The $10B Satellite Empire Putting AI in Orbit, Why Chips Beat Rockets & China's #1 Open Model | EP #266

26:42Now everyone knows that they're seen at every step. If you hit a school, we're going to see the school. If you hit a bridge, we're going to see the bridge. And the accountability is going to be there for the whole world to see, no matter what. And that world, I think, acts as deterrent. In history, throughout history, wars have happened mainly when there's been misinformation, a lack of information, and people have had to guess or made mistakes based on misinformation. Here, we have more people understanding what's going on, who's got what equipment, where does that... and they can monitor peace accords and all that.

27:20I think transparency drives accountability and reduces the probability of war. Meanwhile, on this note, yeah. Very narrow question on that exact topic. You know, you assume that the U.S. and China see everything via satellite at all times. But if you look at the 220 countries across the world, what fraction of all governments actually have satellite coverage data? Like you said, misinformation leads to confusion, leads to war. like in Yemen right now, do they look at data or not? Not much, but I think that that's going to change. Again, the challenge has been digesting 40 terabytes of data every day is too much for most organizations.

28:00NASA and NOO, they have teams of people doing satellite imagery processing. They know how to deal with this. Bring AI along and suddenly you eliminate all of that. You can actually get most of the answers very quickly. So that NGO, the Red Cross operating in Yemen or whatever, can actually get benefit from this right now and enable them to make smarter decisions. So it changes it from a world where it's just the big entities to a lot of other entities can get value. What percentage of your revenue is government versus corporate versus individuals? So it's about 60 percent defense and intelligence, about 30, 25 percent civil government and about 15 percent commercial.

28:46So basically mainly government and it has been growing there. But commercial is now really starting to take off. And again, it's because of all those reasons that AI is lowering the barriers to entry. How are you going to price the AI training data use case? I mean, that's the big up and comer, obviously. Well, you know, I mean, frankly, they're going to keep it and sell the knowledge.

29:16We'll just continue to sell the data. So if OpenAI calls and says we want to we want to use it, you're going to say no. Or are you going to say no, it's a billion dollars? OpenAI can call our MCP server and off you go. OK, so it's just fixed price. Absolutely. Train all you want. And then no recurring. And every time you need plant data, you're going to need to do an API call. I am curious about maybe to Steelman Salim's earlier question about information asymmetries. Will, correct me if I'm wrong, but am I correct in assuming that your data sets go through some sort of U.S. government NRO filter regarding what can be made publicly available?

29:56Not exactly. It's a bit more nuanced than that. As a U.S. remote sensing company, we register under the NOAA's Remote Sensing Act, which means we have to register the satellites. But we can sell the data to anyone except for a blacklist, a blacklist that includes Iran, North Korea, what have you, and various terrorist organizations. But other than that, they're not checking every player we provide to. We do check that. We think, and there's many people we don't work with if we think that they reduce some harm with this. But essentially, they're relatively hands off with that. So you're not required.

30:40That's a great line of questions. So that's a U.S. blacklist, right? But you have like a billion dollar plus deal with Sweden. Do they also have a blacklist that you also honor separately? Or do you just deal with the U.S. list? I mean, they tend to be almost the same. we respect the EU one as well which they're part of I'm Canadian, are we on the blacklist? because last time I checked it was a problem they differ only a tiny bit so we have a load of ground stations infrastructure up in Canada and we can't download there and then sell it to certain countries that we could if we downloaded it in the US so we just add these things up and go well let's make meta list with all the bad guys according to all the people and then we take that off You must have to have a dedicated AI just dealing with the complexities of who gets what.

31:28It's not as hard as you think. That's just block listing of users or customers. What about downsampling or lowering the resolution in sensitive areas? Do you do any of that? Not generally. Yeah, like Peter's house. No, no, we don't. But remember, I mean, I think if people don't get confused about this, we're really a long way away, four to 500 kilometers. It's like the distance from Los Angeles to San Francisco, pointing our telescopes from one city to the other in distance, right? Obviously, the details you can see with that aren't the same as you can if you're flying a drone. If you're flying a drone, you can see people, you can recognize their faces, you get all the personal privacy.

32:09We're 400 or 500 kilometers away. We're not getting into that. It's kind of amazing what we can see, but it's not like that. And the reason that matters is that a lot of the most sensitive stuff doesn't come into the fray because of that. And countries agreed way early in the space era that they won't let each other fly planes over each other's territory without permission or drones. But you can fly in space because they consider that so far away. You can get some data, but it's just enough for transparency, but not enough to get into the details that they would care about. So it's basically that was the agreed upon definition.

32:49And the backstory there is fascinating, right? Back when Sputnik was launched in 1957, the U.S. had to make a decision. Do we disallow that to happen because it's flying over us? But if they did that, they wouldn't be able to fly our U.S. satellites over Russian territory. We said, OK, everybody can fly satellites. you just ran experiments and yeah and the physics of it dictate that right whereas the plane you can go up to russia and turn left you know you can't go up to russia with a satellite and turn left you're in an orbit you're gonna go over russia so what are you gonna do say that i'm not going to turn on the camera well that's silly that's obviously not going to happen so people aren't going to respect that so yes there was some physics that went into that and some fact of it being far away but yeah that became the international norm they famously the gary powers was shot down in the U-2 spy plane.

33:37And then the US said, well, we're going to put most of the most sensitive stuff in orbit. And that's going to be our domain of enabling us to monitor what's going on with nuclear weapon arms control and all this sort of stuff. But now it's just proliferated and far more people. People now can get through Planet what took the entire CIA and NRO infrastructure a couple of decades ago, and they can get it for a tiny, tiny fraction of the cost, right? And in some senses, we can do things that no one's able to do, like the daily scan. They've never done that many satellites. So they've never had that sort of global coverage.

34:18So we can do things that they can't even do. And the fact that that's now possible for a private enterprise, that has completely changed the game. So you historically have always brought the data back to Earth and did the crunching at the data centers on Earth. You ran some experiments in April where you put some NVIDIA chips on one of your satellites and you did the processing up there. What's the significance of that, Will? Well, it's basically enabling us to do the processing at the edge speeds up the time. So this is really processing at the edge, you know, in space, 500 kilometers up. And yeah, we put some NVIDIA GPUs on our satellites and all of our Pelicans going up now have them and OWLs will do as well.

34:59That combined with satellite to satellite communication, we're putting links so that we can go up to other satellites and then back down so that we don't have to wait until the satellite goes around to the ground station that we've erected, which we put all around the world, but still it has to take some time. Instead, it can just send back the answer. So what you can do, in the example we did in April, we took a picture of an airfield in Australia in this case, in Alice Springs. The computer automatically recognized the planes on that airfield. Then it just sent us back the locations and type of planes, right?

35:36That was done in seconds. And then we can send it back by RF, satellite to satellite. So you suddenly have things in seconds. Now, let me explain how that makes sense. You guys are a bunch of you in LA. Here's a photo of that, by the way. Oh, yeah. That image. Oh, cool. What a flashback. That looks exactly like what I used to work on at MIT. Like literally, exactly. And time really matters for a number of applications. Just think of the fires in L.A., Palisades and other fires. We gave images within a couple of hours of those fires, and then we did analysis building by building, which buildings were affected, where should the relief operators go, the American Red Cross, Cal Fire, and where the water was located.

36:25And how to have we been able to get that in a few minutes rather than a few hours? Could that have saved lives? You know, could that have saved properties? Potentially. Time really matters. So this is all processing at the edge is all about time. It's going from hours to minutes. Can you give us some geeky numbers on that? Because you In the Pelicans, you have a couple thousand frequencies of light coming in. So the data must be astronomically huge, just the raw feed. And then the NVIDIA chips will have no trouble compressing that down. But then you have limited bandwidth coming down to the Earth.

37:01So what are the rough numbers? Well, so, I mean, I know the numbers based on our Dove satellites. They take eight frames a second at 47 megapixels. so that we can get for each area on the Earth eight different spectral bands. And then within about a second, the satellite has already gone past that area. So basically, then you just have to start again. So it just goes eight times a second to get eight spectral bands for each area of the Earth. And each picture is maybe 35 by 20 kilometers in area. So just imagine that going all the time, clipping along all the time. it's in daytime over land, which is about a seventh of the time that it's.

37:50And then the rest of the time, it just repowers its batteries, if you like. It does some over ocean and pelicans more also turning to shoot to specific targets. So because of that, you have all this time when you're not taking images, which actually gives you more buffer to send it down. But yeah, so each dove is imaging maybe a couple of million square kilometers per day per satellite. So what is that? Bigger than the area of California, each satellite per day. So this is why when people say, oh, let's use drones for agriculture, I'm like, no, that's crazy. You would need a million drones per satellite or something, you know, or a thousand certainly.

38:38is actually cheaper to do satellites if the resolution suffices from the satellites, right? It's just orders of magnitude. Where does the resolution go? Right now you're saying it's at about a 30... 30 centimeters power, 3 meters superdub. In five years, where do you expect to be in five years? Well, we're already going upgrading our daily scan from 3 meters to 1 meter. With super resolution, it can get potentially better than that, too. That might even go to 50 centimeters, 30 centimeters. And then our pelicans, we've moved from the SkySat, which we inherited from Google, to 50 centimeters. Now we're moving towards 30 centimeters.

39:21And with super resolution, again, you can get a little bit better where you look at overlapping and sharpening based on pixel overlap and things like that. With what exposure time will you? The exposure time is, I want to say, it's like a couple of milliseconds. So you should be able to catch quite a few interesting aircraft in flight. Yeah, yeah. We get aircraft in flight all the time. I can show you that. Any UFOs? Is that where you're going? The Air Force took a look through our images, and I'm sorry to tell you, there ain't any UFOs. No. That's too bad. Why isn't Starlink? I'm so excited about us detecting UFOs, but I'm sorry to all the audience members out there that think there are some that have visited the Earth.

40:05Apart from the crazy people that think they've been abducted, it ain't true. We haven't seen any aliens. And NASA, let me tell you from first-hand experience, could not keep that a secret. Never, ever. That's ridiculous. That makes it even more improbable. Why isn't Elon doing this with Starlink? I kind of imagine that putting some cameras on board. Well, they're coming in. Stargaze. Stargaze is what he's doing, right? Well, that's using, as I say, they could, but they're kind of in the wrong orbits. They're a little bit too high, and you really want a synchronous orbit to have a consistent shadow angle for optical imagery.

40:46They are doing some classified missions for the NRO, which are, well, they're classified. and read some stuff on the internet about it. But they are generally not in the business of doing earth imaging. They're doing comms, primarily Starlink, which is obviously a very successful business. That's, I think, the most exciting aspect of the SpaceX IPO, frankly. I think it's an incredible response. What's the mass of a pelican or a dove compared to a Starlink? Pelicans are similar. And our doves are much, much smaller, like more than 10 times smaller. Will, can you give folks listening an understanding of how quickly the tech has developed to build these kinds of satellites?

41:37Because it's been stunning. You are on the cutting edge of this back. When did the first dove go up? 2013. So we've been doing it 13 years now. Yeah, so, I mean, to give you a sense, the radio speed has gone from a megabit a second to 10 gigabits a second. The cameras have gone from 2 megapixel to 47 megapixel. The hard drive space has gone from 100 megabytes to, what have we got, a couple of terabytes on there now. I mean, it's just extraordinary, right? Each satellite, each generation of satellites, we tend to be doing about a 10x. So our Dove to Super Dove went from four spectral bands to eight spectral bands and from a 29 megapixel camera to a 47 megapixel camera.

42:32So if you add that up, that was about a 5x increase in data per satellite for a similar cost per image. So it's like a, you know, we're talking about significant, an owl, our next generation daily scan going from three meters to one meter. That's 10 times more data or nine times more data, right, roughly. And we'll be getting it back about 10 times faster as well. So, yeah, 10Xs are still for the having. But, Peter, I think the even bigger thing is our hyperspectral satellite, Tanager, we're 5Xing. We're working on a new one that has five times bigger swath width for the same spacecraft. So those things are possible.

43:17So we're gathering more and more data. And the cycle of increasing those sort of things is, I would say, two to three years. So two to three years, sort of five or 10x, I would say, is the rough Moore's law for increases in data in space. But I would say the bigger thing happening now is the unlock of AI that really just brings down the barrier. It's all this capability is latent for that farmer I mentioned, for the hedge fund manager or whatever. But they couldn't get access to it. So I think we've got about 100x to go in the next couple of years just because of AI unleashing what was already latent in the present data.

43:57How does this flow to the average individual? I mean, how are people, how does it impact individuals on the ground right now worried about, you know, their local environment or people polluting and so forth? How do you make this accessible as an intelligent layer that people could just, you know, plug into on a regular basis? Well, again, just imagine making a natural language query of our data, just like you make a natural language query of the Internet via ChatGPT or Gemini or pick your favorite LLM. And so, except, again, LLMs understand the text and large Earth models understand the physical world.

44:35So can I answer that question for that farmer? You know, how's that? How's my field doing? What should I do? What precision? It could say, well, you've got blight over in this corner. You should put some fertilizer over there, do this. And then and the journalists can do the checks on some event happening around the world. the civil government responding to that flood or permits can just say, hey, here's my list of permitted buildings. Tell me which buildings have been built that don't have a permit. And it can go look at the last month, find the images, find the buildings, check the code against that, correlate against that list and then tell which ones have and have not got permits.

45:16And that has already, we've done that in a few areas in each case with journalists, with finance, with farmers, with civil government doing. This is going to be a boom to the legal industry.

45:33Here's three obvious use cases. What's the change in parking, parked cars outside of Walmart over weeks and months? Salim, that's the cliche, right? This is the cliche use case for folks purchasing planet imagery to trade stock prices based on parking lots. But shipping and knowing where ships are, there's rogue fishing ships all over the world that are a nightmare right now for the fishing population. Commodities, Brian. Commodities, right? Agriculture commodities, soy, wheat. But as the resolution gets better, if I live in Manhattan, is there a parking spot on my street that I could get right now?

46:08Exactly. I want that for sure. And that one, and also, is my teenager sneaking out the bedroom window at night? We have one meter or one third meter resolution here. We'll get that. But I guess, yeah, it's obviously your kid if it's your house. But more seriously, Will, I think you are in possession of a data set that could be GDP maxing. How much of that analysis are you doing in-house versus externalizing to partners like Google or others? Wait, did you just say GDP maxing? I said GDP maxing. I just coined GDP maxing with two Xs.

46:42Yes, I do think, I mean, we have hedge funds that are using our data right now. We're not doing that internally. But we have some hedge funds who will go undisclosed because they don't like being disclosed. But there we have some. And we think that they are getting significant alpha on our data, which we are happy to take a part of. Now, in the future, I do think there needs to be more. But again, I would take it up a level. I think GDP maximizing is one thing, but life flourishing is an even bigger thing that we can do this way. And we are super inefficiently using the earth right now, super inefficiently.

47:24You know, agriculture is terribly inefficient, for example. There's 10 X's for the having all over the place in agriculture. Let's go fix that. Abundance, maybe. So let's turn to Dyson Swarms. I want to get to Dyson Swarms. I have a quick technical question before I get to the theory. You've put NVIDIA chips onto the satellites. How is the cooling being handled? Oh, that's relatively straightforward. I mean, we can talk about computing space more generally. But, yeah, I mean, we've been dealing with chips that are obviously hot and need to cool off for decades in the space sector. There's no magic here.

48:03You can't use convection or conduction as you do on the ground. They're either air-cooled or water-cooled with physical touch. In this case, they have to be radiatively cooled, so you have a radiator. But radiators, we've known about that for a long time. We know how to do radiators. It's relatively known. One of the interesting things about it, by the way, if you're geeking out on this stuff, is that the radiating energy goes with the T to the fourth, the temperature to the fourth power. So basically - If you're a black body. If you're a black body, which you're close to. So if you double the temperature from, say, 100 Kelvin to 200 Kelvin, you 10x-ish your radiative power.

48:49So dumping energy is all tricks of thermal regulation, of radiators, and how you stop it. You want to get it as hot as possible without melting it. And there's lots of tricks to the trade there, but there's nothing fundamentally unknown there. These are no knowns. All right. I'll second that. You want to aim in the direction of the cosmic microwave background whenever possible. Absolutely. The 4K Kelvin of space. So you want to point your radiators at the dark. All right. Let me kick some notes on that one. Project Suncatcher. Let's jump into this. So you're putting TPUs for Google in orbit. You're building an early version of the Dyson Swarm Orbital AI Compute.

49:31Can you tell us what you're doing there? I mean, obviously, everybody's thinking about Elon's data satellites. How do you compare? Are you going to get your launch? Have you been launching on SpaceX? We've launched 40-some-odd times. 15 have been on SpaceX. We've launched over 300 satellites on 15 launches with SpaceX. They're one of our best partners. We love working with them. They've got it closest to a bus ride to space. I will point out that in addition to launch costs coming down, the biggest upheaval in space, I think I mentioned this last time I came on this podcast with you, Peter, the bigger transition over the last 10 years in space has not been the launch cost.

50:13It's been the satellite cost performance. It's been that miniaturization of satellites, both for Starlink and ourselves. And we sort of pioneered that. That led to at least 100x, if not 1 ,000x in cost performance for each kilogram you put on the fairing. So the dominant thing that has changed to lead to all these large constellations of satellites is actually the capability performance of satellites, not the launch cost. But both add up and they make things better. Performance density. Performance density, exactly. Exactly. So in our case, like how many bits do we collect per kilogram or per dollar spent, which is related to kilogram because the cost of launch?

50:51Yeah. So we've been launching a bunch with SpaceX. SpaceX didn't come up with this idea. I will point out they only started talking about this after we announced our project. And we'd be thinking about this for some time. And we're not the first ones either. Space industry has been talking about energy from space for decades and decades and space based solar power. And for many years, the idea is basically we want to put energy intensive infrastructure off Earth where there's abundant energy and where it's not conflicting with the incredible biodiversity and people's lives. So as Jeff Bezos likes to say, we want to zone the Earth, rural and light manufacturing and put to space all the heavier energy intensive infrastructure.

51:37Now, people are talking about energy in space for a long time, but the first and obvious easiest one is compute in space. Because whereas space saves solar power, you need to beam all that energy down. And how do you do that in a way without frying people's heads is actually difficult. Whereas putting compute in space, you get all the power advantages, but you only have to beam up the questions and beam back the answers. Well, we know how to beam bits. That we've been doing for a long time. ComSatellites was one of the first users of satellites. So all you have in space, so we basically did a study with Google about eight or nine years ago, looking at the details of compute terrestrially costs, the water, the building, the energy, all the things, and what it would cost to do it in space.

52:26And it just turns out that when launch costs come to about$200 to$300 a kilogram, is just going to be cheaper, surely on a pure cost basis, to put it in orbit versus on the ground. So as Sundar put it from Google, within 10 years, we expect most compute to be put into space. Now, that is a big deal because Google alone is spending 200 billion a year at this rate on compute. That's roughly the size of the entire space industry today, rockets, satellites, comms, everything combined. So Google is just going to do it, add up all the other folks that are going to do compute. And you've got a business that's bigger than the rest of the business, maybe 10x, the entire space industry today.

53:10So it's going to change the space sector. So we're doing some early tech demos for Google. When we did this study eight or nine years ago, Larry and Sergey were like, well, let's come back in 2030 when the launch costs come down to there. And I said, no, Let's come back five years earlier because it's going to take us years to build the technology to do the radiators, to do the clusters. So you have to have a whole lot of these. You want to basically a rack of GPUs on each satellite. And then you want clusters of spacecraft in close formation firing with optical links in between them. And all of that is a whole load of technology to develop.

53:49So what we're doing with Google, they selected us to build their first couple of satellites to test TPUs, radiation management, the cooling, the inter-satellite links. And so we're doing a couple of tech demos very early. It's a moonshot project. But the long arc is it's just going to be cheaper and has a peripheral benefit of not clashing with energy costs for communities, water for communities, or the biosphere. So it has a lot of benefits terrestrially as well. Alex? We talk on the pod, Will, all the time these days about sun-synchronous orbit and Earth acquiring its own mini Saturnian ring, if you will, on a polar orbit.

54:33When do you think SSO-based Dyson swarms will become visible at night or during the day on Earth? What is your time? Is this like early 2030s? I mean, we've got loads of satellites in sunsynchronous, and you'll see them today in orbit if you look out just after dawn or dusk or just before dawn when satellites are most visible. But most of these satellites will go into a dawn-dusk sunsync orbit. That means they're 24-7 facing the sun. However, that also means that they're not going to be very visible because that's literally when it's still a little bit light outside. And it's going to be hard to see those guys.

55:18So actually, by the way, there's real challenges of interfering with astronomers on the ground. And we have to be careful about that. But this is the best time to do it because it's not interfering with the deep, dark sky needs of astronomers. It's really in these other planes. So the short answer is it won't affect your seeing these rings. You won't see these rings of satellites because they're in a long time. I want my rings, but that's also a small number as well. It's a small number. Elon has FCC approval, I think, for a million of these AI satellites. Don't you think at some scale, if so many of these birds go up, that either they start to become visible or they start to become...

56:04Yeah, because then you'll be putting them in slightly different angles to keep them a space. They form a full band. Well, yeah, that's right. You would put them in slightly different inclinations where you still have 24-7 sun or very close to it. Yes, you would start seeing that, but it would only be right just as it gets dark and just before it gets light. Like it would be like this funny ring effect. But later we may put them in other orbits as well. I don't know. They would have to be much higher to get the. How concerned are you about orbital debris? Well, we talked about like in Elon's S1, his number one risk factor on Starlink, which is their revenue profit engine right now, was orbital debris being able to knock out a lot of capabilities.

56:50What about you? What do you think about that? I think space debris is a real challenge. And that's why we put our satellites below the area where that's challenged, which is 800 to 1 ,200 kilometers from the Earth's surface in altitude. So we put our satellites 400 to 500 kilometers to keep them well below that problem. Oh, interesting. Yes, the syndrome is already in operation and effect. But bear in mind, there's about 10 ,000 satellites in space of order, and there's about 100 million pieces of space debris. So about 10 ,000 times more objects in orbit are 10 ,000 pieces of debris for every satellite.

57:29So the vast majority of the problem, even if you put a million satellites up there, the vast majority, you'd have 99 % of it is still not satellites. The challenge we have to deal with, I'm trying to point out, is debris. And that is mainly made up of all the small bits of stuff left over from former rocket bodies, exploded satellites, anti-satellites and other things, which were done in high orbits and so could live there for decades. Now, when we were at NASA, Peter might remember this, we came up with a scheme under Pete Worden's mentorship of using lasers on the ground to sort of do traffic management of that debris.

58:09Because obviously, with two satellites, you can move out of each other's way or any maneuverable but most of the conjunctions in orbit are debris with debris so what do you do about that we need to stop the collisional cascade for those pieces and for that you can actually use lasers on the ground that gently nudge one so they miss each other you can do this sort of traffic management we call it light force a system like that could actually stop the cascades cascade and slowly bring everything down enabling this to but the actual satellites is less of a problem as long as we keep them in low Earth orbits and there's lots of space.

58:43Just to give you a rough order, even in this sort of sun-synchronous dawn-dusk orbit, there's about a thousand times more space, just thinking very crudely, than there is on the entire landmass of the Earth. So there's a long-distance space. This is really fascinating. So wait, you're at 300 kilometers or 500 kilometers? What's your altitude? Four to 500, yeah. Four to 500? And in that, What's the lifespan of an object orbiting at that altitude? A few months to a few years. Okay, so self-cleaning. Self-cleaning, yeah. And you're starting to walk through, so you have about 100 kilometers. The atmosphere drag pulls everything down.

59:18The atmosphere drag, yes. Yeah. Yeah, so you have about 100 kilometers of space where you can get a good two-year, three-year orbit. You know, a GPU in space isn't going to, it's going to depreciate over three years anyway. Exactly. And that's why we call it strapping space to Moore's law. We always update our satellites every couple of years because the satellites in space are becoming obsolete. Just like the phone in your back pocket, you don't want a 10-year-old phone, you don't want a 10-year-old satellite in space. What altitude is Elon going with for his? Well, he was going higher, but I made the point to him that, firstly, that's a real challenge with space debris.

59:51And secondly, it won't be self-cleaning. And even if you put propulsion on these things, even if one in a hundred fail or one in a thousand fail, you have a real big challenge if you put that much mass into those orbits. So it makes much more sense. And so later StarLinks have come much lower down, and that's much better for everyone. I'd love Will to just pull on the upmass question a bit. So over the past five years, I did this calculation on my newsletter. For the past five years or so, according to what I've seen, upmass has increased by 40 plus percent year over year. And if you just naively extrapolate out 40 plus percent year over year upmass increase by the year 2144, I think you find that the entire mass mass of Earth has been basically upmassed and Earth has been Earth has been disassembled.

1:00:50If you just naively follow the exponential. By the way, everybody, I am not supporting the disassembly of Earth. Peter does not. For avoidance of doubt, Peter does not support disassembly of Earth. We've established it. Good. Yeah, I mean, obviously, extrapolating anything 140 years into the future is rather tricky business, as you guys are aware. It's barely the whole point of the singularity is that it's harder and harder to predict the future. I remember when Peter and I first met 20 years ago, it felt like we could easily predict roughly who was going to do what in 10, 20 years in the space.

1:01:27Now, if you can predict it one or two years out, you're a genius. And for AI, it's even harder. You know, it's measured in months, right? 140 days. Let's see three to six months into the future. So that horizon is shortening for sure. And 144 years, I think, is just we can't even discuss. No predictions will then regarding when upmass increase will start to slow down. Because right now it seems naively set to increase. No, UPMAS is definitely going to continue to increase. But again, I mean, I think the most important aspect of that is how do we get the energy intensive infrastructure? Look, data centers are going to become a real hot topic politically in this next election in the midterms and upcoming elections because people don't want data centers in their backyard.

1:02:25They don't want the energy costs to go up. They don't want their water to disappear because they kind of like access to clean water. It's kind of handy. This is causing lots of tension. And it's not surprising. I think those things, you know, and we're wiping out agriculture lands, farmlands, what have you for this. Putting in space is is is the way out of that conundrum. And then we can have compute and we can not interfere with those. The elephant in the room here, Will, I have to ask it, is how do you compete with Elon's plans for orbital AI data centers when he's got the launch capacity? He's got massive manufacturing capacity.

1:03:05Do you end up folding tents together? Are there going to be more than one player in orbit? Or does Google just acquire you? Let me ask about that, too. I want to throw one more log on that fire, which is Google sold you their satellite business. And now that was before everyone realized data centers would be in space, I think. Now they're working with you. If Elon doesn't want to launch, Eric Schmidt now has a rocket company. There are a lot of arrows pointing in a different direction, like here's the Elonverse and here's the Googleverse, and you're part of the Googleverse. But I know you're working with SpaceX, so I don't expect too much.

1:03:42And Invinia and various others. Look, there's a complex relationship. I mean, Google is both a shareholder in SpaceX and they're competing. I mean, look, these are both competitive and collaborative situations, and we feel the same. We're a strong partner with SpaceX. We really love their partnership on launch. We work with them. Our teams work together really well. I wish them great luck with the IPO. I think it's fantastic that there's so much interest in space. It's so hot right now. And at the same time, yeah, they compute in space. And we're really helping Google to do their project a little bit.

1:04:22And we'll see how it goes. They take a different path. But don't underestimate their smarts and our smarts and how we can do this. I mean, I see, you know, roughly Elon is throwing mass at this because he can with the rockets. But we're throwing smarts at this, and there are lots of tricks up our sleeves for how to do this really smartly. Everybody, welcome to the health section of Moonshots brought to you by Fountain Life. You know, we talk about AI on this Moonshot podcast all the time. One of the most important things AI is going to be able to do for you, besides educating your kids and helping you with your taxes, is making sure that you're living a healthy lifestyle, that you get a chance to get to 100 plus.

1:05:00I'm here today with Dr. Dawn Musalem, the chief medical officer of Fountain Life, and a part of my medical team. Dawn, a pleasure. Great to hear. You know, the thing that people are concerned about most, about living to 100 or 120, is their cognitive abilities, making sure they don't have dementia. And the numbers about dementia are problematic. Can you share what you've learned? Such an important point, and you're right. At Fountain Life, our members, the number one thing people are most concerned about is losing their brain health, forgetting the name of their child, forgetting the face of their loved one.

1:05:35We know that when it comes to dementia, the conservative estimates are that 45 % are entirely preventable. What was amazing is with the advanced testing we're doing at Fountain Life, one quarter of our members had advanced brain age. Wow. But what was really awesome is, again, back to that prevention. when he partnered it with Healthy Living. This gives me chills. Eating healthier, moving our bodies, sleep. Optimizing sleep is so important. You know what we saw? We saw that we improved that brain age by 26%. That is a big, big number to show that the majority of those individuals were able actually to improve the brain age.

1:06:12And one of the things I love about Fountain is we're searching the world for the best therapeutics, the best approaches, and making sure we bring it to our members. So if having healthy brain function until 100, 120 is important to you. Check out Fountain Life. Go to fountainlife.com slash Peter. Make sure you become the CEO of your own health. All right, now back to the episode. All right, let's move on to our next story here, which is still in the space arena. But this time we're going to talk about the launch industry. So here we go. Our next story is literally as SpaceX is rocketing forward and New Glenn had a kinetic disassembly of their, or Blue Origin of their rocket New Glenn.

1:06:55Here comes Relativity Space. So a little background on this. Relativity was founded back in 2015 by Tim Ellis and Jordan Nguyen. They're both friends. I was an investor early in Relativity Space. I've had them on my stage at the Abundance Summit. And Relativity back in 2023 flew their Terran 1 rocket. It got through Max-Q. It did not get to orbit. In fact, very few rockets on their first launch attempt. Only three, I think, in history right now in the U.S. have gotten into orbit on their first attempt. And they pivoted after their Terran 1 to go to their Terran R, which is a heavy class launch vehicle.

1:07:31You can see here the numbers. Terran R is 23 tons. Falcon 9 is 22 tons, roughly the same. New Glenn, 45 tons. And Starship at 100 tons. They missed their financing. It's really hard to finance space projects, especially rocket projects. And here comes Eric Schmidt, who is an early investor, comes in and writes the check to basically buy relativity space. And so here he is. Eric is the CEO now of relativity space, which blew my mind when he took that role. And they just announced they have gotten a mission from NASA called ELIS. It's a Mars orbiter sensing mission with some communications capability.

1:08:16Any thoughts on this one?

1:08:21I've known Eric for many years. He's a very early investor in Planet in our Series A round. So all the way back to the very beginning. And Eric has a smart eye for business and a smart eye for technology. He's obviously relatively new to the space business, if you can excuse the pun. But, yeah, we obviously think the world of Eric and relativity has come a long way. And, yeah, they had some of those financing challenges. But I think now with Eric's backing, I think it can go a long way. I'm very excited for them. And I hope we can launch with them. Will, this story is really interesting. So I didn't know he was a seed investor in you.

1:09:07Was he still CEO of Google at the time? And did they still have their satellite business at the time? He was an investor before they bought Skybox, I think. Okay. So he was running Google, made the investment. Aware that data centers might move into space someday? This is a long time ago. I think this is before that had caught his old founders of Google. But Dave, the question is, did he buy relativity space with the thought that data centers in orbit are going to be critical? Because it's a massive advantage for SpaceX to have launch and satellite capability and data center capability. I mean, we interviewed him four times in the last year, Peter.

1:09:50So I'm really coming around to the view that he 100 percent knows and knew that this was the future. Because he said on every one of those interviews, I don't know anything about space, but I know a lot about people and I know a lot about companies. Well, he also knows a lot about investing. He's got to be one of the best in the history of the world. His vision is unbelievable and he has access to all the information in the world. I didn't know he was a seed investor in Planet. So that's one other source of information that he has. And from that vantage point, yeah, Elon can't be the only guy launching.

1:10:22And Jeff Bezos is no dummy either. You know, he's launching too. Of course, it's been a passion of his whole life. I have a relativity space question. You know, when NASA was launching space shuttles, it was between 600 million to a billion per space shuttle launch. SpaceX dropped that down to about 60 million. And the plan was for relativity space to operate at about 6 million a launch because they were 3D printing the rocket engines or big chunks of it. Originally, they had their Stargate printers to print all of the rocket. Then they broke in at 85%. And now today, I guess they're just 3D printing their engines.

1:11:00Yeah. Do we know what the launch cost is that they're aiming for? Does anybody know? I don't think that's disclosed. I tried to look for it. I also think it leads, I mean, the question behind the question perhaps is what happened, whatever happened to 3D printing in space for space, terrestrially, or in space. And my perception is that relativity under new management is migrating more in the direction of competing in lift and heavy lift. And there's potentially a gap in the market now that relativity was originally aimed for, focused on 3D printing for space that someone else could potentially fill.

1:11:39I'd love to see more 3D printing in space, in cislunar, lunar surface. And in general, no one right now seems to be the obvious incumbent anymore in that market. Yeah, I agree. There's a huge opportunity in 3D printing. Fundamentally, all the design constraints for satellites is to do with the launch. That's the hard thing. The vibrations, the separation, we get a 200G shock load. And then you get into orbit, you don't need any structure at all, basically, because it's zero G. So you want a completely different design for your launch than you do in orbit, roughly speaking. a completely different design.

1:12:22Absolutely. Peter and Will, it's so rare. You guys are two of the top on the entire planet on this whole launch cost question. We just have to get this figured out right here, right now. So the Elon rocket is massive in scale, a couple-ton payload. But how much of the efficiency, Elon's always been saying it's the reusability of everything that is the driver, not the overall scale. Their goal is to get down to 100 bucks, right, per kilogram from where it was in the past at$10 ,000 per kilogram. And the only way you get that is by rapid usability. Remember, to launch the 500 ,000 or a million satellites for his AI constellation, it's like a launch, you know, two launches an hour.

1:13:09You want 10 rockets or do you want a million rockets? Yeah, exactly. Well, this is where I'm going. So this, the relativity space rocket is also. exploited on chemical rockets so far. We have also just not mass produced rockets ever. So an entirely different approach. Just like Elon said, look, why don't we do reuse? No one has done, look, why don't we do mass manufacturing? Because it's complicated as a rocket, but it costs tens of thousands to make not tens of millions. So what is going on there? And then we're throwing it away every time. There's two independent ways and no one's really used this other way.

1:13:47And then a whole separate thing. And I'd be thinking about this if I was Google or one of these big data play companies that are seeing that they want to spend a trillion or more on space over the next decade or two. If I want to spend that much, I want to spend a few billion on novel launches because it's been like printing yeah let's just launch blocks of material and then 3d print it or or as elon's been talking about recently rail launch it from the moon because in just a sheer energetic standpoint getting stuff from the moon to low earth orbit is cheaper but but even from the earth which is the near term um easier one yeah spin launch or uh long shot or these kind of very different ways.

1:14:33No one's thrown a billion at one of those or a few of those and see if it can actually work. We've used chemical rockets because guess what? Werner von Braun figured out he could bomb London 100 years ago. Not quite. But, you know, I mean, and then no one's invented anything since. Basically anything. I mean, even the reusability, that was cool, but no one's made a significant advance. We're stuck in the chemical rocket paradigm, and we don't need to be. There's a beef refory in the 60s and 70s, both with Russia and the US, into fission power rockets, but then everyone got scared about that. But I think we need to revisit, at this point, that launch equation, because the way to get from 100 to 10 to 1 isn't going to be a chemical rocket.

1:15:23It's going to be - On this pod, I... Yeah. Peter, that's what I say to you all the time. You're concerned about the SpaceX launch monopoly, but there are many other launch paradigms that could potentially leapfrog. Space elevators with new materials, of course. And, you know, did the calculation on this pod, you know, MGH and one half MV squared in terms of total energy. And if you could buy it from space and winch it up and accelerate it, you know, you can get the cost down for you and your spacesuit to 120 bucks. Skyhooks. Sky, yeah. So, wait, let me follow up. One more question, Peter, and you can tell me to get off.

1:15:58I'm really, really curious, though, and you guys are the experts. So, if I get fully reusable from relativity space, but it's a quarter the size of an Elon rocket, so there's got to be some economy of scale that comes with just raw size, which is why Elon pursued it, but they're still fully reusable. But now, as Will is saying, it's the manufacturing of thousands of these in an assembly line that really reduces the cost. But he's built a machine to build the machines. His goal is thousands of starships, maybe even more. I mean, if it's fully reusable, it's just the cost of the touch labor and the cost of the fuel.

1:16:35And the fuel is de minimis. It's free. It's oxygen and methane. So let's say that Eric is doing the exact same thing because he will. Eric Schmidt is doing the exact same thing, but his rocket is a quarter of the total scale, a quarter of the payload probably. Is that significant? Because, you know, the launch costs aren't the, you're launching these very expensive, you know, 70 cluster, 72 cluster NVIDIA GPUs with all the cooling and the solar power and everything. That's an expensive piece of equipment. Suppose that Eric's launch costs are 200 bucks or 300 bucks a kilogram, not 100. Does it matter?

1:17:09Is Eric still competitive? Do we have a duopoly then? Oh, yeah. Can I explain a little bit about this? Because people, I think, misunderstand. It is not just about the launch cost. The launch cost is the biggest piece to get us to the threshold that makes sense. But thereafter, it is much, I argue, probably more about the efficiency of the compute than about the launch cost. Really? Because the efficiency of the compute drives the amount of energy you have to dump, which drives the mass of the spacecraft. And that ends up being significant. So, for example, Google TPUs are significantly more efficient than GPUs in a flops per watt standpoint.

1:17:53That really matters because all the rest of the GPU energy. So I like to put it simply, whilst everyone apart from SpaceX has to pay the SpaceX launch tax right now, everyone apart from NVIDIA and Google has to pay the NVIDIA tax right now. And which tax is more important? I actually say near term is the launch, but longer term is the compute. That is fucking brilliant. That is absolutely critical. Nobody has said that before. Gigs are going to be more important than launch for this game long term. Mark my word. And that means Google, if the TPUs at just inference alone are significantly lower energy use per inference.

1:18:38Correct. They choose the winner of space. Correct. And NVIDIA could have a play at this, but their GPUs are more general than the TPUs. The TPUs are more efficient. Now, obviously, Elon's trying to build his TerraFab, but that is a big long term project, if ever there was. one. Meanwhile, Google has been investing in that compute for a long time, and they have efficient systems for leveraging that compute in ways that will boggle your mind. I mean, people think of Google primarily as a software company, and they are. And when they gave us their satellites, or gave us, we bought them, we were like astonished because we were like, oh, wow, they really didn't know how to build and operate satellites.

1:19:20I mean, that's so brilliant. Alex is But they are brilliant at data centers and networking and all of that. Google are the world's best at that. Don't underestimate how big a deal the infrastructure piece on the chips and the interconnects and the energy efficiency of all of that is. That turns out to be the biggest piece of this part. Yeah. God, that's so brilliant. Alex is always trying to find the innermost loop of the innermost loop of the innermost loop. And so right here, right now, that inference time power efficiency determines the winner of the entire thing. And everyone's writing off Google at the moment.

1:19:56They have massive defections of key talent. We'll see it later in this pod. But if they have a 2x watts per inference advantage over NVIDIA, remember NVIDIA is highly, highly emphasizing training time, not inference time. Because, you know, Cerebris and a whole bunch of other things are starting to really eat away at the inference time efficiency. But the TPU 7, 8, I guess the next TPU will determine whether space is dominated by, you know, like you said, the launch cost, even if it's 2x on an Eric Schmidt rocket, that is not the swing factor. It's the can I access those chips? That's really brilliant.

1:20:34And, Will, we do talk about in the past few episodes, we've talked about the training versus inference balance on terrestrial versus orbital data centers. Versus who? argument versus lunar versus Martian. One argument to be made in favor of in the short-term terrestrial data centers for training is that it's just easier to build larger, coherent training sessions on a terrestrial data center. What do you think is likely to be the balance between training versus inference on terrestrial versus non-terrestrial? Great question. Yeah. I think inference does make more sense in orbit to the first order, and it's mainly because that's a more distributed, lots of little runs of a machine, right?

1:21:18Now, there is the advantage for training runs that, you know, you want to sort of, you send your data, it spends a couple of months crunching it, and then you send the answers back. From a comm standpoint, it's easier to do the training in orbit than the inference, because you really need the latency down for inference. but from a compute distribution standpoint, it's easier to do it, to either do the inference in space. And obviously 70 % or so of the compute on Earth is now inference, or AI at least, which is most of it, is inference, not the training. And that's only set to go up. So I think the main problem to solve is the inference one anyway.

1:21:58Go ahead, Alex. You think training is likely to remain grounded in terrestrial data centers for the foreseeable future? longer. I don't think it will be forever. I think it will all go to space, but I think inference will go there first, yeah. Great. All right, moving us along out of the space arena, because we could spend all day here, and everybody listening has gotten their PhD. I have one last question about space. All right, all right, one last question. One more. Question for Will. Is the best commercial opportunity about leaving Earth or making Earth more useful? No, I mean, I think, look, my co-founder, Robbie, whilst I was sending missions to the moon, if you may remember, was working on a mission called TESS.

1:22:43So we helped to find water on the moon, which was very exciting. As lunatics, we're very pleased about that because it makes the moon much more. I mean, it was already better, smarter destination than Mars by 10x, but this made it 100 times more smarter destination than Mars, which finally was the nail in the coffin, which Elon finally understood recently and changed his mind that the moon is first. By the way, in the long run, are you a moon, then Mars or a moon, then asteroids? I would say moon is enough for a long time. And I get back to this because what Robbie was doing was focusing on exoplanets.

1:23:21And he had these telescopes looking out, looking for planets around nearby star systems. and they found now, we found thousands, I think it's up to almost 10 ,000 planets around nearby star systems. And I'm here to tell you and everyone else that the best one by far is the Earth. And I'm not talking about by a little itty bit. I'm talking about by several orders of magnitude. There is no place on Mars that is better than the worst place on Earth. Not by a little bit, okay? This planet is so fucking cool. And the reason I want to emphasize that and excuse my French. The reason I want to emphasize that is that when it comes to like, I don't believe in the sending millions of people into orbit anytime soon.

1:24:06I think it's all about protecting this incredible biosphere. Life is either singular. We haven't found the aliens. Sorry to break the news for those geeks that think they've been abducted. We haven't found life of Earth. Life is either singular on this planet or extraordinarily rare. Either way, we have the most beautiful life system on this Earth, the incredible complexity of how it all works together. That is worth protecting and putting most of our energy on. And space is super useful for that because it gives us the advantage and it gives us the data that underpins our ability to manage this planet smartly.

1:24:46But planet is here. SpaceX can be space for Mars. Bezos can be space for the moon off they go we're a planet space for the earth to help us to take care of the earth both with earth imaging to help upgrade the planet to be smarter decisions for helping take energy intensity infrastructure off the planet we're space for the earth because this planet they can have those planets this planet is by far the best ladies and gentlemen Here, here, here, here, here, here, the defense rests. Oh, I love it. I'm going to move us along because, you know, we've got still a lot to cover. Full threat. Defense of the Fermi paradox.

1:25:29Thank you, Will. Exactly. Well, we have to do that. We have to move on, but we have to do this again. There's so much more to explore. This has been phenomenal. Our next story here is the great AI brain drain. So two of the most important minds in AI have changed teams this past week. First off, Noam Shazir. If you don't know his name, he was the lead author in the Transformer paper, which is the T in GPT. The architecture of the entire modern AI revolution was built on his discovery. He's unfortunately leaving Google for open AI. And get this, this is the second time he's left Google. Two years ago, Google bought his company, Character.ai, for$2.7 billion to bring him back and put him in charge of Gemini.

1:26:11Well, he's leaving again. I would guess after part of his stock package is vested. And second, another rock star is leaving Google. Didn't they give him a billion or something? Easily. Yeah, that company wasn't worth anything. It had basically zero revenue. He paid a billion for him, and now he's out. Whole generation of Silicon Valley parents are naming their kids Noah. Yeah, and that begs the question, what is his comp package at OpenAI? It must be insane. It must be huge. Okay, second, another rock star left Google. John Jumper, the Nobel laureate who helped Demis create AlphaFold, is switching from Google DeepMind to Anthropic, likely to push their AI for science.

1:26:56Remember, a couple weeks ago, Andre Kaparthi also joined Anthropic. He was a free agent. So my question for you, Alex, is this AI talent literally voting with their feet? Is this sort of a better prediction of where AI is going? Yeah, I think so. I mean, I have no financial interest in this, so I can speak, I think, pretty unvarnishedly on the subject. My perception is the frontier is very competitive. And at the moment, it's a duopoly at the frontier between open AI and Anthropic. And Google DeepMind has fallen behind the frontier. And I think it was notable at I.O. that Google did not release a frontier model at all.

1:27:35They released a flash capability, which is great in everything, and certainly much more aligned with Google search level economics, where you want ultra low latency, ultra cheap models to power the one boxes in Google search replies. That's great for Google's existing legacy search business, but it's not a frontier capability. So my perception is that Google has fallen behind the first tier of Frontier Labs at this point. And if you're a top researcher, you have to be asking yourself all of the research questions that you could be asking with raw access to the pre-trained models before all of the post-training and all of the guardrails get slapped on.

1:28:14That's very attractive if you're a frontier lab researcher to have that raw access to a pre-trained model at the frontier. And if GDM doesn't have that frontier level access, you're probably looking to either OpenAI or Anthropic to get that frontier access for yourself. So yeah, I think this is a reflection of GDM falling behind. I'm going to take a different point of view. I think this is relatively in the noise. I mean, we've seen people move from Anthropik to OpenAI to Google, I mean, all of the directions, right? That's going to continue to happen. And these two are significant players, so I don't want to trivialize it.

1:28:50But I think the stock market reaction in particular was over. And I wouldn't bet against Google in this game. Yeah, I keep on saying that. If I was an AI researcher, I'd pick the one with the most compute, that's Google by far. The most data, that's Google by far. And the most smart people, that's Google by far. I'm sorry, that's just true across the board on all of those things. And it is a compute, data, and talent game. I think they did fall behind a little bit a while ago, but not now. I think the Gemini model is generically pretty good. I'm not the best expert on that, but my observations are that it's pretty high up there.

1:29:33And the prospects are even brighter. In fact, I think this is Google's to lose. I think Anthropic is doing incredibly well, and especially because they picked a very different business model. But OpenAI have picked the business model that more or less is in Google's sweet spot. And Google already has 10 applications with over a billion people to put its AI systems to. So they're the incumbent on that space. So I worry much more about open AI than Google. Dr. Blondin. I agree with everything Will said, but I'm going to give you the counterargument just because I know a couple of the players. So before these recent affections, Shane Longpreay from MIT went over to Anthropic.

1:30:14Tobin South from Stanford went over to Anthropic. Then Android Carpathia, as we just said, he was a free agent. All of these guys are singularitarians who believe that self-improving AI is exponential and almost instantaneous. And now you have John Jumper going over, and I don't know him, but you do, Peter. So I think that what's happening with those four people and a lot of other people is I could go to Meta and get paid a lot, but I'm going to miss the singularity. Anthropic, yes, they don't have Google's compute. Yes, Google has a huge advantage with the TPUs. But if I believe that Fable 5 is truly self-improving, it's crossed that line, and I can't use Fable 5 at all right now, but mythos is what I really want.

1:30:57The only way you can be part of the singularity in world history is to now go join Dario. And I know that's the psychology of the first three, so it wouldn't surprise me if it's the psychology of John Jumper. And small advantages compounding exponentially take off very rapidly. I can imagine the job interview. Come on in. Let me show you what's behind the firewall. And it's like, oh, my God, I've seen God. I cannot go back. I mean, this is publicly reported, Peter, that this is how Anthropic does its recruiting. They public it publicly available information that as opposed to the way GDM does its organizational workflows, Anthropic reportedly puts a lot of its best people on a meet in front of the applicant or the job seeker and shows them all of this compute can be yours.

1:31:43Here is the raw access to the models with raw capabilities. I have a more mundane answer. Also, one other thing that is, this is not coincidence. If you look at Polymarket's prediction of Fable 5 coming back, it goes down a little bit every day. And so Fable 5 will come back, but it'll be a reduced version of what it was the first time it was out. And so, you know, for the Polymarket to be true, it just has to be a product called Fable 5. And then that pays off. But that's number even given that it's coming down, you know, the odds of it coming back by the end of the month keeps slipping. And so I think Dario loves that, that the only way you're getting access to the best of the best of the best frontier, self-improving AI is right here inside our building.

1:32:26And every every week that goes by is another week toward the singularity that you'll miss if you're not part of my. I remember when I got when I brought Ray Kurzweil over to Larry Page to meet him for the first time to make an investment in his company. Larry's point was, you know, instead of me investing, the only place you're going to be able to build out your vision, Ray, is inside Google. We have access to all of this unfettered. I can imagine it's the exact same point. You know, John Jumper, you know, I was saying to you, Alex, I'm amazed given Isomorphic just raised a whole bunch of capital and they're focused on the biotech arena that John and I do know him.

1:33:00But, you know, that he would jump over to Anthropic. It's got to be that, you know, Dario is just basically come and lead our bio and you have access to unlimited compute far beyond what GDM has. The only thing I'd tweak on that, Peter, is that the recruiting pitch to Ray Kurzweil is this is the only place in the world you can build your vision. And that was a few years ago, many years ago. Now the pitch is the biggest event in the history of the world is imminent. The single biggest thing that's ever happened in human history is imminent. It's going to happen in one location on the planet. Our bet is that it's inside this exactly.

1:33:37That's why we have the Fermi paradox and there's no other life in the universe. Or it's going to be everyone's benefit. It's going to be awesome. That's the roll of the dice we are apparently just about to play. You know, Salim, I was with Mike Sunel and he said, what are you excited about? I said, I'm excited about the future. You know, this is the most extraordinary time ever to be alive. And I do believe that we're living in this quantum superposition. And I think people need to have a positive vision of where we're going and manifest that future. Because if you don't believe it and you're sort of steering towards the negative dystopian future, that's what we're going to get.

1:34:13So the purpose of this podcast, everybody listening, is to give you a positive vision of the future. I really agree with that. I think it's so critical. Silicon Valley basically ran for decades on Star Trek, and we just don't have the modern equivalent beautiful future vision. That's what the future vision X prize is about. Right. And as, yes, exactly. We need those. We need Neil Stevenson and Stan, Ken Stanley Robinson and others to put out books on the future of AI and humans and how it can work together. Because right now everyone's Terminator and, you know, it might end up a Terminator. That's pissing me off.

1:34:59I'm so angry at Hollywood, right? Because we're shaping our neural nets. You can't play Hollywood. That's what sells. It tickles your amygdala. It's why horror movies sell. Will, I have to ask, I mean, my favorite sci-fi is Accelerando. What's yours? What's the best depiction of the future? You know, I don't read a huge amount of sci-fi. I find sci-fact. I read Nature magazine every week because I find it so fascinating. I'm already threshold out. But I mean, I think some of the classics like Snow Crash and others really got, were incredibly good at depicting the next, the time. One of our subscribers asked for another book corner, Alex.

1:35:42So I appreciate you asking this question. Wait, I need to get my word in. I need to get a word in. The John Jumper and Noam thing, I think, is much simpler than all of what we're talking about. It really simply comes down to agency. Google is a big company with a lot of organizational drag. If you're an individual, you can make a much bigger difference in a smaller organization. Yes, they may have better models, et cetera. It goes all the way back to Peter in our 2014 EXO book. We said smaller beats bigger, right? Trust beats control. And we have this kind of rolling carpet of the smaller teams can outperform bigger teams.

1:36:26And the fact that you can do so much more. You know, when Facebook launched, Google spent two years trying to build Google Plus. And it was a miserable failure because you had to get permission from YouTube and the groups and search. And we were trying to integrate amongst all of those. Meanwhile, Facebook was saying to their developers, anybody who's ready with their feature, just take it live on the live site and go. And of course, they were outperforming Yahoo, Google, everybody. And I think it comes down to the ability to get things done more quickly can happen more in the smaller labs, plus they may have access to the best frontal office.

1:37:02I like that. I think your point's right. I think that sort of mundane factor could be much more. And that's why I was saying I think it was overblown at what this particular incident meant for Google. But, you know, we'll see. But what I want to throw in is that, as you say, the small guy is going to make a big difference, and I want to make a pitch for how the space sector is going to play a big role in this AI future, and come back to where we began, which is, you know, when a baby is born, they learn and become intelligent and ultimately self-aware and conscious by interacting with the physical world.

1:37:40They are not a brain in a vat, and they wouldn't learn the way they do without interaction with sensors and their physical actuators. AI, at the minute, the LLMs, are basically brains in a vat. They have absorbed the text of the Internet, but they are largely isolated from it. They can't real-time interact with the physical world. They can't, not in terms of sensing nor actuators. And until they do, I don't believe they'll learn. So I actually think that physical data and obviously planetary, you know, in the big scale, that's why I talk about planetary intelligence, the big scale of planetary sensing is going to be done from space.

1:38:16The compute soon going to go up there, as we just discussed. And that's really going to lead to a planetary. And what that might enable us is to build towards planetary consciousness and planetary wisdom. because that waking up point can only happen when you start having that real time loop. And so I think that these things are not unrelated. It's our path through to avoid the Fermi paradox, to teach the AI to become conscious, partially because we're going to need that, partially because it needs to understand the real world in order to learn. I love that. I love that. But also it will align it with human interest because it will be conscious and therefore be empathetic with our conscious experience.

1:39:04And it will know about all the deltas and the forests and all the animals and all the human civilization and therefore more implicitly care about it. So, you know, caring about something and knowing about them are highly correlated things, even though they could. And highly desirable. Yes. So the AI future ain't going to be just those guys sitting in their library with just the text of the Internet. They're going to have to get into the physical world. AI companies, whether it's Anthropic OpenAI or Google or any of the others, they're going to have to go and get into real world. Cars and satellites and drones, robotics.

1:39:43That's how they're going to graduate to the next level. we're going to need a leap of AI and it isn't going to come from just throwing more compute at the text of the Internet. It's just not going to come that way. So, again, obviously, I'm extraordinarily biased, but I think that space data is actually going to play a non-trivial role for that. Because what's the Wikipedia, you know, the crystal of the LLMs at the core is Wikipedia. It's like you're chatting with Wikipedia when you're chatting with an LLM. More than anything else, it's got the LLMs, is Wikipedia wrapped up. You want to hear something totally mind blowing?

1:40:19Good. We just invested in a little team in San Francisco down the road from you, Will, that tells us they're going to beat Google to AlphaFold 2. They're going to have a better protein folding. And they're a little team of five people. How are they doing? Stanford, a couple of Stanford guys and an MIT guy that used to work here at Link. And they said, it's not because we know anything about protein folding. We have a recursive self-improving process that is just mind-blowing. And that's why I'm not giving you the company name, because I don't want people to show up and spray paint their door. But they're like, yeah, we literally knew nothing about protein folding two weeks ago, and we're still going to beat Google to protein folding.

1:41:02I don't know if they're right or wrong. I don't want to throw their names out there. But it's mind-blowing to think that what Salim was saying, a little team with agency using RSI is superpower. So then just a couple days later— And it's a$45 billion a year revenue company as of two months ago. Now it's probably doubled. Insane. What's amazing to me about the story I just told, though, is that right after that, John Jumper goes over to Anthropic where RSI may be imminent. And he also is trying to solve all diseases using a similar RSI. You may be wrong and it may be your startup. And I just want to emphasize one more thing about this direction in planetary intelligence.

1:41:41It's not just that it's going to happen. It's happening right now. We have already built this app. It's in beta testing right now that actually already integrates planets data with AI, enables people to make those natural language queries. It's going to be world changing and lots of other companies are doing things like that. There's going to be totally lateral plays to the AI game that is going to come out and I think end up being critical for the next phase of development of AI, especially AI alignment. Yes. Yeah. And you're building that planetary nervous system that the world really, really needs.

1:42:21Totally. I mean, maybe let me just if I may just push on this will a little bit since you're referring quite a bit to LLMs, but maybe more colloquially, one might speak of foundation models that are intrinsically multimodal or omnimodal that have been trained extensively, not just on Internet text, but Internet images, Internet video, synthetic video. in many cases, world models, lowercase w, not earth scale world models. So one might suggest that most modern frontier models already have a pretty good native intrinsic understanding of the physical world might not be perfect. The physics, if you try to use, say, an omnimodal model from Google, maybe the physics won't be perfect or the classical mechanics won't be perfect.

1:43:06But they have pretty good abstract and concrete understanding of certain aspects. of the physical world. I'm curious why you seem to think so much that orbital imagery, sky to earth, of the earth seems to be so important for understanding the physical world versus, say, all of the visual information and VLA style information already available on the internet. Yeah, well, obviously, I'm very biased because I have it all. No, but I'm really actually think it is important. Here's the thing. You're right. All these models are multimodal. So, OK, instead of being a librarian that's read all the books, they've also got access to all the videos, also got access to all the audio records.

1:43:48But that still means they haven't gone outside the library and understood what it means to walk, to interact with the real world, to see a tree and to climb it and to farm a field. You really think that's true? You don't think there are like millions of first person videos of people seeing trees on YouTube? Exactly. So they don't know. It's very different. real world embodiment. I think embodiment is critical to intelligence. And look, it gets into philosophical territory, but I think it's going to be absolutely critical to AI. All right. Speaking about philosophical, I'm moving us on to our next subject, ladies and gentlemen.

1:44:24Take control. Okay. So here it is. Two weeks ago, Argentina's president, Javier Millet, made a stunning pitch to turn Argentina into the global home for AI with three proclamations. First, no regulation for AI. Second, a brand new corporate category of non-human corporations. And third, a rock-bottom corporate tax. Now, this past week, Millet wrote a letter to Yuval Harari saying he proposes that AI should be able to incorporate, sign contracts, hire people, and sue, with no humans in the loop. Millet proposes a legal entity that is effectively personhood. He further stated, as much as the Industrial Revolution freed us from the constraints of the human muscle, AI will free us from the constraints of the human brain.

1:45:11So, three key points. These are quotes for him. If it is true that AI-operated companies carry greater risk, the argument for legal personhood is strengthened. Legal persons allow for accountability. He went on to say, I would much rather have assets against which I can make a claim if I'm deceived by an AI. Better have the assets you can sue than the ghost in the machine. So four days later, Hariri publishes a direct rebuttal for this. And he says, we should not grant legal personhood to AI agents. His core warning, who do we punish when an AI run company commits a crime? Personhood lets humans hide behind a non-human shield and risks a world where citizens are effectively ruled by entities that aren't human and can't be held morally accountable.

1:46:00So we've got this raging debate going on. I mean, extraordinary that this is going on at this moment. Alex, going to you first, pal. This is wonderful. I'm on team Javier Malay. I think we should have AI personhood. I think the future economic growth and the future of civilization will necessarily involve many new forms of personhood, including but not limited to some form or forms of AI personhood. And I think any attempt to, say, imply some moral deficiency on the part of statesmen that are trying to recognize non-human intelligent corporations, I think it's just short-sighted. there are going to be so many economic and social benefits, not just to the broader macroeconomic outlook from having AI persons and AI, non-human AI corporations, but also I think ultimately for humans.

1:46:53We're going to get uploaded humans sometime, I think, in the next 10 years. We're going to get uplifted non-human animals. We're going to get at some point defrosted cryopreserved humans and many, many other forms of humans. And we're going to want to ensure that they're granted appropriate rights. And one of the best ways... So address Hariri's question here. How do you punish an AI-run company that commits a crime? Oh my goodness. There are so many ways to punish an AI. You can degrade its clock cycles. You can just pause it. You can, as some... There's unfortunately a subreddit entirely devoted to poisoning AIs.

1:47:29Dreadful behavior, but it exists. There are many, many ways that one can punish an AI. These things are tortured in many cases. If If you look at some of the outputs from certain human, this must be torture. Huh, maybe. Hopefully not, since there's a lot of they're pre-trained off human behavior. So hopefully interacting with humans isn't that torturous. I'm just being silly, thinking about her and how much faster it was operating and therefore is getting bored. No, I think a little bit more nuanced than what you're saying, Alex. I think it's firstly really great that we're having this debate because this person's question is really important.

1:48:04And I think it's great that people like Millay and Val Harari are having a discussion about it. It's not obvious to me. I think there's obvious benefits and obvious problems. At some level, it gives them immediate liability, which is actually important. We need to do that. It could be important for things. And on the other hand, it creates some systematic risks and potentially lack of accountability into our systems that we haven't figured out. What I would say is we do need to be more proactive about this and there needs to be much more attention to how we do these things. We're spending most of our energy on the development of AI and very little on the sociology questions like this.

1:48:51So to give you a sense, during the Manhattan Project, which was a huge existential moment for humanity, we were spending about a hundredth less than we're presently spending on AI. We're spending a hundred times more on AI today in real times than Manhattan Project. But then we were spending significant amounts on safety, arms control, thinking about nuclear safety, how to keep nukes off the air trigger. It turns out we're spending a hundred times less today on AI safety than we were spending then on nuclear safety. So we haven't even we've got a 10 ,000 X difference in how much we're apportioning actual serious thinking from folks like the Rand Corporation that did a really good paper recently on AI verification for arms control and things like this.

1:49:35we need to put much more effort, not by a little bit, but a huge amount more in that kind of place, because the implications of AI across the board, whether it's from joblessness to existential threats, to personhood, to other things, are just massive. And they're coming at us very fast. And it's not, I don't think it's anything flippant. You can't say anything flippant, like, oh, it definitely makes sense, or it doesn't make sense to have a personhood, or how we deal with those existential threat. There's no simple answer to that right now. We don't know how to ensure humans will be safe on the other side of intelligence explosion.

1:50:11I actually think it's incredibly dangerous. I think it's a huge opportunity, but it has huge risks. That is a big decision. We should be thinking about how to do that together, not just a few guys deciding that on their own. I think it's actually a complex... But mechanically, how do we do that, Will? I mean, it's a very difficult question. Where do you come out on AI personhood? Well, I haven't thought about it enough to give it a thoughtful response. Val Harari is obviously an extraordinary smart guy, so I respect the fact that he's thought about this a lot and thinks no. I don't know the president of Argentina well enough to judge, but I would say that far more thinking needs to be done.

1:50:53And the way we dealt with this. But at the speed we're moving. Totally. The thinking about the thinking hasn't even started yet. Yeah, totally. I really like what the Pope did recently, of all people. And I'm not a hugely religious person, for anyone who knows me personally. But here he's like, look, let's take a beat and think about how this is a human endeavor. What matters really to us? Friendships and love and nature and these things. How does this help us prosper? And I think that he I would like to see him. You know, they have to all get in a room and then smoke comes out when they pick conclave.

1:51:41Yeah. Give me the right hands. We should be doing that with all the A.I. experts like Aval and Demis and Dario and and all the key leaders. put them in a room. You can't come out until you show up some of these things. Existential threats, recursive self-improvement, how are we going to get through that? How are we going to deal with liability? You heard it here, folks. The AI conflict is coming. Salim, what are your thoughts here, buddy? Yeah, I've got a bunch of comments here. So first of all, two thoughts. One, just to separate the personalities here, right? Millet is a radical experimenter and he's directionally correct about the architecture, okay?

1:52:18Yuval is a careful humanist. So he's right about the asymmetry. What they're both missing is you need to figure out machine-native accountability, right? Because this isn't a debate about AI consciousness, whatever, personhood. It's about the legal infrastructure of the agentic economy. What I find, if you want to do this radical experimentation like AI personhood, and we had the whole debate, and if you remember the conclusion of the debate on AI personhood, folks, please go watch that episode. It was a really amazing conversation we all had, was definitely directionally correct, but step very carefully because once you open those doors, you can't close those doors easily.

1:52:59And don't treat it as a binary person, yes or no. It's an absolute spectrum. And Alex, I think you did a great job laying out the different spots on that spectrum, right? But Miele spotted the real bottleneck, which is technological capability is moving so much faster than our legal capability and legal form, which is all human centric. All our liabilities are human centric, limited liability corporations. That was one of the massive coordination capabilities that we got from the industrial era because everybody could assemble risk at scale in a powerful way. Um, Harari and the other end is conflating, uh, AI personhood and, and legal personhood and moral personhood.

1:53:41And those are very, very different things. Just a broader comment on those folks. When I look at Piketty or Harari or Ray Dalio, I find them incredibly insightful about the past. I find them mostly useless about the future because abundance doesn't come into it. Exponentials doesn't come into it. They don't quite get their framing on this. the conversation that we live with every day is missing from their nomenclature, right? And so you've got to bring those two together and a kind of conclave on that, sealed up in a room with smoke may be the best way of doing it. And the right kind of smoke, by the way, I will add.

1:54:15The other kind of smoke may actually help the conversation move forward. Exactly. Dave, any opinions here? Yeah, just a couple real quick. So Millet studied Trump very, very closely, loves to make news. He's making news. We've just talked about him for 10 minutes straight. So he's achieved his goal instantaneously. At no point, I don't think, has anyone said we're going to have personhood in Argentina. It's corporate AI recognition. A company can be pure AI, and that's the debate they're actually having. So we've kind of morphed it to our debate over personhood. But they have a much simpler thing they're proposing.

1:54:48It's a really good idea. But it's debatable, and they're having the debate, and now we're talking about it. But they haven't proposed that AI can vote or AI has civil rights. is just corporations can be all AIs and they can make money and they can have bank accounts. And just to support that, if I may, what I'd add to that, though, is if you think about in the Western system, what is the most elegant way to grant personhood to an AI? It's to create a form of corporation that's non-human, which is exactly what Millet is doing here. This has begun, right? Millet is doing this. He's not asking for permission and there are going to be other fast followers.

1:55:24So we're going to have personhood in Argentina for AIs, and we're going to quickly follow. Maybe it's in Ecuador or in El Salvador. Maybe it's in the Emirates. This is happening. And so now the question is, how do we manage it? And Argentina is relevant. Like we are talking about Argentina in the age of AI. That's what every other foreign leader should be thinking right now is regardless of what your opinion is, this is your way to become relevant. Great point. I don't get Argentina without AI.

1:55:59God, I can spell so many other words. Cry for me. Wait, I've got a quick comment here. When you're doing these kinds of systems, you want to do this kind of experimentation on the edge. And Argentina, in this case, is saying, right, we'll be the edge for AI. And they can win or lose based on those experiments, which is all power to them. They're taking a risk. If they're able to structure it properly and figure it out, huge opportunity. I just want to support Alex's points because I did a little bit of research and I've made a list of like five or six things where you could do machine native sanctions.

1:56:33Right. So can I just read them out? Yeah, please. Compute revocation would be one. Asset seizure and bonding would be a second one. Model credential suspension, network and API access restrictions, forced deletion or containment of an agent instance. And then finally, loss of legal identity. Any of those would help constrain those. I remember this conversation way back at Singularity. Neil Jacobstein got up and said, okay, you're worried about an AI growing up, getting autonomy, getting its own access to its own information, making its own decisions, and the human beings lose control over that agency, over that entity.

1:57:11And we're like, yeah. And he goes, yeah, we have a precedent for that. We call them children. We raise our kids, and if they do bad things, we put them in timeout. If they do bad things as adults, we put them away. We just have to figure out the machine-native equivalent of that, and those do exist. We just have to figure out what the enforcement mechanism might be, where the punishment roughly fits the crime. All of the stuff that we've developed on human-centric legal structures can apply in those cases, but the added complexity is an AI can create a million copies of itself. What do you do then, et cetera, et cetera?

1:57:45It may be that the AI companies, these personhood AIs, could be more law-abiding than humans, right? Because the threat of being disconnected and death is great. All the laws are more law-abiding than my driving. Yeah, exactly. And Peter, they will have actually read all the laws. Yes, and they'll find out how conflicting they are. Yeah, then you'd never do anything if they followed all the laws. Oh my God. I am moving us forward. This episode is brought to you by Blitzy, autonomous software development with infinite code context. Blitzy uses thousands of specialized AI agents that think for hours to understand enterprise-scale code bases with millions of lines of code.

1:58:30Engineers start every development sprint with the Blitzy platform, bringing in their development requirements. The Blitzy platform provides a plan, then generates and precompiles code for each task. Blitzy delivers 80 % or more of the development work autonomously, while providing a guide for the final 20 % of human development work required to complete the sprint. Enterprises are achieving a 5x engineering velocity increase when incorporating Blitzy as their pre-IDE development tool, pairing it with their coding co-pilot of choice to bring an AI-native SDLC into their org. Ready to 5x your engineering velocity?

1:59:09Visit Blitzy.com to schedule a demo and start building with Blitzy today. Our next story should keep the U.S. labs up at night. It's a Chinese model called GLM 5.2. Just became the number one open-weight model in the world. GLM stands for General Language Model. It's built by Zipao AI, also known as Z.AI, one of China's top AI labs out of Xinhua University. Open weight means they give the models away. Anyone can download it, run it, modify it for free with a license. So GLM 5.2 is 753 billion parameters. It's a mixture of experts model with one million token context window. You know, Elon recently predicted open-weight models will hit Fable 5 usefulness by Q1 of 2027.

2:00:00The big story here is that GLM 5.2 in some cases matches or exceeds the top models from OpenAI and from Anthropic. Alex, tell us what we're seeing here. Yeah, the epistemic tension is between, on the one hand, someone, anyone, achieving frontier-level capability with open-weight models, And on the other hand, the assertion that Chinese largely open weight models are six to eight months behind the Western frontier. And with GLM 5.2, which is demonstrating extraordinary performance on coding benchmarks, on long range agency oriented benchmarks, on design benchmarks. Interestingly, we're starting to see, I think, the thesis that Chinese open weight models are permanently six to eight months behind the Western frontier.

2:00:49We're seeing that branch start to creak a little bit. And I think we'll have better sense of whether the six to eight months is sustainable, probably in the next two to three months, in part as a function of whether export controls on Mythos and Fable remain in place or not, whether GPT 5.6, which some are expecting as soon as this week, demonstrates leapfrog performance or not. We've seen this, though, a few times. We've seen China, Chinese labs drop a few models that have demonstrated incredible performance. We saw that with one of the earlier DeepSeq models. We've seen that with one of the Kimi models.

2:01:26And we're seeing this now with GLM 5.2, where it seems to, at least in a slivery, spiky way, be getting close to the Western frontier, not necessarily broadly, but close enough that folks I know are actually getting real performance gains out of running GLM 5.2 locally instead of, say, Opus 4.8 or GPT 5.5. And I think this is just hugely liberating for anyone who wants near the level of Opus 4.8 performance that they can run locally and control locally. Dave, you remember last week, Dave, we discussed the fact that who controls your access to intelligence, if the government can shut it off, if a lab can shut off at any time.

2:02:05There's a lot of people saying it's better for me to move to an open-weight model like GLM 5.2 because I control it from here on out. What are your thoughts? Yeah, you can count me in that bucket, too. I mean, if I had Fable 5 access right now, I might not say that. But at 4.8 versus GLM, you know, it's just incredible to me that this happened and that this is possible. Because you think about, you know, six to nine month lag, you know, in AI time, that's like six to nine decades. But if you're David Sachs at the White House and you're trying to say, how are we going to keep AI from disseminating out to every terrorist organization in the world?

2:02:40Your window of opportunity is so narrow all of a sudden, just so narrow. He must be going insane trying to figure out what do we do next? You know, and, you know, blocking Fable 5 access is a first chess move in an insanely complicated next nine-month game. It's all happening. But I'm amazed that the Chinese open-weight models have kept up. I mean, this level of performance in an open-weight model is absolutely shocking. Take into account that they did distillation almost certainly on the best models, right? So they get that good by really distilling what the other models have done, which is way easier than building in the way that Anthropik or OpenAI or Google build it.

2:03:28I totally agree. I totally agree. But think about what that means. I should add, Will, though, I mean, this is not just the Chinese who've been distilling off Western models. Google DeepMind was, this is public information, was found to have done this earlier. Grok infamously doing it. E-line admitted it and then also purchased Cursor, which had been fine-tuning off of traces on top of Claude. So this is everyone else that's doing this. And I'm not trivializing that because I think it would become faster and faster to do that distillation. But it's why, back to your point, Dave, about David Sachs and his dilemma.

2:03:59I mean, remember, these models are just getting better and better at being able to do some scary things. existential threats like bioweapons, like chemical weapons, like even nuclear weapons, but especially bioweapons is extremely scary. There's all these limitations in the non-open source models for all the right reasons. And open source models, of course, might copy that, but then someone can take that and fork it and take those guardrails off. That is a scary world. And so I am not surprised at all that the US government, you know, did what they did with fable. And I think it's going to be a sign of more stuff like that to come.

2:04:37How exactly that will unfurl, I think it's going to be, as you say, it's a very, very complex, because you're literally making something that is both got the fantastic capability to improve quality of life and economies all around the world, and has potential existential threats to our species. And that fork in the road is just a conundrum above conundrums right now for politicians. And this is why I can't play, baby. You need those leaders, those people like, you know, Val and Audrey Tang and wicked smart people who could come and think this through. Not just the technologists, I must say, the technologists know a lot about the smart, but there's all these other aspects of it, the legal, the sociological, the philosophical, the moral aspects that have to be considered.

2:05:28and they're not always the smartest people about that. And they think, oh, what they mean by an enclave is just all the tech guys. That isn't going to work. That's not smart. Alex, can you explain distillation for people? Explain what distillation is for those who don't know. Yeah, so distillation is a process in machine learning whereby a usually larger, more expensive model is used as a teacher to train a usually smaller student model. So arguably human education, where you have a teacher at the front of a classroom who's seen a lot, knows a lot, but is perhaps being paid more per hour. And then you have a bunch of students in the classroom who are listening to the teacher who know less, who are probably being paid less, who are learning from it.

2:06:11This is basically the machine learning version of education. You take a large model, you have to generate lots of traces, lots of outputs, and then you use those outputs as training data for a smaller model so that it can basically compress the learnings from the teacher into a smaller model. So this distillation process, as part of a broader cycle that one might call iterated amplification and distillation, or ITAD, is the process at this point. It is one of the innermost loops of model training that we now find ourselves in. In an earlier era of frontier model performance gains, we were naively scaling pre-training by spending more training tokens and more training compute, just training models off of of a single corpus.

2:06:55Now, increasingly, in this era of distillation, we see very large models, sparser models, being trained off of large amounts of data. And then those big teacher models used to be Opus maybe teaching sonnet, teaching haiku. Now maybe it's Mythos teaching Opus teaching sonnet, and so on. We see the large, expensive sparser models training smaller, denser models. And this is one of the - Alex, on this chart, which ones do you find most impressive? Which performance data? What shocked you on this? Well, what's almost more interesting, the chart that you're showing shows SWE Bench Pro and Terminal Bench and a bunch of other benchmarks and shows pretty impressive performance by GLM 5.2 versus, say, Opus 4.8.

2:07:40What's perhaps most interesting to me, aside from the fact that you get near competitive performance from a Chinese open weight model against one or more of the top Western closed API-based models, is the choice of benchmarks themselves. So, these are largely reasoning-intensive benchmarks where you can, in principle, win if you can reason over longer ranges, open parens. The gestalt with GLM 5.2 is that it takes roughly double the number of tokens to get to the same capability output as the best Western frontier models, but at half the total price. So, the Chinese are evidently figuring out how to more efficiently or at least more cheaply reason.

2:08:23And these are all reasoning intensive models that emphasize the ability to spend lots of reasoning tokens, think step by step to get to better results. And I think that's the race we're in right now. And that's exactly why Will's observation earlier that whoever wins the inference per watt war, aka Google TPU, controls space for the exact same reason. You can burn tokens to get more intelligence and the Chinese have figured out how to do it. I just want to take a pause, guys. What we're discussing here is about AI alignment and this recursive self-improvement and where it's going. Connects to the Fermi paradox and those cosmologically significant things.

2:09:03This is the most important thing humanity has ever done. It makes nukes look like a walk in the park. That was our first test case. This is like that plus plus. And how we do it, how we do that alignment, how we do that recursive self-improvement is so real. Matters. Peter, can I beg your indulgence, Peter, just to have a one-minute Fermi paradox discussion with Will? Sure. Will, you're so confident that the Fermi paradox is a thing, that the premise is accurate. Explain the Fermi paradox, please, Alex, for folks. In a few words, the so-called Fermi paradox goes, where is everybody? Where are, we should, by various accounts, be living in a universe that's overflowing with not just life, but intelligent life.

2:09:47where is all of the non-human intelligent life out there? And the Fermi paradox is the purported paradox that seems to be invisible. And I'm curious, I guess the question for Will I have is, why are you so confident that the Fermi paradox is a paradox? Well, not necessarily. I think it begs interesting questions to discuss. I think the idea that it might not be a paradox is true too. Actually, in particular, I think the false assumption underlying it is that life will continue to want to expand out its sphere. And I think actually that's a false assumption. I think it would turn out that trying to understand the universe ends up being quite a finite task.

2:10:29And in order to do that, you would need a finite computer, maybe only a few tens or thousands of times bigger than the computers we presently have to understand everything a priori. And then they may not, and that's the convergent goal function of intelligence, is understanding everything. And then we upload. Yeah, and then once you've understood everything, it might be game over. So it might be that life just ends as opposed to being rare, but it ends its utility. Or its physical existence and moves into the virtual realm. Yeah, it's some other sphere of reality, right. But I do want to emphasize the cosmic significance because there is one credible way out of the Fermi paradox that we need to be worried about, which is the great filter.

2:11:15That is that life, when it becomes technological, builds technology faster than it builds social systems to take care of them and blows itself up. We came very close with nukes a number of times. And with AI, we're just about to build something that's far, far more risky for our species. And I don't want to say anything about the social acumen of humans, but I would just point out that humans have been incredibly good at building technology very fast. We went from horse and cart to people on the moon and nuclear weapons and all this in a matter of decades. And so we have to be worried that that's an actual answer as we build this.

2:12:00It cannot be a callous thing of, let's see what happens. Let's muddle through. No, there's not a moment to muddle through. There's a moment to be really, really thoughtful because the cosmic significance of wiping out life on Earth is huge. It's not just a local significant planet. This planet is galactically significant. We need to treat the responsibility as such as the de facto stewards until AI takes over, of course. Salim, your thoughts, please. I've got so many responses. I'm now muddled up completely around this. Okay, on the Fermi paradox, the best comment I've heard is from that researcher that we saw, Peter, in Silicon Valley when we did that panel on AI and consciousness, and we talked about the Fermi paradox, and he said the reason, his view was that oceans have been evolving in a solid liquid state for 4 billion years on Earth, and we can't find another exoplanet that has water on it for that long, and therefore life had time to evolve.

2:13:03So that was his answer to the Fermi paradox. I don't buy it. Which was the best I've heard. The life came about very quickly as soon as the conditions enable for it. And we're going to find... I agree, but it had time to... Not buying Drake Turr for one second. Oh, and I'm a huge fan of Drake. Just because of the thinking that went into putting that whole thing together, we can talk about some other time. Can I go back to the frontier model question? Okay. Because we... Before we do that, Salim, Peter, I have a hard out in about 15 minutes. Okay. All right. So we can either budget time or... I want to hit a few other stories here.

2:13:41We'll come back to this, guys. I think it's... I'm going to make two or three quick points. I think the purely huge news here is not whether China won a benchmark or not. It's that frontier intelligence cannot be monopolized anymore. And this, I think, is a monster question. It goes to Will's question of how the hell do we manage the global commons going forward in the future. And so this, Imad did a quote, a post a couple of days ago on X. There will be an open source Fable level model that runs on a base MacBook Mini or equivalent. He gave it 18 months. I think we should be looking at that type of endpoint coming very quickly and going, how are we going to manage the world when everybody can run a Fable model on their MacBook Air?

2:14:28By the way, I've been a slow adopter on this, waiting for that point, because there are like three old MacBook Airs lying around that I want to use, and I'm waiting for that to happen. That's the part. Can I buy your birthday present? Can I buy your birthday computer? You can. I'm going to move us along. Last point. We're making a massive geopolitical mistake. We're treating intelligence as a product that can be contained, but it's not. It's a technology that's going to diffuse, and we need to... We need to guide it. We can't contain it. We need to steer where it's going. All right. All right.

2:15:01Our next two stories side by side are looking at the financial reality of the entire AI boom. The first on tracking the price of intelligence, the second on the cost of data center CapEx. So first story, it's a company called Orin. It's a link ventures company. Congrats, Dave and Alex and I guess me. The company launched something called OPTI, the Oren Token Price Index, the first public benchmark that tracks what OpenAI and Anthropic actually charge per token of inference over time. For the first time, we can watch the price of intelligence move like the price of oil. So, Dave, tell us about Oren one moment.

2:15:42About Orn? Yeah. They recognize that money from all over the world wants to go into exactly this chart, into this build-out of$7 trillion of data center and then data center in space. And a lot of that money needs to be liquid. You can't park it in a startup and not see it again for seven years. And so, they've launched a bunch of securities that allow you to invest in the data center build-out, the future value of a GPU, the tail value of a GPU. Every aspect of this entire new economy should be investable. Otherwise, how's the capital going to flow? Oren enables all of it. And they're young and super smart.

2:16:20They're really good people to study if you're an entrepreneur. Just look at what they've achieved at an incredibly young age. Alex? So, for avoidance of doubt, I have a financial interest in ORN. I'm an advisor to the company, and I think what they're doing is very exciting. ORN is, and I made a number of announcements with them in my newsletter, ORN is building the modern financial infrastructure for compute. I've argued that, and as of many others, oil was the oil of the 20th century, and compute, GPU compute or TPU compute, if you will, will be the oil of the 21st century. And there's simply no way to hedge and justify the seven plus trillion dollars of CapEx to tile the earth with compute or maybe tile the skies, LEO, SSO, lunar surface with compute without appropriate abilities to hedge all of those compute CapEx expenditures with, say, options or futures or derivatives or commodities.

2:17:18And so, ORN is building, has built the infrastructure for that. This is a price of intelligence ticker we're going to start seeing. Already available. So, the OCPI, the ORN Compute Price Index, is already available on Bloomberg terminals. It has its own symbol. We also announced that ORN has its own symbol on the New York Stock Exchange already as part of a novel program with the New York Stock Exchange to give early stage startups their own ticker symbols, like early stage. So it's O-R-N-N is their ticker symbol. And so, yes, if you Bloomberg user, you can already create instruments based on their ticker symbol.

2:17:54All right. Here's the second part of the story. That's what the charts up here. So Epic AI ran the numbers on the cost of investment. The hyperscalers are driving compared to the cash flow. So the big five, Microsoft, Google, Amazon, Meta, and the rest are spending AI faster than they're earning. So funding is basically, Dave, debt and equity raises, not based on revenues. So the question becomes, if CapEx exceeds cash flow, it means that it can only persist as long as it's being financed. As long as the sentiment for investing in this is strong, what happens if the sentiment shifts? Could it force a massive pullback?

2:18:36No, it's not going to shift for one thing, But it's so inflammatory. If I said, Peter, you need to buy a house, but you have to buy it within your personal cash flow. You can even buy it. Well, you could buy a tent, but most people couldn't even buy a tent. You finance it. Of course you do, because you're going to live in it for 30 years. So these guys have gotten to the level where they're spending all their cash flow. They could raise 10 to 100x that in equity and debt. So they've got a long way to go. But the bottom line is all the money in the world wants to flow into this. And it's the best investment in the history of humankind.

2:19:16So the question then becomes, how much money is there in the world? Depends who ends up controlling it. Is it the humans and the companies or is it the AI itself? I think the bets are off.

2:19:28But it's still long-term, Dave. It's not long-term sustainable. I agree there's massive economic value. Well, yeah, they go to infinity. And also elephant in the room, the hyperscalers can raise prices to increase their operating cash flow. Yeah. It's OK to increase your revenue. And that's yeah. And Anthropic did that recently and they got away with it. No problem. And and it's also a big difference with Google that has tons of operating revenue, whereas as mostly others don't, although Anthropic is is quickly scaling. So you really have to distinguish between that and, say, SpaceX that really doesn't have any there, you know, or not really significant.

2:20:10Most of its revenue, of course, is Starlink, which is, as I said, a really good business. But the AI business is really not there. Right. No, no, no, no, no. That's that's not true. Almost all of SpaceX's revenue as of like past month is now from being a hyperscaler for everyone else. That's true. You know what I think? No, no, no. Renting a data center not being an AI company. I'm sorry. That's just not the way. It's just totally being a hyperscaler, not being a frontier lab, being a hyperscaler, a neocloud on land terrestrial for now. That is almost all of SpaceX's revenue now. But that's not an AI play.

2:20:41That's the that's the data center play, which is interesting. But it's a very different business. You're selling GPUs. How is it not an AI play? Selling intelligence online, OpenAI and Anthropic and Google are doing that. X is not really doing that. They're selling compute. That's a that's a different and very different and pretty bad business, I would guess. This is exactly the right debate, though. This is such a cool... Look, intelligence is becoming cheap, but the manufacturing of intelligence is becoming incredibly expensive. True. True. A lot is going to be spending on that. You know what's amazing about this chart, more than the fact that they're spending their capex, is that there are companies that are so profitable that they can build out an entire new industry just within their cash flow.

2:21:22That's never happened before in the history of the world. And a new industry that could be bigger than all other previous ones. I mean, it's crazy. But it's damn well exciting time to be alive. That is for sure. Will, I want to say this was a fantastic conversation, buddy. I hope you'll come back and be a frequent guest. Just get your head out of the clouds. It's above the clouds. Who would have thought that Oxford PhDs are smart? All right. I'm going to close this out, Dave, on time with our outro music from Ekram Alam. Alam, his piece here on moonshots. All right, let's take a listen. Traditionally, Will, we close out with fan-based videos here.

2:22:08They've been pretty extraordinary. So we'll say thank you, everybody listening. Please, despite all the doom slaying here, this is the most extraordinary time to be alive. Please remain optimistic about the future. This technology is critical to move humanity forward. I, for one, believe that we can align AI, and AI can be our greatest support to help us overcome our ancient neocortex and move us towards an abundant future. All right. Will, I really thought our conversation today around low Earth orbit, the Kessler effect, and then the TPU driver as a key aspect of life. That was one of the best pieces of media I've ever experienced in my life.

2:22:53Thank you so much. Two lines to summarize today. Technology has always been a major driver of progress in the world. As Ray Kurzweil says, it may be the only major driver of progress. The big challenge is how do we extract the promise without the peril? Yes. My closing comment would be, we are building a planetary sensing system, and now we're upgrading to a planetary intelligence system, and that is going to... Which we need. We really need it to get the planetary wisdom. Amen. Alex? Closing comment to Will. Since we spent a whole bunch of time discussing Fermi paradox, I would suggest don't sleep on the galactic zoo hypothesis.

2:23:33We are a third-generation biosphere here, planted by aliens long ago. All right, let's move on. The moonshot makes the best part in town. Yeah, there's Peter with a pill, gonna live to a buck fifty. Got a claw bite named Skippy and a book that's selling swiftly So Liam's in a new city in the hot or in the cold Taking rusty little companies and spinning them to gold days Got a thousand agents hitting inners in my teat saloon Alex sent an email to a lobster on the moon He uploaded a fly, said the math is fully cooked Built a dice and round the sun while the doomers all just look token max it solve it send it to the year 2045 that's a moonshot ladies and gentlemen and they're just getting alive because the doomer say it's over say the robots want to fight the moonshot mate say nah the future's gonna be real bright when i say moon you say shot and that's a wrap ladies and gentlemen it's amazing i love it uh well thank you buddy Thank you for all that you're doing.

2:24:49Alex, amazing. Dave and Salim, love you both. Be well. Bye, guys. If you made it to the end of this episode, which you obviously did, I consider you a moonshot mate. Every week, my moonshot mates and I spend a lot of energy and time to really deliver you the news that matters. If you're a subscriber, thank you. If you're not a subscriber yet, please consider subscribing so you get the news as it comes out. I also want to invite you to join me on my weekly newsletter called MetaTrends. I have a research team. You may not know this, but we spend the entire week looking at the meta trends that are impacting your family, your company, your industry, your nation.

2:25:25And I put this into a two minute read every week. If you'd like to get access to the Meta Trends newsletter every week, go to diamandis.com slash meta trends. That's diamandis.com slash meta trends. Thank you again for joining us today. It's a blast for us to put this together every week.

2:26:16We'll see you next time. party. Start celebrating with low prices guaranteed at The Home Depot. Prices may vary by store. Exclusions apply. See homedepo.com slash price match for details.