This is the Lex Fridman podcast. The following is a conversation with Elon Musk. Sponsors include Athletic Greens, ButcherBox, Inside Tracker, Roka, and Eight Sleep. Check them out. Now, here is my conversation with Elon Musk.
Yeah. Make yourself comfortable.
Do you not use headphones?
Okay. How close do I need to get to the microphone? The closer you are, the sexier your sound.
Hey babe, what's up? Can't get enough of you, baby.
I'm going to clip that out. Anytime somebody messages me and thinks I'm sexy, come right out and tell me so.
So good. So good. Okay. Serious mode activate. All right.
Serious mode. Come on. You're Russian. You can be serious. Everyone is serious all the time in Russia.
Yeah. That's right. We'll get there, we'll get there.
It's gotten soft.
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 spaceflight 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 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.
Oh, it's kind of easy to say. I do want to do that. Humanity has obviously a lot of issues and people at times do bad things, but despite all that, I love humanity and I think we should make sure we do everything we can to have a good future and an exciting future, one that maximizes the happiness of the people.
Let me ask about Crew Dragon Demo 2. So that first flight with humans on board, 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.
Yeah, that was extremely stressful, no question. We obviously could not let them down in any way. So extremely stressful, I'd say to say the least. But I was confident that at the time we launched, no one could think of anything at all to do that would improve the probability of success. We racked our brains to think of any possible way to improve it and could not think of anything more, nor could NASA. So that's just the best we could do. We went ahead and launched. Now, I'm not a religious person, but I nonetheless got on my knees and prayed for that mission.
How did it feel when it was a success? First, when the launch was a success, and when they returned back to Earth, it was a great relief.
Yeah, for high stress situations, I find it's not so much elation as relief. And I think once we got more comfortable and proved out the systems, it was definitely a lot more enjoyable with the subsequent astronaut missions. The Inspiration4 mission was actually very inspiring. I'd encourage people to watch the Inspiration4 documentary on Netflix. It's really good. That one I was kind of able to enjoy the actual mission and not just be super stressed all the time. For people that don't know, it's the first all-civilian mission to orbit.
Yeah, and it was the highest orbit in like 30 or 40 years. The only one higher was the Hubble servicing mission, and before that Apollo in '72. It's pretty wild, it's cool. I think as a species we want to keep doing better and reach higher ground. It would be tragic if Apollo was the high water mark for humanity. It's concerning that we're 49 years after the last moon mission and haven't been back. Does that mean we've peaked? We need to get back to the moon, build a science base, then go to Mars and become a spacefaring civilization. But since you're so busy with engineering challenges, are you still able to marvel at the magic of space travel? Or are you overwhelmed? That May 30th launch was an engineering problem at the time, but maybe now it's becoming a historic moment. How many moments from the 21st century will be remembered? Inspiration4 might be seen as early steps of a new age.
During the launch itself, the thing many people don't know is that I'm the chief engineer of SpaceX. I've signed off on all design decisions. So if something goes wrong, it's fundamentally my fault. I'm thinking about all the things that could go wrong and be better. When I see the rocket, I see a readout of risks and problems. That's not what others see when they see the product.
So ask you to analyze Starship the same way. You said see a list of risks. If you could magically solve one engineering problem perfectly, which would it be?
On Starship. Efficiency of engine, weight, complexity, controls?
No, the biggest thing is engine production. Not design. Prototypes are easy, production is hard. We have the most advanced rocket engine ever designed. The best existing engine is the Russian RD-180/170. But an engine only counts if it's reached orbit. Ours hasn't yet, but it's the first to surpass those Russian engines.
You're talking about Raptor. What makes it amazing? What excites you?
Raptor is a full-flow staged combustion engine at very high chamber pressure, 300 bar. The record is 267 bar. The difficulty scales nonlinearly. High chamber pressure allows high power density, thrust-to-weight, and specific impulse. Specific impulse is exhaust velocity. High pressure enables a compact engine with high expansion ratio.
Why is it hard to manufacture at scale?
What does complexity mean? Many components, unique alloys.
It's a materials problem. In full-flow staged combustion, there are many feedback loops. Hot gas flows to many places, with recursive effects. It's hard to control. No one's made this before. We need it for highest efficiency. For full reusability, everything must be best: engine, airframe, heat shield, lightweight avionics, clever controls. We shed mass every way, like catching the booster with a tower instead of landing legs. We'll catch the largest flying object ever with a tower and chopstick arms. Like Karate Kid but much bigger.
I mean pulling something like that probably won't work the first time.
You doubt sometimes if it's even possible.
The possible part? At this point I think we'll get Starship to work. It's a question of timing for full and rapid reusability. It will take many launches. But the physics pencils out. I'm very confident success is in the set of possible outcomes.
For a while I wasn't convinced it was in that set.
That's important. But we were... You're saying there's a chance.
I'm saying there's a chance. Exactly. Not sure how long. But we have a talented team working night and day. The critical thing is a fully and rapidly reusable orbital rocket. None has ever been. It's the holy grail. Many smart people tried and failed.
What's your source of belief when it's so difficult? Experts have failed, public doubts, you know success is non-empty but unlikely. Where do you get strength?
Source of strength? That's not how I think. It's simply something important to get done. We should keep doing it or die trying. I don't need a source of strength.
So quitting is not even...
Not in my nature. I don't care about optimism or pessimism. We're going to get it done.
Going to get it done. Can you introspect your thinking process through engineering problems? First principles thinking?
Physics is law, everything else is recommendation. No one can break physics. For any technology, first ensure no physics violation. First principles analysis: boil down to fundamental principles, establish axiomatic base, reason up, cross-check. Another tool: think in the limit. Scale to large or small numbers. For manufacturing cost, ask if it's still expensive at a million units per year. If yes, volume is not the issue; redesign the part. This is common in rocketry where people blame low volume.
Yeah, manufacturing is underrated. Much harder than design. So you ask if it's foolish design or low volume. If expensive at high volume, it's design.
Exactly. Then focus on reducing complexity.
Do you include supply chain and materials?
Yeah. Another example: compute raw material value of constituents in a rocket. That sets the asymptotic limit for cost, the magic wand number. If you could magically rearrange atoms, that's the lowest cost. Actual cost is how you put atoms in shape.
I talk to Jim Keller. He schooled me on first principles for Tesla bot cost.
Jim did great work. If you're good at manufacturing, at high volume you can make anything for asymptotic cost of raw materials plus IP.
People start with familiar tools. Instead, imagine the platonic ideal of perfect product, then figure how to get atoms in that shape. It sounds absurd but necessary.
People use familiar tools by inertia. That leads to products that can be made, but unlikely the ideal. So think both directions: what we can build now and what the perfect product looks like. Then create new tools. People rarely think that way, but it's powerful.
Siobhan Zillis is here. OK, about Mars. You mentioned moon base, but big leap is Mars. When will SpaceX land a human there?
Best case 5 years, worst case 10 years.
Engineering the vehicle. Starship is the most complex rocket ever. Optimization is minimizing cost per ton to orbit and Mars surface. There's a cost per ton where a self-sustaining city is affordable; above that, not. Right now no amount of money can get a ticket to Mars. We need to make it possible. We don't want another flags-and-footprints. To pass a great filter, we must be multiplanet. Earth will face calamity eventually—human-caused, external, or the sun expanding in 500 million years. This is the first time in 4.5 billion years we can extend life beyond Earth. The window may be open for a short time. Risks include nuclear weapons, pandemics, demographic collapse. Most likely future is good, but even 1% per century of civilization-ending event means being multiplanet is life insurance for life.
It's turned into an infomercial.
Life insurance for life. Yes. We can bring Earth's creatures to Mars. They can't bring themselves. If we don't, they die when the sun expands.
Most difficult aspect of building a civilization on Mars? Terraforming, engineering, financial, human?
They could return. Some will. We need the spaceships back for reuse. They are expensive. Many will stay.
Do you think about terraforming? You're focused on getting there.
Absolutely. If you can't get there, nothing else matters.
So and like I said we can't get...
The current cost of one ton to the surface of Mars is about a billion dollars. That is way too expensive for a self-sustaining civilization. We need to improve that by at least a factor of a thousand.
Yes, ideally less than a million a ton. The key threshold is when the Mars city can survive even if Earth ships stop coming. It needs to be self-sustaining. I'm not sure it will happen in my lifetime, but I hope to see momentum. The minimum tonnage is probably at least a million tons for infrastructure like semiconductor fabs and refineries. Mars is not super hospitable but it is the least inhospitable.
Earth is like easy. Yeah, and also I should clarify in the solar system.
Yes, in the solar system.
There might be nice like vacation spots, there might be some great planets out there but it is hopeless. Too hard to get there.
Yeah, way way way way too hard.
Let me push back on that. Not really a pushback but a quick curveball question. General relativity allows for wormholes. Do you think those can ever be leveraged by humans to travel faster than light?
The whole thing is debatable. We currently do not know any means of going faster than light. There are ideas about warping space, but the energy required is gigantic.
How much innovation is possible with rocket propulsion? Can you get a 10x improvement in efficiency?
The holy grail is a fully and rapidly reusable orbital system. Falcon 9 is the only reusable rocket but we don't get the upper stage back. Marginal cost per flight is about 15 to 20 million dollars. With full reusability, we can reduce cost per ton to orbit by a factor of 100. Starship could cost one to two million dollars per launch and put over 100 tons in orbit.
So the biggest bang for the buck is reusability, not a physics breakthrough.
No brilliant breakthrough. Just extremely difficult engineering.
Got it. No new physics required. Let me ask a philosophical question: on Mars, what form of government and economic system would work best?
It would be a new frontier. I suggest direct democracy, not representative democracy. People vote on laws directly, laws must be short. Absolute transparency. Also need a garbage collection function for rules and regulations. Maybe make it easier to remove laws than add.
You posted a meme about crypto. Do you think smart contracts could implement laws?
I am too dumb to understand smart contracts. My approach is to keep deals short and clear. Just make sure everyone understands.
You said Doge is the people's coin. Will SpaceX put Doge on the moon or Mars? Could Doge be Mars currency?
Mars will need a different currency due to speed of light delay. It would need to be localized. Let the Martians decide. Cryptocurrency is an interesting approach to reduce error in the money database. Money is information. Bitcoin has limited transaction volume and high latency. Dogecoin has higher volume and low fees. Linear inflation encourages spending. I am not the creator of Bitcoin; Nick Szabo is more responsible for the ideas.
Are you sure you are not Satoshi? Would you tell us if you were?
Yes, I would tell. But I am not.
Do you think anonymity is a feature or bug?
The evolution of ideas before Bitcoin is clear. Nick Szabo is probably more responsible than anyone else. Names don't matter.
So singular figures aren't as important as the evolution of ideas.
Yeah. What is a name anyway? Shakespeare said a rose by any other name would smell as sweet.
I got Elon to quote Shakespeare. I accomplished something today.
Shall I compare you to a summer's day?
Tesla Autopilot has been an incredible journey. I was skeptical at first because of computer vision difficulty. But Tesla's approach, especially Andre Karpathy's work on data engine and neural networks, is impressive.
Lane keeping is relatively easy. We wrapped a lot of code around Mobileye. We have many talented people. The Autopilot team is extremely talented.
What insights have you gained about autonomous driving? It was harder than expected?
It was way harder than I thought. To solve self-driving, you need to recreate what humans do: optical sensors and biological neural nets in digital form. Cameras with advanced neural nets. It requires an accurate vector space. The control problem is like a video game; the hard part is perception.
What is the hardest part? Is it the control logic conundrums?
It's a lot of software. Creating an accurate vector space from image space is difficult. The brain does incredible processing. Human vision is flawed but the brain fills in. And then the brain tries to forget and compress to only relevant objects.
The brain's ability to edit and compress into concepts is remarkable.
Human memory is weak. The brain is trying to forget as much as possible. When driving, you only need the relevant vectors. Humans are distracted. So it is harder than it seems.
You'll remember that 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
Well, you need to establish vector space and then actually have predictions for those vector spaces. For example, if you drive past a bus and you see people crossing the intersection before you lose sight of them, you need to remember that they're there and predict their position, even though you can't see them. Tracking through occlusions is very difficult in computer vision. It's like object permanence in humans. Toddlers eventually develop object permanence; before that, peekaboo is fresh every time. But we figure it out. Sometimes I wish we never did figure out permanence. That's an important problem to solve.
Yeah, we're doing it. I understand this. Yeah. So some of it is about object permanence with neural nets.
It's like the same thing that happens with humans. When a toddler grows up, they develop a sense of object permanence. Before that, if you hide a toy, it's a new thing every time. But then they realize the object is not gone. Sometimes I wish we never did figure out permanence. That's an important problem to solve.
That's an important problem to solve. Yes, so an important evolution of the neural nets in the car is memory across both time and space. You need to remember things for a certain duration, but there's a cost to remembering for too long. Even if your time-based memory is, say, five seconds, you still need to remember that pedestrians were waiting at a crosswalk for a minute before the light changes. So it's a hard problem.
And I just think the data engine side of that, getting the data to learn all the concepts you're describing, is an incredible iterative process. It's this hydronet of many
Hydronet. Yeah. We're changing the name to something else.
Okay. I'm sure it'll be equally as Rick and Morty like.
There's a lot of... yeah, we've rearchitected the neural nets in the cars so many times. It's crazy. If you see the full array of neural nets operating in the car, it boggles the mind. We started with simple image recognition on single frames from single cameras, then tried to knit them together. We're primarily running C because C++ has too much overhead. We wrote our own C compiler to get maximum performance, and we just did a new revision that compiles directly to our autopilot hardware. We want to compile everything down for efficiency, handling CPU, GPU, and other processors. There's a lot of hardcore software engineering at a bare-metal level. We're constrained by our full self-driving computer, so we try to maximize frames per second within a finite amount of compute and power. We put a lot of effort into efficiency, and our talented software engineers improve how we use trip accelerators for matrix math. Neural nets are 99% dot products.
And you want to achieve as high frame rates as possible. It's like a video game.
Full resolution, high frame rate, low latency, low jitter. One of the things we're moving toward is no post-processing of the image through the image signal processor. Most cameras do a lot of post-processing to make pictures look pretty, but we don't care about that. We just want the data, so we're moving to raw photon counts. The computer sees much more data than what you'd see on a camera, even in very low light, because it can detect tiny differences in photon counts. It can see in the dark incredibly well, much better than you could imagine. We also save about 13 milliseconds of latency by removing post-processing. With eight cameras, that's roughly 1.6 milliseconds per camera, so bypassing the image processor gains us back 13 milliseconds of latency, which is important. We track latency all the way from photon hitting the camera to output command to the drive unit. Some controllers have an update frequency of 10 hertz, which is slow—you lose 100 milliseconds potentially. We want to update steering and braking controls at 100 hertz instead, getting 10-millisecond latency instead of 100. Jitter is more challenging than latency because you can anticipate fixed latency, but jitter introduces variable delay that makes control decisions harder. You need to interpolate to deal with jitter. The car will fundamentally maneuver better with lower jitter, with superhuman ability and reaction time much faster than a human. Over time, full self-driving will be capable of maneuvers that would make James Bond movies look tame.
That's exactly what I was imagining in my mind as you said it. It's like impossible maneuvers that a human couldn't do. Well, let me ask, looking back the six years and looking into the future, based on your current understanding, how hard do you think this full self-driving problem is? When do you think Tesla will solve level 4 FSD?
It's looking quite likely that it will be next year. Anyone who's been following the full self-driving beta closely will see that the rate of disengagements has been dropping rapidly. Disengagements per million miles have dropped dramatically. The trend looks like next year the probability of an accident on FSD will be less than that of an average human, and then significantly less. But we need to prove it to regulators and ourselves. We want a standard that is at least two or three times safer than a human before we say it's okay. Version 10.6 just came out, 10.7 is on the way, and maybe version 11 is on the way eventually. We were hoping to get 11 out this year, but it involves fundamental rewrites of the neural net architecture and improvements in creating vector space. 11 would be a single stack for everything—one stack to rule them all. There are fundamental changes that will allow much more capability, but initially they'll have issues. It's basically taking a massive amount of C++ code and replacing it with neural nets. Andrej Karpathy says neural nets are eating software. Over time, there's less conventional software and more neural net stuff—more matrix-based, less heuristics. One big change is that currently the neural nets deliver a giant bag of points to the C++ code. We call it the giant bag of points. Each pixel has an associated label, and the C code assembles that into vectors. We want another layer of neural nets on top to distill the giant bag of points into vector space within the neural net, not in heuristics. That's a game changer. Neural nets all the way down. The neural net would output proper vectors to the control code instead of having C construct them. We've done a good job, but we're hitting a local maximum. Also, all networks need to move to surround video, and training needs to move to surround video as well. Efficiency of training needs to improve. We need to move everything to raw photon counts instead of processed images, which is a big reset on training.
So ultimately it's reducing the complexity of the whole thing.
Lines of code will actually go lower.
Yeah, that's fascinating. So you're doing infusion of all the sensors, reducing the complexity of having to deal with these cameras.
