Elon Musk: AI, Space, X, Mars, speed and hardcore

I'd love to kick off with AI. Now, what's your take on where we are in the AI race just now?

There's so much happening, and AI is the fastest advancing technology that I've ever seen of any kind. And I've seen a lot of technology. Barely a week goes by without some new announcement.

So if you look at the amount of AI hardware, the computers coming online that are dedicated to AI, that is increasing what looks like at least by a factor of 10 every year, if not every six to nine months. So when you combine the hardware coming online, really, order of magnitude increase every, call it at least every nine months.

And many, many software breakthroughs. If you look at that curve, it looks insane. So I think we'll,

my guess is that we also will have AI that is smarter than any one human, probably around the end of next year.

And then AI, the total amount of sort of sentient compute of AI, I think, will probably exceed all humans in five years.

Is it the supply of chips? What is it?

And the hardware deployment, to break it down into the three areas of people, data and hardware, starting with hardware. Last year it was about chip supply. People could not get enough, Nvidia chips, particularly.

This year it's starting to transition to a voltage transformer supply. So just actually getting it up, voltage transformers put in place. So my sort of very niche joke is transformers for transformers, because a lot of the AI that's run is called a transformer. So you need transformers to run transformers. And then next, if you look out a year or two, or certainly three years, it's just electricity availability.

So those constraints on the hardware side.

So many of the world's smartest people are doing AI. People that would have done physics before, in fact, or have done physics, for example, have moved into AI because it's just the fastest moving field. So we're seeing a lot of the best talents, a lot of the smartest humans going into AI. And then we see, along with that, algorithmic breakthroughs.

And then you start hitting the wall with the data problem.

So you can fit all books ever written,

just the text, the text in compressed form on one hard drive, or we'll call it one computer.

So when you're looking at, like, tokens to train on,

and you, because you still think of, like, all the books ever written in every, in all languages by all humans. Sounds like a lot. Certainly, it's far more than any one human could ever read. It actually is a small number of training tokens. It's just not enough.

So then, you start having to look at all the videos ever created,

all the podcasts, everything.

And you start even running out of data there. Well, hopefully they will include this podcast. Definitely will include this podcast. What's the biggest challenge you have with XAI?

Well, XAI is still relatively new. So it's not, like,

the limiting factor right now is just training our Grok version 2 model, which should be, we think, better than GPT-4. And that's, we're hoping to complete that in May. So that's the training right now. So it's just really, we're just trying to get enough GPUs online to train it fast enough. And we're hoping to to get that done in May, which I think probably will happen.

And then, and that's with roughly 20,000 H100s. And doing, I think, very efficient training, then the next step would be for Grok 3, which would be, I guess, GPT-5 or beyond, would, you know, requires 100,000 NVIDIA H100s training coherently.

So that's, you know, a half order of magnitude, basically, more training. And then you really start to have running into this data problem where you have to either create synthetic data or use real-world video. So the two sources of kind of, like, unlimited data are synthetic data and real-world video, which I should say Tesla has a pretty big advantage. in real-world video. Tesla has by far the most real-world video of anyone.

But there's still a pretty big leap beyond that. to say, smarter than the machine-augmented human collective. So, like, is it smarter than all humans working together who are also using computers to augment their output? And that, I think, is probably five years away. One way to look at it is to try to assess, like, roughly what is the ratio of digital to biological compute?

the rate at which AI is progressing is faster than probably any regulatory agency can keep up with.

But I do have a comment on what I think is very important for.

achieving safe AI, which is that it's very important to train the AI to be as truthful as possible and not to,

yeah, just to be as truthful as possible.

I think you can get some very dangerous things when you program an AI to be politically correct. Things that may seem relatively innocuous now, but will not be so in the future. if AI has immense power. You can take the Google Gemini example, where it refused to publish, to produce a picture of George Washington as a white man. And any, in fact, any historical figure would automatically be made diverse because it's been programmed to insist on diversity, which sounds, you know, perhaps okay at first, but not if the AI has so much power that it can actually enforce diversity and decide there's too many of one kind of people or too many of one sex and kill off, just kill off enough until the diversity number is what it's programmed to believe is correct.

I mean, the AI right now is

very concentrated in San Francisco and London.

And then, you know, there's this, you know, a lot happening in China, but I don't have insight into what they're doing, except that they will, I'm confident they will not be far behind what is developed in the West.

So,

but mark my words, if we do not program an AI to be as truthful as possible, that is where it will go awry. That is where the danger lies.

So much appreciated.

I think it's, we will, that that electric, that all vehicles will go fully electric. It's only a matter of time. That includes aircraft, ultimately, and boats, obviously trains. The only thing that is ironically difficult to, or you can't really make it electric is rockets, because you need, you can't get away from having to expel mass. Sort of Newton's third law.

But all cars will be electric. It's only a matter of time. And we'll look back on combustion cars in the same way that we look back on steam engines.

That I was, it was inevitable that there would be internal combustion cars, and it's just as inevitable that all cars will go electric.

And there will be some, you know, it's not going to be a completely straight up line. There'll be some ebb and flow in how far electric cars go. But that, but the ultimate victory of electric cars is inevitable. And I think the sooner we get there,

S. or very few are available in the U.S. Some are available in Europe.

But from what my team tells me, they are very good.

So you can say like, well,

I think it's probably on the order of a million tons, maybe, maybe more, but somewhere between probably a million tons and 10 million tons are needed to make Mars.

really, if you have a hundred tons per flight, you need 10,000 flights to get to a million, million tons. And that's a hundred tons landed to the surface of Mars. So in order to get a hundred tons landed to the surface of Mars, you need 500, five times that number in Earth orbit.

So we do a lot of orbital refilling. So launching sort of rockets, tanker ships over and over again, that would replenish the propellant of the ships that would go to Mars. And then you'd need roughly an order of 10,000 of them to get to a million tons.

But we plan to do that.

We think we can get that done within 20 years.

And then we would probably launch several ships and just confirm that they can land okay on Mars. We'll also be doing the Moon simultaneously with that. So I think we'll get people back to the Moon, I should say, within five years. And we'll get uncrewed ships landed on Mars within five years. And then we'll be building up the production rate and improving the design of the booster of the ship.

So the first people on Mars, I think, within seven years or so, seven to nine years.

And from there, we need to rapidly increase. We need massive numbers of ships going. And Earth and Mars only are in the same quadrant of the solar system, roughly for six months every two years, or at least it's only possible to really transfer efficiently from Earth to Mars, say, every six months. But really, there's about a couple of months where it's ideal, every 26 months. So every two years, you would see basically a fleet depart Mars.

I think it would be quite a spectacular thing to see 1,000 ships depart from Mars, all at once, like Battlestar Galactica. What kind of new technology do we need before we'll be self-sufficient here?

Actually, I think we already know all the technology that's necessary for that. We just need to build. So no new physics is needed for this. Why is it so important for you?

I think it's important for consciousness in general.

So, if we wish to maximize the lifespan of consciousness, then being a multi-planet species will result in a much longer existence of consciousness than if we are on one planet. If we're on one planet, we're simply fighting our time until there's eventually a calamity. It could be soon, it could be a long time, but eventually something will happen. It could be a global thermonuclear war. It could be simply that civilization merely subsides.

Our civilization may not die with a bang, it may die with a whimper, just gradually falling into obsolescence. But if we're a multi-planet species, then we've got two planets, and they can support each other. And we can go beyond two planets, ultimately to the moons of Jupiter, beyond to the outer parts of the solar system, and ultimately to other star systems. So this tiny candle of consciousness that we have in this vast darkness can be extended and amplified, and we're just far more likely to survive, for consciousness to.

Well, just to put this in perspective,

So I really can think of it as resource allocation. Do you think it's worth spending half a percent of Earth's resources to ensure that we have redundancy in consciousness, and that we extend consciousness beyond Mars to other planets, to Mars and other planets, and ultimately other star systems? And then also take into account the fact that there are certain things we simply cannot avoid on Earth. Like, is it within your power or mine to stop World War III? I don't think so, if it happens.

And if we have global thermonuclear warfare, our technology level will drop to the Stone Age,

and we may never survive. And then we may get hit by a comet, like the dinosaurs, and if the dinosaurs had spaceships, they'd probably still be around.

And then, if you wait long enough, the Sun will continue to expand, and eventually engulf Earth and destroy it, and destroy all life. So just given a certain amount of time, no matter what you do on Earth, no matter how careful you are, Earth will, all life on Earth will die. That it will happen is a certainty.

And I think we're making good progress there. I mean, I'd call it everything. Anything you want to do, you can do on the X platform. Whether it's text, audio, video, payments, financial stuff,

communications of all kinds. But then also, where there's publicly disseminated information, just to be the best source of truth. And I think it already is that. People may say, there's some piece of misinformation. I say yes, but look at the replies.

The replies correct that misinformation. And look at community notes, and how good the batting average of community notes is. It's extremely good. It's by far the best fact-checking system on the internet.

So, and a lot of people still labor under the illusion that the legacy newspapers that they read are actually true. There's so much nonsense in them. I mean, how many times do you read an article in a newspaper where you know the circumstances of what that article is? And?

So we see very little.

We do see a lot of attempts to influence things, but they seem to be coming from the West,

And there would be to suspend accounts. Immediately, we're given two hours to suspend an account or face massive fines. And the final straw is we were being given demands to suspend sitting members of the parliament and major journalists. And, moreover, we could not tell them that this was at the behest of Alexander Morales. We had to pretend that it was due to our rules of service.

And that was the final straw. And we said no.

Well, I knew it wouldn't be just a total bed of.

We're just rigorously trying to pursue the goal of being the most accurate and truthful place on the internet. And that doesn't mean that what is said is always true or accurate, but it is perhaps another way to frame it as the least inaccurate place on the internet.

But overall, we're trying to serve the people of Earth. And this is sort of an esoteric way of doing it. But to try to be kind of like the group consciousness of Earth. So you can think if each person is like a neuron contributing to the collective brain of Earth, and you want to try to minimize the noise and maximize the signal of every neuron that's connected to the X network. That's basically what is the collective will of humanity?

And how to serve the collective will of humanity and serve the greater good. That's our goal.

Now, there's definitely going to be people who want to manipulate that information and try to have it be the most accurate place, as far as the best of our ability, and have it be kind of a marketplace of ideas where people can propose ideas and debate them. And I think so far, it's working reasonably well in that regard. People that don't like the truth will not like X, or if they want to manipulate.

You're becoming like a real geopolitical force, and a really important one. How do you look at that?

We have to keep civilization going onward and upward as much as possible, and try to minimize the civilizational threats that occur.

We can't get to Mars if civilization collapses. It's not going to happen.

We've got to keep civilization going. And I think we should view our civilization as being much more fragile than we think. We kind of take for granted, oh, it's always going to be there. But actually, if you study history, you realize that there's a rise and fall of civilizations.

I was reading in depth about the ancient Sumerians,

who were arguably the first civilization, if you call civilization like writing and stuff. They were the first to develop writing.

But eventually, they died out, and they were gone.

And then nobody could read their writing at all, and they just faded out as a civilization. But they're pretty impressive in their time. And the ancient Egyptians, the same thing.

And sort of one after another, ancient Greece had its day. China and India will have incredibly impressive populations. But there's been ebbs and flows in China and Indian civilizations over the aeons, the millennia, as well. So, I guess I'm just trying to take the steps that increase the scope and scale of consciousness. That's what I'm trying to do.

It's not that I'm trying to put a political thumb on the scale or anything like that. But I think I'm trying to have the political world go where the people want it to go.

I don't think I manage smart people.

I guess with really smart people, I don't really think of it like managing them. I think that if somebody is very smart and talented, they can go anywhere and do anything, anytime. They don't have to work with me. They could go anywhere. So, I really just say, look, this is the goal we're after, and this is what we're trying to achieve.

Do you agree with this goal? And if you do, then let's try to get it done and provide my opinion along the way. And once in a while, I'll say, look, guys, you've just got to trust me on this one. We've got to do this thing. And if it turns out to be a bad decision, we can all hold that against me in the.

It's just insisting on attention to detail.

If you're trying to make a perfect product, you must have attention to detail. And I haven't actually read the i6 book.

well, I'll ask you some questions from the book then. He talks about the hardcore and ultra-hardcore culture. What is an ultra-hardcore culture? I mean, it's working culture, right? I mean, ultra-hard work.

How hard is that?

And every waking hour, maybe, is an exaggeration, because there are a few hours, obviously with friends and family and critical other things. But 100-hour weeks would be—I've done many, many stretches of 100-hour weeks, like true 100-hour weeks, where roughly six hours per day of sleeping. I would not recommend that. That's for emergencies. It's not all the time.

During very difficult times at Tesla, I've had to do that. And sometimes at the beginning of my earlier startups, I did that, where I just wouldn't leave the office. I would just sleep under my desk and just work seven days a week.

Sometimes it's necessary for success or to avoid.

But do you enjoy being in this crisis mode?

No, I don't. It sucks.

But sometimes it's the difference between success and failure.

Sounds good.

And really, what it did was just go faster. So the power of speed is.

underappreciated as a competitive dimension.

And fast?

How do you know?

No company is going to be going in the right direction all the time, so you have to do course corrections, like a guided missile, you call it course corrections.

But in the case of SpaceX, it's like, okay, our goal is to extend humanity beyond Earth.

And we didn't even know how to frame the question correctly. We just knew that that was the general goal.

We didn't know what propellant we'd use or what the raw materials would be, or how the rocket would be built, how it would be designed, what's actually important.

So, for example, going from our Falcon architecture, which uses refined jet fuel and liquid oxygen in an open cycle gas generator architecture engine, to Starship, which is a liquid methane, liquid oxygen.

propellant in a staged combustion, very high pressure engine.

That's a big architectural change. But we didn't know that we needed to make that architectural change until we were pretty far down the road. Like about halfway, it took us about 10 years to figure out that was even the right architecture. Now we're confident it is.

Well, I mean, which ones do you tolerate and which ones don't you tolerate?

The first three flights of SpaceX failed. The fourth one succeeded.

And if the fourth one had not succeeded, we would have gone bankrupt. We would have had no money. so it was a very close call.

But since then, SpaceX has done very well. It's now the Falcon 9.. Knock on wood, it's the most reliable rocket in the world and launches about every two to three days.

And then Model 3 is sort of half a million per year.

Model Y, over a million per year. So these are all bet the company vehicles. But the reason we could do, for example, a Sabat truck, which was kind of a radical new design, was because it wasn't a bet, the company decision. So I was like, OK, look, let's try something. I want to try something totally crazy.

It's like, what truck would Blade Runner drive?

But actually, by the way, the Sabat truck is doing great.

So one of the things that I think is important for innovation is that you do accept failure.

You have to always look at the incentive structure of an organization and say, is that organization properly incenting innovation? If you do innovation, you're necessarily going to uncharted territory. So there are going to be some mistakes. There are going to be some failures.

Actually,

for SpaceX rocket engine development, I keep telling the team, look, if we're not occasionally blowing up an engine on the test stand, we're not trying hard enough.

Have that learning and wisdom been helpful when you have been planning your companies?

It's hard to say exactly what I've learned from video games, except that I do like playing video games. If I want to take my mind off work, I will typically play a very hard video game. Such as? Which one? Well, over the years, there's been many, many different video games.

When I was a little kid, it was like Pong and the little tank games and things. But if you take a game like, for example, Civilization, it's actually quite a good...

It tells you how civilizations are formed. I remember playing the original Civilization with a technology tree, and how you'd invent different things. You'd invent literacy, and

invent democracy, and invent gunpowder. All these things. And you start to realize, oh wow, there are stages to technology. You can't actually get to democracy without literacy.

These stages of technology development or stages of ideas, that's a helpful framework for a company.

In recent years, there was a game I played that was developed in Sweden called Polytopia, which is actually quite a good game. A lot of people like playing chess, but I think chess is not a great... There's not a lot of transfer learning from chess to the real world, because in chess, you've got only 64 squares. It's a set-piece. battle, same pieces every time.

There are no terrain differences, there's no technology tree, there's no fog of war. But a game like Polytopia has all of those things. Random terrain generation,

differences in attack and defense bonuses, depending on what type of terrain. You've got 16 tribes, I think, each with different abilities. You've got a technology tree that you can choose to develop in different ways, and you've got, of course, fog of war. That, I think, is much closer to reality.

I was playing Diablo for a while, pretty fun.

Diablo, at high levels, gets very complicated. You could call it a spreadsheet with a game attached.

I briefly got the, for about a day, the world record in this abattoir of the year, on a four-person team of clearing the hardest level, which was not bad for someone who's like 53.

. Basically, we'll be 53 soon.

There is still some Twitch element to it, and it's hard to beat kids at games with a Twitch element.

But yeah, I find these games interesting. if you can,

Okay, I can hear you here. Okay, good. Sounds good.

Let's see. Thumbs up. if you can hear me. Let's try again. Okay, sounds good.

So, yeah, I think things are good.

Overall, yeah, I feel pretty good about the future. I mean, there's going to be bumpy quarters from here and there, but I think the long-term future of Tesla is extremely strong. For example.

Hello,

hello, hello.

So, just trying to have the slide of consciousness last as long as possible, and maybe understand more about the nature of the universe, or simulation, or whatever this is.

So, I have a philosophy of curiosity, which is to understand the universe, understand the nature of the universe, or even what questions to ask.

Kind of like, I would subscribe to the Douglas Adams, Hitchhiker's Guide to the Galaxy School of Philosophy, where we're trying to understand what questions to ask about the answer, that is, the universe.