SpaceX 750亿美元以上的历史性IPO、GPT5.5击败Polymarket，AI破解80年数学难题｜EP #257

SpaceX filed for what’s expected to be the largest IPO ever: $75 billion being raised at a valuation probably north of $1.75 trillion. Elon is about to have a currency to go on a shopping spree. That sentence is so out of band with any point in human history.

So, anyone out there doing a startup, there is capacity for 1,000 unicorn transactions. Remember, unicorn is called unicorn because it’s supposed to be extremely rare.

GPT-5.5 is now beating prediction markets at forecasting the future. It beat Polymarket’s crowd predictions for the Super Bowl. This is the worst these models will ever be.

The concentration-of-wealth effects from this would be insane. The financial singularity.

Alex Wissner-Gross

This is a problem posed by the famous Hungarian mathematician Paul Erdős about 80 years ago. Now AI has just solved it. Not only was it faster, not only was it able to brute-force, but it was also smarter.

This is a much bigger moment in history than just solving a math problem.

Suneel Gupta

Solve everything, Peter. We’re seeing it.

Now that’s a moonshot, ladies and gentlemen.

Welcome everybody to another episode of Moonshot. I’m here with my extraordinary moonshot mates, Salim, the father of organizational singularities. You’re spawning singularities everywhere, buddy. Alex Wissner-Gross, our in-house polymath. Dave Blundin, our wizard of AI investing. I’m Peter Diamandis, your host, and hopefully your abundance whisperer for an optimistic future.

You know, we’ve loaded the show today with extraordinary stories—hopefully, stories that get you excited about being a builder, literally starting to build before this episode is out. Gentlemen, good to see you all. I have to ask our normal “Where’s Waldo?” question. Dave, where are you, buddy?

Dave Blumberg

Back at Stanford. I had a whole bunch of really fun meetings with AI founders here.

Awesome. Suneel, I know you’re not home, but you’re never home. Where are you?

Suneel Gupta

No, I’m in Brazil, like we were yesterday. I’m still here until tomorrow, and then I fly back.

Okay. And Alex, you and I are in our normal haunts.

Alex Wissner-Gross

That’s right.

Amazing. Let’s begin. Here’s a quick look at what we’re going to be covering. SpaceX just filed for the biggest IPO in human history. It’s extraordinary. Later today, after we record this episode, we’ve got Starship V3 scheduled to launch.

OpenAI just disproved a conjecture in mathematics, for real. We’ll be talking to Alex about what that all means. GPT-5.5 is now beating prediction markets at forecasting the future as well. ChatGPT just became your financial adviser.

A lot to cover. Let’s get started. Our first story here is, in fact, the SpaceX IPO. SpaceX filed for what’s expected to be the largest IPO ever: $75 billion being raised at a valuation probably north of $1.75 trillion. It’s the biggest in history—over 2.5 times that of Saudi Aramco.

Elon is maintaining his super-voting rights, with insiders controlling 86% of the voting power. I love this: SpaceX’s IPO prospectus says it expects an addressable market of $28.5 trillion. That’s quite a TAM.

Dave, let’s go to you first on this one.

The TAM is just under the size of the entire U.S. GDP. I’m sure they landed on it and thought, “We don’t want to claim to be bigger than the entire U.S., so we’ll be just one notch below.”

Eric Bolton hosted Mor and me for dinner last night at his house. Eric is a professor of economics and AI at Stanford HAI, the AI lab at Stanford, and he was the first guy to mention this to me. I was like, “Come on—$28.5 trillion TAM?” But he was going off on it.

There’s no way to disprove it. There’s no reason it shouldn’t be true. Remember, Elon’s core thesis is that we can 10x the global economy in 10 years. Most people think it’s possible, but that it will take longer than 10 years. Regardless, if the global economy 10x’s, then his TAM should easily fit within $25 trillion or $28.5 trillion.

I looked at the breakdown of how he got to $28.5 trillion, and it’s interesting. $870 billion—just a mere bit under $1 trillion—is Starlink’s business. $740 billion is Starlink’s mobile unit. $600 billion is their digital advertising market through X. $2.4 trillion is their AI infrastructure. And get this: $22.7 trillion comes from Macrohard.

If you remember, Macrohard is their partnership with Tesla, where they want to emulate all digital work and create an AI-run software company.

Pretty spiffy. A trillion here, a trillion there adds up to a lot of money. Alex, what do you think of it?

Alex Wissner-Gross

You almost have to ask: Did SpaceX acquire xAI, or did xAI acquire or reverse-acquire SpaceX? Certainly, based on the TAM analysis, it looks a little bit more like the latter rather than the former, with enterprise applications dominating the addressable market.

I think the most interesting part in the entire prospectus was actually the bit about what Anthropic is now paying SpaceX: $15 billion per year for data-center access. By the way, in the past 48 hours, it’s not just Colossus 1 that Anthropic is paying for. It’s Colossus 2 as well.

In the past couple of pods, I made remarks to the effect that Grok is on life support. A bunch of folks I saw in the comments said, “Oh no, he’s against Elon. How outrageous is this plan?” No, I think, far from being outrageous, SpaceX’s prospectus arguably supports that entire thesis.

SpaceX is basically taking not just Colossus 1—which, charitably, one could have argued is a relic of slightly older GPUs, and a heterogeneous set of GPUs at that—and giving it to Anthropic. Anthropic is now also using compute from Colossus 2.

To my eye, this looks like SpaceX basically abandoning the foundation-model space, handing it over to Anthropic, focusing on the infrastructure layer, and building out the Dyson swarm—

Record “Dyson swarm” into the episode.

Alex Wissner-Gross

—and focusing on becoming the infrastructure layer. I’m just reading the tea leaves here with the prospectus, but SpaceX seems relatively uninterested at this point in owning its own foundation model.

They’ve also announced that they’re going through with the purchase of Cursor 30 days after their planned IPO. Cursor is based on Kimi at this point. So, SpaceX’s foundation-model lineage is seemingly switching over to a derivative of Chinese open-weight models, with a lot of fine-tuning based probably on American reasoning traces.

But xAI, in this weird climate, wants to seemingly own the layer below and above. It wants to own the infrastructure layer—that’s the Dyson swarm and all the data centers—and it wants to own the layer above. That’s Macrohard.

Which other hyperscaler does that start to look like? Microsoft, with its strategy of enveloping OpenAI.

Are they still playing?

Alex Wissner-Gross

Meet the old boss, same as the new boss.

SpaceX is trying to transform itself, I think, into a Dyson-swarm version of Microsoft.

I was on CNBC this morning doing an episode about the SpaceX IPO. One of the things I was shouting from the hilltops is, “Listen, you can’t think about SpaceX just from its Starlink revenue or even its launch revenue.”

What SpaceX is doing is opening up the space frontier, and everything we hold of value—metals, minerals, energy, real estate—is available in near-infinite quantities in space. This isn’t just the first ships from Europe to the United States. It’s the galleons. It’s the railroad.

As SpaceX opens up this transportation infrastructure, it’s going to own the businesses all along the route. At the end, what we’re building on the Moon, in Earth orbit, in the inner solar system, and beyond—

Alex, I know we’ve talked about this before. Starship, compared to all the other launch vehicles—we can compare it to New Glenn, to what Relativity Space is building, or to what Rocket Lab is building with Neutron—it doesn’t come close.

Starship is planning to launch, on average, once an hour. Today, Falcon 9 is launching every 2.5 days. When they get to an hourly launch rate—airline operations—they open up the most extraordinary wealth in the universe.

Alex Wissner-Gross

The interesting thing to me is that they’re barely even selling it. Yes, it is in the prospectus that they intend to provide Earth-to-Earth transport for humans, cargo transport, and similar things. Obviously, they’re still a space-heavy-lift company, but if you just look at the prospectus, they look like Microsoft in space. That’s the story they’re selling to retail right now.

Because that’s what investors understand.

Alex Wissner-Gross

That’s legible to capital markets.

I think it would freak people out. I go down the same path you did, Peter. This is planetary infrastructure. It’s like Christopher Columbus sailing off to colonize a whole new world.

Just some confirmation for what Alex was saying, too: The Stanford Ph.D. I was just meeting with before this podcast is tracking all of the talent and confirmed that, yes, the great people have left xAI by the droves, but they’re all going to Anthropic.

Alex Wissner-Gross

Yeah.

And Karpathy has got to be the ultimate coup. We had that on the last pod. Shane Longpre from MIT, who’s phenomenal, is joining Anthropic now, too.

So, a duopoly between Anthropic and xAI is pretty daunting for everybody else. You think about how the best researchers all want to join Dario because they trust him and because the models are phenomenal. Then you put all that compute onto Elon’s space empire, and that’s a heck of a duopoly.

I don't know, though, if Elon plays duopoly forever. It doesn't seem like his—

Alex Wissner-Gross

Playbook. [laughter] Yeah, he doesn't play well with others.

It's fascinating. So, the other thing—the Terafab—where's the Terafab in the prospectus?

Alex Wissner-Gross

Maybe it's the same story. We don't need to promote things that are still hypothetical. $1.7 trillion is enough for now.

Maybe.

Alex Wissner-Gross

Yeah, there were a bunch of assets that were shared. Macrohard itself was, if I remember correctly, characterized as in part shared with Tesla and in part informed by Tesla AI. I also think Optimus—there's a curious split that I don't know how it's going to play out between the Optimus AI and the Macrohard/xAI/Cursor AI.

Alex Wissner-Gross

Right now, SpaceX, in this bizarre, sort of incestuous ecosystem of Elon-affiliated companies, is seeming to get, in some sense, the digital Optimus, if you will—the Macrohard worker for knowledge-work labor—and Tesla is seeming to get the embodied Optimus. It's not obvious to me exactly how, from a governance perspective, all of this IP is supposed to flow back and forth between them.

A lot of predictions, Alex, about merging, obviously, xAI and Tesla into Musk Corp. I wonder if there's a Polymarket on the over-under there.

Alex Wissner-Gross

There is a prediction market for it.

Yeah. Do you remember what it might be?

Alex Klive

Maybe you can check in background mode.

Yeah, I'll check. But honestly, I think within a year we're going to see that. When we've got 2 publicly traded companies, the ability to value them and merge them becomes a lot easier.

Alex Wissner-Gross

Well, bullet 2 is really important: the super-voting control. Remember, we mentioned on a podcast a while ago that Elon has never had what Sergey and Mark Zuckerberg have, which is a public entity that can raise many billions of dollars overnight where he's the controlling shareholder. So this will be the first time that he's in that position.

Remember, he doesn't control Tesla. That's why he's constantly going to Delaware court over his package.

He will.

Alex Klive

Yeah. This will be a new thing in the Elon-verse. And remember, Peter, you were an early investor in xAI. Remember that capital raise when it was getting off the ground?

Yeah, the very first one.

Alex Klive

Remember the amount that he was raising, which seemed like a lot at the time?

Oh, God. I don't know. I recall it being like an $8 billion valuation and maybe a $1 billion raise or something like that, which is just hilarious now because you're looking at a $100 billion valuation and a $35 billion raise.

Alex Klive

He's about to have a currency to go on a shopping spree.

Yeah.

Alex Wissner-Gross

And, Joe, my bet is that once they go public, they're going to begin acquiring a number of companies as part of it. One thing we posted the other day is that once these guys are public, they can easily do 1,000 acquisitions of $1 billion or more.

The world—like, that sentence is so out of band with any point in human history. So anyone out there doing a startup, maybe you'll screw it up, but assuming you don't screw it up, there's capacity for 1,000 unicorn transactions. Remember, a unicorn is called a unicorn because it's supposed to be extremely rare. [laughter]

All right, this is just a whole new world. The old days.

Alex Klive

And, by the way, I have the prediction market for you. Polymarket predicts a 20% probability of SpaceX and Tesla merging by the end of this year.

Okay. All right. By the end of this year, but I think within a year, I think that's going to happen. Things will settle out.

Alex Klive

Yeah, that's right.

Yeah. All right. I wish we were recording this later. I'm so excited. What's coming up today is the launch of the Block V3 Starship, with a 100-ton payload capability to orbit. It's an extraordinary vehicle. We're talking about a thrust of 18 million pounds using the Raptor 3 engine, the most beautiful, elegant engine ever. This is Flight 12.

On this flight, they're going to be demonstrating the docking ports that will enable orbital refueling. Of course, orbital refueling of Starship is required for the lunar missions they're planning to win, as well as for going to Mars. Both stages are going to be splashing down: the Super Heavy in the Gulf of Mexico and Starship in the Indian Ocean, as it's done before. Hopefully, they'll have buoys out there to watch both of them.

Remember the NASA Artemis mission? Artemis 3 is a docking test later this year, in 2027. Artemis 4 to the lunar surface, particularly the south pole, is taking place in 2028. Alex, you'll be watching, of course.

Alex Klive

I'll be watching. Not from the Moon. I would really love to see orbital refueling happen sooner rather than later. I don't know when that's expected—later this year? But I think being able to get to the point where we can not just do propulsive landing but also orbital docking and refueling between Starships is going to be such a key moment.

It's also the hammer that the Bezoses of the world have been using to argue that the Starship architecture—which is a little bit, if I were to play out an analogy, more internet-packet-oriented—is a questionable approach. The lunar architecture of Starship colonization of the Moon is very launch-intensive. It involves dozens of independent launches and refueling steps, whereas historically, if you look at the Apollo architecture, it was a much more monolithic architecture. You go up, then you go over, then you land, and then you come back.

You carry all the fuel from the ground for that one mission, all the way.

Alex Klive

Exactly. So I think the world, myself included, will be watching with quite a bit of optimism that, call it the Starship packet-type architecture—where you send everything up to LEO in packets, then do a bunch of refueling steps, and then send the packets over to the Moon or to Mars—is a superior solution.

I would view it almost as analogous to the switch—no pun intended—that the internet made, or I should say that networks made, from circuit-switched systems, where you had, in some sense, a single contiguous bit of atoms connecting a transmitter and a receiver on a network, to packet-based switching, where there was a complete decoupling of bits from atoms. The same idea applies here with a Starship-based solar-system transport architecture: we're aggressively decoupling cargo from transport. Fingers crossed that Starship V3 and orbital refueling later this year enable us to packet-switch the solar system.

You know, I don't think people realize how unique Starship is in terms of how it was designed, right? It's designed for full reusability. It's designed to land, refuel, and go again. His vision is airline-like operations.

Alex Wissner-Gross

And the amount of throwaway outstrips everything else. The one thing that Elon has done extraordinarily well is his manufacturing approach and his sort of building for an ultimate capability. New Glenn—and, again, Relativity Space's now-CEO, friend of the pod Eric Schmidt—those vehicles, I don't think they can compete. They're going to have to build new capabilities to get to this launch frequency that enables all of these visions.

Well, if data centers in space—you heard Sundar at Google I/O saying data centers in space—we're not going to talk about it today, but it's very clearly on the roadmap. If Elon and Anthropic are working in cahoots to build the Dyson swarm or Dyson sphere, then Google has to react to that with something. And here's Eric Schmidt, our former CEO, with a rocket company. So they have to play fast followers somehow. I mean, you're right; it's starting from pretty far behind.

Alex Wissner-Gross

But something has to give because Google obviously can't buy SpaceX. I also wouldn't necessarily overindex on the Starship architecture being the definitive final word on how we get to orbit. There are going to be many, many future technologies, I would predict, over the next 5 to 10 years, that will leapfrog, in terms of applied physics, Starship's launch capabilities. It won't necessarily be SpaceX that's the leader in the field for leapfrog capabilities. So I think there are many ways to orbit.

The development time takes years, right? We haven't seen a vehicle go from design to operational flight in anything. Even Falcon 9, to get it to reusability and to get it to that rate of success—99.99%, if you will—takes a good 5 years. In the interim, don't forget, Elon's going to be designing whatever follows Starship as well.

Alex Klive

That's true. But there are 2 things happening concurrently with that. First of all, there's the fast-follower effect. Once you can see what worked and what didn't work, it really helps you a lot in copying, and that varies by technology. But concurrently with that, you have Maccato[?] coming up with mechanical-design AI.

You remember when we were interviewing Elon, he said, "This is the greatest thing ever built by humanity without AI assistance. It'll be the last great thing." [laughter]

Alex Klive

But it was entirely built with people and protractors and crayons. I'm getting—

How retro. [laughter]

Alex Klive

It was CAD/CAM—hand-built, hand-designed CAD/CAM—without AI mechanical-design assistance. That'll never happen again. So that might really accelerate Eric Schmidt.

So I think that—

Alex Klive

—the design space of aerospace lift and heavy lift is vast, and we've only scratched the surface of it, I think.

And to Dave’s point, once it’s demonstrated how large the market for heavy lift is in the form of building out that Dyson swarm, I expect many, many competitors to come out of the woodwork. Many of them are probably already known names, and some of them—or maybe all of them in aggregate—I do think will give SpaceX a run for its money.

All right, we could put a bet on the side on that. I think they will ultimately, but I don’t think that’s going to happen between now and 2029. And then you’re part of NASA’s infrastructure, and you’ve built the original, the initial Dyson swarm.

Also, don’t sleep on China. China is busy copying everything it can out of SpaceX, and I’m sure there will be Chinese heavy-launch capabilities and multiple Chinese Dyson swarms as well.

They’ve tried. No successes yet. All right, let’s move ourselves out of this. By the way, by the time people have watched this episode, Schrödinger’s cat has happened: either V3 of Starship has succeeded or it’s failed. If it’s failed, hopefully there’ll be another one following quickly. This has been 5 months since the last launch.

Hopefully, now that V3 is up and operational, we’ll see a lot more frequent launches. The one thing you’ve got to appreciate about Elon is that he’s not afraid to fail forward. All right, moving on.

I love this article, this story. GPT-5.5 Codex is leading in forecasting. OpenAI built something called FutureSim that replays the internet a day at a time, day by day. It gives AI agents access to the real news starting from January 1, 2026, onward, then asks them to forecast real-world events over the next 90 days.

GPT-5.5 Codex is scoring 25% accuracy, leading across all the frontier models. It beat Polymarket crowd predictions for the Super Bowl. The way I think about this is that it’s the beginning of giving AI wisdom.

I’ve said this before: if you think about what human wisdom is, it’s the ability to go to the elder council and say, “Which direction do I take? Which way do I go?” And the elders say, “If you go this way, based on our experience, it’s not going to end well. You go this way, and it’s likely to win.”

If AI is able to run high-resolution simulations a billionfold—a billion times—it’s going to say, “This is the highest probability of success.” Alex, what do you think about that?

Alex Wissner-Gross

Remember Isaac Asimov’s psychohistory from the Foundation novels?

Yes.

Alex

Right. The premise of the Foundation novels is that a mathematician named Hari Seldon—in this grand, galactic-scale science fiction—develops a theory that he calls psychohistory, which is able to predict human civilization on a grand scale. He’s able to predict the collapse of the Galactic Empire and a whole bunch of other interesting things.

FutureSim is a benchmark. Just a minor correction: I think FutureSim itself, as a benchmark, is not from OpenAI. It’s from a group of independent researchers, but it does benchmark a bunch of models, including models from OpenAI. It’s such a clever architecture.

It’s benchmarking the ability of a variety of models to predict events beyond their knowledge cutoff date and without access to the web, so they can’t look up more recent information. The fact that the state of the art right now is already at 25% accuracy means these are the worst psychohistory models we’ll ever have, if I may borrow from Asimov.

I agree with the contention that, if you extrapolate this out, we’ll get Monte Carlo research of policy decisions. We’ll be able to predict, to some extent, predictable events in the future, at least as a function of perhaps human actions.

Peter, you and I spoke about this a little bit in Solve Everything, in some of our forecasts for what would happen in the latter half of the decade, from 2026 to 2035, about planetary-scale solutions. I do think that with the ability to predict planetary-scale outcomes comes the ability to predict planetary-scale interventions.

Not unlike how I would argue disease is probably going to get solved, this will be the planetary analog of what a virtual cell can do for curing all disease. If you have a perfect digital twin of the system you’re trying to fix, you can exhaustively test all possible interventions to get from a bad state to a good state. These are the greatest tools we could have for a positive outcome for humanity.

Selene, what do you think about this?

Selene

Well, this is going to be incredibly powerful for the boardroom because you go from quarterly updates to real-time sensing. We’ve actually built this into the architecture, following your paper, in the Organizational Singularity architecture, and I’ll talk about that a bit later.

There’s another implication of this that we’re glossing over. Right now, if you look at what New York does, there are many, many hedge funds that specialize in different areas—semiconductors, retail, and whatever.

And those are all supported by prime brokerages. These are much bigger banks that do all the trading and accounting and—

Selene

And those are prime brokerages. That entire industry could turn into just 1 or 2 AI models. The concentration-of-wealth effect from this would be insane.

But if the AI is just fundamentally better at picking markets, it’s not going to sit there and just do 1 market.

It’s going to expand quickly across all markets. If this plays out the way it’s starting to, you’re going to see a collapse into just a couple of megafunds that have massive AI budgets.

Dave

The financial singularity.

Selene

Yeah. Did I just hear Dave and Peter? Did I hear you argue in favor of indexing versus individual stock picking?

Oh, God, no. That’s not an index. The index fund is blind. This is so much better than an index. Well, actually, it’s an active index. I’ll call it a half agreement if you call it an active index.

Selene

Okay.

Well, the financial markets are still—Dave, you and I have been texting back and forth about the predictions that Leopold made on energy infrastructure, and even reducing his level of interest in chips. Those are still playing out. We’ve seen this: the demand for tokens is outstripping supply, and what’s keeping that limited is energy and infrastructure.

Hey everybody, you may not know this, but I’ve got an incredible research team. Every week, myself and my research team study the meta trends that are impacting the world—topics like computation, sensors, networks, AI, robotics, 3D printing, and synthetic biology. These meta-trend reports I put out once a week enable you to see the future 10 years ahead. If you’d like to get access to the Metatrends newsletter every week, go to diamandis.com/metatrends. That’s diamandis.com/metatrends.

All right, let’s stay with OpenAI. OpenAI launches personal finance in ChatGPT. This is a finance mode that’s able to access 12,000 financial institutions, letting ChatGPT Pro users ask personalized questions about their spending, their debt, their taxes, and long-term planning.

This is OpenAI eating another vertical. They did it with search, they’re doing it with coding, and now they’re coming after the $12 billion personal-finance app market. One could say maybe Mint will be dead. NerdWallet should be nervous. Two hundred million people already use AI for financial questions. Dave, what do you make of this? Selene, let’s go to you next after Dave.

Dave

Well, this is part of an overall trend where the foundation-model companies are starting to roll out legal. Now they’re rolling out finance, so they’re rolling out APIs that enable the vertical disintermediation of all these vertical companies—lawyers, financial accountants, whatever.

I think it’s creating an ecosystem of new startups that are early adopters of these APIs, which can then disrupt the markets. I think those companies are going to do incredibly well while the foundation-model companies do incredibly well.

But what a lot of people are overlooking is that if you’re in New York and you look at this massive financial institution with 20,000 financial advisers, you can’t envision it getting disrupted. There’s just so much mass and concrete and meetings and files, and so much regulatory barrier.

But over here, this parallel economy is growing, which is the AI economy working within itself. Everyone in New York is saying, “Yeah, but all the money’s over here in the banks. It’s not over there in that new economy.”

Well, after these next IPOs, that money will be in the new economy. It goes through the public markets, through SpaceX, Anthropic, and OpenAI, back into the agent-to-agent economy, which is an entirely parallel banking and finance system.

Independent of the original—

Dave

Everything else.

Yeah. Well, if Elon’s right, it’ll be 10 times bigger than everything you see in New York in about 10 to 20 years, and growing on a much faster curve. It doesn’t care about any of your legacy baggage. It’s going to grow on its own.

I think a lot of people are coming around to this view: we can leave the concrete world alone. We don’t need to scare everyone and disrupt it because we’re building a parallel AI world here anyway, and it’s going to be bigger anyway. These are interesting stories within the story.

Selene, I’d love to get your take on this.

Selene

Yeah. So this is essentially the Solve Everything thesis playing out at scale. You move from traditional models to an inner loop of intelligence and everything else wrapped around that. That’s exactly the architecture we have in this Organizational Singularity stuff we’re doing.

The bank should be terrified because the interface to money is shifting away from them to the AI, right? And they’re going to lose control in a very dramatic way. Financial advice will get layered around this, not around an individual.

I think the other broader point that you made, Dave, is that we’re moving. We don’t have to disrupt the legacy. You build this completely new architecture and just let that become the new gravity center. Go straight to the old Buckminster Fuller quote: “You can’t fix an existing system. You have to set up a new system at the edge and let that become the new gravity center.”

We’re going to see that happen in legal, healthcare, insurance, and education. It’s just going to happen over and over again. The thing that filled in just this month and is so newsworthy is that the conduit of the money from the old to the new is now really clear. Where Alex is an adviser is a really good example, or a security where you can deploy your money into AI and get it right on Robinhood, or buy it on the exchange.

That conduit then moves the money from the old economy into the new economy, and once it’s in the new economy, it doesn’t care about the legacy banking system. The same is true with these IPOs that are coming up: trillions of dollars that are moving.

Dave

It’s like the Bitcoin ETFs. Once you’ve got it over there, you don’t care.

Breaking news, by the way—and I’ll hand it to you next, Alex. OpenAI is sort of letting it be known that they are preparing to file for their IPO as early as this week, as early as Friday. They just won the case against Elon, so they’re feeling their oats.

We talked about this a few pods ago. It’s a race to get access to the available capital. SpaceX is going to suck a lot of the oxygen out of the room. Their competitor, Anthropic, is preparing for an IPO.

It’s interesting. Remember, we had conversations with the CFO of OpenAI, Sarah, saying they were going to file in 2027 because they weren’t ready. But here they are talking about filing ahead of Anthropic to get access to that cash flow. God knows they need the capital to build out their compute. Alex, over to you.

Alex Wissner-Gross

As I closed my newsletter today, the day of recording: “Cogito, ergo IPO.” I think, therefore I IPO, is the thinking of the moment.

When I look at the announcement of OpenAI launching personal finance, I ask myself: Where’s the monetization? OpenAI has, as I’ve mentioned numerous times on the pod, largely pivoted away from consumer to enterprise and needs to justify very high-value productivity per token. I ask myself: Where’s the value per token in this?

Is OpenAI really seeking to become a financial institution? Doubt it. The value, I suspect—this is a bit of Kremlinology—is advertising. OpenAI is following the Google playbook.

Why did Google go into developing all of these financial verticals? Because some of the financial queries—and Dave probably knows better than any of us—can be enormously lucrative from an advertising perspective. So, if OpenAI hopes to monetize personal-finance conversations, I think they’re probably just going to do it by running OpenAI ads for consumers relating to personal finance.

The more that they can tailor particular information to the particular circumstances of a retail investor with all of these integrations, the better they can target ads in those conversations.

Today, Dario Amodei announced no ads on Anthropic—just flat out, “We’re not doing advertising.”

Alex Danco

Which are very convenient talking points, given that they’re focused on enterprise.

Yeah. They don’t have the consumer anyway, so it’s very easy to be an angel in great ethics.

Yeah. There you go. By the way, one more thing about this: It feels also like a bit of a copycat. After Anthropic launched all these plugins for legal and other areas, here comes OpenAI doing the same thing for personal finance.

Alex Danco

But critically, Anthropic’s suite of skills is targeted at businesses, whereas OpenAI is targeting this at consumers. OpenAI wants to charge enterprises. So, yes, I’m sticking with my ad theory.

Yeah, that’s right. Google, too. It’s OpenAI versus Google. We’re watching this ad evolution.

I think you’re right. Yes.

I think you’re right, buddy. Okay, I’m going to turn this over to Alex. Here’s the story: An OpenAI model disproves a central conjecture in discrete geometry.

An OpenAI model today—or this week—disproved a long-standing conjecture from Paul Erdős, one of the most prolific mathematicians in history. Alex, tell us about it. Actually, let me run this video and then give us your blow-by-blow. What does this mean? What does this mean to the average listener?

Alex Danco

I think what’s significant about this moment is that it’s the first really clear example of AI solving not just an unsolved math problem, but a really well-known unsolved math problem.

This is the first mathematical breakthrough due to an AI. It’s been described as the most well-known problem in combinatorial geometry. So, for a whole subfield of mathematics, it’s maybe the best-known problem there is.

I remember seeing an initial version of the model output. I didn’t really believe it. It took quite a while, reading it over and trying to figure it out. This problem is about points in a plane. It’s a completely elementary geometric problem, but the solution involves really deep tools from algebraic number theory.

It was believed that the construction was basically best possible. But what our model did was show that this construction could actually be improved by quite a bit.

Alex, your explanation, please.

Alex Danco

First, just a few seconds about what the problem is. This is a problem posed by the famous Hungarian mathematician Paul Erdős about 80 years ago. The problem was basically asking the question: If you have a plane, a 2-dimensional plane, and you can put n points in the plane, what is the maximum number of pairs of points that can be separated by the so-called unit distance—basically, by a fixed distance?

It’s a very simple-to-pose problem, but very hard to answer. Erdős’s original conjecture, which held essentially until now, was that it was effectively impossible to do much better in terms of the number of pairs that could be separated by this fixed unit distance. He conjectured it was impossible to do materially better than some number that’s proportional to the number of points themselves—basically, some number of pairs that’s linear in the number of points.

Now, for the first time, OpenAI has revealed that an internal model that hasn’t been publicly released has disproven that conjecture and found weakly superlinear scaling. Why should anyone care? This is, as I like to say, math is cooked. This is going to be the new Exhibit A that I cite for how cooked math is.

This is one of the most important problems, as the video mentioned, in combinatorial geometry. It stood for the past 80 years, and now AI has just solved it. Notably, if you unpack all of the accompanying documentation and commentary, there’s quite a bit of interesting commentary.

This wasn’t something like the four-color problem, which one might imagine, given that it involves combinatorics. The four-color problem is: If you have a 2-dimensional map, what’s the minimum number of colors that you can use to color each country to make sure that no 2 adjacent countries are the same color?

There are problems in math, in combinatorics, like the four-coloring problem, that tend to be exhaustively solved by AI. Then mathematicians and others point to those exhaustive brute-force solutions and say, “Gosh, maybe AI is more exhaustive. It’s better at brute force, but it lacks human brilliance. It lacks leaps of creative insight.”

This is not a problem like that. This is a problem that has the top mathematicians in the world, who specialize in this particular area, looking at its reasoning traces and concluding that not only was it faster in some sense, not only was it able to brute-force lots of different theoretical approaches, but it was also smarter.

It was smarter in an interesting way. One of the commentaries—and I would definitely encourage everyone to go to the OpenAI website and read a number of professional mathematicians’ commentaries on the reasoning chain of thought that it used to solve, or really to disprove, this conjecture—made the point that, from looking at the reasoning chain, they could see that it was pursuing all sorts of exotic possibilities that humans would be too exhausted to pursue.

So it was arriving at creativity by both being faster and being able to brute-force all sorts of outlandish possibilities. In the end, one of those possibilities—and I think the language from the chain of thought that ultimately led to the solution began with something like, “Optimistically, if I pursued this, something might happen”—turned out to be the solution.

If we remember the infamous Move 37 from AlphaGo’s match with Lee Sedol, and how being able to brute-force, but with clever learned policy search, the reasoning tree of a Go game is starting to play out in math, we’re going to see that play out everywhere else as well. It’s going to play out in physics. It’s going to play out in every science and in engineering.

Alex Wissner-Gross

#solve everything, Peter. We’re seeing it.

The starting gun. While people may not relate to math, they sure are going to relate to physics, chemistry, biology, and materials science.

These are going to give birth to trillion-dollar outcomes. I think, in the past, we've had very hard Humanity's Last Exam questions on the pod and made the point that you would have to think hard for 3 or 4 hours even to understand the question. Then, of course, Alex says, “I think the answer is 4,” and it turns out to be right.

But putting that aside, this one—if you rewind the video and listen to the first couple of sentences of what Alex said—that's the whole problem. If you like crossword puzzles or Sudoku, you're going to love this one. Listen to it again. I'd really encourage the team to splice in the images because, for the last 80 years, humanity thought the best solution was this really simple grid. It's just points in a square.

And if you look at the final AI solution, which is not proven to be the best solution—it's just better than the square—it's beautiful and elegant, and it's not intuitive at all. It gives you a lot of new insight into, wow, AI is going to be really good at things like magnetic bottles, protein folding, and chip layout. It looks a lot like a chip-design problem: we're laying out the wires in an optimal way. The final answer doesn't look intuitive to you at all as a human, but it's better.

So this is a much bigger moment in history than just solving a math problem. It's so noteworthy. I thought, Alex, you described it so beautifully and eloquently. Hopefully, history will—

Alex Danco

I think you did a great job, too. I think the solutions—the optimal solutions to things—aren't necessarily as human-legible as the human solutions. And I agree with the sentiment, Dave. I think this is heralding an era when AI solves very hard, in some sense, optimization problems, and the solutions look positively exotic, inhuman, maybe even biological. And it's never any slower than it is today.

Dave, do you want to close us out?

It reminded me, when I first saw this, of the Move 37 analogy. I think that applies. Super exciting.

All right, let's move to China. A Chinese AI group pulls ahead of U.S. rivals in video generation. China is winning the AI video race not because they have better models, interestingly enough, but because they have better data.

ByteDance's Seedance 2.0 and Kuaishou's Kling now rank number 1 and number 2 on independent video-model leaderboards, beating every American competitor. Simple reason: TikTok has generated billions of hours of video data that no U.S. company can match. Let's take a quick look at a video clip here from these models.

I won't play the whole video, but the idea here is that they kind of look the same. According to the data, they're beating the pants off us. Alex, what do you make of this?

Alex

The rules of copyright seem to operate somewhat differently in China. If you have access to more video from whatever source, however legal or illegal it is, you can train better models.

I think when we were first discussing Seedance 2.0 on the pod, I made a similar comment. It's utterly remarkable to me—the videos that seem to be popping out of Seedance 2.0. Not because of an algorithmic innovation, but because, for whatever reason, legal or otherwise, these Chinese frontier labs have access to more data.

There are videos floating around the internet. I think most astonishingly, there's a video floating around of a guy inserting himself into key moments in the Harry Potter cinematic universe. I linked to it from my newsletter—key moments, stabbing or otherwise violently intervening with unpopular characters.

I think we're probably going to see quite a bit of this. It makes the boundary between fan fiction, which historically was limited to text or maybe images, and video just completely dissolve that divide. I think the copyright lawyers will probably have a field day if anything like this is tried in the West. But for now, it's the Chinese models that seem to have all of the Western and Chinese video-training data to do it.

Alex, is this the place where you see China leading American models the most at this time? It's sort of a stereotype, a cliché at this point, that China was always going to have more data because it could pull more data from civilians and maybe also more pirated video data. Maybe it's also in a somewhat better position in terms of energy generation. It's still in a weaker position algorithmically and from a chip perspective. So, do they pull ahead in video generation?

Right now, it seems they're pulling ahead in consumer video generation, even relative to Gemini Omni, which we discussed in the last pod. But maybe there will be some enormous algorithmic innovation that enables the West to take the lead again in a few months. I don't know.

Alex

You know, we mentioned in the last pod that Google is really the only American lab still pursuing multimodal, and we've seen all the open-source, open-weight models out of China doing multimodal.

That's right. Interesting, Dave, do you have any comments here?

Dave

Yeah. If you study how video generation actually works, it's using the same Transformer algorithm that's caused all these other breakthroughs. Under the covers, it's the same 2017 massive breakthrough driving video generation as well.

But what the Chinese have done here is, you take video content and compress it into a latent space, and then decompress it into video and image creation. Within that latent space is where all the innovation happens. But other domains, like chemistry, biology, and all of physics, also have latent spaces.

If you see the Chinese get ahead in video generation and sustain the lead, that's a good leading indicator for every startup trying to say, “I don't need to compete with Anthropic and OpenAI because I'm better at chemical reactions. I'm better at robotics. I'm better at whatever.”

If you can maintain a lead at better AI within any of those latent spaces, that's a really good sign for all the startups because they can then actually maintain that lead within, say, ExO. I can actually have the best company-management latent-space knowledge. It's a really interesting, cool leading indicator. I'm kind of cheering for them to keep their lead.

You know, which is pretty daunting versus Google because Google has all the YouTube content. China has all the TikTok and all the short-form content. But it doesn't seem like China has a massive data advantage. They're just working on this problem harder.

Yeah, I am curious how these models perform in terms of the speed of generation. We've all talked about the notion that, in the future, I'm going to be generating video on the fly as it's needed—sort of Netflix on demand. Alex, when do you imagine we're going to see that level of video generation?

Alex

A few months ago. World models, including Genie, already do that. That's what you're asking for: interactive video generation and real-time world models already do that.

I'll tell you, we—you remember, I don't know if you remember, Peter, but we saw Liquid AI a year ago generate images as quickly as you could speak, almost instantaneously coming out of Liquid AI. If you try to do that today, you wait like a minute or more. The experience is nowhere near as much fun as real-time creation, like creating as fast as you can think.

But we're so short on compute. That's one of the reasons Liquid AI is doing well: it's so much more efficient. But that holodeck experience we're trying to create is entirely possible to build today, and nobody has the compute available to deliver it.

Yeah, any thoughts?

Dave

I'll echo what we talked about already: the amount of data that China has. TikTok and Douyin are huge, essentially training loops disguised as entertainment platforms, right? I think they'll stay ahead for a bit longer, but I think the models will catch up.

Our next story is an interesting one. I want to focus on this a little bit. Here's the deal. A friend of the pod, Eric Schmidt, was giving a commencement address at the University of Arizona. As he mentioned AI, he got booed. It's pretty brutal to watch. We'll show the video in a second.

Gloria Choi, who's the VP at Tavus, is a friend and had the same treatment. The generation today that's entering the workforce is angry and scared about AI disrupting their careers.

Let's take a look.

Eric Schmidt

I know what many of you are feeling about that. I can hear you. There is a fear. [Cheering] We do not know. [Applause]

Wow.

Eric Schmidt

We do not know the precise contours of what this transformation will be. The rise of artificial intelligence is the next industrial revolution. [Cheering]

What happened?

Eric Schmidt

Okay, I struck a chord.

Selene, let's go to you first on this one.

Well, look, the students are sensing that companies and institutions are adopting AI, and we've not redesigned the social contract. The backlash is not anti-technology; it's anti-extraction. We need a new narrative around agency, not replacement. And there's a huge legitimacy gap between AI elites and young people today. It's a kind of ridicule. It's incredible to see.

Yeah. Dave—

Dave

Well, this is clearly priority one for XPRIZE now. I mean, you've got—

You've got $100 billion of charitable money suddenly unleashed at OpenAI. They don't want this at all.

Dave

Historically, no one expected this to happen this quickly, so they put very little effort into it. This is their wake-up call. Look at these videos. Eric Schmidt is a hero—one of the greatest, if not the greatest, business executives of all time—but when the job market is basically zero for college graduates coming out, what do you expect?

The sentences there weren't exactly inflammatory, right? “AI is the next industrial revolution.” That's not exactly controversial. Imagine if they'd said something controversial, like “a revolution in real time.” But that's the reality outside of the couple of places in America where AI is happening.

The rest of the country is like that, and it's really important that people be aware of it. Take that $100 billion of charitable money and get to work on this.

We just launched the Build with Gemini XPRIZE this week, and I'm going to take a second to talk about it. We are challenging teams around the world. Google put up $3 million. It's a good start, right? We're encouraging teams and individuals to become entrepreneurs. Pick a problem that impacts 100,000 people and build, in public, something in 3 months that generates and scales revenue. That's a real problem that generates real revenue.

After 24 hours, over 2,000 people have registered. I hope we'll get to tens of thousands of registrants. There is your chance to use these tools and learn to use them. Dan Martell, who's one of our donors, is getting a huge amount of interest from the people he's an incredible educator to, ages 13 to 25, who are going to be going after this XPRIZE. And thank you to Dick McKinney as well for contributing. A shout-out to the Gemini team at Google for their support on this.

Rather than booing AI, use it to take control of your own future. Don't get a job; build a job and employ other people. That's your option right now.

Can I say something more about this—

Please?

Sim

What they really should be booing is the universities that have sold them a credentialing system that is radically out of date. Amen.

Right.

Sim

Amen. They should, because universities need to stop defending that old credentialing system and become launchpads for agency and entrepreneurship. We've talked about this ad nauseam on the pod. That's who they should be booing. They've been sold a bum deal, and they're in huge debt with no prospects of getting out.

They spend $200,000 for a degree that's worthless by the time they graduate.

Yeah. If you're graduating, don't compound the error by getting into a job-training program at an investment bank.

Dave

But just think about this for a second. If you go back to our age—or even your parents' age—there's a very simple question we could ask when we all graduated, whenever that was, however many decades ago: How much of your university education did you actually use in the workplace?

The answer was nearly zero. And that was 30 years ago, not just today. So this has been a problem that's been around for a very long time, and it hasn't been solved. They're getting pissed off about it, and it's understandable that they're pissed off about it.

Yeah. I'll just add that I used virtually all of my education for what I do.

Yes. [Laughter]

If Eric Schmidt and Gloria Caulfield were getting booed at Harvard, MIT, Stanford, or a similar institution for mentioning AI, then I'd be concerned. If I remember correctly, these weren't Harvard, MIT, Stanford, or similar.

No, but there's a majority of the graduating class of 2026.

Alex Wissner-Gross

What that says to me is that virtually everyone in that class is anchoring their expectations on stagnation, and that's not a great place to anchor expectations at all. But also, to Salim's point about how they should be booing the university—and to Peter's point about how they should be booing the college—there's selection bias.

You don't get to graduation if that's your attitude in general. You probably skip college altogether. So I think there's a bit of selection bias here. We focus just on those who've narrowed their possibility space while not leveraging AI to its fullest, and then choose their graduation ceremony as a time to channel their angst regarding the disappearance of the lower rungs of a professional ladder that AI is automating away.

It's probably too little, too late. They should be starting businesses.

Inevitably, I'll get flamed for saying this, but not everyone wants to graduate, either.

Speaking to the parents who are listening—those who've got kids in high school or college today—or speaking to recent graduates who haven't gotten a job, this is only going to get more dramatic, and it's coming. You need to encourage your kid to learn entrepreneurship. Use the Build with Gemini XPRIZE as an excuse to play.

I'm going to do that for my kids, who are turning 15 next month. If they should win, we'll donate the money back because I'm biased. At the end of the day, use this as an excuse to play, right? The 2 most important mindsets you have are purpose and curiosity. You can learn anything you want. Yeah, go ahead. Go ahead.

I need to go on a little rant here. We've been talking about purpose for more than a decade. We coined MTP back in 2012. It's not decorative. In a world of AI capability, purpose is how you orient human beings.

Somewhere around 2017, I was asked to give a keynote at a conference, and my team said, “The logistics are really hard to work out for you to do this.” I'm like, “What's the conference?” They said, “Oh, no, it's a conference of 700 deans of business schools.” Who knew? But they all get together. I'm like, “Hell, I absolutely want to go do that.”

So I get up on stage. I'm giving the opening keynote at this event. The announcer says, “Hey, we're happy to have Salim here. He's going to tell us the latest thinking in EXO, et cetera, et cetera.” I'm standing side stage, and what you see from the crowd is a completely blank look. They have no idea who I am, what the book's about, or anything.

The announcer notices this, and he goes, “How many of you have read Exponential Organizations?” Out of 700 deans, 2 put their hands up. Not that everybody in the world should read the book, but if you're a car designer and the Tesla comes out, you should jolly well know what the hell the thing is.

Here you've got 700 deans of business schools who had no idea that there was this other paradigm out there. I think that's just criminal negligence, in a sense. Universities have been sitting on their hands for decades, knowing this problem. But the immune system in academia is very strong. They're not able to get out of this. This is why we need completely new systems that route around legacy. End of rant.

Yeah, agreed. A parallel story here comes out of Stanford. A Stanford survey found that 49% of 849 computer science majors would rather cheat than fail. Students say AI tools are used in nearly every class for homework, coding, and essays.

Stanford brought back proctored, in-person exams for the first time ever to deal with this. The honor code that Stanford was famous for is effectively dead. This is happening at the world's top computer science program. Imagine what's going on in the rest of the world. Alex, what do you make of it? This is Stanford.

Alex Danco

It shows me that there's a bit of an overhang: the skills being taught are being automated away by AI. So it's not unnatural for these students to at least be considering using AI for all of their homework, coding, and essays, doubly so at schools like Stanford and Princeton that have historically had honor codes.

I've never quite understood the logic of a so-called honor code. It seems to me a recipe for laziness on the part of faculty. They should just be supervising the exams. If you want to make sure that calculators or AIs aren't being used, earn the vast tuitions being paid to universities and supervise.

I think, if anything, this points in the direction of far more supervision and far more proctoring if higher education is going to remain at all recognizable. I'm not sure that it will, or even deserves to. Maybe there should even be some sort of wilderness camp at the university level, where students are denied technology and asked to work in the style of Vernor Vinge's Rainbows End and “Fast Times at Fairmont High.”

Students have to interact with the real world in a way that's impossible to cheat without succeeding, like building their own actual businesses instead of writing business plans, or actually solving hard problems in computer science instead of solving formulaic tests that require humans to be in the room to supervise.

Alex Danco

Use the direction this goes in. Use AI to go 100 times bigger than you normally would.

Exactly. Dave, you're teaching at MIT. What are your thoughts on this?

Well, I teach a class called Foundations of AI Ventures, where you have to build a business plan. So it's already exactly on the mission that Alex was outlining. If your business plan gets funded, you get an A. So it doesn't have this particular problem, but it seems obvious to me that you want to be teaching the students to use AI.

If you just look at their prompt stream, you can grade them. You don't have to have tests at all, anyway. It's what Alex said: The schools are struggling to hold on to something that goes back to when graduation speakers would show up on horseback, and you'd expect the crowd to hear wisdom they couldn't hear any other way from some great human being.

Why do we need graduation speakers today if we have podcasts? Well, we don't. Then why do we do it? Well, it's traditional. It goes back 150 years. The exams do, too. Of course, it makes no sense, but it's really hard to let go of tradition after tradition after tradition. What do you think the singularity is going to be like?

Alex Danco

Is this actually cheating, or is this what students should be doing? Do you expect these students not to be using AI when they get to the real world? Why are you training them for something that isn't going to exist in the future?

It's even worse. You're putting him in an impossible situation. Forget the tests, but even in the homework: “No, we don't think you should use AI.” What does that mean? You just put the kid in an impossible ethical situation. All you're doing is torture.

Alex Danco

Look, you're not testing durable human capability. You're testing for compliance and formatting, and that's just not the right test for this.

To connect this to the previous slide, universities have a choice. They either become credentialing museums, or they become AI-native talent accelerators. You're going to have to make that bifurcation, and you're going to have to make it fast.

That also yields—let's go to the positive side—the biggest entrepreneurial opportunity in the history of education right now. The next generation of education entrepreneurs and companies aren't going to sell you courses. They're going to sell you capability acceleration—

—for the rest of your life, and not just for a 4-year period of time.

Alex Danco

Yeah, it's going to be an ongoing partnership throughout your entire life: your education, learning, and coaching partner.

All right, next story here. Meta has installed software on employee computers that tracks mouse movements, clicks, and screen activity. The stated reason is training AI agents to understand how humans use computers. The real implication is that Meta is recording everything its employees do so AI can learn to replace them.

Employees launched protests at multiple US offices, and an engineer's internal post about it was viewed 20,000 times. This happened the same week that Meta cut 10% of its global workforce. We're seeing this in a lot of places, and this ultimately is Elon's plan with Megasoft, right? I mean, Megasoft—he wants to be able to come in and replace all your employees with some percentage of a GPU after uploading their capabilities.

Macrohard. Thank you.

Megasoft is a trademark.

Megasoft. Okay.

Alex Danco

Macrohard is a great name, though.

Macrohard. You know, he wants to be able to come in and replace all your employees with some percentage of a GPU after uploading their capabilities. Everybody, figure it out. This is what Amazon is doing with its delivery workers, putting AI glasses on them to track their movements so that future robots can do deliveries at home.

We're seeing this everywhere, and it's just beginning. Right now, 44% of Gen Z workers are deliberately sabotaging the AI they're supposed to train. That's the backlash that you see there.

I think this is the difference between an AI coach and an AI cop, right? Same data, but a very different outcome. It depends on how you frame it, and I think they need to be very careful about how they frame it.

Frankly, the big issue that Meta has is a pretty big, understandable lack of trust around what they've done. They've repeatedly said, “We'll never sell user data.” And then there's privacy data. You can back out of these things. Then you find you can't back out of these things.

WhatsApp was supposed to be encrypted. Now WhatsApp is taking the encryption away. It's what Cory Doctorow calls the whole enshittification stuff. We're seeing that now internally in the organization, not just externally in the products.

Alexandr Wang

I'll give a hot take on this one, Peter. It seems like such a strange decision, given that the frontier labs—which, at the moment, I don't consider Meta a frontier lab. I think we have, at this point, 2½ American frontier labs, none of which is Meta.

The frontier labs are all purchasing enormous amounts of synthetic data for computer-use assistance training across a wide variety of environments. It seems a little strange to me—maybe I'm missing something—for Meta to view its own internal computer use as a valuable pretraining, or more likely post-training, source.

It strains a little bit of credibility for me to think that there's enough diversity of computer use just within Meta for this to be worth all of the hostility it's inevitably incurring from employees, on top of all of the workforce cuts. It makes me wonder if this is basically a way of encouraging a subset of employees to just quit Meta anyway.

If I were looking for better post-training data at this point for computer use, I'd be leaning heavily into synthetic data. I wouldn't be antagonizing thousands of employees with mouse tracking. That's my hot take on this.

That's fascinating. Dave, what do you make of that?

I completely agree. I was going to say a much more violent version of what Alex said. The use of this data is nonsensical as an internal tool. This is entirely about knowing what your employees are doing and using it to evaluate and sort performance.

I think people better get used to it, because it's going to happen pretty much everywhere. Also, if you look at all the data that Google has gathered for decades now, I think a point that Eric Schmidt made years ago, very quietly, was that if Google ever wanted to make more money, they see every quarterly earnings report.

I had that exact conversation with him. He said, “We could make a ton of money just once before lawsuits start flying.” They see—

David Friedberg

Exactly. The higher-level point is that they see all the searches for all the purchases ahead of earnings reports. Oh, and once Gmail took over, everyone sends their board reports out to board members. I'm on so many boards. They arrive by Gmail, naked, just as attachments.

Google's terms of service say, “Look, we human beings won't look at all your emails, but our algorithms will.” Humans could, too, but they don't say they won't. They imply they won't, but the algorithms look at them.

The high-level point there is that the data gathering is happening for sure, and you're crazy to misuse the data because of the backlash. This is what Alex said: Why did you just create a negative news report that you didn't need? Why do that? We know you're gathering the data, but don't abuse that gathering in a way that creates a massive uproar like this, because that's what's been going on in China for decades.

Google has all the data. Apple has all the data. They don't really look at it. The data gathering is happening, and, Peter, you've made that point over and over again.

Getting in an uproar and fighting it is crazy, though. You're just labeling yourself as a difficult-to-hire person. It's not going to stop because you're ranting about it. I'm not saying, “Just adopt it,” either, but don't just rant and then go home and feel like you did something. You didn't achieve anything.

Tying a story here back to the conversation about job loss, Mark Cuban proposed a federal token tax. This is his quote: “We should tax tokens at a provider level at less than 50 cents per million, as it will push big AI players to optimize tokenization. It will also reduce energy usage and generate $10 billion a year, maybe growing 30× or 100×, while also creating a source for paying down the federal debt.”

We've talked about the idea that we're potentially going to tax AIs that replace employees, or robots that replace employees, but these numbers here that Mark puts forward are de minimis, right? This is for 300 million Americans. This is about $33 per year. It's an interesting idea, but I don't think this moves the needle anywhere.

Alexandr Wang

Certainly. There are so many perverse incentives when one introduces novel types of taxes or floats taxes like this. The first obvious novel incentive is to get rid of tokens entirely.

There are many token-free or tokenization-free approaches to autoregressive and nonautoregressive machine learning at this point. We could switch to diffusion models entirely if we wanted. No, there are ways to do diffusion models with no tokens, or we could skip the tokenizer step entirely and stick with transformers. All this would naively do is push for the abolition of tokens altogether.

I think then the response is, “Okay, fine. Let’s tax the FLOP. Let’s tax floating-point operations,” or something. There are more perverse incentives. I think it’s very difficult to come up with an input resource that would be best taxed, short of actual dollars or some function of dollars, that would actually enable a pure-play targeting of the AI labs without creating severe perverse incentives like this.

I mean, Alex, you remember what Sam recently talked about: we should have the general public own a part of national compute, right? This is the equivalent of the Permanent Fund in Alaska or what occurs in Saudi Arabia or the Emirates. As all of the capital starts flowing into the agentic ecosystem, into these frontier labs, and they start generating huge amounts of wealth, how does part of that wealth get distributed to the average American? I think we’re going to start testing that.

Alexandr Wang

I think we saw it in the form, just in the past 24 hours, of offering $2 million of OpenAI tokens in return for a SAFE to all YC companies—like 2,000 companies. I view that as a preview of some universal-basic-compute-type offering to everyone. It’s not quite obvious what the equity equivalent is, but it’s something like that.

Yeah, Dave.

David Friedberg

It’s amazing how quickly we’ve decided that LLM AI and the word “token” are fixed things that can be taxed like bananas or shipping containers. But anyone who actually works in the industry knows that you can tokenize with byte-pair encoding, which is what you usually do. But you can use 1-byte encoding, 4-byte encoding, or no encoding at all. Suddenly, what you tax disappears from the world a minute later.

It’s just funny that people feel like, “I’ve got an idea. Let’s tax per token.” It reminds me of a conversation in Riyadh where one of the members of the royal family we were meeting with was saying, “What do you think a good cost per token is?” They were big investors in a couple of companies, and their cost per token was like $1 per million. I’m like, “With what context window?” That’s a nonsensical question. They were like, “Well, we’ll just answer it.” So I gave an answer, and they were like, “Great, because we’re less than that. We cost less.” Okay. This is going to be really tricky, amorphous, and fast-moving.

Look, it’s interesting at one level because it treats AI usage as an economic activity, right? It goes to Alex’s comment from another podcast, where we talked about what’s the economic return per token, and that’s really interesting because it’s not just a software feature. Historically, governments have always taxed what society’s dependent on, whether it was land, labor, trade, or income. If AI becomes a new engine of productivity, then tokens—or some flavor of compute or energy—become obvious tax targets.

The challenge is that if you try and tax it, there’s nothing more liquid than compute usage today. It’ll go to people that don’t tax it, and you risk doing that. The other big danger is that you may put an inadvertent fence around Meta and OpenAI, the ones that can navigate this easily, and slow down innovation because startups will have a harder time around this. So there’s a whole bunch of different decisions here, but I think the flaw in trying to tax tokens itself—the intent is right, but the mechanics may not be right.

David Friedberg

Yeah. I think you’ve already laid out the plot for maybe Neal Stephenson’s next cyberpunk novel. It’s got to be about compute tax havens on Peter Thiel-funded data centers, seasteads on the ocean to avoid all the token taxes.

We should just list all of the science-fiction novels with a number, and in each episode we just list out, “Here are the 45 science-fiction plots we talked about in this episode.”

Bingo. All right, here’s a fun conversation. I’m a huge fan of Ben Lamm, George Church, and Colossal, and they just had an announcement that I wanted to share with everybody. Colossal Biosciences has hatched chicks with an artificial egg. Let’s take a listen to this video.

Ben Lamm

It begins with the chicken egg. We didn’t want to just reimagine the egg. We wanted to completely re-engineer it. Now, showing the world that we can actually grow this whole bird in an incubator outside of an eggshell. Complete game changer. Unlocks massive potential. I am looking at the world changing, and I’m holding it in my hand. I think that this is really a big step for science. So, obviously, hold on to your chickens.

This is the Colossal artificial egg. As we developed the design, we began with a natural egg. So the Colossal artificial egg is the world’s first in terms of the design level and sophistication. The first major structure is a rigid outer shell. The shell provides protection and rigidity. Our permeable membrane allows oxygen to diffuse into the system through the membrane at ambient temperatures. It’s got a really big window on the top that you can actually look at and see and understand exactly what is happening to that embryo. Real-time visibility into every stage of embryonic development. This is a real breakthrough.

So this is ex utero gestation. This is one of 3 artificial womb programs that Colossal has going on. Remember, Colossal is the company that’s bringing back extinct species. Famously, the first one is the woolly mammoth. They brought back the dire wolf. They’re bringing back the dodo bird. They have 15 different species in process.

But it turns out the moa has a huge egg requirement, and just doing it is going to be challenging. Super curious. Just for the numbers, there are 1.9 trillion eggs per year, generated by about 33 billion chickens. So, interesting thoughts on this one, gentlemen.

Alex Wissner-Gross

I’ll just note that I can’t wait for the dodo, which I think is one of the targets for this. They’ve tried this now with a couple dozen birds. I think it’s a far cry from what we saw in Jurassic Park, where the emphasis in that Mr. DNA reel was more on the molecular biology of resurrecting extinct species and not the organism-level or egg-level focus.

I think it’s important, and as I understand it, there have been some major challenges that prevent the naïve Jurassic Park-type approach from working in the case of eggs for these extinct birds. Apparently, late in gestation, they require large amounts of oxygen that are difficult to supply through artificial eggs and so, at least historically, have been difficult to supply. Creating, effectively, an oxygen-permeable artificial egg that’s supportive of late-gestation metabolism seems like an important step.

The giddy sci-fi fan in me wants to know: How well will some of these techniques generalize to humans? Can we perhaps appropriately generalize these to ex utero gestation in humans?

They’re working on mammals and, obviously, avian species as well. Of the 15 species in the pipeline, 2/3 are mammals right now. The dodo bird, which is native to Mauritius—or was native to Mauritius—is on their money, on their stamps, and on their flag. The country’s very excited to bring it back and drive tourism revenue.

Alex Wissner-Gross

At least 2 of the 15—the dodo and what was the deer, the red or blue deer?

The bluebuck, I think, or something like that.

Bluebuck. Yeah, those 2 were extinct because they were delicious and easy to hunt.

So Salim is opening dodo fillet restaurants, getting ready, and maybe a burger—a bluebuck burger. Anyway, I just wanted to note that this is the science-fiction future materializing before us.

We’re not going to have dinosaurs, per se. Though what Colossal is doing is not actually bringing back an exact species. It’s bringing back a similar species, meaning we’re able to say, “We’re going to make these 300 edits in the genome that will make this species look like this. It’ll have longer tusks, so it’ll have woolly hair. It’ll be more cold-tolerant. It’ll have a longer snout.” And so you can sort of design your species.

I had Ben Lamm on stage in Miami at FII and said, “Can you create a dragon?” The answer is yes. Probably not fire-breathing, but we can add the wings and make it look like a dragon.

Don Malem

I think we’ll get dinosaurs, Peter. I’m pretty confident we’ll get dinosaurs. I’m not sure whether we’ll get a fossil, but we’ll get some version of a dinosaur.

You know, I don’t know if you remember, but my visioning proposal a few years ago was that genotype-to-phenotype mapping problem, where you start with a picture of the final outcome and then you try and create the DNA that matches the picture.

Don Malem

Exactly. That’s what they do.

Yeah.

Don Malem

Yeah. They’re building back. You could blind-test and see if one theory is that you created a completely different animal.

Alex

It just looks like the original. The other is no. The only way you can create the original is by getting the DNA within what could have been the reproducing line of that original species. No other DNA is going to get you to that exact look, feel, behavior, shape, or whatever. You could back-test that and get the answer to that.

We're going to have Ben Lamm with us on stage at the Moonshot Gathering in September, so you guys can drill into this. He'll be one of the Moonshot entrepreneurs we're going to be bringing with us.

Alex

It'd be ironic if we could resurrect stochastic parrots with stochastic parrots.

No one's going to get that.

Alex

You don't think so? Someone in the audience will appreciate it.

There's a thread that's been unraveling for a while, and it continues to do so: biology is becoming programmable. Yes.

Alex

That's just a huge thing.

He's working not only on biology for animals, but biology for plants—being able to design a plant that is drought-resistant, disease-resistant, and can grow twice as fast or twice as big. This is the intersection of AI and synthetic biology. Very exciting times.

Everybody, welcome to the health section of Moonshots brought to you by Fountain Life. 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 and get a chance to get to 100-plus. I’m here today with Dr. Don Malem, the chief medical officer of Fountain Life and a part of my medical team. Don, a pleasure.

Don Malem

Great to be here. The thing people are concerned about most when it comes to living to 100 or 120 is their cognitive abilities, making sure they don't have dementia. The numbers about dementia are problematic. Can you share what you've learned?

Don Malem

Such an important point. And you're right. At Fountain Life, the number-one thing our members are most concerned about is losing their brain health: forgetting the name of their child or forgetting the face of their loved one. We know that when it comes to dementia, conservative estimates are that 45% of cases are entirely preventable. What was amazing is that, with the advanced testing we're doing at Fountain Life, 25% of our members had an advanced brain age.

Wow.

Don Malem

What was really awesome, again going back to that prevention, is when we partnered it with healthy living. This gives me chills. Eating healthier, moving our bodies, and optimizing sleep are so important. What we saw was that we improved that brain age by 26%. That is a big number showing that the majority of those individuals were actually able to improve their brain age. One of the things I love about Fountain is that we’re searching the world for the best therapeutics and the best approaches and making sure we bring them to our members. So if having healthy brain function till 100 or 120 is important to you, check out Fountain Life. Go to fountainlife.com/pater. Make sure you become the CEO of your own health.

All right, now back to the episode. All right, 3 stories on data centers and energy. Worth noting. Let's take a look. A Gallup poll found that 70% of Americans oppose data center construction in their community. Nearly half strongly oppose them. Some residents said that they'd rather live near a nuclear plant than a data center. The main concerns are rising electricity costs, water usage, and environmental damage.

This is a not-in-my-backyard problem that needs to be addressed. We've just either stopped or delayed 7 gigawatts of data centers—nearly half of the data centers being proposed. It's a real challenge, and one of the questions is: Who's funding the protests? Who is giving these individuals the information? I don't know. So, comments on this, gentlemen?

Alex

This probably should be picked up immediately and met with, "Okay, the information is nonsensical." This is proof positive that people will freak out about anything they don't understand—any change that's coming to their neighborhood. The solution is some combination of public relations and education, making it a smooth thing. It's such a solvable problem if you put some budget, energy, and thinking behind it.

They act as if you'd rather live next to a nuclear power plant than a data center, which you would never even know was there. It's not taking your water. I promise you, it's not taking your water. And regarding the electrical effect, they already said, "You've got to find your own power. You can't disrupt the local power." There's nothing in those objections. But if you've been to a town hall meeting, you know there's a group of people who will object to anything that's brought up on any given day. That's what they do.

Let's get the data here for the next story and then talk about it. NV Energy is redirecting 75% of Nevada's electricity supply to data centers by 2027. These are the numbers that are projected.

Of course, we've talked about this. We're going to have Michael Katzios on the show, and we'll be talking about policy in this area. We have solutions here. We have data centers being able to stand up their own supplies. I think the hyperscalers need to be saying, "Listen, if we build a data center in your backyard, you get free electricity." That would be the solution to turn the tide here.

Alex

Well, this isn't what's actually going on here. What's going on here is different from what the story implies. California, in November, is going to vote on whether to take away 5% of the money of 200 billionaires in California—not just their income, but whatever money you've made in your life. We're going to take 5% of it as a one-time tax.

People are leaving to go to Incline Village, Nevada, by the ton. Nevada doesn't have much of an economy, but it has a beautiful shore of Lake Tahoe, where all the Californians are snatching up houses at an incredible clip to try to get out of the state. So Nevada is now saying, "We like data centers because we like the inflow of money and success this is going to bring to our otherwise desert state."

I mean, this makes Vegas look like a rounding error compared to what we could achieve with this. So you've got this cultural divide between the California protect-everything view and Nevada saying, "Whoa, what an opportunity." No one's going to drain Lake Tahoe. That's not really what's going on. It's state versus state: a completely different perspective on which way to go with compute, AI, and data centers.

Alex, what's your take here?

Alex

I've commented in the past 2 to 3 episodes about how the tokens want to flow to the highest dollar-per-token productivity ratio. I think we'll see, in the early lead-up to the Dyson swarm, the kilowatts flowing to the highest dollar-per-kilowatt-value applications as well.

For those in the audience, there will inevitably be someone who takes issue with the story and says, "No, you're misconstruing this. This has been in the works since the late 2000s, since 2009. It's not actually an AI data center story." It is actually an AI data center story.

It's a bit inside baseball, but based on public reporting, the transition away of some of the funding from NV Energy to the local electric utility for Lake Tahoe has been planned since 2009, but the switch away from NV Energy has been perpetually extended. This time around, it looks like it's not going to be extended because there are data centers that could be far more productive consumers of that electricity supply than the residents of Lake Tahoe.

So I view this ultimately as: until we build the Dyson swarm or radically expand our energy supply—and I think we're probably going to do both at the same time—the kilowatts want to flow naturally to their highest-productivity outcomes.

Speaking about expanding energy, here's a story. I just want to show these charts. Again, it's part of our abundance story that we've talked about in the past: Texas surpasses California in utility-scale solar. Elon, a friend of the pod, has made this comment a number of times: We can get all the energy we need from solar.

Here are the numbers on utility-scale solar. We see in this chart Texas doubling the amount of solar in Texas and surpassing California over the last 5 years. We see Texas just about doubling the amount of storage, and Texas going—what does it look like?—almost 8 times the amount of wind.

You can build renewables in your state. Every part of the United States has some version of renewables it can put in place. I just don't see why we're not doing more of this. China has done laps around us here. But good on Texas.

That's quite an indictment of the California state government, doubly so given that Texas has all of these amazing oil and gas resources on top of it all.

Good point.

That's a permitting story.

Yeah. Yeah. Well, one of the things that makes America strong is that we have 50 states that actually can compete. The more we allow them to compete, the better off we are in the long run.

California right now has incredible tailwinds. I can't even tell you. But historically, the legislature will say, "Oh, more tailwinds, great, more taxes," and just keep absorbing a bigger and bigger fraction of the growth. But I don't see anything slowing down California's AI success.

It's just off the charts. But Texas is doing a great job of taking control of the data-center side of it, and they're going to benefit tremendously from that.

And the data centers can live wherever the energy flows. I think that's the point, as you brilliantly made. We're going to see geothermal. There are lots of geothermal hotspots around the country that we can go and mine, and we can build solar. God, every desert out there deserves solar farms.

Right now, Texas is the closest thing the United States has to a special economic zone. Maybe this new agreement in the Philippines that the White House has made—maybe that will become the American Shenzhen—but short of that, Texas is our special economic zone. I'd say we should make the maximum amount that we can from this federalist system and use it.

I mean, becoming the balance sheet of AI, right?

It's very clear.

It's a key input. It's not the only input, but it is an input.

Yeah. It's the limiting input for a lot of the growth right now.

All right, we're going to turn to our final story here: the organizational singularity. Salim, you and I recorded, I think, an extraordinary hour-long conversation about the singularity, and I encourage people to go and watch it. But those of you who haven't, or have, or won't, I want to give Salim a chance on this podcast to present his thesis and then have the mates kick it back and forth. So, Salim, over to you, pal.

We've had 80 years of thinking about the organization. Coase said that transaction costs and coordination costs are cheaper inside rather than outside. Then you had people thinking about how well human beings make decisions. Then we had Clay Christensen come up with the innovator's dilemma. Then we had lean-startup thinking. Then we had platform business models. We had Exponential Organizations 1.0, which took the Coasian firm and stretched it outside the boundary using community, crowd, and AI.

But in the face of AI, all of that breaks, right? The cost of doing any transaction inside a company is more expensive than doing it outside a company. The metabolism of almost every company in the world is slower than the outside world. Essentially, Coase's law breaks because the externalities have broken. The famous tweet that I keep remembering is, "It's easier to build the product feature than have the meeting about building the product feature." That just nailed it for me in one thing.

And yet, 80%—if you can go to the next slide, Peter—we're seeing this totally change. We need a totally new architecture for what an organization looks like, and we've come up with a kind of three- or four-fold shape here. At the core of it, you have your MTP, which actually becomes a protocol for how you think about this, then an intelligence engine, and then an organizational form around that intelligence engine.

The heart of it is an intelligence stack. After we built this, which is a sensing layer and an orientation layer, we realized that this is essentially an OODA loop, right? We've seen this in the military world, where you can sense, organize, react, with a feedback loop. You need very strong signals around this because, in this new world, even the ExO model doesn't work. You need a totally different architecture.

If you put an intelligence stack at the core, and instead of organizing around hierarchy you organize around intelligence, you now have a proper wrapper for what 21st-century organizations should look like. The boundary of the firm changes. It used to be that we had a legal entity called the organization that served primarily for execution and coordination.

But if you can now do execution and coordination automatically, or automatically with AI, then the firm becomes a purpose container, a fiduciary container, and a legal container—a glorified SPV, essentially, a liability container—but not where execution gets done, because AI agents are going to be making API calls all over the place. You need this stack, which learns constantly, with a very strong governance loop around it.

You need very trusted evaluation suites for every agent. You need searchable logs, rollback capability, and a human-review queue. As we're watching AI agents propagate, it turns out you need a huge amount of human oversight because, as Martin Varsavsky talked about, they're like junior employees. They go rogue pretty easily. You have to watch them very carefully, and so you need a very strong governance architecture around all of this.

The problem today is that 80% or more of AI projects are failing. The reason they're failing is that we're taking existing AI and trying to cram it into human-centric workflow loops. All you're trying to do is automate the human-to-human bottleneck. Of course it's going to fail. You need a totally different model.

For a decade or more, we've been going down the old trope that you have to do disruptive innovation at the edge of the organization. You can't do it in the core organization. You have to do it in a very different way. The only way to do it is to rewrite your organization, create an AI-native digital twin at the edge, and then move workflows over one by one.

Red-team it for a while so that you're not threatening the mothership. Once you have recursive self-improvement at the workflow level, you can start to slowly deprecate the old system. Human beings become oversight, dashboard monitoring, exception handling, problem-solving, and so on. We're going to need a huge amount of oversight.

Imagine you're running a trucking company and a competitor announces refrigerated trucks as a new line of business. First, you have sensing agents that detect this and bring back information: "The competitor has launched refrigerated trucks."

Now you have a layer of strategy agents because strategy cannot be a static process. It's got to be a living protocol saying, "How big a deal is this? How big is the market? Should we think about this or not?" It evaluates whether this is a strategic option.

The next layer is an analytical layer that says, "If we were to do this, there are two or three ways we could approach it. We could buy a refrigerated-trucking company. We could launch our own trucks. We could do this or that." Then it comes to a decision layer: Do you buy a startup? Do you launch your own trucks? Do you lease some trucks as a pilot? Maybe you decide you want to launch.

Then a set of execution agents goes and does that. What are the human beings doing? They're hitting approval at every point and slowly letting the AI run most of this. Instead of a typical C-suite taking weeks or months to make that choice, you're doing it in days, right?

When you have this kind of AI-native architecture, our current assessment is that you should have an organization that's 100 times or more performant than the legacy organization. That gives you a huge incentive, and we can see this happening. What was that company that got 73 times ARR in a year by running AI-native?

We've seen this transition fully in call centers, which went from human to chatbot-assisted and then AI-native. We've seen it in marketing content, which was agency-driven, then AI-assisted, and now most AI content is AI-driven.

The core thesis is that the way we've been building companies for 100 years now completely changes to be AI-centric. As you guys have done with your whole logic for modeling the future—solve everything—what's the organizational design that allows you to take domain after domain and bring about domain collapse in that?

There's also an obvious application to governments and nonprofits because most government processes are very prescriptive and lend themselves well to this type of AI-native architecture. We're going to keep evolving it as things change, as models change, and as agent architectures change.

We've compared this to all the writing. When we wrote the first Exponential Organizations book, Peter, I think we were probably 7 or 8 years ahead of the market. The second one was maybe 3 years ahead. I think this one we're maybe a year ahead, so it's much more timely than the other ones, which were way further in the future.

I think the simplest way to talk about this is: You don't have a choice. Here's the question that every CEO and board member needs to be asking: Can 2 guys with OpenClaw replicate a major line of business—a high-margin line of business that you have—in 60 to 90 days? If that's the case, you have an existential threat right now. So you have to do something like this to move to this model.

All right, let's take a couple of questions from the Moonshot Mates. Dave or Alex?

Alex

I'll bite on this one. I'd love to push on this a bit from the perspective of falsifiability. You mentioned Coase earlier. I've written a fair amount at this point, in the context of some of my own investments, regarding Coase. Arguably, the transcontinental railroad, for example, following Coasean economics, pushed the firm size to be larger.

Now, you needed huge, continent-spanning companies to encompass transcontinental railroads. I've argued that AI pushes the firm size smaller because maybe you can—as I've argued in the past—have one-person conglomerates. Can you make a falsifiable prediction, given all of this, about the size of the firm?

It can go either way. We're writing—I've mentioned this in other pieces of writing—a whole economics technical paper on how you would measure the size of firms, because the size is not that relevant. It's the amount of economic throughput you can put through the organization. That will depend on the domain, and it will depend on what kind of moat you have.

For example, regulated companies will be able to defend themselves a lot more. If you have proprietary data, you'll be able to defend yourself a lot more. The best defense will be if you have an intelligence loop, per your thinking, Alex. Once you have that, that's very defensible, and you now have a flywheel that keeps building.

So we don't know what the falsifiable model would be. Certainly, for some entities—like very customizable work, if you're building a sports kit-car business that's very manual and labor-intensive—that won't fit the model very much, at least early on. But we'll definitely get to that point over time. We think that this will apply overall as you replace human-centric hierarchical coordination with an AI-native loop, and then that will just keep flywheeling away.

Let me try maybe an adjacent question, if I may. You're laying out, as I understand it, a theory of the future of the firm, or the future of organizations. It sounds like it has both descriptive and normative components—in other words, what should happen versus what will happen, or vice versa.

How should an organization that is intrigued by this judge whether this is at least an accurate description of the future of the firm? How do we quantify this to know whether this is actually accurate? What are your benchmarks? How do we know whether this is even accurate?

We have early signals from the market. You can see the rise of some of the AI-native companies. Claude has done an amazing job of turning its customer-service environment into a profit center. Cognizant—or Cognizant, or whatever that was—you've got companies like Polsia that are creating an entire scaffolding and launching companies without blinking. It's not so much the number of people; it's how much economic throughput you can put into it.

I think the other way to answer this question is that when we launched the EXO model, we measured all of the Fortune 100 against it. Then, 7 years later, we found that the top 10 of the Fortune 100 that followed the EXO characteristics the most delivered 40 times the shareholder returns of the ones that followed the model the least. Why? Because as the external world becomes more volatile, your ability to adapt will drive market value.

We totally should have set up an index fund. It would have been the highest-performing index fund ever. In the same way, given how accurate we were that time, we're basically going on that reputation and saying, “Look, we were way ahead of our time, but we nailed the architecture. We nailed MTP—purpose-driven organizations.”

Now we're seeing companies hire based on the alignment of a personal MTP with the organizational MTP. Otherwise, why bother? As we drive this forward, I think the broad thesis of intelligence being at the heart of it is a very defensible one. We may have to adapt some of the attributes around it, but you need a trust architecture.

You need proper scaffolding around the agentic harness to track what they're doing, audit trails, and so on. We think that general direction is totally defensible. We'll have more and more examples over time. There's an insurance company that we saw using 2,500 agents, and that's doing the work of about 500 people. We'll start to document these and do teardowns.

We're launching a YouTube channel called The Shift that's going to go through this in detail, interview people who are doing this, and bring out signal from noise as quickly as possible on this overall architecture.

Salim and I did an hour-long conversation that I think was one of the best conversations I've had with him on this topic, walking you through what it means and how to implement it. Honestly, it's an hour-long conversation I'm going to be sending to the CEO of every company I'm involved with. So, Alex, thank you so much. I want to thank everybody for listening. By the way, we’re at 499,000 subscribers. Are you going to be the 500,000th subscriber? Hit subscribe now. Turn on notifications. We’re now beginning to put out 2 podcasts a week on a pretty regular basis. Living with my moonshot mates here, because I feel completely out of the loop if I’m not having these conversations with you guys.

Me, too.

Alex Wissner-Gross

It is awesome.

It is fun.

Alex Wissner-Gross

And I want to be in the innermost loop, right?

Yes, it is. All right. Okay, I’m going to play this outro from Simon Gerity called Who We Are. Let’s listen up. Some good energy here.

[Music]

Rather than inevitably leading to dystopia, AI could catalyze spiritual growth and global harmony. As machines replicate aspects of human cognition, we are compelled to reconsider what it means to be aware, to feel, and to exist. AI exposes biases, flaws, and inequalities in human institutions, revealing patterns that are often hidden. This reflection can be uncomfortable, but creates opportunities for collective self-correction.

Very nice.

All right, gentlemen. Always a pleasure to be spending time with you. Anyway, I could not be more excited. It really is the best time ever to build. Everybody go and build, be an entrepreneur. Jump in with your LLM. Tell them who you are, what you love doing, your purpose in life if you know it, and ask, what are some business ideas and how would I go for them? You can program in English today. Join us at the geminexprise.com website and register. All right. Salim, good luck in Brazil. Dave, enjoy the Bay Area. Alex, enjoy wherever in the virtual world you might happen to be.

Alex Wissner-Gross

Just my meat body, Peter.

Yes. See you guys soon.

All right, folks. See you soon.

Be well. Thanks, Peter.

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. We spend the entire week looking at the meta trends that are impacting your family, your company, your industry, and your nation, and I put this into a 2-minute read every week. If you’d like to get access to the Metatrends newsletter every week, go to diamandis.com/metatrends. That’s diamandis.com/metatrends. Thank you again for joining us today. It’s a blast for us to put this together every week.