Elon Musk: The hardest part of robot manufacturing is the hand, and the performance of the new generation AI5 chip has increased by 40 times

Recently, Elon Musk appeared on the well-known American business technology podcast "All In Podcast," where he shared his views on topics such as the Department of Government Efficiency (DOGE), the Optimus Prime robot, SpaceX's Starship, and AI. During the show, Musk stated that the lesson he learned from his experience with DOGE is: the Trump administration cannot be fixed at all, and we can only hope for AI and robots to solve America's debt problem. Musk referred to the Optimus robot as "the greatest product in human history," stating that the third version of Optimus currently being designed will address hand flexibility, possess an AI brain, and is expected to achieve mass production—three advantages that no other competitors can match. Musk also mentioned that the Tesla AI5 chip has achieved a 40-fold performance improvement compared to AI4, marking the largest upgrade since version 12; the current AI4 chip can already achieve 2-3 times or even 10 times safer autonomous driving than humans. He anticipates that as early as next year, we may have AI that is smarter than any individual human in all aspects. "All In Podcast" is a well-known podcast focusing on investment and business, covering topics such as investment, technology, entrepreneurship, and economics. The show is co-hosted by four top venture capitalists from Silicon Valley: Chamath Palihapitiya, Jason Calacanis, David Sacks, and David Friedberg. Here are the highlights from the discussion: The lesson Musk learned from his experience with DOGE is: the government cannot be fixed at all, and we can only hope for AI and robots to solve America's debt problem. Musk admitted that he has invested more energy in Optimus than in anything else. He is currently completing the design of the third version of Optimus. This robot will be an outstanding robot. If successful, Optimus will become the greatest product of all time. If you want to do everything a human can do, you need a humanoid robot. Musk stated that efforts are currently underway to finalize the design of Optimus hardware, primarily the hands, which will have flexibility comparable to that of humans. The hands and forearms are the main challenges in the entire robotics engineering. The third version of Optimus will have a very complex hand and an AI mind capable of navigating and understanding reality, and it will achieve mass production, which are three key elements missing from other robotics companies. Humanoid robots lack a supply chain, so they must be built from scratch, requiring a lot of vertical integration; all of Optimus's actuators cannot be obtained from the existing supply chain. The development difficulty of Optimus is higher than any previous Tesla product, although Starship is the most challenging. Once the annual production of Optimus reaches 1 million units, the marginal production cost will be around $20,000. The cost of AI chips may be $5,000 or $6,000, or even more If the annual production reaches 1 million units, the production cost will reach $20,000, or even $25,000. The price will depend on demand. The Tesla AI5 chip has achieved a 40-fold performance improvement compared to AI4, with an 8-fold increase in computing power and a 9-fold increase in memory. The current AI4 chip can already achieve 2-3 times or even 10 times safer autonomous driving than humans, and version 14 software will be the largest upgrade since version 12. The most serious limitation of AI4 currently lies in running the softmax operation, which requires about 40 steps to run softmax in simulation mode, while AI5 can complete it natively in just a few steps. At that time, "your car will make you feel like it is conscious." SpaceX will demonstrate the complete reusability of Starship next year, with version three capable of sending over 100 tons of payload into orbit. The biggest technical challenge now is to manufacture a fully reusable orbital thermal protection shield, which is a materials science and engineering problem that has never been solved. Starlink will provide direct satellite service for mobile phones, allowing users to watch videos anywhere in the world, but hardware upgrades will be needed to support the new frequency. In the future, the possibility of acquiring operators to gain more spectrum cannot be ruled out, with the ultimate goal of providing a comprehensive communication solution that integrates home and mobile services. Creating a Grok version of Wikipedia, "Grokipedia," would be interesting. xAI's Grok large model is rewriting and correcting training data using a large amount of reasoning computation, including correcting errors and biases in information sources like Wikipedia. Musk believes that AI may be smarter than any individual human next year, and by 2030, it may be smarter than all humans combined. Colonizing Mars remains Musk's long-term goal, with the key test being whether Mars can be self-sufficient without Earth supplies. He predicts that if cargo shipments during each Mars transfer window (every two years) grow exponentially, Mars could achieve self-sufficiency in about 25 years. The following is a summary of the full dialogue, assisted by AI tools for translation: Musk: I am currently at Tesla's global engineering headquarters in Palo Alto, California, not in Washington, D.C. I haven't been to Washington, D.C. since May. It was truly a challenging experience. Host: Do you have any insights about your experience in Washington, D.C.? Musk: The government is basically unfixable. I support David's noble efforts; having talented people in government is a good thing, but ultimately, look at our national debt, which is incredibly high, and interest payments even exceed the Department of Defense's budget. If artificial intelligence and robotics cannot solve our national debt problem, we are doomed. Musk: I believe Optimus will become the greatest product in human history Host: What is the current progress? How much energy do you dedicate to Optimus? What is the timeline? I think you are on the third version now, maybe the fourth. Tell us everything. Musk: We are refining the design of Optimus's third version, and it will really be an outstanding robot. It will have hand flexibility similar to that of humans, meaning it will have a very complex hand structure. It has an AI brain capable of navigating and understanding reality, and it will be mass-produced. These are the three missing elements. If you look at any other robotics company, they lack these three things. These are three genuinely difficult problems. At this point, I am probably putting more energy into Optimus than any other single project. This involves solving real-world AI problems, all the electromechanical issues of Optimus, supply chain, and production challenges, as there currently is no supply chain for humanoid robots. So we have to recreate it from scratch, which requires a lot of vertical integration. The actuators in Optimus cannot be sourced from the existing supply chain. But I think to be precise, if successful, Optimus will be the largest product ever, with a cost of $20,000 to $40,000 per robot after mass production by 2030. Host: What do you think the cost of the first batch of products will be? When will we be able to buy one to work on a farm? Musk: I think once we reach an annual production of one million units, the marginal production cost will be around $20,000. This somewhat depends on your investment in the robot AI chips. You need to achieve a lot of efficiency improvements in the actuators. Each arm has 26 actuators, including 26 motors, gearboxes, and power electronics. The AI chips will be quite expensive, possibly accounting for $5,000 or $6,000, or even more, in the bill of materials. But I think at a production scale of one million units per year, the production cost will be around $20,000, maybe $25,000, roughly at that level. The price will be determined by demand. Host: Elon, can you explain to everyone why hand design is so important, why your actuator design is so unique, why it is so difficult and no one manufactures it, and why we must start here to correctly build the rest of the robot? Musk: It turns out that the human hand has evolved to be extremely precise and complex. Your hand is actually an amazing thing. Take a close look at your hands and think about all the things you can do with them; it’s really a lot. Your hands are very versatile tools. You can swing a baseball bat, thread a needle, play the piano, or play the violin. You can disassemble or assemble a car. Hands are extremely versatile tools. Most of the muscles in the hand are actually in the forearm. So your hand is a bit like a puppet; it’s mainly a puppet. The muscles come from the forearm, pulling on the tendons, and the design or evolution of human tendons is exceptionally remarkable You have such a tendon network. The human hand has about 27 or 28 degrees of freedom depending on different calculation methods. That's amazing. To create a general-purpose humanoid robot, you have to solve the hand problem. The robot needs hands. Host: Is it like when you initially built Tesla, where the supply chain didn't exist, and now you have to go out and find partners, establish all this vertical integration, and get support? Is it really that you can't find anything at all, and you have to build all these things? Yes, we actually couldn't buy actuators no matter how much we spent. They simply didn't exist. Even though there are 10,000 to 20,000 different sizes and shapes of motors on the market. We had to design each motor, gearbox, and control electronics from scratch, basically starting from the first principles of physics. Host: The good news is that you have a lot of factory experience over the past few decades. So how challenging is this compared to the Cybertruck, Model Y, Model X, and Gigafactory, you know, the manufacturing nightmare known as Model X? Musk: It's much harder than those. Much, much harder. Host: What about Starship? Musk: Starship is harder. Host: So, somewhere between Model X and Starship? Musk: Yes. Host: Hardware and software, which is harder? Musk: Right now, we are having difficulties with the final design of the hardware. As I said, it's mainly the hand design. Not to overlook the other parts of the robot, which are also important, but the hand (including the forearm) accounts for a large portion of the difficulty of the entire robotic engineering. Host: Assuming you overcome the hardware challenges, based on the advancements in large language models, what can you achieve? Can consumers interact directly with the robot, talk to it, ask it to do things, and it will understand and execute? Musk: Yes, no problem. I noticed you spend a lot of time online. Actually, it’s not that long. Maybe I went a bit overboard on Bunning Grom Imagine, but seriously, those characters and robots seem to embody Annie's image. Host: Elon, why choose a humanoid design? You could create something better or simpler than humans to accomplish specific tasks, perhaps even do more than humans. How do you decide to make it human-like? If you want to do everything a human can do, you need a humanoid robot. If you only want to do part of it, it's much simpler. There is a good reason why humans evolved into their current shape and abilities. Having five fingers (four fingers plus a thumb) is indeed valuable, and even the little finger is quite useful. Toes are more important for walking, but fingers are equally important Elon Musk: And humans designed this world; we designed it for ourselves. If you can create humanoid robots, they will immediately be backward compatible with everything we have built for the world. Host: Elon, there’s another part to the robots. Besides large language models and hand drives, there are also the chips that run them. Regarding Dojo, I saw you posted about AI5 and AI6 on X, and you seem very excited about the development direction of the chip layer. Can you tell us what that is? What are we building? Is it a complement to everything that exists in the world, or a potential long-term competitor? Musk: At Tesla, we essentially have two different chip projects. One is Dojo for training, and the other is what we call the AI4 inference chip. AI4 is currently installed in all vehicles. We are finalizing the design of AI5, which will be a huge leap from AI4. By certain metrics, the improvements of AI5 will be 40 times better than AI4. Forty times, not 40%. This is because we are working very closely and meticulously on AI software and AI hardware. We know exactly where the limiting factors are, and the AI hardware and software teams are designing the chips together. Host: A 40-fold improvement in silicon chips, what does that mean for the audience's experience? Does this mean a magnitude increase in the quality of FSD and the safety of the Tesla driving experience, as well as improvements in robot quality? How will these manifest when you actually go into production? Musk: Specifically, the 40 times refers to the most severe limitation of AI4—running the softmax operation. Currently, we have to run softmax in about 40 steps in simulation mode, while in AI5, it can be done natively in just a few steps. The AI5 chip will be able to easily handle mixed precision models and will dynamically manage mixed precision. AI will perform better in many technical aspects. In terms of raw computing power, the computing capability has increased by 8 times, memory has increased by about 9 times, and memory bandwidth has improved by about 5 times. But because we have addressed some core limitations of AI4, the 8-fold increase in computing power is multiplied by another 5-fold improvement, which comes from optimizing the currently suboptimal parts of AI4 at a very fine silicon chip level. That’s where the 40-fold improvement comes from. I believe that the AI4 chip in current cars can achieve at least 2 to 3 times safer autonomous driving than humans, and it may even reach 10 times. The relevant software will be released in the coming months. Version 14 will be the biggest upgrade to Tesla software since version 12. We have increased the number of parameters by an order of magnitude. A lot of reinforcement learning has been used. You can think of AI as a way to compress reality, where some of the compression steps are too lossy, and we have addressed the lossy issues in the compression steps. These are all software updates. By the end of the year, your car will make you feel like it is conscious. Host: I saw in the news that you spent about $17 billion to purchase some spectrum, allowing your satellites and Starlink network to connect directly to mobile phones. What will it look like in a year or two? Will we cancel our Verizon accounts and just expand our Starlink accounts? Musk: We hope so, because Verizon is indeed not very good. How many of you want a Starlink phone? Is this technically feasible? This is a long-term project. It will allow SpaceX to provide high-bandwidth connections directly from satellites to mobile phones. But the phones will need hardware changes. Current phones do not support these frequencies, so the chipsets will need to be modified to add these frequencies, which will take about two years. Therefore, phones capable of using the acquired spectrum will start shipping in about two years. We also need to build satellites that communicate on these frequencies. We are building satellites in parallel and working with mobile phone manufacturers to add these frequencies to the phones. Then the satellites and phones will handshake well for high-bandwidth connections. The end result is that you should be able to watch videos anywhere on your phone. Host: That would be crazy. Will these frequencies work indoors, inside buildings? Just like your phone does now? Musk: If you are in a building with a thick metal roof, then no. But in a regular home, yes. Host: Elon, your vision is that we won't need an AT&T account, and we won't need to roam when we're in the UK or India; we can sign directly with Starlink and ultimately use it worldwide. Not now, but at some point. Is that the ultimate goal? Basically, we won't need regional operators; we have a global operator, which is you? Musk: That will be one of the options. To be clear, we will not put other operators out of business. They will still exist because they have a lot of spectrum. But yes, you should be able to have a Starlink account, just like you have an AT&T, T-Mobile, or Verizon account, and you can have an account that works with your home Starlink antenna, providing free Wi-Fi as well as mobile service. This will be an integrated solution providing high bandwidth for home and mobile direct connections. Host: Will you acquire some operators to gain more spectrum? Maybe you could acquire Verizon? Musk: That possibility cannot be ruled out. I think it could happen. Host: Let's talk about Starship. You just had what looked like a very successful launch. How close is it to being predictable and ready for commercial use? Musk: I think we will recover Starship next year. We have one more Starlink version two launch left, with only one version two design booster and spacecraft remaining. After that will be version three, which is a huge upgrade because it is equipped with Raptor 3 engines Version 3 of the rocket has changed almost everything. Version 3 may have some initial teething problems because it is such a complete redesign. But it is capable of sending over 100 tons of payload into orbit in a fully reusable manner. I believe that unless we encounter significant setbacks, SpaceX will demonstrate full reusability next year, capturing boosters and spacecraft, and being able to send over 100 tons of payload into useful orbit. Host: What is the current capability of the best rocket in the world in terms of payload? Musk: In terms of commercial rockets, there is the Falcon Heavy. With the reuse of side boosters, it can deliver about 40 tons. So that's five times larger... well, it's 2.5 times larger, but Starship will be fully reusable. Host: Got it. So all components will return. Elon, after the failed launch that resulted in an explosion, many claimed there would be environmental and FAA issues, among various other problems, but the speed at which you returned to the launch pad is incredible. How did you get back on track so quickly? Not just in terms of technology and work, but also in regulatory approvals, as they said there would be various issues and scrutiny. How did you manage that? Musk: We faced a lot of issues and scrutiny, but we resolved them one by one. Thanks to the SpaceX team, who worked extremely hard to complete the next flight and booster tests and successfully launched it from the launch pad. I have great respect for the SpaceX team and am very proud of their outstanding recovery work. Creating a fully reusable orbital rocket is one of the most difficult engineering challenges ever, certainly a candidate for the most difficult engineering projects, at least worthy of a podium finish. This has been SpaceX's goal since 2002, and now 23 years have passed. It has been a long journey, with what I believe is the most talented team of rocket engineers ever. I believe we will achieve full reusability next year. Host: Before achieving full reusability, what are the main technical obstacles you are focusing on? Are there some key challenges that you are repeatedly pondering, or is it more about completing a checklist of learnings and integrating them into the next launch? Musk: For the complete reusability of the spacecraft, there is still a lot of work to be done on the thermal protection system. No one has ever made a fully reusable orbital thermal protection system. The thermal protection system of the Space Shuttle required nine months of maintenance after each flight. So, no one has ever made a fully reusable orbital thermal protection system. Host: Is this a materials science issue, an engineering issue, or both? Musk: It is both a materials science and engineering issue. We are really studying fundamental physics, starting from first principles of physics, trying to figure out how to create materials that can withstand high temperatures, are lightweight, and do not conduct heat to the main structure During the ascent, if it encounters rain, the tiles will not dissolve in the rainwater. There are many different issues, and you really need to ensure that these tiles can work properly. You cannot conduct heavy inspection work. It does require tens of thousands of tiles to work properly without needing to be refurbished or inspected one by one like a space shuttle. Host: Now I would like to ask some questions about Grok and xAI. Can you give us an update? You mentioned that the training of the next generation model will not start from a general network and general scraping, but is expected to use a large amount of synthetic data. Please tell us about the development of Grok and why this innovation is so important. Musk: We are using a large amount of reasoning computation to analyze all source data, which is essentially a corpus of human knowledge, and then thinking about each piece of information, adding missing content, correcting errors, and removing false information from the training data. For example, take Wikipedia, but this really applies to books, PDFs, websites, and various forms of information. Grok uses a large amount of reasoning computation to look at Wikipedia pages, determine what is true, partially true, false, or missing in the pages, and then rewrite the pages to correct the content, remove false information, correct half-truths, and add missing background information. Host: Can this content be published directly? Can we create a "Grokipedia"? This would be particularly useful for our profile pages, as Wikipedia is very biased, and it's an ongoing battle. If content is corrected, a group of people will try to revert it within five minutes. Wikipedia has become extremely partisan, filled with activists. Musk: If we really fix Wikipedia as a source of truth and publish it for the whole world to use, that would be great. I will discuss this with the team, such as Grok or other projects. This would be a version of Grok, and it would be interesting. Host: Regarding training Grok 5, you are expanding the supercluster Colossus located in Memphis and also plan to build a second one. Can you give us an update on the progress? As part of this, where do we stand on scaling laws? If we scale up the cluster, can we obtain more powerful AI models? Is there a point of diminishing returns? If we invest twice the computational resources, will we get a 10% better model or a 100% better model? Is this a logarithmic linear relationship? How much potential do you think is left for hardware scaling? Musk: I believe there is a natural logarithmic function relationship between the amount of computation and the increase in intelligence. For example, ten times the computation may double the intelligence. This could be a rough rule of thumb. But it still means that increasing from an IQ of 100 to 200 is still quite a significant improvement. I think we will see intelligence continue to expand until most of the energy of the sun is used for computation, and eventually most of the energy of the galaxy, which corresponds to the computational scale of a Type II or Type III civilization Once you see artificial intelligence not just as a destination to reach, but as part of the overall intelligence upgrade we recognize, human intelligence also expands with population growth, allowing us to store more and more information, thus enhancing human intelligence. However, due to declining population and low growth rates, human intelligence is leveling off and may actually decline. I suspect that as early as next year, we may have AI that is smarter than any individual human in every aspect. By 2030, AI may be smarter than the sum of all humans. Host: Do you think humanity is declining because of the evolution of AI? Do you think there is an ecological evolution happening on Earth that we don't really understand the structure of, but perhaps we have a vague sense that it is coming? Musk: I hope the birth rate can rebound. I am a strong supporter of increasing the birth rate. Host: Have you done anything about it? Musk: I am trying to set a good example. Host: We had an unexpected big discussion at this conference about suicidal sympathy, Western civilization, and declining birth rates. I noticed you are very proactive and open about this, and that opening borders is like letting invaders in. Could these three be the same thing? Musk: It seems that there are many symptoms of suicidal tendencies in the West. The most obvious is that the birth rate is not at replacement level. Clearly, if this situation continues indefinitely, the West will not be able to reproduce enough population to replace itself. But there are other issues. The borders are completely open to the extent that Western culture and social structures are beginning to collapse, which is particularly evident in Europe, where the native cultures of the UK, France, or Germany may begin to be replaced by incoming, differentiated cultures. There is also the crime issue; we see this case on social media, where a young woman named Irina was senselessly murdered on the subway. That alone is horrifying, but beyond that, elite media for some reason refuse to report on it, as if it doesn't exist. So there is a crime problem that is not being addressed or even acknowledged, and we seem to be trying to deny this downward spiral reality. All these data points seem to indicate that the West is suicidal or seems unwilling to defend or perpetuate itself. Host: Look, I think everyone here feels that life is good, right? I mean, life is pretty great, and I think it's worth living. When Alex Karp defended the West here earlier today, it received the loudest applause at the conference. So, I think we may not really understand what is happening. We don't really understand. Elon, what is your view? What do you think about the West's suicide? Musk: I am very concerned about this. I am very concerned. I think the behavior of the West is akin to suicide Host: But at least in the United States, there is usually a sense of optimism. When was the last time you talked to Europeans living in Europe, and they expressed optimistic feelings? Musk: It's been a while. Decades, even rarely seeing a person. So I think unless people have a sense of optimism and purpose about the future, suicide might be something that happens. Having children is an optimistic act towards the future. So I think we need to give people a sense of optimism and excitement about the future, and the belief that the future will be better than the past, so they will be more interested in having children. Host: Elon, can you talk about the integrity issue of simulation? Are we virtual? Musk: Maybe. I don't know whether we will find the answer in AI or space first. I hope more people can support the philosophy of curiosity because I think it's very exciting and essentially optimistic. There is an amazing curiosity about the nature of the universe. When you uncover a secret in the universe, it's magical, and the whole world of understanding opens up. In the past, we didn't even know where all the continents were; maps would say "Here be dragons," and we only knew that when they sailed in that direction, they didn't come back. Host: You also have plans to land on the moon. How is that progressing? Is it still on the agenda? Musk: Yes, I think we want to try to reach new heights of civilization. I feel that a lunar base or this actual moon landing will be an important step in the right direction. I think establishing a lunar base is good, but we should land on the moon to establish a lunar research base. Some parts of the moon may be older than some parts of the Earth, and if we have a scientific base on the moon, we may understand the nature of the universe better. That would be very cool. Then we obviously want to go beyond the moon to Mars and establish a self-sustaining city on Mars. I do believe that humanity's destiny is at a crossroads, and if we can establish a self-sustaining city on Mars, the key test is: if supply ships from Earth stop coming for any reason, will Mars continue to thrive or will it perish? The moment Mars can thrive and develop on its own, the potential lifespan of consciousness will greatly increase because we are no longer reliant on everything going smoothly on Earth. There is always the possibility of self-destruction on Earth, such as world wars, super viruses, or asteroids that could wipe out the dinosaurs. We know from the fossil record that there have been many mass extinction events in history. The question I want to know is: will the arc of civilization continue to rise, allowing us to make Mars self-sustaining before the arc of civilization declines? Because the opportunity window to make life multi-planetary has now appeared for the first time in Earth's 4.5 billion year history. Host: Elon, assuming we get there, and you are there, you would be a senior politician. You would have the moral authority over Mars. How would you govern Mars? Elon Musk: What I want to emphasize is that the self-sufficiency of Mars is more important than the form of governance on Mars or who was there early on. We truly become a multi-planetary species and achieve planetary redundancy. If an extinction event occurs on Earth, the potential lifespan of consciousness would be greatly increased as we become a multi-planetary species. The key test is whether Mars can survive if supply ships stop delivering. The previous Mars missions are not that important. What matters is whether you can deliver enough tonnage to Mars for it to thrive on its own. This means it must have all the elements of civilization. It’s not just about building chip factories or spacecraft manufacturing plants on Mars; it also needs to have the capability to build. Host: Do you have any concept of the time scale? For example, assuming the Starship is in some state around 2026. Musk: Assuming the Starship starts operating in 2026, there will be a lot of testing. Clearly, there will be a lot of early tests, and we only have specific launch windows. Host: (It sounds like) there are a lot of time constraints. In your view, is this a 50-year thing? Or 150 years? Is this something for our generation, or for our children's generation? If conditions are optimal, where do you think that time point is? Musk: If things go in our direction, I think it can be done in 30 years. The premise is that the tonnage delivered to Mars during each Mars transfer window grows exponentially, and transfer windows occur every two years. Every two years, the planets align, and you can transfer to Mars. I think it will take about 15, but it might only take 10, roughly 10 to 15 Mars transfer windows. If the tonnage delivered to Mars during each Mars transfer window grows exponentially, then Mars should be able to achieve self-sufficiency in about 25 years. Host: That's amazing, that's incredible. Alright, ladies and gentlemen, thank you Elon Musk. We will see you again when we get back to town. We miss you, and we will see you in person next time. Musk: Thank you, brother