Knowledge Fight Interactive Search Tool

All Episodes

Previous Next

Aug. 5, 2019 - Lex Fridman Podcast

01:59:15

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

Large

Audio Only |

Scroll

Time	Text
	The following is a conversation with George Hotz.
	He's the founder of Kama AI, a machine learning-based vehicle automation company.
	He is most certainly an outspoken personality in the field of AI and technology in general.
	He first gained recognition for being the first person to carry or unlock an iPhone.
	And since then, he has done quite a few interesting things at the intersection of hardware and software.
	This is the Artificial Intelligence Podcast.
	If you enjoy it, subscribe on YouTube, give it five stars on iTunes, support it on Patreon, or simply connect with me on Twitter at Lex Friedman, spelled F-R-I-D-M-A-N. And I'd like to give a special thank you to Jennifer from Canada for her support of the podcast on Patreon.
	Merci beaucoup, Jennifer. She's been a friend and an engineering colleague for many years since I was in grad school.
	Your support means a lot and inspires me to keep this series going.
	And now, here's my conversation with George Hotz.
	Do you think we're living in a simulation?
	Yes, but it may be unfalsifiable.
	What do you mean by unfalsifiable?
	So if the simulation is designed in such a way that they did like a formal proof to show that no information can get in and out, and if their hardware is designed for anything in the simulation to always keep the hardware in spec, it may be impossible to prove whether we're in a simulation or not.
	So they've designed it such that it's a closed system, you can't get outside the system.
	Well, maybe it's one of three worlds.
	We're either in a simulation which can be exploited, we're in a simulation which not only can't be exploited, but the same thing is true about VMs.
	A really well-designed VM, you can't even detect if you're in a VM or not.
	That's brilliant. So the simulation is running on a virtual machine.
	But now, in reality, all VMs have ways to detect.
	That's the point. I mean, you've done quite a bit of hacking yourself, so you should know that really any complicated system will have ways in and out.
	So this isn't necessarily true going forward.
	I spent my time away from comma.
	I learned a cock.
	It's a dependently typed...
	It's a language for writing math proofs in.
	And if you write code that compiles in a language like that, it is correct by definition.
	The types check its correctness.
	So it's possible that the simulation is written in a language like this, in which case...
	Yeah, but that can't be sufficiently expressive a language like that.
	Oh, it can. It can be?
	Oh, yeah. Okay.
	The simulation doesn't have to be term complete if it has a scheduled end date.
	Looks like it does, actually, with entropy.
	I mean, I don't think that a simulation that results in something as complicated as the universe would have a formal proof of correctness, right?
	Yeah. It's possible, of course.
	We have no idea how good their tooling is, and we have no idea how complicated the universe computer really is.
	It may be quite simple. It's just very large, right?
	It's definitely very large.
	But the fundamental rules might be super simple.
	Yeah, Conway's Game of Life kind of stuff.
	Right. So if you could hack...
	So imagine a simulation that is hackable.
	If you could hack it...
	What would you change?
	How would you approach hacking a simulation?
	The reason I gave that talk...
	By the way, I'm not familiar with the talk you gave.
	I just read that you talked about escaping the simulation or something like that.
	So maybe you can tell me a little bit about the theme and the message there too.
	It wasn't a very practical talk about how to actually escape a simulation.
	It was more about a way of restructuring an us-versus-them narrative.
	If we continue on the path we're going with technology, I think we're in big trouble, like as a species, and not just as a species, but even as me as an individual member of the species.
	So if we could change rhetoric to be more like to think upwards, like to think about that we're in a simulation and how we could get out, already we'd be on the right path.
	What you actually do once you do that, well, I assume I would have acquired way more intelligence in the process of doing that.
	So I'll just ask that. So the thinking upwards, what kind of ideas, what kind of breakthrough ideas do you think thinking in that way could inspire?
	And why did you say upwards?
	Upwards. Into space?
	Are you thinking sort of exploration in all forms?
	The space narrative that held for the modernist generation doesn't hold as well for the postmodern generation.
	What's the space narrative? Are we talking about the same space, the three-dimensional space?
	No, no, space, like going up in space, like building, like Elon Musk, like we're going to build rockets, we're going to go to Mars, we're going to colonize the universe.
	And the narrative you're referring, I was born in the Soviet Union, you're referring to the race to space.
	The race to space, yes. Explore, okay.
	That was a great modernist narrative.
	It doesn't seem to hold the same weight in today's culture.
	I'm hoping for good postmodern narratives that replace it.
	So let's think.
	So you work a lot with AI. So AI is one formulation of that narrative.
	There could be also, I don't know how much you do in VR and AR. Yeah.
	That's another, I know less about it, but every time I play with it in our research, it's fascinating, that virtual world.
	Are you interested in the virtual world?
	I would like to move to virtual reality.
	In terms of your work?
	No, I would like to physically move there.
	The apartment I can rent in the cloud is way better than the apartment I can rent in the real world.
	Well, it's all relative, isn't it?
	Because others will have very nice apartments too, so you'll be inferior in the virtual world as well.
	No, but that's not how I view the world, right?
	I don't view the world...
	I mean, that's a very almost zero-summish way to view the world.
	Say, like, my great apartment isn't great because my neighbor has one too.
	No, my great apartment is great because, like, look at this dishwasher, man.
	You just touch the dish and it's washed, right?
	And that is great in and of itself if I have the only apartment or if everybody had the apartment.
	I don't care. So you have fundamental gratitude.
	The world first learned of Geohat, George Hatz, in August 2007, maybe before then, but certainly in August 2007 when you were the first person to unlock, carry unlock an iPhone.
	How did you get into hacking?
	What was the first system you discovered vulnerabilities for and broke into?
	So, that was really kind of the first thing.
	I had a book in 2006 called Grey Hat Hacking.
	And I guess I realized that if you acquired these sort of powers, you could control the world.
	But I didn't really know that much about computers back then.
	I started with electronics. The first iPhone hack was physical.
	You had to open it up and pull an address line high.
	And it was because I didn't really know about software exploitation.
	I learned that all in the next few years and I got very good at it.
	But back then, I knew about how memory chips are connected to processors and stuff.
	You knew about software and programming.
	You just didn't know.
	Oh, really? So your view of the world and computers was physical, was hardware?
	Actually, if you read the code that I released with that in August 2007, it's atrocious.
	What language was it? A C. And in a broken sort of state machine-esque C, I didn't know how to program.
	Yeah. So how did you learn to program?
	What was your journey?
	Because, I mean, we'll talk about it.
	You've live-streamed some of your programming.
	This chaotic, beautiful mess.
	How did you arrive at that?
	Years and years of practice.
	I interned at Google the summer after the iPhone unlock.
	And I did a contract for them where I built hardware for Street View, and I wrote a software library to interact with it.
	And it was terrible code.
	And for the first time, I got feedback from people who I respected saying, like, don't write code like this.
	Now, of course, just getting that feedback is not enough.
	Mm-hmm. The way that I really got good was I wanted to write this thing that could emulate and then visualize ARM binaries because I wanted to hack the iPhone better.
	And I didn't like that I couldn't single step through the processor because I had no debugger on there, especially for the low level things like the bootrum and the bootloader.
	So I tried to build this tool to do it.
	And I built the tool once and it was terrible.
	I built the tool a second time, it was terrible.
	I built the tool a third time.
	This was by the time I was at Facebook, it was kind of okay.
	And then I built the tool a fourth time when I was a Google intern again in 2014.
	And that was the first time I was like, this is finally usable.
	How do you pronounce this, Kira?
	Kira, yeah. So it's essentially the most efficient way to visualize the change of state of the computer as the program is running.
	That's what you mean by debugger.
	Yeah, it's a timeless debugger, so you can rewind just as easily as going forward.
	Think about if you're using GDB, you have to put a watch on a variable if you want to see if that variable changes.
	In Cure, you can just click on that variable, and then it shows every single time when that variable was changed or accessed.
	Think about it like Git for your computers, the run log.
	So there's like a deep log of the state of the computer as the program runs and you can rewind.
	Why isn't that? Maybe it is.
	Maybe you can educate me. Why isn't that kind of debugging used more often?
	Because the tooling's bad.
	Well, two things. One, if you're trying to debug Chrome, Chrome is a 200 megabyte binary that runs slowly on desktops.
	So that's going to be really hard to use for that.
	But it's really good to use for CTFs and for boot ROMs and for small parts of code.
	So it's hard if you're trying to debug massive systems.
	What's a CTF and what's a boot ROM? A boot ROM is the first code that executes the minute you give power to your iPhone.
	Okay. And CTF were these competitions that I played, Capture the Flag.
	Capture the Flag. I was going to ask you about that.
	I watched a couple videos on YouTube.
	Those look fascinating. What have you learned about maybe at the high level of vulnerability of systems from these competitions?
	The, like... I feel like in the heyday of CTFs, you had all of the best security people in the world challenging each other and coming up with new toy exploitable things over here.
	And then everybody, okay, who can break it?
	And when you break it, you get like, there's like a file on the server called flag.
	And then there's a program running, listening on a socket that's vulnerable.
	So you write an exploit, you get a shell, and then you cat flag, and then you type the flag into like a web-based scoreboard and you get points.
	So the goal is essentially to find an exploit in a system that allows you to run shell, to run arbitrary code on that system.
	That's one of the categories.
	That's like the pwnable category.
	Um, pwnable?
	Yeah, pwnable. It's like, uh, you know, you pwn the program.
	You, uh, it's a program. Oh, uh, Yeah, you know, first of all, I apologize.
	I'm going to say it's because I'm Russian, but maybe you can help educate me.
	Some video game, like, misspelled own way back in the day.
	Yeah, and it's just... I wonder if there's a definition.
	I'll have to go to Urban Dictionary for it.
	It'll be interesting to see what it says. Okay, so what was the heyday of CTF, by the way?
	What decade are we talking about?
	I think like, I mean, maybe I'm biased because it's the era that I played, but like 2011 to 2015, Because the modern CTF scene is similar to the modern competitive programming scene.
	You have people who do drills.
	You have people who practice.
	And then once you've done that, you've turned it less into a game of generic computer skill and more into a game of, okay, you drill on these five categories.
	And then before that, It wasn't, it didn't have like as much attention as it had.
	I don't know, they were like, I won $30,000 once in Korea for one of these competitions.
	Holy crap. So that means, I mean money is money, but that means there was probably good people there.
	Exactly, yeah. Are the challenges human constructed or are they grounded in some real flaws and real systems?
	Usually they're human constructed, but they're usually inspired by real flaws.
	What kind of systems are imagined is really focused on mobile?
	Like what has vulnerabilities these days?
	Is it primarily mobile systems like Android?
	Oh, everything does.
	Yeah, of course.
	The price has kind of gone up because less and less people can find them.
	And what's happened in security is now if you want to like jailbreak an iPhone, you don't need one exploit anymore.
	You need nine. Nine chained together.
	What's the benefit?
	Speaking higher level philosophically about hacking, it sounds from everything I've seen about you, you just love the challenge and you don't want to do anything.
	You don't want to bring that exploit out into the world and let it run wild.
	You just want to solve it and then you go on to the next thing.
	Oh yeah, I mean, doing criminal stuff's not really worth it.
	And I'll actually use the same argument for why I don't do defense for why I don't do crime.
	If you want to defend a system, say the system has 10 holes, right?
	If you find 9 of those holes as a defender, you still lose because the attacker gets in through the last one.
	If you're an attacker, you only have to find 1 out of the 10.
	If you're a criminal, if you log on with a VPN nine out of the ten times, but one time you forget, you're done.
	Because you're caught.
	Okay. Because you only have to mess up once to be caught as a criminal.
	Yeah. That's why I'm not a criminal. Yeah.
	I was having a discussion with somebody just at a higher level about nuclear weapons, actually, why we haven't blown ourselves up yet.
	And my feeling is all the smart people in the world, if you look at the distribution of smart people, smart people are generally good.
	And then this other person, I was talking to Sean Carroll, the physicist, and he was saying, no, good and bad people are evenly distributed amongst everybody.
	My sense was good hackers are in general good people and they don't want to mess with the world.
	What's your sense? I'm not even sure about that.
	Like, I have a nice life.
	Crime wouldn't get me anything.
	But if you're good and you have these skills, you probably have a nice life too, right?
	Right, you can use it for other things.
	But is there an ethical... Is there a little voice in your head that says, well, yeah, if you could hack something to where you could hurt people and you could earn a lot of money doing it, though.
	Not hurt physically, perhaps, but disrupt their life in some kind of way.
	Isn't there a little voice that says...
	Well, two things.
	One, I don't really care about money.
	So, like, the money wouldn't be an incentive.
	The thrill might be an incentive.
	But when I was 19, I read Crime and Punishment.
	That was another great one that talked me out of ever really doing crime.
	Because it's like, that's going to be me.
	I'd get away with it, but it would just run through my head.
	Even if I got away with it, you know.
	And then you do crime for long enough, you'll never get away with it.
	That's right. In the end, that's a good reason to be good.
	I wouldn't say I'm good. I would just say I'm not bad.
	You're a talented programmer and a hacker in a good positive sense of the word.
	You've played around, found vulnerabilities in various systems.
	What have you learned broadly about the design of systems and so on from that whole process?
	You... Learn to not take things for what people say they are, but you look at things for what they actually are.
	I understand that's what you tell me it is, but what does it do?
	Right. And you have nice visualization tools to really know what it's really doing.
	Oh, I wish. I'm a better programmer now than I was in 2014.
	I said, Kira, that was the first tool that I wrote that was usable.
	I wouldn't say the code was great.
	I still wouldn't say my code is great.
	So how was your evolution as a programmer, except practice?
	You started with C. At which point did you pick up Python?
	Because you're pretty big in Python now.
	Now, yeah, in college.
	I went to Carnegie Mellon when I was 22.
	I went back, I'm like, I'm going to take all your hardest CS courses.
	We'll see how I do, right?
	Like, did I miss anything by not having a real undergraduate education?
	Took operating systems, compilers, AI, and their, like, freshman reader math course.
	And... Some of those classes you mentioned are pretty tough, actually.
	They're great. At least the circa 2012 operating systems and compilers were the best classes I've ever taken in my life.
	Because you write an operating system and you write a compiler.
	I wrote my operating system in C and I wrote my compiler in Haskell.
	But somehow I picked up Python that semester as well.
	I started using it for the CTFs actually.
	That's when I really started to get into CTFs.
	And CTFs you're all, it's a race against the clock.
	So I can't write things in C.
	Oh, there's a clock component.
	So you really want to use the programming language so you can be fastest in.
	48 hours. Pwn as many of these challenges as you can.
	Pwn.
	Yeah, you got like 100 points a challenge, whatever team gets the most.
	You were both at Facebook and Google for a brief stint.
	Yeah. With Project Zero, actually, at Google for five months, where you developed, Kira.
	What was Project Zero about in general?
	I'm just curious about the security efforts in these companies.
	Well, Project Zero started the same time I went there.
	What years are you there? 2015.
	2015. So that was right at the beginning of Project Year.
	It's small. It's Google's offensive security team.
	I'll try to give the best public-facing explanation that I can.
	So the idea is basically these vulnerabilities exist in the world.
	Nation states have them.
	Some high-powered bad actors have them.
	Sometimes People will find these vulnerabilities and submit them in bug bounties to the companies.
	But a lot of the companies don't really care.
	They don't even fix the bug. It doesn't hurt for there to be a vulnerability.
	So Project Zero is like, we're going to do it different.
	We're going to announce a vulnerability and we're going to give them 90 days to fix it.
	And then whether they fix it or not, we're going to drop the zero day.
	Oh, wow. We're going to drop the weapon on the textbook.
	That is so cool. I love that.
	Deadlines. Oh, that's so cool.
	Give them real deadlines. Yeah.
	And I think it's done a lot for moving the industry forward.
	I watched your coding sessions that you streamed online.
	You code things up, the basic projects, usually from scratch.
	I would say, sort of as a programmer myself, just watching you, that you type really fast and your brain works in both brilliant and chaotic ways.
	I don't know if that's always true, but certainly for the live streams.
	So it's interesting to me because I'm much slower and systematic and careful.
	And you just move, I mean, probably in order of magnitude faster.
	So I'm curious, is there a method to your madness?
	Or is it just who you are?
	There's pros and cons.
	There's pros and cons to my programming style, and I'm aware of them.
	If you ask me to get something up and working quickly with an API that's kind of undocumented, I will do this super fast, because I will throw things at it until it works.
	If you ask me to take a vector and rotate it 90 degrees and then flip it over the XY plane, I'll spam program for two hours and won't get it.
	Oh, because it's something that you could just do with a sheet of paper, think through, design, and then just...
	You really just throw stuff at the wall and you get so good at it that it usually works.
	I should become better at the other kind as well.
	Sometimes I'll do things methodically.
	It's nowhere near as entertaining on the Twitch streams.
	I do exaggerate it a bit on the Twitch streams as well.
	The Twitch streams, I mean, what do you want to see a game where you want to see actions permitted, right?
	I'll show you APM for programming too.
	I recommend people go to it.
	I think I watched...
	I watched probably several hours of you...
	Like, I've actually left you programming in the background while I was programming.
	Because you made me...
	It was like watching a really good gamer.
	It energizes you.
	Because you're moving so fast.
	It's awesome. It's inspiring.
	It made me jealous that...
	Because my own programming is inadequate in terms of speed.
	I'm twice as frantic on the live streams as I am when I code without them.
	It's super entertaining.
	I wasn't even paying attention to what you were coding, which is great.
	It's just watching you switch windows and Vim, I guess.
	Yeah, Vim and Screen.
	I've developed the workflow at Facebook and stuck with it.
	How do you learn new programming tools, ideas, techniques these days?
	What's your methodology for learning new things?
	So I wrote for Kama The distributed file systems out in the world are extremely complex.
	Like if you want to install something like Ceph.
	Ceph is, I think, the open infrastructure distributed file system.
	Or there's newer ones like SeaweedFS.
	But these are all 10,000 plus line projects.
	I think some of them are even 100,000 line projects.
	And just configuring them was a nightmare.
	So I wrote one.
	It's 200 lines.
	And it uses like Nginx volume servers and has this little master server that I wrote in Go.
	And the way I... Go, wow.
	If I would say that I'm proud per line of any code I wrote, maybe there's some exploits that I think are beautiful and then this.
	This is 200 lines.
	And just the way that I thought about it, I think, was very good.
	And the reason it's very good is because that was the fourth version of it that I wrote.
	And I had three versions that I threw away.
	You mentioned, did you say Go?
	I wrote in Go, yeah. In Go.
	Is that a functional language?
	I forget what Go is. Go is Google's language.
	Right. It's not functional.
	It's... In a way, it's C++, but easier.
	It's strongly typed.
	It has a nice ecosystem around it.
	When I first looked at it, I was like, this is like Python, but it takes twice as long to do anything.
	Now that OpenPilot is migrating to C, but it still has large Python components, I now understand why Python doesn't work for large codebases and why you want something like Go.
	Interesting. So why doesn't Python work for...
	So even most, speaking for myself at least, we do a lot of stuff, basically demo-level work with autonomous vehicles, and most of the work is Python.
	Yeah. Why doesn't Python work for large code bases?
	Because, well, lack of type checking is a big problem.
	So errors creep in. Yeah, and like, you don't know, the compiler can tell you like nothing, right?
	So everything is either, you know, like syntax error is fine, but if you misspell a variable in Python, the compiler won't catch that.
	There's like linters that can catch it some of the time.
	There's no types. This is really the biggest downside.
	And then, well, Python's slow, but that's not related to it.
	Well, maybe it's kind of related to it, so it's lack of...
	So what's in your toolbox these days?
	Is it Python? What else?
	Go? I need to move to something else.
	My adventure into dependently typed languages, I love these languages.
	They just have, like, syntax from the 80s.
	What do you think about JavaScript, ES6, like the modern TypeScript?
	JavaScript is, the whole ecosystem is unbelievably confusing.
	NPM updates a package from 0.2.2 to 0.2.5 and that breaks your Babel linter which translates your ES5 into ES6 which doesn't run on.
	So why do I have to compile my JavaScript again, huh?
	It may be the future, though.
	You think about, I mean, I've embraced JavaScript recently, just because, just like I've continually embraced PHP, it seems that these worst possible languages live on for the longest, like cockroaches never die.
	Yeah. Well, it's in the browser and it's fast.
	It's fast. Yeah.
	It's in the browser and compute might become, you know, the browser.
	It's unclear what the role of the browser is in terms of distributed computation in the future.
	So... JavaScript is definitely here to stay.
	Yeah. It's interesting if autonomous vehicles will run on JavaScript one day.
	I mean, you have to consider these possibilities.
	Well, all our debug tools are JavaScript.
	We actually just open source them.
	We have a tool, Explorer, which you can annotate your disengagements.
	And we have a tool, Cabana, which lets you analyze the can traffic from the car.
	So basically, anytime you're visualizing something about the log, you're using JavaScript.
	JavaScript, yeah. Well, the web is the best UI toolkit by far.
	And then, you know what? You're coding in JavaScript.
	We have a React guy. He's good.
	React. Nice. Let's get into it.
	So let's talk autonomous vehicles.
	You founded Kama AI. At a high level, how did you get into the world of vehicle automation?
	Can you also just, for people who don't know, tell the story of Kama AI? Sure.
	So I was working at this AI startup and a friend approached me and he's like, I don't know where this is going, but the coolest applied AI problem today is self-driving cars.
	I'm like, well, absolutely. Do you want to meet with Elon Musk?
	And he's looking for somebody to build a vision system for autopilot.
	This is when they were still on AP1. They were still using Mobileye.
	Elon back then was looking for a replacement and he brought me in and we talked about a
	contract where I would deliver something that meets mobile eye level performance.
	I would get paid $12 million if I could deliver it tomorrow and I would lose $1 million for
	every month I didn't deliver.
	So I was like, okay, this is a great deal.
	This is a super exciting challenge.
	Yeah.
	you You know what? Even if it takes me 10 months, I get $2 million.
	It's good. Maybe I can finish up in five.
	Maybe I don't finish it at all and I get paid nothing and I'll work for 12 months for free.
	So maybe just take a pause on that.
	I'm also curious about this because I've been working in robotics for a long time.
	I'm curious to see a person like you just step in and sort of somewhat naive, but brilliant, right?
	So that's the best place to be because you basically full steam take on a problem.
	How confident, how from that time, because you know a lot more now, at that time, how hard do you think it is to solve all of autonomous driving?
	I remember I suggested to Elon in the meeting putting a GPU behind each camera to keep the compute local.
	This is an incredibly stupid idea.
	I leave the meeting 10 minutes later and I'm like, I could have spent a little bit of time thinking about this problem before I went in.
	Why is this a stupid idea? Oh, just send all your cameras to one big GPU. You're much better off doing that.
	Oh, sorry. You said behind every camera have a small GPU. I was like, oh, I'll put the first few layers of my conv there.
	Like, why'd I say that?
	That's possible. It's possible, but it's a bad idea.
	It's not obviously a bad idea.
	Pretty obviously a bad idea. But whether it's actually a bad idea or not, I left that meeting with Elon, like, beating myself up.
	I'm like, why did I say something stupid?
	Yeah, you haven't... Like, you haven't at least, like, thought through every aspect fully, yeah.
	He's very sharp, too. Like, usually in life, I get away with saying stupid things, and then kind of, of course...
	Oh, right away, he called me out about it.
	And, like, usually in life, I get away with saying stupid things, and then, like, people will...
	You know, a lot of times people don't even notice.
	And I'll, like, correct it and bring the conversation back.
	But with Elon, it was like, nope. Like, okay, well, that's not at all why the contract fell through.
	I was much more prepared the second time I met him.
	Yeah. But in general, how hard did you think it is?
	Like, 12 months is a tough timeline.
	Oh, I just thought I'd clone Mobileye IQ 3.
	I didn't think I'd solve level 5 self-driving or anything.
	So the goal there was to do lane keeping, good lane keeping.
	I saw, my friend showed me the outputs from a Mobileye, and the outputs from a Mobileye was just basically two lanes and a position of a lead car.
	I'm like, I can gather a data set and train this net in weeks.
	And I did. Well, first time I tried the implementation of Mobileye in a Tesla, I was really surprised how good it is.
	It's going incredibly good.
	Because I thought it's, just because I've done a lot of computer vision, I thought it'd be a lot harder to create a system that's that stable.
	So I was personally surprised.
	I have to admit it because I was kind of skeptical before trying it because I thought it would go in and out a lot more.
	It would get disengaged a lot more and it's pretty robust.
	So how hard is the problem when you tackled it?
	I think AP1 was great.
	Elon talked about disengagements on the 405 down in LA where the lane marks were kind of faded and the mobile eye system would drop out.
	I had something up and working that I would say was the same quality in three months.
	Same quality, but how do you know?
	You say stuff like that confidently, and I love it, but the question is, you can't...
	You're kind of going by feel, because you chest it out.
	You're going by feel, absolutely. I borrowed my friend's Tesla.
	I would take AP1 out for a drive, and then I would take my system out for a drive.
	It seems reasonably the same.
	So, the 405, how hard is it to create something that could actually be a product that's deployed?
	I mean, I've read an article where Elon said something about you saying that to build autopilot is more complicated than a single George Hodge level job.
	How hard is that job to create...
	Something that would work across globally?
	I don't think globally is the challenge.
	But Elon followed that up by saying it's going to take two years in a company of 10 people.
	Yeah. And here I am four years later with a company of 12 people.
	And I think we still have another two to go.
	Two years. So yeah.
	So what do you think about how Tesla's progressing with Autopilot V2, V3? I think we've kept pace with them pretty well.
	I think Navigator and Autopilot is terrible.
	We had some demo features internally of the same stuff and we would test it and I'm like, I'm not shipping this even as like open source software to people.
	Why do you think it's terrible? Yeah.
	Like when it makes a lane change, it does it worse than a human.
	You shouldn't ship things like autopilot, open pilot, they lane keep better than a human.
	If you turn it on for a stretch of highway, like an hour long, it's never going to touch
	a lane line.
	A human will touch probably a lane line twice.
	I don't know if you're grounded in data on that.
	I read your paper. Okay.
	No, but that's interesting.
	I wonder actually how often we touch lane lines in general, like a little bit, because it is...
	I could answer that question pretty easily with the common data set.
	Yeah. I'm curious. I've never answered it.
	I don't know. I just... Two is like my...
	It feels right.
	That's interesting because every time you touch a lane, that's a source of a little bit of stress and kind of lane keeping is removing that stress.
	That's ultimately the biggest value add, honestly, is just removing the stress of having to stay in lane.
	And I don't think people fully realize, first of all, that that's a big value add, but also that that's all it is.
	And that, not only, I find it a huge value add.
	I drove down, when we moved to San Diego, I drove down in an enterprise rent-a-car, and I missed it.
	I missed having the system so much.
	It's so much more tiring to drive without it.
	It is that lane centering that's the key feature.
	Yeah. And in a way, it's the only feature that actually adds value to people's lives in autonomous vehicles today.
	Waymo does not add value to people's lives.
	It's a more expensive, slower Uber.
	Maybe someday it'll be this big cliff where it adds value, but I don't usually believe it.
	It is fascinating. I haven't talked to...
	This is good, because I haven't...
	I have intuitively, but I think we're making it explicit now.
	I actually believe that really good lane keeping Is a reason to buy a car, will be a reason to buy a car, and it's a huge value add.
	I've never, until we just started talking about it, I haven't really quite realized it, that I've felt with Elon's chase of level four is not the correct chase.
	It was on, because you should just say, Tesla has the best, as if from a Tesla perspective, say, Tesla has the best lane keeping.
	Kameyai should say, Kameyai is the best lane keeping.
	And that is it.
	Yeah. Do you think?
	You have to do the longitudinal as well.
	You can't just lane keep.
	You have to do ACC, but ACC is much more forgiving than lane keep, especially on the highway.
	By the way, are you Kamei's camera only, correct?
	No, we use the radar. From the car, you're able to get the...
	We can do a camera only now.
	It's gotten to the point, but we leave the radar there as like a...
	It's fusion now.
	Okay, so let's maybe talk through some of the system specs on the hardware.
	What's the hardware side of what you're providing?
	What's the capabilities on the software side with OpenPilot and so on?
	So OpenPilot...
	As the box that we sell that it runs on, it's a phone in a plastic case.
	It's nothing special. We sell it without the software.
	So you buy the phone, it's just easy.
	It'll be easy setup, but it's sold with no software.
	OpenPilot right now is about to be 0.6.
	When it gets to 1.0, I think we'll be ready for a consumer product.
	We're not going to add any new features.
	We're just going to make the lane keeping really, really good.
	Okay, I got it. So what do we have right now?
	It's a Snapdragon 820.
	It's a Sony IMX298 forward-facing camera.
	Driver monitoring camera.
	Just use a selfie cam on the phone. And...
	A CAN transceiver, a little thing called PANDAS, and they talk over USB to the phone, and then they have three CAN buses that they talk to the car.
	One of those CAN buses is the radar CAN bus.
	One of them is the main car CAN bus, and the other one is the proxy camera CAN bus.
	We leave the existing camera in place so we don't turn AEB off.
	Right now, we still turn AEV off if you're using our longitudinal, but we're going to fix that before 1.0.
	Got it. Wow, that's cool.
	And it's canned both ways.
	So, how are you able to control vehicles?
	So we proxy the vehicles that we work with already have a lane-keeping assist system.
	So lane-keeping assist can mean a huge variety of things.
	It can mean it will apply a small torque to the wheel after you've already crossed a lane line by a foot, which is the system in the older Toyotas, versus like, I think Tesla still calls it lane-keeping assist, where it'll keep you perfectly in the center of the lane on the highway.
	Right. You can control, like with a joystick, the car.
	So these cars already have the capability of drive-by wire.
	So is it trivial to convert a car that it operates with, that OpenPile is able to control the steering?
	Oh, a new car?
	So we have support now for 45 different makes of cars?
	What are the cars in general?
	Mostly Hondas and Toyotas.
	We support almost every Honda and Toyota made this year.
	And then a bunch of GMs, a bunch of Subarus, a bunch of Chevys.
	It doesn't have to be like a Prius.
	It could be a Corolla as well.
	The 2020 Corolla is the best car with OpenPilot.
	It just came out. The actuator has less lag than the older Corolla.
	I think I started watching a video with your...
	The way you make videos is awesome.
	You're just literally at the dealership streaming.
	I had my friend on the phone.
	I'm like, bro, you want to stream for an hour?
	Yeah. And basically, if stuff goes a little wrong, you just go with it.
	Yeah, I love it. Well, it's real.
	Yeah, it's real. That's so beautiful.
	And it's so in contrast to the way other companies would put together a video like that.
	Kind of why I like to do it like that.
	Good. I mean, if you become super rich one day and successful, I hope you keep it that way because I think that's actually what people love, that kind of genuine...
	Oh, it's all that has value to me.
	Yeah. Money has no...
	If I sell out to, like, make money, I sold out.
	It doesn't matter. What do I get?
	A yacht? I don't want a yacht. Yeah.
	And I think Tesla actually has a small inkling of that as well.
	With Autonomy Day, they did reveal more than...
	I mean, of course, there's marketing communications, you can tell, but it's more than most companies would reveal, which is...
	I hope they go towards that direction more, other companies, GM, Ford...
	Oh, Tesla's going to win Level 5.
	They really are. So let's talk about it.
	You think you're focused on level two?
	Currently, currently. We're going to be one to two years behind Tesla getting to level five.
	Okay. We're Android, right?
	We're Android. You're Android.
	I'm just saying once Tesla gets it, we're one to two years behind.
	I'm not making any timeline on when Tesla's going to get it.
	That's right. You did. That's brilliant. I'm sorry, Tesla investors, if you think you're going to have an autonomous robo-taxi fleet by the end of the year...
	I'll bet against that.
	So what do you think about this, the most level four companies They are kind of just doing their usual safety driver, doing full-autonomy kind of testing.
	And then Tesla does basically trying to go from lane keeping to full autonomy.
	What do you think about that approach?
	How successful would it be?
	It's a ton better approach.
	Because Tesla is gathering data on a scale that none of them are.
	They're putting real users behind the wheel of the cars.
	It's, I think, the only strategy that works.
	The incremental. Well, so there's a few components to the Tesla approach that's more than just the incremental.
	What you spoke with is the one is the software, so over-the-air software updates.
	Necessity. I mean, Waymo crews have those too.
	Those aren't. Those differentiate everything from the automakers.
	Right. No lane keeping system, no cars or lane keeping system have that except Tesla.
	And the other one is the data, the other direction, which is the ability to query the data.
	I don't think they're actually collecting as much data as people think, but the ability to turn on collection and turn it off.
	So, I'm both in the robotics world and the psychology, human factors world.
	Many people believe that level 2 autonomy is problematic because of the human factor.
	Like, the more the task is automated, the more there's a vigilance decrement, you start to fall asleep, you start to become complacent, start texting more and so on.
	Do you worry about that?
	Because if you're talking about transition from lane keeping to full autonomy, if you're spending 80% of the time Not supervising the machine.
	Do you worry about what that means to the safety of the drivers?
	One, we don't consider OpenPilot to be 1.0 until we have 100% driver monitoring.
	You can cheat right now, our driver monitoring system.
	There's a few ways to cheat it. They're pretty obvious.
	We're working on making that better.
	Before we ship a consumer product that can drive cars, I want to make sure that I have driver monitoring that you can't cheat.
	What's a successful driver monitoring system look like?
	Is it all about just keeping your eyes on the road?
	Well, a few things.
	So that's what we went with at first for driver monitoring.
	I'm actually looking at where your head is looking.
	The camera's not that high resolution.
	Eyes are a little bit hard to get. Well, head is big.
	Head is good.
	And actually, a lot of it just...
	Psychology-wise, to have that monitor constantly there, it reminds you that you have to be paying attention.
	But we want to go further.
	We just hired someone full-time to come on to do the driver monitoring.
	I want to detect phone in frame, and I want to make sure you're not sleeping.
	How much does the camera see of the body?
	This one, not enough.
	Not enough. The next one, everything.
	What's interesting, Fisheye, because we're doing just data collection, not real time, but Fisheye is a beautiful, being able to capture the body.
	And the smartphone is really the biggest problem.
	I can show you one of the pictures from our new system.
	Awesome. So you're basically saying the driver monitoring will be the answer to that.
	I think the other point that you raised in your paper is good as well.
	You're not asking a human to supervise a machine without giving them the...
	They can take over at any time. Right.
	Our safety model, you can take over.
	We disengage on both the gas or the brake.
	We don't disengage on steering.
	I don't feel you have to. But we disengage on gas or brake.
	So it's very easy for you to take over, and it's very easy for you to re-engage.
	That switching should be super cheap.
	The cars that require...
	Even autopilot requires a double press.
	That's almost, I see, I don't like that.
	And then the cancel.
	To cancel in autopilot, you either have to press cancel, which no one knows what that is, so they press the brake.
	But a lot of times you don't actually want to press the brake.
	You want to press the gas, so you should cancel on gas.
	Or wiggle the steering wheel, which is bad as well.
	Wow, that's brilliant. I haven't heard anyone articulate that point.
	Oh, this is all I think about.
	Because I think actually Tesla has done a better job than most automakers at making that frictionless.
	But you just described that it could be even better.
	I love Super Cruise as an experience once it's engaged.
	I don't know if you've used it, but getting the thing to try to engage...
	Yeah, I've driven Super Cruise a lot.
	So what's your thoughts on the Super Cruise system in general?
	You disengage Super Cruise and it falls back to ACC, so my car's still accelerating.
	It feels weird. Otherwise, when you actually have Super Cruise engaged on the highway, it is phenomenal.
	We bought that Cadillac.
	We just sold it, but we bought it just to experience this.
	And I wanted everyone in the office to be like, this is what we're striving to build.
	GM pioneering with the driver monitoring.
	You like their driver monitoring system?
	It has some bugs. If there's a sun shining back here, it'll be blind to you.
	But overall, mostly, yeah.
	That's so cool that you know all this stuff.
	I don't often talk to people that, because it's such a rare car, unfortunately, currently.
	We bought one explicitly for this.
	We lost like 25k in the deprecation, but if it was worth it.
	I was very pleasantly surprised that GM system was so innovative and really wasn't advertised much, wasn't talked about much, and I was nervous that it would die, that it would disappear.
	Well, they put it on the wrong car.
	They should have put it on the Bolt and not some weird Cadillac that nobody bought.
	I think that's going to be, they're saying at least it's going to be into their entire fleet.
	So what do you think about, as long as we're on the driver monitoring, what do you think about Elon Musk's claim that driver monitoring is not needed?
	Normally, I love his claims.
	That one is stupid. That one is stupid, and he's not going to have his level 5 fleet by the end of the year.
	Hopefully, he's like, okay, I was wrong.
	I'm going to add driver monitoring. Because when these systems get to the point that they're only messing up once every thousand miles, you absolutely need driver monitoring.
	So let me play, because I agree with you, but let me play devil's advocate.
	One possibility is that...
	Without driver monitoring, people are able to self-regulate, monitor themselves.
	You've seen all the people sleeping in Teslas?
	Yeah. Well, I'm a little skeptical of all the people sleeping in Teslas because...
	I've stopped paying attention to that kind of stuff because I want to see real data.
	There's too much glorified.
	It doesn't feel scientific to me.
	So I want to know how many people are really sleeping in Teslas versus sleeping.
	I was driving here sleep deprived in a car with no automation.
	I was falling asleep.
	I agree that it's hypey.
	It's just like, you know what?
	If Elon put driver monitoring, my last autopilot experience was I rented a Model 3 in March and drove it around.
	The wheel thing is annoying.
	And the reason the wheel thing is annoying, we use the wheel thing as well, but we don't disengage on wheel.
	For Tesla, you have to touch the wheel just enough to trigger the torque sensor to tell it that you're there, but not enough as to disengage it.
	Which, don't use it for two things.
	Don't disengage on wheel, you don't have to.
	That whole experience, wow, beautifully put.
	All those elements, even if you don't have driver monitoring, that whole experience needs to be better.
	Driver monitoring, I think would make, I mean, I think Super Cruise is a better experience once it's engaged over autopilot.
	I think Super Cruise's transition to engagement and disengagement are significantly worse.
	There's a tricky thing, because if I were to criticize Super Cruise, it's a little too crude, and I think it's like six seconds or something, if you look off-road, it'll start warning you.
	It's some ridiculously long period of time.
	And just the way, I think it's basically, it's a binary.
	It should be adaptive. Yeah, it needs to learn more about you.
	It needs to communicate what it sees about you more.
	Tesla shows what it sees about the external world.
	It would be nice if Super Cruise would tell us what it sees about the internal world.
	It's even worse than that.
	You press the button to engage and it just says supercruise unavailable.
	Why? Yeah, that transparency is good.
	We've renamed the driver monitoring packet to driver state.
	Driver state. We have car state packet, which has the state of the car, and driver state packet, which has the state of the driver.
	Estimate their BAC. What's BAC? Blood alcohol content.
	You think that's possible with computer vision?
	Absolutely. To me, it's an open question.
	I haven't looked into it too much.
	Actually, I quite seriously looked at the literature.
	It's not obvious to me that from the eyes and so on, you can tell.
	You might need stuff from the car as well.
	You might need how they're controlling the car, right?
	And that's fundamentally, at the end of the day, what you care about.
	But I think, especially when people are really drunk, they're not controlling the car nearly as smoothly as they would look at them walking, right?
	The car is like an extension of the body.
	So I think you could totally detect that.
	And if you could fix people who are drunk, distracted, asleep, if you fix those three...
	Yeah, that's huge.
	So what are the current limitations of OpenPilot?
	What are the main problems that still need to be solved?
	We're hopefully fixing a few of them in 06.
	We're not as good as autopilot at stopped cars.
	So if you're coming up to a red light at like 55, so it's the radar stopped car problem, which is responsible for two autopilot accidents, it's hard to differentiate a stopped car from a signpost.
	Yeah, a static object.
	So you have to fuse. You have to do this visually.
	There's no way from the radar data to tell the difference.
	Maybe you can make a map, but I don't really believe in mapping at all anymore.
	Wait, wait, wait. What? You don't believe in mapping?
	No. So you basically, the OpenPilot solution is saying react to the environment as you see it, just like human beings do.
	And then eventually when you want to do navigate on OpenPilot, I'll train the net to look at Waze.
	I'll run Waze in the background.
	I'll train the content on Waze. Are you using GPS at all?
	We use it to ground truth. We use it to very carefully ground truth the paths.
	We have a stack which can recover relative to 10 centimeters over one minute.
	And then we use that to ground truth exactly where the car went in that local part of the environment, but it's all local.
	How are you testing in general, just for yourself, like experiments and stuff?
	Where are you located?
	San Diego. So you basically drive around there and collect some data and watch the performance?
	We have a simulator now. And our simulator is really cool.
	It's not like a Unity-based simulator.
	Our simulator lets us load in real estate.
	We can load in a drive and simulate what the system would have done on the historical data.
	Ooh, nice. Interesting.
	Right now, we're only using it for testing, but as soon as we start using it for training, that's it.
	What's your feeling about the real world versus simulation?
	Do you like simulation for training, if this moves to training?
	So, we have to distinguish two types of simulators, right?
	There's a simulator that is completely fake.
	I could get my car to drive around in GTA. I feel that this kind of simulator is useless.
	You're never... There's so many...
	My analogy here is like, okay, fine.
	You're not solving the computer vision problem, but you're solving the computer graphics problem.
	Right. And you don't think you can get very far by creating ultra-realistic graphics.
	No. Because you can create ultra-realistic graphics of the road, now create ultra-realistic behavioral models of the other cars.
	Oh, well, I'll just use my self-driving.
	No, you won't.
	You need actual human behavior because that's what you're trying to learn.
	Driving does not have a spec.
	The definition of driving is what humans do when they drive.
	Whatever Waymo does, I don't think it's driving.
	Right. Well, I think actually Waymo and others, if there's any use for reinforcement learning, I've seen it used quite well, I study pedestrians a lot too, is try to train models from real data of how pedestrians move and try to use reinforcement learning models to make pedestrians move in human-like ways.
	By that point, you've already gone so many layers.
	You detected a pedestrian?
	Did you hand code the feature vector of their state?
	Did you guys learn anything from computer vision before deep learning?
	Well, okay. I feel like this is...
	So perception to you is the sticking point.
	I mean, what's the hardest part of the stack here?
	There is no human understandable feature vector separating perception and planning.
	That's the best way I can put that.
	So it's all together, and it's a joint problem.
	So you can take localization.
	Localization and planning, there is a human understandable feature vector between these two things.
	I mean, okay, so I have like three degrees position, three degrees orientation, and those derivatives, maybe those second derivatives, right?
	That's human understandable. That's physical.
	The between perception and planning.
	So Waymo has a perception stack and then a planner.
	And one of the things Waymo does right is they have a simulator that can separate those two.
	They can replay their perception data and test their system, which is what I'm talking about, about the two different kinds of simulators.
	There's the kind that can work on real data and there's the kind that can't work on real data.
	Now, the problem is that I don't think you can hand code a feature vector.
	You have some list of like, oh, here's my list of cars in the scenes.
	Here's my list of pedestrians in the scene.
	This isn't what humans are doing.
	What are humans doing?
	Global. And you're saying that's too difficult to hand engineer?
	I'm saying that there is no state vector.
	Given a perfect, I could give you the best team of engineers in the world to build a perception system and the best team to build a planner.
	All you have to do is define the state vector that separates those two.
	I'm missing the state vector that separates those two.
	What do you mean? So what is the output of your perception system?
	I'll put it in the perception system.
	Okay, well, there's several ways to do it.
	One is the slam component is localization.
	The other is drivable area, drivable space.
	Drivable space, yep. And then there's the different objects in the scene.
	Yep. Different objects in the scene over time, maybe, to give you input to then try to start modeling the trajectories of those objects.
	Sure. That's it.
	I can give you a concrete example of something you missed.
	What's that? So, say there's a bush in the scene.
	Humans understand that when they see this bush, that there may or may not be a car behind that bush.
	Drivable area and a list of objects does not include that.
	Humans are doing this constantly at the simplest intersections.
	So now you have to talk about occluded area.
	Right. Right? But even that, what do you mean by occluded?
	Okay, so I can't see it.
	Well, if it's the other side of a house, I don't care.
	What's the likelihood that there's a car in that occluded area?
	Right? And if you say, okay, we'll add that.
	I can come up with 10 more examples that you can't add.
	Certainly occluded area would be something that a simulator would have because it's simulating the entire, you know, the occlusion is part of it.
	Occlusion is part of a vision stack.
	But what I'm saying is if you have a hand engineered, if your perception system output can be written in a spec document, it is incomplete.
	Yeah. I mean, certainly it's hard to argue with that because in the end that's going to be true.
	Yes. And I'll tell you what the output of our perception system is.
	What's that? It's a 1,024 dimensional vector.
	Oh, you're on that.
	No, it's 1,024 dimensions of who knows what.
	Because it's operating on real data.
	And that's the perception.
	That's the perception state.
	Think about an autoencoder for faces.
	If you have an autoencoder for faces and you say it has 256 dimensions in the middle, and I'm taking a face over here and projecting it to a face over here.
	Can you hand label all 256 of those dimensions?
	Well, no, but those are generated automatically.
	But even if you tried to do it by hand, could you come up with a spec between your encoder and your decoder?
	No, because it wasn't designed.
	No, no, no, but if you could design it, if you could design a face reconstructor system, could you come up with a spec?
	No, but I think we're missing here a little bit.
	I think you're just being very poetic about expressing a fundamental problem of simulators, that they're going to be missing so much that the feature vectors will just look fundamentally different in the simulated world than the real world.
	I'm not making a claim about simulators.
	I'm making a claim about the spec division between perception and planning, even in your system.
	Just in general. Right?
	Just in general. If you're trying to build a car that drives, if you're trying to hand code the output of your perception system, like saying, like, here's a list of all the cars in the scene.
	Here's a list of all the people. Here's a list of the occluded areas.
	Here's a vector of drivable areas.
	It's insufficient. And if you start to believe that, you realize that what Waymo and Cruise are doing is impossible.
	Currently what we're doing is the perception problem is converting the scene into a chessboard.
	And then you reason some basic reasoning around that chessboard.
	And you're saying that really there's a lot missing there.
	First of all, why are we talking about this?
	Because isn't this a full autonomy?
	Is this something you think about?
	Oh, I want to win self-driving cars.
	So your definition of win includes?
	Level four. Level five.
	I don't think level four is a real thing.
	I want to build the alpha go of driving.
	So AlphaGo is really end to end.
	Yeah. Yeah, it's end to end.
	And do you think this whole problem, is that also kind of what you're getting at with the perception and the planning?
	Is that this whole problem, the right way to do it, is really to learn the entire thing.
	I'll argue that not only is it the right way, it's the only way that's going to exceed human performance.
	Well... It's certainly true for Go.
	Everyone who tried to hand code Go things built human inferior things.
	And then someone came along and wrote some 10,000 line thing that doesn't know anything about Go that beat everybody.
	It's 10,000 lines.
	True. In that sense, the open question then that maybe I can ask you is driving...
	Is much harder than go.
	The open question is how much harder.
	So how, because I think the Elon Musk approach here with planning and perception is similar to what you're describing, which is really turning into Not some kind of modular thing, but really do formulate as a learning problem and solve the learning problem with scale.
	So how many years?
	How many years would it take to solve this problem or just how hard is this freaking problem?
	Well, the cool thing is I think there's a lot of value that we can deliver along the way.
	I think that you can build Lane-keeping assist, actually, plus adaptive cruise control, plus, okay, looking at ways, extends to, like, all of driving.
	Yeah, most of driving, right?
	Oh, your adaptive cruise control treats red lights like cars.
	Okay. So, let's jump around.
	You mentioned that you didn't, like, navigate an autopilot.
	Yeah. What advice, how would you make it better?
	Do you think as a feature that, if it's done really well, it's a good feature?
	Reliant on like hand-coded hacks for like how does navigate an autopilot do a lane change?
	It actually does the same lane change every time and it feels mechanical humans do different lane changes
	Humans sometime will do a slow one sometimes do a fast one navigate an autopilot at least every time I use that is it
	the identical?
	lane change How do you learn?
	I mean, this is a fundamental thing actually, is the braking and then accelerating, something that's still, Tesla probably does it better than most cars, but it still doesn't do a great job of creating a comfortable, natural experience.
	And navigate on autopilot is just lane change is an extension of that.
	So how do you learn to do natural lane change?
	So, we have it, and I can talk about how it works.
	So, I feel that we have the solution for lateral, but we don't yet have the solution for longitudinal.
	There's a few reasons longitudinal is harder than lateral.
	The lane change component...
	The way that we train on it very simply is like our model has an input for whether it's doing a lane change or not.
	And then when we train the end-to-end model, we hand label all the lane changes because you have to.
	I've struggled a long time about not wanting to do that, but I think you have to.
	For the training data.
	For the training data, right?
	We actually have an automatic ground truther which automatically labels all the lane changes.
	Is that possible? To automatically label lane changes?
	Yeah. Yeah, I detect the lane.
	I see when it crosses it, right? And I don't have to get that high percent accuracy, but it's like 95, good enough.
	Okay. Now, I set the bit when it's doing the lane change in the end-to-end learning.
	And then I set it to zero when it's not doing a lane change.
	So now if I want it to do a lane change at test time, I just put the bit to a one and it'll do a lane change.
	Yeah. But so if you look at the space of lane change, you know, some percentage, not a hundred percent that we make as humans is not a pleasant experience because we messed some part of it up.
	Yeah. It's nerve wracking to change.
	You have to look, you have to see, you have to accelerate.
	How do we label the ones that are natural and feel good?
	You know, that's the, cause that's your ultimate criticism.
	The current navigating autopilot just doesn't feel good.
	Well, the current Navigator Autopilot is a hand-coded policy written by an engineer in a room who probably went out and tested it a few times on the 280.
	Probably a better version of that, but yes.
	That's how we would have written it.
	Maybe Tesla, they tested it.
	That might have been 200 years.
	200 years, yeah.
	No, but so if you learn the lane change, if you learn how to do a lane change from data, just like you have a label that says lane change and then you put it in when you want it to do the lane change...
	It'll automatically do the lane change that's appropriate for the situation.
	Now, to get at the problem of some humans do bad lane changes, we haven't worked too much on this problem yet.
	It's not that much of a problem in practice.
	My theory is that all good drivers are good in the same way and all bad drivers are bad in different ways.
	And we've seen some data to back this up.
	Well, beautifully put. So you just basically, if that's true, hypothesis, then your task is to discover the good drivers.
	The good drivers stand out because they're in one cluster and the bad drivers are scattered all over the place and your net learns the cluster.
	Yeah, so you just learn from the good drivers, and they're easy to cluster.
	In fact, we learn from all of them, and the net automatically learns the policy that's like the majority.
	But we'll eventually probably have to filter that out.
	If that theory is true, I hope it's true.
	Because the counter theory is there is many clusters, maybe arbitrarily many clusters of good drivers.
	Yeah. Because if there's one cluster of good drivers, you can at least discover a set of policies.
	You can learn a set of policies, which would be good universally.
	That would be nice if it's true.
	And you're saying that there is some evidence that...
	Let's say lane changes can be clustered into four clusters.
	Right. There's a finite level of...
	I would argue that all four of those are good clusters.
	All the things that are random are noise and probably bad.
	And which one of the four you pick, or maybe it's 10 or maybe it's 20...
	You can learn that. It's context dependent.
	It depends on the scene. And the hope is it's not too dependent on the driver.
	Yeah, the hope is that it all washes out.
	The hope is that the distribution's not bimodal.
	The hope is that it's a nice Gaussian.
	So what advice would you give to Tesla how to fix, how to improve navigating an autopilot?
	The lessons that you've learned from Comm.ai?
	The only real advice I would give to Tesla is please put driver monitoring in your cars.
	With respect to improving it...
	You can't do that anymore. I started to interrupt.
	But, you know, there's a practical nature of many of hundreds of thousands of cars being produced that don't have a good driver-facing camera.
	The Model 3 has a selfie cam.
	Is it not good enough? Did they not have put IRL LEDs for night?
	That's a good question. But I do know that it's fisheye and it's relatively low resolution.
	So it's really not designed.
	It wasn't... It wasn't designed for Dermotor.
	You can hope that you can kind of scrape up and have something from it.
	Yeah. But... Why didn't they...
	Put it in today. Put it in today.
	Put it in today. Every time I've heard Carpathie talk about the problem and talking about like software 2.0 and how the machine learning is gobbling up everything, I think this is absolutely the right strategy.
	I think that he didn't write and navigate on autopilot.
	I think somebody else did and kind of hacked it on top of that stuff.
	I think when Carpathie says, wait a second.
	Why did we hand code this lane change policy with all these magic numbers?
	We're going to learn it from data. They'll fix it.
	They already know what to do there. Well, that's Andre's job is to turn everything into a learning problem and collect a huge amount of data.
	The reality is though, not every problem can be turned into a learning problem in the short term.
	In the end, everything will be a learning problem.
	The reality is Like if you want to build L5 vehicles today, it will likely involve no learning.
	And that's the reality is, so at which point does learning start?
	It's the crutch statement that LiDAR is a crutch.
	At which point will learning get up to part of human performance?
	It's over human performance on ImageNet, classification, on driving is a question still.
	It is a question. I'll say this.
	I'm here to play for 10 years.
	I'm here to play for 10 years and make money along the way.
	I'm not here to try to promise people that I'm going to have my L5 taxi network up and working in two years.
	Do you think that was a mistake? Yes.
	What do you think was the motivation behind saying that?
	Other companies are also promising L5 vehicles with very different approaches in 2020, 2021, 2022.
	If anybody would like to bet me that those things do not pan out, I will bet you.
	Even money? Even money?
	I'll bet you as much as you want.
	So, are you worried about what's going to happen?
	Because you're not in full agreement on that.
	What's going to happen when 2022, 2021 come around and nobody has fleets of autonomous vehicles?
	Well, you can look at the history.
	If you go back five years ago, they were all promised by 2018 and 2017.
	But they weren't that strong of promises.
	I mean, Ford really declared pretty...
	I think not many have declared as definitively as they have now these dates.
	Well, okay, so let's separate L4 and L5. Do I think that it's possible for Waymo to continue to kind of like hack on their system until it gets to level four in Chandler, Arizona?
	Yes. No safety driver?
	Chandler, Arizona? Yeah.
	By, sorry, which year are we talking about?
	Oh, I even think that's possible by like 2020, 2021.
	But level four Chandler, Arizona, not level five New York City.
	Level four, meaning some very defined streets it works on really well.
	Very defined streets. And then practically, these streets are pretty empty.
	If most of the streets are covered in Waymos, Waymo can kind of change the definition of what driving is, right?
	If your self-driving network is the majority of cars in an area, they only need to be safe with respect to each other, and all the humans will need to learn to adapt to them.
	Now go drive in downtown New York.
	Already, you can talk about autonomy on farms.
	It already works great because you can really just follow the GPS line.
	What does success look like for Calm AI? What are the milestones where you can sit back with some champagne and say, we did it, boys and girls?
	Well, it's never over.
	Yeah, but you must drink champagne and everything.
	Sure. So what are some wins?
	A big milestone that we're hoping for by mid-next year is profitability of the company.
	And we're going to have to revisit the idea of selling a consumer product.
	But it's not going to be like the comma one.
	When we do it, it's going to be perfect.
	OpenPilot has gotten so much better in the last two years.
	We're going to have a few features.
	We're going to have 100% driver monitoring.
	We're going to disable no safety features in the car.
	Actually, I think it would be really cool what we're doing right now.
	Our project this week is we're analyzing the data set and looking for all the AEB triggers from the manufacturer systems.
	We have a better data set on that than the manufacturers.
	Does Toyota have 10 million miles of real-world driving to know how many times they're AEB triggered?
	So let me give you, because you asked, right, financial advice.
	Yeah. Because I work with a lot of automakers, and one possible source of money for you, which I'll be excited to see you take on, is basically selling the data.
	So, which is something that most people, and not selling in a way where here, here at Automaker, but creating, we've done this actually at MIT, not for money purposes, but you could do it for significant money purposes and make the world a better place by creating a consortia where Automakers would pay in and then they get to have free access to the data.
	And I think a lot of people are really hungry for that and would pay significant amount of money for it.
	Here's the problem with that. I like this idea all in theory.
	It'd be very easy for me to give them access to my servers.
	And we already have all open source tools to access this data.
	It's in a great format. We have a great pipeline.
	But they're going to put me in the room with some business development guy.
	And I'm going to have to talk to this guy.
	And he's not going to know most of the words I'm saying.
	I'm not willing to tolerate that.
	Okay, Mick Jagger.
	No, no, no, no. I think I agree with you.
	I'm the same way, but you just tell them the terms and there's no discussion needed.
	If I could just tell them the terms, then like, all right, who wants access to my data?
	I will sell it to you for, let's say, you want a subscription?
	I'll sell it to you for $100,000 a month.
	Anyone? 100k a month. 100k a month.
	I'll give you access to this data subscription.
	Yeah. Yeah, I think that's kind of fair.
	Came up with that number off the top of my head.
	If somebody sends me like a three-line email where it's like, we would like to pay 100k a month to get access to your data, we would agree to like reasonable privacy terms of the people who are in the data set, I would be happy to do it.
	But that's not going to be the email.
	The email is going to be, hey, do you have some time in the next month where we can sit down and we can, I don't have time for that.
	We're moving too fast. Yeah.
	You could politely respond to that email by not saying I don't have any time for your bullshit.
	You say, oh, well, unfortunately, these are the terms.
	And so we brought the cost down for you in order to minimize the friction, the communication.
	Here's the whatever it is, one, two million dollars a year, and you have access.
	And it's not like I get that email from like, but okay, am I going to reach out?
	Am I going to hire a business development person who's going to reach out to the automakers?
	No way. Yeah. Okay.
	I got you. If they reached into me, I'm not going to ignore the email.
	I'll come back with something like, yeah, if you're willing to pay a hundred K a month for access to the data, I'm happy to set that up.
	That's worth my engineering time.
	That's actually quite insightful of you.
	You're right. Probably because many of the automakers are quite a bit old school, there will be a need to reach out.
	And they want it, but there will need to be some communication.
	You're right. Mobileye, circa 2015, had the lowest R&D spend of any chipmaker.
	And you look at all the people who work for them, and it's all business development people because the car companies are impossible to work with.
	Yeah, so you have no patience for that, and you're a legit android, huh?
	I have something to do, right?
	It's not like I don't mean to be a dick and say I don't have patience for that, but it's like that stuff doesn't help us with our goal of winning self-driving cars.
	If I want money in the short term, If I showed off the actual learning tech that we have, it's somewhat sad.
	It's years and years ahead of everybody else's.
	Maybe not Tesla's. I think Tesla has similar stuff to us, actually.
	I think Tesla has similar stuff, but when you compare it to what the Toyota Research Institute has, you're not even close to what we have.
	No comments. But I also can't...
	I have to take your comments...
	I intuitively believe you, but I have to take it with a grain of salt.
	Because, I mean, you are an inspiration because you basically don't care about a lot of things that other companies care about.
	You don't try to bullshit, in a sense, like, make up stuff to drive a valuation.
	You're really very real and you're trying to solve the problem.
	I admire that a lot. What I don't necessarily fully can't trust you on without due respect is how good it is, right?
	I can only...
	But I also know how bad others are.
	And so... I'll say two things about...
	Don't trust but verify, right?
	I'll say two things about that. One is try...
	Get in a 2020 Corolla and try OpenPilot 0.6 when it comes out next month.
	Okay. I think already you'll look at this and you'll be like, this is already really good.
	And then I could be doing that all with hand labelers and all with the same approach that Mobileye uses.
	When we release a model that no longer has the lanes in it, that only outputs a path, then think about how we did that machine learning.
	And then right away, when you see, and that's going to be an open pilot, that's going to be an open pilot before 1.0.
	When you see that model, you'll know that everything I'm saying is true, because how else did I get that model?
	Good. You know, what I'm saying is true about the simulator.
	Yeah, yeah, yeah. That's super exciting.
	That's super exciting. But, like, you know, I listened to your talk with Kyle.
	And Kyle was originally building the aftermarket system.
	And he gave up on it because of technical challenges.
	Yeah. Of the fact that he's going to have to support 20 to 50 cars.
	We support 45. Because what is he going to do when the manufacturer ABS system triggers?
	We have alerts and warnings to deal with all of that and all the cars.
	And how is he going to formally verify it?
	Well, I got 10 million miles of data.
	It's probably better verified than the spec.
	Yeah. I'm glad you're here talking to me.
	I'll remember this day.
	Yeah. Because it's interesting.
	If you look at Kyle's from Cruise, I'm sure they have a large number of business development folks.
	And he's working with GM. He could work with Argo AI, working with Ford.
	It's interesting because chances that you fail, business-wise, like bankrupt, are pretty high.
	Yeah, I know. And yet, it's the Android model.
	You're actually taking on the problem.
	So that's really inspiring.
	Well, I have a long-term way for Comet to make money, too.
	And one of the nice things when you really take on the problem, which is my hope for Autopilot, for example, is things you don't expect.
	Ways to make money or create value that you don't expect will pop up.
	Oh, I've known how to do it since kind of...
	2017 is the first time I said it.
	Which part? To not do how to do which part?
	Our long-term plan is to be a car insurance company.
	Insurance. Yeah, I love it.
	Yep, yep. I make driving twice as safe.
	Not only that, I have the best data set to know who statistically is the safest drivers.
	And, oh, oh, we see you.
	We see you driving unsafely.
	We're not going to insure you. And that causes a, like, bifurcation in the market because the only people who can't get comma insurance are the bad drivers.
	Geico can insure them. Their premiums are crazy high.
	Our premiums are crazy low.
	We'll win car insurance. Take over that whole market.
	Okay, so... If we win.
	If we win. But that's I'm saying, like, how do you turn Kama into a $10 billion company?
	It's that. That's right. So you, Elon Musk, who else?
	Who else is thinking like this and working like this in your view?
	Who are the competitors? Are there people seriously, I don't think anyone that I'm aware of is seriously taking on lane keeping, you know, like where it's a huge business that turns eventually into full autonomy that then creates other businesses on top of it and so on, thinks insurance, thinks all kinds of ideas like that.
	Do you know anyone else thinking like this?
	Not really. That's interesting.
	I mean, my sense is everybody turns to that in like four or five years.
	Like Ford, once the autonomy doesn't fall through.
	But at this time...
	Elon's the iOS. By the way, he paved the way for all of us.
	I would not be doing comma AI today if it was not for those conversations with Elon.
	And if it were not for him saying like...
	I think he said like, well, obviously we're not going to use LiDAR.
	We use cameras. Humans use cameras.
	So what do you think about that? How important is LIDAR? Everybody else on L5 is using LIDAR. What are your thoughts on his provocative statement that LIDAR is a crutch?
	See, sometimes he'll say dumb things like the driver monitoring thing, but sometimes he'll say absolutely, completely, 100% obviously true things.
	Yeah. Of course LIDAR is a crutch.
	It's not even a good crutch.
	You're not even using it.
	Oh, they're using it for localization.
	Yeah. Yeah. Which isn't good in the first place.
	If you have to localize your car to centimeters in order to drive, Like, that's not driving.
	Yeah, they're not currently not doing much machine learning, I thought, but LiDAR data.
	Meaning, like, to help you in the task of, general task of perception.
	The main goal of those LiDARs on those cars, I think, is actually localization more than perception.
	Or at least that's what they use them for.
	Yeah, that's true. If you want to localize to centimeters, you can't use GPS. The fanciest GPS in the world can't do it.
	Especially if you're under tree cover and stuff.
	It's LiDAR, you can do this pretty easily.
	So you really, they're not taking on, I mean, in some research they're using it for perception, but, and they're certainly not, which is sad, they're not fusing it well with vision.
	They do use it for perception.
	I'm not saying they don't use it for perception, but the thing that, they have vision-based and radar-based perception systems as well.
	You could remove the LIDAR and keep around a lot of the dynamic object perception.
	You want to get centimeter accurate localization?
	Good luck doing that with anything else.
	So what should Cruz Waymo do?
	What would be your advice to them now?
	Oh, I don't know. I mean, Waymo is actually...
	I mean, they're doing...
	They're serious.
	Waymo, out of the ball of them, are quite...
	Serious about the long game, if L5 requires 50 years, I think Waymo will be the only one left standing at the end, given the financial backing that they have.
	They're buku Google bucks.
	I'll say nice things about both Waymo and Cruise.
	Let's do it. Nice is good.
	Waymo is by far the furthest along with technology.
	Waymo has a three to five year lead on all the competitors.
	If the Waymo looking stack works, maybe three-year lead.
	If the Waymo looking stack works, they have a three-year lead.
	Now, I argue that Waymo has spent too much money to recapitalize, to gain back their losses in those three years.
	Also, self-driving cars have no network effect like that.
	Yeah. Uber has a network effect.
	You have a market. You have drivers and you have riders.
	Self-driving cars, you have capital and you have riders.
	There's no network effect.
	If I want to blanket a new city in self-driving cars, I buy the off-the-shelf Chinese knockoff self-driving cars and I buy enough of them in the city.
	I can't do that with drivers.
	And that's why Uber has a first mover advantage that no self-driving car company will.
	Can you disentangle that a little bit?
	You're not talking about Uber, the autonomous vehicle.
	You're talking about the Uber car.
	I'm Uber. I open for business in Austin, Texas, let's say.
	I need to attract both sides of the market.
	I need to both get drivers on my platform and riders on my platform.
	And I need to keep them both sufficiently happy.
	Riders aren't going to use it if it takes more than five minutes for an Uber to show up.
	Drivers aren't going to use it if they have to sit around all day and there's no riders.
	So you have to carefully balance a market.
	And whenever you have to carefully balance a market, there's a great first mover advantage because there's a switching cost for everybody.
	The drivers and the riders would have to switch at the same time.
	Let's even say that, let's say, Luber shows up.
	And Luber somehow agrees to do things at a bigger, we've done it more efficiently.
	Luber only takes 5% of a cut instead of the 10% that Uber takes.
	No one is going to switch because the switching cost is higher than that 5%.
	So you actually can, in markets like that, you have a first mover advantage.
	Yeah. Autonomous vehicles of the Level 5 variety have no first mover advantage.
	If the technology becomes commoditized, say I want to go to a new city, look at the scooters.
	It's going to look a lot more like scooters.
	Every person with a checkbook can blanket a city in scooters, and that's why you have 10 different scooter companies.
	Which one's going to win? It's a race to the bottom.
	It's a terrible market to be in, because there's no market for scooters.
	Because the scooters don't get a say in whether they want to be bought and deployed to a city or not.
	We're going to entice the scooters with subsidies and deals.
	So whenever you have to invest that capital, it doesn't come back.
	That can't be your main criticism of the Waymo approach.
	Oh, I'm saying even if it does technically work.
	Even if it does technically work, that's a problem.
	I don't know if I were to say, I would say, you're already there.
	I haven't even thought about that, but I would say the bigger challenge is the technical approach.
	So Waymo's cruise is- And not just the technical approach, but of creating value.
	I still don't understand how you beat Uber The human driven cars.
	In terms of financially, it doesn't make sense to me that people want to get in an autonomous vehicle.
	I don't understand how you make money.
	In the long term, yes, like real long term.
	But it just feels like there's too much capital investment needed.
	Oh, and they're going to be worse than Ubers because they're going to stop for every little thing everywhere.
	I'll say a nice thing about Cruise.
	That was my nice thing about Waymo. They're three years ahead of it.
	They're three years technically ahead of everybody.
	Their tech stack is great.
	My nice thing about Cruise is GM buying them was a great move for GM. For $1 billion, GM bought an insurance policy against Waymo.
	They put Cruise is three years behind Waymo.
	That means Google will get a monopoly on the technology for at most three years.
	And if technology works, you might not even be right about the three years, it might be less.
	Might be less. Cruise actually might not be that far behind.
	I don't know how much Waymo has waffled around or how much of it actually is just that long tail.
	Yeah, okay. If that's the best you could say in terms of nice things, that's more of a nice thing for GM that that's a smart insurance policy.
	It's a smart insurance policy. I mean, I think that's how...
	I can't see Cruise working out any other...
	For Cruise to leapfrog Waymo would really surprise me.
	Yeah, so let's talk about the underlying assumption of everything.
	We're not going to leapfrog Tesla.
	Tesla would have to seriously mess up for us to be able to leapfrog them.
	Okay, so the way you leapfrog, right, is you come up with an idea or you take a direction, perhaps secretly, that the other people aren't taking.
	And so Cruise, Waymo, even Aurora...
	I don't know Aurora, Zoox is the same stack as well.
	They're all the same codebase even.
	They're all the same DARPA Urban Challenge codebase.
	So the question is, do you think there's a room for brilliance and innovation there that will change everything?
	Okay, so I'll give you examples.
	It could be if revolution in mapping, for example, that allow you to map things, do HD maps of the whole world, all weather conditions somehow really well, or revolution in simulation, to where what you said before becomes incorrect.
	That kind of thing. Any room for breakthrough innovation?
	What I said before about, oh, they actually get the whole thing.
	Well, I'll say this about, we divide driving into three problems.
	And I actually haven't solved the third yet, but I have an idea how to do it.
	So there's the static. The static driving problem is assuming you are the only car on the road.
	Right. And this problem can be solved 100% with mapping and localization.
	Yep. This is why farms work the way they do.
	If all you have to deal with is the static problem, and you can statically schedule your machines, right?
	It's the same as like statically scheduling processes.
	You can statically schedule your tractors to never hit each other on their paths, right?
	Because you know the speed they go at.
	So that's the static driving problem.
	Maps only helps you with the static driving problem.
	Yeah. The question about static driving, you've just made it sound like it's really easy.
	Static driving is really easy. How easy?
	Well, because the whole drifting out of lane.
	When Tesla drifts out of lane, it's failing on the fundamental static driving problem.
	Tesla is drifting out of lane.
	The static driving problem is not easy for the world.
	The static driving problem is easy for one route.
	One route and one weather condition with one state of lane markings and no deterioration, no cracks in the road.
	I'm assuming you have a perfect localizer.
	So that's all for the weather condition and the lane marking condition.
	But that's the problem. How do you have a perfect localizer?
	Perfect localizers are not that hard to build.
	Okay, come on now. With LiDAR?
	LiDAR, yeah. With LiDAR, okay.
	LiDAR, yeah. But you use LiDAR, right?
	Use LiDAR, build a perfect localizer.
	Building a perfect localizer without LiDAR...
	It's going to be hard.
	You can get 10 centimeters without LiDAR.
	You can get 1 centimeter with LiDAR.
	I'm not even concerned about the 1 or 10 centimeters.
	I'm concerned if every once in a while you're just way off.
	Yeah, so this is why you have to carefully make sure you're always tracking your position.
	You want to use LiDAR camera fusion, but you can get the reliability of that system up to...
	100,000 miles, and then you write some fallback condition where it's not that bad if you're way off, right?
	I think that you can get it to the point, it's like ASL-D, that you're never in a case where you're way off and you don't know it.
	Yeah. Okay, so this is brilliant.
	So that's the static. Static.
	We can, especially with LiDAR and good HD maps, you can solve that problem.
	Easy. Easy.
	The static problem is so easy.
	It's very typical for you to say something is easy.
	I got it. It's not as challenging as the other ones.
	Okay. Well, okay, maybe it's obvious how to solve it.
	The third one's the hardest. And a lot of people don't even think about the third one and even see it as different from the second one.
	So the second one is dynamic.
	The second one is like, say there's an obvious example, it's like a car stopped at a red light.
	You can't have that car in your map because you don't know whether that car is going to be there or not.
	So you have to detect that car in real time and then you have to do the appropriate action, right?
	Also, that car is not a fixed object.
	That car may move. And you have to predict what that car will do.
	So this is the dynamic problem.
	So you have to deal with this.
	This involves, again, you're going to need models of other people's behavior.
	Are you including in that?
	I don't want to step on the third one.
	Are you including in that your influence on people?
	Ah, that's the third one.
	That's the third one. We call it the counterfactual.
	Yeah, brilliant. I just talked to Judea Pearl who's obsessed with counterfactuals.
	Oh yeah, yeah, I read his books. So the static and the dynamic, our approach right now for lateral will scale completely to the static and dynamic.
	The counterfactual, the only way I have to do it yet, the thing that I want to do once we have all these cars is I want to do reinforcement learning on the world.
	I'm always going to turn the exploiter up to Max.
	I'm not going to have them explore.
	But the only real way to get at the counterfactual is to do reinforcement learning because the other agents are humans.
	So that's fascinating that you break it down like that.
	I agree completely. I've spent my life thinking about this problem.
	It's beautiful. And part of it, because you're slightly insane, because...
	Not my life.
	Just the last four years. No, no.
	Some non-zero percent of your brain has a madman in it.
	That's a really good feature.
	But there's a safety component to it that I think, sort of with counterfactuals and so on, that would just freak people out.
	How do you even start to think about, just in general, I mean, you've had some friction with NHTSA and so on.
	I am frankly exhausted by safety engineers.
	The prioritization on Safety over innovation to a degree where it kills, in my view, kills safety in the long term.
	So the counterfactual thing, they just actually exploring this world of how do you interact with dynamic objects and so on.
	How do you think about safety?
	You can do reinforcement learning without ever exploring.
	And I said that like, so you can think about your, in like reinforcement learning, it's usually called like a temperature parameter.
	And your temperature parameter is how often you deviate from the argmax.
	I could always set that to zero and still learn.
	And I feel that you'd always want that set to zero on your actual system.
	Gotcha. But the problem is you first don't know very much, and so you're going to make mistakes.
	So the learning, the exploration happens through mistakes.
	We're all ready, yeah, but okay.
	So, the consequences of a mistake.
	OpenPilot and Autopilot are making mistakes left and right.
	We have 700 daily active users, 1,000 weekly active users.
	OpenPilot makes tens of thousands of mistakes a week.
	These mistakes have zero consequences.
	These mistakes are, oh, I wanted to take this exit and it went straight, so I'm just going to carefully touch the wheel.
	The humans catch them. The humans catch them.
	And the human disengagement is labeling that reinforcement learning in a completely consequence-free way.
	So driver monitoring is the way you ensure they keep paying attention.
	How is your messaging?
	Say I gave you a billion dollars, you would be scaling it now.
	Oh, I couldn't scale it with any amount of money.
	I'd raise money if I had a way to scale it.
	Yeah, you're now not focused on scaling.
	I don't know how to do... Oh, I guess I could sell it to more people, but I want to make the system better.
	Better, better. And I don't know how to...
	But what's the messaging here?
	I got a chance to talk to Elon, and he basically said that the human factor doesn't matter.
	You know, the human doesn't matter because the system will perform, there'll be sort of a, sorry to use the term, but like a singular, like a point where it gets just much better.
	And so the human, it won't really matter.
	But it seems like that human catching the system when it gets into trouble is like the thing which will make something like reinforcement learning work.
	So how do you think messaging for Tesla, for you, for the industry in general should change?
	Our messaging is pretty clear.
	At least, like, our messaging wasn't that clear in the beginning, and I do kind of fault myself for that.
	We are proud right now to be a Level 2 system.
	We are proud to be level two.
	If we talk about level four, it's not with the current hardware.
	It's not going to be just a magical OTA upgrade.
	It's going to be new hardware.
	It's going to be very carefully thought out.
	Right now, we are proud to be level two.
	And we have a rigorous safety model.
	I mean, not like, okay, rigorous, who knows what that means.
	But we at least have a safety model.
	And we make it explicit. It's in safety.md in OpenPilot.
	And it says, seriously though, safety.md.
	Oh, this is brilliant. This is so Android.
	So, well, this is the safety model.
	And I like to have conversations like, you know, sometimes people will come to you and they're like, your system's not safe.
	Okay. Have you read my safety docs?
	Would you like to have an intelligent conversation about this?
	And the answer is always no. They just like scream about, it runs Python!
	Okay, what? So you're saying that because Python's not real-time?
	Python not being real-time never causes disengagements.
	Disengagements are caused by, you know, the model is QM. But safety.md says the following.
	First and foremost, the driver must be paying attention at all times.
	I still consider the software to be alpha software until we can actually enforce that statement, but I feel it's very well communicated to our users.
	Okay. Two more things.
	One is the user must be able to easily take control of the vehicle at all times.
	So if you step on the gas or brake with OpenPilot, it gives full manual control back to the user or press the cancel button.
	Step two, the car will never react so quickly, we define so quickly to be about one second, that you can't react in time.
	And we do this by enforcing torque limits, braking limits, and acceleration limits.
	So we have, like, our torque limits way lower than Tesla's.
	This is another potential...
	If I could tweak Autopilot, I would lower their torque limit and I would add driver monitoring.
	Because Autopilot can jerk the wheel hard.
	Yeah. OpenPilot can't.
	We limit, and all this code is open source, readable, and I believe now it's all MISRA-C compliant.
	What's that mean? MISRA is like the automotive coding standard.
	At first, I've come to respect, I've been reading the standards lately, and I've come to respect them.
	They're actually written by very smart people.
	Yeah. They're brilliant people, actually.
	They have a lot of experience. They're sometimes a little too cautious, but in this case, it pays off.
	MISRA is written by computer scientists, and you can tell by the language they use.
	You can tell by the language they use.
	They talk about whether certain conditions in MISRA are decidable or undecidable.
	You mean like the halting problem?
	Yes. All right.
	You've earned my respect.
	I will read carefully what you have to say, and we want to make our code compliant with that.
	All right. So you're proud level two.
	Beautiful. So you were the founder and I think CEO of Come AI. Then you were the head of research.
	What the heck are you now?
	What's your connection to Come AI? I'm the president, but I'm one of those like unelected presidents of like a small dictatorship country, not one of those like elected presidents.
	Oh, so you're like Putin when he was like the, yeah, I gotcha.
	So what's the governance structure?
	What's the future of Comm.ai as a business?
	Are you just focused on getting things right now, making some small amount of money in the meantime, and then when it works, it works and you scale?
	Our burn rate is about 200k a month and our revenue is about 100k a month.
	So we need to forex our revenue, but we haven't tried very hard at that yet.
	And the revenue is basically selling stuff online.
	Yeah. We sell stuff at shop.com.ai.
	Is there other... Well, okay.
	So you'll have to figure out revenues.
	That's our only... See, but to me, that's like respectable revenues.
	Yeah. We make it by selling products to consumers.
	We're honest and transparent about what they are.
	Most actually level four companies, right?
	Because you could easily start blowing up like smoke, like overselling the hype and feeding into getting some fundraisers.
	Oh, you're the guy. You're a genius because you hacked the iPhone.
	Oh, I hate that.
	I hate that. Yeah, I can trade my social capital for more money.
	I did it once.
	I almost regret it doing it the first time.
	Well, on a small tangent, you seem to not like fame, and yet you're also drawn to fame.
	Where are you on that currently?
	Have you had some introspection, some soul searching?
	Yeah. Actually, I've come to a pretty stable position on that after the first time.
	I realized that I don't want attention from the masses.
	I want attention from people who I respect.
	Who do you respect? I can give a list of people.
	So are these like Elon Musk type characters?
	Yeah. Actually, you know what?
	I'll make it more broad than that.
	I won't make it about a person.
	I respect skill.
	I respect people who have skills.
	And I would like to be, I'm not going to say famous, but be known among more people who have real skills.
	Who in cars do you think have skill, not do you respect?
	Oh, Kyle Vogt has skill.
	A lot of people at Waymo have skill.
	And I respect them.
	I respect them as engineers.
	I think about all the times in my life where I've been dead set on approaches and they turn out to be wrong.
	I might be wrong.
	I accept that. I accept that there's a decent chance that I'm wrong.
	And actually, I mean, having talked to Chris Hermsen, Sterling Anderson, those guys, I mean, I deeply respect Chris.
	I just admire the guy.
	He's legit. When you drive a car through the desert when everybody thinks it's impossible, that's legit.
	And then I also really respect the people who are, like, writing the infrastructure of the world, like the Linus Torvalds and the Chris Latt.
	Oh, yeah. They're doing the real work.
	I know. They're doing the real work.
	Having talked to Chris Ladner, you realize, especially when they're humble, it's like you realize, oh, you guys were just using all the hard work that you did.
	Yeah, that's incredible. What do you think, Mr.
	Anthony Lewandowski, he's another mad genius.
	Sharp guy. Oh, yeah. Do you think he might long-term become a competitor?
	Oh, Takama? Well, so I think that he has the other right approach.
	I think that right now, there's two right approaches.
	One is what we're doing, and one is what he's doing.
	Can you describe, I think it's called Pronto AI, he's starting new thing.
	Do you know what the approach is?
	I actually don't know. Embark is also doing the same sort of thing.
	The idea is almost that you want to...
	I can't partner with Honda and Toyota.
	Honda and Toyota are 400,000-person companies.
	It's not even a company at that point.
	I don't personify it.
	I think of it like an object. But a trucker drives for a fleet, maybe, that has...
	Some truckers are independent.
	Some truckers drive for fleets with 100 trucks.
	There are tons of independent trucking companies out there.
	Start a trucking company and drive your costs down.
	Or figure out how to drive down the cost of trucking.
	Another company that I really respect is Nauto.
	Actually, I respect their business model.
	Nauto sells a driver monitoring camera and they sell it to fleet owners.
	If I owned a fleet of cars and I could pay $40 a month to monitor my employees...
	It reduces accidents 18%.
	In the space, that is the business model that I most respect.
	Because they're creating value today.
	Yeah, that's a huge one.
	How do we create value today with some of this?
	And the lane keeping thing is huge.
	And it sounds like you're creeping in or full steam ahead on the driver monitoring too, which I think actually were the short-term value if you can get it right.
	I'm not a huge fan of the statement that everything is to have driver monitoring.
	I agree with that completely, but that statement usually misses the point that to get the experience of it right is not trivial.
	Oh, no, not at all. In fact, like, so right now we have, I think the timeout depends on speed of the car.
	But we want to depend on like the scene state.
	If you're on like an empty highway, it's very different if you don't pay attention than if like you're like coming up to a traffic light.
	And long term, it should probably learn from the driver because I watched a lot of video.
	We've built a smartphone detector just to analyze how people are using smartphones.
	And people are using it very differently.
	Interesting. So texting styles.
	We haven't watched nearly enough of the videos.
	We haven't. I got millions of miles of people driving cars.
	In this moment, I spend a large fraction of my time just watching videos because it never fails to learn.
	I've never failed from a video-watching session to learn something I didn't know before.
	In fact, I usually, like when I eat lunch, I'll sit, especially when the weather is good, and just watch pedestrians with an eye to understand, like from a computer vision eye, just to see, can this model, can you predict what are the decisions made?
	And there's so many things that we don't understand.
	This is what I mean about the state vector.
	Yeah, it's... I'm trying to always think, like, because I'm understanding in my human brain, how do we convert that into...
	How hard is the learning problem here, I guess, is the fundamental question.
	So, something that's...
	From a hacking perspective, this always comes up, especially with folks.
	Well, first, the most popular question is the trolley problem, right?
	So, that's not sort of a serious...
	Problem, there are some ethical questions, I think, that arise.
	Maybe you want to, do you think there's any ethical, serious ethical questions?
	We have a solution to the Charlie problem at ComAI.
	Well, so there is actually an alert in our code, ethical dilemma detected.
	It's not triggered yet. We don't know how yet to detect the ethical dilemmas, but we're a level two system, so we're going to disengage and leave that decision to the human.
	You're such a troll. No, but the trolley problem deserves to be trolled.
	Yeah. That's a beautiful answer, actually.
	I know. I gave it to someone who was like, sometimes people ask like you asked about the trolley problem.
	You can have a kind of discussion about it.
	But when you get someone who's really earnest about it, because it's the kind of thing where if you ask a bunch of people in an office whether we should use a SQL stack or a NoSQL stack, if they're not that technical, they have no opinion.
	But if you ask them what color they want to paint the office, everyone has an opinion on that.
	And that's why the trolley problem is.
	I mean, that's a beautiful answer.
	Yeah, we're able to detect the problem and we're able to pass it on to the human.
	Wow, I've never heard anyone say it.
	That's such a nice escape route.
	Okay, but... Proud level two.
	I'm proud level two.
	I love it. So the other thing that people have some concern about with AI in general is hacking.
	So, how hard is it, do you think, to hack an autonomous vehicle, either through physical access or through the more sort of popular now, these adversarial examples on the sensors?
	Okay, the adversarial examples one.
	You want to see some adversarial examples that affect humans?
	Hmm. Right?
	Oh, well, there used to be a stop sign here, but I put a black bag over the stop sign and then people ran it.
	Right. Adversarial.
	Yeah. Right? There's tons of human adversarial examples, too.
	The question in general about security, if you saw something just came out today, and there are always such hype-y headlines about how Navigate on Autopilot was fooled by a GPS spoof to take an exit.
	Yeah. At least that's all they could do was take an exit.
	If your car is relying on GPS in order to have a safe driving policy, you're doing something wrong.
	If you're relying... And this is why V2V is such a terrible idea.
	V2V now relies on both parties getting communication right.
	This is not even...
	Security is like a special case of safety.
	Safety is like we put a little piece of caution tape around the hole so that people won't walk into it by accident.
	Security is like put a 10-foot fence around the hole so you actually physically cannot climb into it with barbed wire on the top and stuff.
	If you're designing systems that are unreliable, they're definitely not secure.
	Your car should always do something safe using its local sensors.
	And then the local sensor should be hardwired.
	And then could somebody hack into your CAN bus and turn your steering wheel and your brakes?
	Yes, but they could do it before Common AI too.
	Let's think out of the box on some things.
	So do you think teleoperation has a role in any of this?
	So remotely stepping in and controlling the cars?
	No, I think that if the safety operation by design requires a constant link to the cars, I think it doesn't work.
	So that's the same argument you're using for V2I, V2V? Well, there's a lot of non-safety critical stuff you can do with V2i.
	I like V2i. I like V2i way more than V2i because V2i is already like, I already have internet in the car, right?
	There's a lot of great stuff you can do with V2i.
	Like, for example, you can, well, I already have V2i.
	Waze is V2i, right?
	Waze can route me around traffic jams.
	That's a great example of V2i.
	And then, okay, the car automatically talks to that same service.
	So it's improving the experience, but it's not a fundamental fallback for safety.
	No. If any of your things that require wireless communication are more than QM, like have an ASL rating, you shouldn't be.
	You previously said that life is work and that you don't do anything to relax.
	How do you think about hard work?
	What do you think it takes to accomplish great things?
	There's a lot of people saying that there needs to be some balance.
	In order to accomplish great things, you need to take some time off, you need to reflect and so on.
	And then some people are just insanely working, burning the candle on both ends.
	How do you think about that?
	I think I was trolling in the Siraj interview when I said that.
	Off camera, right before, I smoked a little bit of weed.
	Like, you know, come on. This is a joke, right?
	Like, I do nothing to relax. Look where I am.
	I'm at a party, right? That's true.
	So no, no, of course I don't.
	When I say that life is work though, I mean that like, I think that what gives my life meaning is work.
	I don't mean that every minute of the day you should be working.
	I actually think this is not the best way to maximize results.
	I think that if you're working 12 hours a day, you should be working smarter and not harder.
	Well, so work gives you meaning.
	For some people, other source of meaning is personal relationships, like family and so on.
	You've also, in that interview with Suraj, or the trolling, mentioned that one of the things you look forward to in the future is AI girlfriends.
	So that's a topic that I'm very much fascinated by, not necessarily girlfriends, but just forming a deep connection with AI. What kind of system do you imagine when you say AI girlfriend, whether you were trolling or not?
	No, that one I'm very serious about.
	And I'm serious about that on both a shallow level and a deep level.
	I think that VR brothels are coming soon and are going to be really cool.
	It's not cheating if it's a robot.
	I see the slogan already.
	I don't know if you've watched or just watched the Black Mirror episode.
	I watched the latest one, yeah. Yeah, yeah.
	Oh, the Ashley 2 one?
	No, where there's two friends having sex with each other.
	Oh, in the VR game.
	In the VR game. It's the two guys, but one of them was a female.
	Yeah. Which is another mind-blowing concept.
	In VR, you don't have to be the form.
	You can be two animals having sex.
	Yeah. It's weird.
	I mean, I'll see how nice that the software maps the nerve endings, right?
	Yeah, it's weird. I mean, yeah, they sweep a lot of the fascinating, really difficult technical challenges under the rock, like assuming it's possible to do the mapping of the nerve endings.
	I wish, yeah, I saw that the way they did it with the little like stim unit on the head, that'd be amazing.
	So, well, no, no.
	On a shallow level, like, you could set up, like, almost a brothel with, like, real dolls and Oculus quests.
	Write some good software.
	I think it'd be a cool novelty experience.
	But, no, on a deeper, like, emotional level, I mean...
	Yeah, I would really like to fall in love with a machine.
	Do you see yourself having a long-term relationship of the kind monogamous relationship that we have now with a robot, with an AI system even?
	Not even just a robot. So I think about maybe my ideal future.
	When I was 15, I read Eliezer Yudkowsky's early writings on the singularity and that AI is going to surpass human intelligence massively.
	He made some Moore's law-based predictions that I mostly agree with.
	And then I really struggled for the next couple years of my life.
	Like, why should I even bother to learn anything?
	It's all going to be meaningless when the machines show up.
	Right. Maybe when I was that young, I was still a little bit more pure and really clung to that.
	And then I'm like, well, the machines ain't here yet.
	And I seem to be pretty good at this stuff.
	Let's try my best.
	What's the worst that happens?
	But the best possible future I see is me sort of merging with the machine.
	And the way that I personify this is in a long-term monogamous relationship with a machine.
	Oh, you don't think there's a room for another human in your life if you really truly merge with another machine?
	I mean, I see merging.
	I see like the best interface to my brain is like the same relationship interface to merge with an AI, right?
	What does that merging feel like?
	I've seen couples who've been together for a long time and like I almost think of them as one person.
	Like, couples who spend all their time together.
	That's fascinating. You're actually putting...
	What does that merging actually look like?
	It's not just a nice channel.
	Like, a lot of people imagine it's just an efficient link, search link to Wikipedia or something.
	I don't believe in that. But it's more...
	You're saying that there's the same kind of relationship you have with another human.
	That's a deep relationship. That's what merging looks like.
	That's pretty... I don't believe that link is possible.
	I think that that link, so you're like, oh, I'm gonna download Wikipedia right to my brain.
	My reading speed is not limited by my eyes.
	My reading speed is limited by my inner processing loop.
	And to like bootstrap that sounds kind of unclear how to do it and horrifying.
	But if I am with somebody, and I'll use a somebody, Who is making a super sophisticated model of me and then running simulations on that model.
	I'm not going to get into the question whether the simulations are conscious or not.
	I don't really want to know what it's doing.
	But using those simulations to play out hypothetical futures for me, deciding what things to say to me to guide me along a path.
	That's how I envision it. So on that path to AI of superhuman level intelligence, you've mentioned that you believe in the singularity, that singularity is coming.
	Yeah. Again, could be trolling, could be not, could be part...
	All trolling has truth in it.
	I don't know what that means anymore.
	What is the singularity? Yeah.
	So that's really the question.
	How many years do you think before the singularity, what form do you think it will take?
	Does that mean fundamental shifts in capabilities of AI? Does it mean some other kind of ideas?
	Maybe it's just my roots.
	So I can buy a human being's worth of compute for like a million bucks a day.
	It's about one TPU pod v3.
	I think they claim 100 petaflops.
	That's being generous. I think humans are actually more like 20, so that's like five humans.
	That's pretty good. Google needs to sell their TPUs.
	But I could buy GPUs.
	I could buy a stack of like...
	I'd buy 1080 TIs.
	Build data center full of them.
	And for a million bucks, I can get a human worth of compute.
	But when you look at the total number of flops in the world...
	When you look at human flops, which goes up very, very slowly with the population, and machine flops, which goes up exponentially, but it's still nowhere near.
	I think that's the key thing to talk about when the singularity happened.
	When most flops in the world are silicon and not biological, that's kind of the crossing point.
	Like they're now the dominant species on the planet.
	And just looking at how technology is progressing, when do you think that could possibly happen?
	Do you think it would happen in your lifetime?
	Oh yeah, definitely in my lifetime.
	I've done the math. I like 2038 because it's the Unix timestamp rollover.
	Yeah, beautifully put.
	So you've said that the meaning of life is to win.
	If you look five years into the future, what does winning look like?
	So... There's a lot of...
	I can go into technical depth to what I mean by that, to win.
	It may not mean...
	I was criticized for that in the comments.
	Like, doesn't this guy want to save the penguins in Antarctica?
	Listen to what I'm saying.
	I'm not talking about I have a yacht or something.
	I am an agent. I am put into this world.
	And I don't really know what my purpose is.
	But if you're a reinforcement, if you're an intelligent agent and you're put into a world, what is the ideal thing to do?
	Well, the ideal thing, mathematically, you can go back to like Schmidt-Huber theories about this, is to build a compressive model of the world.
	To build a maximally compressive, to explore the world such that your exploration function maximizes the derivative of compression of the past.
	Schmidt-Huber has a paper about this.
	And like, I took that kind of as like a personal goal function.
	Yeah. So what I mean to win, I mean like, maybe this is religious, but I think that in the future I might be given a real purpose, or I may decide this purpose myself.
	And then at that point, now I know what the game is and I know how to win.
	I think right now I'm still just trying to figure out what the game is.
	But once I know, So you have imperfect information, you have a lot of uncertainty about the reward function, and you're discovering it.
	Exactly. That's a better way to put it.
	The purpose is to maximize it while you have a lot of uncertainty around it.
	And you're both reducing the uncertainty and maximizing at the same time.
	So that's at the technical level.
	If you believe in the universal prior, what is the universal reward function?
	That's the better way to put it.
	So that win, it's interesting.
	I think I speak for everyone in saying that I wonder what that reward function is for you, and I look forward to seeing that in five years and ten years.
	I think a lot of people, including myself, are cheering you on, man.
	So I'm happy you exist, and I wish you the best of luck.
	Thanks for talking today, man.

▲