Original text of Musk's latest interview: Tesla FSD, humanoid robots, Starlink, Starship launch plan, and Grok

Recently, Musk had a video connection at the All-In Summit, discussing topics such as humanoid robots, AI, and space exploration around the progress of his core businesses and technologies.

During this connection, Musk revealed:

The third - version Optimus robot needs to overcome three major challenges: hand dexterity, an AI brain, and large - scale production.

The performance of Tesla's AI5 chip is 40 times that of the AI4 chip. The FSD v14 software is planned to be pushed in a few months, and vehicles will have perception capabilities by the end of the year.

Starlink is expected to launch mobile phones supporting direct satellite connection in two years. The goal is to become a global independent operator, and the plan to acquire other operators is not excluded.

The Starship is expected to achieve full reusability next year, solving the key technical obstacle of the heat shield; the payload capacity of the third - version Starship will be increased to 100 tons into orbit.

Grok is optimizing data through a large number of reasoning calculations and considering launching the "Grok Encyclopedia"; the Colossus 2 super - cluster is expanding. A rough estimate shows that a 10 - fold increase in model computing power will bring a 1 - fold increase in intelligence.

It is necessary to establish a lunar research base and promote the construction of a self - sufficient city on Mars. Optimistically, the goal of Mars' self - sufficiency can be achieved within 30 years.

Key points from Musk's interview

1. Challenges and prospects of the Optimus robot

The Optimus robot is in the third - version design stage. The core is to solve three major problems: having the dexterity of a human hand, possessing an AI brain capable of exploring and understanding reality, and achieving large - scale production.

Musk believes that the hand design of the Optimus robot is crucial. The human hand has 27 to 28 degrees of freedom and numerous functions. Realizing the functions of a human hand is a problem that must be solved in the design of a general - purpose humanoid robot. Each arm of the robot requires about 26 customized actuators, which the existing supply chain cannot provide. They must be newly designed from the first principles of physics, which is the main bottleneck in hardware R & D. After large - scale production, when the annual output reaches 1 million units, the marginal production cost of the Optimus robot is about $20,000 to $25,000. If the R & D and production are successful, Optimus will become the greatest product of all time.

2. Disruptive upgrade of Tesla's AI chips

The next - generation AI5 inference chip will achieve a 40 - fold comprehensive performance improvement compared with AI4. It can efficiently process mixed - precision models and significantly reduce the steps of key operations (such as the Softmax function).

The AI4 chip currently used in cars will make autonomous driving at least 2 - 3 times safer than human driving, and may even reach 10 times. Version 14 will be pushed via OTA in the next few months. It is the biggest upgrade since version 12, with an order - of - magnitude increase in the number of parameters and the use of reinforcement learning. The team is effectively solving the problem of lossy compression of reality by AI. Musk expects that by the end of this year, vehicles will have perception capabilities.

3. Starlink's direct - to - phone plan

SpaceX will use new spectra to achieve high - bandwidth communication between satellites and mobile phones. Mobile phones need hardware support for new frequency bands. The phones are expected to be launched in two years, and dedicated satellites will be built simultaneously to achieve the goal of watching videos on mobile phones anywhere in the world.

Musk's vision is to make Starlink a global independent operator, providing a comprehensive solution of "home antenna + direct phone connection". He will consider acquiring operators to obtain spectrum resources, including Verizon, the largest communication operator in the United States, but emphasizes that he will not squeeze out existing operators.

4. Progress of the Starship and the breakthrough of reuse technology

The Starship will achieve full reusability next year. The third - version Starship will be equipped with Raptor 3 engines, and its payload capacity will be increased to 100 tons into orbit, 2.5 times higher than the current most powerful commercial rocket.

The key technical obstacle to achieving full reusability of the Starship is the heat shield, which requires solving problems in materials science and engineering. It is necessary not only to manufacture a fully reusable orbital heat shield that can withstand high temperatures, is lightweight, does not conduct heat, has stable tiles and does not dissolve in rain, but also to ensure the reliable attachment of hundreds of thousands of heat tiles without maintenance.

Musk plans to establish a lunar research base and hopes to build a self - sufficient city on Mars. The key indicator is that Mars can still thrive after cutting off its dependence on Earth. If the Starship progresses smoothly, with the payload increasing exponentially at each two - year Mars transfer window, it is highly likely that Mars can achieve self - sufficiency in 25 to 30 years.

5. Development of Grok AI and general artificial intelligence

Grok is correcting and enhancing training data through reasoning calculations, deleting false information and supplementing background information. Musk will discuss with the team the possibility of launching the "Grok Encyclopedia".

The team is expanding the Colossus 2 super - cluster. Musk believes that there is a natural logarithm relationship between model computing power and intelligence. A 10 - fold increase in computing power may double the intelligence. It is expected that next year, there may be an AI that surpasses humans in a single task, and around 2030, AI may exceed the total intelligence of all humans.

Original text of the interview

Host: Hello. Where are you?

Musk: Palo Alto

Host: You're in Palo Alto.

Musk: Yes. I'm at Tesla's global engineering headquarters in Palo Alto.

Host: How many R & D cycles are you devoting specifically to the Optimus robot? How's the progress? What's the time schedule? I think you're on the third version, maybe the fourth.

Musk: No, everything takes a long time. We're finalizing the design of the third - version Optimus, and it'll be a really amazing robot. It'll basically have the dexterity of a human hand, which means it'll have a very complex hand, and an AI brain that can explore and understand reality, and it can be mass - produced. These are the three things that Optimus is missing, and they're really difficult. I think these will take more cycles than any single other aspect of Optimus, like getting the AI closer to the real world, all the electromechanical issues of Optimus, and its supply - chain and production challenges. Because there's currently no supply chain for humanoid robots, we have to create it from scratch, which requires a lot of vertical integration. None of Optimus' actuators can be obtained from the existing supply chain. But I think it's fair to say that if successful, Optimus will be the most influential product of all time.

Host: What about the cost after scaling up? $20,000, $30,000, $40,000 per robot? What do you think the cost of the first batch will be, and when can we buy one to work on the ranch?

Musk: I think the marginal production cost will be around $20,000 once you reach one million units per year. It depends largely on how much you spend on the AI chip in the robot. There also needs to be a lot of efficiency improvements in the actuators. Each arm of Optimus has about 26 actuators, like 26 electric motors, gearboxes, and power electronics. That AI chip will be quite expensive, maybe $5,000 or $6,000 per robot, maybe more. But I think in terms of volume, at one million units per year, the production cost will be around $20,000, maybe $25,000, around that range. And the price will depend on demand.

Host: Elon, can you explain to everyone why getting the robot's hand right is so important? And why is the actuator design so unique? Why is it so difficult? Why hasn't anyone been able to do it? And why do you almost have to start there to build the rest of the robot correctly?

Musk: It turns out that human hands are incredible. They've evolved into such complex machines. Your hand is actually a really amazing thing. Look closely at your hand and think about all the things you can do with it. There are a lot. Your hand is a very versatile tool.

Host: Yes. You can high - five.

Musk: Very versatile. You can swing a baseball bat, you can thread a needle, you can put a thread through a needle hole, you can play the piano, the violin, you can disassemble or assemble a car. The hand is an extremely versatile tool. And most of the muscles in the hand are actually in the forearm. So most of your hand is a bit like a puppet, with these muscles coming from the forearm and pulling tendons. The design of human tendons, or the evolution of human tendons, is extremely excellent. So you have this network of tendons. A human hand has about, depending on how you count, 27 or 28 degrees of freedom. It's amazing. So, to create a general - purpose humanoid robot, the problem of the hand, the hand challenge, must be solved.

Host: Is it a bit like when you first founded Tesla, when the supply chain didn't exist, and you had to go out and find partners, and you know, build all this vertical integration for support? It's really that you couldn't find anything, and you had to build from scratch.

Musk: Yes, we can't buy those actuators no matter how much money we spend. They simply don't exist. Even though there are thousands of different sizes and shapes of electric motors on the market. We've had to design the gearboxes for each electric motor and the control electronics from scratch, basically starting from the first principles of physics.

Host: And the good news is you've had a lot of experience in factories over the years.

Musk: Yes.

Host: How much more difficult is this than the Cybertruck, Model Y, Model X, and Tesla's Gigafactory in Nevada?

Musk: It's more difficult than any of those.

Host: Okay. Much more difficult? How about compared to the Starship?

Musk: No, the Starship is more difficult.

Host: So it's somewhere between the Model X and the Starship?

Musk: Yes.

Host: Which is more difficult, hardware or software?

Musk: Currently, we're struggling with the final design of the hardware. As I said, it's really mainly the hand. Not to downplay the rest of the robot, which is also very important, but together with the forearm, it makes up the majority of the engineering difficulty of the whole robot.

Host: And let's assume you've overcome the hardware challenges. Based on all the progress of large - language models (LLMs) currently, how much can we get? You know, will consumers be able to interact with it, talk to the robot, and make it do things.

Musk: That'll be no problem.

Host: You've spent a lot of time with Ani. I noticed that online.

Musk: Not that long. Maybe I over - hyped Ani when promoting Grok Imagine. You might get a physical embodiment of Ani

Host: Why the humanoid form, Elon? You could make something that might be better at certain tasks than humans, or something simpler for specific tasks, and maybe even better and more capable than humans at what it can do. How did you decide to make it look exactly like a human?

Musk: If you want it to do everything a human can do, it turns out you need a humanoid robot. So if you want to do a subset, it's much easier. But it turns out that humans have evolved into this form and with these capabilities for good reasons, like having five fingers. That's valuable. Four fingers and a thumb, and even the little finger is actually quite useful. Toes are more of a question mark.

Host: Humans also designed the world for ourselves. If you can build a humanoid robot, it'll be immediately backward - compatible with the world we've built.

Musk: That's exactly right.

Host: Elon, there's also other stuff with the robot. So there are large - language models (LLMs), there's the hand actuators, but there's also the silicon chip that runs it, and you know, Tesla's Dojo supercomputer. I remember you posted on X about the AI5 and AI6 chips. It seems you're very excited about the direction of silicon - layer development. Can you tell us what that is and what we're building here? Is it a supplement to everything that exists in the world? Could it potentially be a long - term competitor?

Musk: At Tesla, we basically have two different chip projects. One is for Dojo, which is for training, and then there's what we call the AI4. This is our inference chip. The AI4 is being deployed in all vehicles, and we're finalizing the design of the AI5, which will be a huge leap from the AI4. By some metrics, the AI5 will be 40 times better than the AI4. This is because we're closely collaborating on a very fine - grained level between the AI software and AI hardware teams. So we know exactly where the limitations are, and as a result, the AI hardware and software teams are actually co - designing the chip.

Host: So a 40 - fold improvement in the silicon chip. I think, as everyone here has experienced, is this just an order - of - magnitude improvement in the quality of Full Self - Driving (FSD) and the safety experienced as a Tesla driver? What about the quality of the robot? Where will these improvements ultimately show up?

Musk: Yes, exactly. The 40 - fold improvement is, if you take, for example, the worst limitation of AI in running the Softmax function. Currently, in simulation mode, we need about 40 steps to run the Softmax function, while the AI5 will do it in just a few steps. The AI5 should also be able to handle