Elon Musk: SpaceX, Mars, Tesla Autopilot, Self-Driving, Robotics, and AI | Lex Fridman Podcast #252

Lex Fridman 0:00 The following is a conversation with Elon Musk, his third time on this, the Lex Fridman podcast.

Lex Fridman 0:07 Yeah, make yourself comfortable.

Elon Musk 0:08 Boo.

Lex Fridman 0:08 Uh, no, wow, okay.

Elon Musk 0:10 (laughs)

Lex Fridman 0:10 No. (laughs)

Elon Musk 0:11 Do you, you don't do the headphone thing?

Lex Fridman 0:13 No.

Elon Musk 0:13 Okay.

Lex Fridman 0:13 No.

Elon Musk 0:14 I mean, how close do I get, need to get this thing to you?

Lex Fridman 0:15 The closer you are, the sexier you sound.

Elon Musk 0:17 Hey, babe. What's up?

Lex Fridman 0:18 Yeah.

Elon Musk 0:18 Can't get enough of your love, baby. (laughs)

Lex Fridman 0:21 (laughs) I'm gonna clip that out. Anytime somebody messages me about Elon, I'll just respond with that.

Elon Musk 0:27 If you want my body and you think I'm sexy, come right out and tell me so. Do, do, do, do, do.

Lex Fridman 0:36 (laughs) So good. So good. Okay.

Elon Musk 0:39 (laughs)

Lex Fridman 0:40 Serious mode activate. All right.

Elon Musk 0:43 Serious mode. Come on, you're Russian, you can be serious.

Lex Fridman 0:45 Yeah, I know, right?

Elon Musk 0:46 Everyone's serious all the time in Russia.

Lex Fridman 0:47 Yeah.

Elon Musk 0:48 (laughs)

Lex Fridman 0:48 Yeah, but we'll get there, we'll get there. It's taken America too long.

Elon Musk 0:52 Y- y- yeah.

Lex Fridman 0:52 It's gotten soft.

Lex Fridman 0:55 Allow me to say that the SpaceX launch of human beings to orbit on May 30th, 2020 was seen by many as the first step in a new era of human space exploration. These human space flight missions were a beacon of hope to me and to millions over the past two years as our world has been going through one of the most difficult periods in recent human history. We saw, we see the rise of division, fear, cynicism, and the loss of common humanity, right when it is needed most. So first, Elon, let me say thank you for giving the world hope and reason to be excited about the future.

Elon Musk 1:32 Oh, that's kind of you to say. I do want to do that. Humanity has, uh, obviously a lot of issues. And, and, uh, you know, people at times do, do bad things. But, you know, despite all that, um, you know, I, I love humanity and I think we should, uh, make sure we do everything we can to have a good future and, and an exciting future and one where, that maximizes the happiness of the people.

Lex Fridman 1:58 Let me ask about, uh, Crew Dragon Demo-2. So that, that first flight with humans on board, um, how did you feel leading up to that launch? Were you scared? Were you excited? What was going through your mind? So much was at stake.

Elon Musk 2:14 Yeah. No, that was extremely stressful. No question. Um, we obviously could not, um, let them down in any way. Um, so extremely stressful, I'd say, uh, to say the least. But we did... I was confident that, at the time that we launched, that no one could think of anything at all to do that would improve the probability of success. Um, and we, we racked our brains to think of any possible way to improve the probability of success. We could not think of anything more, and, and nor could NASA. And so then that, that's just the best that we could do, so then we, we had, we went ahead and launched. Now, I'm not a religious person. Um, but I nonetheless got on my knees and prayed for that mission.

Lex Fridman 3:03 Were you able to sleep?

Elon Musk 3:05 No. (laughs)

Lex Fridman 3:07 How did it feel when it was a success? First when the launch was a success, and when they returned back home, or back to Earth?

Elon Musk 3:16 It was a great relief. Yeah. It, it... For, for high-stress situations, I find it's, it's not so much elation as relief. Um, and, um, you know, I think once... A- a- as we got more comfortable and proved out the systems, 'cause, you know, we really, um... You know, you've got to make sure everything works. Um, I was, it was definitely a lot more enjoyable with the subsequent, uh, astronaut tra- missions. And I thought the, the Inspiration mission was actually very inspiring, um, the Inspiration4 mission. Um, I'd e- I'd encourage people to watch the Inspiration documentary on Netflix. It's actually really good. Um, and it really is inspi- I was actually inspired by that. Um, and I, I, I... So that one I felt I was kind of able to enjoy the actual mission and not just be super stressed all the time.

Lex Fridman 4:11 So for people that somehow don't know, it's the all civilian, first time all civilian out to space, out to orbit.

Elon Musk 4:19 Yeah, and it was the h- I think the highest orbit that, uh, in like, I don't know, 30 or 40 years or something. The only one that was higher was the s- one shuttle, uh, sorry, Hubble, uh, Servicing Mission. Um, and then before that, it would have been, um, Apollo in '72. It's pretty wild. So it's cool, it's c- Y- you know, I think, uh, as, you know, as a species, like we want to be, you know, continuing to do better and, and reach higher ground. And, and, like, I think it would be tragic, extremely tragic if, um, Apollo was the high water mark for humanity, you know? And that, and that's as far as we ever got. Um, and it's, um, it's concerning that here we are, um, 49 years after the last mission to the moon, and so almost half a century, uh, and we've not been back. Um, and that's worrying. It's like, is that... Does that mean we've peaked as a civilization or what? So, like I think we gotta get back to the moon and build a base there, you know, a science base. I think we could learn a lot about the nature of the universe if we have a proper science base on the moon. Um, you know, like we have a science base in Antarctica and, you know, many other parts of the world. And, um, so that, that's like I think the n- the next big thing. We've gotta have like a, a serious like moon base, um, and then get people to Mars and, you know, get, get out there and be a spacefaring civilization.

Lex Fridman 5:52 I'll ask you about some of those details, but since you're so busy with the hard engineering challenges of everything that's involved, are s- are you still able to marvel at the magic of it all, of space travel, of every time the rocket goes up, especially when it's a crewed mission?... or are you just so overwhelmed with the, all the challenges that you have to solve? And actually, sort of, to add to that, the reason I, I wanted to ask this question of May 30th, it's, it's been some time, so you can look back and think about the impact already. It's already, at the time, it was an engineering problem, maybe. Now, it's becoming a historic moment. Like, it's a moment that ... How many moments will be remembered about the 21st century? To me, wha- that, or something like that, maybe Inspiration4 or one of those, will be remembered as the early steps of a new age of, uh, space exploration.

Elon Musk 6:44 Yeah. Uh, I mean, d- during the launches itself ... So, I mean, the, I think, I think maybe s- some people know, but a lot of people don't know is, like, uh, I'm actually the chief engineer of SpaceX. So, um, the, you know, I've signed off on pretty much all the design decisions. Um, and, you know, so if, if there's something that goes wrong with that, uh, vehicle, it's, it's fundamentally my fault, you know? So, um, so I'm really just thinking about all the things that ... Like, so, so when I see the rocket, I see all the things that could go wrong and the things that could be better, and the same with the Dragon spacecraft. It's, uh, like other people see, "Oh, this is a, a spacecraft or a rocket and that's, this looks really cool." I'm like ...

Lex Fridman 7:28 (laughs)

Elon Musk 7:28 I've, I've, like, a readout of, like, this is the, these are, these are the risks, these are the pro- problems. That's what I see. Like ... (laughs) So it's not what other people see when they see the product,

Elon Musk 7:40 you know?

Lex Fridman 7:40 So, let me, uh, ask you then to analyze Starship in that same way. Uh, I know you have, you'll talk about, in more detail about Starship in the near future, perhaps. You had that-

Elon Musk 7:50 Yeah, we can talk about it now if you want.

Lex Fridman 7:51 Um, but, just in that same way, like you said you see when you see a, uh, when you see a rocket, you see a sort of a list of risks. In that same way, you said that Starship was a really hard problem. So, there's many ways I can ask this, but if you magically could solve one problem perfectly, one engineering problem perfectly, which one would it be?

Elon Musk 8:13 On Starship?

Lex Fridman 8:14 On, on, sorry, on Starship. So is it maybe related to the efficiency, the, uh, the engine, the weight of the different components, the complexity of various things, maybe the controls of the, the crazy thing it has to do to land?

Elon Musk 8:26 No, it's (laughs) actually, the, by far the, the biggest thing absorbing my time is, uh, uh, engine production.

Lex Fridman 8:34 Hm.

Elon Musk 8:35 Not, not the design of the engine, but ... (laughs) I, I co- I- I've often said prototypes are, are easy, production is hard. Um, so we have the most advanced rocket engine that's ever been designed. Um, the ... 'Cause I'd say currently the, the, the best rocket engine ever is probably the RD-180 w- or RD-170, the, um ... That, that's the Russian engine, basically. Um, and, um, and still, it ... I think an engine should only count if it's gotten something to orbit. Um ...

Lex Fridman 9:09 (laughs)

Elon Musk 9:09 So our engine has not gotten anything to orbit yet. Um, but it is, it's the first engine that's actually better than, than the, the, the Russian RD engines, which are, were a amazing design.

Lex Fridman 9:22 So you're talking about Raptor engine. What makes it amazing? What, what are the different aspects of it that make it ... Like ...

Elon Musk 9:30 Yeah.

Lex Fridman 9:30 ... what are you the most excited about, uh, if the whole thing works, in terms of, uh, efficiency, all those kinds of things?

Elon Musk 9:37 Well, it's ... The Raptor is a, a full-flow, uh, staged combustion, um, engine, and it's a- a- at operating at a very high, uh, chamber pressure. So one of the key figures of merit, or p- perhaps the key f- key figure of merit, is, um, what is the chamber pressure at which the rocket engine can operate? That's the combustion chamber pressure. Um, so Raptor is, uh, designed to operate at, uh, 300 bar, possibly maybe higher. That's 300 atmospheres. So, um, the record right now for operational engine is the RD engine that I mentioned, the Russian RD, which is, I believe, around 267, uh, bar. Um, and the, the, the, the difficulty of the chamber pressure is increases on a non-linear basis. So 10% more chamber pressure is more like, uh, 50% (laughs) more difficult. Um, but that, that chamber pressure is, that, that, that is what allows you to get a very high, uh, power density for, uh, for the engine. Um, so, uh, enabling, um, a, a very high, uh, thrust-to-weight ratio, um, and, um, a very high specific impulse. So specific impulse is like a measure of the efficiency of a rocket engine, or, um ... It's, it's really the, the, the, the, uh, exhaust, the, the effective exhaust velocity of, of the gas coming outta the engine. Um, so, uh, with, with a, with a very high chamber pressure, you can have, um, a, a, a compact engine that nonetheless has a high expansion ratio, which is the ratio between the, uh, um, exit nozzle, uh, and the, uh, throat. So y- you know, engine's got ... Like, you see a rocket engine's got like, sort of like a ba- like a hourglass shape. It's like a chamber and then it necks down, and then there's a nozzle. And the ratio of the, the, the exit diameter to the, the throat, uh, is the expansion ratio.

Lex Fridman 11:46 So why is it such a hard engine to manufacture at scale?

Elon Musk 11:51 Uh, it's very complex. Um-

Lex Fridman 11:53 So a lot of com ... What does complexity mean here? Is it a lot of components involved?

Elon Musk 11:56 There's a lot of, a lot of components and a lot of, uh, unique materials that, uh ... So we had to invent a ho- um, several alloys that don't exist in order to make this engine work. Um, and-

Lex Fridman 12:11 So it's a materials problem too.

Elon Musk 12:14 So, a materials problem, and, um, in- in a- in a staged combustion, a full-flow staged combustion, there- there are many, uh, feedback loops in the system. So, y- uh, basically, you've- you've got, uh, propellant and- and- and, uh, hot gas flowing, (laughs) um, to, simultaneously to so many different places on the engine. Um, and, uh, they- they all have a recursive effect on each other. So, you change one thing here, it has a recursive effect here, changes something over there. And- and it's- it's- it's- it's quite hard to control. Um, there, like there's a reason no one's made this before. Um, but, (sighs) um ... And the reason we're doing, um, a staged combustion, uh, full flow is- is because it- it has the highest poten- the highest, uh, theoretical possible, uh, efficiency. Um, so in- in- in order to make a fully reusable rocket, um, which tha- that's the really, the holy grail, uh, of orbital rocketry. Um, you have to have ... E- everything's gotta be the best. Uh, it's gotta be the best engine, the best airframe, the best heat shield, um, extremely light, uh, avionics. Um, ver- you know, very clever control mechanisms. Um, you've got to shed mass i- i- in- in any possible way that you can. Um, for example, we are, instead of putting landing legs on the booster and ship, we are going to catch them with a tower to save the weight of the landing leds- legs. So, that's like, I mean, we're talking about catching the largest flying object ever made, uh, with, on a giant tower with- with chopstick arms. It's like Karate Kid with the fly-

Lex Fridman 14:08 Yeah.

Elon Musk 14:08 ... but much bigger. (laughs)

Lex Fridman 14:12 I mean, pulling something like-

Elon Musk 14:13 This probably won't work the first time. (laughs) Uh, and anyway, so this is bananas. This is banana stuff.

Lex Fridman 14:19 So, you mentioned that you doubt ... Well, not you doubt, but there- there's days or moments when you doubt that this is even possible, it's so difficult.

Elon Musk 14:30 The possible part is ... Well, at- at this point, (clears throat) we'll, I think we will- we will get Starship to work. Um, um, there's a question of timing. How long will it take us to do this? Uh, how long will it take us to actually achieve, uh, full and rapid reusability? Um, 'cause it will take probably many launches before we are able to have full and rapid reusability. Um, but I can say that- that the physics pencils out. Like, the- like, we're not, uh ... Like, at this point, I'd say we- we're confident that w- that se- like, let's say we- I'm- I'm very confident sec- success is in the set of all possible outcomes. (laughs)

Lex Fridman 15:14 Hmm. All right. It's not a null set of- (laughs)

Elon Musk 15:16 For- for- for a while there, I was not convinced that success was in the set of possible outcomes. (laughs)

Lex Fridman 15:21 (laughs)

Elon Musk 15:21 So, which is very important actually. But so we're, um, um ... And- and but- but it-

Lex Fridman 15:28 So, you're saying there's a chance.

Elon Musk 15:29 I'm saying there's a chance, exactly.

Lex Fridman 15:31 Okay.

Elon Musk 15:31 Um, uh, just not sure how- how- how long it will take. Uh, but we have a very- a very talented team, they're working night and day to make it happen. Um, and, uh, li- like I said, the- the- the critical thing to achieve for the revolution in space flight and for humanity to- to be a space-faring civilization is to have a fully and rapidly reusable rocket, orbital rocket. Um, there's not even been any orbital rocket that's been fully reusable ever. And this has always been the- the- the- the holy grail of rocketry. Um, and, uh, many smart people, very smart people, uh, have tried to do this before and they've not succeeded, so, um, 'cause it's s- such a hard problem.

Lex Fridman 16:17 What's your source of belief in situations like this? When the engineering problem is so difficult, there's a lot of experts, many of whom you admire who have failed in the past.

Elon Musk 16:29 Yes.

Lex Fridman 16:30 And, um, a lot of people, you know, the, a lot of experts, maybe journalists, all the kind of, you know, the public in general have a lot of doubt about whether it's possible. And you yourself know that, uh, even if it's a non-null set, non-empty set of success, it's still unlikely or d- very difficult! Like, where do you go to, both personally, um, intellectually, as an engineer, as a team, like, for source of strength needed to sort of persevere through this? And to, uh, keep going with the project, take it to completion?

Elon Musk 17:18 A source of strength. Hmm. I- I, doesn't really not how I think about things. Um, I mean, for me, it's simply this- this is something that is important to get done. Um, and we- we should just keep doing it, um, or die trying. And I- I don't need a source of strength.

Lex Fridman 17:36 So, quitting is not even, like, um ...

Elon Musk 17:39 That's not in- it's not in my nature.

Lex Fridman 17:40 Okay.

Elon Musk 17:41 And I- I don't care about optimism or pessimism. Fuck that, we're gonna get it done.

Lex Fridman 17:47 Gonna get it done. (sighs)

Lex Fridman 17:51 Can you, uh, then zoom back in to specific problems with Starship, or any engineering problems you work on? Can you try to introspect y- your particular biological neural network, your thinking process, and describe how you think through problems, through different engineering and design problems? Is there like a systematic process? You've spoken about first principles thinking-

Elon Musk 18:12 Yeah.

Lex Fridman 18:12 ... but is there a kinda-

Elon Musk 18:12 Absolutely.

Lex Fridman 18:13 ... process to it?

Elon Musk 18:14 Well, um ...You know, I like saying, like, like, "Physics is the law, and everything else is a recommendation."

Lex Fridman 18:21 Mm-hmm.

Elon Musk 18:21 Um, like, I've met a lot of people who can break the law, but, uh, I haven't ha- met anyone who could break physics. (laughs) So, uh, so the first for- you know, any kind of technology problem, you have to s- sort of just make sure you're not violating physics. Um, and, you know, uh, first principles analysis I think is something that can be applied to really any walk of life, uh, any- anything, really. It's ju- it's- it's really just saying, um, you know, "Let's- let's boil something down to the most fundamental, uh, principles, the things that we are most confident are true at a foundational level." And that sets your a- your, sets your axiomatic base, and then you reason up from there, and then you crosscheck your conclusion against the- the axiomatic truths. Um, so, um, you know, some basics in physics would be like, "Are you violating conservation of energy or momentum?" Or something like that, you know? Then you- you're- it's not gonna work. (laughs) Um, so, uh, that's th- you know. So, that- that's just to establish, is it- is it- is it possible... And then a- a- another good physics tool is thinking about things in the limit. If you, if you take a particular thing and you, uh, sc- scale it to a very large number or to a very small number h- how does, how do things change? Um...

Lex Fridman 19:46 Both, like, tempor- like, in number of things you manufacture, something like that, and then i- in time?

Elon Musk 19:51 Yeah, like, let's take, say, take an example of, like, um, l- like manufacturing, which I think is, uh, just a very underrated problem. Um, and, and, uh, like you said, it's- it's much harder to take a- a- a- a advanced technology product and bring it into volume manufacturing than it is to design it in the first place. My words

Narrator 20:13 (laughs)

Elon Musk 20:14 So, um, so let's say you're trying to figure out, is, um, like, why is this, this, uh, part or product, uh, expensive? Is it, um, because of something fundamentally foolish that we're doing, or is it because our volume is too low? And so then you say, "Okay, well, what if our volume was a million units a year? Is it still expensive?" That's what I mean by thinking about things in the limit. If it's still expensive at a million units a year, then volume is not the reason why your thing is expensive. There's something fundamental about the design.

Lex Fridman 20:44 And then you then can focus on the com- reducing complexity or something like that, and then desig-

Elon Musk 20:48 You could change the design to, change the- change the part to be something that is, uh, uh, not fundamentally expensive. But- but it- it, like, that's a common thing in rocketry, 'cause the- the unit volume is- is relatively low, and so a common excuse would be, "Well, it's expensive because our unit volume is low." Um, and if we were in, like, automotive or something like that, or consumer electronics, then our cost would be lower. I'm like, I'm like, "Okay, so let's say we scale... N- now you're making a million units a year. Is it still expensive?" If the answer is yes, then, uh, economies of scale are not the issue.

Lex Fridman 21:22 Do you throw... Into manufacturing, do you throw, like, supply chain? You talked about resources and materials and stuff like that. Do you throw that into the calculation of trying to reason from first principles, like, "How are we gonna make the supply chain work here?"

Elon Musk 21:34 Yeah, yeah.

Lex Fridman 21:35 And then the cost of materials, things like that? Or is that too much...

Elon Musk 21:38 Yeah. Uh, e- exactly. So, um, like, a- another, like, a good example of thinking about things, uh, in the limit is, um, if you take any, uh, you know, any- any product, any machine or whatever, um... Like, take a rocket or whatever, uh, and say, um, if you've got... If- if you look at the r- raw materials in the rocket, um, so you're gonna have, like, uh, I don't know, aluminum, steel, titanium, Inconel, uh, special, uh, specialty alloys, um, copper. And- and you say, "What are the, how... What- what- what's the weight of the constituent elements, o- of each of these elements, and what is their raw material value?" And that sets the asymptotic limit for how, uh, low the cost of the vehicle can be, unless you change the- the materials. So, and then when you do that, uh, call it, like, maybe the magic wand number or something like that, so that would be like if you had the, you know, uh, uh, uh, like, just a- a pile of these raw materials here and you could wave a magic wand and rearrange the atoms into the final shape. Um, that would be the lowest possible cost that you could make this thing for, unless you change the materials. So then... And that is always a- a, usually, almost always a- a very low number. Um, so then the- the... What's actually causing things to be expensive is how you put the atoms into the desired shape.

Lex Fridman 23:06 Yeah, I actually... If you don't mind me taking a tiny tangent. I had a... I often talk to Jim Keller, who's somebody who worked with you as a, as a friend.

Elon Musk 23:13 Oh, yeah. Yeah. J- Jim was, uh, yeah. Did great work at Tesla.

Lex Fridman 23:17 So, um, I suppose he carries the flame of the same kind of thinking...

Elon Musk 23:22 Mm-hmm.

Lex Fridman 23:22 ... that you're- you're talking about now. Um, and I- I guess I see that same thing at- at Tesla and- and, uh, SpaceX. Folks who work there, they kind of learn this way of thinking, and it kinda becomes obvious, almost. But, anyway, I had a argument... Not argument. Uh, he educated me about how cheap it might be to manufacture a Tesla bot. We just, we had an argument. "What is, how can you reduce the cost of scale of producing a robot?" Because I'd gotten the chance to interact quite a bit, um, obviously in- in the academic circles with humanoid robots, and then Boston Dynamics...

Elon Musk 24:00 Sure.

Lex Fridman 24:00 ... and stuff like that. And they- they're very expensive to- to build.

Elon Musk 24:04 Mm-hmm.

Lex Fridman 24:04 And then, uh, Jim kind of schooled me on saying like, "Okay, like, this kind of first principles thinking of, 'How can we get the cost of manufacture down?'" Um, I suppose you do that, you have done, uh, that kind of thinking for Tesla bot and for all kinds of, all kinds of complex... Systems that are traditionally seen as complex, and you say, "Okay, how can we simplify everything down?"

Elon Musk 24:27 Yeah. I mean, I think if you're- if you are really good at manufacturing, you can basically make, at high volume, you can basically make anything for a cost that asymptotically approaches the raw ma- raw material value of the constituents, plus any i- intellectual property that you need to license.

Lex Fridman 24:46 (laughs) Right.

Elon Musk 24:46 Anything.

Lex Fridman 24:47 Right.

Elon Musk 24:49 But it's hard. It's not like-

Lex Fridman 24:50 Yeah.

Elon Musk 24:50 ... that's a very hard thing to do, but- but it is possible for anything. Anything in volume can be made, uh, like I said, for a cost that asymptotically approaches its raw material, uh, constituents, plus intellectual property license rights. So what will often happen in trying to design a product is- is people will start with the tools and- and parts and methods that they are familiar with, um, and then, and- and try to create a product using their existing tools and methods. Um, the other way to think about it is, uh, actually imagine the, try to imagine the platonic ideal of the perfect product or technology, whatever it might be, um, and say, "What is the s- what is- what is the perfect arrangement of atoms that would be the- the best possible product? And now, let us try to figure out how to get the atoms in that shape."

Lex Fridman 25:42 (laughs) I mean, it s- it sounds, um, uh, it's almost like Rick and Morty absurd until you start to really think about it, and it, you really should think about it in this way 'cause-

Elon Musk 25:56 Yeah.

Lex Fridman 25:56 ... everything else is kind of, uh, uh, if you- if you think, uh, you- you might fall victim to the momentum of the way things were done in the past, unless you think in this way.

Elon Musk 26:06 Well, just as a function of inertia, people will, uh, want to use the same tools and methods that they are familiar with. Um, they just, that's what they'll do by default.

Lex Fridman 26:15 Yeah.

Elon Musk 26:15 Um, and then that- that will lead to an outcome of things that can be made with those tools and methods but is unlikely to be the, um, platonic ideal of the perfect product. Um, so then, so that's why it- it's good to think of things in both directions like, what can we build with the tools that we have? But then, but- but also what is the, what is the perfect, the theoretical perfect product look like? And- and that- that theoretical perfect product is gonna be a moving target 'cause as you learn more, the definition of, or- or for that perfect product will- will change 'cause you don't actually know what the perfect product is, but you can successfully approximate, uh, a- a- a more perfect product. Um, so think about it like that, and then saying, "Okay now, what tools, methods, materials, whatever do we need to create in order to get the atoms in that shape?" But- but- but people var- rar- rarely think about it that way, but it's a powerful tool.

Lex Fridman 27:12 I should mention that the brilliant Shivon Zilis is hanging (laughs) , hanging out with us, uh, in case you hear a voice of, uh, wisdom from, uh, from- from outside, from up above.

Lex Fridman 27:25 Uh, okay, so let me ask you about Mars. You mentioned it would be great for science to put, um, a base on the moon to do some research, but the truly big leap, again, in this category of seemingly impossible, is to put a human being on Mars. When do you think SpaceX will land a human being on Mars?

Elon Musk 27:50 Hmm. Best case is about five years. Worst case, 10 years.

Lex Fridman 28:17 What are the determining factors, would you say, from an engineering perspective? Or is that, that not the bottlenecks?

Elon Musk 28:24 Uh, no, it's- it's fundamentally, um, you know, engineering the- the vehicle. Um, I mean, Starship is the most compl- complex and advanced rocket that's ever been made by, I don't know, order of magnitude or something like that. It's a lot. It's really next level. So, um, and the fundamental optimization of Starship is minimizing cost per ton to orbit, and ultimately cost per ton to the surface of Mars. Um, this may seem like a mercantile objective, but it is actually the thing that needs to be optimized. Um, like there is a certain cost per ton to the surface of Mars where we can afford to establish a self-sustaining, uh, city, um, and the- and- and then above that, we cannot afford to do it. Um, so right- right now, you couldn't fly to Mars for a trillion dollars. There's no amount of money could get you a ticket to Mars. So, we need to get that above, uh, you know, to get that like something that is actually possible at all. Um, um, but- but then, but that's- that's- w- we don't- we don't just want to have, you know, with Mars, flags and footprints and then not come back for a half century like we did with the moon. Uh, in- in order to pass a very important grade filter, I think we- we need to be a multi-planet species. Um, this may sound somewhat esoteric to- to a lot of people, but, uh, like eventually, given enough time, uh, there's something, Earth is likely to experience some calamity, um, that could be, uh, something that humans do to themselves or an external event like happened to the dinosaurs. Um, and, um, but- but, uh, you know, eventually, and- and if- if nothing- if none of that happens, and somehow magically we- we keep going, uh, then the- the sun will ex- the sun is gradually expanding, um, and will engulf the Earth, um, and probably Earth gets too hot for, uh, life in, uh-... about 500 million years. It's a long time, but that's only 10% longer than Earth has been around. And so if you think about like the, the current situation is really remarkable, um, and kind of hard to believe. But, uh, Earth's been around for four and a half billion years, and this is the first time in four and a half billion years that it's been possible to extend life beyond Earth. And that window of opportunity may be open for a long time, and I hope it is, but it also may be open for a short time. And we should, uh, I think it was wise for us to, uh, act quickly while the window is open, just in case it, it closes.

Lex Fridman 31:13 Yeah, the existence of nuclear weapons, pandemics, all kinds of threats-

Elon Musk 31:19 Yeah.

Lex Fridman 31:19 ... should, uh, should kind of, um, give us some motivation.

Elon Musk 31:25 I mean, civilization could get, um, could die with a bang or a whimper. You know, if it's a, if it dies of demographic collapse, then it's more of a whimper, (laughs) obviously. Um, but, and if it's World War III, it's more of a bang. Um, but, but these are all risks. Um, I mean, it's important to think of these things and just, you know, think of things as like probabilities, not certainties. Um, there's a certain probability that something bad will happen on, on Earth. I, like, I think most likely the future will be good. Um, but there's like, let's say for argument's sake, um, a one percent chance per century of, of a civilization-ending event. Like that was Stephen Hawking's estimate. Um, I think he's, he might be right about that. Uh, so then, uh, you know, we, we should basically think of this like being a multi-planetary species is like taking out insurance for life itself. Like, life insurance for life. (laughs)

Lex Fridman 32:24 (laughs)

Elon Musk 32:25 Um-

Lex Fridman 32:26 Wow.

Elon Musk 32:26 ... so-

Lex Fridman 32:27 This turned into an infomercial real quick.

Elon Musk 32:29 Life insurance for life, yes. (laughs)

Lex Fridman 32:31 (laughs)

Elon Musk 32:31 Um, and, you know, uh, we, we can bring the, the, the creatures from, uh, you know, plants and animals from Earth to Mars and breathe life into the planet. Um, and, and have a second planet with, with life. Um, that would be great. Um, they can't bring themselves there, you know, so if we don't bring them to Mars, then they will just, for sure all die when the sun expands anyway, and then

Elon Musk 32:55 that'll be it.

Lex Fridman 32:56 What do you think is the most difficult aspect of building a civilization on Mars? Terraforming Mars, like from an engineering perspective, from a financial perspective, human perspective, to get, to get a large number of folks there who will never return back to Earth?

Elon Musk 33:15 Uh, no, they could certainly return. Some will return back to Earth.

Lex Fridman 33:18 They will choose to stay there-

Elon Musk 33:19 Yeah.

Lex Fridman 33:19 ... for the rest of their lives.

Elon Musk 33:21 Yeah, many will. Um, but, uh, we, you know, we, (laughs) we need the spaceships back, like the ones that go to Mars-

Lex Fridman 33:29 Yeah.

Elon Musk 33:29 We need them back, so-

Lex Fridman 33:30 Yeah.

Elon Musk 33:30 ... you can hop on if you want, you know?

Lex Fridman 33:31 Yeah.

Elon Musk 33:32 It's like, but we can't just not have the spaceships come back with, those things are expensive. We need them back, like to come back and do another trip.

Lex Fridman 33:38 I mean, do you think about the terraforming aspect, like actually building? Are you so focused right now on the spaceships part that's so critical-

Elon Musk 33:45 Yeah, yeah.

Lex Fridman 33:45 ... to get to Mars?

Elon Musk 33:45 It's just, we absolutely, if you can't get there, nothing else matters. So, and, and like I said, you, we can't get there with, at some extraordinarily high cost. I mean, the current cost of, um, let's say one ton to the surface of Mars is on the order of a billion dollars. So, 'cause you don't just need the rocket and the launch and everything. You need like heat shield. You need, you know, guidance system. You need, uh, deep space communications. Uh, you need some kind of landing system. So, like rough approximation would be a billion dollars per ton to the surface of Mars right now. Um, this is obviously, um, way too expensive to create a self-sustaining civilization. Um, so we need to improve that by at least a factor of a thousand.

Lex Fridman 34:36 A million per ton?

Elon Musk 34:38 Yes. Ideally less than, much less than a million per ton, but if it's not, like it's gotta be... Say, you have to say like, "Well, how much can society afford to spend or want to, want to spend on a self-sustaining city on Mars?" The self-sustaining part is important. Like it's just, the key threshold, um, the great filter will, will have been passed when the city on Mars can survive even if the spaceships from Earth stop coming for any reason. Doesn't matter what the reason is, but if they stop coming for any reason, will it die out or will it not? And if there's even one critical ingredient missing, then it still doesn't count. It's like, you know, if you're on a long sea voyage and you've got everything except vitamin C. (laughs) And it's only a matter of time, you know you're gonna die. (laughs) So, so we gotta get Mars, a Mars city to the point where it's self-sustaining. Um, I'm not sure this will really happen in my lifetime, but I, I hope to see it at least have a lot of momentum, and, and then you could say, "Okay, what is the minimum tonnage necessary to have a self-sustaining city?" Um, and there's a lot of uncertainty about this. You could say like, I don't know, it's probably at least a million tons, um, 'cause you have to set up a lot of infrastructure on, on Mars. Um, like I said, you can't be missing any, anything that, in order to be self-sustaining, you can't be missing, like you need, you know, semiconductor fabs. You need iron ore refineries. Like you need lots of things, you know? Uh, so, um, and Mars is not super hospitable. It's, it's the least inhospitable planet, but it's definitely a fixer-upper of a planet.

Lex Fridman 36:18 Outside of Earth.

Elon Musk 36:19 Yes. (laughs)

Lex Fridman 36:20 Earth is pretty good.

Elon Musk 36:20 Earth is like easy. (laughs) Yeah.

Lex Fridman 36:22 And also, I should, we should clarify, in the solar system.

Elon Musk 36:25 Yes, in the solar system.

Lex Fridman 36:26 There might be nice, like vacation spots.

Elon Musk 36:29 There might be some great planets out there, but, uh, it's hopeless.

Lex Fridman 36:32 Too hard to get there?

Elon Musk 36:33 Yeah, way, way, way, way too hard, (laughs) to say the least.

Lex Fridman 36:37 Let me push back on that. Not really a pushback, but quick, uh, curveball

Lex Fridman 36:41 of a question. So, you did mention physics as the first starting point. So, um...... general relativity allows for wormholes. Uh, they technically can exist. Do you think, um, those can ever be leveraged by humans to travel faster than the speed of light?

Elon Musk 36:59 Well... (sighs)

Lex Fridman 37:01 Are you saying

Elon Musk 37:01 That the wormhole thing is, is debatable. Uh, the, the, the, we currently do not know of any means of going faster than the speed of light. Um, the, there is like, like (sighs) te- (sighs) there, there are some ideas about having space like so, so... (sighs) You can only move at the speed of light through, through space. But if you can make space itself move that, that, that's like that, that's w- warping space. Um, space is, is capable of moving faster than the speed of light. (laughs)

Lex Fridman 37:39 (laughs) Right.

Elon Musk 37:40 Uh, like the universe in the Big Bang, the univ- the universe expanded at much gr- much more than the speed of light, by a lot.

Lex Fridman 37:46 Yeah.

Elon Musk 37:47 Um, so, um, but the... If this is possible, the, the amount of energy required to warp space is so gigantic, it's boggles the mind.

Lex Fridman 38:03 So all the work you've done with propulsion, how much innovation is possible with rocket propulsion? Is this, um... I mean, you've seen it all and you're constantly innovating in every aspect. How much is possible? Like how much can you get 10x somehow? Is there something in there in physics that you can get significant improvement in terms of efficiency of engines and all those kinds of things?

Elon Musk 38:24 Well, as I was saying, like the, the, really the holy grail is a, a fully and rapidly reusable orbital system. Um, so, uh, right now, uh, the Falcon 9 is the only reusable rocket out there. Uh, but it, but the, the booster comes back and lands. I'm sure you've seen the videos. Uh, and we get the nose conal fairing back, but we do not get the upper stage back. So, uh, that means that we have a minimum cost of, of, uh, uh, building an upper stage. Um, you can think of like a two-stage rocket of, of sort of like two airplanes, like a big airplane and a smaller airplane. Um, and we get the big airplane back, but not the smaller airplane. And so it still costs a lot, you know? So that upper stage is, you know, at least $10 million. Um, and then the degree of the, the, the booster is not as reus- it's not as rapidly and completely reusable as we'd like and nor are the fairings. So, you know, our, our kind of minimum marginal cost in our county overhead for per flight is on the order of $15 to $20 million maybe. Um, so, uh, that's, that's extremely good for... (sighs) It's by far better than any rocket ever in history. Um, but, uh, with full and rapid reusability, we can reduce the cost per ton to orbit by, uh, a factor of 100. Just think of it like, um, like imagine if you had an aircraft or something or a, a car. Oh, yeah, um, and if you had to buy a new car every time you went for a drive, that would be very expensive. It'd be silly, frankly.

Lex Fridman 40:07 Mm-hmm.

Elon Musk 40:08 But, um, but you, in fact, you just refuel the car or recharge the car and that's, uh, makes your trip, uh, p- like, (laughs) I don't know, a thousand times cheaper. (laughs) So it's the same for rockets. Uh, if you l- it's, uh, very difficult to make this complex machine that can go to orbit. And so if you cannot reuse it and have to, have to throw even any part of, any significant part of it away, that massively increases the cost. So, you know, Starship in theory could do a cost per launch of like a million, maybe $2 million or something like that. Um, and, uh, and put over 100 tons in orbit, which is crazy.

Lex Fridman 40:53 Yeah.

Elon Musk 40:54 So...

Lex Fridman 40:55 That's incredible. So you're saying like it's, uh, by far the biggest bang for the buck is to make it fully reusable versus like some kind of brilliant breakthrough in theoretical physics?

Elon Musk 41:05 No, no. There's no, there's no brilliant break- no, there's no... (laughs) It just me- you wanna make the rocket reusable.

Lex Fridman 41:10 Yeah.

Elon Musk 41:11 This is, this is an extremely difficult engineering problem.

Lex Fridman 41:13 Got it.

Elon Musk 41:13 Uh, but no, no new physics is required.

Lex Fridman 41:18 (sighs) Just brilliant engineering.

Lex Fridman 41:19 Let me ask a slightly philosophical fun question. Gotta ask. I know you're focused on getting to Mars, but once we're there on Mars, what do you... What form of government, economic system, political system do you think would work best for an early civilization of humans? Is- do- I mean, the in- the interesting reason to talk about this stuff, it also ma- helps people dream about the future. I know you're really focused about the short-term engineering dream, but it's like, I don't know, there's something about imagining an actual civilization on Mars that gives people-

Elon Musk 41:53 Sure.

Lex Fridman 41:54 ... really gives people hope.

Elon Musk 41:55 Well, it would, it would be a new frontier and an opportunity to rethink the whole nature of government, uh, just as was done in the creation of the United States. So, uh, I mean, I would suggest, um, having, uh, direct democracy, like people vote directly on things as opposed to representative democracy. So, uh, representative democracy I think is too, uh, subject to a special interest and c- you know, a coercion of the politicians and that kind of thing. Um, so I, I'd recommend, uh, that, that there's just, um, direct democracy. People vote on laws, the population votes on laws themselves, and then the laws must be short enough that people can understand them. (laughs)

Lex Fridman 42:46 Yeah, and then like keeping a well-informed populace, like really being transparent about all the information about what they're voting for.

Elon Musk 42:53 Yeah, absolute transparency.

Lex Fridman 42:54 Yeah. And not make it as annoying as those cookies where you have to accept the cook-

Elon Musk 42:58 (laughs) Accept cookies. I always like, uh, you know, there's like always like a slight amount of trepidation when you click accept cookies. Like, I feel as though there's like perhaps like a, like a very tiny chance that'll open a portal to hell, or something like that. (laughs)

Lex Fridman 43:12 (laughs) That's exactly how I feel.

Elon Musk 43:13 It's like, "Why? Why do they, why, why do they keep, why do we need to accept the, what do they want with this cookie?" Uh, like, somebody got upset with accepting cookies or something somewhere. I mean, who cares? Like, so annoying to kee- keep accepting all these cookies.

Lex Fridman 43:26 To me, this is just a-

Elon Musk 43:28 (laughs)

Lex Fridman 43:28 ... a great-

Elon Musk 43:28 I'll gladly accept, uh, yes, you can have my damn cookie. I don't care. Whatever. (laughs)

Lex Fridman 43:32 You heard it from Elon first. He accepts all your damn cookies.

Elon Musk 43:36 Yeah. (laughs)

Lex Fridman 43:37 (laughs)

Elon Musk 43:37 And stop asking me. (laughs)

Lex Fridman 43:40 Uh-

Elon Musk 43:40 It's annoying.

Lex Fridman 43:41 Yeah, it's, uh, it's one example of, um, im- implementation of a good idea done really horribly.

Elon Musk 43:50 Yeah. It's, it's somebody who was like, there's some good intentions of, like, privacy or whatever, but now everyone just has to take, accept cookies and it's not, you know, you have billions of people who have to keep clicking accept cookie and it's super annoying. Then we, just accept the damn cookie, it's fine. There, there is like, um, I think a fundamental problem that we're, be- because we've not really had a, uh, a, a major, uh, like a world war or something like that in a while, and obviously we'd, we'd like to not have world wars, um, the, there's not been a cleansing function for rules and regulations. Um, so w- wars did have, uh, you know, some sort of lining in that there would be a, a reset on rules and regulations, uh, after a war. Uh, so World Wars I and II, there were huge resets on rules and regulations. Um, now, as, if what, if the soci- society does not have a war, the, the, and there's no cleansing function or garbage collection for rules and regulations, then rules and regulations will accumulate every year, 'cause they are immortal. There's no actual, humans die, but the laws don't. Uh, so the, we, we need a garbage collection function for rules and regulations. They should not just be immortal, um, 'cause some of the rules and regulations that are put in place will be counterproductive. Uh, done with good intentions, but counterproductive. Sometimes not done with good intentions. So, um, if you just, if rules and regulations just accumulate every year, um, and you get more and more of them, then eventually you won't be able to do anything. You're just like Gulliver with, you know, tied down by thousands of little strings. And we ha- we see that in, um, you know, US and like, like basically all, all, all economies that, uh, have been around for, for a while, uh, and, and regulators and legislators create new rules and regulations every year, but they don't put effort into removing them, and I think that's very important that we put effort into removing rules and regulations. Um, but it gets tough, 'cause you get special interests that then are dependent on, like they, they have a, you know, a, uh, a vested interest in that whatever rule and regulation, and that they, then they fight to not get it removed. Um-

Lex Fridman 45:57 Yeah, so it, I mean, I guess the problem with the Constitution is it's, it's kinda like C versus Java, 'cause it doesn't have any garbage collection built in. I think there should be, I, I, when you first said the, the, the metaphor of garbage collection, I loved it.

Elon Musk 46:11 Yeah, it's from a coding standpoint.

Lex Fridman 46:12 From a coding standpoint, yeah, yeah. I, it would be intere- interesting if the laws themselves kinda had a built in thing where they kinda die after a while-

Elon Musk 46:20 Yeah.

Lex Fridman 46:20 ... and somebody explicitly publicly defends them.

Elon Musk 46:23 Mm-hmm, yeah.

Lex Fridman 46:23 So that, that's sort of, it's not like somebody has to kill them, they kinda die themselves. They disappear.

Elon Musk 46:29 Yeah. Um-

Lex Fridman 46:32 Not to defend Java or anything, but, you know. The, the C++, you know, you could also have gar- great garbage collection in Python and so on.

Elon Musk 46:40 Yeah. So yeah, something's go- something needs to happen or, or just the, the civilization's arteries, arteries just harden over time, and, and, uh, you can just get less and less done because there's just r- a rule against everything. Um, so, so I think like, I don't know, for Mars or whatever, I'd say, or even for, you know, obviously for Earth as well, like I think there should be an active process for removing, uh, rules and regulations and questioning their existence. Just, um, like if, if we've got a function for creating rules and regulations, 'cause rules and regulations you can also think of as like, they're like software or lines of code for operating, uh, civilization. That's rules and regulations. Um, so it's not like we shouldn't have rules and regulations, but the, you have, you have code accumulation but no code removal. Um, and so it just gets to be- become ar- basically archaic bloatware after a while, um, and, and it's just, it, it makes it hard for things to progress. So I don't know, maybe Mars you'd have like an, uh, uh, you know, any given law must have a sunset, you know, and, and, and, uh, and re- require active voting to keep, restore, to keep it up there, you know? Um, and I actually also say like, and these are just, I don't know, recommendations or thoughts, um, ultimately will be up to the people on Mars to decide, but I, I think, um, it should be easier to remove a law than to add one because of the, just to overcome the inertia of laws. So maybe it's like, uh, for argument's sake you need like say 60% vote to have a law take effect, but only a f- 40% vote to remove it.

Lex Fridman 48:23 So let me be the guy, you, you posted a meme on Twitter recently where there's th- there's, there's like a, a row of urinals, a guy just walks all the way across-

Elon Musk 48:32 Oh, sure yeah.

Lex Fridman 48:33 ... and he tells you about crypto.

Elon Musk 48:34 (laughs)

Lex Fridman 48:35 So, so, so-

Elon Musk 48:35 This, this, I mean, that's happened res- so many times, I think maybe even literally. Uh (laughs) -

Lex Fridman 48:40 Yeah. Do you think, technologically speaking, there's any room for ideas of smart contracts or so on, 'cause you mentioned laws. Um, that's an interesting implement- use of things like smart contracts to implement the laws by which governments function.... like, something built on Ethereum or maybe a dog coin that enables smart contracts somehow.

Elon Musk 49:04 I never, I don't quite understand this whole smart contract thing, um, you know. I, I mean ... (laughs)

Lex Fridman 49:11 (laughs) So it's a-

Elon Musk 49:12 I'm too dumb to understand smart contracts.

Lex Fridman 49:14 (laughs) That's a good line. (laughs)

Elon Musk 49:17 (laughs) I mean, my general approach to any kind of, like, deal or whatever is just make sure there's clarity of understanding. That's the most important thing.

Lex Fridman 49:25 Right.

Elon Musk 49:25 Um, and, and just keep any kind of deal very, very short and simple, plain language, um, and just make sure everyone understands this is the deal. Does everyone... Is it clear? Um, and, uh, and, and what are the consequences if various things don't happen? Um, but u- usually deal, deals are, um, you know, business deals or whatever, are way too long and complex and overly lawyered, and pointlessly.

Lex Fridman 49:52 You mentioned that, uh, Doge is the, the people's coin.

Elon Musk 49:57 Yeah.

Lex Fridman 49:57 Um, and you said that you were literally going, SpaceX may consider literally putting, uh, a Dogecoin on the moon.

Elon Musk 50:07 Yeah.

Lex Fridman 50:07 Is it, is this something you're still considering? Uh, Mars perhaps? Uh, do you think there's some chance... We've talked about political systems on Mars, that, uh, Dogecoin is the, the official currency of Mars at some point in the future?

Elon Musk 50:22 (laughs) Well, I, I think Mars itself will need to have a different currency because you can't synchronize due to speed of light.

Lex Fridman 50:31 Hmm.

Elon Musk 50:31 Or not easily. Um...

Lex Fridman 50:32 So it must be completely standalone from Earth?

Elon Musk 50:36 Well, yeah, 'cause the, the... Mars is r- At closest approach, it's four light minutes away roughly. And then at furthest approach, uh, it's roughly 20 light minutes away. Um, maybe a little more. Um, so you can't really have, uh, something synchronizing, you know, if you've g- if the, if you've got a 20 minute speed of light issue. If it's got a one minute blockchain, uh, it's, it's not gonna synchronize properly. Um, so Mars would n- would... I, I don't know if Mars would have a cryptocurrency as a thing, but probably. Seems likely. Um, but it would be some kind of localized, uh, thing on Mars. Um...

Lex Fridman 51:12 And you let the people decide?

Elon Musk 51:14 Yeah. (laughs) Absolutely.

Lex Fridman 51:16 (laughs)

Elon Musk 51:17 It's... The f- the future of Mars should be up to the Martians. Uh, but yeah, so, um,

Elon Musk 51:24 I mean, I think the cryptocurrency thing is an interesting approach to reducing the, um, error in the, the database that is called money. Um, you know, I think I have a pretty deep understanding of the, of what money actually is on a practical day-to-day basis because of PayPal. Um, you know, we really got in deep there. Um, and right, right now the money system, (laughs) actually for practical purposes, is, is, is really a bunch of, uh, heterogeneous, uh, mainframes running, uh, old COBOL.

Lex Fridman 52:06 (laughs) Okay, you mean literally? That's-

Elon Musk 52:09 Literally.

Lex Fridman 52:09 That's literally what's happening?

Elon Musk 52:10 In batch mode. Okay.

Lex Fridman 52:13 (laughs) In batch mode.

Elon Musk 52:14 Yeah. Uh, uh, pity the poor bastards who have to maintain that code, okay? That's a, that was a d- pain dr- that's pain.

Lex Fridman 52:22 Not even Fortran, it's COBOL, yep.

Elon Musk 52:24 It's COBOL. Like... And they still, the ba- banks were still buying mainframes in 2021 and running ancient COBOL code. Uh, and, uh, you know, the, the Federal Reserve is, like, probably even older than the, what the banks have, and they have an old COBOL mainframe. (laughs) And so now the... And, and so the, the government effectively has editing privileges on the, on the money database. Um, and they use those editing privileges to, um, make more money (laughs) whenever they want. And this increases the error in the database that is money. So if... I think money should really be viewed through the lens of, uh, information theory. And, uh, and so it's, uh, yeah, kind of like, uh, like an internet connection. Like, what's the bandwidth? Uh, you know, uh, to- total bit rate. Uh, what is the latency, jitter, uh, packet drop? Uh, you know, errors in, errors in network, uh, communication. So, just think of money like that, basically. Um, I think that's probably the right way to think of it. And, and then say what, what system, uh, from an information theory standpoint, allows an economy to function the best. Uh, and, you know, um, crypto is an attempt to reduce the, the error, uh, in, uh, in, in money that is contributed by, uh, governments, uh, d- diluting the money supply, uh, as basically a pernicious f- pernicious form of taxation.

Lex Fridman 53:58 So, both policy in terms of with inflation and actual, like, technological COBOL, like, cryptocurrency takes us into the 21st century in terms of the actual systems that allow you to do the transactions, to store wealth, all those kinds of things.

Elon Musk 54:16 Like I said-

Lex Fridman 54:17 Okay.

Elon Musk 54:17 ... just think of money as information. People, um, uh, often will think of money as having power in and of itself. Um, it does not. Money is, uh, is information. And it, it does not have power in and of itself. Uh, the, like... The, you know, again, applying the, the physics tools of thinking about things in the limit is helpful. If you are stranded on a tropical island, um, and, uh, you have a trillion dollars, it's useless, 'cause there's no, there's no resource allocation. M- money is a database for resource allocation. But there's no resources to allocate except yourself, so money is useless. Um... (sighs) Uh, if you're stranded on a desert island with no food, you'd, uh, all the Bitcoin in the world will not stop you from starving.

Lex Fridman 55:11 Yeah.

Elon Musk 55:12 So, um, so like, just, just think of money as, as a database for resource allocation, um, across time and space. And, um, and then what, what, what system, uh, i- is, what, what, in what form should that, that database or data system... What, what would be most effective? Now, the- there's a, there is a fundamental issue with, um, say, Bitcoin in its current form, uh, in that it's, the transaction volume is very limited. Um, and, uh, the latency, it's- the latency for, for a properly confirmed transaction is too, is too long, much longer than you'd like. So, it's not, it's actually not great from, um, a transaction volume standpoint or a latency standpoint. Um, uh, so it is perhaps useful as, as to s- to solve an aspect of the money database problem, uh, whi- which is a sort of store of wealth or an ac- an accounting of relative obligations, I suppose. Um, but it is not useful as a currency, as a day-to-day currency.

Lex Fridman 56:30 But people have proposed different technological solutions-

Elon Musk 56:33 Like Lightning.

Lex Fridman 56:34 ... yeah, Lightning Network and the Layer 2 technologies on top of that. I mean, it's, it's all, it seems to be all kind of a trade-off, but the point is, it's kinda brilliant to say that just think about information, think about what kind of database, what kind of infrastructure enables that exchange of information.

Elon Musk 56:48 Yeah, just say like if you're operating an economy, um, and you need to have something that a- allows for the efficient, to, to have efficient, uh, value ratios between products and services. So you've got this massive number of products and services, and you need to... You can't just bar- barter. (laughs) It's just like, that would be extremely unwieldy. Uh, so, you need something that gives you the, the, a, a ratio of exchange between goods and services. Um, and, and then something that allows you to, uh, shift obligations across time, like debt. Debt and equity, shift obligations across time. Then what does, what, what does the best job of that? Um,

Elon Musk 57:33 part of the reason why I think there's some, um, merit to Dogecoin, even though it was obviously created as a joke, um, is that it, it actually does have a much higher, uh, transaction volume capability than Bitcoin, um, and the, you know, the co- the cost of doing a transaction, the, the, the Dogecoin fee is, is very low. Like right now, if you wanted to do a Bitcoin transaction, the price of doing that transaction is very high, so you could not use it effectively for most things. Um, and nor could it even scale to a high volume. Um, uh, and when Bitcoin was, you know, started, I guess, what, around 2008 or something like that, um, the internet connections were much worse than they are today, like order of magnitude. I mean, there's no... Way, way worse, you know, in 2008. So, so like having a, you know, a small, uh, block size or whatever is, you know, and a long synchronization time is... Made sense in 2008, but to, you know, 2021 or fast-forward 10 years, it's like, it's like comically low, you know? It's, uh... So, um, and I think there's some value to having a linear increase in the amount of currency that, uh, is generated. Um, so because some amount of the currency... Like, like if, if a cur- if a currency is too deflationary, or like, uh, or I should say, if, if, if a cur- if a currency is expected to increase in value over time, there's reluctance to spend it. 'Cause you're like, "Oh, I, if I, I'll just hold it and not spend it because its scarcity is increasing with time, so if I spend it now, then I will regret spending it, so I will just, you know, hodl it." (laughs)

Lex Fridman 59:26 Mm-hmm.

Elon Musk 59:26 Um, but if there's some dilution of the currency occurring over time, that's, that's more of an incentive to use it as a currency. So, um, Dogecoin, somewhat randomly, has, uh, a, um, just a, a fixed number of, of sort of coins or hash strings that, uh, are generated every year. So there's, there's some inflation, but it's not a percentage base, it's a, it's... So the, the... It's a fixed number, so the percentage of inflation will necessarily decline over time. Um, so it, it just... I, I'm not saying that it's like the ideal system for a currency, but I think it actually is, uh, just fundamentally better than anything else I've seen. Just by accident. Um, so...

Lex Fridman 1:00:16 I like how you said, um, around 2008. So you're not, uh, you know, some people suggested you might be Satoshi Nakamoto, you previously said you're not. Let me ask-

Elon Musk 1:00:27 No, I'm not.

Lex Fridman 1:00:27 You're not, for sure.

Elon Musk 1:00:28 No, 100%.

Lex Fridman 1:00:28 Would, would you tell us if you were?

Elon Musk 1:00:30 Yes.

Lex Fridman 1:00:30 Okay. (laughs) Uh, do you think it's a feature or bug that he's anonymous, or she, or they? It's an interesting kind of quirk of human history, that there is a particular technology that is a completely anonymous inventor, or creator.

Elon Musk 1:01:03 Well, I mean, you can- you can look at the, (sighs) um, evolution of ideas, um, before the launch of Bitcoin and see who wrote, uh, y- you know, about those ideas. Um, and then, uh, like, I don't know exact- obviously, I don't know who- who created Bi- Bitcoin for practical purposes, but y- the evolution of ideas is- is pretty clear before that. And, like, it seems as though, like, Nick Zavo, uh, is probably, more than anyone else, uh, responsible for the evolution of those ideas. So, you know, he claims not to be Sak- Sa- sa- sa- Nakamoto, but I'm not sure that's- that's neither here nor there. Uh, but he- he seems to be the one more responsible for the ideas behind Bitcoin than anyone else.

Lex Fridman 1:01:51 So, it's not perhaps, like, singular figures aren't even as important as the- the figures involved in the evolution of ideas that led to a thing, so...

Elon Musk 1:01:58 Yeah.

Lex Fridman 1:01:59 Yeah. It's, you know, and most ...

Elon Musk 1:02:01 (clears throat)

Lex Fridman 1:02:01 P- perhaps it's sad to think about history, but maybe most names will be forgotten anyway.

Elon Musk 1:02:06 What is a name anyway? It's a name- a name attached to an idea. What does it even mean really?

Lex Fridman 1:02:13 I think Shakespeare had a thing about roses and stuff, whatever he said.

Elon Musk 1:02:17 (laughs) A rose by any other name would smell as sweet.

Lex Fridman 1:02:21 (laughs) I got Elon to quote Shakespeare. I feel- I feel like I accomplished something today.

Elon Musk 1:02:26 Shall I compare thee to a summer's day? (laughs)

Lex Fridman 1:02:30 (laughs)

Elon Musk 1:02:30 Blah.

Lex Fridman 1:02:30 I'm gonna clip that out.

Elon Musk 1:02:31 (laughs)

Lex Fridman 1:02:31 Instead of the whole thing. Um-

Elon Musk 1:02:34 None more temperate or more fair. (laughs)

Narrator 1:02:38 Being balding.

Lex Fridman 1:02:38 Uh, Autopilot. Tesla Autopilot.

Elon Musk 1:02:42 (laughs)

Lex Fridman 1:02:44 (laughs) Um, Tesla Autopilot has been through an incredible journey over the past six years, um, or perhaps even longer in the minds of- in your mind and the minds of many involved. Uh-

Elon Musk 1:02:57 Yeah, I think that's where we first, like, connected really was the Autopilot stuff.

Lex Fridman 1:03:00 Yeah.

Elon Musk 1:03:00 Uh, autonomy and-

Lex Fridman 1:03:01 Well, it's, the whole journey was incredible to me to watch. I was, um ... because I knew, well, part of it is I was at MIT, and I- I knew the difficulty of computer vision.

Elon Musk 1:03:12 Yeah.

Lex Fridman 1:03:13 And I knew the whole, I had a lot of colleagues and friends, about the DARPA challenge, I knew how difficult it is.

Elon Musk 1:03:18 Mm-hmm.

Lex Fridman 1:03:18 And so there was a natural skepticism. When I first drove a Tesla with, uh, the initial system based on Mobileye-

Elon Musk 1:03:24 Yeah.

Lex Fridman 1:03:25 ... I thought, there's no way th- so, at first, when I- when I got in, I thought, there's no way this car could maintain, um, like stay in the lane and create a comfortable experience. So, my intuition initially was that the lane keeping problem is way too difficult to solve. And-

Elon Musk 1:03:41 Oh, lane keeping? Yeah, that's relatively easy.

Lex Fridman 1:03:43 No, I, well ...

Elon Musk 1:03:44 Yeah.

Lex Fridman 1:03:44 But, like, uh, but not the s- but solve in the way that we just, we talked about previous is prototype versus a thing that actually creates a pleasant experience over hundreds of thousands of miles and millions.

Elon Musk 1:03:57 Sure.

Lex Fridman 1:03:57 I, yeah, so-

Elon Musk 1:03:58 We- we- we-

Lex Fridman 1:03:58 ... I was proven wrong pretty quickly.

Elon Musk 1:03:59 ... we had to wrap a lot of code around the- the MobileEye thing.

Lex Fridman 1:04:01 Yes.

Elon Musk 1:04:02 It's- it doesn't just, didn't just work by itself. (laughs)

Lex Fridman 1:04:03 Yes. I mean, th- that's part, that's part of the story of how you approach things sometimes.

Elon Musk 1:04:07 Yeah.

Lex Fridman 1:04:07 Sometimes you do things from scratch, sometimes at first you kind of see what's out there, and then you decide to do it from scratch. That was one of the boldest decisions I've seen is both on the hardware and the software to decide to eventually go from scratch. I thought, again, I was skeptical whether that's going to be able to work out-

Elon Musk 1:04:24 Sure.

Lex Fridman 1:04:24 ... 'cause it's such a, such a difficult problem. And so it was an incredible journey. What I see now with, um, everything, the hardware, the compute, the sensors, the, uh, the things I maybe care and love about most is the, the stuff that Andrej Karpathy is leading with the dataset selection, the whole data engine process, the neural network architectures, the- the way that's in the real world that network is tested, validated, all the different test sets. Uh, you know, versus the ImageNet model of computer vision, like what's in academia is like real-world artificial intelligence.

Elon Musk 1:04:59 Yeah.

Lex Fridman 1:05:00 So, and-

Elon Musk 1:05:01 Um, and A- Andrej is awesome, and obviously plays an important role, but we have a- a lot of really talented people driving things.

Lex Fridman 1:05:07 Yes.

Elon Musk 1:05:07 So, um, and, uh, Ash- Ashok is actually the- the head of Autopilot engineering. Um, uh, uh, Andrej is sort of director of AI.

Lex Fridman 1:05:16 AI stuff, yeah, yeah, yeah.

Elon Musk 1:05:17 Yeah.

Lex Fridman 1:05:17 So yeah, there's, I'm aware that there's an incredible team of just a lot going on.

Elon Musk 1:05:22 Yeah, just, uh, you know, obviously people- people will give, often will give me too much credit, and they'll give Andrej too much credit. So ...

Lex Fridman 1:05:28 And people should realize how much is going on under the-

Elon Musk 1:05:32 Yeah.

Lex Fridman 1:05:32 ... under the hood.

Elon Musk 1:05:32 There's just a lot of really talented people. Um, r- the Tesla Autopilot AI team is extremely talented. It's like some of the smartest people in the world. Um, so yeah,

Elon Musk 1:05:44 we're getting it done.

Lex Fridman 1:05:45 Wha- what are some insights you've gained over those five, six years of Autopilot about the problem of autonomous driving? So, you leaped in having some sort of, uh, first principles kinds of intuitions, but nobody knows how difficult the- the pro-

Elon Musk 1:06:04 Yeah.

Lex Fridman 1:06:04 ... like, the problem

Elon Musk 1:06:05 ... I thought- I thought the self-driving problem would be hard, but it's- it was harder than I thought. It's not like I thought it would be easy. I thought it would be very hard, but it was actually way harder than- than even that. So, I mean, what it comes down to at the end of day is, to solve self-driving, uh, you have to solve, uh, you- you basically need to recreate, um, what y- what humans do to drive, which is humans drive with optical sensors, eyes, and biological neural nets. Um, and so in order to, that- that's how the entire road system is designed to work with- with, uh, pa- basically passive optical and neural nets, um, biologically. Um, and now that we need to, so for actually for full self-driving to work, we have to recreate that in digital form. Um, so we have to, um, that- that means cameras with, uh, advanced, uh, neural nets in silicon form, uh-... and, and then you, it will obviously solve for full self-driving. That's the, that's the only way. I don't think there's any other way.

Lex Fridman 1:07:10 But the question is, what aspects of human nature do you have to encode into the machine, right? So you have to solve the perception problem, like detect, and then you first, uh, well, realize what is the perception problem for driving, like all the kinds of things you have to be able to see. Like, what- what do we even look at when we drive? There's, uh, I just recently heard André talked about at MIT about, like, car doors. I think it was the world's greatest talk of all time about car doors.

Elon Musk 1:07:37 Yeah.

Lex Fridman 1:07:37 Um, the- the- f- you know, the fine details of car doors. Like, what- what is even an open car door, man? So, like, the- the ontology of that, that's a perception problem. We humans solve that perception problem, and- and Tesla has to solve that problem. And then there's the control and the planning coupled with the perception. You have to figure out, like, what's involved in driving, like, especially in all the different edge cases. Um, and- and then the, uh, I mean, maybe you can comment on this. How much game theoretic kinda stuff needs to be involved, you know, at a four-way stop sign? You know, our, as humans when we drive, our actions affect the world. Like...

Elon Musk 1:08:18 Sure.

Lex Fridman 1:08:18 It changes how others behave. Most autonomous driving, if you, you're usually just responding, um, to the scene, as opposed to, like, really, um, asserting yourself in the scene. Do you think...

Elon Musk 1:08:33 I think these ... So, actually, I think- I think these con- these sort of contro- control logic conundrums are not com- are not the hard part. Um, the, you know ... Let's see, um ...

Lex Fridman 1:08:45 What do you think is the hard part of, in this whole, um, beautiful complex problem?

Elon Musk 1:08:50 So, it's a lot of friggin' software, man, a lot of smart lines of code. Um, uh, f- for sure, in- in order to have, um ... create an accurate vector space. Uh, so like, if- if you're- you're coming from image space, which is like this- this flow of, um, photons co- you know, going to the camera- cameras and- (clears throat) and then, uh ... So- so you have this massive bitstream, um, in- in image space, uh, and then you have to, uh, ef- effectively compress, uh, the, uh, uh, ma- a massive bitstream, uh, corresponding to photons that knocked off an electron in- in a camera sensor, uh, and- and turn that bitstream into- into a vector space. Um, uh, by- by vector space, I mean like, uh, you know, you've got cars and- and humans, and, uh, lane lines, and curves and, uh, traffic lights, and that kind of thing. Um, o- once you, uh, have an accurate vector space, um, the control problem is similar to that of a video game, like a Grand Theft Auto or Cyberpunk, um, if you have accur- accurate vest- vector space. It's the control problem is, it's- it's, I wouldn't say it's- it's trivial, it's not trivial, but it's, um, like, it's- it's- it's a- it's- it's not like some insurmountable thing. It's just a, it's ... But- but having accurate vector space is very difficult.

Lex Fridman 1:10:32 Yeah, I think we humans, uh, don't give enough respect to how incredible the human perception system is, to- to mapping the raw photons to the vector space representation in our heads.

Elon Musk 1:10:44 Your brain is doing an incredible amount of processing, um, and- and giving you an image that is a very cleaned up image. Like, when we look around here, we see c-, like, you see color in the corners of your eyes, but actually your eyes have very few, uh, uh, cones, like the, uh, cone receptors in the peripheral vision. Your- your- your eyes are painting color in the peripheral vision. You don't realize it, but they're, eyes are actually painting color. And your eyes also have like this blood vessels and all sorts of gnarly things, and there's a blind spot, but do you see your blind spot? No. Your- your- your brain is painting in the missing, the blind spot. You're gonna do these like, see these things online where you look, look here and look at this point, and- and then look at this point, and it's, if- if it's in your blind spot, it- it, the- your brain will just fill in the- the missing bits.

Lex Fridman 1:11:33 That's so cool. The peripheral vision's so cool.

Elon Musk 1:11:35 Yeah.

Lex Fridman 1:11:35 Makes you realize all the illusions for vision science and so it makes you realize just how incredible the brain is.

Elon Musk 1:11:40 The brain's doing crazy amount of post-processing on the vision signals from your eyes. Um, it's insane. So, um, and then, and then even once you get all those vision signals, uh, your- your- your brain is constantly trying to for- to- to forget as much as possible. So, human memory is p- perhaps the weakest thing about the brain, is memory. So, because memory is so expensive to a brain and so limited, your brain is trying to forget as much as possible, and distill the things that you see into, uh, the smallest am- smallest amounts of information possible. So, your brain is trying to, not just get to a vector space, but get to a vector space that is the smallest possible vector space of only relevant objects. Um, and I think, like, you- you can sort of look inside your brain, or at least, you know, I can. Like, when you drive down the road and- and try to think about what your brain is actually doing ...

Lex Fridman 1:12:37 Yeah.

Elon Musk 1:12:37 ... consciously. And it's- it's conscious ... It's- it's tr- it's- it's, it's like you'll see a car that's ... You could, because your- you- you're, you don't have cameras ... You- you, I don't have eyes in the back of your head or the side, you know? So you say like-... uh, having arguments (laughs) you know, is like, um, so, so then h- like, say like, like, uh, like when's the last time you looked right and left, and, you know, or, and, and rearward, um, or even diagonally, uh, uh, you know, forward, uh, to actually refresh your vector space. So, you're, you're glancing around and what your mind is doing is, is g- is trying to distill, um, the re- relevant vectors, basically objects with a position and motion, uh, and, a- and, and then, and, and then, uh, editing that down to the least amount t- that's necessary for you to drive.

Lex Fridman 1:13:50 It does seem to be able to, uh, edit it down or compress it even further into things like concepts. So it's not ... It's like it goes beyond. The human mind seems to go sometimes beyond vector space to, to sort of space of concepts, to where y- you'll see a thing. It's no longer represented spatially somehow. It's almost like a concept that you should be aware of. Like if this is a s- a school zone, you'll remember that-

Elon Musk 1:14:13 Yeah.

Lex Fridman 1:14:13 ... as a concept, which is a weird thing to represent. But perhaps for driving, you don't need to fully represent those things, or maybe you get those kind of, um ...

Elon Musk 1:14:24 Well, you e- e-

Lex Fridman 1:14:25 ... indirectly.

Elon Musk 1:14:25 You n- you need to like establish vector space and then actually have predictions for, uh, that ve- those vector spaces. So like, um, you know, like if uh ... You know, like you drive past, say, say a s- a b- a, a bus and the, and you see that there's, there's people ... If, before you drove past the bus you saw people crossing, and just, like, or some just ... Imagine there's like a, a large truck or something blocking sight. Um, but you s- before you came up to the truck you saw that there were some kids about to cross the road-

Lex Fridman 1:15:00 Mm-hmm.

Elon Musk 1:15:00 ... in front of the truck. Now you can no longer see the kids, but you, you'd need to be able ... But, but you would now know, "Okay, those kids are probably s- gonna pass by the truck and cross the road," even though you cannot see them. So you have to have, um, memory, uh, that, that you had need to remember that there were kids there and you need to have some f- forward prediction of what their, uh, position will be-

Lex Fridman 1:15:23 It's a really hard problem.

Elon Musk 1:15:24 ... at the time of relevance.

Lex Fridman 1:15:25 So with, with occlusions and computer vision, when you can't see an object anymore, even when it just w- walks behind a tree and reappears, that's a really, really ... I mean, at least in academic literature, it's tracking through occlusions, it's very difficult.

Elon Musk 1:15:40 Yeah, we're doing it. Um-

Lex Fridman 1:15:42 I, I understand this. (laughs)

Elon Musk 1:15:43 Yeah (laughs) .

Lex Fridman 1:15:43 So, some of it-

Elon Musk 1:15:44 It's like object permanence. Like same thing happens with the humans with ne- neural nets. Like if it, if it, when like a toddler grows up, like there's a, there's a point in time where, uh, they develop, they have a sense of object permanence. So before a certain age, if you have a ball, uh, or a toy or whatever, and you put it behind your back, and you p- pop it out, if they don't ... Before they have object permanence, it's like a new thing every time. It's like, "Whoa, this toy went poof, disappeared, and now it's back again."

Lex Fridman 1:16:09 Right.

Elon Musk 1:16:09 And they can't believe it, and then they can play peekaboo all day long because this peekaboo is fresh every time. (laughs)

Lex Fridman 1:16:14 (laughs)

Elon Musk 1:16:16 But then we figure out o- object permanence, then they realize, "Oh no, the, the object is not gone. It's just behind your back." Um ...

Lex Fridman 1:16:22 Sometimes I wish we never did figure out-

Elon Musk 1:16:24 (laughs) Object permanence.

Lex Fridman 1:16:26 ... object permanence (laughs) .

Elon Musk 1:16:26 Yeah, so that's, uh ...

Lex Fridman 1:16:29 So that's an important problem to solve.

Elon Musk 1:16:31 Yes. So, so in a, so in a, like in important evolution of the neural nets in the car is, uh, memory acos- uh, memory across both time and space. Um, so, uh, now you can't remember ... Like you have to say like how long do you want to remember things for? And, and it's, it, there's m- there's a cost to remembering things for a long time. So you could, you, you like run out of mem- memory to, if you f- try to remember too much for too long. Um, and, and then you also have things that are stale if, if, if they're tr- if you remember 'em for too long. And then you also need things that are mem- uh, remembered over time. So even if you, like say have like, for argument's sake, five seconds of memory, uh, o- on a time basis. But like let's say you, you, you're parked at a light, and you s- and you saw ... You use a pedestrian example, that people were waiting to cross the, cross the road, and you can't, you can't quite see them because of an occlusion. Uh, and, but they might wait for a minute before the light changes f- for them to cross the road. You still need to, to remember that they, that that's where they were.

Lex Fridman 1:17:36 Mm-hmm.

Elon Musk 1:17:37 Um, and that they're probably going to cross the road type of thing. Um, so even if that exceeds your, your, your time-based memory, it should not exceed your space m- uh, memory.

Lex Fridman 1:17:48 And I, I just think the data engine side of that, so getting the data to learn all of the concepts that you're saying now is an incredible process. It's this iterative process of just ... It's this, this hydranet of m- many-

Elon Musk 1:18:00 (laughs) Hydranet.

Lex Fridman 1:18:01 Yeah.

Elon Musk 1:18:02 Wait for-

Lex Fridman 1:18:03 (laughs)

Elon Musk 1:18:03 We're changing the name to something else.

Lex Fridman 1:18:05 Okay. Uh, I'm sure it'll be, uh, equally as-

Elon Musk 1:18:08 Yeah.

Lex Fridman 1:18:08 ... Rick and Morty, like.

Elon Musk 1:18:09 There's a lot of ... There's a ... Yeah.

Lex Fridman 1:18:11 (laughs)

Elon Musk 1:18:11 We've rearchitected the, the neural net, uh, the neural nets in the cars so many times it's crazy.

Lex Fridman 1:18:18 Oh, so every time there's a new major version you'll rename it to something more ridiculous, or-

Elon Musk 1:18:22 Uh ...

Lex Fridman 1:18:23 ... or memorable and beautiful? Sorry. Not ridiculous, of course.

Elon Musk 1:18:26 (laughs) If y- if you see the full, the full like, uh, array of neural nets that, that, that are operating in the car, it's g- it kind of boggles the mind. There's so-

Lex Fridman 1:18:36 Yeah.

Elon Musk 1:18:36 There's so many layers it's crazy. Um, so yeah.

Lex Fridman 1:18:42 (laughs)

Elon Musk 1:18:42 Um, but and, and we're, we, we started off with, uh, simple neural nets that were, uh, basically i- image recognition on a single frame from a single camera. Uh, and then, uh, trying to knit those together with, uh, you know, in, with the-... C. Uh, uh, uh, I should say we, (laughs) we're- we're really primarily running C here, 'cause C++ is, uh, s- too much overhead. And we have our own C compiler. So, to get maximum performance, we actually wrote- wrote our own C compiler and are continuing to optimize our C compiler, uh, for, uh, maximum efficiency. In fact, we've just recently, uh, done a new rev on a- on a C compiler that'll compile directly to our autopilot hardware. Um-

Lex Fridman 1:19:27 So you wanna compile the whole thing down and- with your own compiler-

Elon Musk 1:19:30 Yeah.

Lex Fridman 1:19:30 ... like, so efficiency here-

Elon Musk 1:19:32 Absolutely.

Lex Fridman 1:19:32 ... ca- 'cause there's all kinds of compute. There's CPU, GPU, there's like-

Elon Musk 1:19:35 Yeah.

Lex Fridman 1:19:35 ... basic types of things th- th- and you have to somehow figure out the scheduling across all of those things. And so you're compiling the code down-

Elon Musk 1:19:42 Yeah.

Lex Fridman 1:19:42 ... that does all... okay. This is... yeah, so that's why there's a lot of people involved. (laughs) 'cause it-

Elon Musk 1:19:47 There's- there's a lot of hardcore, uh, software engineering-

Lex Fridman 1:19:50 Yeah.

Elon Musk 1:19:50 ... at a very, sort of, bare metal level. Uh, 'cause you- you- you- we're trying to do a lot of compute, uh, that's constrained to the- th- y- you know, our full self-driving computer. So- and we wanna try to have the highest frames per second, um, possible. Um, with- with- in a f- in a s- very finite amount of compute, um, and power. So, um, we really put a lot of effort into the efficiency of our compute. Um, and- and, uh, so there's actually a lot of work done by some very talented software engineers at Tesla that, uh, at- at a- at a very foundational level to improve the efficiency of compute and how we use the- th- the trip accelerators, uh, which are basically, um, do- you know, uh, doing matrix math, dot- dot products, uh, like a bazillion dot products. You know, it's like (laughs) what- what- what- what a neural net, so it's like compute-wise, like, 99% dot products. (laughs) So, you know, um...

Lex Fridman 1:20:57 And you wanna achieve as many high frame rates, like a video game. You want-

Elon Musk 1:21:01 Yeah.

Lex Fridman 1:21:01 ... full- full resolution-

Elon Musk 1:21:03 Mm-hmm.

Lex Fridman 1:21:03 ... high frame rate.

Elon Musk 1:21:05 High frame rate, low latency, um, low jitter. Uh, so, um... I think one of the things we're, um, moving towards now is no post-processing of the image through the, um, uh, the image signal processor. So, um, like, for- for what happens for cameras is that, well, most cameras is they, um... there- there's a lot of post-processing done in order to make pictures look pretty.

Lex Fridman 1:21:39 Mm-hmm.

Elon Musk 1:21:40 Uh, and so we don't care about pictures looking pretty. Um, we (laughs) we- we just want the data. We- we... so we're- we're moving to just r- raw photon counts. So the system will... like, the image that- th- that the computer sees is actually much more than what you'd see if you represent it on a camera.

Lex Fridman 1:21:59 Ah.

Elon Musk 1:21:59 It's got much more data. Uh, and even in very low light conditions, you can see that there's a small photon count difference between, you know, this spot here and that spot there, which means that... so it c- it can see in the dark incredibly well. Um, because it- it can detect these tiny differences in photon counts.

Lex Fridman 1:22:17 That's cool.

Elon Musk 1:22:17 Like ma- much better than you'd p- possibly imagine. Um, so... and- and then we also save, uh, 13 milliseconds on latency, uh, so, um...

Lex Fridman 1:22:29 From removing the post-processing and the image?

Elon Musk 1:22:31 Yes.

Lex Fridman 1:22:31 Yeah.

Elon Musk 1:22:32 It's like, um... 'cause we've got, you know, eight cameras, and- and then there's, uh, roughly, I don't know, one and a half milliseconds or so, maybe 1.6 milliseconds of latency, um, for each camera. And so, it... like, uh, um, going to just... uh, it- it basically... bypassing the i- uh, image processor, uh, gets us back 13 milliseconds of latency, which is important. Um, and- and we track latency all the way from, you know, photon hits the- the camera to, you know, a- all the steps that it's gotta go through to get... you know, go through the, um... the- the various neural nets and the- the C code and... uh, and there's a little bit of C++ there as well. Um, well, like, m- maybe a lot, but it... the- the core stuff is- i- the heavy-duty compute is all in C. Um, and, uh... and so- so we track that latency all the way to an output command to the, um, drive unit to accelerate, uh, the brakes to- to slow down, steering, you know, turn left or right. Um, so 'cause you've got to output a command, that's gotta go to a controller, and, like, some of these controllers have an update frequency that's maybe, uh, 10 hertz or something like that, which is slow. That's like now you lose 100 milliseconds potentially.

Lex Fridman 1:23:50 Mm.

Elon Musk 1:23:50 So, um... so then we wanna update the- the drivers on the, like, say, steering and braking control to have, um, more like, uh, 100 hertz instead of 10 hertz. So then you get a 10 millisecond latency instead of 100 milliseconds worst case latency. And- and actually, ji- jitter is more of a- a challenge than- than- than latency. 'Cause latency is like... you can- you can- you can anticipate and predict, but if you're... but if you've got a stack up of things going from the camera to the c- to the computer, through then a- a series of other computers, and finally to an actuator on- on the c- the car-

Lex Fridman 1:24:21 Mm-hmm.

Elon Musk 1:24:22 ... if you have a stack up of, uh, of tolerances, of timing tolerances, then you can have quite a variable latency, which is called jitter. And- and that makes it, uh, hard to- to- to anticipate exactly what... how you should turn the car or accelerate, because, you know, if you've got maybe 150, 200 milliseconds of jitter, then you could be off by, you know, up to 0.2 seconds. And this can make... this could make a big difference. Uh-

Lex Fridman 1:24:47 So you have to interpolate somehow to- to- to s- to, uh, deal with the effects of jitter, so th- that you can make, like, robust-

Elon Musk 1:24:55 W-

Lex Fridman 1:24:55 ... control decisions? You have- then you have to... uh, so the jitter is in the sensor information? Or is... it... the jitter can occur at any s- stage in the pipeline.

Elon Musk 1:25:05 You can... f- if you have just... if you have fixed latency, you can anticipate, um...... and, and, uh, like say, okay, we know that, uh, our information is, for argument's sake, 150 milliseconds stale. Like so for, for, um, 150, for argument's sake, 150 milliseconds from photon second camera to, um, where you can measure a, a change in the acceleration of the vehicle. Um, so then, uh, then you can say, okay well we're gonna... And to, we, we know it's 150 milliseconds, so we're going to take that into account and, uh, and, and compensate for that latency. However, if, if you've got then 150 milliseconds of latency plus 100 milliseconds of jitter, that... Which could be anywhere from zero, zero to 100 millis- milliseconds on top, so, so then your latency could be from 150 to 250 milliseconds. Now you've got 100 milliseconds that you don't know what to do with. And, and, uh, that's basically random. So, getting rid of jitter is extremely important.

Lex Fridman 1:26:03 Hmm. And that affects your control decisions and all those kinds of things? Okay. Um-

Elon Musk 1:26:09 Yeah. The, the car's just going to fundamentally maneuver better with lower jitter. Um-

Lex Fridman 1:26:13 Got it.

Elon Musk 1:26:13 Uh, and the, the, the cars will maneuver with superhuman ability and reaction time much faster than a human. Um, I mean, I think over time the autopilot full self-driving will be capable of maneuvers that, um, you know, uh, you know, are, are, are far more than what like James Bond could do in like the best movie type of thing.

Lex Fridman 1:26:36 (laughs) That's exactly what I was imagining in my mind-

Elon Musk 1:26:38 Yeah.

Lex Fridman 1:26:38 ... a- as you said it. Um-

Elon Musk 1:26:40 It's like im- impossible maneuvers that a human couldn't do, you know, so...

Lex Fridman 1:26:45 Well, let me ask sort of, uh, looking back the six years, looking out into the future, based on your current understanding, how hard do you think this, this full self-driving problem... When do you think Tesla will solve level 4 FSD?

Elon Musk 1:27:01 I mean, it's looking quite likely that it will be next year.

Lex Fridman 1:27:05 And what does the solution look like? Is it the current pool of FSD beta candidates, they start getting greater and