Musk's latest interview: 10 questions tell you that the first - principles thinking is a superpower.
Some time ago, at the AI Startup School hosted by YC, Garry Tan, the CEO of YC, invited Elon Musk for an online conversation. This conversation was very special. The audience below the stage were a group of very young entrepreneurs, aged between 18 and 25. Some of them have already made a name for themselves in the AI field.
Facing these rising stars, Musk was more candid than ever. The content he talked about was very detailed and down - to - earth. For example, in 1995, he drilled holes in the floor of Zip2 to connect the network cable; in 2001, he flew to Russia to buy intercontinental missiles; in 2008, both SpaceX and Tesla were on the verge of bankruptcy...
Another example is that he talked about how to "simply and crudely" turn the situation around with first - principles thinking, push the cost of rockets to the limit, and build a computing center with 100,000 chips in six months. He also shared various predictions and expectations for AI and the future of humanity.
These contents break the narrative of a "genius entrepreneur" and also show Musk's consistent subversive thinking and pragmatic spirit, which are very inspiring.
We have sorted out and extracted some of the contents and viewpoints for you, hoping to inspire you:
Entrepreneurship is just about making something useful
Garry Tan: Nowadays, entrepreneurs are getting younger and younger. Can you recall the situation when you were 18 or 19, learning programming and even conceiving your first entrepreneurial direction, Zip2? Maybe we can talk about the very beginning.
Musk: Back in 1995, I actually had two choices in front of me: either go to Stanford to pursue a doctorate in materials science and study supercapacitors related to electric vehicle range; or do something in the then - unknown Internet field.
I talked to a materials science professor and asked him, "Can I postpone my enrollment for a quarter?" Thinking that this thing would probably fail and I might have to come back to study. But he said, "This might be our last conversation." And he was right.
In 1995, I wrote some of the earliest maps, navigation systems, online yellow pages, and white pages on the Internet. I did it all by myself. I couldn't afford a server or the cost of a T1 line. The first office was on Sherman Avenue in Palo Alto, and there was an Internet service provider downstairs. So I drilled a hole in the floor and pulled a network cable to the ISP.
At first, I didn't think about starting a business. I even sent my resume to Netscape, but there was no response. I even wandered around in Netscape's lobby trying to socialize, but I was too socially awkward. At that time, I thought, well, why not try writing software by myself. My original intention was just to do something useful on the Internet. Since I couldn't get into an Internet company, I had to start one myself.
The first entrepreneurial lesson: Let technology speak
Musk: We poured our hearts and souls into Zip2 and developed very powerful software technology. But from my perspective, this technology never really played its role. At that time, media companies such as The New York Times and the Hearst Corporation were investors, customers, and also board members. They would look at things from the perspective of traditional media and ask you to do things that seemed reasonable but were incompatible with new technologies. The first entrepreneurial lesson I learned was not to let traditional media companies hold too many shareholder and director seats.
Zip2 was quite successful at that time and was sold for $300 million, which was a huge sum of money back then. Unlike now, there are too many unicorns. The investment in the AI field is very hot. Some companies that have been established for less than a year are valued at $1 billion or even tens of billions of dollars. Some may succeed, but the valuations of some companies are really astonishing.
Directly target consumers
Musk: After exiting Zip2, I got $20 million. At that time, I was sharing a house with four roommates, and I only had $10,000 in my bank account, and then suddenly I got a check. Later, I invested almost all my funds in X.com, leaving all my chips on the table.
At Zip2, we were restricted by our customers, just like having our wings clipped. So what if there were no restrictions and we directly targeted consumers? That's how PayPal was born. X.com and Confinity merged to form PayPal. There were so many geniuses in the team at that time. The PayPal mafia may have spawned the most startups in the 21st century.
SpaceX has done something that only a startup can achieve
Musk: On the Long Island Expressway, my college roommate, Adeo Ressi, asked me what I was going to do after leaving PayPal. I said that I might do some public - service work in space. At that time, I thought it was impossible to do business in space. I was a bit curious about when we could send people to Mars. I searched on NASA's website and found that there was no manned Mars mission.
My initial idea was to send a small greenhouse to Mars. In 2001 and 2002, I did something quite crazy. I went to Russia and told their military high - ups, "I want to buy a few missiles." Since they were going to destroy these missiles due to arms reduction, I thought, how about we take two? Remove the nuclear warheads, add an upper stage, and send it into space. As a result, the negotiation price became more and more outrageous. Later, I realized that it wasn't that humans didn't have the will to go to Mars, but that the existing technology would exceed the budget for a single trip, and even NASA couldn't afford it.
So in 2002, I decided to found SpaceX. The probability of success was less than 10%, maybe only 1%. But if a startup didn't do this, big military - industrial companies definitely wouldn't, because they had to meet the government's needs, and the government only wanted to do traditional projects . So, either a startup would achieve this, or it wouldn't happen at all. A tiny chance is better than no chance at all.
At about the same time, we were also working on Tesla. In the summer of 2008, SpaceX's third launch failed, and Tesla's financing also failed. The company was on the verge of bankruptcy. Fortunately, the fourth launch was successful, but it wasn't enough. We needed a big contract from NASA to survive. It was around December 22nd, before Christmas. NASA suddenly called and said, "We've decided to award you a contract for space - station resupply." I almost blurted out, "I love you guys." At 6 p.m. on December 24th, 2008, in the last hour of the financing deadline, we finally finalized Tesla's financing. 2008 was really a tough year.
Many people fail because "the ratio of ego to ability exceeds 1"
Garry Tan: From PayPal and Zip2 to now, you've always been able to find the right talent. If you went back to the time when you hadn't done any management, what would you say to yourself?
Musk: I think we should try our best to be "useful". "Useful" is a bit like "useful work" in physics. For example, in terms of your final product, if it succeeds, how much value can it bring to how many people? That's what I define as "useful work".
Whether you're a CEO or in any role in a startup, you have to be useful to the company. You must put aside your ego and crush your self - conceit. Many people fail because "the ratio of ego to ability exceeds 1". If this ratio is too high, it will cut off your feedback loop to reality. In the context of AI, it means "breaking the reinforcement - learning loop". If you want to maintain a strong feedback loop, you have to be brave enough to take responsibility and put aside your ego.
So I prefer the word "engineering" to "research". I don't want to call X.ai a laboratory; I just want to call it a company. The simpler and less egotistical the wording, the closer you are to success. You have to keep yourself close to reality.
First - principles thinking applies to all fields
Garry Tan: Everyone regards you as a model of first - principles thinking. How do you use first - principles thinking to build your understanding of reality? How do you teach this way of thinking to others?
Musk: First - principles thinking essentially involves breaking things down into the most basic axiomatic elements and conducting rigorous scientific deductions on this basis, without relying on experience or analogy. First - principles thinking applies to any field, like a superpower.
For example, how much does it cost to build a rocket? With traditional thinking, you might refer to historical data, and the cost of a new rocket you calculate will inevitably be similar to the past. But with first - principles thinking, you'll go back to the raw materials of the rocket: aluminum, copper, carbon fiber, steel, etc. What's the weight of the rocket? How much does each component material weigh? What's the price of each? What you calculate is the actual lower limit of the cost, infinitely close to the cost of the raw materials. As a result, you'll find that without considering recycling technology, the raw - material cost of a rocket may only account for 1% - 2% of the historical cost, which means the manufacturing process is very inefficient.
Another example in the field of AI. Last year, X.AI tried to build a training super - cluster that required 100,000 H100 chips. The suppliers said it would take 18 - 24 months to complete the construction. But if we couldn't finish it within six months, we'd lose our competitiveness. Others said it was impossible, but we broke the problem down into "site, power supply, cooling, and power stabilization" and solved them one by one:
Since we didn't have time to build a new factory, we found a disused factory in Memphis. The original power supply of the factory was 15 megawatts, and we needed 150 megawatts. So we rented generators and placed them on one side of the factory; we also rented a quarter of all the mobile cooling equipment in the United States and placed them on the other side of the factory.
That wasn't all. During training, the voltage and power fluctuated greatly, and they might drop by 50% within 100 milliseconds, which the generators couldn't keep up with. We installed Tesla's giant battery packs, Megapacks, and modified the software to smooth out the fluctuations. Then we worked in four shifts around the clock to complete all the cable wiring. I even slept in the data center and participated in the wiring myself.
High - quality tokens are almost used up
Garry Tan: In the competition of large AI models, is pre - training still effective? Does the scale law still hold?
Musk: Your talent strength, hardware scale, and the ability to use hardware efficiently are all very important. It's not just about buying a bunch of GPUs; you have to train them stably in a coordinated way. What unique data access do you have? How do you distribute it? For example, how do users access your AI? These are all key factors.
Many people believe, and my friend Ilya (former co - founder and chief scientist of OpenAI) also said, that we've almost used up the pre - training data generated by humans. Especially high - quality tokens are almost used up. What we need to do next is essentially to synthesize data and accurately judge the authenticity of the synthesized data. Are they real? Or are they hallucinations that don't match reality? It's very difficult to make AI rooted in reality. We're working on this. For example, Grok 3.5, which we're currently training, focuses very much on reasoning ability.
There will be more humanoid robots than all other robots combined
Musk: In the future, there will be humanoid robots of various sizes and forms. Their number will be an order of magnitude larger than the sum of all other types of robots, and the gap will be very large.
Jensen Huang has brought a large number of humanoid robots from different companies onto the stage, about a dozen types. I was a bit resistant before. In recent years, I've been very slow in the field of AI and humanoid robots because I didn't want to make "The Terminator" a reality. But later, I realized that whether I do it or not, this will happen. You really only have two choices: watch or participate. Of course, I choose the latter.
Human consciousness is like a faint candle flame in the vast darkness
Garry Tan: Looking at the next 10 years, 20 years, or 100 years, how do you view the relationship between AI, robotics, and the multi - planet plan? What drives you to pursue the multi - planet goal?
Musk: I really hope that human civilization will still exist in 100 years. I predict that by then, the number of humanoid robots will be at least five times, maybe ten times, that of humans.
Using the Kardashev scale to measure the civilization index: a Type - I civilization can utilize the energy of an entire planet, a Type - II civilization can utilize the energy of a star, and a Type - III civilization can utilize the energy of a galaxy. Currently, we may only be using 1% - 2% of the Earth's energy, and we're still far from a Type - I civilization. We're in the very early stage of the "intelligence explosion".
I think we can transport enough supplies to Mars within 30 years to make it self - sufficient. Even if the Earth stops supplying, Mars can continue to develop, increasing the probability of human survival.
The Fermi Paradox puzzles me: Why haven't we seen any aliens? Maybe intelligent life is extremely rare, and we might be the only civilization in the Milky Way. Human consciousness is like a faint candle flame in the vast darkness, and we should do our best to prevent it from going out. Becoming a multi - planet species is the first step to reaching other star systems. Once you can set foot on two planets, there will be a driving force for the evolution of space - travel technology, and ultimately spread human consciousness across the stars.
I hope AI can help us understand the essence of the universe
Garry Tan: Where should we go in the future? These things that seem like science fiction may be realized by the people present here. What advice do you have for the smartest technology talents of this generation?
Musk: As I said at the beginning, if you're doing something useful, that's really great. Creating value for others is doing good. For example, focusing on super - realistic artificial intelligence is the most important thing for AI safety. I think a very important point in building AI is to strictly adhere to the truth, even if this truth is politically incorrect.
Where are the aliens? What's the origin and end of the universe? What unknown questions should we ask? Are we living in a simulated world? Ultimately, I hope AI can help us understand the essence of the universe...
Musk's innovation doesn't come from daydreaming but from insights into real - world problems. Whether it's AI, robotics, or interstellar exploration, the most important thing is always to do something useful - create value, stay true, and break through the boundaries of cognition. His sharing reminds us that instead of waiting for the future, we should build it with our own hands. I hope every explorer can find their own direction.
This article is from the WeChat official account "Hundun University" (ID: hundun - university), author: Hundun Academy. It is published by 36Kr with authorization.