other hundreds of cryptocurrencies winning out?
00:43:51 Is there, like, what’s your intuition about the space?
00:43:53 Or are you just sitting back and watching the chaos
00:43:55 and look who cares what emerges?
00:43:57 Oh, I don’t.
00:43:58 I don’t speculate.
00:43:59 I don’t really care.
00:43:59 I don’t really care which one of these projects wins.
00:44:02 I’m kind of in the Bitcoin as a meme coin camp.
00:44:05 I mean, why does Bitcoin have value?
00:44:07 It’s technically kind of, you know,
00:44:11 not great, like the block size debate.
00:44:14 When I found out what the block size debate was,
00:44:16 I’m like, are you guys kidding?
00:44:18 What’s the block size debate?
00:44:21 You know what?
00:44:22 It’s really, it’s too stupid to even talk.
00:44:23 People can look it up, but I’m like, wow.
00:44:27 You know, Ethereum seems,
00:44:28 the governance of Ethereum seems much better.
00:44:31 I’ve come around a bit on proof of stake ideas.
00:44:35 You know, very smart people thinking about some things.
00:44:37 Yeah, you know, governance is interesting.
00:44:40 It does feel like Vitalik,
00:44:44 like it does feel like an open,
00:44:46 even in these distributed systems,
00:44:48 leaders are helpful
00:44:51 because they kind of help you drive the mission
00:44:54 and the vision and they put a face to a project.
00:44:58 It’s a weird thing about us humans.
00:45:00 Geniuses are helpful, like Vitalik.
00:45:02 Yeah, brilliant.
00:45:06 Leaders are not necessarily, yeah.
00:45:10 So you think the reason he’s the face of Ethereum
00:45:15 is because he’s a genius.
00:45:17 That’s interesting.
00:45:18 I mean, that was,
00:45:21 it’s interesting to think about
00:45:22 that we need to create systems
00:45:25 in which the quote unquote leaders that emerge
00:45:30 are the geniuses in the system.
00:45:33 I mean, that’s arguably why
00:45:35 the current state of democracy is broken
00:45:36 is the people who are emerging as the leaders
00:45:39 are not the most competent,
00:45:40 are not the superstars of the system.
00:45:43 And it seems like at least for now
00:45:45 in the crypto world oftentimes
00:45:47 the leaders are the superstars.
00:45:49 Imagine at the debate they asked,
00:45:51 what’s the sixth amendment?
00:45:53 What are the four fundamental forces in the universe?
00:45:56 What’s the integral of two to the X?
00:45:59 I’d love to see those questions asked
00:46:01 and that’s what I want as our leader.
00:46:03 It’s a little bit.
00:46:04 What’s Bayes rule?
00:46:07 Yeah, I mean, even, oh wow, you’re hurting my brain.
00:46:10 It’s that my standard was even lower
00:46:15 but I would have loved to see
00:46:17 just this basic brilliance.
00:46:20 Like I’ve talked to historians.
00:46:22 There’s just these, they’re not even like
00:46:23 they don’t have a PhD or even education history.
00:46:26 They just like a Dan Carlin type character
00:46:30 who just like, holy shit.
00:46:32 How did all this information get into your head?
00:46:35 They’re able to just connect Genghis Khan
00:46:38 to the entirety of the history of the 20th century.
00:46:41 They know everything about every single battle that happened
00:46:46 and they know the game of Thrones
00:46:51 of the different power plays and all that happened there.
00:46:55 And they know like the individuals
00:46:56 and all the documents involved
00:46:58 and they integrate that into their regular life.
00:47:02 It’s not like they’re ultra history nerds.
00:47:03 They’re just, they know this information.
00:47:06 That’s what competence looks like.
00:47:08 Yeah.
00:47:09 Cause I’ve seen that with programmers too, right?
00:47:10 That’s what great programmers do.
00:47:12 But yeah, it would be, it’s really unfortunate
00:47:15 that those kinds of people aren’t emerging as our leaders.
00:47:19 But for now, at least in the crypto world
00:47:21 that seems to be the case.
00:47:23 I don’t know if that always, you could imagine
00:47:26 that in a hundred years, it’s not the case, right?
00:47:28 Crypto world has one very powerful idea going for it
00:47:31 and that’s the idea of forks, right?
00:47:35 I mean, imagine, we’ll use a less controversial example.
00:47:42 This was actually in my joke app in 2012.
00:47:47 I was like, Barack Obama, Mitt Romney,
00:47:49 let’s let them both be president, right?
00:47:51 Like imagine we could fork America
00:47:52 and just let them both be president.
00:47:54 And then the Americas could compete
00:47:56 and people could invest in one,
00:47:58 pull their liquidity out of one, put it in the other.
00:48:00 You have this in the crypto world.
00:48:02 Ethereum forks into Ethereum and Ethereum classic.
00:48:05 And you can pull your liquidity out of one
00:48:07 and put it in another.
00:48:08 And people vote with their dollars,
00:48:11 which forks, companies should be able to fork.
00:48:16 I’d love to fork Nvidia, you know?
00:48:20 Yeah, like different business strategies
00:48:22 and then try them out and see what works.
00:48:26 Like even take, yeah, take comma AI that closes its source
00:48:34 and then take one that’s open source and see what works.
00:48:38 Take one that’s purchased by GM
00:48:41 and one that remains Android Renegade
00:48:43 and all these different versions and see.
00:48:45 The beauty of comma AI is someone can actually do that.
00:48:47 Please take comma AI and fork it.
00:48:50 That’s right, that’s the beauty of open source.
00:48:53 So you’re, I mean, we’ll talk about autonomous vehicle space,
00:48:56 but it does seem that you’re really knowledgeable
00:49:02 about a lot of different topics.
00:49:03 So the natural question a bunch of people ask this,
00:49:06 which is how do you keep learning new things?
00:49:09 Do you have like practical advice
00:49:12 if you were to introspect, like taking notes,
00:49:15 allocate time, or do you just mess around
00:49:19 and just allow your curiosity to drive?
00:49:21 I’ll write these people a self help book
00:49:23 and I’ll charge $67 for it.
00:49:25 And I will write, I will write,
00:49:28 I will write on the cover of the self help book.
00:49:30 All of this advice is completely meaningless.
00:49:32 You’re gonna be a sucker and buy this book anyway.
00:49:34 And the one lesson that I hope they take away from the book
00:49:38 is that I can’t give you a meaningful answer to that.
00:49:42 That’s interesting.
00:49:44 Let me translate that.
00:49:45 Is you haven’t really thought about what it is you do
00:49:51 systematically because you could reduce it.
00:49:53 And there’s some people, I mean, I’ve met brilliant people
00:49:56 that this is really clear with athletes.
00:50:00 Some are just, you know, the best in the world
00:50:03 at something and they have zero interest
00:50:06 in writing like a self help book,
00:50:09 or how to master this game.
00:50:11 And then there’s some athletes who become great coaches
00:50:15 and they love the analysis, perhaps the over analysis.
00:50:18 And you right now, at least at your age,
00:50:20 which is an interesting, you’re in the middle of the battle.
00:50:23 You’re like the warriors that have zero interest
00:50:25 in writing books.
00:50:27 So you’re in the middle of the battle.
00:50:29 So you have, yeah.
00:50:30 This is a fair point.
00:50:31 I do think I have a certain aversion
00:50:34 to this kind of deliberate intentional way of living life.
00:50:39 You’re eventually, the hilarity of this,
00:50:41 especially since this is recorded,
00:50:43 it will reveal beautifully the absurdity
00:50:47 when you finally do publish this book.
00:50:49 I guarantee you, you will.
00:50:51 The story of comma AI, maybe it’ll be a biography
00:50:56 written about you.
00:50:57 That’ll be better, I guess.
00:50:58 And you might be able to learn some cute lessons
00:51:00 if you’re starting a company like comma AI from that book.
00:51:03 But if you’re asking generic questions,
00:51:05 like how do I be good at things?
00:51:07 How do I be good at things?
00:51:10 Dude, I don’t know.
00:51:11 Do them a lot.
00:51:14 Do them a lot.
00:51:15 But the interesting thing here is learning things
00:51:18 outside of your current trajectory,
00:51:22 which is what it feels like from an outsider’s perspective.
00:51:28 I don’t know if there’s advice on that,
00:51:30 but it is an interesting curiosity.
00:51:33 When you become really busy, you’re running a company.
00:51:38 Hard time.
00:51:40 Yeah.
00:51:41 But there’s a natural inclination and trend.
00:51:46 Just the momentum of life carries you
00:51:48 into a particular direction of wanting to focus.
00:51:51 And this kind of dispersion that curiosity can lead to
00:51:55 gets harder and harder with time.
00:51:58 Because you get really good at certain things
00:52:00 and it sucks trying things that you’re not good at,
00:52:03 like trying to figure them out.
00:52:05 When you do this with your live streams,
00:52:07 you’re on the fly figuring stuff out.
00:52:10 You don’t mind looking dumb.
00:52:11 No.
00:52:14 You just figure it out pretty quickly.
00:52:16 Sometimes I try things and I don’t figure them out quickly.
00:52:19 My chest rating is like a 1400,
00:52:20 despite putting like a couple of hundred hours in.
00:52:23 It’s pathetic.
00:52:24 I mean, to be fair, I know that I could do it better
00:52:26 if I did it better.
00:52:27 Like don’t play five minute games,
00:52:29 play 15 minute games at least.
00:52:31 Like I know these things, but it just doesn’t,
00:52:34 it doesn’t stick nicely in my knowledge stream.
00:52:37 All right, let’s talk about Comma AI.
00:52:39 What’s the mission of the company?
00:52:42 Let’s like look at the biggest picture.
00:52:44 Oh, I have an exact statement.
00:52:46 Solve self driving cars
00:52:48 while delivering shippable intermediaries.
00:52:51 So longterm vision is have fully autonomous vehicles
00:52:56 and make sure you’re making money along the way.
00:52:59 I think it doesn’t really speak to money,
00:53:00 but I can talk about what solve self driving cars means.
00:53:03 Solve self driving cars of course means
00:53:06 you’re not building a new car,
00:53:08 you’re building a person replacement.
00:53:10 That person can sit in the driver’s seat
00:53:12 and drive you anywhere a person can drive
00:53:14 with a human or better level of safety,
00:53:17 speed, quality, comfort.
00:53:21 And what’s the second part of that?
00:53:23 Delivering shippable intermediaries is well,
00:53:26 it’s a way to fund the company, that’s true.
00:53:28 But it’s also a way to keep us honest.
00:53:31 If you don’t have that, it is very easy
00:53:34 with this technology to think you’re making progress
00:53:39 when you’re not.
00:53:40 I’ve heard it best described on Hacker News as
00:53:43 you can set any arbitrary milestone,
00:53:46 meet that milestone and still be infinitely far away
00:53:49 from solving self driving cars.
00:53:51 So it’s hard to have like real deadlines
00:53:53 when you’re like Cruz or Waymo when you don’t have revenue.
00:54:02 Is that, I mean, is revenue essentially
00:54:06 the thing we’re talking about here?
00:54:07 Revenue is, capitalism is based around consent.
00:54:11 Capitalism, the way that you get revenue
00:54:13 is real capitalism comes in the real capitalism camp.
00:54:16 There’s definitely scams out there,
00:54:17 but real capitalism is based around consent.
00:54:19 It’s based around this idea that like,
00:54:20 if we’re getting revenue, it’s because we’re providing
00:54:22 at least that much value to another person.
00:54:24 When someone buys $1,000 comma two from us,
00:54:27 we’re providing them at least $1,000 of value
00:54:29 or they wouldn’t buy it.
00:54:30 Brilliant, so can you give a whirlwind overview
00:54:32 of the products that Comma AI provides,
00:54:34 like throughout its history and today?
00:54:38 I mean, yeah, the past ones aren’t really that interesting.
00:54:40 It’s kind of just been refinement of the same idea.
00:54:45 The real only product we sell today is the Comma 2.
00:54:48 Which is a piece of hardware with cameras.
00:54:50 Mm, so the Comma 2, I mean, you can think about it
00:54:54 kind of like a person.
00:54:57 Future hardware will probably be
00:54:58 even more and more personlike.
00:55:00 So it has eyes, ears, a mouth, a brain,
00:55:07 and a way to interface with the car.
00:55:09 Does it have consciousness?
00:55:10 Just kidding, that was a trick question.
00:55:13 I don’t have consciousness either.
00:55:15 Me and the Comma 2 are the same.
00:55:16 You’re the same?
00:55:17 I have a little more compute than it.
00:55:18 It only has like the same compute as a B, you know.
00:55:23 You’re more efficient energy wise
00:55:25 for the compute you’re doing.
00:55:26 Far more efficient energy wise.
00:55:29 20 petaflops, 20 watts, crazy.
00:55:30 Do you lack consciousness?
00:55:32 Sure.
00:55:33 Do you fear death?
00:55:33 You do, you want immortality.
00:55:35 Of course I fear death.
00:55:36 Does Comma AI fear death?
00:55:38 I don’t think so.
00:55:39 Of course it does.
00:55:40 It very much fears, well, it fears negative loss.
00:55:42 Oh yeah.
00:55:43 Okay, so Comma 2, when did that come out?
00:55:49 That was a year ago?
00:55:50 No, two.
00:55:52 Early this year.
00:55:53 Wow, time, it feels like, yeah.
00:55:56 2020 feels like it’s taken 10 years to get to the end.
00:56:00 It’s a long year.
00:56:01 It’s a long year.
00:56:03 So what’s the sexiest thing about Comma 2 feature wise?
00:56:08 So, I mean, maybe you can also link on like, what is it?
00:56:14 Like what’s its purpose?
00:56:15 Cause there’s a hardware, there’s a software component.
00:56:18 You’ve mentioned the sensors,
00:56:20 but also like what is it, its features and capabilities?
00:56:23 I think our slogan summarizes it well.
00:56:25 Comma slogan is make driving chill.
00:56:28 I love it, okay.
00:56:30 Yeah, I mean, it is, you know,
00:56:33 if you like cruise control, imagine cruise control,
00:56:35 but much, much more.
00:56:36 So it can do adaptive cruise control things,
00:56:41 which is like slow down for cars in front of it,
00:56:42 maintain a certain speed.
00:56:44 And it can also do lane keeping.
00:56:46 So staying in the lane and doing it better
00:56:48 and better and better over time.
00:56:50 It’s very much machine learning based.
00:56:53 So this camera is, there’s a driver facing camera too.
00:57:01 What else is there?
00:57:02 What am I thinking?
00:57:02 So the hardware versus software.
00:57:04 So open pilot versus the actual hardware of the device.
00:57:09 What’s, can you draw that distinction?
00:57:10 What’s one, what’s the other?
00:57:11 I mean, the hardware is pretty much a cell phone
00:57:13 with a few additions.
00:57:14 A cell phone with a cooling system
00:57:16 and with a car interface connected to it.
00:57:20 And by cell phone, you mean like Qualcomm Snapdragon.
00:57:25 Yeah, the current hardware is a Snapdragon 821.
00:57:29 It has wifi radio, it has an LTE radio, it has a screen.
00:57:32 We use every part of the cell phone.
00:57:35 And then the interface with the car
00:57:37 is specific to the car.
00:57:38 So you keep supporting more and more cars.
00:57:41 Yeah, so the interface to the car,
00:57:42 I mean, the device itself just has four CAN buses.
00:57:45 It has four CAN interfaces on it
00:57:46 that are connected through the USB port to the phone.
00:57:49 And then, yeah, on those four CAN buses,
00:57:53 you connect it to the car.
00:57:54 And there’s a little harness to do this.
00:57:56 Cars are actually surprisingly similar.
00:57:58 So CAN is the protocol by which cars communicate.
00:58:01 And then you’re able to read stuff and write stuff
00:58:04 to be able to control the car depending on the car.
00:58:06 So what’s the software side?
00:58:08 What’s OpenPilot?
00:58:10 So I mean, OpenPilot is,
00:58:11 the hardware is pretty simple compared to OpenPilot.
00:58:13 OpenPilot is, well, so you have a machine learning model,
00:58:21 which it’s in OpenPilot, it’s a blob.
00:58:24 It’s just a blob of weights.
00:58:25 It’s not like people are like, oh, it’s closed source.
00:58:27 I’m like, it’s a blob of weights.
00:58:28 What do you expect?
00:58:29 So it’s primarily neural network based.
00:58:33 You, well, OpenPilot is all the software
00:58:36 kind of around that neural network.
00:58:37 That if you have a neural network that says,
00:58:39 here’s where you wanna send the car,
00:58:40 OpenPilot actually goes and executes all of that.
00:58:44 It cleans up the input to the neural network.
00:58:46 It cleans up the output and executes on it.
00:58:49 So it connects, it’s the glue
00:58:50 that connects everything together.
00:58:51 Runs the sensors, does a bunch of calibration
00:58:54 for the neural network, deals with like,
00:58:58 if the car is on a banked road,
00:59:00 you have to counter steer against that.
00:59:02 And the neural network can’t necessarily know that
00:59:03 by looking at the picture.
00:59:06 So you do that with other sensors
00:59:08 and Fusion and Localizer.
00:59:09 OpenPilot also is responsible
00:59:11 for sending the data up to our servers.
00:59:14 So we can learn from it, logging it, recording it,
00:59:17 running the cameras, thermally managing the device,
00:59:21 managing the disk space on the device,
00:59:23 managing all the resources on the device.
00:59:24 So what, since we last spoke,
00:59:26 I don’t remember when, maybe a year ago,
00:59:28 maybe a little bit longer,
00:59:30 how has OpenPilot improved?
00:59:33 We did exactly what I promised you.
00:59:34 I promised you that by the end of the year,
00:59:36 where you’d be able to remove the lanes.
00:59:40 The lateral policy is now almost completely end to end.
00:59:46 You can turn the lanes off and it will drive,
00:59:48 drive slightly worse on the highway
00:59:49 if you turn the lanes off,
00:59:51 but you can turn the lanes off and it will drive well,
00:59:54 trained completely end to end on user data.
00:59:57 And this year we hope to do the same
00:59:58 for the longitudinal policy.
01:00:00 So that’s the interesting thing is you’re not doing,
01:00:03 you don’t appear to be, maybe you can correct me,
01:00:05 you don’t appear to be doing lane detection
01:00:08 or lane marking detection or kind of the segmentation task
01:00:12 or any kind of object detection task.
01:00:15 You’re doing what’s traditionally more called
01:00:17 like end to end learning.
01:00:19 So, and trained on actual behavior of drivers
01:00:24 when they’re driving the car manually.
01:00:27 And this is hard to do.
01:00:29 It’s not supervised learning.
01:00:32 Yeah, but so the nice thing is there’s a lot of data.
01:00:34 So it’s hard and easy, right?
01:00:37 It’s a…
01:00:37 We have a lot of high quality data, yeah.
01:00:40 Like more than you need in the second.
01:00:41 Well…
01:00:42 We have way more than we do.
01:00:43 We have way more data than we need.
01:00:44 I mean, it’s an interesting question actually,
01:00:47 because in terms of amount, you have more than you need,
01:00:50 but the driving is full of edge cases.
01:00:54 So how do you select the data you train on?
01:00:58 I think this is an interesting open question.
01:01:00 Like what’s the cleverest way to select data?
01:01:04 That’s the question Tesla is probably working on.
01:01:07 That’s, I mean, the entirety of machine learning can be,
01:01:09 they don’t seem to really care.
01:01:11 They just kind of select data.
01:01:12 But I feel like that if you want to solve,
01:01:14 if you want to create intelligent systems,
01:01:16 you have to pick data well, right?
01:01:18 And so do you have any hints, ideas of how to do it well?
01:01:22 So in some ways that is…
01:01:25 The definition I like of reinforcement learning
01:01:27 versus supervised learning.
01:01:29 In supervised learning, the weights depend on the data.
01:01:32 Right?
01:01:34 And this is obviously true,
01:01:35 but in reinforcement learning,
01:01:38 the data depends on the weights.
01:01:40 Yeah.
01:01:41 And actually both ways.
01:01:42 That’s poetry.
01:01:43 So how does it know what data to train on?
01:01:46 Well, let it pick.
01:01:47 We’re not there yet, but that’s the eventual.
01:01:49 So you’re thinking this almost like
01:01:51 a reinforcement learning framework.
01:01:53 We’re going to do RL on the world.
01:01:55 Every time a car makes a mistake, user disengages,
01:01:58 we train on that and do RL on the world.
01:02:00 Ship out a new model, that’s an epoch, right?
01:02:03 And for now you’re not doing the Elon style promising
01:02:08 that it’s going to be fully autonomous.
01:02:09 You really are sticking to level two
01:02:12 and like it’s supposed to be supervised.
01:02:15 It is definitely supposed to be supervised
01:02:16 and we enforce the fact that it’s supervised.
01:02:19 We look at our rate of improvement in disengagements.
01:02:23 OpenPilot now has an unplanned disengagement
01:02:25 about every a hundred miles.
01:02:27 This is up from 10 miles, like maybe,
01:02:32 maybe maybe a year ago.
01:02:36 Yeah.
01:02:37 So maybe we’ve seen 10 X improvement in a year,
01:02:38 but a hundred miles is still a far cry
01:02:41 from the a hundred thousand you’re going to need.
01:02:43 So you’re going to somehow need to get three more 10 Xs
01:02:48 in there.
01:02:49 And you’re, what’s your intuition?
01:02:52 You’re basically hoping that there’s exponential
01:02:54 improvement built into the baked into the cake somewhere.
01:02:56 Well, that’s even, I mean, 10 X improvement,
01:02:58 that’s already assuming exponential, right?
01:03:00 There’s definitely exponential improvement.
01:03:02 And I think when Elon talks about exponential,
01:03:04 like these things, these systems are going to
01:03:06 exponentially improve, just exponential doesn’t mean
01:03:09 you’re getting a hundred gigahertz processors tomorrow.
01:03:12 Right? Like it’s going to still take a while
01:03:15 because the gap between even our best system
01:03:18 and humans is still large.
01:03:20 So that’s an interesting distinction to draw.
01:03:22 So if you look at the way Tesla is approaching the problem
01:03:26 and the way you’re approaching the problem,
01:03:28 which is very different than the rest of the self driving
01:03:31 car world.
01:03:32 So let’s put them aside is you’re treating most
01:03:35 the driving task as a machine learning problem.
01:03:37 And the way Tesla is approaching it is with the multitask
01:03:40 learning where you break the task of driving into hundreds
01:03:44 of different tasks and you have this multiheaded
01:03:47 neural network that’s very good at performing each task.
01:03:51 And there there’s presumably something on top that’s
01:03:54 stitching stuff together in order to make control
01:03:59 decisions, policy decisions about how you move the car.
01:04:02 But what that allows you, there’s a brilliance to this
01:04:04 because it allows you to master each task,
01:04:08 like lane detection, stop sign detection,
01:04:13 the traffic light detection, drivable area segmentation,
01:04:19 you know, vehicle, bicycle, pedestrian detection.
01:04:23 There’s some localization tasks in there.
01:04:25 Also predicting of like, yeah,
01:04:30 predicting how the entities in the scene are going to move.
01:04:34 Like everything is basically a machine learning task.
01:04:36 So there’s a classification, segmentation, prediction.
01:04:40 And it’s nice because you can have this entire engine,
01:04:44 data engine that’s mining for edge cases for each one of
01:04:48 these tasks.
01:04:49 And you can have people like engineers that are basically
01:04:52 masters of that task,
01:04:53 like become the best person in the world at,
01:04:56 as you talk about the cone guy for Waymo,
01:04:59 the becoming the best person in the world at cone detection.
01:05:06 So that’s a compelling notion from a supervised learning
01:05:10 perspective, automating much of the process of edge case
01:05:15 discovery and retraining neural network for each of the
01:05:17 individual perception tasks.
01:05:19 And then you’re looking at the machine learning in a more
01:05:22 holistic way, basically doing end to end learning on the
01:05:27 driving tasks, supervised, trained on the data of the
01:05:31 actual driving of people.
01:05:34 They use comma AI, like actual human drivers,
01:05:37 their manual control,
01:05:38 plus the moments of disengagement that maybe with some
01:05:44 labeling could indicate the failure of the system.
01:05:47 So you have the,
01:05:48 you have a huge amount of data for positive control of the
01:05:52 vehicle, like successful control of the vehicle,
01:05:55 both maintaining the lane as,
01:05:58 as I think you’re also working on longitudinal control of
01:06:01 the vehicle and then failure cases where the vehicle does
01:06:04 something wrong that needs disengagement.
01:06:08 So like what,
01:06:09 why do you think you’re right and Tesla is wrong on this?
01:06:14 And do you think,
01:06:15 do you think you’ll come around the Tesla way?
01:06:17 Do you think Tesla will come around to your way?
01:06:21 If you were to start a chess engine company,
01:06:23 would you hire a Bishop guy?
01:06:26 See, we have a,
01:06:27 this is Monday morning.
01:06:29 Quarterbacking is a yes, probably.
01:06:36 Oh, our Rook guy.
01:06:37 Oh, we stole the Rook guy from that company.
01:06:39 Oh, we’re going to have real good Rooks.
01:06:40 Well, there’s not many pieces, right?
01:06:43 You can,
01:06:46 there’s not many guys and gals to hire.
01:06:48 You just have a few that work in the Bishop,
01:06:51 a few that work in the Rook.
01:06:52 Is that not ludicrous today to think about
01:06:55 in a world of AlphaZero?
01:06:57 But AlphaZero is a chess game.
01:06:58 So the fundamental question is,
01:07:01 how hard is driving compared to chess?
01:07:04 Because, so long term,
01:07:07 end to end,
01:07:08 will be the right solution.
01:07:10 The question is how many years away is that?
01:07:13 End to end is going to be the only solution for level five.
01:07:15 For the only way we’ll get there.
01:07:17 Of course, and of course,
01:07:18 Tesla is going to come around to my way.
01:07:19 And if you’re a Rook guy out there, I’m sorry.
01:07:22 The cone guy.
01:07:24 I don’t know.
01:07:25 We’re going to specialize each task.
01:07:26 We’re going to really understand Rook placement.
01:07:29 Yeah.
01:07:30 I understand the intuition you have.
01:07:32 I mean, that,
01:07:35 that is a very compelling notion
01:07:36 that we can learn the task end to end,
01:07:39 like the same compelling notion you might have
01:07:40 for natural language conversation.
01:07:42 But I’m not
01:07:44 sure,
01:07:47 because one thing you sneaked in there
01:07:48 is the assertion that it’s impossible to get to level five
01:07:53 without this kind of approach.
01:07:55 I don’t know if that’s obvious.
01:07:57 I don’t know if that’s obvious either.
01:07:58 I don’t actually mean that.
01:08:01 I think that it is much easier
01:08:03 to get to level five with an end to end approach.
01:08:05 I think that the other approach is doable,
01:08:08 but the magnitude of the engineering challenge
01:08:11 may exceed what humanity is capable of.
01:08:13 But what do you think of the Tesla data engine approach,
01:08:19 which to me is an active learning task,
01:08:21 is kind of fascinating,
01:08:22 is breaking it down into these multiple tasks
01:08:25 and mining their data constantly for like edge cases
01:08:29 for these different tasks.
01:08:30 Yeah, but the tasks themselves are not being learned.
01:08:32 This is feature engineering.
01:08:35 Yeah, I mean, it’s a higher abstraction level
01:08:40 of feature engineering for the different tasks.
01:08:43 Task engineering in a sense.
01:08:44 It’s slightly better feature engineering,
01:08:46 but it’s still fundamentally is feature engineering.
01:08:49 And if anything about the history of AI
01:08:51 has taught us anything,
01:08:52 it’s that feature engineering approaches
01:08:54 will always be replaced and lose to end to end.
01:08:57 Now, to be fair, I cannot really make promises on timelines,
01:09:02 but I can say that when you look at the code for Stockfish
01:09:05 and the code for AlphaZero,
01:09:06 one is a lot shorter than the other,
01:09:09 a lot more elegant,
01:09:09 required a lot less programmer hours to write.
01:09:12 Yeah, but there was a lot more murder of bad agents
01:09:21 on the AlphaZero side.
01:09:24 By murder, I mean agents that played a game
01:09:29 and failed miserably.
01:09:30 Yeah.
01:09:31 Oh, oh.
01:09:32 In simulation, that failure is less costly.
01:09:34 Yeah.
01:09:35 In real world, it’s…
01:09:37 Do you mean in practice,
01:09:38 like AlphaZero has lost games miserably?
01:09:40 No.
01:09:41 Wow.
01:09:42 I haven’t seen that.
01:09:43 No, but I know, but the requirement for AlphaZero is…
01:09:47 A simulator.
01:09:48 To be able to like evolution, human evolution,
01:09:51 not human evolution, biological evolution of life on earth
01:09:54 from the origin of life has murdered trillions
01:09:58 upon trillions of organisms on the path thus humans.
01:10:02 Yeah.
01:10:03 So the question is, can we stitch together
01:10:05 a human like object without having to go
01:10:07 through the entirety process of evolution?
01:10:09 Well, no, but do the evolution in simulation.
01:10:11 Yeah, that’s the question.
01:10:12 Can we simulate?
01:10:13 So do you have a sense that it’s possible
01:10:15 to simulate some aspect?
01:10:16 MuZero is exactly this.
01:10:18 MuZero is the solution to this.
01:10:21 MuZero I think is going to be looked back
01:10:23 as the canonical paper.
01:10:25 And I don’t think deep learning is everything.
01:10:26 I think that there’s still a bunch of things missing
01:10:28 to get there, but MuZero I think is going to be looked back
01:10:31 as the kind of cornerstone paper
01:10:34 of this whole deep learning era.
01:10:37 And MuZero is the solution to self driving cars.
01:10:39 You have to make a few tweaks to it,
01:10:41 but MuZero does effectively that.
01:10:42 It does those rollouts and those murdering
01:10:45 in a learned simulator and a learned dynamics model.
01:10:50 That’s interesting.
01:10:51 It doesn’t get enough love.
01:10:51 I was blown away when I read that paper.
01:10:54 I’m like, okay, I’ve always said a comma.
01:10:57 I’m going to sit and I’m going to wait for the solution
01:10:58 to self driving cars to come along.
01:11:00 This year I saw it.
01:11:01 It’s MuZero.
01:11:05 So.
01:11:06 Sit back and let the winning roll in.
01:11:09 So your sense, just to elaborate a little bit,
01:11:12 it’s a link on the topic.
01:11:13 Your sense is neural networks will solve driving.
01:11:16 Yes.
01:11:17 Like we don’t need anything else.
01:11:18 I think the same way chess was maybe the chess
01:11:21 and maybe Google are the pinnacle of like search algorithms
01:11:25 and things that look kind of like a star.
01:11:28 The pinnacle of this era is going to be self driving cars.
01:11:34 But on the path of that, you have to deliver products
01:11:38 and it’s possible that the path to full self driving cars
01:11:42 will take decades.
01:11:44 I doubt it.
01:11:45 How long would you put on it?
01:11:47 Like what are we, you’re chasing it, Tesla’s chasing it.
01:11:53 What are we talking about?
01:11:54 Five years, 10 years, 50 years.
01:11:56 Let’s say in the 2020s.
01:11:58 In the 2020s.
01:11:59 The later part of the 2020s.
01:12:03 With the neural network.
01:12:05 Well, that would be nice to see.
01:12:06 And then the path to that, you’re delivering products,
01:12:09 which is a nice L2 system.
01:12:10 That’s what Tesla’s doing, a nice L2 system.
01:12:13 Just gets better every time.
01:12:14 L2, the only difference between L2 and the other levels
01:12:16 is who takes liability.
01:12:17 And I’m not a liability guy, I don’t wanna take liability.
01:12:20 I’m gonna level two forever.
01:12:22 Now on that little transition,
01:12:25 I mean, how do you make the transition work?
01:12:29 Is this where driver sensing comes in?
01:12:32 Like how do you make the, cause you said a hundred miles,
01:12:35 like, is there some sort of human factor psychology thing
01:12:41 where people start to overtrust the system,
01:12:43 all those kinds of effects,
01:12:45 once it gets better and better and better and better,
01:12:46 they get lazier and lazier and lazier.
01:12:49 Is that, like, how do you get that transition right?
01:12:52 First off, our monitoring is already adaptive.
01:12:54 Our monitoring is already seen adaptive.
01:12:56 Driver monitoring is just the camera
01:12:58 that’s looking at the driver.
01:13:00 You have an infrared camera in the…
01:13:03 Our policy for how we enforce the driver monitoring
01:13:06 is seen adaptive.
01:13:07 What’s that mean?
01:13:08 Well, for example, in one of the extreme cases,
01:13:12 if the car is not moving,
01:13:14 we do not actively enforce driver monitoring, right?
01:13:19 If you are going through a,
01:13:22 like a 45 mile an hour road with lights
01:13:25 and stop signs and potentially pedestrians,
01:13:27 we enforce a very tight driver monitoring policy.
01:13:30 If you are alone on a perfectly straight highway,
01:13:33 and this is, it’s all machine learning.
01:13:35 None of that is hand coded.
01:13:36 Actually, the stop is hand coded, but…
01:13:39 So there’s some kind of machine learning
01:13:41 estimation of risk.
01:13:42 Yes.
01:13:43 Yeah.
01:13:44 I mean, I’ve always been a huge fan of that.
01:13:45 That’s a…
01:13:47 Because…
01:13:48 It’s difficult to do every step into that direction
01:13:53 is a worthwhile step to take.
01:13:55 It might be difficult to do really well.
01:13:56 Like us humans are able to estimate risk pretty damn well,
01:13:59 whatever the hell that is.
01:14:01 That feels like one of the nice features of us humans.
01:14:06 Cause like we humans are really good drivers
01:14:08 when we’re really like tuned in
01:14:11 and we’re good at estimating risk.
01:14:12 Like when are we supposed to be tuned in?
01:14:14 Yeah.
01:14:15 And, you know, people are like,
01:14:17 oh, well, you know,
01:14:18 why would you ever make the driver monitoring policy
01:14:20 less aggressive?
01:14:21 Why would you always not keep it at its most aggressive?
01:14:23 Because then people are just going to get fatigued from it.
01:14:25 Yes.
01:14:26 When they get annoyed.
01:14:27 You want them…
01:14:28 Yeah.
01:14:29 You want the experience to be pleasant.
01:14:30 Obviously I want the experience to be pleasant,
01:14:32 but even just from a straight up safety perspective,
01:14:35 if you alert people when they look around and they’re like,
01:14:39 why is this thing alerting me?
01:14:41 There’s nothing I could possibly hit right now.
01:14:42 People will just learn to tune it out.
01:14:45 People will just learn to tune it out,
01:14:46 to put weights on the steering wheel,
01:14:48 to do whatever to overcome it.
01:14:49 And remember that you’re always part
01:14:52 of this adaptive system.
01:14:53 So all I can really say about, you know,
01:14:55 how this scales going forward is yeah,
01:14:57 it’s something we have to monitor for.
01:14:59 Ooh, we don’t know.
01:15:00 This is a great psychology experiment at scale.
01:15:02 Like we’ll see.
01:15:03 Yeah, it’s fascinating.
01:15:04 Track it.
01:15:04 And making sure you have a good understanding of attention
01:15:09 is a very key part of that psychology problem.
01:15:11 Yeah.
01:15:12 I think you and I probably have a different,
01:15:14 come to it differently, but to me,
01:15:16 it’s a fascinating psychology problem
01:15:19 to explore something much deeper than just driving.
01:15:22 It’s such a nice way to explore human attention
01:15:26 and human behavior, which is why, again,
01:15:30 we’ve probably both criticized Mr. Elon Musk
01:15:34 on this one topic from different avenues.
01:15:38 So both offline and online,
01:15:39 I had little chats with Elon and like,
01:15:44 I love human beings as a computer vision problem,
01:15:48 as an AI problem, it’s fascinating.
01:15:51 He wasn’t so much interested in that problem.
01:15:53 It’s like in order to solve driving,
01:15:56 the whole point is you want to remove the human
01:15:58 from the picture.
01:16:01 And it seems like you can’t do that quite yet.
01:16:04 Eventually, yes, but you can’t quite do that yet.
01:16:07 So this is the moment where you can’t yet say,
01:16:12 I told you so to Tesla, but it’s getting there
01:16:17 because I don’t know if you’ve seen this,
01:16:19 there’s some reporting that they’re in fact
01:16:21 starting to do driver monitoring.
01:16:23 Yeah, they shift the model in shadow mode.
01:16:26 With, I believe, only a visible light camera,
01:16:29 it might even be fisheye.
01:16:31 It’s like a low resolution.
01:16:33 Low resolution, visible light.
01:16:34 I mean, to be fair, that’s what we have in the Eon as well,
01:16:37 our last generation product.
01:16:38 This is the one area where I can say
01:16:41 our hardware is ahead of Tesla.
01:16:42 The rest of our hardware, way, way behind,
01:16:43 but our driver monitoring camera.
01:16:46 So you think, I think on the third row Tesla podcast,
01:16:50 or somewhere else, I’ve heard you say that obviously,
01:16:54 eventually they’re gonna have driver monitoring.
01:16:57 I think what I’ve said is Elon will definitely ship
01:16:59 driver monitoring before he ships level five.
01:17:01 Before level five.
01:17:02 And I’m willing to bet 10 grand on that.
01:17:04 And you bet 10 grand on that.
01:17:07 I mean, now I don’t wanna take the bet,
01:17:08 but before, maybe someone would have,
01:17:09 oh, I should have got my money in.
01:17:10 Yeah.
01:17:11 It’s an interesting bet.
01:17:12 I think you’re right.
01:17:16 I’m actually on a human level
01:17:19 because he’s been, he’s made the decision.
01:17:24 Like he said that driver monitoring is the wrong way to go.
01:17:27 But like, you have to think of as a human, as a CEO,
01:17:31 I think that’s the right thing to say when,
01:17:36 like sometimes you have to say things publicly
01:17:40 that are different than when you actually believe,
01:17:41 because when you’re producing a large number of vehicles
01:17:45 and the decision was made not to include the camera,
01:17:47 like what are you supposed to say?
01:17:49 Like our cars don’t have the thing
01:17:51 that I think is right to have.
01:17:54 It’s an interesting thing.
01:17:55 But like on the other side, as a CEO,
01:17:58 I mean, something you could probably speak to as a leader,
01:18:01 I think about me as a human
01:18:04 to publicly change your mind on something.
01:18:07 How hard is that?
01:18:08 Especially when assholes like George Haas say,
01:18:10 I told you so.
01:18:12 All I will say is I am not a leader
01:18:14 and I am happy to change my mind.
01:18:17 And I will.
01:18:17 You think Elon will?
01:18:20 Yeah, I do.
01:18:22 I think he’ll come up with a good way
01:18:24 to make it psychologically okay for him.
01:18:27 Well, it’s such an important thing, man.
01:18:29 Especially for a first principles thinker,
01:18:31 because he made a decision that driver monitoring
01:18:34 is not the right way to go.
01:18:35 And I could see that decision.
01:18:37 And I could even make that decision.
01:18:39 Like I was on the fence too.
01:18:41 Like I’m not a,
01:18:42 driver monitoring is such an obvious,
01:18:47 simple solution to the problem of attention.
01:18:49 It’s not obvious to me that just by putting a camera there,
01:18:52 you solve things.
01:18:54 You have to create an incredible, compelling experience.
01:18:59 Just like you’re talking about.
01:19:01 I don’t know if it’s easy to do that.
01:19:03 It’s not at all easy to do that, in fact, I think.
01:19:05 So as a creator of a car that’s trying to create a product
01:19:10 that people love, which is what Tesla tries to do, right?
01:19:14 It’s not obvious to me that as a design decision,
01:19:18 whether adding a camera is a good idea.
01:19:20 From a safety perspective either,
01:19:22 like in the human factors community,
01:19:25 everybody says that you should obviously
01:19:27 have driver sensing, driver monitoring.
01:19:30 But that’s like saying it’s obvious as parents,
01:19:36 you shouldn’t let your kids go out at night.
01:19:39 But okay, but like,
01:19:43 they’re still gonna find ways to do drugs.
01:19:45 Like, you have to also be good parents.
01:19:49 So like, it’s much more complicated than just the,
01:19:52 you need to have driver monitoring.
01:19:54 I totally disagree on, okay, if you have a camera there
01:19:58 and the camera’s watching the person,
01:20:00 but never throws an alert, they’ll never think about it.
01:20:03 Right?
01:20:04 The driver monitoring policy that you choose to,
01:20:08 how you choose to communicate with the user
01:20:10 is entirely separate from the data collection perspective.
01:20:14 Right?
01:20:15 Right?
01:20:15 So, you know, like, there’s one thing to say,
01:20:20 like, you know, tell your teenager they can’t do something.
01:20:24 There’s another thing to like, you know, gather the data.
01:20:27 So you can make informed decisions.
01:20:28 That’s really interesting.
01:20:29 But you have to make that,
01:20:30 that’s the interesting thing about cars.
01:20:33 But even true with common AI,
01:20:35 like you don’t have to manufacture the thing
01:20:37 into the car, is you have to make a decision
01:20:40 that anticipates the right strategy longterm.
01:20:44 So like, you have to start collecting the data
01:20:46 and start making decisions.
01:20:47 Started it three years ago.
01:20:49 I believe that we have the best driver monitoring solution
01:20:52 in the world.
01:20:54 I think that when you compare it to Super Cruise
01:20:57 is the only other one that I really know that shipped.
01:20:59 And ours is better.
01:21:01 What do you like and not like about Super Cruise?
01:21:06 I mean, I had a few Super Cruise,
01:21:08 the sun would be shining through the window,
01:21:12 would blind the camera,
01:21:13 and it would say I wasn’t paying attention.
01:21:14 When I was looking completely straight,
01:21:16 I couldn’t reset the attention with a steering wheel touch
01:21:19 and Super Cruise would disengage.
01:21:21 Like I was communicating to the car, I’m like, look,
01:21:22 I am here, I am paying attention.
01:21:24 Why are you really gonna force me to disengage?
01:21:26 And it did.
01:21:28 So it’s a constant conversation with the user.
01:21:32 And yeah, there’s no way to ship a system
01:21:33 like this if you can OTA.
01:21:35 We’re shipping a new one every month.
01:21:37 Sometimes we balance it with our users on Discord.
01:21:40 Like sometimes we make the driver monitoring
01:21:41 a little more aggressive and people complain.
01:21:43 Sometimes they don’t.
01:21:45 We want it to be as aggressive as possible
01:21:47 where people don’t complain and it doesn’t feel intrusive.
01:21:49 So being able to update the system over the air
01:21:51 is an essential component.
01:21:52 I mean, that’s probably to me, you mentioned,
01:21:56 I mean, to me that is the biggest innovation of Tesla,
01:22:01 that it made people realize that over the air updates
01:22:04 is essential.
01:22:06 Yeah.
01:22:07 I mean, was that not obvious from the iPhone?
01:22:10 The iPhone was the first real product that OTA’d, I think.
01:22:13 Was it actually, that’s brilliant, you’re right.
01:22:15 I mean, the game consoles used to not, right?
01:22:17 The game consoles were maybe the second thing that did.
01:22:18 Wow, I didn’t really think about one of the amazing features
01:22:22 of a smartphone isn’t just like the touchscreen
01:22:26 isn’t the thing, it’s the ability to constantly update.
01:22:30 Yeah, it gets better.
01:22:31 It gets better.
01:22:35 Love my iOS 14.
01:22:36 Yeah.
01:22:38 Well, one thing that I probably disagree with you
01:22:41 on driver monitoring is you said that it’s easy.
01:22:46 I mean, you tend to say stuff is easy.
01:22:48 I’m sure the, I guess you said it’s easy
01:22:52 relative to the external perception problem.
01:22:58 Can you elaborate why you think it’s easy?
01:23:00 Feature engineering works for driver monitoring.
01:23:03 Feature engineering does not work for the external.
01:23:05 So human faces are not, human faces and the movement
01:23:10 of human faces and head and body is not as variable
01:23:14 as the external environment, is your intuition?
01:23:17 Yes, and there’s another big difference as well.
01:23:20 Your reliability of a driver monitoring system
01:23:22 doesn’t actually need to be that high.
01:23:24 The uncertainty, if you have something that’s detecting
01:23:27 whether the human’s paying attention and it only works
01:23:29 92% of the time, you’re still getting almost all
01:23:31 the benefit of that because the human,
01:23:33 like you’re training the human, right?
01:23:35 You’re dealing with a system that’s really helping you out.
01:23:39 It’s a conversation.
01:23:40 It’s not like the external thing where guess what?
01:23:43 If you swerve into a tree, you swerve into a tree, right?
01:23:46 Like you get no margin for error there.
01:23:48 Yeah, I think that’s really well put.
01:23:49 I think that’s the right, exactly the place
01:23:54 where comparing to the external perception,
01:23:58 the control problem, the driver monitoring is easier
01:24:01 because you don’t, the bar for success is much lower.
01:24:05 Yeah, but I still think like the human face
01:24:09 is more complicated actually than the external environment,
01:24:12 but for driving, you don’t give a damn.
01:24:14 I don’t need, yeah, I don’t need something,
01:24:15 I don’t need something that complicated
01:24:18 to have to communicate the idea to the human
01:24:22 that I want to communicate, which is,
01:24:23 yo, system might mess up here.
01:24:25 You gotta pay attention.
01:24:26 Yeah, see, that’s my love and fascination is the human face.
01:24:32 And it feels like this is a nice place to create products
01:24:38 that create an experience in the car.
01:24:40 So like, it feels like there should be
01:24:42 more richer experiences in the car, you know?
01:24:47 Like that’s an opportunity for like something like On My Eye
01:24:51 or just any kind of system like a Tesla
01:24:53 or any of the autonomous vehicle companies
01:24:56 is because software is, there’s much more sensors
01:24:59 and so much is on our software
01:25:00 and you’re doing machine learning anyway,
01:25:02 there’s an opportunity to create totally new experiences
01:25:06 that we’re not even anticipating.
01:25:08 You don’t think so?
01:25:10 Nah.
01:25:10 You think it’s a box that gets you from A to B
01:25:12 and you want to do it chill?
01:25:14 Yeah, I mean, I think as soon as we get to level three
01:25:16 on highways, okay, enjoy your candy crush,
01:25:19 enjoy your Hulu, enjoy your, you know, whatever, whatever.
01:25:23 Sure, you get this, you can look at screens basically
01:25:26 versus right now where you have music and audio books.
01:25:28 So level three is where you can kind of disengage
01:25:31 in stretches of time.
01:25:34 Well, you think level three is possible?
01:25:37 Like on the highway going for 100 miles
01:25:39 and you can just go to sleep?
01:25:40 Oh yeah, sleep.
01:25:43 So again, I think it’s really all on a spectrum.
01:25:47 I think that being able to use your phone
01:25:50 while you’re on the highway and like this all being okay
01:25:53 and being aware that the car might alert you
01:25:55 and you have five seconds to basically.
01:25:57 So the five second thing is you think is possible?
01:25:59 Yeah, I think it is, oh yeah.
01:26:00 Not in all scenarios, right?
01:26:02 Some scenarios it’s not.
01:26:03 It’s the whole risk thing that you mentioned is nice
01:26:06 is to be able to estimate like how risky is this situation?
01:26:10 That’s really important to understand.
01:26:12 One other thing you mentioned comparing KAMA
01:26:15 and Autopilot is that something about the haptic feel
01:26:20 of the way KAMA controls the car when things are uncertain.
01:26:25 Like it behaves a little bit more uncertain
01:26:27 when things are uncertain.
01:26:29 That’s kind of an interesting point.
01:26:31 And then Autopilot is much more confident always
01:26:34 even when it’s uncertain until it runs into trouble.
01:26:39 That’s a funny thing.
01:26:40 I actually mentioned that to Elon, I think.
01:26:42 And then the first time we talked, he wasn’t biting.
01:26:46 It’s like communicating uncertainty.
01:26:48 I guess KAMA doesn’t really communicate uncertainty
01:26:51 explicitly, it communicates it through haptic feel.
01:26:55 Like what’s the role of communicating uncertainty
01:26:57 do you think?
01:26:58 Oh, we do some stuff explicitly.
01:26:59 Like we do detect the lanes when you’re on the highway
01:27:01 and we’ll show you how many lanes we’re using to drive with.
01:27:04 You can look at where it thinks the lanes are.
01:27:06 You can look at the path.
01:27:08 And we want to be better about this.
01:27:10 We’re actually hiring, want to hire some new UI people.
01:27:12 UI people, you mentioned this.
01:27:14 Cause it’s such an, it’s a UI problem too, right?
01:27:17 We have a great designer now, but you know,
01:27:19 we need people who are just going to like build this
01:27:21 and debug these UIs, QT people.
01:27:23 QT.
01:27:24 Is that what the UI is done with, is QT?
01:27:26 The new UI is in QT.
01:27:29 C++ QT?
01:27:32 Tesla uses it too.
01:27:33 Yeah.
01:27:34 We had some React stuff in there.
01:27:37 React JS or just React?
01:27:39 React is his own language, right?
01:27:41 React Native, React is a JavaScript framework.
01:27:44 Yeah.
01:27:45 So it’s all based on JavaScript, but it’s, you know,
01:27:48 I like C++.
01:27:51 What do you think about Dojo with Tesla
01:27:55 and their foray into what appears to be
01:28:00 specialized hardware for training your own nets?
01:28:05 I guess it’s something, maybe you can correct me,
01:28:07 from my shallow looking at it,
01:28:10 it seems like something like Google did with TPUs,
01:28:12 but specialized for driving data.
01:28:15 I don’t think it’s specialized for driving data.
01:28:18 It’s just legit, just TPU.
01:28:20 They want to go the Apple way,
01:28:22 basically everything required in the chain is done in house.
01:28:25 Well, so you have a problem right now,
01:28:27 and this is one of my concerns.
01:28:31 I really would like to see somebody deal with this.
01:28:33 If anyone out there is doing it,
01:28:35 I’d like to help them if I can.
01:28:38 You basically have two options right now to train.
01:28:40 One, your options are NVIDIA or Google.
01:28:45 So Google is not even an option.
01:28:50 Their TPUs are only available in Google Cloud.
01:28:53 Google has absolutely onerous
01:28:55 terms of service restrictions.
01:28:58 They may have changed it,
01:28:59 but back in Google’s terms of service,
01:29:00 it said explicitly you are not allowed to use Google Cloud ML
01:29:03 for training autonomous vehicles
01:29:05 or for doing anything that competes with Google
01:29:07 without Google’s prior written permission.
01:29:09 Wow, okay.
01:29:10 I mean, Google is not a platform company.
01:29:14 I wouldn’t touch TPUs with a 10 foot pole.
01:29:16 So that leaves you with the monopoly.
01:29:19 NVIDIA? NVIDIA.
01:29:21 So, I mean.
01:29:22 That you’re not a fan of.
01:29:23 Well, look, I was a huge fan of in 2016 NVIDIA.
01:29:28 Jensen came sat in the car.
01:29:31 Cool guy.
01:29:32 When the stock was $30 a share.
01:29:35 NVIDIA stock has skyrocketed.
01:29:38 I witnessed a real change
01:29:39 in who was in management over there in like 2018.
01:29:43 And now they are, let’s exploit.
01:29:46 Let’s take every dollar we possibly can
01:29:48 out of this ecosystem.
01:29:49 Let’s charge $10,000 for A100s
01:29:51 because we know we got the best shit in the game.
01:29:54 And let’s charge $10,000 for an A100
01:29:57 when it’s really not that different from a 3080,
01:30:00 which is 699.
01:30:03 The margins that they are making
01:30:05 off of those high end chips are so high
01:30:08 that, I mean, I think they’re shooting themselves
01:30:10 in the foot just from a business perspective.
01:30:12 Because there’s a lot of people talking like me now
01:30:14 who are like, somebody’s gotta take NVIDIA down.
01:30:19 Yeah.
01:30:19 Where they could dominate it.
01:30:21 NVIDIA could be the new Intel.
01:30:22 Yeah, to be inside everything essentially.
01:30:26 And yet the winners in certain spaces
01:30:30 like autonomous driving, the winners,
01:30:33 only the people who are like desperately falling back
01:30:36 and trying to catch up and have a ton of money,
01:30:38 like the big automakers are the ones
01:30:40 interested in partnering with NVIDIA.
01:30:43 Oh, and I think a lot of those things
01:30:44 are gonna fall through.
01:30:45 If I were NVIDIA, sell chips.
01:30:49 Sell chips at a reasonable markup.
01:30:52 To everybody.
01:30:53 To everybody.
01:30:53 Without any restrictions.
01:30:54 Without any restrictions.
01:30:56 Intel did this.
01:30:57 Look at Intel.
01:30:58 They had a great long run.
01:30:59 NVIDIA is trying to turn their,
01:31:01 they’re like trying to productize their chips
01:31:04 way too much.
01:31:05 They’re trying to extract way more value
01:31:07 than they can sustainably.
01:31:09 Sure, you can do it tomorrow.
01:31:10 Is it gonna up your share price?
01:31:12 Sure, if you’re one of those CEOs
01:31:13 who’s like, how much can I strip mine this company?
01:31:15 And I think, you know, and that’s what’s weird about it too.
01:31:17 Like the CEO is the founder.
01:31:19 It’s the same guy.
01:31:20 Yeah.
01:31:21 I mean, I still think Jensen’s a great guy.
01:31:22 He is great.
01:31:23 Why do this?
01:31:25 You have a choice.
01:31:26 You have a choice right now.
01:31:27 Are you trying to cash out?
01:31:28 Are you trying to buy a yacht?
01:31:30 If you are, fine.
01:31:32 But if you’re trying to be
01:31:34 the next huge semiconductor company, sell chips.
01:31:37 Well, the interesting thing about Jensen
01:31:40 is he is a big vision guy.
01:31:42 So he has a plan like for 50 years down the road.
01:31:48 So it makes me wonder like.
01:31:50 How does price gouging fit into it?
01:31:51 Yeah, how does that, like it’s,
01:31:54 it doesn’t seem to make sense as a plan.
01:31:57 I worry that he’s listening to the wrong people.
01:31:59 Yeah, that’s the sense I have too sometimes.
01:32:02 Because I, despite everything, I think NVIDIA
01:32:07 is an incredible company.
01:32:09 Well, one, so I’m deeply grateful to NVIDIA
01:32:12 for the products they’ve created in the past.
01:32:13 Me too.
01:32:14 Right?
01:32:15 And so.
01:32:16 The 1080 Ti was a great GPU.
01:32:18 Still have a lot of them.
01:32:18 Still is, yeah.
01:32:21 But at the same time, it just feels like,
01:32:26 feels like you don’t want to put all your stock in NVIDIA.
01:32:29 And so like Elon is doing, what Tesla is doing
01:32:32 with Autopilot and Dojo is the Apple way is,
01:32:37 because they’re not going to share Dojo with George Hott’s.
01:32:40 I know.
01:32:42 They should sell that chip.
01:32:43 Oh, they should sell that.
01:32:44 Even their accelerator.
01:32:46 The accelerator that’s in all the cars, the 30 watt one.
01:32:49 Sell it, why not?
01:32:51 So open it up.
01:32:52 Like make, why does Tesla have to be a car company?
01:32:55 Well, if you sell the chip, here’s what you get.
01:32:58 Yeah.
01:32:59 Make some money off the chips.
01:33:00 It doesn’t take away from your chip.
01:33:02 You’re going to make some money, free money.
01:33:03 And also the world is going to build an ecosystem
01:33:07 of tooling for you.
01:33:09 Right?
01:33:09 You’re not going to have to fix the bug in your 10H layer.
01:33:12 Someone else already did.
01:33:15 Well, the question, that’s an interesting question.
01:33:16 I mean, that’s the question Steve Jobs asked.
01:33:18 That’s the question Elon Musk is perhaps asking is,
01:33:24 do you want Tesla stuff inside other vehicles?
01:33:28 Inside, potentially inside like a iRobot vacuum cleaner.
01:33:32 Yeah.
01:33:34 I think you should decide where your advantages are.
01:33:37 I’m not saying Tesla should start selling battery packs
01:33:39 to automakers.
01:33:40 Because battery packs to automakers,
01:33:41 they are straight up in competition with you.
01:33:43 If I were Tesla, I’d keep the battery technology totally.
01:33:46 Yeah.
01:33:46 As far as we make batteries.
01:33:47 But the thing about the Tesla TPU is anybody can build that.
01:33:53 It’s just a question of, you know,
01:33:54 are you willing to spend the money?
01:33:57 It could be a huge source of revenue potentially.
01:34:00 Are you willing to spend a hundred million dollars?
01:34:02 Anyone can build it.
01:34:03 And someone will.
01:34:04 And a bunch of companies now are starting
01:34:06 trying to build AI accelerators.
01:34:08 Somebody is going to get the idea right.
01:34:10 And yeah, hopefully they don’t get greedy
01:34:13 because they’ll just lose to the next guy who finally,
01:34:15 and then eventually the Chinese are going to make knockoff
01:34:17 and video chips and that’s.
01:34:19 From your perspective,
01:34:20 I don’t know if you’re also paying attention
01:34:21 to stay on Tesla for a moment.
01:34:24 Dave, Elon Musk has talked about a complete rewrite
01:34:28 of the neural net that they’re using.
01:34:31 That seems to, again, I’m half paying attention,
01:34:34 but it seems to involve basically a kind of integration
01:34:39 of all the sensors to where it’s a four dimensional view.
01:34:44 You know, you have a 3D model of the world over time.
01:34:47 And then you can, I think it’s done both for the,
01:34:52 for the actually, you know,
01:34:53 so the neural network is able to,
01:34:55 in a more holistic way,
01:34:56 deal with the world and make predictions and so on,
01:34:59 but also to make the annotation task more, you know, easier.
01:35:04 Like you can annotate the world in one place
01:35:08 and then kind of distribute itself across the sensors
01:35:10 and across a different,
01:35:12 like the hundreds of tasks that are involved
01:35:15 in the Hydro Net.
01:35:16 What are your thoughts about this rewrite?
01:35:19 Is it just like some details that are kind of obvious
01:35:22 that are steps that should be taken,
01:35:24 or is there something fundamental
01:35:26 that could challenge your idea
01:35:27 that end to end is the right solution?
01:35:31 We’re in the middle of a big rewrite now as well.
01:35:33 We haven’t shipped a new model in a bit.
01:35:34 Of what kind?
01:35:36 We’re going from 2D to 3D.
01:35:38 Right now, all our stuff, like for example,
01:35:39 when the car pitches back,
01:35:40 the lane lines also pitch back
01:35:43 because we’re assuming the flat world hypothesis.
01:35:47 The new models do not do this.
01:35:48 The new models output everything in 3D.
01:35:50 But there’s still no annotation.
01:35:53 So the 3D is, it’s more about the output.
01:35:56 Yeah.
01:35:57 We have Zs in everything.
01:36:00 We’ve…
01:36:00 Zs.
01:36:01 Yeah.
01:36:02 We had a Zs.
01:36:03 We had a Zs.
01:36:04 We unified a lot of stuff as well.
01:36:06 We switched from TensorFlow to PyTorch.
01:36:10 My understanding of what Tesla’s thing is,
01:36:13 is that their annotator now annotates
01:36:15 across the time dimension.
01:36:16 Mm hmm.
01:36:19 I mean, cute.
01:36:22 Why are you building an annotator?
01:36:24 I find their entire pipeline.
01:36:28 I find your vision, I mean,
01:36:30 the vision of end to end very compelling,
01:36:32 but I also like the engineering of the data engine
01:36:35 that they’ve created.
01:36:37 In terms of supervised learning pipelines,
01:36:41 that thing is damn impressive.
01:36:43 You’re basically, the idea is that you have
01:36:47 hundreds of thousands of people
01:36:49 that are doing data collection for you
01:36:51 by doing their experience.
01:36:52 So that’s kind of similar to the Comma AI model.
01:36:55 And you’re able to mine that data
01:36:59 based on the kind of edge cases you need.
01:37:02 I think it’s harder to do in the end to end learning.
01:37:07 The mining of the right edge cases.
01:37:09 Like that’s where feature engineering
01:37:11 is actually really powerful
01:37:14 because like us humans are able to do
01:37:17 this kind of mining a little better.
01:37:19 But yeah, there’s obvious, as we know,
01:37:21 there’s obvious constraints and limitations to that idea.
01:37:25 Carpathia just tweeted, he’s like,
01:37:28 you get really interesting insights
01:37:29 if you sort your validation set by loss
01:37:33 and look at the highest loss examples.
01:37:36 Yeah.
01:37:37 So yeah, I mean, you can do,
01:37:39 we have a little data engine like thing.
01:37:42 We’re training a segment.
01:37:43 I know it’s not fancy.
01:37:44 It’s just like, okay, train the new segment,
01:37:48 run it on 100,000 images
01:37:50 and now take the thousand with highest loss.
01:37:52 Select a hundred of those by human,
01:37:54 put those, get those ones labeled, retrain, do it again.
01:37:57 And so it’s a much less well written data engine.
01:38:01 And yeah, you can take these things really far
01:38:03 and it is impressive engineering.
01:38:06 And if you truly need supervised data for a problem,
01:38:09 yeah, things like data engine are at the high end
01:38:12 of what is attention?
01:38:14 Is a human paying attention?
01:38:15 I mean, we’re going to probably build something
01:38:17 that looks like data engine
01:38:18 to push our driver monitoring further.
01:38:21 But for driving itself,
01:38:22 you have it all annotated beautifully by what the human does.
01:38:26 Yeah, that’s interesting.
01:38:27 I mean, that applies to driver attention as well.
01:38:30 Do you want to detect the eyes?
01:38:31 Do you want to detect blinking and pupil movement?
01:38:33 Do you want to detect all the like face alignments
01:38:36 or landmark detection and so on,
01:38:38 and then doing kind of reasoning based on that?
01:38:41 Or do you want to take the entirety of the face over time
01:38:43 and do end to end?
01:38:45 I mean, it’s obvious that eventually you have to do end
01:38:48 to end with some calibration, some fixes and so on,
01:38:51 but it’s like, I don’t know when that’s the right move.
01:38:55 Even if it’s end to end, there actually is,
01:38:58 there is no kind of, you have to supervise that with humans.
01:39:03 Whether a human is paying attention or not
01:39:05 is a completely subjective judgment.
01:39:08 Like you can try to like automatically do it
01:39:11 with some stuff, but you don’t have,
01:39:13 if I record a video of a human,
01:39:15 I don’t have true annotations anywhere in that video.
01:39:18 The only way to get them is with,
01:39:21 you know, other humans labeling it really.
01:39:22 Well, I don’t know.
01:39:26 If you think deeply about it,
01:39:28 you could, you might be able to just,
01:39:30 depending on the task,
01:39:31 maybe a discover self annotating things like,
01:39:34 you know, you can look at like steering wheel reverse
01:39:36 or something like that.
01:39:37 You can discover little moments of lapse of attention.
01:39:41 I mean, that’s where psychology comes in.
01:39:44 Is there indicate,
01:39:45 cause you have so much data to look at.
01:39:48 So you might be able to find moments when there’s like,
01:39:51 just inattention that even with smartphone,
01:39:54 if you want to detect smartphone use,
01:39:56 you can start to zoom in.
01:39:57 I mean, that’s the gold mine, sort of the comma AI.
01:40:01 I mean, Tesla is doing this too, right?
01:40:02 Is they’re doing annotation based on,
01:40:06 it’s like a self supervised learning too.
01:40:10 It’s just a small part of the entire picture.
01:40:13 That’s kind of the challenge of solving a problem
01:40:17 in machine learning.
01:40:18 If you can discover self annotating parts of the problem,
01:40:24 right?
01:40:25 Our driver monitoring team is half a person right now.
01:40:27 I would, you know, once we have,
01:40:29 once we have two, three people on that team,
01:40:33 I definitely want to look at self annotating stuff
01:40:35 for attention.
01:40:38 Let’s go back for a sec to a comma and what,
01:40:43 you know, for people who are curious to try it out,
01:40:46 how do you install a comma in say a 2020 Toyota Corolla
01:40:51 or like, what are the cars that are supported?
01:40:53 What are the cars that you recommend?
01:40:55 And what does it take?
01:40:57 You have a few videos out, but maybe through words,
01:41:00 can you explain what’s it take to actually install a thing?
01:41:02 So we support, I think it’s 91 cars, 91 makes the models.
01:41:08 We’ve got to 100 this year.
01:41:10 Nice.
01:41:11 The, yeah, the 2020 Corolla, great choice.
01:41:16 The 2020 Sonata, it’s using the stock longitudinal.
01:41:21 It’s using just our lateral control,
01:41:23 but it’s a very refined car.
01:41:25 Their longitudinal control is not bad at all.
01:41:28 So yeah, Corolla, Sonata,
01:41:31 or if you’re willing to get your hands a little dirty
01:41:34 and look in the right places on the internet,
01:41:35 the Honda Civic is great,
01:41:37 but you’re going to have to install a modified EPS firmware
01:41:40 in order to get a little bit more torque.
01:41:42 And I can’t help you with that.
01:41:43 Comma does not officially endorse that,
01:41:45 but we have been doing it.
01:41:47 We didn’t ever release it.
01:41:49 We waited for someone else to discover it.
01:41:51 And then, you know.
01:41:52 And you have a Discord server where people,
01:41:55 there’s a very active developer community, I suppose.
01:42:00 So depending on the level of experimentation
01:42:04 you’re willing to do, that’s the community.
01:42:07 If you just want to buy it and you have a supported car,
01:42:11 it’s 10 minutes to install.
01:42:13 There’s YouTube videos.
01:42:15 It’s Ikea furniture level.
01:42:17 If you can set up a table from Ikea,
01:42:19 you can install a Comma 2 in your supported car
01:42:21 and it will just work.
01:42:22 Now you’re like, oh, but I want this high end feature
01:42:24 or I want to fix this bug.
01:42:26 Okay, well, welcome to the developer community.
01:42:29 So what, if I wanted to,
01:42:31 this is something I asked you offline like a few months ago.
01:42:34 If I wanted to run my own code to,
01:42:39 so use Comma as a platform
01:42:43 and try to run something like OpenPilot,
01:42:46 what does it take to do that?
01:42:48 So there’s a toggle in the settings called enable SSH.
01:42:51 And if you toggle that, you can SSH into your device.
01:42:54 You can modify the code.
01:42:55 You can upload whatever code you want to it.
01:42:58 There’s a whole lot of people.
01:42:59 So about 60% of people are running stock comma.
01:43:03 About 40% of people are running forks.
01:43:05 And there’s a community of,
01:43:07 there’s a bunch of people who maintain these forks
01:43:10 and these forks support different cars
01:43:13 or they have different toggles.
01:43:15 We try to keep away from the toggles
01:43:17 that are like disabled driver monitoring,
01:43:18 but there’s some people might want that kind of thing
01:43:21 and like, yeah, you can, it’s your car.
01:43:24 I’m not here to tell you.
01:43:29 We have some, we ban,
01:43:31 if you’re trying to subvert safety features,
01:43:32 you’re banned from our Discord.
01:43:33 I don’t want anything to do with you,
01:43:35 but there’s some forks doing that.
01:43:37 Got it.
01:43:39 So you encourage responsible forking.
01:43:42 Yeah, yeah.
01:43:43 We encourage, some people, yeah, some people,
01:43:46 like there’s forks that will do,
01:43:48 some people just like having a lot of readouts on the UI,
01:43:52 like a lot of like flashing numbers.
01:43:53 So there’s forks that do that.
01:43:55 Some people don’t like the fact that it disengages
01:43:57 when you press the gas pedal.
01:43:58 There’s forks that disable that.
01:44:00 Got it.
01:44:01 Now the stock experience is what like,
01:44:04 so it does both lane keeping
01:44:06 and longitudinal control all together.
01:44:08 So it’s not separate like it is in autopilot.
01:44:11 No, so, okay.
01:44:12 Some cars we use the stock longitudinal control.
01:44:15 We don’t do the longitudinal control in all the cars.
01:44:17 Some cars, the ACCs are pretty good in the cars.
01:44:19 It’s the lane keep that’s atrocious in anything
01:44:21 except for autopilot and super cruise.
01:44:23 But, you know, you just turn it on and it works.
01:44:27 What does this engagement look like?
01:44:29 Yeah, so we have, I mean,
01:44:30 I’m very concerned about mode confusion.
01:44:32 I’ve experienced it on super cruise and autopilot
01:44:36 where like autopilot, like autopilot disengages.
01:44:39 I don’t realize that the ACC is still on.
01:44:42 The lead car moves slightly over
01:44:44 and then the Tesla accelerates
01:44:46 to like whatever my set speed is super fast.
01:44:48 I’m like, what’s going on here?
01:44:51 We have engaged and disengaged.
01:44:53 And this is similar to my understanding, I’m not a pilot,
01:44:56 but my understanding is either the pilot is in control
01:45:00 or the copilot is in control.
01:45:02 And we have the same kind of transition system.
01:45:05 Either open pilot is engaged or open pilot is disengaged.
01:45:08 Engage with cruise control,
01:45:10 disengage with either gas brake or cancel.
01:45:13 Let’s talk about money.
01:45:14 What’s the business strategy for Kama?
01:45:17 Profitable.
01:45:18 Well, so you’re.
01:45:19 We did it.
01:45:20 So congratulations.
01:45:23 What, so basically selling,
01:45:25 so we should say Kama cost a thousand bucks, Kama two?
01:45:29 200 for the interface to the car as well.
01:45:31 It’s 1200, I’ll send that.
01:45:34 Nobody’s usually upfront like this.
01:45:36 Yeah, you gotta add the tack on, right?
01:45:38 Yeah.
01:45:39 I love it.
01:45:39 I’m not gonna lie to you.
01:45:41 Trust me, it will add $1,200 of value to your life.
01:45:43 Yes, it’s still super cheap.
01:45:45 30 days, no questions asked, money back guarantee,
01:45:47 and prices are only going up.
01:45:50 If there ever is future hardware,
01:45:52 it could cost a lot more than $1,200.
01:45:53 So Kama three is in the works.
01:45:56 It could be.
01:45:57 All I will say is future hardware
01:45:59 is going to cost a lot more than the current hardware.
01:46:02 Yeah, the people that use,
01:46:05 the people I’ve spoken with that use Kama,
01:46:07 that use open pilot,
01:46:10 first of all, they use it a lot.
01:46:12 So people that use it, they fall in love with it.
01:46:14 Oh, our retention rate is insane.
01:46:16 It’s a good sign.
01:46:17 Yeah.
01:46:18 It’s a really good sign.
01:46:19 70% of Kama two buyers are daily active users.
01:46:23 Yeah, it’s amazing.
01:46:27 Oh, also, we don’t plan on stopping selling the Kama two.
01:46:30 Like it’s, you know.
01:46:31 So whatever you create that’s beyond Kama two,
01:46:36 it would be potentially a phase shift.
01:46:40 Like it’s so much better that,
01:46:42 like you could use Kama two
01:46:44 and you can use Kama whatever.
01:46:45 Depends what you want.
01:46:46 It’s 3.41, 42.
01:46:48 Yeah.
01:46:49 You know, autopilot hardware one versus hardware two.
01:46:52 The Kama two is kind of like hardware one.
01:46:53 Got it, got it.
01:46:54 You can still use both.
01:46:55 Got it, got it.
01:46:56 I think I heard you talk about retention rate
01:46:58 with the VR headsets that the average is just once.
01:47:01 Yeah.
01:47:02 Just fast.
01:47:02 I mean, it’s such a fascinating way
01:47:03 to think about technology.
01:47:05 And this is a really, really good sign.
01:47:07 And the other thing that people say about Kama
01:47:09 is like they can’t believe they’re getting this 4,000 bucks.
01:47:12 Right?
01:47:12 It seems like some kind of steal.
01:47:17 So, but in terms of like longterm business strategies
01:47:20 that basically to put,
01:47:21 so it’s currently in like a thousand plus cars.
01:47:27 1,200.
01:47:28 More, more.
01:47:30 So yeah, dailies is about, dailies is about 2,000.
01:47:35 Weeklys is about 2,500, monthlys is over 3,000.
01:47:38 Wow.
01:47:39 We’ve grown a lot since we last talked.
01:47:42 Is the goal, like can we talk crazy for a second?
01:47:44 I mean, what’s the goal to overtake Tesla?
01:47:48 Let’s talk, okay, so.
01:47:49 I mean, Android did overtake iOS.
01:47:51 That’s exactly it, right?
01:47:52 So they did it.
01:47:55 I actually don’t know the timeline of that one.
01:47:57 But let’s talk, because everything is in alpha now.
01:48:02 The autopilot you could argue is in alpha
01:48:03 in terms of towards the big mission
01:48:05 of autonomous driving, right?
01:48:07 And so what, yeah, is your goal to overtake
01:48:11 millions of cars essentially?
01:48:13 Of course.
01:48:15 Where would it stop?
01:48:16 Like it’s open source software.
01:48:18 It might not be millions of cars
01:48:19 with a piece of comma hardware, but yeah.
01:48:21 I think open pilot at some point
01:48:24 will cross over autopilot in users,
01:48:26 just like Android crossed over iOS.
01:48:29 How does Google make money from Android?
01:48:31 It’s complicated.
01:48:34 Their own devices make money.
01:48:37 Google, Google makes money
01:48:39 by just kind of having you on the internet.
01:48:42 Yes.
01:48:43 Google search is built in, Gmail is built in.
01:48:45 Android is just a shill
01:48:46 for the rest of Google’s ecosystem.
01:48:48 Yeah, but the problem is Android is not,
01:48:50 is a brilliant thing.
01:48:52 I mean, Android arguably changed the world.
01:48:55 So there you go.
01:48:56 That’s, you can feel good ethically speaking.
01:49:00 But as a business strategy, it’s questionable.
01:49:04 Or sell hardware.
01:49:05 Sell hardware.
01:49:06 I mean, it took Google a long time to come around to it,
01:49:08 but they are now making money on the Pixel.
01:49:10 You’re not about money, you’re more about winning.
01:49:13 Yeah, of course.
01:49:14 No, but if only 10% of open pilot devices
01:49:18 come from comma AI.
01:49:19 They still make a lot.
01:49:20 That is still, yes.
01:49:21 That is a ton of money for our company.
01:49:22 But can’t somebody create a better comma using open pilot?
01:49:27 Or are you basically saying, well, I’ll compete them?
01:49:28 Well, I’ll compete you.
01:49:29 Can you create a better Android phone than the Google Pixel?
01:49:32 Right.
01:49:32 I mean, you can, but like, you know.
01:49:34 I love that.
01:49:35 So you’re confident, like, you know
01:49:37 what the hell you’re doing.
01:49:38 Yeah.
01:49:40 It’s confidence and merit.
01:49:43 I mean, our money comes from, we’re
01:49:44 a consumer electronics company.
01:49:46 Yeah.
01:49:46 And put it this way.
01:49:48 So we sold like 3,000 comma twos.
01:49:51 2,500 right now.
01:49:54 And like, OK, we’re probably going
01:49:59 to sell 10,000 units next year.
01:50:01 10,000 units, even just $1,000 a unit, OK,
01:50:04 we’re at 10 million in revenue.
01:50:09 Get that up to 100,000, maybe double the price of the unit.
01:50:12 Now we’re talking like 200 million revenue.
01:50:13 We’re talking like series.
01:50:14 Yeah, actually making money.
01:50:15 One of the rare semi autonomous or autonomous vehicle companies
01:50:19 that are actually making money.
01:50:21 Yeah.
01:50:22 You know, if you look at a model,
01:50:24 and we were just talking about this yesterday.
01:50:26 If you look at a model, and like you’re AB testing your model,
01:50:29 and if you’re one branch of the AB test,
01:50:32 the losses go down very fast in the first five epochs.
01:50:35 That model is probably going to converge
01:50:37 to something considerably better than the one
01:50:39 where the losses are going down slower.
01:50:41 Why do people think this is going to stop?
01:50:43 Why do people think one day there’s
01:50:44 going to be a great like, well, Waymo’s eventually
01:50:46 going to surpass you guys?
01:50:49 Well, they’re not.
01:50:52 Do you see like a world where like a Tesla or a car
01:50:55 like a Tesla would be able to basically press a button
01:50:59 and you like switch to open pilot?
01:51:01 You know, you load in.
01:51:04 No, so I think so first off, I think
01:51:06 that we may surpass Tesla in terms of users.
01:51:10 I do not think we’re going to surpass Tesla ever
01:51:12 in terms of revenue.
01:51:13 I think Tesla can capture a lot more revenue per user
01:51:16 than we can.
01:51:17 But this mimics the Android iOS model exactly.
01:51:20 There may be more Android devices,
01:51:22 but there’s a lot more iPhones than Google Pixels.
01:51:24 So I think there’ll be a lot more Tesla cars sold
01:51:26 than pieces of common hardware.
01:51:30 And then as far as a Tesla owner being
01:51:34 able to switch to open pilot, does iPhones run Android?
01:51:40 No, but it doesn’t make sense.
01:51:42 You can if you really want to do it,
01:51:43 but it doesn’t really make sense.
01:51:44 Like it’s not.
01:51:45 It doesn’t make sense.
01:51:46 Who cares?
01:51:46 What about if a large company like automakers, Ford, GM,
01:51:51 Toyota came to George Hots?
01:51:53 Or on the tech space, Amazon, Facebook, Google
01:51:58 came with a large pile of cash?
01:52:01 Would you consider being purchased?
01:52:07 Do you see that as a one possible?
01:52:10 Not seriously, no.
01:52:12 I would probably see how much shit they’ll entertain for me.
01:52:19 And if they’re willing to jump through a bunch of my hoops,
01:52:22 then maybe.
01:52:22 But no, not the way that M&A works today.
01:52:25 I mean, we’ve been approached.
01:52:26 And I laugh in these people’s faces.
01:52:28 I’m like, are you kidding?
01:52:31 Yeah.
01:52:31 Because it’s so demeaning.
01:52:33 The M&A people are so demeaning to companies.
01:52:36 They treat the startup world as their innovation ecosystem.
01:52:41 And they think that I’m cool with going along with that,
01:52:43 so I can have some of their scam fake Fed dollars.
01:52:46 Fed coin.
01:52:47 What am I going to do with more Fed coin?
01:52:49 Fed coin.
01:52:50 Fed coin, man.
01:52:51 I love that.
01:52:52 So that’s the cool thing about podcasting,
01:52:54 actually, is people criticize.
01:52:56 I don’t know if you’re familiar with Spotify giving Joe Rogan
01:53:00 $100 million.
01:53:01 I don’t know about that.
01:53:03 And they respect, despite all the shit
01:53:08 that people are talking about Spotify,
01:53:11 people understand that podcasters like Joe Rogan
01:53:15 know what the hell they’re doing.
01:53:17 So they give them money and say, just do what you do.
01:53:21 And the equivalent for you would be like,
01:53:25 George, do what the hell you do, because you’re good at it.
01:53:28 Try not to murder too many people.
01:53:31 There’s some kind of common sense things,
01:53:33 like just don’t go on a weird rampage of it.
01:53:37 Yeah.
01:53:38 It comes down to what companies I could respect, right?
01:53:43 Could I respect GM?
01:53:44 Never.
01:53:46 No, I couldn’t.
01:53:47 I mean, could I respect a Hyundai?
01:53:50 More so.
01:53:52 That’s a lot closer.
01:53:53 Toyota?
01:53:54 What’s your?
01:53:55 Nah.
01:53:56 Nah.
01:53:57 Korean is the way.
01:53:59 I think that the Japanese, the Germans, the US, they’re all
01:54:02 too, they’re all too, they all think they’re too great.
01:54:05 What about the tech companies?
01:54:07 Apple?
01:54:08 Apple is, of the tech companies that I could respect,
01:54:11 Apple’s the closest.
01:54:12 Yeah.
01:54:12 I mean, I could never.
01:54:13 It would be ironic.
01:54:14 It would be ironic if Comma AI is acquired by Apple.
01:54:19 I mean, Facebook, look, I quit Facebook 10 years ago
01:54:21 because I didn’t respect the business model.
01:54:24 Google has declined so fast in the last five years.
01:54:28 What are your thoughts about Waymo and its present
01:54:32 and its future?
01:54:33 Let me start by saying something nice, which is I’ve
01:54:39 visited them a few times and have ridden in their cars.
01:54:45 And the engineering that they’re doing,
01:54:49 both the research and the actual development
01:54:51 and the engineering they’re doing
01:54:53 and the scale they’re actually achieving
01:54:55 by doing it all themselves is really impressive.
01:54:58 And the balance of safety and innovation.
01:55:01 And the cars work really well for the routes they drive.
01:55:07 It drives fast, which was very surprising to me.
01:55:10 It drives the speed limit or faster than the speed limit.
01:55:14 It goes.
01:55:16 And it works really damn well.
01:55:17 And the interface is nice.
01:55:19 In Chandler, Arizona, yeah.
01:55:20 Yeah, in Chandler, Arizona, very specific environment.
01:55:22 So it gives me enough material in my mind
01:55:27 to push back against the madmen of the world,
01:55:30 like George Hotz, to be like, because you kind of imply
01:55:36 there’s zero probability they’re going to win.
01:55:38 And after I’ve used, after I’ve ridden in it, to me,
01:55:43 it’s not zero.
01:55:44 Oh, it’s not for technology reasons.
01:55:46 Bureaucracy?
01:55:48 No, it’s worse than that.
01:55:49 It’s actually for product reasons, I think.
01:55:51 Oh, you think they’re just not capable of creating
01:55:53 an amazing product?
01:55:55 No, I think that the product that they’re building
01:55:58 doesn’t make sense.
01:56:01 So a few things.
01:56:03 You say the Waymo’s are fast.
01:56:05 Benchmark a Waymo against a competent Uber driver.
01:56:09 Right.
01:56:09 Right?
01:56:10 The Uber driver’s faster.
01:56:11 It’s not even about speed.
01:56:12 It’s the thing you said.
01:56:13 It’s about the experience of being stuck at a stop sign
01:56:16 because pedestrians are crossing nonstop.
01:56:20 I like when my Uber driver doesn’t come to a full stop
01:56:22 at the stop sign.
01:56:22 Yeah.
01:56:23 You know?
01:56:24 And so let’s say the Waymo’s are 20% slower than an Uber.
01:56:31 Right?
01:56:33 You can argue that they’re going to be cheaper.
01:56:35 And I argue that users already have the choice
01:56:37 to trade off money for speed.
01:56:39 It’s called UberPool.
01:56:42 I think it’s like 15% of rides are UberPools.
01:56:45 Right?
01:56:46 Users are not willing to trade off money for speed.
01:56:49 So the whole product that they’re building
01:56:52 is not going to be competitive with traditional ride sharing
01:56:56 networks.
01:56:56 Right.
01:56:59 And also, whether there’s profit to be made
01:57:04 depends entirely on one company having a monopoly.
01:57:07 I think that the level four autonomous ride sharing
01:57:11 vehicles market is going to look a lot like the scooter market
01:57:14 if even the technology does come to exist, which I question.
01:57:18 Who’s doing well in that market?
01:57:20 It’s a race to the bottom.
01:57:22 Well, it could be closer like an Uber and a Lyft,
01:57:25 where it’s just one or two players.
01:57:28 Well, the scooter people have given up
01:57:31 trying to market scooters as a practical means
01:57:34 of transportation.
01:57:35 And they’re just like, they’re super fun to ride.
01:57:37 Look at wheels.
01:57:38 I love those things.
01:57:39 And they’re great on that front.
01:57:40 Yeah.
01:57:41 But from an actual transportation product
01:57:43 perspective, I do not think scooters are viable.
01:57:46 And I do not think level four autonomous cars are viable.
01:57:49 If you, let’s play a fun experiment.
01:57:51 If you ran, let’s do a Tesla and let’s do Waymo.
01:57:56 If Elon Musk took a vacation for a year, he just said,
01:58:01 screw it, I’m going to go live on an island, no electronics.
01:58:05 And the board decides that we need to find somebody
01:58:07 to run the company.
01:58:09 And they did decide that you should run the company
01:58:11 for a year.
01:58:12 How do you run Tesla differently?
01:58:14 I wouldn’t change much.
01:58:16 Do you think they’re on the right track?
01:58:17 I wouldn’t change.
01:58:18 I mean, I’d have some minor changes.
01:58:21 But even my debate with Tesla about end
01:58:25 to end versus SegNets, that’s just software.
01:58:29 Who cares?
01:58:30 It’s not like you’re doing something terrible with SegNets.
01:58:33 You’re probably building something that’s
01:58:35 at least going to help you debug the end to end system a lot.
01:58:39 It’s very easy to transition from what they have
01:58:42 to an end to end kind of thing.
01:58:45 And then I presume you would, in the Model Y
01:58:50 or maybe in the Model 3, start adding driver
01:58:52 sensing with infrared.
01:58:53 Yes, I would add infrared camera, infrared lights
01:58:58 right away to those cars.
01:59:02 And start collecting that data and do all that kind of stuff,
01:59:04 yeah.
01:59:05 Very much.
01:59:06 I think they’re already kind of doing it.
01:59:07 It’s an incredibly minor change.
01:59:09 If I actually were CEO of Tesla, first off,
01:59:11 I’d be horrified that I wouldn’t be able to do
01:59:13 a better job as Elon.
01:59:14 And then I would try to understand
01:59:16 the way he’s done things before.
01:59:17 You would also have to take over his Twitter.
01:59:20 I don’t tweet.
01:59:22 Yeah, what’s your Twitter situation?
01:59:24 Why are you so quiet on Twitter?
01:59:25 Since Dukama is like what’s your social network presence like?
01:59:30 Because on Instagram, you do live streams.
01:59:34 You understand the music of the internet,
01:59:39 but you don’t always fully engage into it.
01:59:41 You’re part time.
01:59:42 Well, I used to have a Twitter.
01:59:44 Yeah, I mean, Instagram is a pretty place.
01:59:47 Instagram is a beautiful place.
01:59:49 It glorifies beauty.
01:59:49 I like Instagram’s values as a network.
01:59:53 Twitter glorifies conflict, glorifies shots,
02:00:00 taking shots of people.
02:00:01 And it’s like, you know, Twitter and Donald Trump
02:00:05 are perfectly, they’re perfect for each other.
02:00:08 So Tesla’s on the right track in your view.
02:00:12 OK, so let’s try, let’s really try this experiment.
02:00:16 If you ran Waymo, let’s say they’re,
02:00:19 I don’t know if you agree, but they
02:00:21 seem to be at the head of the pack of the kind of,
02:00:25 what would you call that approach?
02:00:27 Like it’s not necessarily lighter based
02:00:29 because it’s not about lighter.
02:00:30 Level four robotaxi.
02:00:31 Level four robotaxi, all in before making any revenue.
02:00:37 So they’re probably at the head of the pack.
02:00:38 If you were said, hey, George, can you
02:00:42 please run this company for a year, how would you change it?
02:00:47 I would go.
02:00:47 I would get Anthony Levandowski out of jail,
02:00:49 and I would put him in charge of the company.
02:00:56 Well, let’s try to break that apart.
02:00:58 Why do you want to destroy the company by doing that?
02:01:01 Or do you mean you like renegade style thinking that pushes,
02:01:09 that throws away bureaucracy and goes
02:01:11 to first principle thinking?
02:01:12 What do you mean by that?
02:01:14 I think Anthony Levandowski is a genius,
02:01:16 and I think he would come up with a much better idea of what
02:01:19 to do with Waymo than me.
02:01:22 So you mean that unironically.
02:01:23 He is a genius.
02:01:24 Oh, yes.
02:01:25 Oh, absolutely.
02:01:26 Without a doubt.
02:01:27 I mean, I’m not saying there’s no shortcomings,
02:01:30 but in the interactions I’ve had with him, yeah.
02:01:34 What?
02:01:35 He’s also willing to take, like, who knows
02:01:38 what he would do with Waymo?
02:01:39 I mean, he’s also out there, like far more out there
02:01:41 than I am.
02:01:41 Yeah, there’s big risks.
02:01:43 What do you make of him?
02:01:44 I was going to talk to him on this podcast,
02:01:47 and I was going back and forth.
02:01:48 I’m such a gullible, naive human.
02:01:51 Like, I see the best in people.
02:01:53 And I slowly started to realize that there
02:01:56 might be some people out there that, like,
02:02:02 have multiple faces to the world.
02:02:05 They’re, like, deceiving and dishonest.
02:02:08 I still refuse to, like, I just, I trust people,
02:02:13 and I don’t care if I get hurt by it.
02:02:14 But, like, you know, sometimes you
02:02:16 have to be a little bit careful, especially platform
02:02:18 wise and podcast wise.
02:02:21 What do you, what am I supposed to think?
02:02:23 So you think, you think he’s a good person?
02:02:26 Oh, I don’t know.
02:02:27 I don’t really make moral judgments.
02:02:30 It’s difficult to.
02:02:30 Oh, I mean this about the Waymo.
02:02:32 I actually, I mean that whole idea very nonironically
02:02:34 about what I would do.
02:02:36 The problem with putting me in charge of Waymo
02:02:38 is Waymo is already $10 billion in the hole, right?
02:02:41 Whatever idea Waymo does, look, commas profitable, commas
02:02:44 raised $8.1 million.
02:02:46 That’s small, you know, that’s small money.
02:02:48 Like, I can build a reasonable consumer electronics company
02:02:50 and succeed wildly at that and still never be able to pay back
02:02:54 Waymo’s $10 billion.
02:02:55 So I think the basic idea with Waymo, well,
02:02:58 forget the $10 billion because they have some backing,
02:03:00 but your basic thing is, like, what can we do
02:03:04 to start making some money?
02:03:05 Well, no, I mean, my bigger idea is, like,
02:03:07 whatever the idea is that’s gonna save Waymo,
02:03:10 I don’t have it.
02:03:11 It’s gonna have to be a big risk idea
02:03:13 and I cannot think of a better person
02:03:15 than Anthony Levandowski to do it.
02:03:17 So that is completely what I would do as CEO of Waymo.
02:03:20 I would call myself a transitionary CEO,
02:03:22 do everything I can to fix that situation up.
02:03:24 I’m gonna see.
02:03:25 Yeah.
02:03:27 Yeah.
02:03:28 Because I can’t do it, right?
02:03:29 Like, I can’t, I mean, I can talk about how
02:03:33 what I really wanna do is just apologize
02:03:35 for all those corny, you know, ad campaigns
02:03:38 and be like, here’s the real state of the technology.
02:03:40 Yeah, that’s, like, I have several criticism.
02:03:42 I’m a little bit more bullish on Waymo
02:03:44 than you seem to be, but one criticism I have
02:03:48 is it went into corny mode too early.
02:03:50 Like, it’s still a startup.
02:03:52 It hasn’t delivered on anything.
02:03:53 So it should be, like, more renegade
02:03:56 and show off the engineering that they’re doing,
02:03:59 which just can be impressive,
02:04:00 as opposed to doing these weird commercials
02:04:02 of, like, your friendly car company.
02:04:07 I mean, that’s my biggest snipe at Waymo is always,
02:04:10 that guy’s a paid actor.
02:04:11 That guy’s not a Waymo user.
02:04:12 He’s a paid actor.
02:04:13 Look here, I found his call sheet.
02:04:15 Do kind of like what SpaceX is doing
02:04:17 with the rocket launches.
02:04:18 Just put the nerds up front, put the engineers up front,
02:04:22 and just, like, show failures too, just.
02:04:25 I love SpaceX’s, yeah.
02:04:27 Yeah, the thing that they’re doing is right,
02:04:29 and it just feels like the right.
02:04:31 But.
02:04:32 We’re all so excited to see them succeed.
02:04:34 Yeah.
02:04:35 I can’t wait to see when it won’t fail, you know?
02:04:37 Like, you lie to me, I want you to fail.
02:04:39 You tell me the truth, you be honest with me,
02:04:41 I want you to succeed.
02:04:42 Yeah.
02:04:44 Ah, yeah, and that requires the renegade CEO, right?
02:04:50 I’m with you, I’m with you.
02:04:51 I still have a little bit of faith in Waymo
02:04:54 for the renegade CEO to step forward, but.
02:04:57 It’s not, it’s not John Kraftik.
02:05:00 Yeah, it’s, you can’t.
02:05:02 It’s not Chris Hormiston.
02:05:04 And those people may be very good at certain things.
02:05:07 Yeah.
02:05:08 But they’re not renegades.
02:05:10 Yeah, because these companies are fundamentally,
02:05:12 even though we’re talking about billion dollars,
02:05:14 all these crazy numbers,
02:05:15 they’re still, like, early stage startups.
02:05:19 I mean, and I just, if you are pre revenue
02:05:21 and you’ve raised 10 billion dollars,
02:05:23 I have no idea, like, this just doesn’t work.
02:05:26 You know, it’s against everything Silicon Valley.
02:05:28 Where’s your minimum viable product?
02:05:29 You know, where’s your users?
02:05:31 Where’s your growth numbers?
02:05:33 This is traditional Silicon Valley.
02:05:36 Why do you not apply it to what you think
02:05:38 you’re too big to fail already, like?
02:05:41 How do you think autonomous driving will change society?
02:05:45 So the mission is, for comma, to solve self driving.
02:05:52 Do you have, like, a vision of the world
02:05:54 of how it’ll be different?
02:05:57 Is it as simple as A to B transportation?
02:06:00 Or is there, like, cause these are robots.
02:06:03 It’s not about autonomous driving in and of itself.
02:06:05 It’s what the technology enables.
02:06:09 It’s, I think it’s the coolest applied AI problem.
02:06:12 I like it because it has a clear path to monetary value.
02:06:17 But as far as that being the thing that changes the world,
02:06:21 I mean, no, like, there’s cute things we’re doing in common.
02:06:25 Like, who’d have thought you could stick a phone
02:06:26 on the windshield and it’ll drive.
02:06:29 But like, really, the product that you’re building
02:06:31 is not something that people were not capable
02:06:33 of imagining 50 years ago.
02:06:35 So no, it doesn’t change the world on that front.
02:06:37 Could people have imagined the internet 50 years ago?
02:06:39 Only true genius visionaries.
02:06:42 Everyone could have imagined autonomous cars 50 years ago.
02:06:45 It’s like a car, but I don’t drive it.
02:06:47 See, I have this sense, and I told you, like,
02:06:49 my longterm dream is robots with which you have deep,
02:06:55 with whom you have deep connections, right?
02:06:59 And there’s different trajectories towards that.
02:07:03 And I’ve been thinking,
02:07:04 so I’ve been thinking of launching a startup.
02:07:07 I see autonomous vehicles
02:07:09 as a potential trajectory to that.
02:07:11 That’s not where the direction I would like to go,
02:07:16 but I also see Tesla or even Comma AI,
02:07:19 like, pivoting into robotics broadly defined
02:07:24 at some stage in the way, like you’re mentioning,
02:07:27 the internet didn’t expect.
02:07:29 Let’s solve, you know, when I say a comma about this,
02:07:32 we could talk about this,
02:07:33 but let’s solve self driving cars first.
02:07:35 You gotta stay focused on the mission.
02:07:37 Don’t, don’t, don’t, you’re not too big to fail.
02:07:39 For however much I think Comma’s winning,
02:07:41 like, no, no, no, no, no, you’re winning
02:07:43 when you solve level five self driving cars.
02:07:45 And until then, you haven’t won.
02:07:46 And you know, again, you wanna be arrogant
02:07:48 in the face of other people, great.
02:07:50 You wanna be arrogant in the face of nature, you’re an idiot.
02:07:53 Stay mission focused, brilliantly put.
02:07:56 Like I mentioned, thinking of launching a startup,
02:07:58 I’ve been considering, actually, before COVID,
02:08:01 I’ve been thinking of moving to San Francisco.
02:08:03 Ooh, ooh, I wouldn’t go there.
02:08:06 So why is, well, and now I’m thinking
02:08:09 about potentially Austin and we’re in San Diego now.
02:08:13 San Diego, come here.
02:08:14 So why, what, I mean, you’re such an interesting human.
02:08:20 You’ve launched so many successful things.
02:08:23 What, why San Diego?
02:08:26 What do you recommend?
02:08:27 Why not San Francisco?
02:08:29 Have you thought, so in your case,
02:08:31 San Diego with Qualcomm and Snapdragon,
02:08:33 I mean, that’s an amazing combination.
02:08:36 But.
02:08:37 That wasn’t really why.
02:08:38 That wasn’t the why?
02:08:39 No, I mean, Qualcomm was an afterthought.
02:08:41 Qualcomm was, it was a nice thing to think about.
02:08:42 It’s like, you can have a tech company here.
02:08:45 Yeah.
02:08:45 And a good one, I mean, you know, I like Qualcomm, but.
02:08:48 No.
02:08:49 Well, so why San Diego better than San Francisco?
02:08:50 Why does San Francisco suck?
02:08:51 Well, so, okay, so first off,
02:08:53 we all kind of said like, we wanna stay in California.
02:08:55 People like the ocean.
02:08:57 You know, California, for its flaws,
02:09:00 it’s like a lot of the flaws of California
02:09:02 are not necessarily California as a whole,
02:09:03 and they’re much more San Francisco specific.
02:09:05 Yeah.
02:09:06 San Francisco, so I think first tier cities in general
02:09:09 have stopped wanting growth.
02:09:13 Well, you have like in San Francisco, you know,
02:09:15 the voting class always votes to not build more houses
02:09:18 because they own all the houses.
02:09:19 And they’re like, well, you know,
02:09:21 once people have figured out how to vote themselves
02:09:23 more money, they’re gonna do it.
02:09:25 It is so insanely corrupt.
02:09:27 It is not balanced at all, like political party wise,
02:09:31 you know, it’s a one party city and.
02:09:34 For all the discussion of diversity,
02:09:38 it stops lacking real diversity of thought,
02:09:42 of background, of approaches, of strategies, of ideas.
02:09:48 It’s kind of a strange place
02:09:51 that it’s the loudest people about diversity
02:09:54 and the biggest lack of diversity.
02:09:56 I mean, that’s what they say, right?
02:09:58 It’s the projection.
02:10:00 Projection, yeah.
02:10:02 Yeah, it’s interesting.
02:10:02 And even people in Silicon Valley tell me
02:10:04 that’s like high up people,
02:10:07 everybody is like, this is a terrible place.
02:10:10 It doesn’t make sense.
02:10:10 I mean, and coronavirus is really what killed it.
02:10:13 San Francisco was the number one exodus
02:10:17 during coronavirus.
02:10:18 We still think San Diego is a good place to be.
02:10:21 Yeah.
02:10:23 Yeah, I mean, we’ll see.
02:10:24 We’ll see what happens with California a bit longer term.
02:10:29 Like Austin’s an interesting choice.
02:10:32 I wouldn’t, I don’t have really anything bad to say
02:10:33 about Austin either,
02:10:35 except for the extreme heat in the summer,
02:10:37 which, but that’s like very on the surface, right?
02:10:40 I think as far as like an ecosystem goes, it’s cool.
02:10:43 I personally love Colorado.
02:10:45 Colorado’s great.
02:10:47 Yeah, I mean, you have these states that are,
02:10:49 like just way better run.
02:10:51 California is, you know, it’s especially San Francisco.
02:10:55 It’s not a tie horse and like, yeah.
02:10:58 Can I ask you for advice to me and to others
02:11:02 about what’s it take to build a successful startup?
02:11:07 Oh, I don’t know.
02:11:08 I haven’t done that.
02:11:09 Talk to someone who did that.
02:11:10 Well, you’ve, you know,
02:11:14 this is like another book of years
02:11:16 that I’ll buy for $67, I suppose.
02:11:18 So there’s, um.
02:11:20 One of these days I’ll sell out.
02:11:24 Yeah, that’s right.
02:11:24 Jailbreaks are going to be a dollar
02:11:26 and books are going to be 67.
02:11:27 How I jailbroke the iPhone by George Hots.
02:11:32 That’s right.
02:11:32 How I jail broke the iPhone and you can too.
02:11:35 You can too.
02:11:36 67 dollars.
02:11:37 In 21 days.
02:11:39 That’s right.
02:11:39 That’s right.
02:11:40 Oh God.
02:11:41 Okay, I can’t wait.
02:11:42 But quite, so you have an introspective,
02:11:44 you have built a very unique company.
02:11:49 I mean, not you, but you and others.
02:11:53 But I don’t know.
02:11:55 There’s no, there’s nothing.
02:11:56 You have an introspective,
02:11:57 you haven’t really sat down and thought about like,
02:12:01 well, like if you and I were having a bunch of,
02:12:04 we’re having some beers
02:12:06 and you’re seeing that I’m depressed
02:12:08 and whatever, I’m struggling.
02:12:09 There’s no advice you can give?
02:12:11 Oh, I mean.
02:12:13 More beer?
02:12:13 More beer?
02:12:15 Um, yeah, I think it’s all very like situation dependent.
02:12:23 Here’s, okay, if I can give a generic piece of advice,
02:12:25 it’s the technology always wins.
02:12:28 The better technology always wins.
02:12:30 And lying always loses.
02:12:35 Build technology and don’t lie.
02:12:38 I’m with you.
02:12:39 I agree very much.
02:12:40 The long run, long run.
02:12:41 Sure.
02:12:42 That’s the long run, yeah.
02:12:43 The market can remain irrational longer
02:12:44 than you can remain solvent.
02:12:46 True fact.
02:12:47 Well, this is an interesting point
02:12:49 because I ethically and just as a human believe that
02:12:54 like hype and smoke and mirrors is not
02:12:58 at any stage of the company is a good strategy.
02:13:02 I mean, there’s some like, you know,
02:13:04 PR magic kind of like, you know.
02:13:07 Oh, hype around a new product, right?
02:13:08 If there’s a call to action,
02:13:09 if there’s like a call to action,
02:13:10 like buy my new GPU, look at it.
02:13:13 It takes up three slots and it’s this big.
02:13:14 It’s huge.
02:13:15 Buy my GPU.
02:13:16 Yeah, that’s great.
02:13:17 If you look at, you know,
02:13:18 especially in the AI space broadly,
02:13:20 but autonomous vehicles,
02:13:22 like you can raise a huge amount of money on nothing.
02:13:26 And the question to me is like, I’m against that.
02:13:30 I’ll never be part of that.
02:13:31 I don’t think, I hope not, willingly not.
02:13:36 But like, is there something to be said
02:13:40 to essentially lying to raise money,
02:13:44 like fake it till you make it kind of thing?
02:13:47 I mean, this is Billy McFarland in the Fyre Festival.
02:13:50 Like we all experienced, you know,
02:13:53 what happens with that.
02:13:54 No, no, don’t fake it till you make it.
02:13:57 Be honest and hope you make it the whole way.
02:14:00 The technology wins.
02:14:01 Right, the technology wins.
02:14:02 And like, there is, I’m not used to like the anti hype,
02:14:06 you know, that’s a Slava KPSS reference,
02:14:08 but hype isn’t necessarily bad.
02:14:13 I loved camping out for the iPhones, you know,
02:14:17 and as long as the hype is backed by like substance,
02:14:21 as long as it’s backed by something I can actually buy,
02:14:23 and like it’s real, then hype is great
02:14:26 and it’s a great feeling.
02:14:28 It’s when the hype is backed by lies
02:14:30 that it’s a bad feeling.
02:14:32 I mean, a lot of people call Elon Musk a fraud.
02:14:34 How could he be a fraud?
02:14:35 I’ve noticed this, this kind of interesting effect,
02:14:37 which is he does tend to over promise
02:14:42 and deliver, what’s the better way to phrase it?
02:14:45 Promise a timeline that he doesn’t deliver on,
02:14:49 he delivers much later on.
02:14:51 What do you think about that?
02:14:52 Cause I do that, I think that’s a programmer thing too.
02:14:56 I do that as well.
02:14:57 You think that’s a really bad thing to do or is that okay?
02:15:01 I think that’s, again, as long as like,
02:15:03 you’re working toward it and you’re gonna deliver on it,
02:15:06 it’s not too far off, right?
02:15:10 Right?
02:15:11 Like, you know, the whole autonomous vehicle thing,
02:15:14 it’s like, I mean, I still think Tesla’s on track
02:15:18 to beat us.
02:15:19 I still think even with their missteps,
02:15:21 they have advantages we don’t have.
02:15:25 You know, Elon is better than me
02:15:28 at like marshaling massive amounts of resources.
02:15:33 So, you know, I still think given the fact
02:15:36 they’re maybe making some wrong decisions,
02:15:38 they’ll end up winning.
02:15:39 And like, it’s fine to hype it
02:15:42 if you’re actually gonna win, right?
02:15:44 Like if Elon says, look, we’re gonna be landing rockets
02:15:47 back on earth in a year and it takes four,
02:15:49 like, you know, he landed a rocket back on earth
02:15:53 and he was working toward it the whole time.
02:15:55 I think there’s some amount of like,
02:15:57 I think when it becomes wrong is if you know
02:15:59 you’re not gonna meet that deadline.
02:16:00 If you’re lying.
02:16:01 Yeah, that’s brilliantly put.
02:16:03 Like this is what people don’t understand, I think.
02:16:06 Like Elon believes everything he says.
02:16:09 He does, as far as I can tell, he does.
02:16:12 And I detected that in myself too.
02:16:14 Like if I, it’s only bullshit
02:16:17 if you’re like conscious of yourself lying.
02:16:21 Yeah, I think so.
02:16:22 Yeah.
02:16:23 Now you can’t take that to such an extreme, right?
02:16:25 Like in a way, I think maybe Billy McFarland
02:16:27 believed everything he said too.
02:16:30 Right, that’s how you start a cult
02:16:31 and everybody kills themselves.
02:16:33 Yeah.
02:16:34 Yeah, like it’s, you need, you need,
02:16:36 if there’s like some factor on it, it’s fine.
02:16:39 And you need some people to like, you know,
02:16:41 keep you in check, but like,
02:16:44 if you deliver on most of the things you say
02:16:46 and just the timelines are off, yeah.
02:16:48 It does piss people off though.
02:16:50 I wonder, but who cares?
02:16:53 In a long arc of history, the people,
02:16:55 everybody gets pissed off at the people who succeed,
02:16:58 which is one of the things
02:16:59 that frustrates me about this world,
02:17:01 is they don’t celebrate the success of others.
02:17:07 Like there’s so many people that want Elon to fail.
02:17:12 It’s so fascinating to me.
02:17:14 Like what is wrong with you?
02:17:18 Like, so Elon Musk talks about like people shorting,
02:17:21 like they talk about financial,
02:17:23 but I think it’s much bigger than the financials.
02:17:25 I’ve seen like the human factors community,
02:17:27 they want, they want other people to fail.
02:17:31 Why, why, why?
02:17:32 Like even people, the harshest thing is like,
02:17:36 you know, even people that like seem
02:17:38 to really hate Donald Trump, they want him to fail
02:17:41 or like the other president
02:17:43 or they want Barack Obama to fail.
02:17:45 It’s like.
02:17:47 Yeah, we’re all on the same boat, man.
02:17:49 It’s weird, but I want that,
02:17:51 I would love to inspire that part of the world to change
02:17:54 because damn it, if the human species is gonna survive,
02:17:58 we should celebrate success.
02:18:00 Like it seems like the efficient thing to do
02:18:02 in this objective function that we’re all striving for
02:18:06 is to celebrate the ones that like figure out
02:18:09 how to like do better at that objective function
02:18:11 as opposed to like dragging them down back into the mud.
02:18:16 I think there is, this is the speech I always give
02:18:19 about the commenters on Hacker News.
02:18:21 So first off, something to remember
02:18:23 about the internet in general is commenters
02:18:26 are not representative of the population.
02:18:29 I don’t comment on anything.
02:18:31 You know, commenters are representative
02:18:34 of a certain sliver of the population.
02:18:36 And on Hacker News, a common thing I’ll see
02:18:39 is when you’ll see something that’s like,
02:18:42 you know, promises to be wild out there and innovative.
02:18:47 There is some amount of, you know,
02:18:49 checking them back to earth,
02:18:50 but there’s also some amount of if this thing succeeds,
02:18:55 well, I’m 36 and I’ve worked
02:18:57 at large tech companies my whole life.
02:19:02 They can’t succeed because if they succeed,
02:19:05 that would mean that I could have done something different
02:19:07 with my life, but we know that I couldn’t have,
02:19:09 we know that I couldn’t have,
02:19:10 and that’s why they’re gonna fail.
02:19:11 And they have to root for them to fail
02:19:13 to kind of maintain their world image.
02:19:15 So tune it out.
02:19:17 And they comment, well, it’s hard, I, so one of the things,
02:19:21 one of the things I’m considering startup wise
02:19:25 is to change that.
02:19:27 Cause I think the, I think it’s also a technology problem.
02:19:31 It’s a platform problem.
02:19:33 I agree.
02:19:33 It’s like, because the thing you said,
02:19:35 most people don’t comment.
02:19:39 I think most people want to comment.
02:19:42 They just don’t because it’s all the assholes
02:19:45 who are commenting.
02:19:46 Exactly, I don’t want to be grouped in with them.
02:19:47 You don’t want to be at a party
02:19:49 where everyone is an asshole.
02:19:50 And so they, but that’s a platform problem.
02:19:54 I can’t believe what Reddit’s become.
02:19:56 I can’t believe the group thinking, Reddit comments.
02:20:00 There’s a, Reddit is an interesting one
02:20:02 because they’re subreddits.
02:20:05 And so you can still see, especially small subreddits
02:20:09 that like, that are a little like havens
02:20:11 of like joy and positivity and like deep,
02:20:16 even disagreement, but like nuanced discussion.
02:20:18 But it’s only like small little pockets,
02:20:21 but that’s emergent.
02:20:23 The platform is not helping that or hurting that.
02:20:26 So I guess naturally something about the internet,
02:20:31 if you don’t put in a lot of effort to encourage
02:20:34 nuance and positive, good vibes,
02:20:37 it’s naturally going to decline into chaos.
02:20:41 I would love to see someone do this well.
02:20:42 Yeah.
02:20:43 I think it’s, yeah, very doable.
02:20:45 I think actually, so I feel like Twitter
02:20:49 could be overthrown.
02:20:52 Yashua Bach talked about how like,
02:20:55 if you have like and retweet,
02:20:58 like that’s only positive wiring, right?
02:21:02 The only way to do anything like negative there
02:21:05 is with a comment.
02:21:08 And that’s like that asymmetry is what gives,
02:21:12 you know, Twitter its particular toxicness.
02:21:15 Whereas I find YouTube comments to be much better
02:21:18 because YouTube comments have an up and a down
02:21:21 and they don’t show the downvotes.
02:21:23 Without getting into depth of this particular discussion,
02:21:26 the point is to explore possibilities
02:21:29 and get a lot of data on it.
02:21:30 Because I mean, I could disagree with what you just said.
02:21:34 The point is it’s unclear.
02:21:36 It hasn’t been explored in a really rich way.
02:21:39 Like these questions of how to create platforms
02:21:44 that encourage positivity.
02:21:47 Yeah, I think it’s a technology problem.
02:21:49 And I think we’ll look back at Twitter as it is now.
02:21:51 Maybe it’ll happen within Twitter,
02:21:53 but most likely somebody overthrows them
02:21:56 is we’ll look back at Twitter and say,
02:22:00 can’t believe we put up with this level of toxicity.
02:22:03 You need a different business model too.
02:22:05 Any social network that fundamentally has advertising
02:22:07 as a business model, this was in The Social Dilemma,
02:22:10 which I didn’t watch, but I liked it.
02:22:11 It’s like, you know, there’s always the, you know,
02:22:12 you’re the product, you’re not the,
02:22:15 but they had a nuanced take on it that I really liked.
02:22:17 And it said, the product being sold is influence over you.
02:22:24 The product being sold is literally your,
02:22:27 you know, influence on you.
02:22:29 Like that can’t be, if that’s your idea, okay.
02:22:33 Well, you know, guess what?
02:22:35 It can’t not be toxic.
02:22:37 Yeah, maybe there’s ways to spin it,
02:22:39 like with giving a lot more control to the user
02:22:42 and transparency to see what is happening to them
02:22:44 as opposed to in the shadows, it’s possible,
02:22:47 but that can’t be the primary source of.
02:22:49 But the users aren’t, no one’s gonna use that.
02:22:51 It depends, it depends, it depends.
02:22:54 I think that the, you’re not going to,
02:22:57 you can’t depend on self awareness of the users.
02:23:00 It’s a longer discussion because you can’t depend on it,
02:23:04 but you can reward self awareness.
02:23:09 Like if for the ones who are willing to put in the work
02:23:12 of self awareness, you can reward them and incentivize
02:23:16 and perhaps be pleasantly surprised how many people
02:23:20 are willing to be self aware on the internet.
02:23:23 Like we are in real life.
02:23:24 Like I’m putting in a lot of effort with you right now,
02:23:26 being self aware about if I say something stupid or mean,
02:23:30 I’ll like look at your like body language.
02:23:32 Like I’m putting in that effort.
02:23:33 It’s costly for an introvert, very costly.
02:23:36 But on the internet, fuck it.
02:23:39 Like most people are like, I don’t care if this hurts
02:23:42 somebody, I don’t care if this is not interesting
02:23:46 or if this is, yeah, it’s a mean or whatever.
02:23:48 I think so much of the engagement today on the internet
02:23:50 is so disingenuine too.
02:23:53 You’re not doing this out of a genuine,
02:23:54 this is what you think.
02:23:55 You’re doing this just straight up to manipulate others.
02:23:57 Whether you’re in, you just became an ad.
02:23:59 Yeah, okay, let’s talk about a fun topic,
02:24:02 which is programming.
02:24:04 Here’s another book idea for you.
02:24:05 Let me pitch.
02:24:07 What’s your perfect programming setup?
02:24:09 So like this by George Hots.
02:24:12 So like what, listen, you’re.
02:24:17 Give me a MacBook Air, sit me in a corner of a hotel room
02:24:20 and you know I’ll still ask you.
02:24:21 So you really don’t care.
02:24:22 You don’t fetishize like multiple monitors, keyboard.
02:24:27 Those things are nice and I’m not gonna say no to them,
02:24:30 but did they automatically unlock tons of productivity?
02:24:33 No, not at all.
02:24:34 I have definitely been more productive on a MacBook Air
02:24:36 in a corner of a hotel room.
02:24:38 What about IDE?
02:24:41 So which operating system do you love?
02:24:45 What text editor do you use IDE?
02:24:49 What, is there something that is like the perfect,
02:24:53 if you could just say the perfect productivity setup
02:24:57 for George Hots.
02:24:57 It doesn’t matter.
02:24:58 It literally doesn’t matter.
02:25:00 You know, I guess I code most of the time in Vim.
02:25:03 Like literally I’m using an editor from the 70s.
02:25:05 You know, you didn’t make anything better.
02:25:07 Okay, VS code is nice for reading code.
02:25:09 There’s a few things that are nice about it.
02:25:10 I think that you can build much better tools.
02:25:13 How like IDA’s xrefs work way better than VS codes, why?
02:25:18 Yeah, actually that’s a good question, like why?
02:25:20 I still use, sorry, Emacs for most.
02:25:25 I’ve actually never, I have to confess something dark.
02:25:28 So I’ve never used Vim.
02:25:32 I think maybe I’m just afraid
02:25:36 that my life has been like a waste.
02:25:39 I’m so, I’m not evangelical about Emacs.
02:25:43 I think this.
02:25:44 This is how I feel about TensorFlow versus PyTorch.
02:25:47 Having just like, we’ve switched everything to PyTorch now.
02:25:50 Put months into the switch.
02:25:51 I have felt like I’ve wasted years on TensorFlow.
02:25:54 I can’t believe it.
02:25:56 I can’t believe how much better PyTorch is.
02:25:58 Yeah.
02:25:59 I’ve used Emacs and Vim, doesn’t matter.
02:26:01 Yeah, it’s still just my heart.
02:26:03 Somehow I fell in love with Lisp.
02:26:04 I don’t know why.
02:26:05 You can’t, the heart wants what the heart wants.
02:26:08 I don’t understand it, but it just connected with me.
02:26:10 Maybe it’s the functional language
02:26:11 that first I connected with.
02:26:13 Maybe it’s because so many of the AI courses
02:26:15 before the deep learning revolution
02:26:17 were taught with Lisp in mind.
02:26:19 I don’t know.
02:26:20 I don’t know what it is, but I’m stuck with it.
02:26:22 But at the same time, like,
02:26:23 why am I not using a modern ID
02:26:25 for some of these programming?
02:26:26 I don’t know.
02:26:27 They’re not that much better.
02:26:28 I’ve used modern IDs too.
02:26:30 But at the same time, so to just,
02:26:32 well, not to disagree with you,
02:26:33 but like, I like multiple monitors.
02:26:35 Like I have to do work on a laptop
02:26:38 and it’s a pain in the ass.
02:26:41 And also I’m addicted to the Kinesis weird keyboard.
02:26:45 You could see there.
02:26:46 Yeah, yeah, yeah.
02:26:48 Yeah, so you don’t have any of that.
02:26:50 You can just be on a MacBook.
02:26:51 I mean, look at work.
02:26:53 I have three 24 inch monitors.
02:26:55 I have a happy hacking keyboard.
02:26:56 I have a Razer Death Hatter mouse, like.
02:26:59 But it’s not essential for you.
02:27:01 No.
02:27:02 Let’s go to a day in the life of George Hots.
02:27:04 What is the perfect day productivity wise?
02:27:08 So we’re not talking about like Hunter S. Thompson drugs.
02:27:12 Yeah, yeah, yeah.
02:27:13 And let’s look at productivity.
02:27:16 Like what’s the day look like, like hour by hour?
02:27:19 Is there any regularities that create
02:27:23 a magical George Hots experience?
02:27:25 I can remember three days in my life.
02:27:28 And I remember these days vividly
02:27:30 when I’ve gone through kind of radical transformations
02:27:36 to the way I think.
02:27:37 And what I would give, I would pay $100,000
02:27:40 if I could have one of these days tomorrow.
02:27:42 The days have been so impactful.
02:27:44 And one was first discovering Eliezer Yudkowsky
02:27:47 on the singularity and reading that stuff.
02:27:50 And like, you know, my mind was blown.
02:27:54 The next was discovering the Hutter Prize
02:27:57 and that AI is just compression.
02:27:59 Like finally understanding AIXI and what all of that was.
02:28:03 You know, I like read about it when I was 18, 19,
02:28:05 I didn’t understand it.
02:28:06 And then the fact that like lossless compression
02:28:08 implies intelligence, the day that I was shown that.
02:28:12 And then the third one is controversial.
02:28:14 The day I found a blog called Unqualified Reservations.
02:28:17 And read that and I was like.
02:28:20 Wait, which one is that?
02:28:21 That’s, what’s the guy’s name?
02:28:22 Curtis Yarvin.
02:28:24 Yeah.
02:28:25 So many people tell me I’m supposed to talk to him.
02:28:27 Yeah, the day.
02:28:28 He looks, he sounds insane.
02:28:30 Definitely. Or brilliant,
02:28:31 but insane or both, I don’t know.
02:28:33 The day I found that blog was another like,
02:28:35 this was during like Gamergate
02:28:37 and kind of the run up to the 2016 election.
02:28:39 And I’m like, wow, okay, the world makes sense now.
02:28:42 This is like, I had a framework now to interpret this.
02:28:45 Just like I got the framework for AI
02:28:47 and a framework to interpret technological progress.
02:28:49 Like those days when I discovered these new frameworks were.
02:28:52 Oh, interesting.
02:28:53 So it’s not about, but what was special about those days?
02:28:57 How did those days come to be?
02:28:58 Is it just, you got lucky?
02:28:59 Like, you just encountered a hotter prize
02:29:04 on Hacker News or something like that?
02:29:09 But you see, I don’t think it’s just,
02:29:11 see, I don’t think it’s just that like,
02:29:13 I could have gotten lucky at any point.
02:29:14 I think that in a way.
02:29:16 You were ready at that moment.
02:29:17 Yeah, exactly.
02:29:18 To receive the information.
02:29:21 But is there some magic to the day today
02:29:24 of like eating breakfast?
02:29:27 And it’s the mundane things.
02:29:29 Nah.
02:29:29 Nothing.
02:29:30 Nah, I drift through life.
02:29:32 Without structure.
02:29:34 I drift through life hoping and praying
02:29:36 that I will get another day like those days.
02:29:38 And there’s nothing in particular you do
02:29:40 to be a receptacle for another, for day number four.
02:29:46 No, I didn’t do anything to get the other ones.
02:29:48 So I don’t think I have to really do anything now.
02:29:51 I took a month long trip to New York
02:29:53 and the Ethereum thing was the highlight of it,
02:29:56 but the rest of it was pretty terrible.
02:29:57 I did a two week road trip
02:29:59 and I got, I had to turn around.
02:30:01 I had to turn around driving in Gunnison, Colorado.
02:30:06 I passed through Gunnison
02:30:08 and the snow starts coming down.
02:30:10 There’s a pass up there called Monarch Pass
02:30:12 in order to get through to Denver,
02:30:13 you gotta get over the Rockies.
02:30:14 And I had to turn my car around.
02:30:16 I couldn’t, I watched a F150 go off the road.
02:30:20 I’m like, I gotta go back.
02:30:21 And like that day was meaningful.
02:30:24 Cause like, it was real.
02:30:26 Like I actually had to turn my car around.
02:30:28 It’s rare that anything even real happens in my life.
02:30:31 Even as, you know, mundane as the fact that,
02:30:34 yeah, there was snow, I had to turn around,
02:30:36 stay in Gunnison and leave the next day.
02:30:37 Something about that moment felt real.
02:30:40 Okay, so actually it’s interesting to break apart
02:30:43 the three moments you mentioned, if it’s okay.
02:30:45 So I always have trouble pronouncing his name,
02:30:48 but Alousa Yurkowski.
02:30:53 So what, how did your worldview change
02:30:57 in starting to consider the exponential growth of AI
02:31:02 and AGI that he thinks about
02:31:05 and the threats of artificial intelligence
02:31:07 and all that kind of ideas?
02:31:09 Can you, is it just like, can you maybe break apart
02:31:12 like what exactly was so magical to you?
02:31:15 Is it transformational experience?
02:31:17 Today, everyone knows him for threats and AI safety.
02:31:20 This was pre that stuff.
02:31:22 There was, I don’t think a mention of AI safety on the page.
02:31:25 This is, this is old Yurkowski stuff.
02:31:27 He’d probably denounce it all now.
02:31:29 He’d probably be like,
02:31:29 that’s exactly what I didn’t want to happen.
02:31:32 Sorry, man.
02:31:33 Is there something specific you can take from his work
02:31:37 that you can remember?
02:31:38 Yeah, it was this realization
02:31:40 that computers double in power every 18 months
02:31:45 and humans do not, and they haven’t crossed yet.
02:31:50 But if you have one thing that’s doubling every 18 months
02:31:52 and one thing that’s staying like this, you know,
02:31:55 here’s your log graph, here’s your line, you know,
02:31:58 calculate that.
02:31:59 And then the data opened the door
02:32:03 to the exponential thinking, like thinking that like,
02:32:06 you know what, with technology,
02:32:07 we can actually transform the world.
02:32:11 It opened the door to human obsolescence.
02:32:13 It opened the door to realize that in my lifetime,
02:32:16 humans are going to be replaced.
02:32:20 And then the matching idea to that of artificial intelligence
02:32:23 with the Hutter prize, you know, I’m torn.
02:32:27 I go back and forth on what I think about it.
02:32:30 Yeah.
02:32:31 But the basic thesis is it’s a nice compelling notion
02:32:36 that we can reduce the task of creating
02:32:38 an intelligent system, a generally intelligent system
02:32:41 into the task of compression.
02:32:43 So you can think of all of intelligence in the universe,
02:32:46 in fact, as a kind of compression.
02:32:50 Do you find that, was that just at the time
02:32:52 you found that as a compelling idea
02:32:53 or do you still find that a compelling idea?
02:32:56 I still find that a compelling idea.
02:32:59 I think that it’s not that useful day to day,
02:33:02 but actually one of maybe my quests before that
02:33:06 was a search for the definition of the word intelligence.
02:33:09 And I never had one.
02:33:10 And I definitely have a definition of the word compression.
02:33:14 It’s a very simple, straightforward one.
02:33:18 And you know what compression is,
02:33:19 you know what lossless, it’s lossless compression,
02:33:21 not lossy, lossless compression.
02:33:22 And that that is equivalent to intelligence,
02:33:25 which I believe, I’m not sure how useful
02:33:27 that definition is day to day,
02:33:28 but like I now have a framework to understand what it is.
02:33:32 And he just 10X the prize for that competition
02:33:36 like recently a few months ago.
02:33:37 You ever thought of taking a crack at that?
02:33:39 Oh, I did.
02:33:41 Oh, I did.
02:33:41 I spent the next, after I found the prize,
02:33:44 I spent the next six months of my life trying it.
02:33:47 And well, that’s when I started learning everything about AI.
02:33:51 And then I worked at Vicarious for a bit
02:33:53 and then I read all the deep learning stuff.
02:33:55 And I’m like, okay, now I like I’m caught up to modern AI.
02:33:58 Wow.
02:33:59 And I had a really good framework to put it all in
02:34:01 from the compression stuff, right?
02:34:04 Like some of the first deep learning models I played with
02:34:07 were GPT basically, but before transformers,
02:34:12 before it was still RNNs to do character prediction.
02:34:17 But by the way, on the compression side,
02:34:19 I mean, especially with neural networks,
02:34:22 what do you make of the lossless requirement
02:34:25 with the Hutter prize?
02:34:26 So, you know, human intelligence and neural networks
02:34:31 can probably compress stuff pretty well,
02:34:33 but it would be lossy.
02:34:35 It’s imperfect.
02:34:36 You can turn a lossy compression
02:34:37 to a lossless compressor pretty easily
02:34:39 using an arithmetic encoder, right?
02:34:41 You can take an arithmetic encoder
02:34:42 and you can just encode the noise with maximum efficiency.
02:34:45 Right?
02:34:46 So even if you can’t predict exactly
02:34:48 what the next character is,
02:34:50 the better a probability distribution,
02:34:52 you can put over the next character.
02:34:54 You can then use an arithmetic encoder to, right?
02:34:57 You don’t have to know whether it’s an E or an I,
02:34:59 you just have to put good probabilities on them
02:35:01 and then, you know, code those.
02:35:03 And if you have, it’s a bits of entropy thing, right?
02:35:06 So let me, on that topic,
02:35:07 it’d be interesting as a little side tour.
02:35:10 What are your thoughts in this year about GPT3
02:35:13 and these language models and these transformers?
02:35:16 Is there something interesting to you as an AI researcher,
02:35:20 or is there something interesting to you
02:35:22 as an autonomous vehicle developer?
02:35:24 Nah, I think it’s overhyped.
02:35:27 I mean, it’s not, like, it’s cool.
02:35:29 It’s cool for what it is, but no,
02:35:30 we’re not just gonna be able to scale up to GPT12
02:35:33 and get general purpose intelligence.
02:35:35 Like, your loss function is literally just,
02:35:38 you know, cross entropy loss on the character, right?
02:35:41 Like, that’s not the loss function of general intelligence.
02:35:44 Is that obvious to you?
02:35:45 Yes.
02:35:47 Can you imagine that, like,
02:35:51 to play devil’s advocate on yourself,
02:35:53 is it possible that you can,
02:35:55 the GPT12 will achieve general intelligence
02:35:58 with something as dumb as this kind of loss function?
02:36:01 I guess it depends what you mean by general intelligence.
02:36:05 So there’s another problem with the GPTs,
02:36:07 and that’s that they don’t have a,
02:36:11 they don’t have longterm memory.
02:36:13 Right.
02:36:13 So, like, just GPT12,
02:36:18 a scaled up version of GPT2 or GPT3,
02:36:22 I find it hard to believe.
02:36:26 Well, you can scale it in,
02:36:28 so it’s a hard coded length,
02:36:32 but you can make it wider and wider and wider.
02:36:34 Yeah.
02:36:36 You’re gonna get cool things from those systems,
02:36:40 but I don’t think you’re ever gonna get something
02:36:44 that can, like, you know, build me a rocket ship.
02:36:47 What about solved driving?
02:36:49 So, you know, you can use Transformer with video,
02:36:53 for example.
02:36:54 You think, is there something in there?
02:36:57 No, because, I mean, look, we use a GRU.
02:37:01 We use a GRU.
02:37:02 We could change that GRU out to a Transformer.
02:37:05 I think driving is much more Markovian than language.
02:37:09 So, Markovian, you mean, like, the memory,
02:37:11 which aspect of Markovian?
02:37:13 I mean that, like, most of the information
02:37:16 in the state at T minus one is also in state T.
02:37:19 I see, yeah.
02:37:20 Right, and it kind of, like, drops off nicely like this,
02:37:22 whereas sometime with language,
02:37:23 you have to refer back to the third paragraph
02:37:25 on the second page.
02:37:27 I feel like.
02:37:28 There’s not many, like, you can say, like,
02:37:30 speed limit signs, but there’s really not many things
02:37:32 in autonomous driving that look like that.
02:37:33 But if you look at, to play devil’s advocate,
02:37:37 is the risk estimation thing that you’ve talked about
02:37:39 is kind of interesting.
02:37:41 Is, it feels like there might be some longer term
02:37:45 aggregation of context necessary to be able to figure out,
02:37:49 like, the context.
02:37:51 Yeah, I’m not even sure I’m believing my devil’s advocate.
02:37:55 We have a nice, like, vision model,
02:37:58 which outputs, like, a one or two,
02:38:00 four dimensional perception space.
02:38:03 Can I try Transformers on it?
02:38:04 Sure, I probably will.
02:38:06 At some point, we’ll try Transformers,
02:38:08 and then we’ll just see.
02:38:09 Do they do better?
02:38:09 Sure, I’m.
02:38:10 But it might not be a game changer, you’re saying?
02:38:12 No, well, I’m not.
02:38:13 Like, might Transformers work better than GRUs
02:38:15 for autonomous driving?
02:38:16 Sure.
02:38:16 Might we switch?
02:38:17 Sure.
02:38:18 Is this some radical change?
02:38:19 No.
02:38:20 Okay, we use a slightly different,
02:38:21 you know, we switch from RNNs to GRUs.
02:38:23 Like, okay, maybe it’s GRUs to Transformers,
02:38:24 but no, it’s not.
02:38:26 Yeah.
02:38:27 Well, on the topic of general intelligence,
02:38:30 I don’t know how much I’ve talked to you about it.
02:38:32 Like, what, do you think we’ll actually build
02:38:36 an AGI?
02:38:38 Like, if you look at Ray Kurzweil with Singularity,
02:38:40 do you have like an intuition about,
02:38:43 you’re kind of saying driving is easy.
02:38:45 Yeah.
02:38:46 And I tend to personally believe that solving driving
02:38:52 will have really deep, important impacts
02:38:56 on our ability to solve general intelligence.
02:38:59 Like, I think driving doesn’t require general intelligence,
02:39:03 but I think they’re going to be neighbors
02:39:05 in a way that it’s like deeply tied.
02:39:08 Cause it’s so, like driving is so deeply connected
02:39:11 to the human experience that I think solving one
02:39:15 will help solve the other.
02:39:17 But, so I don’t see, I don’t see driving as like easy
02:39:20 and almost like separate than general intelligence,
02:39:23 but like, what’s your vision of a future with a Singularity?
02:39:26 Do you see there’ll be a single moment,
02:39:28 like a Singularity where it’ll be a phase shift?
02:39:30 Are we in the Singularity now?
02:39:32 Like what, do you have crazy ideas about the future
02:39:34 in terms of AGI?
02:39:35 We’re definitely in the Singularity now.
02:39:38 We are?
02:39:38 Of course, of course.
02:39:40 Look at the bandwidth between people.
02:39:41 The bandwidth between people goes up, right?
02:39:44 The Singularity is just, you know, when the bandwidth, but.
02:39:47 What do you mean by the bandwidth of people?
02:39:48 Communications, tools, the whole world is networked.
02:39:51 The whole world is networked
02:39:52 and we raise the speed of that network, right?
02:39:54 Oh, so you think the communication of information
02:39:57 in a distributed way is an empowering thing
02:40:00 for collective intelligence?
02:40:02 Oh, I didn’t say it’s necessarily a good thing,
02:40:03 but I think that’s like,
02:40:04 when I think of the definition of the Singularity,
02:40:06 yeah, it seems kind of right.
02:40:08 I see, like it’s a change in the world
02:40:12 beyond which like the world be transformed
02:40:14 in ways that we can’t possibly imagine.
02:40:16 No, I mean, I think we’re in the Singularity now
02:40:18 in the sense that there’s like, you know,
02:40:19 one world and a monoculture and it’s also linked.
02:40:22 Yeah, I mean, I kind of share the intuition
02:40:24 that the Singularity will originate
02:40:27 from the collective intelligence of us ands
02:40:31 versus the like some single system AGI type thing.
02:40:35 Oh, I totally agree with that.
02:40:37 Yeah, I don’t really believe in like a hard take off AGI
02:40:40 kind of thing.
02:40:45 Yeah, I don’t even think AI is all that different in kind
02:40:49 from what we’ve already been building.
02:40:52 With respect to driving,
02:40:53 I think driving is a subset of general intelligence
02:40:56 and I think it’s a pretty complete subset.
02:40:58 I think the tools we develop at Kama
02:41:00 will also be extremely helpful
02:41:02 to solving general intelligence
02:41:04 and that’s I think the real reason why I’m doing it.
02:41:06 I don’t care about self driving cars.
02:41:08 It’s a cool problem to beat people at.
02:41:10 But yeah, I mean, yeah, you’re kind of, you’re of two minds.
02:41:14 So one, you do have to have a mission
02:41:16 and you wanna focus and make sure you get there.
02:41:19 You can’t forget that but at the same time,
02:41:22 there is a thread that’s much bigger
02:41:26 than that connects the entirety of your effort.
02:41:28 That’s much bigger than just driving.
02:41:31 With AI and with general intelligence,
02:41:33 it is so easy to delude yourself
02:41:35 into thinking you’ve figured something out when you haven’t.
02:41:37 If we build a level five self driving car,
02:41:39 we have indisputably built something.
02:41:42 Yeah.
02:41:43 Is it general intelligence?
02:41:44 I’m not gonna debate that.
02:41:45 I will say we’ve built something
02:41:47 that provides huge financial value.
02:41:49 Yeah, beautifully put.
02:41:50 That’s the engineering credo.
02:41:51 Like just build the thing.
02:41:53 It’s like, that’s why I’m with Elon
02:41:57 on go to Mars.
02:41:58 Yeah, that’s a great one.
02:41:59 You can argue like who the hell cares about going to Mars.
02:42:03 But the reality is set that as a mission, get it done.
02:42:07 Yeah.
02:42:08 And then you’re going to crack some problem
02:42:09 that you’ve never even expected
02:42:11 in the process of doing that, yeah.
02:42:13 Yeah, I mean, no, I think if I had a choice
02:42:16 between humanity going to Mars
02:42:17 and solving self driving cars,
02:42:18 I think going to Mars is better, but I don’t know.
02:42:21 I’m more suited for self driving cars.
02:42:23 I’m an information guy.
02:42:24 I’m not a modernist, I’m a postmodernist.
02:42:26 Postmodernist, all right, beautifully put.
02:42:29 Let me drag you back to programming for a sec.
02:42:32 What three, maybe three to five programming languages
02:42:35 should people learn, do you think?
02:42:36 Like if you look at yourself,
02:42:38 what did you get the most out of from learning?
02:42:42 Well, so everybody should learn C and assembly.
02:42:45 We’ll start with those two, right?
02:42:47 Assembly?
02:42:48 Yeah, if you can’t code an assembly,
02:42:49 you don’t know what the computer’s doing.
02:42:51 You don’t understand like,
02:42:53 you don’t have to be great in assembly,
02:42:54 but you have to code in it.
02:42:56 And then like, you have to appreciate assembly
02:42:58 in order to appreciate all the great things C gets you.
02:43:02 And then you have to code in C
02:43:03 in order to appreciate all the great things Python gets you.
02:43:06 So I’ll just say assembly C and Python,
02:43:07 we’ll start with those three.
02:43:09 The memory allocation of C and the fact that,
02:43:14 so assembly gives you a sense
02:43:16 of just how many levels of abstraction
02:43:18 you get to work on in modern day programming.
02:43:20 Yeah, yeah, yeah, yeah, graph coloring for assignment,
02:43:22 register assignment and compilers.
02:43:24 Like, you know, you gotta do,
02:43:25 you know, the compiler,
02:43:26 the computer only has a certain number of registers,
02:43:28 yet you can have all the variables you want in a C function.
02:43:31 So you get to start to build intuition about compilation,
02:43:34 like what a compiler gets you.
02:43:37 What else?
02:43:38 Well, then there’s kind of a,
02:43:41 so those are all very imperative programming languages.
02:43:45 Then there’s two other paradigms for programming
02:43:47 that everybody should be familiar with.
02:43:49 And one of them is functional.
02:43:51 You should learn Haskell and take that all the way through,
02:43:54 learn a language with dependent types like Coq,
02:43:57 learn that whole space,
02:43:58 like the very PL theory, heavy languages.
02:44:02 And Haskell is your favorite functional?
02:44:04 Is that the go to, you’d say?
02:44:06 Yeah, I’m not a great Haskell programmer.
02:44:08 I wrote a compiler in Haskell once.
02:44:10 There’s another paradigm,
02:44:11 and actually there’s one more paradigm
02:44:12 that I’ll even talk about after that,
02:44:14 that I never used to talk about
02:44:15 when I would think about this,
02:44:15 but the next paradigm is learn Verilog of HDL.
02:44:20 Understand this idea of all of the instructions
02:44:22 execute at once. If I have a block in Verilog
02:44:26 and I write stuff in it, it’s not sequential.
02:44:29 They all execute at once.
02:44:33 And then think like that, that’s how hardware works.
02:44:36 To be, so I guess assembly doesn’t quite get you that.
02:44:40 Assembly is more about compilation,
02:44:42 and Verilog is more about the hardware,
02:44:44 like giving a sense of what actually
02:44:46 is the hardware is doing.
02:44:48 Assembly, C, Python are straight,
02:44:50 like they sit right on top of each other.
02:44:52 In fact, C is, well, C is kind of coded in C,
02:44:55 but you could imagine the first C was coded in assembly,
02:44:57 and Python is actually coded in C.
02:45:00 So you can straight up go on that.
02:45:03 Got it, and then Verilog gives you, that’s brilliant.
02:45:06 Okay.
02:45:07 And then I think there’s another one now.
02:45:09 Everyone, Carpathia calls it programming 2.0,
02:45:12 which is learn a, I’m not even gonna,
02:45:16 don’t learn TensorFlow, learn PyTorch.
02:45:18 So machine learning.
02:45:20 We’ve got to come up with a better term
02:45:21 than programming 2.0, or, but yeah.
02:45:26 It’s a programming language, learn it.
02:45:29 I wonder if it can be formalized a little bit better.
02:45:32 It feels like we’re in the early days
02:45:34 of what that actually entails.
02:45:37 Data driven programming?
02:45:39 Data driven programming, yeah.
02:45:41 But it’s so fundamentally different
02:45:43 as a paradigm than the others.
02:45:44 Like it almost requires a different skillset.
02:45:50 But you think it’s still, yeah.
02:45:53 And PyTorch versus TensorFlow, PyTorch wins.
02:45:56 It’s the fourth paradigm.
02:45:57 It’s the fourth paradigm that I’ve kind of seen.
02:45:59 There’s like this, you know,
02:46:01 imperative functional hardware.
02:46:04 I don’t know a better word for it.
02:46:06 And then ML.
02:46:08 Do you have advice for people that wanna,
02:46:13 you know, get into programming, wanna learn programming?
02:46:16 You have a video,
02:46:19 what is programming noob lessons, exclamation point.
02:46:22 And I think the top comment is like,
02:46:24 warning, this is not for noobs.
02:46:27 Do you have a noob, like a TLDW for that video,
02:46:32 but also a noob friendly advice
02:46:38 on how to get into programming?
02:46:39 We’re never going to learn programming
02:46:41 by watching a video called Learn Programming.
02:46:44 The only way to learn programming, I think,
02:46:46 and the only one is that the only way
02:46:48 everyone I’ve ever met who can program well,
02:46:50 learned it all in the same way.
02:46:51 They had something they wanted to do
02:46:54 and then they tried to do it.
02:46:56 And then they were like, oh, well, okay.
02:47:00 This is kind of, you know, it’d be nice
02:47:01 if the computer could kind of do this.
02:47:02 And then, you know, that’s how you learn.
02:47:04 You just keep pushing on a project.
02:47:09 So the only advice I have for learning programming
02:47:10 is go program.
02:47:12 Somebody wrote to me a question like,
02:47:14 we don’t really, they’re looking to learn
02:47:17 about recurring neural networks.
02:47:19 And he’s saying, like, my company’s thinking
02:47:20 of using recurring neural networks for time series data,
02:47:24 but we don’t really have an idea of where to use it yet.
02:47:27 We just want to, like, do you have any advice
02:47:28 on how to learn about, these are these kind of
02:47:31 general machine learning questions.
02:47:33 And I think the answer is, like,
02:47:36 actually have a problem that you’re trying to solve.
02:47:39 And just.
02:47:40 I see that stuff.
02:47:41 Oh my God, when people talk like that,
02:47:42 they’re like, I heard machine learning is important.
02:47:45 Could you help us integrate machine learning
02:47:47 with macaroni and cheese production?
02:47:51 You just, I don’t even, you can’t help these people.
02:47:54 Like, who lets you run anything?
02:47:55 Who lets that kind of person run anything?
02:47:58 I think we’re all, we’re all beginners at some point.
02:48:02 So.
02:48:03 It’s not like they’re a beginner.
02:48:04 It’s like, my problem is not that they don’t know
02:48:07 about machine learning.
02:48:08 My problem is that they think that machine learning
02:48:10 has something to say about macaroni and cheese production.
02:48:14 Or like, I heard about this new technology.
02:48:17 How can I use it for why?
02:48:19 Like, I don’t know what it is, but how can I use it for why?
02:48:23 That’s true.
02:48:24 You have to build up an intuition of how,
02:48:26 cause you might be able to figure out a way,
02:48:27 but like the prerequisites,
02:48:29 you should have a macaroni and cheese problem to solve first.
02:48:32 Exactly.
02:48:33 And then two, you should have more traditional,
02:48:36 like the learning process should involve
02:48:39 more traditionally applicable problems
02:48:41 in the space of whatever that is, machine learning,
02:48:44 and then see if it can be applied to mac and cheese.
02:48:47 At least start with, tell me about a problem.
02:48:49 Like if you have a problem, you’re like,
02:48:50 you know, some of my boxes aren’t getting
02:48:52 enough macaroni in them.
02:48:54 Can we use machine learning to solve this problem?
02:48:56 That’s much, much better than how do I apply
02:48:59 machine learning to macaroni and cheese?
02:49:01 One big thing, maybe this is me talking
02:49:05 to the audience a little bit, cause I get these days
02:49:07 so many messages, advice on how to like learn stuff, okay?
02:49:15 My, this is not me being mean.
02:49:18 I think this is quite profound actually,
02:49:20 is you should Google it.
02:49:22 Oh yeah.
02:49:23 Like one of the like skills that you should really acquire
02:49:29 as an engineer, as a researcher, as a thinker,
02:49:33 like one, there’s two complementary skills.
02:49:36 Like one is with a blank sheet of paper
02:49:39 with no internet to think deeply.
02:49:41 And then the other is to Google the crap
02:49:44 out of the questions you have.
02:49:45 Like that’s actually a skill people often talk about,
02:49:49 but like doing research, like pulling at the thread,
02:49:52 like looking up different words,
02:49:53 going into like GitHub repositories with two stars
02:49:58 and like looking how they did stuff,
02:49:59 like looking at the code or going on Twitter,
02:50:03 seeing like there’s little pockets of brilliant people
02:50:05 that are like having discussions.
02:50:07 Like if you’re a neuroscientist,
02:50:09 go into signal processing community.
02:50:11 If you’re an AI person going into the psychology community,
02:50:15 like switch communities.
02:50:18 I keep searching, searching, searching,
02:50:19 because it’s so much better to invest
02:50:23 in like finding somebody else who already solved your problem
02:50:27 than it is to try to solve the problem.
02:50:30 And because they’ve often invested years of their life,
02:50:34 like entire communities are probably already out there
02:50:37 who have tried to solve your problem.
02:50:39 I think they’re the same thing.
02:50:40 I think you go try to solve the problem.
02:50:44 And then in trying to solve the problem,
02:50:46 if you’re good at solving problems,
02:50:47 you’ll stumble upon the person who solved it already.
02:50:50 But the stumbling is really important.
02:50:52 I think that’s a skill that people should really put,
02:50:54 especially in undergrad, like search.
02:50:57 If you ask me a question,
02:50:58 how should I get started in deep learning, like especially?
02:51:04 Like that is just so Googleable.
02:51:07 Like the whole point is you Google that
02:51:10 and you get a million pages and just start looking at them.
02:51:13 Start pulling at the threads, start exploring,
02:51:16 start taking notes, start getting advice
02:51:19 from a million people that already like spent their life
02:51:22 answering that question, actually.
02:51:25 Oh, well, yeah, I mean, that’s definitely also, yeah,
02:51:26 when people like ask me things like that, I’m like, trust me,
02:51:28 the top answer on Google is much, much better
02:51:30 than anything I’m going to tell you, right?
02:51:32 Yeah.
02:51:34 People ask, it’s an interesting question.
02:51:38 Let me know if you have any recommendations.
02:51:39 What three books, technical or fiction or philosophical,
02:51:43 had an impact on your life or you would recommend perhaps?
02:51:49 Maybe we’ll start with the least controversial,
02:51:51 Infinite Jest, Infinite Jest is a…
02:51:57 David Foster Wallace.
02:51:58 Yeah, it’s a book about wireheading, really.
02:52:03 Very enjoyable to read, very well written.
02:52:07 You know, you will grow as a person reading this book,
02:52:11 its effort, and I’ll set that up for the second book,
02:52:14 which is pornography, it’s called Atlas Shrugged,
02:52:17 which…
02:52:21 Atlas Shrugged is pornography.
02:52:22 I mean, it is, I will not defend the,
02:52:25 I will not say Atlas Shrugged is a well written book.
02:52:28 It is entertaining to read, certainly, just like pornography.
02:52:31 The production value isn’t great.
02:52:33 You know, there’s a 60 page monologue in there
02:52:36 that Ann Rand’s editor really wanted to take out.
02:52:38 And she paid, she paid out of her pocket
02:52:42 to keep that 60 page monologue in the book.
02:52:45 But it is a great book for a kind of framework
02:52:53 of human relations.
02:52:54 And I know a lot of people are like,
02:52:55 yeah, but it’s a terrible framework.
02:52:58 Yeah, but it’s a framework.
02:53:00 Just for context, in a couple of days,
02:53:02 I’m speaking for probably four plus hours
02:53:06 with Yaron Brook, who’s the main living,
02:53:10 remaining objectivist, objectivist.
02:53:13 Interesting.
02:53:14 So I’ve always found this philosophy quite interesting
02:53:19 on many levels.
02:53:20 One of how repulsive some percent of,
02:53:24 large percent of the population find it,
02:53:26 which is always, always funny to me
02:53:29 when people are like unable to even read a philosophy
02:53:32 because of some, I think that says more
02:53:36 about their psychological perspective on it.
02:53:40 But there is something about objectivism
02:53:45 and Ann Rand’s philosophy that’s deeply connected
02:53:48 to this idea of capitalism,
02:53:50 of the ethical life is the productive life
02:53:56 that was always compelling to me.
02:54:00 It didn’t seem as, like I didn’t seem to interpret it
02:54:03 in the negative sense that some people do.
02:54:05 To be fair, I read that book when I was 19.
02:54:07 So you had an impact at that point, yeah.
02:54:09 Yeah, and the bad guys in the book have this slogan
02:54:13 from each according to their ability
02:54:15 to each according to their need.
02:54:17 And I’m looking at this and I’m like,
02:54:19 these are the most cart,
02:54:20 this is team rocket level cartoonishness, right?
02:54:22 No bad guy.
02:54:23 And then when I realized that was actually the slogan
02:54:25 of the communist party, I’m like, wait a second.
02:54:29 Wait, no, no, no, no, no.
02:54:31 You’re telling me this really happened?
02:54:34 Yeah, it’s interesting.
02:54:34 I mean, one of the criticisms of her work
02:54:36 is she has a cartoonish view of good and evil.
02:54:39 Like the reality, as Jordan Peterson says,
02:54:44 is that each of us have the capacity for good and evil
02:54:47 in us as opposed to like, there’s some characters
02:54:49 who are purely evil and some characters that are purely good.
02:54:52 And that’s in a way why it’s pornographic.
02:54:55 The production value, I love it.
02:54:57 Like evil is punished and there’s very clearly,
02:55:01 there’s no, just like porn doesn’t have character growth.
02:55:06 Well, you know, neither does Alice Shrugged, like.
02:55:09 Really, well put.
02:55:10 But at 19 year old George Hots, it was good enough.
02:55:14 Yeah, yeah, yeah, yeah.
02:55:15 What’s the third?
02:55:16 You have something?
02:55:18 I could give, these two I’ll just throw out.
02:55:21 They’re sci fi.
02:55:22 Perputation City.
02:55:24 Great thing to start thinking about copies of yourself.
02:55:26 And then the…
02:55:27 Who’s that by?
02:55:28 Sorry, I didn’t catch that.
02:55:29 That is Greg Egan.
02:55:31 He’s a, that might not be his real name.
02:55:33 Some Australian guy, might not be Australian.
02:55:35 I don’t know.
02:55:36 And then this one’s online.
02:55:38 It’s called The Metamorphosis of Prime Intellect.
02:55:43 It’s a story set in a post singularity world.
02:55:45 It’s interesting.
02:55:46 Is there, can you, either of the worlds,
02:55:49 do you find something philosophical interesting in them
02:55:51 that you can comment on?
02:55:53 I mean, it is clear to me that
02:55:57 Metamorphosis of Prime Intellect is like written by
02:56:00 an engineer, which is,
02:56:03 it’s very almost a pragmatic take on a utopia, in a way.
02:56:12 Positive or negative?
02:56:15 That’s up to you to decide reading the book.
02:56:17 And the ending of it is very interesting as well.
02:56:21 And I didn’t realize what it was.
02:56:23 I first read that when I was 15.
02:56:25 I’ve reread that book several times in my life.
02:56:27 And it’s short, it’s 50 pages.
02:56:29 Everyone should go read it.
02:56:30 What’s, sorry, it’s a little tangent.
02:56:33 I’ve been working through the foundation.
02:56:34 I’ve been, I haven’t read much sci fi my whole life
02:56:37 and I’m trying to fix that the last few months.
02:56:40 That’s been a little side project.
02:56:42 What’s to you as the greatest sci fi novel
02:56:46 that people should read?
02:56:47 Or is that?
02:56:49 I mean, I would, yeah, I would say like, yeah,
02:56:51 Permutation City, Metamorphosis of Prime Intellect.
02:56:53 I don’t know.
02:56:54 I didn’t like Foundation.
02:56:56 I thought it was way too modernist.
02:56:58 You like Dune and all of those.
02:57:00 I’ve never read Dune.
02:57:01 I’ve never read Dune.
02:57:02 I have to read it.
02:57:04 Fire Upon the Deep is interesting.
02:57:09 Okay, I mean, look, everyone should read,
02:57:10 everyone should read Neuromancer.
02:57:11 Everyone should read Snow Crash.
02:57:12 If you haven’t read those, like start there.
02:57:15 Yeah, I haven’t read Snow Crash.
02:57:16 You haven’t read Snow Crash?
02:57:17 Oh, it’s, I mean, it’s very entertaining.
02:57:19 Go to Lesher Bach.
02:57:20 And if you want the controversial one,
02:57:22 Bronze Age Mindset.
02:57:25 All right, I’ll look into that one.
02:57:27 Those aren’t sci fi, but just to round out books.
02:57:30 So a bunch of people asked me on Twitter
02:57:34 and Reddit and so on for advice.
02:57:36 So what advice would you give a young person today
02:57:39 about life?
02:57:40 In other words, what, yeah, I mean, looking back,
02:57:47 especially when you were younger, you did,
02:57:50 and you continued it.
02:57:51 You’ve accomplished a lot of interesting things.
02:57:54 Is there some advice from those,
02:57:57 from that life of yours that you can pass on?
02:58:01 If college ever opens again,
02:58:03 I would love to give a graduation speech.
02:58:07 At that point, I will put a lot of somewhat satirical effort
02:58:11 into this question.
02:58:12 Yeah, at this, you haven’t written anything at this point.
02:58:15 Oh, you know what?
02:58:16 Always wear sunscreen.
02:58:18 This is water.
02:58:19 Pick your plagiarizing.
02:58:21 I mean, you know, but that’s the,
02:58:23 that’s the like clean your room.
02:58:26 You know, yeah, you can plagiarize from all of this stuff.
02:58:28 And it’s, there is no,
02:58:35 self help books aren’t designed to help you.
02:58:37 They’re designed to make you feel good.
02:58:40 Like whatever advice I could give, you already know.
02:58:44 Everyone already knows.
02:58:45 Sorry, it doesn’t feel good.
02:58:50 Right?
02:58:51 Like, you know, you know,
02:58:53 if I tell you that you should, you know,
02:58:56 eat well and read more and it’s not gonna do anything.
02:59:01 I think the whole like genre
02:59:03 of those kinds of questions is meaningless.
02:59:07 I don’t know.
02:59:08 If anything, it’s don’t worry so much about that stuff.
02:59:10 Don’t be so caught up in your head.
02:59:12 Right.
02:59:13 I mean, you’re, yeah.
02:59:14 In a sense that your whole life,
02:59:16 your whole existence is like moving version of that advice.
02:59:20 I don’t know.
02:59:23 There’s something, I mean,
02:59:25 there’s something in you that resists
02:59:27 that kind of thinking and that in itself is,
02:59:30 it’s just illustrative of who you are.
02:59:34 And there’s something to learn from that.
02:59:36 I think you’re clearly not overthinking stuff.
02:59:41 Yeah.
02:59:42 And you know what?
02:59:42 There’s a gut thing.
02:59:43 Even when I talk about my advice,
02:59:45 I’m like, my advice is only relevant to me.
02:59:47 It’s not relevant to anybody else.
02:59:48 I’m not saying you should go out.
02:59:49 If you’re the kind of person who overthinks things
02:59:51 to stop overthinking things, it’s not bad.
02:59:54 It doesn’t work for me.
02:59:54 Maybe it works for you.
02:59:55 I don’t know.
02:59:57 Let me ask you about love.
02:59:59 Yeah.
03:00:02 I think last time we talked about the meaning of life
03:00:05 and it was kind of about winning.
03:00:08 Of course.
03:00:10 I don’t think I’ve talked to you about love much,
03:00:13 whether romantic or just love
03:00:15 for the common humanity amongst us all.
03:00:18 What role has love played in your life?
03:00:21 In this quest for winning, where does love fit in?
03:00:26 Well, the word love, I think means several different things.
03:00:29 There’s love in the sense of, maybe I could just say,
03:00:32 there’s like love in the sense of opiates
03:00:34 and love in the sense of oxytocin
03:00:37 and then love in the sense of,
03:00:43 maybe like a love for math.
03:00:44 I don’t think it fits into either
03:00:45 of those first two paradigms.
03:00:49 So each of those, have they given something to you
03:00:55 in your life?
03:00:56 I’m not that big of a fan of the first two.
03:01:00 Why?
03:01:03 The same reason I’m not a fan of,
03:01:06 the same reason I don’t do opiates and don’t take ecstasy.
03:01:09 And there were times, look, I’ve tried both.
03:01:14 I liked opiates way more than I liked ecstasy,
03:01:18 but they’re not, the ethical life is the productive life.
03:01:24 So maybe that’s my problem with those.
03:01:27 And then like, yeah, a sense of, I don’t know,
03:01:29 like abstract love for humanity.
03:01:32 I mean, the abstract love for humanity,
03:01:34 I’m like, yeah, I’ve always felt that.
03:01:36 And I guess it’s hard for me to imagine
03:01:39 not feeling it and maybe there’s people who don’t.
03:01:41 And I don’t know.
03:01:43 Yeah, that’s just like a background thing that’s there.
03:01:46 I mean, since we brought up drugs, let me ask you,
03:01:51 this is becoming more and more a part of my life
03:01:54 because I’m talking to a few researchers
03:01:55 that are working on psychedelics.
03:01:57 I’ve eaten shrooms a couple of times
03:02:00 and it was fascinating to me that like the mind can go,
03:02:04 like just fascinating the mind can go to places
03:02:08 I didn’t imagine it could go.
03:02:09 And it was very friendly and positive and exciting
03:02:12 and everything was kind of hilarious in the place.
03:02:16 Wherever my mind went, that’s where I went.
03:02:18 Is, what do you think about psychedelics?
03:02:20 Do you think they have, where do you think the mind goes?
03:02:24 Have you done psychedelics?
03:02:25 Where do you think the mind goes?
03:02:28 Is there something useful to learn about the places it goes
03:02:32 once you come back?
03:02:33 I find it interesting that this idea
03:02:38 that psychedelics have something to teach
03:02:40 is almost unique to psychedelics, right?
03:02:43 People don’t argue this about amphetamines.
03:02:46 And I’m not really sure why.
03:02:50 I think all of the drugs have lessons to teach.
03:02:53 I think there’s things to learn from opiates.
03:02:55 I think there’s things to learn from amphetamines.
03:02:56 I think there’s things to learn from psychedelics,
03:02:58 things to learn from marijuana.
03:03:02 But also at the same time recognize
03:03:05 that I don’t think you’re learning things about the world.
03:03:07 I think you’re learning things about yourself.
03:03:09 Yes.
03:03:10 And, you know, what’s the, even, it might’ve even been,
03:03:15 might’ve even been a Timothy Leary quote.
03:03:17 I don’t wanna misquote him,
03:03:18 but the idea is basically like, you know,
03:03:20 everybody should look behind the door,
03:03:21 but then once you’ve seen behind the door,
03:03:22 you don’t need to keep going back.
03:03:26 So, I mean, and that’s my thoughts on all real drug use too.
03:03:29 Except maybe for caffeine.
03:03:32 It’s a little experience that is good to have, but.
03:03:37 Oh yeah, no, I mean, yeah, I guess,
03:03:39 yes, psychedelics are definitely.
03:03:41 So you’re a fan of new experiences, I suppose.
03:03:43 Yes.
03:03:44 Because they all contain a little,
03:03:45 especially the first few times,
03:03:47 it contains some lessons that can be picked up.
03:03:49 Yeah, and I’ll revisit psychedelics maybe once a year.
03:03:55 Usually smaller doses.
03:03:58 Maybe they turn up the learning rate of your brain.
03:04:01 I’ve heard that, I like that.
03:04:03 Yeah, that’s cool.
03:04:04 Big learning rates have pros and cons.
03:04:07 Last question, and this is a little weird one,
03:04:09 but you’ve called yourself crazy in the past.
03:04:14 First of all, on a scale of one to 10,
03:04:16 how crazy would you say are you?
03:04:18 Oh, I mean, it depends how you, you know,
03:04:19 when you compare me to Elon Musk and Anthony Levandowski,
03:04:21 not so crazy.
03:04:23 So like a seven?
03:04:25 Let’s go with six.
03:04:27 Six, six, six.
03:04:29 What?
03:04:31 Well, I like seven, seven’s a good number.
03:04:32 Seven, all right, well, I’m sure day by day it changes,
03:04:36 right, so, but you’re in that area.
03:04:42 In thinking about that,
03:04:43 what do you think is the role of madness?
03:04:45 Is that a feature or a bug
03:04:48 if you were to dissect your brain?
03:04:51 So, okay, from like a mental health lens on crazy,
03:04:57 I’m not sure I really believe in that.
03:04:59 I’m not sure I really believe in like a lot of that stuff.
03:05:02 Right, this concept of, okay, you know,
03:05:05 when you get over to like hardcore bipolar and schizophrenia,
03:05:09 these things are clearly real, somewhat biological.
03:05:13 And then over here on the spectrum,
03:05:14 you have like ADD and oppositional defiance disorder
03:05:18 and these things that are like,
03:05:20 wait, this is normal spectrum human behavior.
03:05:22 Like this isn’t, you know, where’s the line here
03:05:28 and why is this like a problem?
03:05:31 So there’s this whole, you know,
03:05:33 the neurodiversity of humanity is huge.
03:05:35 Like people think I’m always on drugs.
03:05:37 People are saying this to me on my streams.
03:05:38 And I’m like, guys, you know,
03:05:39 like I’m real open with my drug use.
03:05:41 I’d tell you if I was on drugs and yeah,
03:05:44 I had like a cup of coffee this morning,
03:05:45 but other than that, this is just me.
03:05:47 You’re witnessing my brain in action.
03:05:51 So the word madness doesn’t even make sense
03:05:55 in the rich neurodiversity of humans.
03:05:59 I think it makes sense, but only for like
03:06:04 some insane extremes.
03:06:07 Like if you are actually like visibly hallucinating,
03:06:11 you know, that’s okay.
03:06:15 But there is the kind of spectrum on which you stand out.
03:06:17 Like that’s like, if I were to look, you know,
03:06:22 at decorations on a Christmas tree or something like that,
03:06:25 like if you were a decoration, that would catch my eye.
03:06:28 Like that thing is sparkly, whatever the hell that thing is.
03:06:35 There’s something to that.
03:06:37 Just like refusing to be boring
03:06:42 or maybe boring is the wrong word,
03:06:43 but to yeah, I mean, be willing to sparkle, you know?
03:06:52 It’s like somewhat constructed.
03:06:54 I mean, I am who I choose to be.
03:06:57 I’m gonna say things as true as I can see them.
03:07:01 I’m not gonna lie.
03:07:04 But that’s a really important feature in itself.
03:07:06 So like whatever the neurodiversity of your,
03:07:09 whatever your brain is, not putting constraints on it
03:07:13 that force it to fit into the mold of what society is like,
03:07:18 defines what you’re supposed to be.
03:07:20 So you’re one of the specimens
03:07:22 that doesn’t mind being yourself.
03:07:27 Being right is super important,
03:07:31 except at the expense of being wrong.
03:07:37 Without breaking that apart,
03:07:38 I think it’s a beautiful way to end it.
03:07:40 George, you’re one of the most special humans I know.
03:07:43 It’s truly an honor to talk to you.
03:07:44 Thanks so much for doing it.
03:07:45 Thank you for having me.
03:07:47 Thanks for listening to this conversation with George Hotz
03:07:50 and thank you to our sponsors,
03:07:52 Four Sigmatic, which is the maker
03:07:54 of delicious mushroom coffee,
03:07:57 Decoding Digital, which is a tech podcast
03:07:59 that I listen to and enjoy,
03:08:02 and ExpressVPN, which is the VPN I’ve used for many years.
03:08:07 Please check out these sponsors in the description
03:08:09 to get a discount and to support this podcast.
03:08:13 If you enjoy this thing, subscribe on YouTube,
03:08:15 review it with Five Stars and Apple Podcast,
03:08:17 follow on Spotify, support on Patreon,
03:08:20 or connect with me on Twitter at Lex Friedman.
03:08:24 And now, let me leave you with some words
03:08:27 from the great and powerful Linus Torvalds.
03:08:30 Talk is cheap, show me the code.
03:08:33 Thank you for listening and hope to see you next time.