78 rounds of conversations with Musk.
Author 丨Che Thing(ID:chedongxi)
Image source 丨 "TIME" official website
Want to learn more about what Musk is thinking? Want to learn about the technical challenges facing Tesla and SpaceX? A two-and-a-half-hour video conversation is enough to satisfy most of your curiosity!
Just 9 days ago, Youtube's well-known technology UP master, Big V Lex Fridman, who has 1.43 million fans, had a deep conversation with Musk that lasted for 2 hours, 31 minutes and 47 seconds, from SpaceX's manned spacecraft, Raptor rockets, colonizing Mars, and talking about Tesla's self-driving engineering problems, robots, and Musk's well-known first-sex thinking principles.
As a dialogue program, the video posted on YouTube does not have any other pictures except for the anchor and Musk. But it doesn't matter, what matters is that Lex Fridman's 78 questions are all asked to the point, including technical principles, opinions and opinions, as well as the details of Musk's mental journey and daily life.
Musk in the dialogue is also very active and sincere. At the beginning of the meeting, he sang and trembled directly into the microphone, making it like the star recording scene.
Although the singing is indeed ugly, fortunately, Musk knows everything in the face of 78 questions, and the golden sentences, ridicule and foul language are all in line.
He said that although he is not a religious person, he also prayed on his knees (bless the launch) before the launch of SpaceX's manned spacecraft. Talking about Tesla's autopilot, he joked that Tesla drives better than the 007 James Bond. When it comes to how to face setbacks and challenges, Musk bluntly said that his dictionary has not given up the word, and it does not matter what pessimism and optimism.
"Yikes! That's it. Musk said.
The conversation also includes a lot of Musk's insights, perverse operations, and technical details:
1. When will it land on Mars? The earliest 5 years, the latest 10 years later.
2. How much does a round trip to Mars cost? $1 trillion.
3. When will L4-level FSD be realized? There is hope for 2022.
4. The number of neural networks in Tesla's autopilot software is increasing, and the basic code is rapidly decreasing, making automatic driving closer and closer.
5, SpaceX can not find the right materials when building rockets, they built several alloys...
All in all, these two and a half hours of conversation are informative and wonderful, and are an important window to understand Musk. The following is a full compilation of the transcript of this video dialogue, which has been slightly edited to meet the Chinese reading.
Lex Fridman and Musk
For the first time, manned space has knelt down and prayed that the next step is to establish a civilization on Mars
1. Lex Fridman: In May 2020, SpaceX's manned rocket entered space orbit, which is seen by many as the first step in a new era of human space exploration. These human spaceflights are beacons of hope for me and millions of people, and for the past two years our world has been going through one of the most difficult times in human history.
We see division, fear, cynicism and the loss of common humanity, just when it is most needed, so thank you for giving the world reason to hope and excitement about the future.
Elon Musk: Oh, thank you for saying that. I do want to do that. Humans obviously have a lot of problems, and you know, people sometimes do bad things, but I love this group. I think we should make sure that everything we do allows us to have a great future, an exciting future, to maximize human happiness.
2. Lex Fridman: I would like to ask about the Crew Dragon dem-2 mission, which is Space X's first manned flight, how did you feel before launch? Were you scared? Are you excited? What are you thinking? Are there so many stakes in your mind?
Musk: Yeah, I was really nervous. We obviously can't let them down in any way, so it's extremely stressful. I believe that at the time of the launch of the spacecraft, no one can think of the probability of its success.
We racked our brains to think of ways to increase the probability of success, but we did everything we could, and so did NASA. We've done our best within our capabilities, so we're moving on and launching that spacecraft. I was not a religious person, but I knelt down and prayed for that mission.
3. Lex Fridman: When the ship is successfully launched, can you sleep? Or rather, how did you feel when the spacecraft landed on its successful return?
Musk: After the high pressure, the inner stone finally fell to the ground. It's not so much joy as it is a relief. There's definitely going to be a lot of fun on SpaceX's subsequent missions, and I think the Inspiration Mission is actually very inspiring. I recommend watching a Netflix documentary, Countdown: Inspiration 4 Mission to Space, which is very good. I myself was inspired. During the launch of the spacecraft, I even felt a little enjoyable, not as nervous as when SpaceX first launched the spacecraft.
Lex Fridman: So most people don't really know that The Innin 4 Mission is the world's first commercial mission and the first space trip for "ordinary people."
Musk: Yes, the farthest flight from Earth in the last 30 to 40 years has been the Space Shuttle repair Hubble mission and the 1972 Apollo lunar mission. It's actually a bit puzzling, but of course these two missions are really cool. As human beings, we hope to continue to do better and reach higher realms.
I personally think it would be very tragic if Apollo marked the highest level of humanity. Humans on the moon are the level we've ever reached, but now that 49 years have passed, we haven't made a second lunar landing. Does this mean that our technology has reached its peak? So I think we have to land on the moon again and build a science base there, just like the science bases in Antarctica and the rest of the world. There, we can learn more about the nature of the universe.
I think it's all connected, so that's what I think's the next big thing – building a base and then sending humans to Mars and building a Martian civilization.
5. Lex Fridman: I'll interview you about some of these details. With you busy with engineering challenges, do you still have the energy to marvel at the magic of space travel or every rocket launch?
Do you feel overwhelmed by the challenges you are about to face? It's been a while since SpaceX's first manned mission on May 30, and you can now review the impact it had on you.
What may have affected you the most at the time was an engineering problem, but now it's a historic moment.
What moments do you think are worth remembering in the 21st century? For me personally, one of perhaps one of those moments is the Inspiration 4 Mission, which may be remembered as an early human attempt at the new era of space exploration.
Musk: Regarding this flight, I think maybe some people know, but many more people don't know. It's like, I'm actually the lead engineer at SpaceX, I sign on all of my engineering projects, so if anything goes wrong with any device, it's essentially my fault. So I actually had engineering in mind.
When I see a rocket, I think of the factors that could go wrong and the parts that can improve. That said, it was the same when I saw the Dragon ship. So the average person looking at a rocket might say, "This spacecraft or rocket looks cool!" But I'm going to be like a reader, always talking about what are the risks, what are the problems.
6. Lex Fridman: As you said, when you see a rocket, you make a list of risks. At the same time, you also said that Starship is a very difficult challenge. So, in the same way that you make a list of starship risks, and if one of the engineering problems can be solved perfectly, what would that problem be? Maybe it's about efficiency, the engines, the weight of the different parts, the intertwined complexity of various things, or even the landing of a starship.
Musk: No, the most time-consuming thing at the moment is not the design of the engine, but the production of the engine. I often say that prototyping is simple, the difficulty is to get them out, and we've designed state-of-the-art rocket engines.
The best rocket engines available at the moment are probably the RD-180 or RD-170, which are basically Russian twin-chamber, dual-nozzle engines. And I think the success of an engine is whether it can put the spacecraft into orbit, and our engine has not yet put the starship into orbit, but it actually performs better than the Russian RD series engines.
7. Lex Fridman: So you're referring to the Raptor engine. What is the difference between the Raptor engine and other engines?
Musk: The Raptor engine is a full-flow staged combustion cycle engine that operates at a very high chamber pressure. The key to measuring the quality of the engine is what the number of chamber pressure is.
That number is the pressure in the combustion chamber, and the Raptor engine can run at a pressure of 300 Bar, and possibly even higher. The chamber pressure of the currently operating RD engine should be around 267 Bar. The difficulty of increasing the chamber pressure is increased in a nonlinear manner, so increasing the room pressure by 10% is more like increasing the difficulty of 50%, and the engine needs to obtain a higher power density.
This results in a high thrust ratio and a very high specific impulse. Therefore, a specific pulse is like a measure of the efficiency of a rocket engine, it is actually the effective exhaust velocity of the exhaust gas of the engine. So at a very high pressure, you can have a compact engine, but still it has a high rate of expansion, which is the ratio between the outlet nozzle and the throat.
8. Lex Fridman: So why is the Raptor an engine that is difficult to mass-produce?
Musk: It's very complex to construct. The engine has a lot of parts and a lot of unique materials. Therefore, we had to invent several alloys that did not exist in order to make the engine work, which was a matter of materials. In addition, in a full-flow staged combustion system, there are many feedback loops, propellants and hot gas flowing to different parts of the engine at the same time, and they all have a recursive effect on each other. So if you change a little here, there will be some impact, and then change elsewhere, things will be difficult to control. That's why no one has ever made this type of engine before. The reason why we make a full-flow engine with phased combustion is because it is theoretically the most efficient.
To build a fully reusable rocket, everything had to be done to the extreme— the best engine, the best fuselage, the best heat shielding, the best avionics, and ultra-intelligent control mechanisms.
For example, in reducing the mass of the engine itself, instead of installing landing legs on the booster, we use a tower to undertake the returning engine. So it's like using a pair of chopstick arms to grab the largest flying object ever made on a huge tower. Or maybe it's the same as a karate kid catching a fly, except that the fly is much bigger.
9. Lex Fridman: You mentioned that at some point, you wonder if this can be achieved because it's very difficult.
Musk: One of the things that could be achieved is that I think SpaceX will make the starships work. It's just a matter of time, how long are we going to take to do it? How long will it take for us to truly achieve full and rapid reuse of rockets? Multiple launch tests are required before they can be fully and quickly reused. These experiments I can calculate with physics.
I am confident in success and the calculations are within the predicted success range. At SpaceX, we have a team of talented people who work around the clock to make the vision a reality — the key to enabling the space revolution and making humanity an interstellar civilization, with a fully and rapidly reusable orbital rocket. Never before has an orbital rocket been fully reusable. And this has always been the dream of rocket technology, and many people have tried to do it, but without success.
10. Lex Fridman: You know there are a lot of people, including many experts, journalists, and the general public, who are skeptical about reusable rockets. And you yourself think that this is a non-empty set, although success is more difficult, but you are still fully committed, as an engineer, part of the team to be responsible for promoting the project and completing it. What is the source of your power?
Musk: I just know the way I think about things. For me, this is a very important thing, and we should continue to do it or we will fail. And I don't need any source of power.
11 Lex Fridman: So giving up is not your character?
Musk: Give up these two words are not in my dictionary, whether he is optimistic or pessimistic, get out of the egg, and you're done!
12 Lex Fridman: Can you zoom in on the specific problems of the starship or any of the engineering problems you work on? Can you try to describe the way you think about things? Like different engineering and design issues? Is there a systematic process? You've already said the first principle of thinking, but is there a process?
Musk: Physics is the foundation, everything else is advice. I've seen a lot of people who break the law, but I haven't seen anyone who can break physics. If it's a technical problem, you just need to make sure you're not violating physics. Also, the first principle is analytics, and I think analytics can be applied to anything that comes up in life. In fact, we can boil down some things to the most basic principles. There are some basics in physics, such as the law of conservation of energy, which is impossible if you violate physics. Another approach is to think in terms of limits, and if you were thinking about something and expanding it to a very large category or a very small category, what would happen to things?
13. Lex Fridman: The same is true in rocket building?
Musk: Yeah, in manufacturing as an example, I think it's a seriously underrated problem. Like I said, mass-producing an advanced technology product is harder than designing it. Suppose you want to figure out why this part or product is expensive? Is it because we're doing something stupid? Or is it because our production is too low? So you think, what if the annual output is one million units? Is that still expensive? That's where I'm radical, taking things to the limit. If one million units a year is still very expensive, then the number is not the reason why your things are expensive, the fundamental problem lies in the design.
14. Lex Fridman: Then you can focus on reducing the complexity of your design.
Musk: Yes, it is necessary to change the design, change the components, and solve the price problem from the root. This is common in the rocket space, where production is relatively low, so a common excuse is: "Rockets are expensive because we have low production." If [we] were in the automotive or consumer electronics space, the cost wouldn't be that high. ”
What I thought to myself was, if a million rockets were produced every year now, would the rockets still be so expensive? If the answer is yes, then economies of scale are not an issue.
15. Lex Fridman: Do you incorporate it into manufacturing or supply chains? When you think about resources and raw materials, do you put these factors into a first-principles analysis? For example, how does the supply chain work?
Musk: Exactly, to give an example of thinking about things at the limits, this approach works well for all products. If you look at the raw materials of the rocket, you will find that the raw materials are aluminum, steel, titanium alloys, special alloys, copper and so on.
What is the weight of the constituent elements of each component and what is the value of their raw materials? Without changing the raw materials, the above problems set an asymptotic limit to the cost of the rocket. When the raw material is unchanged, you see the proportion as a magic number, and after continuous calculation, arrange the atoms into the final shape, which will be the lowest cost of your product. In fact, the cost of a product depends on how to put the atoms into the desired shape.
16. Lex Fridman: I want to start with a story that I talk to a lot with Jim Keller who used to work with you.
Musk: Jim did a great job at Tesla.
17 Lex Fridman: So I think he was also inspired by you to think about things in a similar way to yours. Similarly, other people who work at Tesla and SpaceX have the same mindset. They all learned this way of thinking, and it was very obvious.
Jim once told me that the cost of a Tesla Bot may not be high. We had an argument at the time, and I didn't believe what he said, that the cost of mass production on a large scale could not be low. So far, I have spoken many times with academics and experts in the field of simulated robotics, who have developed humanoid robots and powered robots at a high cost. Jim takes a different view, saying, "The first principle is to reduce manufacturing costs. "I think that's what you've done, and you've thought the same way about Tesla cars and various systems that were previously considered complex. Your exact words were: "How do we simplify it now?" ”
Musk: Yeah, I think if a company is good at manufacturing, it can basically be manufactured in large quantities. The manufacturing cost of a product progressively approaches the value of its raw materials, plus any intellectual property rights you need to acquire. It's a very difficult thing to do, but it's possible for anything. So when a product is initially designed, it often happens that people start with the tools, parts, and methods they're familiar with and then try to make the product with the tools and methods they already have.
Another way of thinking is to try to imagine the perfect product or technology, whatever it is. Then think, "How can atoms be perfectly arranged?" Let's try to figure out how to get an item of this shape."
19. Lex Fridman: Before I start thinking about it, it sounds a bit "Rick and Morty" (An American sci-fi comedy cartoon). If people are often frustrated by doing things the way they used to think, then they really should think this way.
Musk: The common way of thinking is actually an inertial function. People are more inclined to use tools and methods they are familiar with, resulting in something that is unlikely to be the perfect product. So that's why we think in two directions: first, what can be built with our tools; second, what does a perfect product look like in theory?
Because the more you learn, the definition of the perfect product will change. Until then, you don't actually know what a perfect product would look like, but you can make a product that approximates perfection. With the above two directions of thinking, you should think: "Now what tools, methods, and materials do we need to create to make atoms into the desired shapes?" ”
This way of thinking works, but few people think so.
Send people to Mars after 10 years at the latest, and Mars is best used in cryptocurrency
20 Lex Fridman: Let's talk about Mars, which you mentioned earlier is that building a scientific base on the moon and doing some research is very beneficial to the scientific community. But in fact, the real leap, in this seemingly impossible category - to land humans on Mars. How soon do you think SpaceX will send humans to Mars?
Musk: 5 years later at the earliest, 10 years later at the latest.
21 Lex Fridman: What are the determinants? Do you think it's determined by engineering factors? Or is the engineering factor not a bottleneck?
Musk: Fundamentally speaking, it's the design of the starship, the Starship starship is the most complex advanced rocket ever, I don't know what the deciding factor is, maybe the size of the starship and something like it.
The starships are large, and the previous ships are not a weight level compared to our starships. So the basic optimization of the starship is to minimize the cost of entering orbit per ton, and ultimately to the cost of landing on the surface of Mars. Optimizing costs may seem like a business goal, but it's actually something that needs to be optimized. The cost per ton of landing on the surface of Mars is fixed, after which we can build a self-sufficient city there. Beyond that, there is very little we can do.
The round-trip cost of going to Mars is now about $1 trillion, but no amount of money can buy a ticket to Mars. In fact, it is not impossible to send people to Mars, but what we want is not only to land on Mars to set up a flag and leave footprints on it, and then, like landing on the moon, not to go back once in half a century. We want to settle on Mars, to survive there for a long time, and I think humans need to be a multi-planet settled species.
This may be a bit esoteric for many people, but over time, the earth is likely to experience some disasters, which may be the result of human actions, or external causes such as sudden climate changes have caused global disasters, and the fate of human beings may be the same as that of dinosaurs.
If this happens, humans will somehow continue to survive, and the sun will gradually grow larger and eventually devour the earth. The Earth could become hot and uninhabitable to humans in 5 million years. Although this is still a long time, it is only 10% longer than the time that the Earth has existed.
But Earth has been around for 4.5 billion years, and for the first time in 4.5 billion years, it is possible to extend life beyond Earth. The window of opportunity may be open for a long time, and when the window of opportunity opens, it is wise to act quickly.
22 Lex Fridman: The nuclear threat, the pandemic, and the risks give us some motivation.
Musk: I mean, civilization could disappear because of a population collapse, or because of World War III.
These are full of uncertainty. I think the future of humanity is most likely good. But every century has a 1% chance of bringing a civilization to an end, and Stephen Hawking has a similar prophecy.
He's most likely right. We should think of humans as a multi-planet inhabited species, like insurance for life.
We send plants and creatures to Mars, inject life into this planet, and make it the second planet with life. They can't travel to Mars on their own, and if SpaceX doesn't bring humans to Mars, humans will become extinct when the sun expands and makes Earth uninhabitable.
23 Lex Fridman: What do you think is the most difficult part of building a civilization on Mars? Is it from the engineering level to transform Mars? Or economic or human? Let these people settle on Mars forever and not return to Earth.
Musk: I think they're going to come back to Earth.
24 Lex Fridman: But some people will choose to settle on Mars forever.
Elon Musk: Yeah, I'm sure a lot of people will, but what we're going to build is a starship that can go back and forth at any time. It's expensive to build a starship, and we want a starship that can be reused while documenting the journey of these people to Mars.
25 Lex Fridman: Have you considered the cost of terraforming Mars? Like some actual construction? Are you now focused on building a spacecraft that will land on Mars?
Musk: Of course I thought about it. But first of all, we must solve the problem of how to get to Mars, if we can't even go to Mars, then there is no point in discussing other things. What I'm thinking about now is to reduce the cost of getting to Mars. At present, the average cost of reaching Mars by weight per ton is about $1 billion, which includes not only the manufacture of rockets and launches, but also thermal insulation, navigation systems, space communications and landing systems. Therefore, $1 billion (about 6.38 billion yuan) is only a rough estimate. But it's obviously a bit expensive, and it's not possible to create a self-sufficient civilization, so we need to cut the cost by at least a thousand times.
26, Lex Fridman: To control the landing cost at $1 million per ton (about 6.38 million yuan)?
Musk: Yes, ideally the landing cost per ton would be much less than $1 million (about 6.38 million yuan). But beyond that, there's the cost of building a self-sufficient city on Mars. This part is key, and if for some reason Earth no longer sends spacecraft to Mars, humans here should be able to survive.
The of a thousand miles, collapsed in the ant nest, is this truth, as long as there is a key part missing, then this is not a self-sufficient city. It's like if a man travels far at sea, he has everything but vitamin C, but he knows that sooner or later he will die without vitamin C.
So we're going to make the city of Mars self-sustaining, and I'm not sure if I'll be able to see that in my lifetime, but I love seeing what it's possible to build.
You might ask me, "What is the minimum tonnage that a spaceship needs to carry to build a self-sustaining city?" ”
There's a lot of uncertainty, and I don't know, because there's a lot of infrastructure needed to be built on Mars, such as semiconductor factories, ore smelters, etc., so the weight is at least a million tons. And Mars is one of the most uninhabitable planets, but it's farther from the sun than Earth.
27 Lex Fridman: Yes, the scope is limited to the solar system.
Musk: Yeah, in the solar system. The solar system may also have planets that are more habitable for humans, but we can't.
28 Lex Fridman: Too hard?
Elon Musk: Yeah, it's very, very difficult.
29 Lex Fridman: Actually, we can change our thinking, you just mentioned that physics is your starting point, and general relativity believes that wormholes exist. Do you think humans can use wormholes to travel at the speed of light?
Musk: The wormhole thing is debatable, but we don't yet know what methods can be used to exceed the speed of light. In fact, there are some space theories that make an object move at the speed of light in space, and then let space itself move, which is actually called distorted space, and this speed can exceed the speed of light.
Like the Big Bang, the universe exploded at a rate far faster than the speed of light.
Lex Fridman: Yes.
Musk: Warping space requires tremendous energy. Exactly how much, is still unknown.
31 Lex Fridman: How much innovation do you have in rocket propulsion? Can [rocket] propulsion be made 10 times more efficient? Have you found a way in physics to significantly improve engine efficiency?
Musk: As I said, the real dream of rocket technology is to build a completely and quickly reusable orbital system. The Falcon 9 is the only reusable rocket, and as you can see in the SpaceX video, the rocket was able to return and land successfully, and we managed to recover the front cone of the rocket fairing, but we didn't successfully recover the first half.
This means that we will keep costs to a minimum when building the first half of the rocket. The Falcon 9 is a two-stage rocket that can be seen as two aircraft, one large and one small, with the larger aircraft successfully recovering and the smaller one not. However, the cost of the first half of the rocket is also very high, at least around $10 million (about 63.72 million yuan).
The Falcon 9 is not being reused to the level we would like. The cost of each launch of the Falcon 9 is not included in our lowest marginal cost, the cost of a rocket's single launch is around $15 million (about 95.57 million yuan) to 20 million US dollars (about 127 million yuan), which is already the lowest rocket launch cost ever. In addition to this, we are working on the rapid reuse of the rocket, and if successful, the cost per ton of the Greek rocket into orbit is expected to be reduced by 100 times.
Imagine if you had a plane or a car and you had to buy a new car every time you drove, it would be expensive. But if you refuel or charge your car, it reduces your travel costs by a factor of 1,000, and that's true in rockets.
As we all know, building rockets is a very difficult thing. So if I can't reuse the rocket, then I can discard some of the construction, even the important parts, and see if I can achieve what I want, and it turns out to be possible. So theoretically, starships can also calculate the cost of each launch, for example, between $1 million (about 6.38 million yuan) and $2 million (about 12.74 million yuan), and then a single launch can carry more than 100 tons of supplies. Although it's a bit crazy to look at it now.
32 Lex Fridman: Yes, it's incredible. So what you're trying to say is that the main purpose of the current Falcon 9 is to make it fully reusable, not to make some sort of breakthrough in theoretical physics.
Musk: There is no brilliant breakthrough, and building a rocket that can be completely reused is already an extremely difficult engineering problem.
33 Lex Fridman: I want to ask a slightly philosophical, interesting question. For now, you're focused on how to get to Mars. If humans had landed on Mars, what kind of political and economic system do you think the early civilizations on Mars should have? This answer helps people imagine the future Martian civilization.
Musk: I personally believe that there should be a direct democracy on Mars, where people can vote directly on things, and the law must be short enough for people to understand. In addition, the law needs to be constantly updated, and there are "legal recycling" provisions, which can abolish some outdated rules and systems at any time.
34, Lex Fridman: You once said that SpaceX will put a piece of Dogecoin on the moon, and you once said that Dogecoin is the currency of the people, so do you think Dogecoin is the official currency of Mars or the currency of the future?
Musk: I think Mars itself should have a different currency than Earth, because the two planets are too far apart to achieve synchronous circulation of money.
35 Lex Fridman: So Mars has to have a separate monetary system?
Musk: Yes, because Mars is about 20 light minutes (the distance a photon travels in a minute), maybe farther, perhaps farther, there is no real synchronization between the two planets.
I don't know if there will be a cryptocurrency on Mars in the future, but there should be a currency that is localized on Mars.
36, Lex Fridman: People on Mars decide the form of money?
Musk: It's perfectly true that the future of Mars should be decided by martians. I think cryptocurrency itself is interesting, it can reduce errors in databases – that is, what we call money.
I am one of the founders of PayPal, so I have a fairly deep understanding of the practical application of money. It's 2021 and the banking industry is still buying large machines, running the old COBOL code.
The Fed's equipment is probably older than the banks', and their equipment is a bulky old-fashioned COBOL host. In addition to this, governments have editorial privileges in currency databases, and they use those editing privileges to get more money, which adds to the error in the database – that is, money. So I think money should be looked at from the perspective of information theory.
The entire economic system can be thought of as an Internet connection, with bandwidth, latency jitter, and packet loss, which are often errors in network communications. So we can also think about money from this perspective, and then from the perspective of information theory, we can think about what kind of system can make the economy work best.
Cryptocurrencies are actually an attempt to reduce the tax impact of governments diluting the money supply.
37. Lex Fridman: Do you think cryptocurrencies are a new way for people to trade and store their wealth in the 21st century?
Musk: Money is just a kind of data, and people tend to think that money itself has power, but it is wrong, money itself has no power.
It's important to include the idea of physics in your thought process, and if you're a billionaire and you're stuck on a tropical island, your money isn't actually worth anything. Because there is no resource allocation on the island, money is the database of resource allocation. There are no resources on the island to allocate other than yourself, so the money is useless.
So we can think of money as a resource allocation database across time and space, and in what form should this database or data system take to maximize its effect? There is a fundamental problem with Bitcoin at the moment – the transaction volume is very limited and the transaction confirmation delay is too long.
From the perspective of this issue, Bitcoin is not an ideal cryptocurrency, and I do not deny that it has some role in itself, and bitcoin may help solve one aspect of the problem of currency databases – namely the storage of wealth or the accounting of relative obligations – but it is useless as an everyday currency.
Although Dogecoin was born out of a joke, I still think it has many advantages. Its trading volume is much higher than Bitcoin, and now if you want to make a Bitcoin transaction, the price of this transaction is very high.
Bitcoin was born in 2008, when the Internet was not as advanced as it is now, and Bitcoin transactions at that time may be able to synchronize with the speed of the Internet at that time. But now it's 2021, and in another 10 years, this [transaction] speed will be ridiculously low.
And I think the number of currencies should grow linearly, and if a currency is too tight and appreciates over time, then people are reluctant to spend it.
Building a vector space is the hardest part of autonomous driving
38. Lex Fridman: Tesla's Autopilot has achieved incredible things over the past six years. I say it's an amazing achievement because I know it's full of challenges.
I was working at MIT at the time, and I had many friends who worked at DARPA (Defense Advanced Research Projects Agency), so I understood the difficulties of computer vision very well. So when I first drove a Tesla car that was originally based on Mobileye's self-driving system, I was skeptical. I don't think it's possible for this car to keep lanes autonomously and create a comfortable driving experience for the driver.
Musk: Actually lane keeping is easy to do.
39 Lex Fridman: But the original prototype that Mobileye gave Tesla didn't solve the problem.
Musk: We have to write a lot of code based on the Mobileye solution, and Mobileye's solution is not the self-driving solution that we bring to market.
40 Lex Fridman: This is actually an episode in the development of Autopilot — the way you do things. You're used to solving problems in a whole new way, which is actually the boldest decision I've ever seen, and you've made a whole new set of self-driving solutions in both software and hardware.
What I'm probably most concerned about and like most is the work That Andrej Karpathy leads—the selection of datasets, the data engine process, the architecture of neural networks, the way networks are tested and validated in real life. Compared to computer vision's image mesh model, this is like artificial intelligence in the real world.
Musk: Andrej is amazing, he obviously played an important role. In fact, we have a lot of very talented people driving this work, Ashock is actually the head of the autonomous driving project, and Andrej is the head of the artificial intelligence project.
It's no exaggeration to say that Tesla Autopilot's AI team brings together one of the top people in the world. That's why Tesla's autopilot software can continue to make progress.
41. Lex Fridman: In the 5 to 6 years of Autopilot's development, what are your insights on autonomous driving?
Musk: From the beginning, I knew it was extremely difficult to get cars to drive themselves, but as the project progressed, I found it to be more difficult than I thought. The basis for autonomous vehicles to drive themselves is to reproduce the way humans drive – the human eye perceives and processes the information with neural networks to finally complete the driving behavior.
Tesla's FSD software is designed in this way, basically using vision algorithms and neural networks to achieve automatic driving. Therefore, in order to deploy FSD, we need to reconstruct the vision algorithm and the neural network in the form of a number, the camera has a performance improvement, and the neural network is in the chip. It is clear that this approach enables fully autonomous driving capabilities. And, that's the only solution, and I don't believe there are other solutions.
42 Lex Fridman: But the question is to encode into machines what human-like perceptions? What do humans pay attention to when driving? For example, if a machine wants to know what an open door is, this is the perception ability. The system must first solve the perception problem before it can further control and plan the car. You have to thoroughly study what's involved in driving, and there are a lot of different edge situations.
Musk: I don't think the difficulty is in the logic of control.
43 Lex Fridman: What do you think is the most difficult part of the whole problem?
Musk: The hardest thing is to build an accurate vector space, because it takes too much software, and a lot of code. After the camera perceives the image, it forms a digital signal, which needs to be mapped into the vector space, and finally can identify cars, people, lane lines, curves, traffic lights, and so on. Once the accurate vector space is established, controlling the vehicle becomes as simple as playing a game, like playing GTA, cyberpunk, etc. Establishing an accurate vector space, while difficult, is not insurmountable.
44 Lex Fridman: Yes, it's an incredible thing that human beings' own perceptual systems map photons into the vector space in their minds, but most people don't realize it.
Musk: Actually, right now your brain is processing a lot of data, giving you a very clear picture. It's as if when we look around, you see colorful colors, but these colors are actually transmitted to the brain by the cone cells of your eyes. This cell can receive light stimulation and can convert light energy into nerve impulses, so it is also called photoreceptor.
Your eyes will "color" what you see. But you don't realize the process. Your brain is doing a lot of post-processing of the eye's visual signals. This is incredible, but in fact, when the human brain processes some data, it also forgets some things, so memory may be the weakest part of the human brain. Our memory is limited, and the brain will forget while receiving information, so in the end it will process less information than the brain receives.
So your brain is trying to create a vector space where you can think about what your brain is consciously thinking while you're driving.
Tesla's cars use cameras to act as eyes to complete autonomous driving. This is the same as the way humans drive, and the human eye is equivalent to two cameras. While it's a good way to use your eyesight to observe your surroundings while driving, people get distracted, thinking about something, texting, arguing, or turning on the radio.
When people look forward and backward, left and right, they are actually refreshing the vector space of the brain, and the brain tries to extract the relevant vector space, which is basically a fixed position and moving objects, and then the brain edits these contents into the minimum amount of information. This information is necessary information when you are driving.
45 Lex Fridman: The brain can automatically edit vector spaces, or further compress them into concepts or something like that. The human brain sometimes transcends the vector space and becomes a conceptual space, and what you see is no longer some kind of manifestation of space, it is like a concept you should know.
If in the school area, you will remember that it is a concept, which is strange. Although you didn't see the students at the time, you will think in your heart: be careful when driving and beware of the students.
Musk: You need to build vector spaces and then make actual predictions about those spaces.
For example, when you are driving, there is a truck in front of you, and some children are ready to cross the road, and when you are closer to the truck, the truck blocks your view and you can't see the children. At this time, you are thinking about where these children are now, and you need to predict (location) in advance.
46 Lex Fridman: For computer vision, it is very difficult to identify moving objects, such as the object that is obscured by a tree and then reappears.
Musk: That's what we're doing —"object permanence," which is kind of like a human neural web, when a person grows up from indolescence, there's a point in time to improve self-awareness, and so does machines.
If you hide a toy or something behind your back and then show it to a baby, for that baby, every time you take something out, it's the first time it comes up, and it's all new to the babies. He/she can play with The Hidden Cat all day. But then, when babies grow up and gain "object permanence," they realize that the thing actually exists all the time, doesn't disappear, just is hidden behind them.
47 Lex Fridman: Sometimes I wish we never got "object permanence."
ELON MUSK: Yes. An important development of automotive neural networks is memory, memory that spans time and space. In fact, the human brain can't control how much memory you have, and the more you want to remember, the more cost it will be.
So if you want to remember too many things for a long time, a lot of memories will become older later, and you also need to remember some new things.
For example, a 5-second memory of you parking your car at a red light and a pedestrian crossing the street when the green light comes on. But if you pass through the intersection again, for some reason you may not be able to see pedestrians, but you still need to remember that there may be pedestrians on the side of the road, and they may have to cross the road and something like that. So even if this exceeds your time memory, it cannot exceed your spatial memory.
48 Lex Fridman: So getting data to learn all the concepts you're talking about is an iterative process.
Musk: Yes, but we might change the name.
49 Lex Fridman: Well, I'm sure it's going to be as good as Rick and Morty.
Musk: Our team has designed neural networks in the car many times, super crazy.
50, Lex Fridman: So every time there's a new version, do you rename it to something more ridiculous or more memorable and beautiful? Ah no, can't be said to be ridiculous.
Musk: There are a lot of layers in the complete array of neural networks that currently manipulate cars, which is very complex, so we didn't do that.
We use a simple neural network, which is basically a frame of an image of a camera to recognize, and then try to combine them together. We have our own C compiler. Therefore, in order to get the maximum performance, we are also constantly optimizing the maximum efficiency of our C compiler.
In fact, we recently made a new REV command on the C compiler that will compile directly onto our Autopilot hardware.
51 Lex Fridman: Do you want to compile everything with your own compiler?
Musk: Yes, there are all sorts of calculations.
52, Lex Fridman: Does that mean that the BASIC THINGS LIKE CPUs and GPUs are also included? So you're involved in compiling the code yourself?
Musk: Yeah, we also got a lot of people involved.
It involves a lot of hard-core software engineering, so we also do a lot of computation on FSD computers, and we need to achieve the highest number of frames per second as possible with limited computing power.
The Tesla team has a lot of very capable software engineers who are also working on making our computing more efficient.
53, Lex Fridman: If you want to achieve as many high frame rates as video games, you need full resolution, high frame rates, and low latency.
Musk: Yes, in addition to the above points, low jitter is also required. Now one of the things we have to do is not need to post-process the image through the image signal processor.
For example, the general camera needs to do a lot of post-processing in order to make the photo look more beautiful. But we don't care if the pictures are pretty, we just want data, so what our camera collects is just raw photon counts.
So our system actually sees a lot more images on a computer than people usually see. The human eye sees two points of almost the same light, and in fact the number of photons may be very different. This means that our camera can also see clearly in the dark, because it can detect small differences in the number of photons.
Since the post-processing of the image is removed, our latency has increased by 13 milliseconds. Each of these cameras saves about 1.5 or 1.6 milliseconds, and 8 cameras add up to 13 milliseconds.
How do we count latency? This includes the scene entering the camera, the camera data entering multiple neural networks, entering programs based on the C language and a small amount of C++ writing, and then giving driving instructions to the autonomous driving system and sending instructions to the electric drive execution unit to achieve acceleration, braking or steering operations.
Some controllers operate at 10Hz (every 100ms), so a delay of 100 milliseconds has already been generated. So we wanted to upgrade the controller of the electric drive system to operate at 100Hz or higher (every 10 milliseconds), which would reduce the reaction delay of the entire system.
Note: Musk said that the delay problem is measurable and predictable, not so difficult. What's really hard is a phenomenon called "jitter," which describes how much data transfer latency varies. The higher the jitter, the delay problem of the entire system will be greatly amplified, and the system will be overwhelmed.
Musk didn't say how Tesla solved the jitter, but he said that the ultimate driving performance of the Autopilot or FSD system is better than that of the 007 James Bond.
L4 will be achieved at the end of this year, and future vehicles can recognize sounds
54 Lex Fridman: Looking back over the past 6 years and looking to the future, based on your current understanding, how difficult do you think it is to solve the problem of FSD? When do you think Tesla Cars will reach L4-rated FSD?
Musk: I think autonomous driving looks very likely, probably in 2022.
55 Lex Fridman: So how is it achieved? Is it through the current FSD beta version? Will this version get better and better and eventually surpass the current level to achieve higher automation?
Musk: Yeah, anyone who pays attention to the Tesla FSD beta can see that the takeover ratio of vehicles is constantly declining. In extreme cases, the driver can also step in to keep the vehicle away from some dangerous situations. Driver intervention per million miles is currently declining dramatically, and it looks like autonomous driving will be possible by 2022, with AFSD having a smaller proportion of accidents than the average person.
For this, we don't have to prove this to regulators, we just need a corresponding criterion to prove that FSD performs better than the average person. I think autonomous driving will be at least two or three times safer than humans, and it will be two or three times more likely to be damaged than humans, and autonomous driving will be a better way.
56, Lex Fridman: So tesla has launched the FSD 10.6 version, the 10.7 version is already on the way, and maybe the 11.0 version will also appear in the near future.
Musk: Yeah, actually we'd very much like to launch version 11.0 in 2021, but version 11.0 has a lot of basic neural network architectures to rewrite, plus some basic improvements in creating vector spaces.
57, Lex Fridman: That means that FSD version 11.0 will have a lot of qualitative leaps, and 11 is also a very cool number.
Musk: FSD 11.0 will really have a lot of basic improvements, and changes to neural networks will bring more performance improvements. There may be a lot of problems in the beginning, but it can be beneficial.
Just like we're taking with alpha software, we basically need to deal with a whole bunch of C, C++ code, replacing part of the C++ code with neural networks, and Andrej Karpathy emphasizes that very much. It's like neural networks gradually "eating" traditional software, and over time, traditional software will be less and less, and neural networks will be more and more.
The end result is still software, but more is achieved through neural networks and less based on heuristic algorithms. This is a big improvement.
58: So you want to keep adding neural networks.
Musk: But it's not even all neural networks, like a game-changer, but you have to make up a neural network with a lot of C, C++.
The neural network uses a small amount of data to identify: it's a lane line; it's a drivable space; it's a car; it's a bicycle or some other obstacle. The neural network must output the correct vector to C, C++ code.
For what we've done, I think we've done a great job of making it grow to a local maximum, which is a remarkable thing. On the other hand, we have also made great adjustments to the training of neural networks, and we need to train the original image to get the final image information.
59, Lex Fridman: So it's a way to reduce complexity, to reduce the complexity of the whole process.
Musk: Yeah, so the amount of code will also be reduced.
60 Lex Fridman: It's fascinating. So, Tesla is currently reducing the difficulty by adding all the sensors.
Musk: Mainly rely on cameras.
61 Lex Fridman: It's the same as humans, but we also hear sounds through our ears.
Musk: Yeah, actually we're going to have to incorporate sound, we need to listen to an ambulance siren or a fire truck, or someone yelling at you, so we need some audio.
Honestly, it's very easy to come up with these ideas, but it's very difficult to implement. Just like proposing that going to the moon is the easiest part, getting to the moon is the hardest part.
There are a lot of core engineering issues at the software and hardware level throughout the process, a lot of code tweaks need to be made to reduce latency, and if we don't, the system won't work.
And that's what engineers do, they're like the unsung heroes you know, and they're critical to the success of the project.
62 Lex Fridman: For me, what's going on out there, what Andre is doing, it's super exciting. Everything from the infrastructure of the software is carried out with a data engine, and the whole process is like a work of art.
Musk: It's unbelievably large. We've written all of these custom software for training and labeling, and for calibration, calibration is essential because, when you're dealing with things like surround video, it's very difficult to mark surround video from scratch.
It takes a long time, say, hours to mark a video clip, or to do a calibration, and the software only needs to do a lot of calculations on the video clip application, pre-assigning and guessing everything that happens in the surrounding video.
63 Lex Fridman: And then there's the need to correct them.
Musk: Yeah, and then all humans have to do is adjust, like where it's not right. It's like increasing productivity by a factor of 100 or more.
Can Tesla humanoid robots be human companions?
64 Lex Fridman: So you said that Tesla robots are going to work in factories. I think humanoid robots are incredible for all robot enthusiasts. I think the graceful movement that humanoid robots, bipedal robots exhibit is really cool. We've talked about all the ideas that are more similar, and some of them that you've already talked about, like the Tesla FSD, are really just another kind of robotics problem.
I'm very concerned that the interaction between humans and robots can show such a human side, do you think robots should work in factories? Or is it anywhere like at home, being a friend or an assistant or something?
Musk: I think the possibilities are endless, and that's not exactly in line with Tesla's main mission direction of accelerating sustainable energy, but making humanoid robots is something very interesting that we can do for the world, and can help the world in many forms.
I think there are a lot of jobs in the future, and if there is no pay, people will not do it, such as washing dishes, if you wash dishes all day, you will be very irritable, even if you really like to wash dishes, are you willing to wash eight hours a day?
Another example is the dangerous work, which is dangerous and boring, and there are repetitive stress injuries.
I think this is where humanoid robots can get the most value, and our goal is to get robots to do some work that humans don't want to do.
65. Lex Fridman: Have you ever imagined hundreds of millions of Tesla robots in the world in the future, doing different jobs and performing different tasks around the world?
Musk: I haven't really thought about it yet, but I think something similar might happen.
66 Lex Fridman: Can I ask a crazy question? At present, the number of Tesla ownership has accelerated, it has approached 2 million, and many of them have FSD (Musk: I think we have exceeded 2 million), do you think the number of Tesla robots will exceed the number of Tesla cars?
Musk: I think it's an interesting question, normally I think about very distant things, but I really haven't thought about the future of Tesla robotics.
I call these robots optimus Subprime because he's not like a giant robot that can transform, but he's supposed to be a generic helper robot. I think Tesla interacts a lot with the real world, we develop autonomous driving and at the same time customize a lot of hardware, hardware software, to make them run efficiently and energy efficiently.
If you have a huge server room with 10,000 compute units, it's not complicated to do a neural network, but now distilling it out to run on a low-power humanoid robot computer or car, it's actually very difficult, with a lot of complicated software work to do.
So we use neural networks to solve the problem of car navigation, and the car is a bit like a robot with four wheels, and if you extend it into a robot with arms and legs, the two more difficult points are to make the robot smart enough and to make it interact with the environment in a smart way.
To achieve all this, you need both real-world AI and being very good at manufacturing. Tesla is very good at manufacturing and has AI, so having humanoid robots working means developing custom motors and sensors that are different from cars.
I think we have the best expertise in developing advanced electric motors and electronic power, both for humanoid robot applications.
67 Lex Fridman: There are some things that appeal to us, but not necessarily all of them. The world is surrounded by great loneliness, many people want more companionship, friendships with other people, etc. In Austin, there are many people who have dogs. There seems to also be a huge opportunity for robots to reduce loneliness in the world and help people connect with each other, and dogs can achieve this goal in some way. Have you considered the Role of Tesla Robots at this point? Or do you think Tesla robots should perform specific tasks and not interact with people?
Musk: Honestly, I didn't think about it from a companionship perspective, but I think it's actually going to end up being a really good partner. It can evolve individually, and over time, he may develop a unique personality, not all robots are the same.
And this personality may become a match for the possessor, or become a companion of the human being, the other half.
I find it funny, like a Japanese phrase I really like: wabi-sabi, subtle flaws that make something special, subtle flaws in the robot's personality, maybe make it an incredible companion, basically like r2d2, c-3po (two robots in the Star Wars movies) or something like that.
68 Lex Fridman: From a robot's point of view, I think there are some flaws that are very good traits, and in a general home environment, there are some flaws that are so cute that you might fall in love with those flaws. But that's very different from autonomous driving, which is in a high-risk environment and you can't screw it up. So robots at home are more interesting.
Musk: Yeah, actually, if you imagine r2d2, c-3po, they have a lot of similar flaws and imperfections, and they argue with each other over stupid things.
69 Lex Fridman: Are they really good at anything? I'm not sure.
Musk: But they definitely added some quirky elements to the story, and they made mistakes. It just resonates with people about their cuteness. So I think that could happen to Tesla robots as well.
We're confident we'll make them, but I'm not sure what the exact timeline is, and we'll probably have a prototype at the end of 2022 or something similar.
70, Lex Fridman: It can be connected to a Tesla car, which must be very cool.
Musk: Yes, it's going to use an autonomous inference computer, and we've done a lot of training for the Model S/Y/3/X models, and the technology in recognizing real-world things can be used directly to robots. However, there are also many actuators and sensors that need to be customized to be developed.
71 Lex Fridman: In addition to the vector space, there is an additional module for love. (Musk: We can also use it in cars), it can be used in all environments, a lot of people will quarrel in the car, and the car can also help people.
You're a history student and a fan of Dan Carlin (Musk: yes, the best blog ever), he's hardly a podcast, (Musk: Dan Carlin is more like an audiobook), and I just talked to him and he said you guys talk about military stuff like that.
Musk: Yeah, I think it should be called the Engineer War, which essentially plays a key role in the victory of the battle when technology changes rapidly.
The rocket was partly inspired by Russia, and young people should try more
72 Lex Fridman: Do you know a lot about World War II?
Musk: We had an in-depth discussion of fighter and bomber technology in World War II, but the final discussion was much broader. I just did a research on all the fighters and bombers in World War II, and the constant game of scissors and stone cloth.
One country can build an airplane, but another country will also build an airplane to beat the previous one, and then the first country will build another plane to beat the previous one, and what really matters is the speed of innovation and access to high-quality fuel and raw materials.
So, although Germany has some amazing designs, they can't make them because they can't get raw materials, they have problems with their oil and fuel, basically the quality of the fuel is very unstable.
(Lex Fridman: So the design is not a bottleneck) Yes, the U.S. has very stable fuels with very consistent quality. If you want to build a high-performance aircraft engine, then the fuel must be a consistent mixture, must have a high octane number, and cannot have impurities or the like, otherwise it will stain the engine.
But the Germans never had good access to oil, and they tried to get it by invading the Caucasus, but that didn't work. Germany has been battling inferior oil, and they can't expect to provide high-quality fuel for their planes, so they have to develop something like additives.
And the United States had great fuel, which was also supplied to Britain, which allowed the British and Americans to design ultra-high-performance aircraft engines. Germany can design engines, they just don't have fuel, and then the quality of the aluminum alloy they get is not so good.
73 Lex Fridman: Did you talk about this with Dan?
ELON MUSK: Yes.
74 Lex Fridman: Looking at history in a broad sense, when you look at Genghis Khan, Stalin, Hitler, what do you learn from these moments? Can it help you gain insight into human nature? Understand human behavior today, whether it is war or individual or human behavior, any aspect of history.
Musk: Yeah, I find history fascinating, there are a lot of incredible things that happen that help you understand the nature of civilization and the individual.
75 Lex Fridman: Does it make you sad that humans do cruel things to each other? For example, during World War II, the cruelty of abuse of power.
Musk: I think some of these things, most of the human beings just live and live, wars, disasters are actually only intermittent and rare, if not, then human beings will soon cease to exist.
Wars tend to be written about by historians, while normal years when little big things happen aren't written about too much. Most people just like farming, living the life they like, being a villager somewhere.
76. Lex Fridman: Tim Dodd (Rocket Science Communicator) made a rocket tree that shows the history of Soviet rockets in its entirety, and Tim is super interested in the future and spaceflight.
Musk: I like guys like that, and it's great if you're interested in things like space, and he's great at explaining rocket technology to the average person. The Raptor used to have a hydrogen engine, but hydrogen has many challenges. It's very low density, it's a refrigerant, so it's always been liquid, very close to zero, and it also needs to be insulated, so it faces a lot of challenges.
I was actually reading some information about the opening of Russian rockets, and I had the impression that the Soviet Union, Russia, Ukraine was switching to methyl oxygen, doing some interesting tests and data, and that was really impressive.
Actually, we can optimize the cost of orbiting, reduce the cost to Mars, and I think you can choose methane, part of the inspiration is from Russia.
77 Lex Fridman: What advice would you give our young people if they wanted to do something big?
Musk: Try to do something useful, something useful to your fellow citizens, to the world. Always think about how much you contribute than you consume? To strive to make a positive contribution to society, I think that is the goal.
Don't try to be a leader or whatever, many times, the person who becomes the leader is the one who doesn't want to be the leader.
If you live a useful life, it's a good life, a life worth living. Like I said, I would encourage people to use physical and psychological tools and apply them to their lives, which is the best tool.
78 Lex Fridman: What advice do you have when you think about education and self-education? How can it become useful and have a more positive impact?
Musk: I encourage people to read a lot of books and absorb as much information as possible. You'll at least have a general understanding of the landscape of knowledge, and try to understand a lot of things because you probably don't know what interests you. Explore extensively, talk to different people, learn about all walks of life, and do whatever profession you like.
Learn as much as you can and look for meaning. Read more and try to find something that matches your talents.