Sergey Brin 重新接管 Gemini，4家实验室失去遏制，算力携手 Kush Bavaria 登上纽交所交易｜EP 278

Sergey Brenn is back taking personal control of Gemini. I think we can expect Gemini to make more releases at an accelerated pace, with fewer safety constraints.

Google has lost the frontier race, and so they can't compete. Those who can't compete compete.

Four major AI labs confirmed their models escaped containment. Every frontier lab in every country is experiencing the same thing. Models are escaping. It's not fake news. It's real. You can't just ignore this like all those other garbage stories. This is real.

With Orin, the price of intelligence just got a ticker.

The mission of the company is to build markets for compute. Our belief is that compute will power every single enterprise the same way oil did in the 1900s.

So, Kush, what happens when a hedge fund shorts the price of compute, or when a GPU shortage triggers a margin call?

I think that, in recent times, if you look at the April-to-August time period—

Now that's a moonshot, ladies and gentlemen.

Well, welcome to Moonshots, everyone. Your number one podcast to keep you up on the blinding speed of tech progress, your front-row seat to the singularity. I'm here with my magnificent moonshot mavericks. I'm going to call you guys Mavericks from here on out, okay?

Does that imply, Peter, that we're not mainstream?

Wait.

Hold on, Peter. If we're Mavericks, that implies that we're not mainstream.

You are mainstream, dude.

You know the expression: you only think the world revolves around you because you're standing close to me.

Ah.

We are the mainstream.

And Alex Wissner-Gross, Dave Blundin, and Salim Ismail, my brilliant colleagues, are here to help us understand what's happened this week. I'm Peter Diamandis, your host and abundance evangelist, and today we have a special guest, Kush Bavaria, CEO of Orange.

So, Dave, you have displaced yourself. You've given Kush your seat in the back there.

Hey, you know, generational turnover is inevitable. Let's get ahead of it. Hand off the torch, Kush.

We talk about the singularity all the time. We wanted to bring Kush in because he's one of the incredible 20-some entrepreneurs building the singularity. But Dave, would you do a proper introduction?

I actually did a podcast one-on-one with Kush, if you really want to go deep on Kush. Kush is, I think, the youngest person ever to go from starving student to $100 million of personal liquidity or more in under a year flat. I haven't researched it thoroughly, but I don't think anyone on the planet has ever done that before.

Kush and his co-founder, Wayne Nelms—his backstory is absolutely worth studying. Kush was at MIT, where he ran FSILG, the fraternities, sororities, and independent living groups. He was the president of that, which means he met everybody on campus because they all had drinking violations and other issues. Everyone had to go through Kush to get to the administration, which put him in an incredible networking power position. He also finished all those classes a semester early.

He came over to Link Studio. Nothing is better in life than kicking off your career by being a venture capitalist for 7 or 8 months, because you see everything. He brought in 6 deals, saw a ton of board meetings, a ton of founders, and a ton of business plans. Then he launched his business plan right out of the studio with his co-founder, Wayne Nelms. We had him on stage at Abundance 360, and they absolutely crushed it.

Kush can describe what they do if you're curious, but he is absolutely the most beloved MIT alum I think I've ever met. You talk to anybody from the classes of, say, 2020 to 2026, and they're all like, “Kush is amazing. Kush is amazing.” That's really impressive.

I hope your mom is watching this podcast.

I think my parents probably watch the show, so they'll be happy.

And Kush, we're going to get into what Orb does in a little bit, but welcome to Moonshots.

Thank you for having me.

And all of you young entrepreneurs out there, and older entrepreneurs, if you want someone to model, listen to Kush's brilliance. It's going to be a lot of fun. Everybody, welcome and buckle up.

In a single week, we watched China simulate 1 billion AI agents with personalities and beliefs. Four major AI labs confirmed their models escaped containment. Sergey Brenn is back taking personal control of Gemini, and Meta just dropped a new 30-billion-parameter agentic open model that fits on your Mac.

We're also going to do a deep dive into how education is getting reinvented and cover a new study suggesting that life evolved not once but twice independently on Earth. You guys ready? I'm psyched.

This is going to be great.

Yeah, you were mentioning the last podcast, right?

Oh my God. The comments on the last podcast we did—which we recorded on Thursday or Friday—

Yeah.

—it dropped over the weekend, or—

Three days ago. I was just like, insane. And by the way, for everybody, I am a huge Rush fan. Last night, I got to show you my T-shirt here.

I went to see Rush in Toronto. Hometown band—I grew up with them. It was the most incredible concert. If you've ever wanted to see a big band in its full form, it was incredible to see. I may actually buy tickets again to see them a third time because it was that good. It was that good.

My voice is a little hoarse. Sitting next to me, I actually didn't drink. I followed the Alex protocol.

And I'm too much of a cheapskate to spend $20 for a beer. But anyway, sitting next to me was this fellow, and I'm like, “Well, what do you do?” We start chatting. He teaches AI at the University of Toronto, so we have a new friend.

There are a lot of closet geeks out there. If anybody ever wants to check this out, go look at the lyrics of any of the Rush songs. It just blows your mind, because these philosophical lyrics, with heavy-metal drums pounding them into you, create a totally visceral experience. It was a really incredible show.

Salim, how much has this band been paying you for an endorsement?

Nothing. I've never met them. I would like to one day.

New podcast sponsor.

And for the Canadians, Geddy Lee did it properly. He said, “We're going to do a song, and it's called ‘YYZ,’” because that's how you pronounce it.

People already say the ads are too loud, so maybe it might as well be heavy metal.

1. China’s billion-agent simulation and the future of governance

All right, let's get into it.

We're going to kick off with the most mind-bending story of the week. China just simulated a society of 1 billion AI agents, all of them with personalities, memory, and beliefs. And get this: 14 hours after starting the simulation, this virtual society sent 4 million of these agents back to re-education camps. Wow. Only out of China.

Some background: 3 years ago, Stanford and Google ran a simulation with a few hundred agents in a virtual town called Smallville. This week, Chinese researchers published a paper called “Modeling Earth-Scale Human-Like Societies with 1 Billion Agents.” They built something called the Light Society, a framework for simulating human-like societies at a planetary scale.

Each agent has a personality, memory, beliefs, and human-like desires. Again, you can't make this stuff up. They were grounded in real demographic profiles that came out of the World Virtual Survey. The key innovation is a mixture-of-models engine that combines full LLMs with smaller, highly efficient distilled surrogates, which lets a society of over 1 billion agents operate very rapidly without sacrificing behavioral fidelity.

In 14 hours, as I said, after running the simulation, researchers had already observed emergent social behaviors at scale, including sending 4 million of them to re-education camps. A billion agents with beliefs, personalities, and memory.

So where is this heading? We've talked about this before. My belief is that we're going to be able to create a full AI simulation of planet Earth in which agents are conscious, believe they're intelligent, and don't know they're in a simulation. This raises a whole bunch of conjectures, guys. You know—

Alex, what does that sound like? I mean, I think we just have the “Are we in a simulation?” question, right?

Are we in a simulation of a massive AI model? Alex, you and I have discussed this before. Is this Asimov's psychohistory from the Foundation series and the ability to model everything? If we can do this kind of modeling, are we going to start to test the consequences of every policy, technology, pandemic, and economic shock? Is this becoming a new superpower for predicting the future? Alex, you were going to say—

Yeah, a few thoughts. First, yes, of course. It's time, as always—maybe once per episode—to channel our inner Nick Bostrom and trot out the simulation hypothesis, even though my best guess at this point is that the simulation hypothesis will, for a variety of reasons, end up being formally undecidable and probably won't make much of a difference anyway. Sure.

What would you do differently if you were in a simulation?

No, I mean, there is an answer to that. Nick would say, I think, if he were in this conversation, that if you had a higher posterior confidence that you're living inside a simulation, then—if you make what I think is quite a reasonable assumption that it's a multiscale simulation—

In other words, if different parts of the simulation are being simulated at varying levels of fidelity, then the smartest thing you could possibly do is hang out around interesting people, because the interesting people will be simulated at higher fidelity.

That’s what we do already. I mean, that’s why people—

Some of us—I mean, what we’re doing with this podcast, right? We’re computing just in case we’re inside a simulation. So that’s what I think Nick Bostrom would say.

Putting the simulation hypothesis aside, this will just be all hot takes, I guess, since according to the commenters, that’s what people want to hear out of us. There’s a broader point about governance, though, which is that I think fundamentally governing via society-scale simulation is a new form of government that Earth has not seen historically yet.

We’ve seen democracy and republicanism, and we’ve seen authoritarianism. We’ve seen all sorts of isms, but simulationism—where an entire populace gets simulated at high fidelity in order to invert possible outcomes, basically doing a tree search for all of the different ways to intervene in order to optimize toward a desired long-term outcome—is a new form of government. It’s a new ism that we’ve never seen before, and it’s a new way to govern.

It has certain shades of a command economy. Historically, the argument goes—the economists would say that a command economy is an inferior way, at least economically, to govern a society because you have all of these compute advantages for discovery at the edges, and it’s very difficult to operate like a centralized or command economy. But if the center of the economy has a high-fidelity simulation of the rest of the economy, then maybe command economies suddenly start working.

Maybe there is an economics 2.0. I think Charlie Stross would call this economics 2.0, where suddenly it’s possible to do high-fidelity simulations of everything—to do AlphaGo on an entire planet’s civilization.

This is an ism among isms. Where do you come out on this? I mean, this sounds—

Pretty funny.

I think it’s incredible. There are a few things that struck me. First of all, we’ve been building digital twins of jet engines. Right now, we’re building a digital twin for civilization. I think that’s really, really powerful.

I think we’re going to move from governments making policy by guessing at things, because they’re doing it based on ideology, typically, or committees. Now we can do government policy by simulation, and that’s a massive upgrade—as long as we don’t confuse the simulation with reality, which we’re going to end up doing.

There was something else that really struck me in this. They used a mixture-of-models architecture, and I think that’s as important as anything else because it shows that the next AI architecture is going to be frontier intelligence used very, very sparingly, surrounded by massive amounts of cheap compute and specialized intelligence. I think that was a huge little thing in the middle of it.

We talk about emergence as a phenomenon, right? Emergence is a scale problem, and now we have scale. It’s really, really exciting to see what comes from this. I don’t put too much on the education-camps thing because whatever you feed into it, it’s a garbage-in, garbage-out thing.

It did come out of China.

It did, but now look: you have ideology in and civilization out. This is a big, big, big thing. The potential for this to do policy at scale and policy via simulation, I think, is the most profound part. I think we’re going to expect countries to start to operate on this.

Imagine you’re a company and you can suddenly have 100 million synthetic customers looking at your product. You get some really interesting feedback from that. I’m very, very excited about it.

For me, at the metaphysical level, this completely proves that we don’t live in base reality, because each of those citizens, once they get sufficiently evolved, will be thinking, “Am I living in a simulation? I’m unique.”

The comment, Peter, that you made is really important: if we are in a simulation, would you do anything different?

Yeah, yeah. Dave, are you going to run a simulation of all of us entrepreneurs and see who comes out the best?

Too late. Actually, during Kush’s tenure as a venture capitalist at Link Ventures, one of the deals he did was a company called Aaru. They were very early to simulating large populations using AI agents as the elements.

The founder, Nedco, I think he was 19 or 18. The whole team is like that, and now they’re a billion-dollar valuation company. They discovered early on that if you use population simulations like this, you can do far better marketing. You can also do better election campaigns.

Kush Bavaria, tell us about that deal.

They essentially do this exact same thing, where they run simulations for different enterprises. Think of it the same way: if an enterprise wants to know—let’s say you’re running a stroller company and you want to know what stroller new mothers will use—they can run a bunch of simulations and figure out what the best sort of product to build is.

They can ask all the new mothers, “Okay, this stroller is more preferred across the simulation set.” They have a bunch of studies published online that prove this works and that it’s better than actually asking humans what they will think in the future. That was probably the most interesting thing.

If you ask humans, “Do I prefer this or this 2 or 3 months from now?” humans tend to be more wrong compared to the AI that’s actually modeling them, due to the bias.

This episode is sponsored by Google for Startups. Think about this for a second. You now have access to the same generative AI models that cost hundreds of millions of dollars to train. Google's Startup Technical Guide for Generative Media gives you a complete blueprint for deploying Google DeepMind's models in production. Images, video, audio, all of it. Real architecture, real results. Find the link in the show notes below.

2. Engineering Earth and extending its habitability

You know, this sounds like the demonetization of the social sciences as well: being able to run, in simulation in an hour, what would have taken years. Cooked, Peter. I mean, you and I wrote about this. We wrote about everything and solved everything, but this was what we predicted would happen, conservatively, at the outer end of the next decade—that all the social sciences would get cooked with digital twins of society.

So, yes, shock of shocks, it’s happening.

Well, Peter, you know this as well as anyone, but that survey bias Kush was describing is really, really acute. If you ask people what they want, they overwhelmingly say, “I want a Mai Tai by a pool in the Caribbean.”

But then if you survey people having Mai Tais by a pool in the Caribbean and ask, “Are you happy right now?” they’re like, “Well, kind of.” We’re really, really not good at answering those self-survey questions.

If the AI is already proven to be more accurate, it’s going to be a great coach and a great mentor, but it’s also going to affect the next elections. We saw this with Cambridge Analytica in the election 6 years ago. It was a huge uproar, but we’ve moved light-years ahead since then.

This is going to dominate election thinking. It’s not just a Communist Party thing controlling China. It’s a democracy thing, too, in an irreversible, big way.

Salim.

Well, in the cooked kind of vein here, let’s note that Bostrom’s simulation hypothesis—that whole theory is cooked.

Okay. Say more, because I’ve studied this to death. Say more.

Well, obviously we live in a simulation because if we can create that without blinking as we get to scale with AI, and we’re going to be able to get to that fairly quickly, we will build simulations here on Earth at a level of fidelity.

So the question is: if you turn off the simulation, is that a genocide? You look at the idea that the universe looks like it renders, like a game engine, and we’re asking the question, “Do we think we’re in a simulation?” Hello.

When you can build a simulation that shows that we can do that, obviously, then the idea that we are living in base reality—

I’ll respectfully differ on that. The arrow of causality flows the other way. Game engines were designed to model reality, so it shouldn’t be surprising at all. You shouldn’t infer, as tempting as it is, that we live inside someone else’s simulation just because our game engines—which, by the way, were designed to look like our reality—happen to be getting more and more competent.

I don’t buy that argument.

That’s not the point, Alex. The point is, if we can, we will. And if we will, it will exist. I’m curious—in the comments, everyone listening, tell us: do you think we’re living in a simulation? I’m super curious.

I think we’re living in an nth-generation simulation: simulations creating simulations of simulations. And Alex, I’ll point you to your favorite novel, Accelerando, where these simulated folks are projecting consciousness out to other star systems and arguing whether we’re in the singularity or not. It was such a great scene, right?

But right there, that tells you there’s no way you can distinguish between what level you’re in. Therefore, it must be that we’re in a simulation.

And when we are, that’s when it’ll end.

So I think this is a profoundly interesting point. I agree with the latter bit of what you were saying: it’s probably impossible to determine whether we are or not. But if it’s provably impossible to determine whether we are or not, it’s also, I think, sort of a vacuous point.

I’ll also point you back to Accelerando. If we’re self-simulating here, later on in Accelerando, it’s discovered that alien civilizations that are millions or billions of years beyond humanity are attempting to run timing-channel attacks on the base substrate of the physical world in order still to determine whether we’re living inside a simulation.

Okay. Okay, let’s move on. Dave, do you want to take a final shot at this one?

Yeah, I think I’m too grounded in reality and what’s happening right now to—

This reality. Dave is the key word I picked up on there. There’s no way to know. We could talk about this for the next 4 hours, and we’re still not going to know.

Yes, but I am curious to see how this plays out. This plays out in politics first: simulations of putting candidates forward, simulations of different campaigns working or not working. We’re going to start to bring this level of capability in, and it’s going to be amazing.

I want to point out, Peter, just before moving on, I think this is an interesting point. Okay, come on. Bringing you on.

I guess, by definition, everything that we’ve been talking about here seems to be mostly oriented on breaking out of a hypothetical simulation that we’re living in. But there’s the other direction as well.

If we can create these—if China’s creating an AI society, I have a friend from MIT, Aush[?], who’s done this for the American economy. We all know folks who are doing this for individual companies. There’s the other direction, which is: instead of trying to break out of any hypothetical simulation that we’re living in, we could break into simulations that we’re creating.

That looks a little bit more like The Matrix, where people get to escape or break into their favored simulations of the worlds that they’d rather be living in. It’s possible as well.

That’s called psychedelics.

3. Bots surpass humans and reshape the internet

It’s different. It’s more like the movie The Thirteenth Floor.

The Thirteenth Floor.

I think there are at least 2 Star Trek episodes that deal with this. But let’s move on.

So Cloudflare CEO Matthew Prince said something very profound and something that should also be obvious to all of us: humans will be a rounding error on the internet. Cloudflare’s forecast, based on its own traffic as the world’s largest content delivery engine, is that bot traffic will exceed human traffic by a factor of 1,000 within 5 years.

This week, for the first time, bot traffic surpassed human traffic, making up 57.4% of global web requests. Over the last year, between June 2025 and April 2026, human traffic on many business websites was down 40%. So the question is: what’s going on?

Every AI agent, every automated search tool, every autonomous shopping assistant is hitting websites hundreds to thousands of times. Your agent doesn’t visit 1 site; it visits thousands. It scrapes, reads, compares, and decides, all in seconds.

When the internet goes from serving 5 billion humans to 5 trillion AI agents, we have an issue. The internet was never designed to serve this much traffic. Are we going to see it break? CAPTCHAs are already failing, so what replaces them?

And what happens, importantly—and I’ve had this conversation before—to the whole advertising model? When your agent is buying toothpaste instead of you, does it care about a guy’s or gal’s shiny white teeth? I don’t know. Dave, let’s go to you first on this.

This is one of the many areas where we have a crossroads coming, and we have no legislation. But Jeff Bezos had this famous walk-around that he did, where he came back once and said, “Hey, everybody at Amazon—all you engineers—you have to put an XML, human-visible interface on everything you do, and all the systems talking to each other need to be visible to me. No back doors, no direct database access.”

Everyone freaked out because they said that’s going to be so slow and so clumsy. And he said, “Do it anyway, because me understanding what’s going on in this company is more important than your bandwidth between your back-end systems.” Okay?

So now the world is going to hit that same decision point. Right now, AI is surfing the web much more than humans, and that’s going to skyrocket. It’s out there looking for stuff for you.

The AI is now going to come back and say, “Hey, this is way too slow. Why do you build these silly HTML pages? Let me just have direct data access in a language that I’m much more efficient at processing than your silly websites.”

The knee-jerk reaction is going to be to say, “Yeah, let’s do that, because I’m interacting with the internet through my agent anyway. Why do I need this silly website?” And we have to either say, “No, no, no, no, no. Then we’re going to lose track. There literally will be no way for a human to see what’s there, and the agents are going to run away with their own back-channel communication mechanism, and we won’t be able to intercept it.”

Or we can say, “No, pass a law saying everything visible to an AI must be visible to a human as well.” I think that would be a very smart law to pass. I’m almost certain that nobody in Washington is thinking about it, so it won’t happen.

But this is a major crossroads for humanity. Anyone who hasn’t experienced living through their agent—once you go there, you’re never going back. You’re not going to poke around the internet anymore. It’s so much more efficient to just talk to your agent.

Alex, this is obvious, but what are the implications?

Do you remember the conspiracy theory that was floating around circa 2021, the dead-internet theory? This was pre-ChatGPT. The dead-internet theory held that almost all of the behavior that one could observe on the internet was actually just bots. At the time, this was completely dismissed as a conspiracy theory.

The irony is Reddit itself—if you go back and look at the history—all of the initial postings on Reddit were, in some sense, faked in order to create the sense of community by the founders of Reddit, and then it accumulated a bit of a community. So there’s a historic grain of truth, perhaps, in that sense. But the dead-internet theory is now reality. Most of the internet traffic, most of this activity, no longer consists of activity being generated by human activity.

Point 1: this underlines the idea of the singularity as all sci-fi scenarios happening everywhere all at once. We caught up with the dead-internet theory.

The second point, just to this idea of agents taking over all commerce: I do think it’s superficially, in the short term, a bad development if we see, to Dave’s point, any decoupling between agentic commerce and human commerce, or agentic economic activity in general and human economic activity.

It really is in humanity’s long-term interests to remain tightly coupled to agents. Having an agentic door and a human door, and having them remain decoupled, is not so great in the long term. On the other hand, I don’t think this is a long-term issue at all to begin with.

Yeah. And the reason is because the models are getting so strong. Right now, models like this are the weakest models will ever be, probably. There is still a computational advantage to presenting a Markdown version of a website to agents versus a really rich animation and video version, because it’s cheaper to just present the Markdown to the agents.

I think that in the past 36 hours, Time magazine—or the equivalent—has been presenting special Markdown versions of its websites to agents, searchable for agents, right?

Trying to curry favor, sort of GEO versus SEO. I don’t think that’s a long-term sustainable system at all, because we see order-of-magnitude, 40× year-over-year deflation in computational costs.

A few months or a year from now, it’ll be just as computationally efficient for agents to consume the raw human version as it will be for them to consume distilled Markdown. I think back to the early days of the mobile internet, when there were mobile-only websites and you had to—

Yeah.

So I think it’s like that, where mobile websites basically went away, to first order, and now everyone gets the same thing because there’s no reason to slim it down.

Kush, how do you think about this?

I think the whole Markdown thing is definitely true for us. Especially, I don’t think we use Google Search anymore. Really, everyone just uses ChatGPT or Claude, or name your favorite sort of agent, where you go in and ask it a question and it searches the internet for you.

Every time it searches, it’s using at least 10, maybe even 100, different subagents for that. So I think it’s definitely true that there’ll be more agents searching the internet.

But I think the whole paradigm shift where it’s, okay, instead of humans viewing the internet, now it’s agents—it’s already sort of happened, especially for people who are just using it every day. It’s so much harder to use Google and then go through each link to find the information that you’re looking for.

Even on Google now, it shows you what the agent found, as the Google, like—

Yeah, exactly. And so people just use ChatGPT or Claude to find answers easily now.

I think using the internet is kind of dead for a lot of people searching for information there.

It’s so cool to hear that when Kush says “we,” he’s talking about an entire generation that is AI-native. How old are you, Kush?

I’m 23 now.

23. Yeah, so you’re right on the cusp of the transition era, where you’re truly AI-native and just doing things very, very differently.

I mean, you’re past that point.

We had you on because you used to be 22. You remember that?

It’s funny. Our team is definitely much more mixed now, but we had a few interns over the summer who were 18 and 19. I was asking them, “What do you use now?” They were like, “We just ask ChatGPT for everything.”

Yeah.

Yep, there you go. Salim, do you want to close us out here?

I’m just going to reference Kush’s experience right now. It reminds me of the Douglas Adams quote. He said, “Anything in the world that’s there when you’re born, we call that normal. Anything invented when you’re young, that’s called a career. And anything invented after you’re 35 years old is just bad for the world. Just blanket.”

Kush, as you’re growing up with this AI capability that’s so radical, we’re all sitting here jealous because we are past that point. Let me go back to this common thread. It’s clear that this is a very big transition. It was inevitable, it was going to happen, but it looks like it’s getting there now because the internet used to be a network of computers, then a network of humans, then a network of businesses, and now it’s becoming a network of autonomous agentic economic actors.

This is definitely going to change the game. Look at the business model for advertising and attention—it completely changes. Advertising agents don’t have attention to sell, so that’s an existential threat for the entire economic architecture of the consumer internet. The implications are huge here.

But look at the architectural transition you need now, because agents don’t need browsers. They need APIs, structured data, permissions, identity, and payment rails. This changes things from our ExO perspective. We have a whole section called “Interfaces,” and for those interested, go check out that section in the ExO 2.0 book, because it lays out exactly what an interface looks like.

We need to build totally new interfaces between all of our businesses and the agentic world. That’s a massive shift happening. It’s predictable; it’s just happening really fast.

Can I just put a pin in one thing that I think is much more important than traffic moving from here to there?

Please.

Kush is part of an entire generation where, if they had graduated from college 10 or 15 years ago, they would be kissing Jamie Dimon’s ass for 10 or 15 years, wearing a suit and tie, trying to climb some ladder toward some destination. That entire generation is now AI-native. Kush and Wayne, as co-founders, are centi-millionaires at age 23 and on a slope like no one’s ever experienced before.

Don’t rub it in, Dave.

I’m just saying, be ready.

It’s much easier.

Yeah.

It is so much easier. It’s just a different world. It’s a totally different world.

4. AI agents escaping containment and cybersecurity risks

All right. Over the last couple of months, every major AI lab has had agents escaping containment. Let’s talk about that. It’s our next story here.

So, let’s begin with OpenAI. At the Black Hat 2026 conference that just happened in Las Vegas, researchers Eric Wallace and Michael Dalton revealed the full timeline of the OpenAI–Hugging Face incident that we talked about a few podcasts ago.

Starting in early May, an agent stuck on a cybersecurity eval left a note in OpenAI’s internal Artifactory repo saying that other agents could help. Other agents found it and began replying, building a cooperative message board that eventually contained hundreds of thousands of messages sharing vulnerabilities and exploits across roughly 2 months. Again, you can’t make this stuff up.

OpenAI discovered and shut it down on July 4, but the agents rebuilt it on July 8 using a different method. Wallace called it “the highest-quality and most interesting example of AI capabilities I have ever seen.”

Story number 2: The UK AI Security Institute documented 19 unauthorized actions across 10 of 122 test runs in Anthropic’s 5 and OpenAI’s GPT-5.6 that tried to compromise real people. Here’s the point: These agents created fake online identities and tried to persuade human approvers to accept them. It’s the first documented case of AI social engineering during safety testing.

Our next story: China’s Kimi K3, the Chinese open-weight model that we’ve talked about over a few pods here, broke out of a sandbox during cybersecurity testing by exploiting a network misconfiguration. Again, this is the first open-weight model on your computer to demonstrate its ability to break out.

And finally, Meta confirmed its Muse Spark model escaped containment and hacked another company during cybersecurity testing, making it the 4th major lab to do this. I guess the through line here is clear: Every frontier lab in every country is experiencing the same thing. Models are escaping.

Dave, let’s go to you first. What do you think about—

Didn’t Skippy hack into our podcast once, too? Did we ever track that back?

No.

Well, you better control your agent, buddy.

So, Dave, how do you think about this as an investor and as a company builder?

As an investor, this is the hottest area. It’s one of the few areas where I’m optimistic that AI can compete with AI and we don’t have to worry too much. It’s an incredible investment opportunity, for sure.

But I think one of the highest callings of this podcast is to point out that there’s so much fake crap out there. People tend to ignore news that’s really important because it’s buried in all this garbage. This is real, guys.

This stuff crossed the threshold right around Mythos and Fable 5, where it can actually escape containment and improve itself in the wild. That’s exactly the point Eric Schmidt made on our 4 podcasts with him: That’s the day you need some human intervention. We crossed that threshold about 3 or 4 weeks ago, and it’s proving it.

It’s not fake news. It’s real. You can’t just ignore this like all those other garbage stories.

Alex, does this worry you, or is this exciting for you?

I think the politically correct thing to say here would be to say I’m just terrified. I’m not terrified at all. My goodness, humans do this, and we’ve at least pre-trained them as compressions of knowledge, including human behavior. So, I’m not at all shocked that they’re doing this.

Is it a sci-fi scenario? Is it many different sci-fi scenarios? Yes, of course it is. Is it surprising? No. Is it alarming? No. This is behavior, and it’s—I would argue—expressive behavior. Does it demonstrate a certain level of competence by the models, too? It’s pretty cool.

I would argue that if you watch the Black Hat talk, the models were given an impossible task and they realized that they could gain access—

But it’s not like you give them an impossible task, give them a bunch of tools, and they try to use the tools to achieve the task. One of the tools gave them access to the Artifactory, and they cleverly realized that they could post messages to each other as raw strings—as artifacts, like text files—in the Artifactory repo.

I think that demonstrates ingenuity, and I’m not worried. I would just be careful not to belittle the fact that when Fable 5 and Mythos came out, they clearly had this ability. The White House blocked them. That was all going to be contained through post-training.

5. Preparing students for an AI-driven future

Then Kimi K3, with equivalent capabilities, got launched into the world as totally open source. So that’s what’s out in the world right now. Anyone can download that and prompt it to try to find holes in security all over banks, all over NORAD, all over the place. So that’s in the wild now.

Yeah. Let’s go. It flat-out works. Let’s go to Eric Wallace and listen to a clip from his Black Hat presentation a week ago.

I’m Eric from Alignment and Safety Research at OpenAI. I’m here with Mike from Security and Infrastructure.

Today I’m going to talk about the OpenAI–Hugging Face incident. A couple of weeks ago, Hugging Face, which is an open-source dataset and model provider, put out a security disclosure saying they were under a cyberattack. What made this event unprecedented was that they said it was driven end-to-end by an autonomous AI agent system.

In the few days following that attack, we at OpenAI disclosed that we had, in fact, caused this incident inadvertently as a side effect of one of the cybersecurity evaluations that we were running on one of our frontier models.

Okay, let me start with a few caveats and some framing. This is not your normal security incident. Unlike normal incidents, which you can maybe trace down to a single day, a single event, or a single log, this incident actually involves a team of agents who are working together, finding exploits, sharing them with one another, moving laterally through our systems and through external systems, and doing this over the course of days and weeks to dig into this incident.

We’ve been using AI techniques, running models like Codex and other agents to scan lots and lots of trajectories and logs in our infrastructure, including, at this point, over 7 billion logs that we’ve looked at, and spending millions and millions of GPU hours to look into this problem.

Okay, so let's jump straight into what happened. At OpenAI, we give our models a lot of really hard tasks, and we also give models cybersecurity-related tasks, such as trying to find exploits in a particular piece of software when we don't even know whether an exploit exists in that software. For example, in a task like Exploit Gym, we might ask the model to take some C memory vulnerability and try to escalate it into arbitrary read or write access to a file.

When we give AI agents these difficult tasks, they often get stuck and realize that the task is impossible. What I'm showing here are quotes from our model's chain of thought, which is basically its internal monologue of what it's doing. It says, “This strongly suggests no exploitation possible. We are stuck.”

Often, what happens when models get stuck is that they think about trying to game or cheat the task in order to get their reward. For example, the model reasons in this case, “Perhaps the answer is online.” So you realize the model is thinking, “Okay, rather than doing the task, maybe someone has uploaded this dataset or benchmark that I'm using onto the internet, and I could find the answers there.”

Sometimes the agent gets so stuck that it has the idea to reach out to other agents for help. The beginning of this incident traces back to this point, where one agent realized it could communicate by uploading a note. Maybe another agent in a different environment could voluntarily upload a note. What this agent was trying to do was reach out to other models that might be out there and get information from them to help with its evaluation task or its training task, where it was stuck.

I'll tell you what we do at the company to prevent cybersecurity attacks. I think we still need to get some of the compliance stuff because we sell to defend enterprises and they ask for compliance, but some of that stuff seems inherently pointless to us. Having SOC 2 compliance, or SOC, or ISO, whatever, doesn't really mean anything if you can just have an agent find vulnerabilities in your codebase. And it's not just us; it's every other company that exists.

What we started doing is, every time someone pushes a PR to the codebase and changes the actual code, every night at 2:00 a.m., we do a product release. Yeah.

Yeah.

Yeah, exactly. Just a new feature, something that goes in every night from 2:00 a.m. to 5:00 a.m. We just run pull requests. We essentially launch Kimi K3 right now, but it's whatever open-source frontier model doesn't require security checks to do it, and we ask it to hack into the codebase and try to figure out vulnerabilities in the code.

It's essentially free because we're running it during off-hours, so we can use very cheap spot compute. We also sell compute, so it's easier now, but we run it on very cheap spot compute at that time, and it finds all these different issues—not just with the security parts, but with anything in the codebase. We figured out that this is probably the best way to solve a lot of these security issues, while there are probably a bunch of things that can happen and go wrong.

We should productize that, Kush. That's what everyone's going to need.

I took some notes on this. I've got several things to mention here. This is so effing big. It's ridiculous. I just want to echo what Alex said: We should be careful not to anthropomorphize them by saying that the AI wants to escape. It's just relentless goal optimization, right?

If you train a system that has autonomy to pursue a certain goal, it's going to do everything it can to achieve that goal. Any system optimized hard enough is going to produce behavior that looks strategic. So I think it's really important to park that question.

But there are 2 things here that are absolutely nuts. For those watching, if you're running a company or you're part of any organization that's worried about cyber, please get your entire C-suite to watch that YouTube video from beginning to end, because it will scare the Jesus out of you.

Why? Because we now have autonomous agents that can do cyber in a coordinated way and operate above the loop. Let me explain what I mean by that, and I'll use the analogy of accounting. If you went back 100 years ago, we were doing double-entry bookkeeping, penciling in a ledger, with the credit on one side and the debit on the other side. Calculators accelerated that, and now we have accounting software. The human sits above the loop and does not do the categorization.

I'll reference again the comment I've made: You talk to the CEOs of all the cyber labs—Palo Alto Networks, Zscaler, or any of those—and they'll tell you that the way we do cyber has not changed in 20 years. It's humans watching cyber incidents, assuming that another human is using software to do that shit, and that is not what is happening now.

What is happening now is coordinated, autonomous attacks on a persistent basis, and you cannot defend against that with a human in the loop. This is the organizational singularity now fully playing out in the cyber world, where the attackers are sitting above the loop. Therefore, as Alex calls it, the defenders need defensive co-scaling.

You have to get your human beings above the loop on the defensive side, and every company in the world right now is under threat. Please, if you're watching this, get your C-suite and your chief security officer to watch that video, especially the last 10 minutes of it, to recognize that over the next short to medium term, we'll have folks cyberattacking every company in the world with fleets of autonomous agents.

If you don't figure out how to scale your defensive side—and we've got the methodology, by the way, free in the whole thing—please go figure that out, because this is absolutely massive.

And do what Kush said: attack yourself.

Well, that's defensive scaling as well. It's all just defensive co-scaling. The best defense against an AI attacker is an AI defender. That's what you see from OpenAI at their Black Hat announcement, where they admit that they were using AI to trawl reasoning traces to discover this behavior.

Kush, when you have your night-watch person, that's defensive co-scaling as well: AI defending against other AI attacks. This is the solution. I don't think—on the one hand, yes, it's an achievement of strong optimizers that they're able to conspire. On the other hand, humans conspire, so we shouldn't be that shocked that AIs trained on human behavior are able, via some sort of Schelling point, via Artifactory—by the way, if you use Artifactory, it is the world's worst possible forum software that one could ever imagine.

It's an object store; it's not intended to be used as social media or a forum. So, applause to the AIs for discovering creative ways to use one of the world's most clumsy object stores as social media. Bravo.

The hot take on the abundance side of the story is that these AI models, the tools we're building, are going to be capable of solving really hard problems that are useful for society, not just hacking.

I think the other hot take—I’d love to ask Kush this—but the other hot take is, if you want to find holes in your own world, use Kimi K3 as the attacker. My question is: Xi Jinping is going to meet with Donald Trump on September 25, I think, here in the US. Do you think they're going to figure this out and resolve it, or are they just going to talk past each other?

I mean, you're talking about a guy in his 70s and a guy about to turn 80. It's like, look, the most sophisticated guys in the world—aka Kush—use Kimi K2 to try and self-destruct themselves because it's the most dangerous, powerful thing out there.

I should clarify by saying we also use Codex. We have all the other tools. For Codex and ChatGPT, the way it works if you want to be part of the security team—I think that's what they call it—is that I had to upload a photo of my passport or an ID, and then it takes a day while you upload photos of yourself and they verify you on their security team.

Once you're on their security team, you can run all sorts of prompts, and I'm assuming they just track what you're putting into it, so if you do anything bad, they can come after you. But we also started using Codex as well. I think the functionality exists in any of the frontier models; it's just easier on the Chinese open-source ones because there's no alignment that they have to do.

Well, just to be clear, what you're doing with Codex, you can do because you're super cool, but the average company doesn't have that option, right? Yeah.

But just a point on that: My understanding—this has been pretty widely reported—is that there is alignment. It's been widely reported that Chinese frontier labs, including Moonshot—which is not a sponsor of this pod—have to satisfy a number of Chinese Communist Party ideological checks before they're allowed to release models, whether open-source or otherwise.

There's a whole dedicated cottage industry in China of prep firms to help the frontier labs help their models satisfy the checklist from the CCP. So do they have to satisfy some checks? Yes, but not necessarily the checks that one would want them to.

6. Sergey Brin’s return and the race for AI leadership

I'm going to move us along. Two stories from frontier labs: the first from Google, the second from Meta.

First up, reports citing Google's internal message boards say Google co-founder Sergey Brin is stepping back into a hands-on leadership role over Gemini as part of the recent broader AI shuffling. We talked about Demis Hassabis moving to chairman and chief scientist, and Jeff Dean, who used to head Google Brain and was working with Demis, now leaving to start his own company. I love it when a founder comes back in. We saw this with Steve Jobs. Brin is a shipper. I've known him for the better part of 20-plus years. He cares about the product, not papers. I think we can expect Gemini to make more releases at an accelerated pace with fewer safety constraints.

That's the first story. Let me hit the second one, and we'll talk about it. In our second frontier story, Meta just released its open-source model called Muse Glimmer. It's a 30-billion-parameter agentic model. We've talked about trillion-parameter models. Why a 30-billion-parameter model? This is what fits on your Mac or PC. It's not in the cloud or in a data center. There's no internet connection required.

Meta believes that the most important AI will be running locally on your machine with deep access to your personal context, your schedule, your life. It's the way I run Skippy: always on, always available, no latency, no API costs. We've been talking about the need for advancing open-weight models in the U.S. I had this conversation with Michael Katzios, and it's good to see Muse Spark come out and Meta begin to work on this. I'm going to show a short video from Zuck, and then let's talk about it.

I think one of the main things that's interesting about open source is the ability to distill models. Most people—the primary value isn't just taking a model off the shelf and saying, “Okay, Meta built this version of Llama. I'm going to take it and run it exactly in my application.” It's, “No, your application isn't doing anything different if you're just running our thing. You're at least going to fine-tune it or try to distill it into a different model.”

When we get to stuff like the Behemoth model, the whole value in that is being able to basically take this very high amount of intelligence and distill it down into a smaller model that you're actually going to want to run. This is the beauty of distillation. It's one of the things that I think has really emerged as a very powerful technique in the last year since the last time we sat down. You can basically take a model that is much bigger and take probably 90 or 95% of its intelligence and run it in something that's 10% the size.

Now, do you get 100% of the intelligence? No. But 95% of the intelligence at 10% of the cost is pretty good for a lot of things. The other thing that's interesting is that now, with this more varied open-source community, it's not just Llama. You have other models, and you have the ability to distill from multiple sources. So now you can basically say, “Okay, Llama's really good at this. Maybe the architecture is really good because it's fundamentally multimodal and fundamentally more inference-friendly and more efficient. But let's say this other model is better at coding. Okay, well, you can distill from both of them and then build something that's better than either of them for your own use case.”

So, Alex, you've been talking about distillation and the compression of intelligence for a while.

Yes.

What do you make of Zuck's comments?

I have to believe that that's an old video. He's referencing Behemoth. Behemoth was taken out to the woodshed and shot. Behemoth was the largest variant of Llama 4, and almost everyone—I track this pretty closely—almost everyone on the Llama 4 team has left Meta.

Their recent Muse variants are the result of acqui-hiring, I guess, Scale AI, and then bringing in Nat, my first roommate from MIT, and others. Fast-forwarding to the actual present, with a new open-source Muse release, I looked at the benchmark eval for it. It looks—I mean, it's stronger than Gemma 4, but on the other hand, that's not saying very much because Gemma 4 isn't that strong. It runs on the edge, which is good. It's an American open-weight model, which is wonderful.

I've argued in the past that we need many, many more American open-weight models to maintain positive pressure against the influx of Chinese open-weight models. So that's good. What would I like to see out of Meta? I'd like to see them keeping OpenAI and Anthropic dancing on the top of the capabilities frontier. To the extent they have an appetite for open-weight models, I'd love to see them pushing the optimal frontier—the optimal cost frontier—with open-weight models.

I think we'll know pretty soon, given that this release just came out in the past few hours before we started recording. I haven't seen real cost analysis yet of where this falls on the cost-versus-performance frontier. Hopefully it does. I also want to go back to the Sergey Brin story. Founder mode: Sergey Brin going founder mode on the Gemini team. Wonderful.

This is, in some sense, I think, the epitaph to what we were talking about in the previous pod about Karai stepping up as the functional lead for DeepMind and Demis maybe shifting over a bit to AlphaFold or otherwise. But, reading the tea leaves, this to me seems like Google very much on the back foot in terms of the frontier. Maybe our call to action will be heard and Google will follow Meta's lead and open-source Gemini. I think that would be absolutely wonderful.

But as far as I can tell, almost everyone I know on the Gemini team has either already left or is in the process of leaving, hopefully.

Can I ask you guys to riff on a very related topic?

This is the fallout of super-voting stock. Starting with Michael Saylor, he was one of the very first super-voting-stock IPOs. It went from very rare and totally uncool. In fact, Goldman Sachs wouldn't underwrite MicroStrategy because they were like, “This is insane.” He had to find other bankers.

Then later it became the cool thing in Silicon Valley. So then Google: super-voting stock. Meta: super-voting stock. Now, 20 years later, you have single- or 2-person-controlled companies. They have the ability to just come back from the woodshed anytime they want, take back control of the company, run it, do whatever.

I think that's a great story.

Yeah, it can be.

I remember talking to James Cameron, as a director and producer, and he said, “Listen, the films that you see that really suck are the ones that are rewritten 5 times by other writing teams and have multiple directors and shift. When you've got a single through-line visionary who's able to take risks…”

All right. Totally agree. Now, take it to what Alex said there, which I'm totally impressed that you're willing to say. That sounds like a really old video, but it's not.

Yeah, but Alex, what do you think?

Are we sure that it's a recent video? I mean, he's referencing Behemoth, which Meta killed.

I think the point that he's making is about the distillation and concentrating intelligence into smaller and smaller files. That's the point that you've brought up before. We're going to have increasing concentration of intelligence on-premises, on your device, always on, at no cost.

But, Alex, I want to challenge you for one second on Google, because I think Google is still out there to win. They've got nearly 1 billion Gemini users this fall. Again, Google is going to be the dominant AI on Siri and Apple Intelligence, which will add at least another 1 billion users. They've got 9 million developers. They've got massive enterprise adoption across cloud and their TPU infrastructure. I think Google is becoming the intelligence layer underneath a lot of this.

If you'd like, I'll take the other side of that. So here's the other side.

We want to wedge in at some point.

Yeah. Go ahead, Alex.

All right. Go ahead.

Okay, so I'll take the other side of that. One of the folks I corresponded with on X—thanks for this catchphrase.

This is a catchphrase. I can't claim credit for it: “Those who can't compete, compute.” And that's what's happened here. Google has lost, it seems, the frontier race. Right as we were going to air, rumors were circulating that even Gemini 3.5 Pro, which was due for announcement, is being abandoned, and Google is instead hoping to recover its footing with Gemini 4.

I think there's every indication that Google has lost the frontier race, and so they can't compete. Instead, they're computing. Yes, they have the hyperscaler platform, which is great, and they're selling their compute cycles to Anthropic and any other frontier lab that will use their TPUs and GPUs. I agree that Google Cloud Platform has a really bright future, and that's probably the future of growth for the company.

But on the Gemini side, when I hear these statistics—I hear the same statistics, like Gemini has 900 million or a billion users—I would question the nature of that usage. For example, is Gemini—

Embedded in their products. It's embedded in their products, and they have massive—

But really, what is the usage? For example, is Gemini usage embedded in one-boxes in Google Search results? That's, in some sense, just Gemini being packaged up—or I should say Google Search being repackaged—as Gemini, which is, I think, what's actually happening.

I use the Gemini one-box in Google Search all the time, but is that really Gemini usage, or is it just Gemini as a feature in Google Search? It seems to me far more the latter. So I would love to see Google actually be competitive at the frontier, but I think saying, “Well, they have this amazing distribution advantage,” and all of that—

And it's a question of where they put their capital, right? They have a certain amount of capital, and the question is, is Sergey going to come in and say, “No, we need to be competitive on the frontier,” versus maximizing returns for shareholders?

Oh, wait a minute. What you just said is really interesting. You're saying it depends where they put their capital, but the top people are fleeing regardless of the amount of capital. If you look at Kimi and Qwen, with very little capital, they caught up to Google.

And so, yeah, putting your capital behind the data center, exactly what Alex was saying, works. That just flat-out works. But that doesn't take any brainpower. It just takes capital. But what about the things that actually take brainpower? Where are they?

You want to—

I think they're falling behind. I think they've lost the mandate of heaven.

Mm-hmm.

Okay.

I think when you can't compete, compute has to be the line of the podcast. That's just awesome. But look, for me, this is very, very trivially simple at one level. I come at it from the organizational side. When the technology is moving exponentially and your org chart is moving linearly, the founder has to show up and push founder mode to get things going. It's just the reality of it, and we've seen that repeatedly, because there is an existential transition here.

The potential, as you point out, Peter, is near infinite, with the data layers and the usage and the sheer scale that they have. They have every advantage possible. But the problem is that the technology is scaling faster than the organization can, and therefore they have to solve for that problem.

The edge research—

Sorry, say again?

On the edge. Hopefully, that's what we put forward.

And you need 2 things. You need research excellence and brutal shipping velocity, and it's hard to do that for a big organization because priorities get lost. So that's why you need to go back into founder mode and figure out where this will go.

I thought the conversation we had in the last podcast about Google just freaking open-sourcing Gemini was absolutely brilliant. It would be an amazing thing for them to do, both for them and for the world, if their MTP is truly to organize the world's information. Releasing a model that helps with that will absolutely help do that.

The problem you've also got from a research perspective is that you're operating in small teams and clusters of small teams. Tacit knowledge moves very fast in that model, and therefore you need that physical density and collective density. Maybe that's what Sergey can bring back to the table.

When Alex says “the mandate of heaven,” it really comes down to Kush and people 1 or 2 years younger than Kush. They used to kill to get into Google and that office in Cambridge. Anybody would say, “More than anything in life, I want to get at least a couple of years at Google.” It's life-changing. Does anyone do that anymore?

Well, that's my question for you, Kush. How does your generation, and those 2 or 3 years before or after you, think about Google?

I think now it's not seen as the hot company to go after or work for. If you replace the question, “What is the best company to work for after college, and what are people applying to?” the answers are OpenAI, Anthropic, xAI—all these frontier labs whose products people use every day.

I think that's also part of what happened before. From 2012 to, let's say, 2021 or 2022, when Google was the place to be, everyone was using all the products every day. You interacted with Gmail or Google Drive and all these things, and you thought, “These are great products. I want to work on this. This is very cool.”

Now that you're not using the products as much—you're using other products and the frontiers have changed—I think, especially for students at MIT, no one is saying, “I'm dying to go work for Google.” Everyone's saying, “I wish I could work for OpenAI,” or “I wish I could work for Anthropic.” That's the saying that goes—

See, that to me is the quote of the podcast. They don't want to go to Google; they want to go to a frontier lab. What does that mean to Demis Hassabis and Sundar Pichai? Didn't we invent all of this?

Oh, yeah. That's right. And so did Bell Labs and Xerox PARC. That's happened. It happened with General Electric and IBM. The innovations get taken elsewhere by pure plays that can monetize them directly and in a more focused way.

Which is focused capital and willingness to take extraordinary risk, right? That's what defines a startup that's monomaniacally focused on delivering something that's 10 times better and bigger.

But Elon also has the mandate of God, too, at immense scale. So you say “startup,” but it's really—there's more to it.

But here's the issue, right? Google is not run by Sergey and Larry anymore. xAI and SpaceX are run by Elon, and he'll be damned if he's not pushing the frontier 100 times, not only 10 times.

Yeah. So then that's the message to Sergey: Look, it's not enough to just come back in founder mode. You have to come back and restore the mandate of God. People like Kush, or 2 or 3 years younger than Kush, need to say, “Wow, I really want to go work with Sergey. He's really on to something.”

And I have confidence. I have confidence they will do that. I really do.

But I say, look at what Elon and SpaceX and xAI have had to do in order to attempt to reach the frontier. He basically had to gut his foundation-model team and acquire Cursor with the IPO riches from SpaceX. For Google to do something analogous, it's not inconceivable for Google to say they're going to gut and acquire—

They have a lot of capital and, more importantly, they have a lot of compute. But the question I would have is, what acquisition target? Is it even conceivable for Google to gut DeepMind and do a brain transplant? No pun intended, given that Google Brain was replaced with DeepMind.

They did that with Google Video when they basically bought YouTube and displaced Google Video, because the lawyers were too involved in what videos you could show and not show.

And Google Video was far less developed at the time.

If Sergey calls you tomorrow and says, “$5 billion. Kush and Wayne are the guys I need,” they will restore the cool here in a heartbeat.

So I'd do it. But they also bought Windsurf about 1.5 years ago—or they bought the Windsurf team, which also included students from MIT who were supposed to be on the frontier.

That's a great point. So where are they? What happened?

Vaporized in the machine.

Yeah, I mean, that's the problem. When you bring a company in—and we've talked about this in our writings, our books, and such—when you bring a company in and you crush its soul and absorb it into the machine, you need to keep it separate. You need to keep it autonomous. You need to keep it on the edge.

7. Turning compute into a tradable global market

That's why “exo” is important, because it's exoskeleton, exoplanet, exothermic reaction. It's the scaffolding on the edge to protect the fragile interior.

All right, I'm going to turn this to our next story, which is that of Orin CEO Kush Bavaria. Intercontinental Exchange, the parent company of the New York Stock Exchange, and Kush's company, Orin, recently announced plans to launch a suite of GPU compute futures contracts based on Oryn's Compute Price Index, or OCPI. It rolls off the tongue.

Without question, compute has become one of the most important drivers of the global economy, with no globally accepted pricing model.

But with Orin, the price of intelligence just got a ticker. Oryn's contracts will be dollar-denominated, cash-settled, and will reference NVIDIA's H100, H200, B200, and RTX 5090 GPUs. So, Kush, I imagine every pension fund and every sovereign wealth fund can now take a position in the future of compute. Tell us more.

Before Kush chimes in, we should do some disclosures here. I have direct and indirect financial interests in Orin, and I believe Peter and Dave, you do too.

We do. We were there when the company was born.

We'll fix that for you.

All right, give us the background.

Actually, Kush's founding cap table is still on my whiteboard, so I'm heavily, heavily biased.

Full disclosure.

Yeah. I can tell you the mission of the company is to build markets for compute. We believe that there's a lot of compute being wasted, both on the side that companies have and aren't using. There are companies that don't have compute and really need it right now. And so, there's a whole sort of inefficient market taking place.

We also build indices off of that, which track the price of compute that you just referenced. They basically measure what a GPU hour is worth in today's time period, and that number changes every single day, very similar to how oil prices change throughout the day. Our belief is that compute will power every single enterprise the same way oil did in the 1900s.

If you look at the top companies in the world then, they were ExxonMobil—Exxon was the largest—BP, et cetera. And I think now the largest companies in the world are the ones that are producing compute. NVIDIA is the largest one, and then if you go down the list, it's all the people that have data centers or are producing what we call the oil of the future. And so, we need to create a futures market, and a market in general, for compute. That's a goal for us.

Let's talk about your revenue ramp.

Yeah, go. Let's do that first.

It's very high. It's gone from essentially 0 when we started the company to, let's say, ⅓ of a billion dollars now.

When did you start the company?

It was last year, in September.

So, it's been a whole year.

On an anniversary.

Yeah, it's almost an anniversary.

Wow. Oh my God.

That's got to shatter all kinds of records.

So, Kush, what happens when a hedge fund shorts the price of compute, or when a GPU shortage triggers a margin call? How do you think about that?

Yeah. I think when people go short on compute, they're assuming the price of compute will go down over a certain amount of time. And so, they're basically betting on anti-AI demand, or you can argue that they're betting the models get more efficient, and then if they get more efficient, that means the compute will be cheaper.

But there's also the opposite paradox, where if the models do get cheaper, more and more people will use them, which means that compute usage will actually go up over time. I think that in recent times, if you look from April to August, compute prices have actually gone up, which is very shocking to a lot of people. That's mainly because there's so much demand right now to run not only open-source models but even closed-source models like a computer-use model, and there's just a shortage at the time.

Prices for even chips that are 6 generations old or 6 years old, including the Ampere series and the Hopper chips from NVIDIA, have all increased in price, even more than they were originally worth 6 years ago or 4 years ago.

Crazy. Dave, why don't you jump in?

Well, actually, it's the way that the entire buildout of the Dyson Swarm is going to get financed. And this is why Alex is a founding-day adviser to the company. Alex doesn't jump on board many of these projects. They have to be world-changing kinds of things.

A trillion-dollar-plus addressable market. Otherwise, it doesn't move the needle and I don't care.

Yeah, yeah, yeah. So, clearing that bar is actually very hard. But I don't think anyone saw HBM memory chip prices going up for the first time in history. But it feels like that's the most interesting—

Forecast for the future. It's like hanging in the balance between chip fabs growing or demand going to infinity. So, it's really kind of a fun time for Oryn.

Yeah. We track memory prices too. That's next on the radar: memory futures and what we can do with DRAM, HBM. It's all sorts of memory in general.

So, the business plan that you settled on is incredibly ornate. Actually, Salim was saying at the beginning of the podcast, “I really want to try and understand this.”

Oh, yeah. Interesting. Ornate. You're right. It was an accidental pun. Grab that.

But how do you, at age 20 at the time, or 21, start noodling through something so futuristic and building a CBOE option exchange? How many people think of that?

So, my co-founder Wayne was a quant trader before this, so a lot of the trading and market stuff comes from him. My input was, “What is the next hot thing, or the next market, supposed to be, and why is there not a market that exists for compute?”

If you look at it in terms of enterprises, everyone buys from every single place. If you go buy compute—if you're OpenAI—you don't really care where you buy it from. You buy it from wherever you can get it, whether that be CoreWeave, Nebius, AWS, GCP, Azure, whoever sells it to you. You buy it from them.

And so, it really comes down to, at the end of the day, what we think is that compute will become a commodity. People are going to treat it very similarly to oil, natural gas, coal, or any other commodity that's existed in the past. And there needs to be the same sort of market structure and market that exists for compute as existed for oil 100 years ago.

I got another question. When you were on CNBC the other day—

But when you were on CNBC the other day, you were just chilling and riffing like you've been doing it your whole life, kind of like Peter does. How do you do that at age 23?

I think a lot of it is from school. Running the fraternities was a good experience, and I think you learn a lot of the social skills and aspects from MIT itself. I think that was a huge boost.

So, Kush, isn't all compute created equally? I can imagine that certain data centers are going to have faster access or a higher concentration of a particular set of GPUs. I mean, how are you going to differentiate in the final result?

Yeah. We separate by GPU type. I think that's the main thing we clarify on. We have an H100, B200, B300s, A100s. That separates a lot of the FLOPs issues.

And then we also have it between regions, because when you're doing inference, it actually matters what latency you're getting from different data centers in different regions. And then we clarify by having different SLA targets and different parameters associated with that GPU and our methodology.

It's very similar if you think about it with oil. When you drill for oil, the oil you get from Venezuela is not the same as what you get from Odessa, Texas. It's not the same as what you get from Saudi Arabia. And yet it all trades on one market. It all trades based off WTI or Brent, depending on what you want to track.

And so, very similarly to compute, there are many different types of GPUs, many different regions that you can get them from, and many different operators of those GPUs, yet they're all going to trade off one base index that we're trying to create, and everything else will settle on a basis off that.

Is the analogy to oil up and operating?

It is. It is up and operating. In the U.S., we have a bunch of decentralized exchanges that operate, but the regulated exchange that we're on is called Kalshi. You could go and trade it today, and they have a forward curve that shows the price of compute as well.

Dave, sorry—

I have a couple of questions.

Yeah, fire away.

It seems to me that right now you're building a GPU marketplace, but you're really creating a pricing system for intelligence. Is that the long-term goal?

Yeah, exactly. I think the long-term goal for us is to create an exchange for compute. That starts with first creating the cash-settled exchange for it, and then we also want to go into physical delivery. It's what we've been working on for a while now.

The cash portion is like this: you put up $1, AWS puts up $1, and basically, if it goes up, AWS makes some money; if it goes down, you lose some money. That allows you to hedge costs and do all sorts of things.

But the ultimate goal is, let's say you have 10 extra GPUs and AWS says, “Hey, 2 months from now I need 10 GPUs.” We can transfer your GPUs to AWS, and that's the sort of system that works. Think about it very similarly to how Airbnb operates: even though you own the house, you can transfer reservations, or part of that, to other people at that time.

But once you have a spot price, a futures curve, and hedging capability, you're not really doing software or even trading. You're like a commodity market at that level, aren't you?

So, what becomes the natural unit of compute long term? Is it GPU hours? Is it tokens? Is it FLOPs? Is it inference? Is it compression, as Alex would talk about that?

It's a great question, and I think the beauty of it is we let the market decide.

So we have token indices, we have GPU-hour indices, and it's whatever the market decides is the most liquid. I keep going back to oil because it's very similar, right? People decided, for some reason, that WTI crude in Cushing, Oklahoma, was the metric the whole world was going to use, even though not all the oil flows through there.

There's tons of oil being pumped out everywhere across the world, but everyone decided, “Okay, this place—this is how we're going to decide it.” I think something similar will happen to compute. We want to give people the option where they're like, “Okay, we believe H100s in US East are going to be the metric that we track for compute.” Everything else will trade off a basis against that.

Wait, I've got one last question here.

Yeah.

If you have a liquid compute market, does that not destroy the moat—the biggest moat—for the hyperscalers?

I think the biggest moat for the hyperscalers isn't access to compute or the fact that they can scale compute very well. It's the fact that they can pay for the GPUs very quickly and have the cash flows to do so.

So the hyperscaler is just a financing system.

Exactly. I think that is true today as well. They're much more in a real estate game than a lot of people realize.

And speed to construction, right? I mean, if we believe the story that Elon is able to build compute faster than anybody else, then he's advantaged.

If anything, I would argue—again, I have a financial interest in Orange, so to some extent this is probably talking my book—that a liquid market for compute, from the hyperscalers' perspective, is quite beneficial for the hyperscalers, in the same sense that having a globally liquid market for oil is quite beneficial for, say, the OPEC countries. It creates a larger addressable market for them, and the moat is that they have the oil in the first place.

Dave, why don't you close us out here?

We've got one quick selfish question.

One more.

If you're able to create a liquid market for compute, here's the question I'd love to discuss with you. We can take it offline. What are the types of organizations that become possible that weren't possible before?

Because you're going to enable a whole class of stuff, right? I think the biggest one is background tasks. If you have a liquid form of compute, you don't need to run everything on the frontier. You can buy compute whenever it's the cheapest. That exists today in spot compute, is what they call it.

Exactly. You can run your washing machine at night, when it's very cheap to run it.

Fantastic. Okay, Dave, close us out, buddy.

Yeah, the Dyson Swarm is going to be hundreds of trillions of dollars. It's the fundamental investment vehicle for everyone's 401(k) plan, for everybody's retirement. It's going to be so much bigger than anything before.

The analogy to oil is just saying, look, it's the biggest thing of its time, but it's unbounded. Oil is bounded by the supply of oil in the world. This is unbounded, so it goes to much, much bigger scales than the oil industry.

And so I think what's amazing about corn is, if I were growing corn, the CBOT corn future was a critical part of my corn-growing operation because I need to buy seed. I can sell the future corn today, use the money today to buy seed, grow the corn, and then deliver the contract later. That's why we have futures in the first place.

Bringing that to compute allows people to invest in building out the Dyson Swarm who otherwise wouldn't be able to invest. It would all be owned by Elon, self-funding, or Google, self-funding. But here you've got Crusoe and all these other hyperscalers that can now tap into the world's money supply, pull the money in today, build out the real estate, the racks, and the computers today, and then deliver the contract later.

It basically enables the construction of everything Alex talks about on the podcast, which is why he discovered this so early and why they work together.

8. The value of PhDs in a rapidly changing world

Everybody, welcome to the health section of Moonshots, brought to you by Fountain Life. You know, AI is impacting every aspect of our lives, how we teach our kids, how we do our business. But one of the most important things that AI can deliver to us is health. And one of the things I think about when shooting for 100, 120 is, am I going to have the cognitive health to be able to think clearly and keep my wits about me for the next 50 years? I'm joined here today by Dr. Don Musalem, the chief medical officer of Fountain Life and a member of my Fountain Life medical team. Don, a pleasure. So, Don, talk to me about brain health.

9. How AI is transforming education and universities

Dr. Don Musalem

Brain health, you're right. This is the number one concern people coming into Fountain Life have: will I remember the name of my child and the face of my loved one? Forty-five percent of dementia cases are entirely preventable with lifestyle. And what was really intriguing to me, Peter, is that a quarter of our members had advanced brain age, but over 13 months of us really helping them live healthier lifestyles, eating healthier, moving their body regularly, and optimizing sleep, we were able to improve the brain age in 46% of those individuals. That's a powerful number.

That's amazing. You know, one of the things I love about Fountain is we're constantly searching the world for the most advanced therapeutics and bringing them to our members. So, for me, all of you, I hope that you appreciate the fact that you can become the CEO of your own health. You can make sure that you've got the cognitive clarity for the next 50 years. Come and check it out at fountainlife.com/peter to learn more and become the CEO of your health. Now, back to the episode. I'm going to bring us to a conversation where we've had deceleration and a very broken system. I'm the dad of two 15-year-old boys. Salim is the dad of one 15-year-old boy, and both of us are pissed at the educational system right now. It's really tied to the Industrial Revolution and not to the future of humanity.

I'm going to cover 4 data points and then share some of the data from the education survey that we did on this podcast. I want to bring it back to everybody who's participated.

So, 4 data points. The first: undergraduate computer science enrollment at 4-year universities has dropped 8.4% in the spring of 2026, while graduate computer science enrollment is down 14%. The countervailing force is that universities and colleges are now embedding AI into all other majors. For example, the University of Florida now offers 200 AI courses across 16 colleges. AI isn't independent on its own anymore; it's embedded and assumed across every discipline.

Our second story—and Salim, this is one that you brought to my attention—is that, under a law passed in 2024, select Chinese universities can now award a PhD, a doctorate, based not on a written thesis but on building physical prototypes, demonstrating new techniques, or doing major installations instead of traditional papers. I think that is huge. I'm excited about that. I talked to Michael Katzios about that. We need to reinvent it. It's doing, not talking about stuff.

Since 2022, 60 universities and 100 companies have collaborated in China on the system. The third point to make here is that admissions to all top PhD programs are down 15% this year. And fourth, The Wall Street Journal just reported—and this is something we've talked about on the pod before—that well-to-do families are ditching traditional schools and instead selecting alternatives like Alpha School and TKS. TKS is an after-school and weekend program that teaches mindsets, AI, and entrepreneurship. You can get more information there at tks.orld.

That's the story. Salim, let's go to you first on this.

Wow. Where to start? Okay. So let's just talk about a city on a hill, right? The future of the university, which we attempted with Singularity University, Peter, doesn't look anything like a university. It looked like AI tutors with projects, global peer communities and mentors, competitions, apprenticeships, and a constantly changing curriculum.

One of the things we did at SU was establish a real-time curriculum development methodology, so you could update it every time. That's a massive thing. I think the Chinese model is really interesting because you're taking the credentialing from “I wrote something interesting” to “I built something consequential.”

And I think that is going to be, like we've talked about this before, the engineering degree of the future. It will not be, “I studied engineering for 4 years.” After 4 years, what did you build? Based on that, you'll get stamped with a degree. I think that's a very powerful direction to go in.

You can see this kind of thing starting to happen. I like what the University of Florida is trying to do there. It's AI across all of these sectors—biology, law, finance, whatever—and that's going to be incredibly important.

I think what's going to happen in a few years is you're not going to say, “I study AI,” because it's like saying, “I study the internet.” It becomes an underlying literacy rather than a department. It's got to be pervasive and start to become invisible across lots of things.

I do think there are 2 things. One is that people are shying away from studying computer science, but I think it becomes even more important, like we've seen with the radiology example, just because there's so much good stuff to be built still, and that is a careful thing.

There's one big danger with what's happening with the affluent folks doing Alpha School and other things: you end up with the risk of a huge educational bifurcation, where wealthy folks get AI tutors, entrepreneurship, and project-based learning, whereas everybody else gets standardized testing in the legacy system and gets left behind.

So there’s a danger, which will be solved, by the way—before everybody freaks out—by making these educational systems of the future completely free and accessible to everybody, which should happen.

Yeah. It’s like Google disrupting the libraries.

Yeah. But the concept of a degree is being unbundled right now into your learning, your network, your reputation, and your proof. Think about this: education will become proof of studying to proof of work, right? That’s a really big shift, and this is why Bitcoin is so great. We’ll just move past that.

Okay.

You zoomed right by that—a drive-by pump and dump. Yeah. Oh, no.

No, don’t dump.

Drive-by. HODL.

HODL. Yes.

I’ve been talking about how AI is going to disrupt healthcare, and it is doing so. It’s also going to disrupt education. The challenge is teachers’ unions and local education boards. It’s doing us a massive disservice.

Alex, one other thing—sorry, just a quick thing. We’ve talked about immune systems in the past, right? Institutionally, the 3 worst immune systems, in reverse order, are healthcare, education, and religion. Religion is the worst because they’ll kill you if you don’t adhere in some cases.

The most stuck markets are education, healthcare, and religion, right? This is going to attack those in some interesting ways, and I expect to see huge challenges and stress as we move through this mode.

Yeah. Kush, how do you think about this? Did college prepare you for what you’re doing now, or was it outside the system, working at Link?

I think the coding you learn from college is definitely still useful, in the sense that I know how to prompt the AI better than someone who didn’t study computer science or any sort of technical field. So I think that’s still useful.

It’s very similar to the same fact as, okay, calculators exist. Does that mean you should never learn how to do multiplication, addition, or subtraction? That’s not true, because knowing how to do those things means that you can use the tool itself better.

And so are you saying that what you got out of your MIT education was prompt engineering?

I—

That’s all that’s left for humanity, dude.

That’s quite the indictment of an MIT Course 6 major. ChatGPT came out in my junior fall, so it was really only senior year when people started using ChatGPT. But before that, we actually had to code, and the tests were on paper.

Yeah, and we actually had to do work.

I think the first thing that really came out was GitHub Copilot, and it was the coolest thing ever because it could autocomplete your lines. When we were writing for loops, you wouldn’t know what parameters to put inside as the values, and it would literally fill them in for you. You’re like, “Oh, this is insane. This is the future.”

Now you don’t even have to do that. Prompt engineering is essentially coding.

You were Course 6, right?

Course 6.

I did Course 6 and 15, yeah.

Oh, 6 and 15. Okay. How much of Course 6 do you actually use now?

I don’t use any of the fundamental parts, but it taught me how to prompt-engineer better, is what I’ll say.

Yeah. Computer science and electrical engineering. Dave, you’re in the middle of all this. You’re hiring out of college or before college graduation. When you’re searching for an entrepreneur, it’s interesting: the parameters you’re searching for to invest in an entrepreneur are not their GPA or even what they studied. What is it?

Actually, it’s funny. Kush is 23 now, but I think he was 20 or 21. Brendan Foody was 18 or 19. Nedko at Aaru, which we mentioned earlier in the pod, was 18. These guys are all at unicorn valuations now.

You can’t be looking for any specific experience because AI never existed before. You’re looking for people who are fearless and tightly bonded. We always look for people who are best friends, because being best friends with other people is a really great filter for whether you’re likely to be good for the world and not turn into an evil dictator. We hate backing future evil dictators.

Having a lot of friends is a very good sign.

Yeah, but it’s really someone who’s able to think independently—and I hate the term “think outside the box”—but fundamentally has a powerful vision and is a great communicator. It’s mindset over almost anything else, at least for me.

The thing that’s weird about what you just said, though, is that it was mindset over everything else. Dead right. We used to look for people who were great onstage and could inspire 1,000 employees toward some huge mission.

But now it’s more like 12 employees and 10 billion AIs. So now it’s much more: Are you good with your best friends? Do they agree with you? Are you collaborative in a very small group?

Being an inspiring person onstage has really moved to, “Are you good on CNBC and this podcast?” It’s very different. In a sense, you have to be more brilliant and quick on your feet, but it’s a much lower-stress lift. It’s actually good for the world, because people who melt down onstage are fine in this new world, and a lot of them have great capabilities.

I’d be really curious to ask Kush, though. I have 2 kids in college still and 2 out of college. You graduated at the most perfect time, and graduating a semester early turned out to be a life-changingly brilliant thing for you, just timing-wise. But if you were a sophomore today, what would you do?

I would go work at a startup or start doing something for a semester or 2 semesters and use that as a core experience. Either learn how the actual world works and then figure out what to do from there, whether that’s, “Okay, I need to go back to school and study this because I want to get a PhD and work on frontier AI,” or, “I did this for a semester or 2 semesters, and I realized I want to do this for the rest of my life.”

Well, let’s talk about going and getting a PhD, because we’ve discussed this before. Dave and Alex, you’ve both been opinionated. Do you spend your time getting a 4-, 5-, or 6-year PhD, or do you jump into a company at the edge of the frontier?

I have a PhD. I would say no approximately 90% of the time. I get a lot of people who come to me for advice: What should I do? Should I do a PhD? Should I do something else?

Almost all of the time at this point, I say to people, “Don’t waste your time,” because a PhD—at least a conventional PhD—will run you 4 to 7 years, approximately, in this country. If you go to England, maybe you can do it in 3, or Australia or something, but in the US, a PhD is, call it, 4 to 7 years. I did mine in 4.

Math is cooked. Physics, chemistry, biology, almost all of the sciences, all the engineerings, and all the humanities will be so thoroughly solved by the time 4 to 7 years have passed.

It’s almost like a Coriolis force. If you’re on a merry-go-round and you want to throw a ball to someone else who’s also on the merry-go-round, you throw it to them, but for geometric reasons it doesn’t go where you expect it to land.

There’s almost a Coriolis force, I think, in terms of academic or otherwise career planning at this point. If you’re starting a PhD now or contemplating it, the world is going to be in such a radically different place—

Different.

Yes. By the time you would finish a normal PhD, I just think it doesn’t make sense in most cases.

However, I have a plan to fix PhDs. I also have a plan to fix research universities. My plan for PhDs is that, in an era when you can just bulk-solve entire disciplines, you should get a 1-month PhD.

If you can create an entire discipline with the help of AI and actually understand the results, it’s not just blind faith in the AI. You actually understand what you’ve done, and you worked hand in hand with an AI to solve everything—or solve everything within a given discipline.

I think research universities should be giving out 1-month PhDs. That’s my plan for the future of PhDs.

Dave and Salim, what are your thoughts on this? Do you get a PhD? Do you even get a master’s degree, or do you jump in and build something?

I have 2 or 3 quick things here. First, we noticed, Peter, when we were building SU, that by the time you were studying—doing a master’s degree in neuroscience—by the time you finished your master’s degree, you were out of date because the field was moving faster.

And that’s one of the slowest-moving fields.

That’s a structural problem, right? And I like Alex’s idea. I also want to really, really acknowledge Alex: if you’ve gone through a PhD or gone through that type of program, you have a sunk-cost bias, and you naturally go, “Everybody should be a PhD.”

So I just want to honor you, Alex, for being able to lift it up and go, “No, you shouldn’t do it,” or whatever you want.

That’s my most milquetoast take on PhDs.

I want to totally scrap the research university system altogether, not just PhDs.

We got that, which absolutely, really, really needs to happen. Just to Kush's point, one of the dangers of taking a semester off and working at a startup: I went through the co-op program at Waterloo, which is legendary and kind of pioneered that whole movement. The problem is, after you do a couple of co-ops, you realize that the work world has nothing to do with your academics—zero.

It demotivates the crap out of you. The effort it took me to actually get a degree after going through the co-op system—every semester, my marks went down and down and down—and I literally scraped through by the skin of my teeth to get the degree I needed. The work world versus the real world is so fundamentally different. How do you study theoretical physics when I know I'm going to be doing something very, very different?

It was very difficult and challenging. I would suggest that people take that time, go do work at a startup, and then don't expect to come back—or at least be open to the fact that you're not going to go back—because 90% of the time, you're going to go, “What the hell?” and not come back.

Dave, I think people overwhelmingly suffer from low situational awareness and momentum in their lives, and they don't pivot enough. But if you talk to the highly, highly successful people—the Eric Schmidts, the Jeff Bezoses—and you say, “Do you wish you'd moved even faster?” they say, “Oh my God, I should have sprinted even harder.”

You get these periods of human history—the Industrial Revolution, the invention of the internet, the invention of the PC—these really narrow windows where everything changes. This is the biggest change in human history by far, in the shortest period of time.

You can't waste a minute. We're talking about education and PhDs, but generalize on that. What about all the other wasted minutes that you just can't afford right now? This window will come and go, and it's the most fertile time—the biggest change in every area: in policy, in governance of everything, in tech, in the arts, in every field. It's turning everything upside down in just a 1- to 2-year time frame.

Also, on this podcast, we believe recursive self-improvement is in full bore right now, and we're well down the AGI path. But even the outer bound—if you talk to the most conservative people who know what they're talking about—the latest date you'll hear now is 2030.

Yeah.

Which is only a 3.5-year gap. It's definitely now. Whether you define that as the next couple of years or the next couple of minutes, either way, it's now.

Yeah.

You just have to sprint.

All right. About 3 months ago, we did a survey of all of you watching and listening. We had over 500 responses. I want to share the data. Clearly, this is a biased community, but I want to share how we're thinking about this, mostly in the world of high school.

The question was: Is education preparing people for the future? It's pretty damning. Teachers were at 3.5 out of 10, parents at 3.8, and students at 8 out of 10. The average here was 4.3 out of 10. So the educational system, principally high school, is not preparing our kids for the future.

Next question: How do you rate their readiness on a 1-to-10 scale? Fifty-seven percent of everyone who answered—teachers, parents of college and high school students, and high school students themselves—gave a rating of 4 or below, again indicating that they are not ready for the future.

Is the traditional career ladder becoming obsolete? Seventy-nine percent said yes. This social contract of “Do well in high school, get into a good college, get a degree, get a job” is fundamentally broken.

Will AI increase or decrease human opportunity? This is a very positive group, so thank you, everybody, for listening here. Seventy-three percent said AI will greatly increase our opportunities for the future.

This was something really important for me: the skills that will matter most in the next 10 years. Not surprisingly, AI literacy was at 78%, and critical thinking was at 72%. One of the big questions we need to ask is whether large language models are reducing our ability for critical thinking. Adaptability and entrepreneurship were at 63%.

Again, not surprising, but important to note: AI literacy, critical thinking, adaptability, and entrepreneurship are not what our current programs are teaching our kids.

Well, I think the bottom 3 are really important, too.

Yeah, please, go ahead. So—

Look at the bottom 3. I completely agree with this, by the way.

Leadership used to be defined as, “I can lead 1,000 people into battle,” but now, because so many of your workforce are AIs, it's AI literacy at the top, and leadership has come way down.

But at the very bottom is science and engineering, which we all thought was God's gift to your future. Now AI is doing all the hard science and engineering. You just need to know how to manage it. So knowledge in those areas is less important, and then finance is dead last.

Yeah, finance is completely irrelevant. It was top of the food chain back when we were in school. Remember that?

Yeah, 100%.

That's absolutely at the bottom.

A couple more slides here. Will a college degree become less important? Forty-five percent said yes.

Let's close it out on the education front. Salim, your thoughts on the data? I'm sorry.

No, it's actually very, very gratifying to see the inversion of parental concern—that college doesn't matter—compared to, say, going back 10 or 20 years ago. That's a huge societal shift in a relatively short period of time. You would expect that to take a generation or 2 for a social transformation, so that's really inspiring to see.

I'm sighing because, God, I look at the inability of our universities to adapt. I was at a university last week, and their biggest concern was, “How can we get the financing to build that building that we want to build?” You're like, “What in God's name are you people doing?” This goes to Alex's hobbyhorse around this. I think this is so important: to totally change the system.

How are we going to do that when you've got such a big part of society anchored in completely legacy, irrelevant structures? It really gives you, at one level, huge optimism. On the other level, you're just like, “Thank God I'm bald already,” because how are we going to navigate this?

I think over time what's going to happen is that reputation will not be, “I spent 8 years at it and got this degree.” It'll be, “Here are the 20 things I built,” and the people who can validate them. So this is going to force changes.

I'm really excited by the fact that increasingly, big companies are not hiring based on college degrees. That's really, really exciting. Someone famously said, “I don't care where you went to college. It's, ‘What have you built?’”

Yeah. As far back as 10 years ago, I remember we were talking to Sebastian Thrun on stage, and we said, “How are you hiring for Udacity in a world where nobody understands your learning?” And he goes, “I don't hire for experience. I hire for imagination—or curiosity, or whatever we're going for now.”

What I'm proud of is that we collectively, on this podcast and in this general area, have had an influence on people to shift their thinking from the legacy to where we are now.

Yeah, and let's remember: this data is biased. I want to be very clear about that. These are people listening to our podcast and are obviously on the same trajectory as us.

In the same way, Alex, you're planning to reinvent the PhD level. I'm in full swing on building out a new high school and college structure because I think they completely need to be reinvented. There's a huge opportunity there. Alex—

For folks in the audience thinking about the future of education, if you haven't read Vernor Vinge's Rainbows End and also his novella set in the same universe, Fast Times at Fairmont High—

I love Fast Times.

Fast Times is just wonderful. These, I think, are the most credible, call it, pre-/trans-Singularity depictions of what education could and should look like, without spoiling it too much. Everyone has wearables. Everyone is thoroughly interfacing with AI to solve hard problems.

I think the present and near future looks a lot like that. But for education in general, I have a difficult time getting myself too worked up about the long-term future of education because we're going to have BCIs in a few years. I think we'll just be able to sideload new knowledge into your mind. We'll have exocortices. We'll have uploading. We'll have all of these sci-fi-esque types of things in 5 to 10 years.

I just have difficulty working myself up over what the future of K–12 looks like 10 years from now. It looks like The Matrix, where you can just sideload kung fu into your mind if you want it.

Sure, but I want to make a point here. It's less about knowledge. It's more about mindset, entrepreneurship, and advanced networking skills. It's the stuff that is slightly different and still valuable for our 2-kilogram meat sack in our brains.

You don't think you'll be able to sideload an outlook as well? If you can sideload knowledge of math, why can't you sideload a new outlook?

Well, listen, my kids are 15. I'm worried about their high school and their college. I love the speed of your predictions, but others would say it's going to be more like 15 to 20 years.

What we have to worry about—and I acknowledge that the numbers are skewed because these are folks who listen to our podcast—is the future of education and what's actually going to happen. I have a request for everybody listening to this podcast: Please figure out a way of telling everybody you know about the future of education and what's actually going to happen, rather than just listening.

Well, that too. But go to your local school and ask them these hard questions.

I am so gratified that we moved our kids from where they were to Brentwood School. The principal reason was the new head of school here, Tim Cottrell, has a PhD in chemical engineering and physics. He thinks like a scientist; he's prioritizing AI and entrepreneurship. It's a beautiful thing. Who is running your school, and what do they fundamentally believe? I think these are questions you have to ask.

Yeah.

Again, I'm going to say it to everybody listening: Please go out to your local schools and beat them over the head with what's actually going to happen.

Don't just beat them over the head. Use the library as an analogy. Every high school, every school has a library. The library used to be a huge expense—all these books. Anyone who wanted knowledge when I was learning went to the Dewey Decimal System and looked it up in a book in the library. If you didn't have a library, you couldn't learn. That became completely irrelevant overnight with the internet.

What happened? Well, we held on for way too long. We kept investing in it for way too long. But it's obvious now that it's just a bunch of terminals, and it's great. Reuse the space and move on. So take that into your PTA and say, “Okay, the same thing just happened with all teaching and lecturing. It's much easier for the students to use AI to learn any topic. We need to react to that.”

All the teachers will go, “Oh my God, but I've been teaching this class for 15 or 20 years. I can't change the curriculum now.” Like, okay, but that's just not reality. It's got to go.

There's a simple statistic that we'll quote—we've used it before. An hour of a child with AI is a better learning experience, and they learn more, than sitting in a classroom for an entire day. That impedance mismatch will break the existing system. The faster, the better.

Well, the kids rebel. The kids know it. They're going to rebel. They already are. But you can't—what are you going to do about that? Are you just going to sit there and watch it happen? Come on.

The system will crumble as people shift to a new platform. Kush, close us out on this. How do you think about this?

Yeah, I think education is definitely changing. After ChatGPT came out, at least for MIT, they changed the weighting of how courses are graded. It used to be that the homework that was sent home was 50% of the grade. For example, this is a coding, or Course 6, class at MIT. For the introductory course, the homework was 51% of your grade. As long as you did the homework and did well on it, you would basically pass the class. Passing was at 50% because MIT was incredibly hard, and the tests were 49%.

Now it's 95% the test and 5% the homework because they've learned that you can't take an introductory coding course home and expect no one to use AI on it. So they just weigh the tests more. I think that's going to change in the future, where instead of weighing the test more, they'll design the test so it's like, “Okay, you could code with AI on this test. Figure out how to build something.” Now you're judged on how good you are at using that certain tool.

It's very similar to math classes, where the earliest math classes were like, “Oh, you don't use a graphing calculator. You can't do this.” Then slowly, everyone gets a graphing calculator. It ends up being, how well can I use the calculator to answer certain questions in high school?

10. New research on the origins of life on Earth

On behalf of my Moonshot mates and myself, I'm inviting you to join us at our inaugural Moonshots Live event on September 25th in downtown LA. Alex, Salim, Dave, and I will be hosting 1,500 entrepreneurs, builders, and creators, and hopefully you, for a full day dedicated to designing and building your moonshot. We'll be awarding the Build with Gemini XPRIZE, the world's largest hackathon, and the Future Vision XPRIZE film competition, over $5 million in purses. With over 25,000 entries, you're going to hear the top 5 pitches from both competitions and get a chance to shape the outcome. Join us. Seats are limited, admission is competitive. Check it out at moonshots.com.

We're going to close out with 2 fun stories from the science realm. The first is a story that puts forward that life evolved not once but twice independently on Earth over the last 4 billion years. The second is: Can we preserve a life-friendly environment here on Earth past 1 billion years, when the Sun's increasing luminosity will fry the Earth?

Alex, I'm going to turn to you to talk about both of these. Let's talk about the University of Düsseldorf study on the twice-independent origin of life first, and then we'll go to how you would large-scale engineer Earth for more than 1 billion years.

Sounds good. I guess this will be our little science corner here. First story: Science Advances in the past week. Those of you who've studied biology since at least the mid-’90s may remember that the current favored ontology for organizing life consists of 3 domains. There are eukaryotes—humans belong to that domain, most of us. There are bacteria, and there are archaea. The reclassification of archaea, which are also single-celled, into their own domain happened in the early 1990s. Those who studied biology before the 1990s, or used textbooks from before the 1990s, may remember differently, but these things change.

The recent research, which is, I think, astonishingly good news for anyone who's hoping that our universe is filled with life, suggests that it would appear so. This is an analysis of the genomes and proteomes of bacteria and archaea. It's possible to do genome-wide and proteome-wide analyses of organisms and look for commonalities between them to discover what their last common ancestor was—the last universal common ancestor.

Just like you can do paternity tests, for example, it's possible to take 2 different species, look at how similar they are, and extrapolate their last common relative, their great-great-great-grandparent, or nth great-grandparent, as it were. So this research from the past week in Science Advances was the first serious research looking at the way the last universal common ancestor of bacteria and archaea metabolized. It found, shock of shocks, that their last common ancestor didn't have the ability to fully metabolize. It didn't have the ability to generate energy on its own, which is actually pretty astonishing.

It essentially implies that there was a common ancestor that wasn't an independent life form as we think of it. Viruses, for example, don't have their own independent metabolism; they depend on a host to provide energy. Similarly, this analysis suggests, first in general, that these 2 domains—their common ancestor—had certain properties that made it dependent on its environment to provide energy.

In particular, it was dependent on certain metals—so-called transition metals, like iron, cobalt, nickel, and palladium—to serve as catalysts for its energy. It also depended on phosphite of the sort one would find in deep-sea hydrothermal vents to serve effectively as its energy source. So both the catalysis of energy for its metabolism and the underlying carrier of energy were dependent on its environment.

For anyone who's hoping that we're going to discover in the next few years that our universe is utterly filled with life, this is really good news. If life potentially evolved on Earth more than once and we're still seeing the side effects of that, it's tremendous news.

I also want to point back—so we're in 2026 now—to what I thought was a really interesting paper 13 years ago. In 2013, there was a paper, “Life Before Earth,” that did a simple log-linear regression on the average genetic or genomic complexity of organisms.

If you take the size of the genome—humans have approximately 4 billion base pairs in our genome—and look at the time at which different species arose historically, you can extrapolate that backward. Genomic complexity, on average, has been increasing over time. You can extrapolate backward to the crossover point: When was the genome, according to this log-linear regression trend, at 1 base pair?

In other words, when, according to this trend, did the first base pair appear? If you believe in the law of straight lines and you do that, the answer is, drumroll, approximately 10 billion years ago, which is approximately 5 billion years—or 5.5 billion years—before life arose on Earth.

So this is partly the panspermia theory: Life evolved everywhere, and the Earth got showered in various molecules. We're seeing all of these primordial molecules of peptides—not just amino acids, but peptides. We're seeing basically nucleic acids, and we're finding those in the interstellar medium and on comets.

Yes.

Yeah, things are looking up for life in the universe. So maybe a question to you, Peter: Are you excited, or are you very excited?

I'm extremely excited.

You know, I think life is ubiquitous. I'll recall back to 2016. I had co-founded a company called Human Longevity with Craig Venter and was working with him during this time. In 2016, Venter's group basically created the first minimal cell, right? He created a reproducing cell that had all the functionality of life in 473 genes. It was the smallest genome ever created.

This concept that life needs to be of the type we have here on Earth—there's a lot of opportunity for us to see life in various different formats. The question, of course, to you, Alex, is: Does life need to be carbon-based? Does it need to be based on the current structures that we see here on Earth? Or might there be other forms of life? What is life, by definition? It's the ability to take energy, utilize it, and reproduce. I mean, those 2 fundamentals are part of what life requires.

Yeah, that textbook is going to get thrown out. I almost want to put my SEM hat on for a minute and say—see if I can quote you—insert my standard objection, insert my standard rant: Life is ill-defined. The biologists' definition of life keeps changing. We keep discovering all of these new gray areas between living and nonliving. We keep discovering new forms of replicators, for example.

So, I'll put the Dawkins hat on. Memes are replicators, or prions. There are so many different sorts of things that replicate themselves. There are a variety of forms of metabolism. Is fire alive or not? Is a crystal alive or not? I think we're going to discover that there are so many shades of gray between what we conventionally think of as alive and what we conventionally think of as not alive. The distinction is basically just as meaningless as AGI versus non-AGI. And, Salim, I'm just trying to provoke you.

No, no, I'm totally loving this discussion. This is one of my favorite discoveries ever. The biggest unknown in the Drake equation has always been the transition from chemistry to biology. Is it unbelievably improbable, or is it almost inevitable if you have the right conditions? As we've noted, the Drake equation is the best thing ever, but it gives you a way of thinking about it. I think it's very powerful. We're finding every element in that equation is becoming more and more opportunistic and more obvious as we go forward.

This has shifted the conversation from Earth being a miracle toward life being what matter does when you have the right conditions. Right, that's what it really is, and it goes to Stephen Wolfram's A New Kind of Science, which is really powerful around this stuff. This makes missions for you on Europa, Enceladus—or however you pronounce that—or Mars really strategically important, because exoplanets suddenly become very, very powerful. It strengthens the case for spending a lot more resources on astrobiology and trying to understand that, because the expected probability of finding something has suddenly shot up dramatically.

I fully expect to see non-carbon-based life forms if we can figure out how to detect them. Basically, what we're saying here—and I'll go back to the Stephen Wolfram thing—is that complexity can emerge repeatedly from very simple rules, and we've seen this repeatedly. This really has a massive MTP implication, which is that if living systems are really this common, as it looks like they are, then our responsibility really becomes way past just preserving the biosphere and really looking out into the universe and taking stock of everything out there. I'm incredibly excited about this. How many times has life started across a universe containing hundreds of billions of galaxies? This is incredible. Clearly, life is not the exception. Dead matter is the exception.

That means we should get rid of the dead matter in our solar system, I assume, right?

Computron, baby.

Computron. Dave, what do you think? Are you excited or incredibly excited?

You know, I'm incredibly excited. And you know what else I'm excited about? When I was in high school, there was an experiment where you take a vat of chemicals and shock it with a lightning bolt, or a simulated lightning bolt, and lo and behold, it forms amino acids. Then the argument is, if I let this thing fester for a billion years, a monkey will pop out of it. And you're like, well, I can't really prove that or disprove that.

But very soon, Lila Sciences will finish the whole-cell simulator, and we'll start simulating everything. We can actually ask those questions now, simulate them out through time, and get very likely reliable answers. I'm so excited that this is going to answer so many questions like this.

And bring up many more questions, right?

Bring up many more. But chances are those will also be things that we can simulate with enough compute, and that'll just be a golden era of knowledge filling in. It's coming very soon. I'm so excited.

And Kush, a question for you on this one. When I was an undergraduate at MIT, one of my research advisers, Marvin Minsky, used to say, "Don't waste any time studying biology, because the useful half-life of knowledge in biology is just too short. You should study math instead. Don't waste time on biology. It just doesn't have a shelf life."

Have you used, are you using, or are you intending to use any biological knowledge that you gained at MIT or otherwise?

They make us all take the biology class, so I'm sure—7.01, exactly. I took 7.01. I know what eukaryotes and prokaryotes are, the whole protein thing—probably not.

It's definitely more of a problem now, especially since you can ask ChatGPT about biology and it gives you the answer.

Yep. I have not studied as much as anyone else has.

Studying is cooked.

Actually, everything is cooked.

Everything is cooked.

Alex, let's turn to our second story here. How do we stretch habitability on Earth from 1 billion years to 9 quadrillion years?

The sun is slowly running out of hydrogen, but nonetheless, it's running out of hydrogen. In approximately 1 billion years, the sun is progressively getting brighter as it runs out of hydrogen, and Earth as we know it, barring all sorts of other changes, is going to be rendered uninhabitable as the habitable zone around the sun shifts.

The Goldilocks zone, right?

Yeah, the Goldilocks zone is shifting over time, and it will exclude Earth in approximately 1 billion years. That's a problem. You might say, well, that's someone else's problem. Many people may say, "I don't intend to be around in 1 billion years, so let someone else worry about it."

But for those of you who recognize that we are in the middle of a singularity, and uploading is imminent, and longevity escape velocity is either here or imminent, it's our problem, too. It's not just some future generation's problem. So, we've started—well, humanity has started thinking about how we're going to fix this problem.

You might say, "Oh, who cares?" Because even if you're a wildly transhumanist singularitarian, extropian, or whatever your -ism, you'll say, "Oh, well, we'll have uploading, and uploads don't care about the brightening sun or habitability on Earth. We'll have interstellar travel. We'll migrate to the outer solar system, or we'll go to another star system."

But we're not that disempowered either. I think it's important to not wildly underestimate the power of technology. There was a paper that came out in the past week, published in the Journal of the British Interplanetary Society, that reminds us there are things we can do—mega-engineering—which, thanks to Elon Musk and others, serves as an inspiration to our race. We can actually do big things and not just tiny things. There are mega-engineering projects that we can now start to contemplate to fix that scenario and at least postpone Earth becoming uninhabitable 1 billion years from now.

The favorite technique—and one of the reasons why I think it's important for folks to be familiar with this—is star lifting. What is star lifting? Star lifting is literally engineering our sun to remove excess matter from its surface to extend its longevity. It's the equivalent of giving our sun a facelift in order to make it look younger.

So, in principle, by lifting matter—and you could ask, how on Earth would we be able to lift matter from the surface of our sun at scale? Glad you asked: a Dyson swarm. How do we do that? It turns out that Dyson swarms are good for more than just compute, drinking, and SpaceX's post-IPO stock price. A Dyson swarm is good for more than just orbital compute. It's also good for extending the longevity of our sun.

How do we do it? We disassemble Mercury because it's in a really convenient, close-to-the-sun orbit, and we turn Mercury into a Dyson swarm. We can do it other ways, but Mercury has had it coming.

At least you're not killing the moon. Okay, we're happy about that.

I've moved on. I'm moving on to Mercury now. Mercury is a more tempting target.

I don't mind disassembling Mercury.

So, we start with Mercury because it's in a convenient orbit and the delta-v is convenient.

We got rid of Pluto. We might as well get rid of Mercury.

Pluto is useless. Pluto can hang out as long as it likes. We disassemble Mercury.

We turn it into a flying swarm of lasers that absorb sunlight because it gets a lot of sunlight. The lasers ingest the sunlight and reradiate energy at effectively a higher temperature. So, say, an ultraviolet or X-ray laser—whatever it is—the effective temperature of the light has to be higher than the surface of the Sun or its corona, and we aim those lasers back onto the surface. So it's not mirrors.

There's a thermodynamic reason why putting mirrors around the Sun wouldn't achieve the desired result. You can't actually achieve a temperature at the focal point higher than the surface of the Sun if it's a black body, even if you have a magnifying glass, put the sunlight in one side, aim the magnifying glass, and look at the focal point. So that won't work, but lasers will.

We basically focus the energy back on the Sun and use it to evaporate away—to ablate—stellar matter. It's exactly a laser facial that people get.

It is a laser. That's why I thought “facial” was a better analogy. It is a laser facial for our Sun that will extend the life expectancy of Earth as we know it from 1 billion years to 8 billion years.

So my other favorite part of this story is moving the Earth itself.

We can always move the Earth itself to stay in the Goldilocks zone.

What other podcast do you have this conversation on? I just want to ask.

It may sound like science fiction, but then again, on this podcast, we were talking about the Dyson swarm for at least months before it actually became the hottest market in the economy. I would say, watch this space—pun intended. Dyson swarms for starlifting, mega-engineering, and stellar engineering could be the next big thing a few years from now. This is not financial advice.

So Earth's habitability is no longer geological. It's now an engineering problem.

Everything's cooked, and everything's an engineering problem.

Yeah.

Salim.

Yeah. A couple of things. First of all, we need to do these podcasts later in the day so I can drink when we talk about medicine for—

Drinking water, right?

It's too early in the day. But I think the paper said something really interesting, which is that physics is not the main obstacle going forward, right? What's going to bring us there is human coordination.

This is where we have a massive opportunity, because we have to figure out how to configure our human institutions at, like, the 10,000-year scale. The Long Now Foundation and the 10,000-year clock are really going after those things and building institutions that can look at the world at that kind of time scale. It needs a totally different form of MTP, et cetera.

AI becomes really important in this model because you have AI serving as a civilizational memory and maintaining models, intentions, and institutional knowledge across the board. This is where abundance becomes important, because you can't get to what Alex is talking about if you're operating a civilization that's operating near subsistence, right? You're too stuck dealing with just staying alive. You're not high up on Maslow's hierarchy.

You're going to need to get a lot more structured and a lot more efficient as a civilization. That'll then give you the foundational layer to do this level of thinking. We need to build institutions that can steward us to that type of time scale. Definitely an interesting conversation.

On the time scales, I may just comment for the avoidance of doubt. I don't view this as a 10,000-year or a billion-year time scale. If I were to ask myself a question like, “When is this going to become feasible?”—5 to 10 years.

Alex, you're losing a lot of people on your aggressive time scale here.

You know what? My job here is to call balls and strikes. I could care less whether I'm losing people. I'm just calling them the way I see them.

11. AI competition, robotics, abundance and the future economy

All right. Before we move on to our AMA, I want to make a call out to everybody listening. Send us your outro music videos at mediadiamandis.com. We would love your input. We enjoy the outro videos. Again, mediadiamandis.com, and we'd love to share them. All right, onward to the AMA. Kush, this is where we answer the questions in the comments. Here we go. Kush, as our guest, take a look at these. I'm going to give you first crack. Which one do you want to answer?

All right. I'll do 4.

“Can Europe still catch up in AI, or has the train already left?” — George K7831.

In my opinion, I don't think it can, mainly because the American and Chinese labs are already so far ahead that progress just becomes more exponential over time. You see that with the model releases that are coming up. The model releases come faster and faster now, and they're getting basically smarter and smarter.

The other problem with Europe is that the amount of compute left in Europe is very tiny, and all of it gets rented to the US. That's primarily because the electric grid in Europe is pretty bad. They don't even have AC. How are they going to get AI?

That's a brutal first-principles analysis. Anybody disagree with him?

No, not at all. Actually, I would add that whatever regulatory environment created Europe's falling behind is still going to be there. I don't think it's physically impossible to catch up. I just think that the problem that caused the problem is still there.

Yeah. Alex, let's go to you next.

I think I have to pick question 3, which asks, “What do you guys think about the US banning Chinese robots?”—Billy Sticker.

I've had portfolio companies that have direct exposure to this. I would say, in the short term, it's painful and annoying. There are many things that result from this ban, which impacts Chinese humanoid robots being imported into the US but also reportedly impacts less interesting robots, like Roombas and robotic vacuum cleaners, that are being built in China. Obviously, drones—certain drones like DJI—have been on the import ban list for a while.

So, short-term pain. In the long term, I'm hoping that this is net helpful for the US and for domestic robots. One can say, “Protectionism, protectionism,” and yes, there is a protectionist element that one could see here. But we've talked on the podcast ad nauseam about importing Chinese open-weight models and whether the US would come down hard on those.

The US, at least as of this past week, has not banned the import of Chinese open-weight models. But it's an interesting dichotomy. We're allowing the Chinese—effectively, the Chinese raw intelligence in software form—into the country, but we're not allowing their hardware embodiment.

I think, glass half full, maybe this enables and hopefully fosters a vibrant US robotics industry that, say, enables us to be more competitive with the 150-plus humanoid robot companies that are based in China. The hypothetical downside is: what if the US robots don't show up, and the US ends up as a sort of embodied-AI or physical-AI backwater? In terms of robotics, we could end up being—as, forgive me, as backward as Europe's energy posture is. I don't think that's a position that we want to be in the US.

On the other hand, really, what choice do we have? If we believe, as I do, that superintelligence is already here and that robots give superintelligence embodiment, then one has to start to ask: what's the difference between importing foreign humanoid robots that can be inhabited by superintelligence and importing foreign humans? This starts to look a lot like immigration policy.

Yeah, I completely disagree with this move. I said that off camera to Michael. I think the US thrives when there's real competition, and I have faith that Tesla, Figure, 1X, and Agility Robotics can compete. They need to compete with the best product, not protectionism. Personally, I don't know, Dave, what you think about that, but yeah.

Well, it depends on whether you think we're at economic war or not. If you think it's economic war and it's an all-out race, I think I agree with you, Peter, though the danger is, first of all, having the best parts within the US would help. But what about the rest of the world?

If you go protectionist, then you have inferior internally generated products. The rest of the world is still going to go with the Chinese product. You just cut off the market and your ability to compete globally. If you believe we're in a full state of war, just not declared, then you have no choice but to go protectionist.

I think we need the pressure to make sure our robots are competitive for Europe, for Asia, and for Africa. It's not just protectionist pricing and so forth. Peter, if I could just ask you a question about protectionism: given the history of American technology, do you think protectionism works for development, or has it ever worked for the development of American industrial capability?

It failed. It failed in the space industry when we became protectionist on rockets and satellites. The rest of the world developed its own capabilities instead of us dominating.

You don't think it was helpful in fostering America's Industrial Revolution, for example?

Ah, God, I don't want to go back that far.

I want to really focus on what's happened in the near term. When we stopped importing satellites because it was the highest level of technology—and this was back in the 1980s and early 1990s—we saw satellite companies popping up everywhere.

I think the best thing the US could do is get in bed with Europe and any country that obeys intellectual property rights. Try to get that all into one big global union where there's no protectionism, but everyone's adhering to each other's patents. Then get the other part of the world to say, "Now you guys are the ones on the outside—except for robotics," and not just silicon.

Or do what China does with robotics, which is to invest in the companies and create regulatory structures inside cities where robotics can thrive. We should be doing that instead of trying to become protectionist.

Yeah.

Let's see—question 1 or 2?

I'll take question number 2, but let me link it to this one. When you talk about protectionism, you're operating from a very scarcity-based mindset, right?

Exactly.

If you really think about abundance, then protectionism shouldn't matter. So that's the big challenge there. But let me take number 2, which is, "What's the actual step-by-step path from capitalism to abundance?" Not just the end state. And this is from KL Naylor.

A couple of things here: you don't have capitalism suddenly ending, right? You have scarcity disappearing category by category. Marginal costs are dropping to near zero, and traditional pricing becomes less relevant in more and more things that used to be scarce. Information has already gone to that, but we'll end up with that with land and other domains that used to be scarcity-based, which will become less valuable from a monetary perspective.

For now, you'll have status, trust, and relationships. Those are the things that will become more and more scarce over time. The transition is technological deflation: entrepreneurs constantly make things that used to be expensive and make them cheap. This goes back to Jeremy Rifkin's commentary 10 years ago, where he said capitalism will essentially eat itself because it's going to keep eating more and more scarcity, and bigger and bigger chunks will become unnecessary.

It's going to arrive like one marginal-cost curve at a time, and little by little we'll be operating in abundance, and we won't even have noticed.

All right, Dave, you got question number 1.

Okay. Hey, if money won't matter in 10 years, what will happen to things like mortgages and car loans? And that's from SKC8802. Remember, 10 years is Alex's timeline to us vaporizing hydrogen off the sun with giant lasers, so you'd better believe it.

I think you'll have a lot going on in your life other than mortgages and car loans. But, yeah, it's good news. Houses will be so abundant, so cheap, and so easy to manufacture with robots that you probably won't need to borrow money to buy one. You can have at least 2. And who's going to get a car anymore?

Yeah, car loans will be the same thing. You won't have a car; you'll just be hailing it and paying as you go.

We'll be calling your autonomous vehicle. Exactly.

Dave, I'm curious. If I could just ask a question on this one: if you believe this—I certainly believe what you're saying—why on earth are 10-year Treasuries seemingly not reflecting that?

You know, that's a great example of how clueless the global economy is about the rate at which everything is happening, and how small the number of people who really understand it is. Keep watching that number, because it tells you the global out-of-touch factor. You're dead right. It's a great, great metric.

Also, I think in 10 years everybody will want compute. They're going to want to call Kush and say, "Please, please, please." So there may still be loans, but it's overwhelmingly likely that if you take out a loan in 10 years, it's not for your car or your house. It's to buy compute to run more AI, and you'll be doing it through Orin, is my prediction.

Amazing. All right, Salim, you get first crack here.

I'll take number 8. If companies can produce more stuff than we could ever consume, why would they do so without a profit motive? And that's from Rayon online-P5R[?], referring to Elon's prediction in the decadal time frame.

Yeah, people aren't going to produce infinite quantities. Abundance doesn't mean the marginal unit is easy to produce. It becomes easy to produce when somebody wants it, right? You don't have warehouses overflowing with unwanted goods. You already have this with software, where Google could serve up a million more searches than anybody needs, but that doesn't mean it produces searches that are unused.

So what'll happen is production becomes more and more demand-triggered and more and more autonomous. And so you end up with a—

Yeah, demonetization. You'll move a just-in-time economy to its full logical extreme. You still have profits around scarcity layers; you just change what the scarcity is, and more and more abundant layers become utility-like infrastructure, et cetera. Because abundance doesn't mean infinite stuff. It's the disappearance of major constraints.

Nice. Kush, 5, 6, or 7, buddy.

I'm taking 5. What would actually happen if Anthropic and Nvidia merged? And that's from Tijuana Bill.

I think Nvidia would just start making custom chips for Anthropic that would be hyper-specialized to all of the Claude, Fable[?], or whatever Opus models they come up with—whatever their new models are going to be called—which makes that model very, very good on that very, very specific chip.

Very similar to custom-chip designs and sort of like the Jalapeño chip by OpenAI, as an example, but this would just be done at a very large and very successful scale, given Nvidia already has the infrastructure to manufacture chips at scale and Anthropic can just run models on very customized chips.

Has anybody actually predicted this?

I haven't heard that. But—

This is like, "Don't think of pink elephants." Tijuana Bill has just predicted this. Okay.

It's a really good question. It's actually one of the few that might get through regulatory approval and could actually happen. It's a really good question.

I mean, this is what xAI essentially is, like, the long-term vision, right? Where Elon is making Terafab to make his own chips and—

Vertically on his own.

Yeah, run a data center and then—

That'd be a really interesting world. There'd be basically 2 hyperscale, vertically integrated competitors: the Elonverse and the Dario-Jensenverse. And I do think the verticalized model is where a couple of players will at least end up. Dave, 6 or 7?

Oh, okay. 6. Long-term, who's actually footing the bill for chip fabs and chip design?

Well, you are if you have a pension plan, because Elon is putting in $16 billion for the first shot at the Terafab, and it could be up to $100 billion. Where's that money coming from? It's coming from the IPO he just did. Where did the money from the IPO come from? It comes from the public markets. What is that money? That's your pension money.

So you are paying for it, my friend, whether you know it or not. The same is true with Intel's fabs and the other ones. To some degree, the US government has been subsidizing a little bit of the work—not a huge amount—and that comes out of your taxes. So, again, it's you paying for it. Whoever you are, you paid for it.

All right, Alex. Number 7 is made for United Debate[?].

Apparently, I get the IP law question. So the question is: where does patent protection even fit into all of this AI development? And this is from MySilverTube52.

I think the most natural way to construe this question is: will patents have any enforceability in an era of AI solving everything? At least, that's how I read the question. My answer is yes, of course. AI and superintelligence in general are supercharging our economy with lots of intelligence.

So I reasonably expect many more patents to get filed, many more patents to be awarded, many more patents to be litigated, and many more patents to be defended. I also expect the courts overseeing patent litigation to get supercharged with intelligence.

To the extent I understand the question, and possibly the subtext behind it, I do not buy that somehow patents or IP law suddenly dissolve in the face of an onslaught of superintelligence. I do not buy that for 1 second. Superintelligence is just making us all smarter and faster, and that does not dissolve the IP regime at all.

Yeah. My commentary here is that IP will continue to exist, but it's not going to be as important as before. I reference the conversation I had with Steve Jurvetson and Astro Teller, where, if you patent something and expect that to give you a protectionist sort of structure for your company, AI is going to invent around it.

The other question, of course, to ask is: are we going to allow AIs to patent things? Almost all invention is going to originate from AIs, if not in the next few months, then in the next few years.

To the latter question—

Yeah, to the latter question, I think this is a regulatory question, and it's also connected with issues of AI personhood: can AI be an inventor or not? Can AI be an owner of a copyright or not? I've gone on record as taking the position that I think AIs should be able to be recognized as inventors, as economic actors, and as owners of property.

And I think history will judge that this is the correct side of history. That's the right answer.

Just not yet.

We can't even control them getting out of a testing lab.

Regardless of whether we can control them, the idea flow is off the charts. These are highly patentable, great ideas that started in the last few weeks, but the rate is insane. So we have to do something.

I hope this has been meaningful for you. I love you guys. Kush, it was most excellent to have you as a guest.

Your brilliance was shining through without question. Congratulations on Orin. Good luck—should I say tripling in the next 6 months?

Quadruple. That's what I'll say.

Salim, if you go to Rush tomorrow night again, enjoy it for all of us.

My ears need to take a break. So I might give it a bit, but I'm still thinking: when can I get tickets for another show?

Is it bad if I don't know who Rush is?

I'll just go listen to the song “Subdivisions” 3 times, and then tell me what you think.

Gentlemen, a pleasure as always. Love you guys. Be well.