Today, I listened to Elon Musk's interview at the All-In Summit 2024, especially the part about artificial intelligence (AI) left a deep impression on me. Coincidentally, I've been reading Yuval Noah Harari's "Homo Deus," where Harari mentions many concerns about the future, particularly the potential social impacts of AI. However, Elon seems to have a more optimistic view of AI.
Castbox podcast link 🔗: https://castbox.fm/episode/id6075589-id735784004
Part:
Host: Do you have any ideas about the main application areas for AI? Will it be used in autonomous driving? Or for robotics? Or perhaps transforming entire industries? I think that currently, the significant impact of AI on business is still in its early stages. What are your thoughts on this?
Elon: I think it's undeniable that right now the spending on AI probably far exceeds the revenue. But the pace of advancement in AI is faster than any technology I've ever seen by a very significant margin. For example, the Turing Test used to be this benchmark, and now you can run an open-source random large language model on a Raspberry Pi, and it might pass the Turing Test. So, I think the good future with AI will lead to tremendous prosperity. It will be an era of abundant resources without shortages of goods and services, where everyone gets what they want except for things we artificially define as scarce, like certain artworks. But for any manufactured goods or provided services, I think with the progress of AI and robotics, the cost will approach zero. I'm not saying they'll be completely free, but the costs will be so low that almost everyone can have whatever they want. It's a great future. From my perspective, there's about an 80% chance of such a great future. There's a 20% chance it leads to destructive consequences.
Host: What does that 20% look like?
Elon: I don't know, honestly. I have to deliberately put aside my concerns about AI to sleep well. Even so, I think the real issue may be how we find meaning in a world where AI can do everything better than us. That might be the bigger challenge. However, more and more people are enjoying life after retirement, so maybe there will be a crisis of meaning in the future because computers can do everything we do better. But that might just be one challenge. In fact, what you need are terminal factors - autonomous cars and general-purpose humanoid robots. Once you have general-purpose humanoid robots and autonomous vehicles, you can actually build anything. I think the scale of the economy is actually unlimited. Obviously, the mass of the Earth is a limit, but the essence of the economy is the average productivity per person multiplied by the number of people. If you have humanoid robots and there's no practical limit to the number of humanoid robots, and they can work very intelligently, then there's no actual limit to the scale of the economy.
Host: The Colossus you just launched is currently the largest private GPU computing cluster, right?
Elon: Yes, it's the most powerful supercomputer of any type.
Host: This confirms what David and Peter said, that the economic value of AI today is almost entirely going to Nvidia. But there are some alternatives, and your company is an example because Dojo is mainly used for processing images and large amounts of video.
Elon: Yes, Tesla's problem is different from the issues with large language models; it requires a different kind of intelligence. As you mentioned earlier, Tesla’s challenge lies in its need to process very long context lengths. Our context data volume is extremely large. You know, Tesla's system has seven cameras, and if you have a few minutes of HD video, the data size is measured in GBs. So Tesla's problem is how to compress this massive context into truly important pixel data. Not only do you need to compress pixels in the spatial dimension, but also in the temporal dimension. Then, you must complete the inference on a relatively small computing device, such as Tesla's AI inference chip, which consumes only a few hundred watts. This inference chip is still the best available; we can't buy better ones from suppliers. Tesla's AI design team is excellent.
Host: But one day, perhaps other companies that need to process videos or images could use this excess computing power, making you a competitor to Nvidia, even though it wasn’t specifically intended.
Elon: Yes, right now we have two projects: Dojo is the training computer, and the other is the inference chip we use in every car. Currently, we only have Dojo 1, but by the end of next year, we should be able to roll out Dojo 2 at scale. We believe Dojo 2 will perform comparably to B200-type training systems. In the future, it may also be offered as a service to other companies. Although my confidence in Dojo has increased, we might not truly know how good it is until Dojo 3. Generally speaking, technology needs three major iterations to reach an excellent level. We will only have our second major iteration next year. The third iteration might come around the end of 2026.
Host: How is the Optimus project progressing? I remember the last time we discussed it, you publicly mentioned that it was undergoing some lightweight testing within the factory, so it has already started to play a role. What about its material cost? For example, if it were mass-produced like the Model 3, and the production line could produce one million units per year, would its cost be $20,000, $30,000, or $40,000?
Elon: I’ve come to the conclusion that as long as you have a large enough production volume, the cost of anything asymptotically approaches the material cost. Of course, I should say some things are limited by intellectual property costs, like patent royalties, so a big chunk of the cost of chips is paying for the copyright and then the depreciation of the chip factory. But the marginal cost of chips is actually very low. Optimus is obviously humanoid robot, it's much lighter than a car and much smaller in volume. So you can expect that at high volumes, I would say... You probably need three more production versions of iteration. You need at least three major design improvements before you get to scale production of over a million units per year. And I think by that time, the labor and material cost of Optimus will not be more than ten thousand dollars.
Host: That could be a decade-long process.
Elon: Basically, you can think of Optimus as costing less than a compact car. So after three major technical iterations, when scaling up production, if a compact car costs twenty-five thousand dollars, Optimus might cost around twenty thousand dollars. It’s a humanoid robot that can serve you like R2D2 and C3PO. Honestly, I think people will become very dependent on their humanoid robots. Just like in Star Wars, everyone loves R2D2 and C3PO. They're great and have personalities. All R2D2 can do is beep, it can't speak English, and C3PO translates those beeps.
Host: If you iterate every two to three years, it might take a decade to reach scale.
Elon: I would say each major iteration cycle is less than two years. It will probably take about five years, maybe six, to reach a production volume of one million units per year.
Host: By that point, everyone on Earth can afford an android. Eventually, if we're sitting here in 30 years, what do you think the ratio of robots to humans will be on Earth?
Elon: I think the number of robots will vastly exceed the number of people, vastly exceed. You've got to admit, who wouldn't want a robot companion? Everyone would want a robot companion. Especially if it could take care of your dog, walk it, it could mow your lawn, look after your kids, even teach them.
Host: We could also send it to Mars, send a lot of robots to Mars, to do the work that needs to be done, help make Mars a habitable planet.
Elon: Mars is already a robot planet, there are lots of robots there, like rovers and helicopters.
Host: Only robots.
Elon: Yes, Mars will be full of robots. I think the opportunity for humanoid robots is the biggest opportunity ever. Because if you assume... I think the ratio of humanoid robots to humans will be at least 2 to 1, maybe 3 to 1. Everyone will want a robot, and then there are all the robots that you don't see making goods and providing services.
Host: Do you think it will be a general-purpose robot that learns different tasks?
Elon: Yes, after all, we humans are also general-purpose beings. We're just made of meat.
Host: Yes, we are "meat" general-purpose robots.
Elon: I'm "operating" my meat robot. By the way, as we design Optimus, we are learning more and more about why humans have this form. For example, why we have five fingers, why the pinky is smaller than the index finger. Of course, we know why we have opposable thumbs, but we are also learning why the main muscles of the hand are actually in the forearm. The movement of our fingers is mainly controlled by the muscles in the forearm. Most of the strength in the fingers comes from the forearm; the fingers operate through tendons — which are like little strings. In the current design of Optimus's hands, the actuators are located in the hand, with only 11 degrees of freedom, so it cannot be as flexible as a human hand, which has about 25 degrees of freedom (depending on how you calculate it). And the hand lacks sufficient strength because the actuators must fit within the size constraints of the hand. A prototype for the next generation of Optimus hand design already exists, where the actuators are moved to the forearm, just like in humans, and the fingers are operated via cables. This next-generation hand design will have 22 degrees of freedom, and we believe this is enough to perform almost any task that a human can do.
Host: It's said that X and Tesla might collaborate to offer some services. My first reaction is, if you equip robots with Grok, they will have personalities and be able to handle voice, video, and images.
Elon: Yes.