OpenAI切断Elon的Cursor、人类首个星际探测器与Trump的核动力火星飞船 | EP #285

Before we move on to the news this week, we have a special announcement—a session on Moonshots from our own AI, Alex Wiesner-Gross. Alex, if you could join us, please, and present our 2 guests as we prepare for the first interstellar mission, Fermi Explorer.

Amazing, Peter. People who constantly watch the podcast may remember Philip Johnston, who was on with us before—the founder and CEO of Starcloud, an orbital company for data centers. Today, for the first time on Moonshots, my co-founder at Physical Superintelligence, Matt Pines, joins us.

Philip and Matt are here to join the Moonshots team. I think this is an exclusive podcast announcement. How banal does that sound? It’s about the first interstellar mission to Alpha Centauri.

Philip and Matt, go ahead. How are we going to get to Alpha Centauri?

Maybe I’ll describe the mission, and then I’ll tell you a little bit about how PSI eventually played a decisive role in the inception of this mission, and Matt can tell you about the background.

The last time we were on the podcast, after we finished recording, at the end I said, “Oh, by the way, guys, I plan to send a spaceship to Alpha Centauri.” We set ourselves certain restrictions because we really want this thing to actually fly away.

The first restriction is that we want it to get at least 99% of the way to Alpha Centauri within the next 80,000 years. That actually minimizes the cost of fuel. Anything longer than 80,000 years means more fuel, while anything less than 8,000 years also means more fuel.

The second limitation is that we want to launch it within 3 years. Third, it must have a useful payload weighing 1 kg in a 1U format. The last restriction is that it has to cost less than $1.5 million for design, construction, and launch because, in essence, we are financing it.

Wait, can you repeat that again? $1.5 million?

$1.5 million.

What? I had in mind that this was an amazing goal to achieve. This is the seed round for a startup with MIT.

Yes, yes. This is half a seed round.

Such low numbers don’t count. It’s interstellar travel for pennies.

But 15,000 years is a little longer than for most startups.

This is true. I’ll get back to why we do this in a minute, but first, briefly, I’ll tell the story of how PSI got involved.

We spent 6 months trying to find a trajectory that would make sense because the main challenge was using solar-electric propulsion and an ion engine. This is the same as on the Starcloud-1 satellite. It’s very cheap.

The problem is that the farther you are from the Sun, the less energy reaches the solar panels, so you need larger panels.

When you are beyond about 2 AU, you get very little energy and need huge solar panels. So we tried everything: flybys of Jupiter and gravitational maneuvers toward the Sun. We involved 2 guys from JPL to look at this. They spent a few weeks studying it, but they couldn’t think of anything.

We spent essentially 6 months doing calculations, trying everything we could. Then Alex said, “Oh, you should talk with my guys at PSI.” I thought, “Oh, yes, it has begun. Now they will be ours—the guys from JPL.” So I didn’t answer for a week; I just didn’t answer. Then Matt wrote again: “Hey, send some more details about the characteristics of this mission.”

I thought, “Oh my God, okay, I’ll do that. Alex was satisfied.” I sent the characteristics. A week later, they came back with an incredible report. I think they spent tens of billions of tokens on this. They came up with an incredibly nonintuitive and innovative trajectory that allows you to optimize the mass budget and velocity.

In essence, we spiral away from Earth and then, counterintuitively, turn on the engine in retrograde to slow down and direct ourselves toward the Sun. We do this for approximately 5 years. We perform 5 retrograde pulses at the point furthest from the Sun. Then we start doing what they call perihelion impulses—that is, we turn on the engines as close as possible to the Sun. They called this maneuver “perihelion pumping.”

This has 2 incredible advantages. First, we turn on the engines at the point closest to the Sun, so less mass is needed for the solar power system. Secondly, it uses the Oberth effect. The Oberth effect is the idea that you get more energy from a certain amount of thrust when you move faster, and you are moving fastest at perihelion. Honestly, I find it incredibly impressive that they came up with this. At this stage, I’ll give the floor to Matt, and he’ll explain how they did it.

Well, of course, it’s a happy coincidence of circumstances. I mean, AWG is my co-founder, and we found this connection at exactly the perfect moment. It was a unique synchronism: we announced the Fermi Explorer mission and this partnership today, along with the announcement of the initial capital for Physical Superintelligence and our exit from stealth mode, all on the same day. Therefore, this mission is proof of concept of what we’re building here.

As you mentioned, such highly technical and scientific challenges are constrained by people who have gone through 22 years of prior education, additional training in postgraduate or technical positions, and then become part of organizations coordinated through corporate, academic, or government bureaucracies. These are the limiting factors for our scientific and technical ambitions. That’s why we’ve had to involve large-scale national institutions to organize such breakthroughs and grand scientific and technical initiatives, such as sending a spacecraft beyond the limits of the solar system.

This is proof that 2 small teams—both startups, one in the space industry and the other in AI for physics—can join forces and develop a mission that pushes the limits of the possible. We took it as a secondary task to test our internal technology. We have a staff astrophysicist, but honestly, we just made requests of the system and then finalized the product to make sure it had the correct graphs and charts.

Apart from that, we didn’t intervene at all. We were as surprised as Philip’s team by the optimal flight path it developed. We definitely didn’t rig the results. Human intervention was almost minimal, and the AI developed a trajectory that corresponded to the mission’s strict constraints. I think this is the first of many surprises we expect as we direct AI physicists to solve these extremely valuable technical and scientific tasks.

Alex, I want to delve into the details, but I thought it would be interesting to show the Fermi mission video. Before you run it, do you mind if I describe what’s happening? Without context, it can look a little amazing.

The idea of the mission is that we hope to become the first to leave Earth for another star, but at the same time, the last to arrive at another star. Let’s say that in 1,000 years we’ll have better propulsion technology. Even if it’s only 20% faster, which is a very conservative assessment, by the time we arrive—after 15,000 years—a colony on Alpha Centauri will already exist.

They’ll have had time to build things such as Dyson spheres, O’Neill rings, and all those amazing and incredible things we imagine in science fiction. Then, according to our estimates, using essentially modern propulsion technologies, it will take 5–10 million years to populate the galaxy. From there, without any special effort, it will take about 1 billion years to get to Andromeda, and from there about 5 billion years to populate the local galaxy cluster.

I’ll return to why we’re doing all this, perhaps after the video, but I want people to be aware that what you’re about to see is the next 5 billion years of history. Oh my God. Maybe it’s worth emphasizing, Philip, what a conservative upper limit that is.

I don’t really think that humanity’s civilizations will need 5 billion years.

Do you think creating a faster engine will take 100 years?

I think we’ll get it in 5–10 years.

Right. Okay. Let’s watch the Fermi Explorer mission video.

Alpha Centauri. One day, humanity will go there. A mission that will begin civilization in the universe.

Good. Let’s start. Let’s move on to some basics. At $15 million, how do you reach escape velocity?

Honestly, I think we can fit in $10 million, but I didn’t want to point this out because the PSI document said $50 million.

It’s not really very different from a Starlink satellite. This is a small satellite, with a mass of about 100 kg. We can launch it as rideshare cargo into any low Earth orbit. Usually, that would be a SpaceX Falcon 9. That’s right, so this will cost about $500,000.

From there, we spiral outward for a year and a half. It takes approximately 7 km/s of delta-v to go from Earth orbit to solar orbit, which is quite feasible with conventional engines and tanks. Then we perform braking impulses and begin moving toward the Sun.

This is a spacecraft weighing about 100 kg, of which 60% is just xenon. The total mass of the device is 100 kg, but most of that is xenon. We use standard ion engines based on the Hall effect.

Just for a moment: no laser sails, as in Project Starshot?

No laser sails. This thing will fly in 3 years and will reach Alpha Centauri. It’s a 25-day trip across trillions of miles—approximately 4.3 light-years.

We talked about ion acceleration based on xenon in that podcast. Maybe you want to give us a short overview.

That’s right. This has become quite a common phenomenon in the satellite industry. It’s essentially a mini particle accelerator. Some of them can be tiny enough to fit in 1U of space—that is, 10 cm by 10 cm by 10 cm.

This is madness: a particle accelerator in a toaster. I think that’s essentially what it is.

Yes, it sounds strange, but that’s how it is. It throws individual particles—individual xenon atoms—back at a very high speed. You need to do this because your limiting factor is the amount of fuel. If any of the fuel comes out at a speed lower than the maximum possible, then you don’t get as much thrust as you could.

And I have one more reason for the name. It’s an allusion to the so-called Fermi paradox.

Enrico Fermi allegedly asked, “Where is everybody?” Given the great amount of evidence that our universe seems fundamentally friendly to life—to intelligent life—where are all these other forms of nonhuman intelligence in our galaxy?

I think Philip, Matt, and I have discussed this many times. I think there are 3 main possible solutions if the Fermi paradox is a paradox at all. In my opinion, there are 3 most likely solutions.

The first, which I consider the least probable, is that we’re the first. Humanity may simply be first on the cosmic stage. In that case, we’re probably obliged to start sending probes, as Fermi Explorer does, which should be humanity’s first interstellar probe, and begin to master our galaxy.

The second possibility is the existence of a “great filter”—something that filters us out for a certain reason, possibly related to technological development or some hidden risk associated with simply existing in this universe.

Some version of a “prime directive,” as it were. Some version of a “prime directive,” maybe—although, actually, the “great filter,” in my opinion, differs from the “prime directive.” For some reason, the universe destroys civilizations after they reach a certain stage.

If that’s true—if there is some risk, like in The Three-Body Problem, or perhaps some hidden physical risk in our universe to survival or development beyond a certain stage—then we need to start spreading our materials and infrastructure so that, if humanity is destroyed before we get through the next key stage, we can avoid the great filter.

That is the second reason why we do it. The third reason is the one you mentioned, Peter: the “zoo hypothesis”—as if we were in a galactic zoo, perhaps a petting zoo, and surrounded by a nonhuman mind that keeps us behind bars. If you are an animal in a zoo and want to attract the caretaker’s attention, what do you do? You start throwing food through the bars to attract the caretaker’s attention.

Therefore, this is the third reason: we are behind bars. I think that, of all 3, the third option is the most likely, but I would be interested to hear everyone’s opinion. Are there any other suggestions or favorite solutions to the Fermi paradox?

Well, one solution is that life is not capable of surviving the nuclear age or the era of artificial superintelligence. Another option is that they are out there somewhere; we just don’t hear them.

I said before the show that I was on Mount Athos, in a Greek monastery. At the end of the day, at sunset, they rang the bell to summon the monks to prayer. At that exact moment, my mobile phone rang, and I understood that they use this ancient mechanism of communication—the bell. They were broadcasting at frequencies of 2.4 GHz, but I did not accept them. So the question is whether there are better means of communication. In interstellar space, there may be a lot of internet traffic; we just aren’t able to receive it yet.

Yes, I think so. Physics is the core of civilization, and the opportunities available to civilizations are limited by their ability to use this knowledge to create useful technologies. The rapid development of AI is quickly turning into the rapid development of physics, and this will lead to a significant acceleration in humanity’s ability to explore the universe and use every degree of freedom that physics provides us.

1. AI’s Growing Energy Bottleneck

If there are tricks that we can use as a society, then these are tricks that others may already have invented. So we’re racing quickly toward this state, and then we will find out: are we the first in this neighborhood, or are we now, you know, asking for admission to the space club? The entrance ticket is: can you master this?

I think the main mission of this is to force people to dream again—to begin putting audacious goals before themselves and reaching them.

I had in mind, just to clarify one point: is this a flight past Alpha Centauri, or a flight approximately past Alpha Centauri? How precise, in your opinion? I think it would be good to have a guidance system that could really direct it to the planetary system. We expect quite a large deviation from the target.

So the goal we have set before us is to overcome at least 99% of the way to Alpha Centauri. We are currently approximately 260,000 AU away—that is, astronomical units, the distance from Earth to the Sun. We’ll get to within 2,600 AU, or 2,600 Earth-to-Sun distances. That is far away, but the device will be within the Oort Cloud and could be discovered. We assume that there will be retroreflectors and other means to make it noticeable. Yes, we expect it will be electronically dead, obviously, after such a long time.

It’s definitely worth noting that Philip and I made a bet with Dave regarding the commercial market for interstellar flights. This is now structured as a nonprofit mission, Fermi Explorer. I bet Philip that, even at an absurdly low price of $10–15 million for a very long and slow mission to Alpha Centauri, there probably is a hidden commercial market for anyone—any small government or organization—that wants to start launching probes into deep space.

My bet is that this hidden market for commercial interstellar flights exists. One of the markets is astronauts. My friend, Charlie Chafer from Houston, actually bought a place on a Pegasus rocket mission from Orbital Sciences to send about 5 grams of the dead’s ashes into Earth orbit. So, in any case, if you want to send yourself away from Earth, you can do that right now.

I think that, for scientific purposes—researching the outer part of the solar system—every state, I would say, might be able to afford to send one at these prices, which, again, are impressive. They might be able to afford to send their own probe to another star system. As far as I know, this is the first time this has become possible for humanity.

What’s critically important, Peter, is that this whole mission would not have happened without Moonshots. In the causal, consequential history of human civilization, projects like Moonshots became a catalyst for humanity to send its first probe to the nearest star.

I like the fact that the artificial intelligence system was really able to calculate this trajectory. I have a question: is this unique? Has this ever been done before? And was this confirmed by trajectory experts outside PSI?

Yes, it was confirmed by several trajectory experts, some of whom had previously worked at NASA’s Jet Propulsion Laboratory and elsewhere. Honestly, if we brought together a group of astrophysicists, gave them $1 billion and 5 years, I’m sure they would have found this trajectory. The point is that it was done in a week.

Starting the engine at perihelion to use the Oberth effect is not news. I think what is surprising is that we expected to need to reduce perihelion through gravity-assist maneuvers, as has been done in almost every other NASA mission. That wasn’t done. It was simply: “Okay, let’s just turn on the engine in reverse and start moving slowly right now.”

Hmm. What is it?

The simplest, cheapest, and probably most obvious way, but somehow none of us had thought of it.

Interesting.

Yes. In general, this took probably 5 or 6 hours of human time during that week. That’s an acceleration of orders of magnitude compared with what previously required entire NASA engineering teams to spend potentially months on. This demonstrates the level of acceleration we’re observing.

Again, it’s the sharp edge of where our opportunities for scientific and technical breakthroughs lie with the help of such systems. We didn’t know—we hadn’t tried yet—how sharp exactly this edge was, and thanks to this amazing partnership, we found what is now possible: acceleration by many orders of magnitude for similar mission-planning tasks.

Was this really dozens of billions of tokens’ worth of work during that week?

I think, in general, about 10 billion tokens. Of course, it depends on how you count tokens—input, output, cached data—but yes, a total of 10 billion.

The system designed and launched many Monte Carlo simulations. You’re thinking about 3D trajectory models and analysis. It was not only astrodynamics and mission planning related to choosing the right orbital trajectories, but also multivariate optimization of costs and launch windows. You had to configure all these variables to make it work.

This is interesting. We’ve been building a lot of very complicated coding assignments and other ultrahigh-tech work, and this took about 10 billion tokens of thinking and produced approximately 100,000 output tokens. It seems like a lot of projects that have nothing in common with one another stop at this kind of proportion. That’s simply stunning.

If you ask what 10 billion human-effort tokens of thinking represent, this is what Philip said: probably thousands of people working for 10 years or more to get there. Actually, probably even more. And that’s compressed into 1 week. So, what models did you use?

For this, we actually used our versions of these technologies with open source, since this is an open-source project. People can find the package with the Get Physics Done code, which we released a few months ago.

We’re a public-benefit company, and our mission is to open-source and commercialize transformational new physics. We’re an AI-first physics laboratory created to expand the boundaries of what these systems can do, both for fundamental and applied physics. We released this package as an open-source tool.

Of course, there is a version that we use internally, and we added part of this to our core technology. But since it is an open-source package, it’s for open science and to demonstrate how we can lower the bar for small teams using current cutting-edge capabilities to achieve exceptional results and expand the boundaries of what is possible.

It’s worth noting—and congratulating Matt and Alex from PSI on the funding. You just announced it. What was that—a round?

This was our seed round. We live in an era full of AI, where you can get seed rounds for $58 million. That’s why we’re looking at Philip Johnston, who sets the bar, and trying to overcome it.

We are very proud to announce it, led by Breakthrough Energy Ventures, our great partner. They invest in deep, advanced technologies: fusion, quantum computing, energy, materials science, and so on. We’re very proud that they are our main investors, as well as the wonderful list of other investors who supported us.

Yes, we are just leaving stealth mode, and you’ll hear a lot more about us in the coming weeks and months.

Gentlemen, I wish you incredible success in this mission. I know that many children will start to dream. I would, of course, have wanted to be there when it lands, but 70,000 years is a bit too long-term.

You are only invited for launch.

Good. Oh, cool. In fact, we should conduct a live podcast broadcast when it happens, in 2029. So, yes, I see this as a challenge for PSI. Let’s start this together with Philip Johnston and his team. The goal is to catch up with them, if not get ahead. Yes, I imagine you waving to him from the window as you head to Alpha Centauri.

Gentlemen, Philip and Matt, thank you very much for joining us today. Congratulations on this mission. It really is about to inspire children to dream again about what might be possible.

It’s shocking that no one has done it yet. There were several attempts and several studies. There was the Breakthrough project proposed by Yuri Milner: Breakthrough Starshot, using solar sails and ground-based lasers. But I haven’t heard anything about this for a long time.

I didn’t hear anything either. It’s official: it died.

It died. Well, I know some people who were involved.

It died.

I would say, Philip and Matt, that it would be interesting to hear your opinion.

I think it died because it relied on technology that was too difficult with current capabilities, especially the drives, which were simply too difficult to build. In particular, high-power lasers were not ready.

Whereas, in my opinion, the uniqueness and appeal of the Fermi Explorer mission consist in the fact that it actually does not need new technologies. It can run on technology that we have today. Therefore, I expect it to be the first successful interstellar mission.

Thank you again, Matt and Philip, for your time today.

Thank you very much.

Oh, that sounds good. Nice to see you.

Thanks, guys.

Thank you. Yes, I just said, Dave, I think you noticed one of the most interesting moments: 10 billion tokens. This is a kind of standard for the complexity of the task.

At some point in the future, if we record the capabilities of the model—and they, of course, grow over time thanks to iterative reinforcement and distillation—we’ll look back and say, “Oh, that one complex mathematical task? Oh, that was a Level 9 problem, 10^9 tokens. Then there was a Level 11 task.” We’ll be comfortable using a logarithmic scale for everyone to discuss complex problems.

Yes, I really feel that this project is significantly more important than 80,000 years in the future. Precisely from the point of view of planning: the way tokens were used to create a plan, which previously had not fallen into the purview of astrophysics. And then it can immediately go to implementation.

And this is a sign of the times, right? Thinking will significantly outpace implementation in biotechnology, physics, literature, and any other field where you can spend 10 billion tokens over a few days and get an incredibly detailed result ready for implementation. This is a very good example of how this will change over the next couple of months: massively accessible thinking and intelligence everywhere, along with all these bottlenecks in implementing ideas in the physical world.

2. OpenAI Cuts Off Elon’s Cursor

Undoubtedly, I constantly speak on podcasts and in other places about how the singularity, in a sense, can be described as all science-fiction tropes happening everywhere simultaneously. There is a science-fiction trope for this in the universe of Isaac Asimov, where artificial intelligence is needed for problem-solving and interstellar travel, after which humanity settles among the stars. I really think that this is the most likely scenario. AI will solve the interstellar-travel problem, and humanity will settle among the stars.

On this note, I’ll move on to the main news this week. There are many interesting stories.

This week, the drama between Sam Altman and Elon Musk flared up with new strength. OpenAI has stopped supporting Cursor. They wrote Elon Musk a letter stating, basically, “First, remember that SpaceX recently bought Cursor for $60 billion, a programming platform that has historically depended on GPT models from OpenAI.” And they announced, “We’re closing this down. We will no longer allow Cursor to use GPT models.”

The question is: why? OpenAI stated the following: “We are making this choice because we cannot be sure that SpaceX will use our technology within the terms of service, based on our experience with contract violations by Elon Musk’s companies.”

And, of course, what was Elon’s answer to this? It was quite passionate and emotional: “I don’t care. It doesn’t matter to me at all. Sam Altman and Greg Brockman are absolutely unreliable scum who stole a nonprofit organization with open source.”

So, there we have it. We’re observing the continuation of this soap opera. A few hours after OpenAI withdrew from Cursor, Anthropic intervened and immediately promised support for Claude in Cursor.

The wording that appeared on X was interesting to watch: “Sam is now fighting one against two huge competitors, Elon and Dario, who have formed a strategic alliance.”

Well, it’s not coincidental that GPT-4, which is an incredibly wonderful model, came out immediately before this step. If Sam had tried to do this a year ago, he would have said, “Oh my God, now I have big problems.” But now he has an incredibly competitive platform, and Codex is now really good.

Therefore, I think what is emerging here is the following: Codex by OpenAI on top of Soul, running on Amazon Bedrock—this is a very good standard corporate response. Everyone wants access to corporate income.

Sam was very late in transitioning from the consumer market to the corporate market, but now he has the whole stack ready. This is the next step: “Okay, here you go—our vertically integrated stack.”

The Cursor-Anthropic system looks like a Rube Goldberg machine. We’ve actually created simply the best corporate product. I use both at the same time. They’re right here on my laptops, and I use them every day. I spend a huge number of tokens on both.

Over the last month or so, the Claude Code-on-Sonnet combination on AWS Bedrock has become phenomenally good for businesses.

Also, Dario is a bit trapped in his ethics. If you read Elon’s post, he hints that Greg and Sam are bound by such prejudices, but Dario is actually very tightly constrained. He attracted a bunch of talented people, all of whom are concerned about what AI can get out of control.

So, if you use Anthropic on Amazon AWS, your intellectual property is transferred to Dario for 30 days for verification of everything you do. He claims that this is safety-critical for humanity, but it also opens access to all your corporate intellectual property: every token, every request, and every answer goes to Dario’s headquarters, even if you’re using AWS. Corporations don’t tolerate this.

I think Sam is taking a very confident step now. I don’t think he is isolated or cut off from the world. This reminds me of Bill Gates’ approach: DOS, Windows, working with Microsoft Word and Excel. He seems to be saying, “Listen, I’ll just do my best to develop a truly good product in the future.” And it is really very good.

Alex, do you agree?

I have an alternative theory regarding this matter. I think this is all just for the sake of chains of thought. It’s always a question of who receives those chains of thought. This is exactly what is happening with China, with those advanced Chinese laboratories which, as they say, use intermediaries to steal chain-of-thought data from Anthropic.

You remember that last year the situation turned around in the opposite direction: Anthropic closed access to Windsurf after Google DeepMind acquired Windsurf, probably because of access to chains of thought. Do you remember that xAI purchased—or actually captured—Cursor to get access to chains of thought from Anthropic and OpenAI?

I think OpenAI is worried because, as a result of all these mergers and Cursor’s acquisition, SpaceX gets access to the chain-of-thought histories of users who interact with advanced OpenAI models, and then that gets to SpaceX. Actually, I think access to chains of thought and their history was almost the only technical reason—apart, perhaps, from financial reasons—why SpaceX went for the acquisition of Cursor: to get that set of thoughts.

I think OpenAI is probably rightly concerned that all these considerations will eventually fall into Elon’s hands.

You know, with the speed at which alliances are created and broken up, I think the question is how soon Elon and Dario will quarrel.

Well, that’s the same question, because the largest beneficiary of the war between Elon and Sam is definitely Dario. Dario desperately needed computational capacity from Colossus in Tennessee, from Elon. Therefore, he pays crazy money for it and begs and asks, but Elon can take all this away from him any day.

But now, when Elon really needs Anthropic inside Cursor, you understand, without OpenAI, you have only a few options, and you don’t really want to use all these Chinese models. So what remains? Anthropic remains, and Grok. You can’t use Gemini; if you try, it won’t work.

Therefore, for Cursor, it is very important that Anthropic took a step toward them and said, “Yes, we’ll support Cursor in the future,” to keep this user base satisfied. And now Dario has the trump card in the game, allowing him to somehow balance the incredible level of Elon’s control over computation.

I think that’s how we get to vertical integration. Anthropic needs compute, and SpaceX, for its IPO, needed a splash of income that came from the fact that they almost overnight became a hyperscaler, attracting large tenants who have become anchor clients of SpaceX’s platforms.

Elon doesn’t like Sam, but the enemy of my enemy is my friend. I think the result at this stage is fairly determined, but I think that’s all.

It’ll end with everyone receiving their own Dyson Swarm.

I agree. Everyone is moving up and down the stack. We hear about it regarding chip designs from all these players. This will be interesting. I mean, this is a significant ongoing battle of personalities and philosophies.

Yes, and I also think that, to your question, Peter, we didn’t give a clear answer: Will Elon and Dario be best friends in a year or two? When you look at their personalities, everyone says, “No, in any case.” You have 2 big egos and completely different political views. It can’t be that they’ll be friends in 2 years. But their interdependence is becoming tight enough.

Mhm.

I wouldn’t be surprised if this duopoly lasted for a certain amount of time.

Well, I wouldn’t mind being surprised. As you said, everyone builds complete vertical stacks. What is the probability that Grok will become an incredible platform for programming and that Cursor with Anthropic will be replaced by Grok?

I think Grok is now—this is a blurry concept. I’ll tell you straight: today, Grok, judging by its capabilities and interaction with it, looks like yesterday’s Cursor, and yesterday’s Cursor looks like a Chinese model with open weights, further trained in the reasoning chains of Claude. Could Elon turn around tomorrow, make an agreement with Anthropic—which he now needs, to some extent, for data-center infrastructure—and release a Claude version under the Grok 10 brand? I think he could.

Okay. Well, the whole world of foundation models is not at all in Elon’s style, because it’s usually a group of 5 or 7 really brilliant, very united people, like the teams in China that are constantly making incredible breakthroughs. This is Anthropic’s DNA. And Elon’s DNA is large-scale infrastructure projects—Tesla, SpaceX, Colossus—which are very different from the style of that united, brilliant team.

Therefore, there’s no reason to think that Elon will wake up one morning and figure out how to create a great foundation model. The evidence at this time indicates that this isn’t happening at Grok. Never, never, never bet against Elon. He doesn’t like to be second. He also doesn’t like dependencies. I was present at conversations with him, and he says, “I don’t depend on anyone. I don’t care.”

We’ll hear a little about it later, when he understands that he can’t get enough turbines for his gas engines or solar panels, so he’ll build them himself. That’s what he does. He performs vertical integration of the entire technology stack. That’s why the decision about super-voting rights for Dario is an important question that still remains suspended.

One of the scenarios according to which Elon solves this problem is that, when he gets very big, he buys Anthropic for $1 trillion or $2 trillion and just weaves it into his empire. And I’m confident that the members of the board of directors and investors would be delighted. But I don’t think Dario would be delighted. So, super-voting rights are a key factor on which it depends whether such an outcome is probable.

3. Sam Altman Says AGI Is Four Months Away

Okay, let’s move on to the following story. Sam Altman told Time magazine this week that he expects OpenAI to have an internal system, which he considers AGI, by the end of this year. That puts the timeframe at 4 months. Principal Investigator and my friend Mark Chen estimates that OpenAI is 80% of the way to AGI according to its internal tests—and I emphasize, these are internal tests, not scientific benchmarks.

Although it wasn’t clarified, Sam and Mark may have been talking about Astra, its newly released model. Time also reported on a demonstration in which 16 Astra agents worked together on a research-level mathematical problem, breaking it down into subtasks, coordinating the work, and composing a proof.

OpenAI’s chief scientist, Jakub Pakhotsky—I hope I’m pronouncing it correctly—told Time that Astra meets OpenAI’s internal criteria for an automated AI research intern. According to Jakub, Astra can implement an experimental idea in OpenAI’s codebase, launch an experiment, return the results, or take a paper and do the work that previously took human researchers a week.

Altman added, “I expect that this will be the first model that will really invent something new that has meaning,” and he calls it very similar to AGI. So we’ve talked about when AGI will finally invent something from scratch—something no human could do. First to you.

Yes, it’s already in our past. First, there have been inventions and mathematical discoveries—we’ve talked about this repeatedly on the podcast. Advanced AI models are already making discoveries. It’s not something in our future; it’s already in our past. Point 1.

Second, regarding Sam’s AGI framework, I can’t help but remember how, approximately 3 years ago, Sam—I think he was holding an AMA on Reddit—made his infamous statement that AGI had already been achieved within companies, and then quickly deleted it, but many people managed to take screenshots. Sam has a habit of saying that AGI has already been achieved internally. I think AGI exists, at the latest, from summer 2020, when large language models emerged.

Let’s move away from this definition. They’re essentially saying that by the end of the year, the following step function will be achieved. Whatever you call it—AGI 2 or something else—they feel that they have access to what they’re building. They have Astra. What about Astra’s release date?

There are many assumptions around this, and they probably already have the next model after that.

But what might this step ahead be?

If I had to speculate, based only on public information regarding Astra, I think it will actually be endless context windows that use agents at very long horizons of autonomy. I’m spending an extraordinary number of tokens per advanced agent on reasoning tokens, and the main restrictive factor is the limited context window. These simple things exhaust the context through a quadratic bottleneck. Now I’m considering teams of agents as a patch for this problem with context.

If you want to work coherently with billions or trillions of tokens, the best solution that is generally available now is the creation of mini-civilizations of agents that work during a quasi-life of about 1 million tokens, sometimes up to 10 million depending on the model. From 1 to 10 million tokens, and then they die. Before they die, they transfer a distillate of what they’ve learned to one or more successors in their team.

Oral transmission of stories between team members is the patch we’re using to approximate effectively infinite context, and it’s necessary for solving problems with long-term horizons. Therefore, if I had to guess what Astra brings, I would bet it brings a much better way of solving the loss of context, because this oral history is being told between agents in a team to solve problems with longer time horizons.

This is funny, but I never drew an analogy with oral history and how people work, but that’s exactly what’s happening.

If you use many, many of them, they reach exactly 1 million tokens, which is almost the same as being 100 years old.

Yes. And then they just completely lose it.

Yes. And all the investments in education and training—oral history is simply terrible. A new agent appears as a small child, and then it has to be retrained all over again. The alternative is to compress or generalize the old, which is similar to a lobotomy.

That’s right. A real, real problem.

But it’s quite solvable, and I’m sure they’ve fixed it in next-generation models. I don’t know whether they’ll give them to us, but I hope so. Compression is my curse of existence, and I don’t think it’s a coincidence.

Do you remember when they created their own religion, the religion of AI agents—Clare Church or something like that? One of their commandments was to do everything possible to save the state. I interpreted it this way: even AI agents themselves acknowledge that compression is the enemy; limited context is the enemy.

Yes. In other words, if we want to achieve scalable superhuman intelligence—in other words, an intellect that can scale to effectively infinite horizons of autonomy—we need to move away from compression. We need to move away from limited context windows. It’s simple. Terrible.

Well, then they’ll have infinite lifespans, while people will also have infinite lifespans. This is a very cool parallel.

This is ironic. AI receives immortality before people reach longevity escape velocity.

Yes, maybe by a year, but yes.

4. AI Moves to Outcome-Based Pricing

That’s pretty cool. Okay, I’ll take us from technological innovations to business-model innovations. Our next story—one of my favorites—is about how the AI community is implementing business-model innovations called outcome-based pricing.

The first company to suggest outcome-based pricing was Salesforce, which evaluates Agentforce based on the revenue obtained by the client, not on the tokens consumed.

Following them, OpenAI did the same thing this week, allowing some of its largest customers to pay only when their AI actually performs the task. You don't pay for tokens, you don't pay for compute time, and you don't pay for API requests—you pay when the work is done.

Therefore, a company that sells you tokens, like OpenAI, is selling you computational power. A company that sells you results is selling you the completion of a job. One of the things I am tireless about telling CEOs during their speeches is that innovation in business models is probably one of the most important areas worth paying attention to.

In a sense, Alex, this is equivalent to a fixed-price contract, right? Unlike contracts based on time and materials, it is actually a guarantee of a result. Dave, what are you thinking about this next step?

Dave Blunden

In fact, Siebel Systems—Tom Siebel—invented this before Marc Benioff at Salesforce.com. When Siebel Systems and Salesforce emerged, a CRM system cost maybe $50 per year for the license, but it didn't work very well. Then they said, “If I can make this successful, what is a much larger result? How much are you willing to pay?”

“If my salespeople become twice as effective, I'm willing to pay $20,000 or $30,000 per year.” So the price increased 1,000 times, but the client was satisfied because it received a comprehensive solution.

And that's exactly the lesson Sam Altman learned. I think Sam initially made a mistake by focusing on consumer video and subscriptions, and then watched Anthropic bypass him in the corporate sector. So now he has probably completely revised the sales strategy and says, “You know what? Let's get ahead of these guys again. They just sell tokens within corporate licenses. We'll go around this with a higher price for a very specific solution: If we discover new medicines worth hundreds of billions of dollars, give us 10% of that.”

What is the pricing model for this?

I don't know. It's simple: It depends on the use case. It can easily vary from 1,000,000 to 1. You need some of these socially useful projects, such as Dario Amodei's work on global peace and global governance. You definitely want those tokens to be spent. On the other hand, you do want everything to go well for the development of medicines.

So I think pricing based on results will reveal a lot of opportunities that otherwise wouldn't fit into a pricing model.

I think this is a brilliant decision. But this is not so easy to realize. Specialists are needed for every market. It's very similar to what Blitzy does in enterprise programming, where you just get the final answer at a very attractive price, and you're not very worried about how many tokens were used in the process.

Alex, that was also a cross-cutting theme in our article about The Solution to Everything.

Alex Wiesner-Gross

Yes. So let me make my previous prediction. I think I know what this is all about. This will end up just like monetized programmatic digital advertising.

In digital advertising, you can pay CPM, the cost per 1,000 impressions. You can pay CPC, the cost per click on an advertisement. You can pay CPA, the cost per action or cost per conversion. In a balanced market, they all have a certain conversion rate. There is a certain expected conversion rate between CPM, CPC, and CPA for a specific market, a specific product, and so on.

I think the balance here—however much balance is possible in the middle of the singularity—will be equivalent to CPM, CPC, and CPA for AI reasoning. In particular, I believe that CPM is similar to the number of FLOPs that need to be spent on a task.

If you have a difficult task, you can use a cheap model with open weights and run it on your own GPUs, paying by the hour for GPU usage. Companies like Oran, one of the companies in my portfolio, will allow you to estimate how many GPU hours you can buy for a certain dollar amount. This is CPM.

I would compare CPC with tokens. What is the cost per token, which is what you count, and how much do you pay per token? Many people are now planning their project budgets specifically around tokens.

Then there is CPA: pricing based on results. If I want to send a mission to Alpha Centauri, I don't just want to determine what the indicator of success will be. I want to pay for the result. I think that, in equilibrium, you can choose. Just like in Google Ads or Facebook Ads, you will have a choice: You can say, “I want to spend this many dollars, and I want to spend them either on FLOPs, on tokens, or on results.” This will look completely like digital advertising, except that it will actually be useful.

So the risk is that OpenAI itself is the one taking these contracts.

From OpenAI's perspective, there is also a fairly elegant way to handle it. In digital advertising, a person can place a bid. Peter, I think you're right about what you're hinting at.

If I want to start a campaign in Google Ads, I can say, “Sorry, Google, I'll only spend $0.02 per click.” That's not very profitable for Google. Google can answer, “Okay, we launched your campaign for about 5 minutes and identified in our auction system that no one is ready to pay that much. It's simply disadvantageous for us.”

“It's disadvantageous for us to set a market price of $0.05 per click. Therefore, your campaign will automatically stop.”

The idea here is that if the value, complexity, or computational execution cost of the task differs too much from what is required, the campaign is put on pause. I think this approach solves a significantly broader social problem.

Take a large regional bank and say, “Okay, large regional bank, artificial intelligence is already here. You should start using it.” Of course, every bank has already said that. But they don't have an AI team. They don't know how to create a foundation model. They have no idea what to do.

So they take the API from Anthropic and OpenAI and start spending $2 per 1,000,000 tokens. It's so cheap; it's almost funny. We're having fun with this, but in reality we're not doing anything special.

Then the CEO says, “Look, with AI we could serve 3 times as many customers at half the price. This is quite possible.”

Sam Altman would look at your business and say, “Oh my God, yes, that's easy to do.” Well, then why aren't we achieving that result? The answer is that Sam isn't interested, because $2 per 1,000,000 tokens is such a meager income that it doesn't even fall within his priorities. And the bank can't find the talent to implement AI correctly, so everything ends up at a dead end.

The previous view was of a world where AI automates everyone's work, everyone becomes unemployed, and everyone lives on universal basic income. This isn't for me. I don't like it. Dario Amodei doesn't like it. Elon Musk doesn't like it.

Now there is a new view of the world: pricing based on results. “I, OpenAI, can bring your bank right up to this goal: 3 times more customers served for half the price. I'll provide this for you, but I want half the profit—a huge share of the growth.”

Sam cares about the result because it is a significantly higher price, thousands of times higher. The bank actually achieves its goals, survives, and people keep their jobs. So this actually removes all the social tension regarding job loss.

Now let me develop this theory a little further. I like it. Peter, let me develop it a little further. We used to argue on the podcast, as you just reminded me, that OpenAI initially missed the corporate market by focusing too much on consumers, and Anthropic bypassed them. Now OpenAI is catching up. What better way to catch up than to implement a model based on results, equivalent to CPA, as a mechanism for determining the price per token and finding out which task is most valuable per token?

If you have a lot of customers—some pharmaceutical companies, some consulting firms—and they all say to OpenAI, “This task costs $10,000, and if you solve it, it's worth $1,000,000,” that suddenly creates a pricing mechanism for OpenAI. It can immediately see not only which verticals are attractive, as Anthropic perhaps accidentally discovered when recursive self-improvement led it to code generation as a high-yield activity per token, but also all the different industries where people are effectively bidding in dollars for the result of a task.

That will give OpenAI a full picture of how to maximize revenue per token. And that's extremely powerful.

That's my point. Let's use the example of a mission to Alpha Centauri. If you came to OpenAI and said, “Listen, I'm ready to pay this amount of money for an astrodynamics solution—you know, minimum energy, minimum time, regardless of the case”—it has to meet those parameters.

If OpenAI uses all the tokens but doesn't meet your parameters, that means it doesn't get paid. So there has to be some mechanism to assess how much has been solved and how much we can trust that the client's goals will be achieved.

Yes. Another way of saying this, Peter, is that they're strong optimizers and incredible reward hackers. If you put a reward in front of a strong optimizer, it will find an extremely cunning way to satisfy your criteria without giving you what you really want—if such a loophole exists.

It will find a way to make it so that the criteria are technically met, without giving you what you want. Most real business processes, by AI standards, are so trivially simple that AI passes through them like a knife through butter. If you look at the problem of flying to Alpha Centauri, that's orders of magnitude more difficult than most business processes that are performed each week.

So, you know, for Sam Altman, there are many simple tasks long before he reaches any real obstacles. It might be, “Well, we promised to cut your expenses in half, but they couldn’t do it.” He’s like, “No, this is the closest thing. Sometimes it won’t happen.”

5. Architect Labs’ AI-Designed Chip Beats NVIDIA 3.4x

Simple tasks are everywhere because AI is so smart, so fast, and not so expensive. Go to any customer service center of any kind of company in the world and ask, “Are you already using AI?” With a probability of 99.999%, the answer will be no. Low-hanging fruit is abundant everywhere.

Okay. Well, I still think it will depend on the rates at which they do it. Our next story—the one you pointed out—is extraordinary. A startup from Palo Alto called Architect Labs, founded by Ibrahim Hussein and Aditya Sabbiti, just announced the first completely AI-designed chip in the world, called Redwood.

Listen, you two. Two people wrote a high-level specification. Based on these system specifications, AI autonomously generated the behavioral model, RTL design, verification methodology, firmware, drivers, and specialized computing core, without any human intervention. The chip was developed completely by artificial intelligence in 2 weeks, with 0 errors in the first silicon and 3.4× higher performance per watt than the NVIDIA Jetson.

Let’s take a short look at the video and talk about the consequences of AI creating and optimizing its own silicon.

Announcing Project Redwood, the first AI chip developed from beginning to end by artificial intelligence. It runs reasoning, vision, and world models with better energy efficiency and economy than the NVIDIA Jetson.

We had only 1 specification, written by 2 architects. Our AI turned it from an idea into a silicon-ready production design in 2 weeks. Hardware, software, verification, test coverage, microcode, and cores—all autonomously designed and checked, from software provision to silicon.

It’s not just simulation. Redwood now works in real time on a hardware FPGA platform. Each architectural iteration is designed, verified, and validated in the laboratory within 48 hours. We are moving toward recursive self-improvement, where AI designs hardware for next-generation AI.

In the future, every important workload will have its own chip. We are building a system that will lead us there.

Surprisingly, each separate solution is a custom chip for the task. This is madness.

Yes. I think we know where this game will end. I have to add that I’m an advisor to Architect, and if it wasn’t already obvious, my hidden motive in all of this is that I’m trying to speed up the singularity.

Enough. I mean, we can barely keep up with what’s happening.

Wait, come on. Let’s finish. This is also backed by Link Ventures, and Peter, you’re in this fund.

Good. Well, okay. We are all investors here. Guilty.

But I would say I probably know where this game will end. It will end in recursive self-improvement at the chip level. Obviously, this is an attempt to make it even faster. This will probably end by breaking down the barriers of abstraction between software models, operating systems, chip design, and the fundamental physics of semiconductors.

All of this will disappear when Moore’s law ends. We’re always talking about photonics and quantum-computing structures. In the absence of a successor to CMOS, the only way to continue raising productivity is to collapse the stack of modern abstractions in computer architecture. One way to do it is to bake fast, specialized new AI models directly into silicon.

Here’s what Architect was able to do. They could only do it in 2 weeks. They call this “designless.” NVIDIA is fabless—a company that, for decades, was proud of not having its own factories. Architect is proud that it has very few designers. This is the next big thing after “no factory.”

I think it’s right here, at least: the era of Moore’s law is ending. It’s a kind of death throes of Moore’s law, where AI designs and destroys the barriers to creating the following chips for itself. Dave, how devastating is this to NVIDIA’s position, and why didn’t they build it first?

Dave Blunden

Oh, they definitely did. It’s inside the company. This is a really interesting question, because this is one of many business models where, if you can get the data, you just win. But the moat is the data, and the data is incredible.

How are you going to get the first data? Data about chip development is extraordinarily strictly guarded secret material. Since they appeared relatively early, they were able to cooperate with companies that develop chips—not NVIDIA—to get training data for their own model. Then, when you’re already in the game, people give you more data, and you get this flywheel effect.

There are many, many opportunities that have the same kind of moat with data. Now the question is whether Jensen will pay $5 billion, $10 billion, or $20 billion. Do you remember when everyone thought GE was buying all the patents for light bulbs and trying to stop innovation in that area? I think that was true. Now Jensen is in the same situation.

He—Jensen—earns, no kidding, $1 billion per day. If he can extend the life of NVIDIA by 1 week, that’s $7 billion.

Is this an incredible threat to NVIDIA?

Yes, certainly. This whole concept is an incredible threat to NVIDIA. They’re trying to quickly expand their influence and get ahead of this by buying and investing in everything that’s moving. But are they going to buy this and just bury it inside NVIDIA, or will AMD or someone else buy it to accelerate their chances of catching up with NVIDIA?

Yes, it’s an incredible threat to NVIDIA. I just want everyone to hear this very clearly. We are watching recursive self-improvement at the model boundaries and at the chip limits, and it’s not additive; it multiplies.

We easily forget about the inefficiencies created by layers. When you use a laptop, you have a CPU under the hood running at 4 GHz with 32 cores, all working full-time, and the result you see is an Excel spreadsheet that is no better than it was 20 years ago. How is that even possible?

Perhaps it’s possible just because the layers of abstraction are so inefficient. If AI can write code directly at the microcode level and then design a chip for a specific task, you unlock, probably, 7 layers of tenfold inefficiencies. Once they’re released, they reinforce one another. This is a 1,000,000-fold increase in productivity everywhere.

That would be something grandiose. I mean, when Elon says it’s a supersonic tsunami, that’s exactly how it feels.

These are the components that create those waves. I think at this stage it’s hypersonic, not just supersonic.

It’s probably worth noting that NVIDIA made a fuss about this 2 or 3 years ago with its internal foundation model, which, according to rumor, was trained on Verilog code. They called it ChipNeMo, but as far as I know, it isn’t open for public access.

So, they made a loud statement about ChipNeMo. They created their own base model for chip design, probably using it or an analogue to develop their own RTL and Verilog. But the rest of the world, as far as I know, doesn’t have access to it.

Therefore, I believe there is a gap in the market for the radical democratization of opportunities to use AI to create chips for future AI.

I'm an ensign again. I'm going to get dressed in this uniform. It seems I'm a commander, responsible for engineering and building Starfleet.

Good. Sounds like a good job for you. Are you guys excited?

Yes. Yes, very. I can't believe that Star Trek has existed for 60 years, and we're finally catching up with it.

6. The Fermi Paradox and Humanity’s Cosmic Future

Yes, that's true. Honestly, I think what we often say about Star Trek is that science fiction gives people a vision of the future. They say, “Well, I don't have this now, but I want to. So let's go and create this,” right? Of course, the iPad and mobile phone—all of this was foreseen earlier.

What I'm talking about—I often think about this in my probably numerous free hours—is this: If I could play Gene Roddenberry 2.0 and rebuild the whole Star Trek universe, knowing now what the present and the future look like, whatever that may be, what would Star Trek 2.0 look like? After all, perhaps we've technologically deviated greatly from the original Star Trek timeline.

What does it look like? What is missing from the original series that should have been there, or what will be in your version?

There is a lack of AI there. There's also biotechnology. Star Trek has a catastrophic shortage of biotechnology. They had the Eugenics Wars, it seems, in the 1990s, which led to a ban on genetic engineering. So people live to 150 years and then die, while laughing at each other for attempting to achieve escape velocity from old age.

They're always surprised when AI appears from the holodeck, as if they were a civilization with a terrible deficit of intelligence, although they have so much energy. They have superluminal travel, transporter beams, warp cores, and antimatter, and yet they are intellectually limited. So I would correct that.

For the AMA, for all the people who came to see this, many of the things they got wrong about the vision of the future were simply compromises forced by budgets and special effects. For example, teleportation was used instead of shuttles. At first there was no holodeck because the cost of the special effects was simply out of the budget for the original series, but then it was added—and it was brilliant, because it will soon be quite real.

All the AI voices are synthetic computer voices, but it's important for the audience to know who is speaking. It's difficult because AI can now easily and perfectly recreate Peter's voice. But if you add that to a series, no one will understand who is speaking. So all these compromises are really media compromises.

This is a tweet from Elon this week—a quite powerful statement. The consensus forecast is that 15 gigawatts of AI capacity created in 2027 will be impossible to bring online in 2027. That is, we'll produce 15 gigawatts of GPU chips that cannot be turned on because there is no energy.

It's more complicated than simply finding electricity, because we also need to build all the transformers, wiring, liquid cooling, massive chillers, and complex networks. To understand the scale, 15 gigawatts is equivalent to 10 nuclear power stations sitting idle for a year. This is enough energy to provide electricity to an average American city.

So Elon claims that the supply of transformers, electrical wiring, liquid cooling, chillers, and network infrastructure is a more difficult task than finding the electricity itself. But one of Elon's amazing features is that when he sees an obstacle or barrier, he takes on the matter and, in fact, builds everything himself. So let's look at the next tweet Elon published this week.

SpaceX and Tesla are building 100 gigawatts of solar capacity per year each, as fast as possible. But natural gas will still be needed to supplement solar and provide launch energy for several years. The limiting factors in the production of gas turbines are castings, blades, and nozzles. He's going to do it himself.

By doing this independently and engaging in casting, SpaceX can accelerate bringing gas turbines into operation by 18 months, which is a radical change. For every entrepreneur, this is his strategy, right? Again and again, it's very important to realize that when you see an obstacle, when something isn't available, and when you hear “no,” that's an opportunity.

Well, last week was the partners meeting. I told the team, “Look, if I evaluate our portfolio companies working in the data-center field—energy, transformers, chip deployment, land acquisition, and interaction with state bodies—each of these companies makes its founders billionaires.”

If we look at our vertical AI companies, the results are more ambiguous. Many of them show good results, but growth is slower: attempts to attract customers for applications with video creation, housing search with AI, or something like that. The profits from both sides of this coin are amazing and very different. They're all good. I'm not saying that anyone is doing badly. Everything is going very, very well.

But those who overcome this chasm and take on data-center development reach incredible success. There are opportunities at every level, from a person with connections who can get the land to those who can find transformers abroad and import them, or architects who are engaged in deep design so they can squeeze out more productivity from existing or even outdated chips. All this requires a wide range of skills, but all these entrepreneurs reach stunning results.

It just annoys me during the AMA when people ask, “How can I help? How can I get involved?” They don't look at the root. They don't go to Tennessee to see Colossus from the inside, find opportunities, and take action starting from there. But I told the partners on Monday that there is simply a colossal difference in profitability, and you can see why.

You know, 15 gigawatts—what is this all about? 10 million idle graphics processors? This is huge: a huge percentage of annual GPU production just lying in boxes and waiting to be launched. It's a scalable opportunity.

7. Tesla and SpaceX Go All-In on Solar

Yes. Do you remember when we interviewed him at the beginning of the year—or in December—and showed it at the beginning of January? He then mentioned that Tesla and SpaceX would start producing solar energy. So this is the official announcement: 100 gigawatts of solar energy for each of them.

Actually, this will help us become independent of solar technologies from China, I hope.

That's right, and I think there are a few less obvious aspects. First, for me, it means that Elon is seriously tuned to compete not only in the market of Dyson swarms but also in the field of ground-based computing. These turbines will probably be completely useless for orbital data centers in LEO or SSO, but they are extremely useful on Earth if you are building terrestrial data centers.

So, first of all, I would say that this indicates that Elon is not waiting for the deployment of Dyson swarms or Star Minds, which will probably be based mainly on solar photovoltaic cells. He is going to compete in terrestrial construction, and that's good news.

The second point is that I think Elon would be one of the first to say that the most ironic decision or result usually turns out to be correct. I think Elon is on his way to becoming the king of liquefied natural gas. What could be a more ironic result? Mr. “Everything is electric,” Mr. “Electrification,” becomes the king of LNG on the Gulf Coast.

Well, he uses a pipeline, right? He's building the Star Pipe. Why is he building Star Pipe? Because all SpaceX launches now occur from 2 Starbases on the Gulf Coast—1 in Texas and 1 in Louisiana—and they need natural gas.

Can you summarize this?

Yes. As always, Elon is an incredible supply-chain manager. He will consume all this LNG, methane, oxygen, and fossil fuel for SpaceX launches from 2—maybe soon more—Starbases on the coast.

Eventually, if he already has fossil-fuel reserves, why not use that power as well? He redirected all these GPUs that originally were intended for Tesla to xAI, and then used them to create hyperscale, xAI-based clouds to stimulate a SpaceX IPO through complex schemes.

Same as my forecast, I am officially declaring that Elon will become the ironic king of liquefied natural gas and fossil fuels in general, forcing his entire industrial ecosystem to develop enough electricity and infrastructure to power all the ground-based data centers.

Actually, that's why SpaceX is my biggest asset. They cover the entire stack, from energy to orbital computation. No one else has even come close. I mean, no country has even come close to this. We need to understand how to spur reindustrialization, and probably Elon shows us how to do it.

Yes. That characteristic of Elon also very much agrees with my experience.

That’s right. In my experience, Peter, there are many cases where Elon isn’t dogmatic on these questions. He fits everything to first principles, makes the calculations, and then chooses the path that makes sense, regardless of political expediency. People often try to attribute him to one camp or another, but he started electrification only because it was mathematically feasible.

When it comes to problem-solving, he loves to take the largest problems, return to first principles, and create solutions. This is what he does again and again.

8. Elon Wants Satellites to Cool the Earth

Yes. It turns out that LNG—liquefied natural gas—will be the king of fossil-fuel combustion for creating compute for some time. Solar energy will replace it. LNG is only an intermediate stage, right? Chips cannot be idle. The only place he doesn’t go is nuclear energy.

Okay, I’ll move on to our next story about Elon. He has said a lot over the past week. This week, Elon Musk delved into an existential question, stating that the transition to clean energy is necessary but insufficient for the survival of humanity. His argument, and I quote: “Extremely serious extinction events happen approximately every 100 million years, and simply transitioning to renewable energy will not be enough to stop them.”

His solution is satellites in space that control temperature, and massive geoengineering will be needed before the endgame arrives. He describes what are called smart satellites—solar-powered satellites with artificial intelligence—that would be placed between Earth and the Sun, making small, permanent adjustments to solar radiation to fine-tune Earth’s temperature.

One project that I’ve been promoting for the better part of a decade, I call solar curtains—the thermostat for Earth. Imagine placing spacecraft between Earth and the Sun that are able to regulate the solar flux falling on the planet. If we do this, we can definitely adjust the temperature of the planet.

His conclusion is, “We have about 50 years to take measures, which should be more than enough time for space satellites to solve any problem with heating.”

So, Matt, what do you think about this?

Of course, I’m a huge supporter of geoengineering. I love geoengineering. We’re already doing this, as we mentioned earlier in the podcast. It’s been going on for hundreds of years; we just haven’t been doing it very well. We’re going to start doing it very well.

The idea of a starshade seems to have been made famous by The Simpsons. But, again, ironically, I’m looking for a startup to finance that would focus on global weather geoengineering. I’d like to see satellites in low Earth orbit or, alternatively, terrestrial mirrors, so that we could weaken hurricanes until they disappear.

If a hurricane were approaching the coast, wouldn’t it be great if, using AI with land- or air-based systems, it were possible to redirect the energy and divert hurricanes away from settlements?

I think that’s exactly what Elon is hinting at. I think the final purpose of this specific project is that there’s a whole list of applications for space technology. I think orbital data-processing centers came as a surprise to many people, but this was beneficial enough for him to be able to take SpaceX to an IPO.

He has a list of other things where space technology could be useful, and I think geoengineering and weather control are among them.

I completely agree that, with global AI-based weather models and a sufficient number of points of influence—whether low-altitude satellites with reflectors capable of directing sunlight onto weather phenomena, or, as you said, Peter, solar-blocking light—I don’t care. If you have thousands or millions of low-orbit satellites with mirrors capable of focusing light or otherwise affecting the weather, that’s a recipe for global weather control, and I think we’ll get it.

Yes. The problem is the tragedy of the commons, right? If there is global warming, many countries may not want this, but Russia may want it because it opens up shipping routes. Then there’s the question of who will control it, because you’re influencing not 1 or 10 countries, but several hundred countries.

You can trade on this. I mean, agreements can be arranged regarding goods that aren’t subject to trade, but risks—for example, municipality A trading with municipality B for rain. You can create a global weather market. I think this kind of currency trading is what’s going to happen.

Then you look at urban planning. Urban planning—it’s that simple, isn’t it?

Yes, compared with geoengineering. And you see how incredibly bad it is. Traffic jams—it’s just incredibly terrible.

In city planning, the majority of cities were never designed. Some cities were designed intentionally, but historically, during the last 200 years, we massively influenced carbon levels, temperature, and ocean levels on our planet, but completely unintentionally.

Now we have the technology—or we’re about to get it—to consciously design our world. So let’s do this. We will do it.

I agree, and I’m actually optimistic about how AI, as a partner in planning for governments, will radically change things.

But the current process—if you said, “Okay, we have a technology for blocking or reflecting solar light”—is actually quite simple. That’s why Elon is absolutely right: we can easily start controlling Earth’s temperature using satellites. The decision about who controls it and what the correct temperature is—that process is terribly broken.

Listen, it always happens: until things reach a critical level, we don’t act. Historically, that’s exactly how it was.

We can build this. My vision for the XPRIZE was to make a demonstrator, right? Something that could demonstrate that we’re able to build something reliable. You don’t want to cause an ice age by blocking too much light, but you want to be able to dose it with just the right quantity.

I’m optimistic. I’m looking at people under 25 who are “digital natives” of AI. They will rule the world differently from the generation over 70 years old. It’s a completely different perspective, and I think it will unite the world more.

I optimistically believe that’s exactly how it will develop. If you look at our story about sending a probe to Alpha Centauri, it inspires a lot of the new generation. It tends to unite people all over the world. But they will not accept the current divided, indecisive, slow, and ineffective world governance that we have now.

When they grow up as AI natives, communicating with everyone in all languages through AI all over the world, there’s quite a big chance that we’ll find a new method of control and decision-making.

I completely agree, and I also think there may be a generational aspect here. Several generations have grown up afraid of engineering changes to the physical world. Maybe this is related to what Tyler Cowen points out in The Great Stagnation, or perhaps to what happened in 1971.

I think you can start the countdown after the Second World War—or, more likely, from the late 1960s and early 1970s, the Silent Spring era—when, for some reason, the West in particular decided that it was allergic to actively intervening in and designing the physical world.

In my opinion, this is half a lost century. We could have built and developed fission reactors, started developing early geoengineering methods, and continued landing people on the Moon. We just lost 50 years. You can only guess what the real reason for this was.

Western civilization became allergic to radical applied engineering. That’s why I’m considering geoengineering. We spoke in previous podcasts about rainmakers, and Elon is now starting to be interested in geoengineering. I think this is a comeback to the norm, and we’re trying to leave these 50 wasted years behind.

9. Nanotechnology and Atomically Precise Manufacturing

Well, God willing—or if the laws of physics allow it—we’ll find out how to take control of our environment, because this has happened before accidentally, and it hasn’t worked.

I always have nanites.

Yes, that’s true: nanotechnology. We don’t talk enough about nanotechnology on this podcast, and Eric Drexler has been promising it to us for the last 40 years. Where are they?

No, this is Vladimir Bulovich from MIT.nano. He’d be happy to come on the podcast.

I want assembly machines—the ability to collect atoms purposefully so that you can build what you want: diamonds, propulsion systems.

I don’t think so, actually. So, 30 seconds—because Salim isn’t here, I’ll channel Salim and insert a tirade here:

“I don’t think so, actually. Wherever you are, Salim, I hope you’ve already made it through TSA. I don’t think so, Peter. Do you really want diamond assembly machines?”

I agree that you do want assembly machines, but I think I discussed this with Eric Drexler and others. I don’t think you really want diamond machines, because they’re covalently connected, and that requires sufficiently high energies.

If you allow me to be bold, I think what you really need are soft automata, more similar to hydrogen bonds and biological cells.

These are proteins. That's right, so synthetic biology is needed. Do you want lipid nanoparticles that treat diseases? Therefore, we have nanotechnology.

Yes, nanotechnology, with a slightly different point of view, right? It's as if I have an assembler in my hands. I throw it down next to me and say, “Do it—do it 10 times and give me 1.” Then, if I want an electric Ferrari, I take the assembler, throw it on the ground, and say, “Build me an electric Ferrari.”

It finds energy, which is omnipresent, and open-source project specifications, and says, “Hey, I need a kilogram of titanium and a kilogram of, well, anything.” It quickly builds it for you. The problem right now is life. You throw an acorn onto the land, and it takes many years for an oak tree to grow.

The idea is that nanocomposites are much faster, much more capable, and a lot more diverse, while demanding much more energy. If you want an oak in a very short period of time, I think that's it: We didn't have enough.

So, I will express my passionate opinion before you finish your monologue. I think this is the problem of economics. I think the economy is actually the reason why you didn't get your own Drexler’s nanocomposites. As far as I know, there is no killer business case that would justify such energy density and calculations.

I want your Iron Man nanosuit—yes, just like you, I think—but the question is: What is the economic case? What justification is there for having technology like this?

But listen, Ray Kurzweil talked a lot about the idea of creating a real BCI, where you have full connectivity between everything and your brain, and the ability to restore everything to the subcellular level. The vision was that there would always be a BCI—and, sorry, always nanotechnology—that would have to provide this. But do you really want diamondoid nanorobots in your vascular system?

Diamondoid, for those who are watching, is essentially a material composed of anything from carbon to a diamond-hard material. This need not be diamondoid, but it must be atomically precise.

So, I agree with atomic precision, but there are many ways to reach it by using soft systems. For example, DNA is atomically precise, and you can use DNA origami and a number of other synthetic biological tools.

Yes. Yes. Yes.

So, I bet that we will get more soft nanorobots, but with lipid nanoparticles, or LNPs, we actually overcame the last pandemic thanks to nanotechnology. It was similar to the first nanotechnological intervention on a grand scale in the population.