Balaji谈AI/AGI现状、Bitcoin与美国即将到来的崩溃——与Dave和Salim对谈 | EP #191

And so your first point is that there is no AGI. There are many AGIs.

What I think AI takes us toward is that everybody’s a CEO. Everybody’s a manager. It’s a lot like interacting with an employee, especially a junior employee, and you have to check their work. As intelligent as the AI is, it’s far, far better with a human interacting with it.

Yes.

This means the smarter you are, the smarter the AI is working with you. It reverses the extreme specialization of the 20th century. Your wingspan increases.

At the end of the day, Balaji, do you believe there are going to be multiple players out there providing equal or similar capabilities, compared with what you just said, Dave—that there’s going to be someone who will be supreme over everybody else?

Here’s what I think about that.

I’m here with my Moonshot mates, Dave Blundin and Salim Ismail. Gentlemen, welcome. Today we’re joined by one of the most extraordinary, outspoken thinkers on the future of AI, crypto, and society at large. Balaji, it’s a pleasure to have you.

If you already know his work, you know how extraordinary he is. If you don’t, you’re in for a real treat. Balaji is an entrepreneur, investor, and futurist; formerly the CTO of Coinbase and a general partner at Andreessen Horowitz. He’s also a deep thinker in biotech and longevity, the founder of Counsyl, one of the earliest consumer genomics companies, an investor in biotech startups at Andreessen Horowitz, and he taught genomics at Stanford, which I found really cool.

Most recently, Balaji is the author of The Network State and the founder of the Network School, offering a bold blueprint for how online communities can evolve into self-governing sovereign entities. I want one of those. I definitely want one around abundance.

Today we’re going to speak about his recent paper, “AI Is Polytheistic, Not Monotheistic.” Super important point. We’ll dive into Bitcoin and crypto, make a right turn down longevity lane, and end up talking about how insanely different the world is going to look in 2035.

Balaji, I think it’s good evening for you. You’re in—

Yes.

Singapore today, at the Network School.

That’s right. I am.

We’ll go there. And Salim, Dave, you’re in your normal—

Normal spots.

Normal habitat.

There is no normal spot, but yeah, as close as you can get.

All right. Well, it’s been another incredible week across the board. Balaji, the last time you and I saw each other physically was in Saudi Arabia at the FII Summit, but before that was in New York. We were talking about longevity, and we were just getting the idea of the longevity, the health-span XPRIZE, up and going. It’s since launched, and we have 630 teams competing to add 20 or 30 healthy years onto your life.

I want to talk about that, and about how the future of extended human health span is going to mix in with Bitcoin, crypto, network states, and all of that. But I think one of the most important papers you put out recently was this conversation around AI being polytheistic and not monotheistic. You put out 10 points, and I want to dive into that.

Sure.

I was looking through some of your podcasts and listening to them, Balaji, and I really love the idea of fighting for Mars, Bitcoin, and life extension. It covers a lot of the bases we look at here. Really awesome.

That’s right: infinite frontier, immutable money, eternal life. What could possibly be better?

Let me start with a broad-ranging question. Do you think the world is going to be recognizable 10 years from now, in 2035?

I mean, in the sense that the continents will exist, yes. But in terms of the superstructure, the societal layer above it, I think very few institutions that predated the internet will survive the internet. I look at it as, in a sense, the single most important force in the world, yet it’s still underestimated.

It’s upstream of AI. It’s upstream of crypto in a real sense, upstream of China, upstream of robotics, upstream of what we’re doing right now. It’s in front of your face when you wake up, and it’s on your watch when you go to sleep. And yet, it’s in front of politics.

Twitter elected Trump, and Twitter deplatformed Trump. Then X reelected Trump, and that happened in every other country. Twitter caused Brexit, and Twitter caused every political fracas we’ve seen. That’s just one rivulet of the internet.

I think the world is going to be radically different. You guys are a little bit older than me, but I think the world was relatively static from 1980 to 2015 or so, and then it really started changing quite rapidly over the last 10 years. I don’t know if you guys feel the same way.

I had Ray Kurzweil on our stage at the Abundance Summit, and he said we’re going to see as much change between now and 2035 as we saw between 1925 and 2025—a century’s worth of progress.

I think that’s right.

Yeah. While the continents are going to be the same, and probably our biology may be the same, society’s structures—when you say those will all disappear, do you mean political systems, financial systems, or what systems?

Yeah. My view on the world is internet-first, and I think that, essentially—okay, I’m going to give a very quick macro structure and then go from there.

Right now, the 2 growing powers in the world, in my view, are China and the internet. China is everything physical; the internet is everything digital. The internet has disrupted blue America and taken over media and money. China has disrupted red America and taken over manufacturing and the military.

Chinese robots and Chinese drones—that’s on the physical side. Then AI in media and crypto in money—that’s on the digital side. These 2 forces are just growing stronger and stronger.

A lot of the 2010s was about trying to resist them with the trade war and the techlash, respectively. But the resistance was actually overcome, and China and the internet are winning. I think the future is China and the internet.

Again, that is communism versus capitalism, much as, for example, in the early 1900s, Europe was the center of the world and then buckled in. It was American capitalism versus Soviet communism, splitting Europe down the center, with the West Germans on the capitalist side and the East Germans on the communist side.

I think America was the center of the world in the early 2000s, and it’s sort of involuting. You’ll have the Chinese on one side and the internet on the other side, with a lot of Chinese Communist Party factions and Bitcoin maximalist factions. You’ll have lots of different startup societies around the world.

One of the reasons for this is that, since you asked for my view on this, people talk about what money is. People talk about it as a unit of account, a medium of exchange, and a store of value. But Andreas Antonopoulos had a good line about the fourth purpose of money, which is a system of control.

That is to say, you can freeze accounts, seize accounts, direct printed money, and give banking licenses or not. That’s actually, in some ways, an even more powerful tool than the police system, because the police aren’t as scalable as the monetary system. The monetary system is everywhere, and it directs people silently. There’s not really a vote over who gets a banking license or not, or what have you.

That entire system—the Fed’s control over it, including the US dollar, but also the Canadian dollar, the Japanese yen, the British pound, the German mark, and so on—is under pressure. If you look at the yields for all Western currencies, they’re rising like this. The yield for China is going down like this, because all kinds of money is going into China. Conversely, Bitcoin is also appreciating.

All the capital is going into RMB and BTC. That means it’s not just the US dollar, but also the Canadian dollar, the British pound, and all the things that are in the US sphere—what you might call the American empire, the West, or the international community; what Putin calls the golden billion.

All of that kind of crashes at the same time in real terms. That means control is lost over a very large swath of formerly wealthy territory.

Hold on—2 clarifying questions very quickly. First, when you say “the internet,” I’d love for you to describe what you see as the internet, because that’s very amorphous.

When you say “the internet” and “China,” would you also equate centralization and decentralization in that equation?

Roughly, yes. Correct. So China is roughly—when I say China and the internet, that’s shorthand for China itself, but it’s also the physical world.

It’s the past, it’s the land, it’s gold, right? It’s bricks, both in the sense of gold bricks and brick-and-mortar; it’s minerals, it’s mining, it’s roads, it’s the giant cities, and so on, coming out of the ground, right? And what is the internet? The internet is the virtual, right? In a sense, actually, China is the 19th century, right?

It’s the past. Or you could also say it’s like 1950s America, with the girders and stuff coming out of the ground. The internet is the virtual. Really, when I say “the internet,” I mean the free internet, the Anglospheric internet, the unregulated internet. That is obviously cryptocurrency and AI.

Of course, once you define these poles, there’s obviously a Chinese internet. In many ways, that’s like the front lines, right? The front lines are: What happens to the Chinese tech founders? Are founders like Jack Ma purged by China, or do they go abroad and start startup cities—a Chinese Capitalist Party to compete with the Chinese Communist Party? I think we’ll see some of that. Or do they integrate with the Chinese regime? I think that’s a very important sort of thing.

It’s like how Germany was a border zone for communism versus capitalism. That’s one very important border zone. Another border zone is India, where there’s physical India, which is in bricks, and there’s internet India and the Indian diaspora, right? In this way, you can identify border zones between the physical and digital all over the world.

I think that if you apply that lens to many things, that’s also state versus network, right? It’s land versus cloud, state versus network, physical versus digital, past versus future, gold versus digital gold. But what’s going away is basically the “rules-based order,” the postwar order. There’s another very important graph.

I’m going to stop and jump in. I’m going to slow this down because I’m sure our listeners are spinning, as well as I am. I’m probably going to come back and listen to this at 0.8 speed just to hear it one more time. Dave, you wanted to say something?

Yeah. Well, the whole time that I was growing up, Japan was going to take over the world. The Japanese economy was skyrocketing. It was about to catch up to the US and bypass us in per-capita GDP any day. My brother actually went to Japan to learn Japanese because of the imminent takeover.

Then it all stopped cold, largely because of an aging demographic that no one had factored into their math. Now, China’s per-capita GDP is still way behind the US, but total country GDP is catching up fast. If you look at the underlying aging demographic, it’s even worse than Japan—by far, it’s the worst in world history.

The reason the US has stayed strong for so long is largely immigration. We have incredibly talented people from all over the world, including China, coming into the country and creating new companies and new ideas, and China has none of that. They have virtually no immigration, so I’m curious to get your take on how that’s going to play out.

If you view the 10-year future of the world as the internet and China—yeah—but is China going to grind to a halt like Japan did?

Totally. Let me actually take all your points in reverse order. First of all, I’m very sympathetic to your worldview. I like that America; that’s the America that we both grew up in and spent a lot of time in. So, I’m only saying these things from an analytical lens—not “this should be the thing,” but “this is the thing.” That’s okay.

First, in reverse order, China has just introduced a so-called K visa, which they’re using to recruit global tech talent. The US is cutting off H-1B visas, researcher visas, student visas, and even tourist visas. Europeans and Australians are reporting being strip-searched; tourism has crashed. The US nationalist faction is cutting off visas, and unfortunately, a lot of socialists are also very much against tech. They’re not really going to be that enthusiastic about more tech guys coming in.

Regarding Japan, the reason that they lost in the 1980s was actually because of something called the Plaza Accord. At that time, remember, the US still occupied Japan. It had the US military in Japan, it nuked Japan, wrote its constitution, and so on and so forth. The Plaza Accord, among other things, stopped Japan from actually using the JPY to go and buy US property and so on, and forced it instead to buy worthless Treasuries and money-printing stuff. Those policies began the Japanese Great Stagnation.

One way of contrasting the big difference between Japan and China is that China is genuinely sovereign. What does that mean? For example, in 2017, the New York Times reported that, in 2010, the Chinese had gone and killed many American spies in China. They had rolled up a bunch of rings there. Now, of course, spy versus spy happens. China obviously spies on America, and there’s all kinds of this thing going back and forth.

But the reason that’s an important contrast is that Japan could never kill American spies, right? It could never do that. It could never have communications or operations that weren’t seen by America. It wasn’t truly sovereign. America had root control over Japan in a way it doesn’t over China.

In fact, that’s the common thread of Chinese policy over the last 50 years. You know how certain founders, like Zuck, may have sold some of their company but never gave up control? He always wanted to maintain voting rights, even if he gave up some financial rights. In the same way, at many times, if you study the last 50 years, China has taken worse financial terms to retain control. They have always pushed for sovereignty.

Their culture is the Century of Humiliation, and they were crushed for many years, so they can never give up sovereignty. They’ll never do that again, no matter how much it costs. It’s very similar to a founder who’s very motivated to keep control of their company.

As a consequence, China is sovereign, and it has surveillance-proof deliberations in a way that Japan doesn’t. As one concrete example of that, last year—or the year before—when Gina Raimondo, who was the then commerce secretary, went to China, they launched this Huawei phone on the day that she arrived. It had chips that showed that China had at least partially gotten out from underneath the chip ban.

That kind of thing could never be done by Japan because Japan couldn’t do anything in stealth, right? China can do things in stealth. Japan can’t.

Finally, I would say that within tech, we often see something that kind of half works. It’s a bubble, then people discount it, and then it comes back 5, 10, or 15 years later and really works. For example, Webvan didn’t work, but then DoorDash did. The PalmPilot kind of worked, and then the iPhone did, right? World of Warcraft kind of worked, and then Fortnite really did.

If you think of China as the real version of that, where it’s actually working, then that’s another thing where past performance may not guarantee future results. Let me pause there.

Can I ask you about timelines on this, too? A lot of the trend lines are 10-year trend lines, and Peter’s question was a 10-year question, but AI is like a 2- or 3-year thing. If I use the analogy of World War II, Pearl Harbor Day was December 1941. The war was effectively over in 1944 and truly over in 1945. From start to finish for the US, it was only a 3- or 4-year journey.

I think AI is on that same kind of trajectory, where the declaration of war was essentially the chip ban against China. We’re about a year into it now, and the timeline from here to a resolution, I think, is going to be similarly very, very short because of the exponential rate of change. When we look at demographic trends—long-term demographic trends—that’s always going to be 10, 15, or 20 years to resolve, but I think the actual action is on a much, much shorter timeline. Do you agree?

Okay. So, I’ve got 4 more counterarguments here.

And then I’m going to move us to the paper that started this conversation, but I really want to hit on that pretty hard.

Of course. I’ll be brief. I have 20 questions, but I’ll wait.

Okay, great. So, first is that China’s demographic problems have robotic solutions. Why did the South lose the Civil War? Because the North industrialized. The North didn’t have slave labor. Many times, if you look at societies that have lots of, for example, illegal aliens—not slaves, but people who are often paid under-the-table wages—they’re exploited in different ways.

If you have a society that’s got an abundance of so-called cheap labor, it often fails to industrialize, fails to automate. China is using its demographic pressure to automate, automate, automate, automate, as a 1% shock every year to force them to automate. If you’ve ever been to China, it’s already ahead on that kind of thing. They already have delivery drones, hotel robots, and all of that kind of stuff.

Point number 2 is, if you actually look at other metrics besides dollars, one of the issues is that dollars are controlled by the US government. So, you have to actually look at physical production. China, for example, has 1 shipyard—as the US secretary of the Navy, Carlos Del Toro, said—that produces more ships than the entire US Navy combined. Hegseth, who’s the current secretary of defense, has said Chinese hypersonics can sink all US aircraft carriers.

The Raytheon CEO said that the US can’t decouple from China. There’s a Govini study commissioned by the Pentagon—a $400 million study—that basically showed the US military is made in China. If you look at the Tomahawk or the JDAM, the famous American weapons, their suppliers are all Chinese factories.

So the war is actually already over. It’s already been lost because China can just hit a button and shut off the supply chain, right? That’s why the rare-earth elements issue was going on, and so on and so forth.

In a sense, the war was fought with the trade war in 2015, and the US lost because what China did was diversify its revenue streams outside of the West. Only 15% of its revenue comes from the US; 85% comes from non-US sources. So even if the tariffs chopped it in half, they still got 90% to 95% of their revenue left.

I think that even the chip ban, for example, is challenging the Chinese to a test of quantum mechanics and semiconductor production. When all 4 of the major chip companies look at it and ask, “Can the Chinese, as a culture, do chipmaking?” Of course. It’s like challenging the Italians to a pizza-making contest. This was right in their wheelhouse: doing math and computer science, and manufacturing for national pride and Chinese sovereignty.

It would be much better to challenge them on, for example, arguing in English. That’s not their strong point—making international media their only success. Go ahead.

No, that’s funny. That’s funny. I’m going to—I’m going to—because I really want to hit where our audience wants to hear about.

Of course, I want to dive into AI and crypto in particular here.

One thing I will say—I’ll just say one last thing—is I do believe the internet can balance China. MAGA is trying to be more China than China. We have to be more America than America. That’s what the internet represents. It represents free speech, free markets, decentralization—things that China can’t clone internally without breaking its whole system apart, right?

I think that a lot of what MAGA is doing is China envy, trying to outmanufacture China, even having a mimetic rivalry over Taiwan. But we can play our own game, which is the internet. So I’m hopeful, but I think the hope is on a different axis than people think. Let me pause.

I have a quick question before you switch topics. Peter, Bal, I’ve heard you talk about China, the internet, Democrats or Republicans, right? Yes.

I’m assuming in that overall equation, Europe is irrelevant, India is mostly irrelevant, and Brazil is irrelevant.

It’s much more minor. I think Eastern Europe—so let me be more granular—is actually fairly bullish. I’m extremely bearish on Western Europe. I think it’ll be the opposite of the 20th century, roughly.

Every country that did poorly in the 20th century because they suffered through communism and socialism, they’re now capitalist and/or rationally nationalist. Conversely, every country that was wealthy in the 20th century became degenerate, and they started fighting amongst themselves.

So I actually think Latin America—El Salvador especially, and maybe Argentina—is on the upswing because they’ve sort of gotten immunized against a century of drugs and so on. I’m not saying everywhere gets there, but I think Bukele could be the Simón Bolívar of Latin America. Bukele is young in his career. He could be the Simón Bolívar of Latin America. I’m actually very bullish on him.

I’m actually very bullish on Indians and moderately bullish on India on many metrics. If you look at nuclear and steel, China is far and away number 1, but India is a real, though distant, number 2. It has come up a lot. It’s radically different from when I was young. In my youth, it really was a third-world country. It’s really not that anymore in big pieces of it, right?

So I’m moderately bullish on India and extremely bullish on Indians. I’m bullish on the Middle East in the form of Dubai and Riyadh. I’m bullish on Eastern Europe. It’s basically just an involution, almost a flip of the 20th century.

All right, let’s go to the AI post.

So you wrote a paper, which I read and enjoyed, called “AI Is Polytheistic and Not Monotheistic.” We’ll put it in the notes here. You identified 10 key points, and I think it’s important for people to understand them because some of them are very fundamental to where things are going.

There was a lot of fear early on that we would have a hard takeoff of superintelligence and there’d be 1 runaway technology that ruled them all. Your first point is: There is no AGI. There are many AGIs. I think that’s critically important. Why don’t you dive into that?

Sure. Let me preface it by saying I actually take the superintelligent stuff seriously, in the sense that I think it’s worth engaging with. Some people will dismiss it as, “Oh, that’s obviously not a thing,” but clearly intelligent human beings exist, dolphins exist, and octopuses exist. Intelligence could exist outside of a human brain, right? Obviously, it could.

I think that actually existing AI over the last 2½ years, as it’s settled down a bit—and there may be another dogleg up in the future, right—

There may be another evolutionary tree that brings us beyond what current systems can do.

Yes, exactly. Because right now, LLMs, and more generally probabilistic models, are actually very good at System 1 thinking—the instinctive thing. Andrew Ng used to say machine learning is what a human can do in 1 second, and AI is maybe what a human can do in 1 minute or 10 minutes, right? It’s like extending the time.

Conversely, computers historically have been very good at System 2 thinking: logical chains of reasoning for long periods of time. So it’s not impossible to imagine something like Toolformer that could use System 1 thinking to route tools for System 2, and some of the stuff with Replit and code generation is kind of there.

I think we’re going to go into a degree of—not exactly an AI winter, that’s maybe too strong—but let’s talk about AI being polytheistic, not monotheistic. That post is really about a constrained view of AI, like AI is actually constrained in ways that people haven’t been talking about.

For example, what didn’t happen was a hard-takeoff scenario of just summoning the demon and this 1 god. By the way, that was not impossible to imagine. Someone could have had that. For example, the US invented the nuclear weapon, the nuclear bomb, and for a period it had it before anybody else. Had the US been like the Nazis or the communists, it would have conquered the whole world and blown it all up. But because it was a genuinely good actor, it didn’t do that. It didn’t use that power for as much evil as it could have.

AI isn’t like that. AI very rapidly, on the grand scheme of things, decentralized, even faster than nuclear weapons. I even think the analogy of nuclear weapons only takes you so far.

Now you have the Chinese open models—which, by the way, was counterintuitive because it’s expensive to build these models—Kimi, Qwen, and DeepSeek, and so on, as well as Llama. Llama 3 was actually a really good thing that Facebook put out there. Llama 4 has not been as good, but I give a lot of props to Zuck for doing it in the first place.

These are expensive things to put out there. All of those—I mean, there are so many good AI models now, right? At a minimum, there’s ChatGPT, obviously; there’s Claude; there’s Grok, which is pretty good, at least on the benchmarks; there’s Gemini; then there are Perplexity’s models; and there are all the Chinese models. There’s just more every day.

But I think the point you made—and I think an important point for people to understand here—is that these models are all leapfrogging each other, and no one is deviating by orders of magnitude. So we’re likely to have, at the end of this, at least—you know, Dave, what’s your guess? How many different large-scale frontier models are there going to be globally 5 years from now, or 2 years from now?

I think the natural dynamic actually is 1 at the end of the day.

But we’re in this golden era right now of a few months, maybe a year, maybe 2 years, where there are on the order of 10.

Why? Because ASI just depresses everybody else.

Well, we haven’t hit true self-improvement. That started about a month ago. It’ll go on for the next, say, 4, 5, 6 months, and so we’ll know if there’s a true singularity—a single winner—about 5 or 6 months from now.

So, you’re saying this is where GPT-5 pulled their punches and they’re using their capability for self-improvement?

There’s no doubt. Actually, that was a lot of research after that last podcast a week ago, because they definitely sandbagged. Imad mentioned it on the podcast, and we researched it, and it’s like, yep, there’s no doubt that a huge fraction of the compute is going back to internal self-improvement all of a sudden across all the foundation model companies. And they’re depriving the consumer experience of that same compute. The backlog of things they want to build and launch is just much, much bigger than the available compute. Well, we’ll find out tomorrow when we’re up talking to Kevin.

We’ll try. That’s interesting. The thing is, where can it self-improve, and what are the biggest bottlenecks? One of my points in that post was that the bottlenecks on AI—AI doesn’t do it end to end. It does a middle-to-middle right now, because the bottlenecks are prompting and verifying.

A prompt is like—if you think of a very fast spaceship, you could give it a vector on the 2-dimensional surface of a sphere, like a theta and a phi, to point it in a direction. And even if it were very fast, you still need to point it in a direction to get it to go there. That’s similar to a prompt: a prompt is like a high-dimensional vector. You’re pointing this very fast spaceship, this AI, in this direction, right? So, the prompt is one bottleneck, and specifying that is one thing, because obviously you can prompt it to do whatever, and then verifying the output on the other end.

I think for visual stuff, one of the points I made in the post is that for visual things, like frontend, images, and video, we have very good GPUs, so we can visually see whether something looks off. Visually. Go ahead.

Yeah, you know, this is the topic that really matters.

I want to get away from the idea that somebody at OpenAI knows the answer, somebody at Anthropic knows the answer, and they’re not telling us. It’s just not true. They don’t know either. Nobody knows, and we’re all going to figure it out very, very soon.

Nobody knows. If I had said a year ago, “Who’s going to win the AI race?” everybody would have said, “Well, it’s going to take a lot of NVIDIA chips and training-time compute and training-time thinking.” Oh, wait—Balaji just pointed out a really important point. It’s turning out that inference-time compute and reprompting are turning out to be more important than training-time compute. So, that’s a massive shift in who’s going to win and why in just the last few months.

Nobody knows how the next 6 months are going to play out. We do know that it’s massively self-improving, but we don’t know which components. Right now, Sam is racing to get Stargate done because he needs that $100 billion, soon to be $500 billion, worth of NVIDIA hardware to be pointed back at the self-improvement problem to maintain that “I’m ahead and I’m going to stay ahead” loop. China’s probably doing something very similar.

And we don’t know if that’s the winning move versus inference-time compute from AMD or from some other new source, versus just raw brainpower innovation on the algorithms. It’s all going to be moving very quickly.

This may turn out to be wrong in 12 months or a year, but I think I’m very skeptical that it’s just a matter of spending more money, because we’ve seen diminishing returns. These companies are highly incentivized to make a lot of money, and we have not seen radical improvements over the last year and a half despite a lot of spending.

In fact, we have seen very disappointing results. “Disappointing” is a weird way to put it, because it’s like a hedonic treadmill where this stuff is amazing, right? But I’m very skeptical that it’s just a matter of spending more money at this point. Number 1, and number 2 is the path the Chinese are taking with DeepSeek and other things is actually in part because of the chip ban or the chip restrictions, which have been partially lifted with this NVIDIA deal.

They took a chip-constrained view. So, they are using fewer chips to get the same—or less advanced chips to get the same result. The other thing they’re doing is a lot more in the way of physical AI, like deployment of physical robots. Self-driving is obviously working in both the U.S. and in China, but that kind of thing is where they are focusing a lot of their energies.

Insofar as some embodiment is necessary—and others have talked about this, like the feedback loops that you’re mentioning—it may not be just a purely digital form. The prompts on the computer may not be enough. Maybe you just need constant haptic feedback, walking around the world to perceive the world, build spatial intuition, and essentially do physics experiments. Let me pause there.

So, Balaji, one of the points that you make in that first point—and I think it’s an important one that I want to understand, and Dave, you may have an opposite point of view here—is that there is no single—let’s put it forward beyond AGI—no single ASI at the end of the day. Balaji, do you believe there are going to be multiple players out there that are going to be providing equal or similar capabilities compared with what you just said, Dave, which is that there’s going to be someone who will be supreme over everybody else?

Now, we’ve heard you. So, you believe there’ll be multiple players. How many players do you imagine?

Here’s why. I think that the hard-takeoff scenario was theoretically possible, right? But I think that at this point, first of all, there’s so much money, so much espionage, so much surveillance, and so much mimesis in this space that I think it’s very difficult for something like that to actually happen at this point. Number 1.

Number 2 is, what does recursive self-improvement actually mean? Is it just cogitation within a computer and the thing reflecting on itself?

It may mean new physics, new intelligence models, new capabilities.

I know, I know, but what I mean is—okay, let me be more precise. Eliezer has put out this thought experiment: imagine if there were a computer that could think for every second that you acted. It could think for a billion years and compute every move that you could possibly think of. You thought you’d do this, but it computed it out, and so on.

That actually does work for things like chess, deterministic games where you can enumerate the game-board state and things like that. But it doesn’t work for chaotic, let alone turbulent, systems, or for cryptographic systems where there are fundamental mathematical constraints on how far out you can forecast with finite-precision arithmetic.

You can easily construct systems. For example, you can have turbulent clocks, like how people will consume randomness and inject randomness for randomized algorithms. You could literally have physically turbulent clocks that can generate things that are mathematically impossible for an AI to predict. And so, I don’t see people—I mean, I should probably write a paper on this or something, right? But you can actually show mathematically that—I mean, it’s kind of obvious.

For example, if you ask AI what the solution is to the equations for forging a digital signature, or to invert a hash function and get the preimage, it can’t do that. That’s what it can’t do. Crypto is what AI can’t do. Chaos is what AI can’t do. Turbulence is what AI can’t do.

Because it can’t do those things, it cannot be this recursively super-self-improving superintelligence that just wins and dominates everything. Those phenomena exist in the world, and there are aspects of predictability where it’s fundamentally, physically, and mathematically constrained. Let me pause there.

Yeah. The first thing you said really resonates, which is not necessarily a race to the biggest GPU farm and the $100 billion budget. And we know that because we believe there’s at least a 100-to-1,000x performance gain in the algorithms themselves.

One of the things that drives my friends at MIT nuts, especially the academic types, is that the transformer algorithm driving all of this AI is not that complicated. And it frustrates the hell out of them because they’ve been researching all these arcane nooks and crannies for decades, and none of it mattered in what’s happening right now.

Yet it’s clearly intelligent. It’s clearly got an IQ of 148. It’s clearly coming up with its own ideas. If you believe there’s a 100-to-1,000x performance gain inside the algorithms, the narrow definition of self-improvement is: do we now have an AI Alec Radford that Leopold Aschenbrenner was talking about in his incredibly brilliant paper? Alec Radford, a researcher at OpenAI—the guy who comes up with the next idea and implements it. Can AI do that completely self-contained now? I think the answer is yes, plus or minus a few weeks.

I really disagree with that. Okay, so let me admit that that is potentially possible, but it’s similar to, for example, in 2015—let me qualify myself, and then let me argue with you a little bit, Dave.

In 2015, I remember writing a note internally, I think, at a16z, when the chatbot thing was big. If you remember chatbots in 2015, I remember saying, “What people seem to want from chatbots is iterated Google.” That is to say, the difficulty of doing chatbots in the 2015 format was that you could give a query, get back Google results, and then iterate on that. Such a thing, if it existed, would be more valuable than Google.

Now, I’m not saying that’s not possible, but that was so far beyond what anybody at that time was attempting with chatbots, which were just if-then-else decision trees. What it turned out, as often happens, is that the entire chatbot bubble helped inspire, among other things, Greg Brockman, who went on to do OpenAI. He’s written about how he was inspired by the chatbot thing as a technical problem to solve, but he took it more seriously than anybody else did.

My view is that, as of right now, unless there’s something that bridges System 1 and System 2 thinking—which is certainly possible, given that computers are now good at both—AI doesn’t have it. It still can’t tell you which is bigger, 9.11 or 9.8, right? Anything that’s spatial, logical, or deterministic—AI doesn’t have that verification. It’s just symbol generation, probabilistic. It doesn’t have another way of putting it.

With chemical reactions, if you have a chemical reaction that only goes 0.95, and you try to iterate 20 times, you have 0.95 to the 20th power. You have a huge loss.

Yeah. Exactly. Right.

The counterargument to this is, “People are saying the task duration of AI is improving and the benchmarks are improving.” But when you use it in practice, it’s just not there.

I’m going to deviate us into the next subject, but here’s an important point for us to discuss in the future. There are 2 outcomes here. One is that we have a polytheistic outcome: multiple AGIs and multiple ASIs. Dave, you’re taking the position that there’s going to be 1 primary ASI in the final result—1 digital god to rule them all.

We can make that a long bet, and we’ll see what happens. Dave, your timeline for when you think we’re going to know that is what?

I think most likely 6 months; out of bounds, 2 years.

All right. This is our conversation. I also wouldn’t take that bet, because I think government involvement is what’s going to force a somewhat polytheistic system. I think the natural dynamic is winner-take-all, but government involvement will hopefully preserve some balance.

And, Dave, your point is that there’s so much compute going into the recursive self-learning stuff now that, over time, 1 will emerge and shoot past everybody else. I’m telling you, I use this stuff every day, and it went from smart to freaking brilliant—mind-blowingly brilliant—in the last few weeks. There’s no doubt in my mind it can self-improve now. Just playing with it is incredible.

All right, I’m going to move on. I’m going to cherry-pick the points on your paper, Balaji. I have lost control of this podcast for sure.

Sorry. Sorry. I just want to say 1 thing. I think what is really underestimated is that AI is not good at time-varying, adversarial systems like markets and politics.

Today? So—

But there was just a paper published about AI’s accuracy in predictions.

That was extraordinarily good.

I saw that, and I need to dig into it. I just saw it in my peripheral vision today. But let me explain what I mean by this.

First, when you have a labeling of cats and dogs, the English language doesn’t vary that much. That conceptual label between the image and the text doesn’t vary that much. When you have the rules of a game of chess, that isn’t time-varying in the same way. But markets, politics, media, and so on are inherently time-varying and adversarial, in the sense that a post or a trade from a day ago, a year ago, or 10 years ago will not necessarily make money or get likes in the same way that it will today.

The moment that anybody has an AI there, they’ll start using it against other people. So you’ll have a bunch of AIs each competing with each other, right? And so, because of that—

You’re saying it’s self-centering?

Yeah, exactly. Another way of putting it is that there are such strong incentives to decentralize AI at this point, with all kinds of crazy corporate surveillance, corporate espionage, people refreshing the papers, and so on. How can you have that advantage beyond a few days or months? People will just copy it very, very quickly, which is what we’ve been seeing.

All right.

Peter, tell me to shut up, because in this conversation I’m just going to go for it. There’s so much here. Give me your counter-counterargument. I’m going to go to 1 of your next points. I’m going to cherry-pick from the 10 points in your paper. The next one is, “AI is amplified intelligence, not artificial intelligence.”

I think this is really important, because there’s a tendency to have tremendous fear because of the view that AI is out there and is something that’s going to compete directly with me and diminish my role and abilities on the planet.

Yes. Let me explain what I mean by this.

Right now, people think of AI as completely agentic, like it can do a task end to end completely by itself. For many tasks, it is really middle-to-middle. I want to clarify that self-driving has gotten to the end-to-end point. It is now literally end to end: it can pick you up and drop you off completely.

It took 20 years since the DARPA Grand Challenge, but it was solvable. However, that took billions of dollars, certainly millions of miles driven, enormous CAPEX, and a lot of time to get there. There’s also a lot of warning and transition because the real world has friction.

So I am somewhere in between. I don’t think AI takes all the jobs right away, because what happens right now is that you find yourself—or at least I find myself—in the actual reality of AI, without naming the tools. For example, there are a fair number of founders who will send over AI-slop slide decks, and I’ll be like, “It’s like the new Lorem ipsum, right? It’s Lorem AI-psum. Okay, AI ipsum, right?” It’s placeholder text.

It’s like the new midwit is a superintelligence. It’s a superintelligence, yet midwit, right? I mean, because it’s garbage in, garbage out. You get a midwit prompt in and you get midwit text out.

That’s what I mean by amplified intelligence. If you’re really smart, yes, you can prompt it to figure out new corners of physics or math. And, crucially, you can verify it by eye with your domain knowledge. I’m sure Terry Tao can search this and find interesting areas of math that they couldn’t find before, but they’re capable of looking at the stream of symbols coming back, parsing them, and figuring out if it’s just gibberish or not.

Other people will talk to it, and it’ll tell them it’s discovering all this great physics and so on. But without actually being able to do the experiment, do the calculations, and do the verification step, they’re convinced they’ve discovered some quantum-mechanical insight, and it’s just all—

Gibberish.

So, the point you made here that I think is very important for people to hear is that the smarter you are, the smarter the AI is working with you. The other half of this, which I keep trying to shout from the rooftops, is: ask great questions.

Yes. Basically, if you’re good at AI, it means everybody’s a CEO and everybody’s a manager, right? It’s a lot like interacting with an employee, especially a junior employee, and you have to check their work. You simply cannot shove it straight to production.

Fundamentally, as a CEO or a manager, you’re interfacing with the market—especially as a CEO or founder, right? AI is not interfacing with the market. So it can throw slop over the end because it only has to satisfy you. It doesn’t have to satisfy the unforgiving market. The market will just set it to 0.

And, Dave, that’s a great point you make: when you’re using a certain type of AI, and you have an improved model, you just slot in the new model, and you stay middle-to-middle.

Yeah, that’s right. AI doesn’t take your job. It takes the job of the previous AI.

Empirically, once you have a slot in your budget for an AI text generator, an AI image generator, AI video generation, AI code generation, and so on and so forth, what you find yourself doing is swapping out Stable Diffusion for Midjourney, swapping out the old Cursor for Claude Code, and swapping out this for that.

You really don’t want to use more than 1, honestly. Usually, right? Sometimes you’ll use more than 1, but what I mean by that is, once you get in the flow of Midjourney, do you want to switch over to Higgsfield or Google Veo? You can.

Actually, these are all built such that a prompt in 1 will basically work with another one, but they’re also subtly different. It’s a little bit like different social networks. They’re kind of similar, with the hearts, the posts, and so on, but they have some differences.

Usually, once you’ve invested in a tool, you don’t want to swap it out right away, but you can swap it out right away, right? So what that actually means is that these things are beating each other’s brains in for the AI budget.

I do think that every company will have AI teams, just like you have a social media team. Twenty years ago, it was considered, “Oh, my God, people just waste time on social media.” Ten years ago, it was novel-ish to have a social media team. Now it’s obviously a core function of every company that creates content for social media.

So every company will have an AI team that’s doing constant process optimization, verification of output, and so on and so forth. So it does actually create jobs as well. The other thing it creates jobs in is proctoring and verification because of what AI is doing.

Look, we all see the positive parts of AI. A lot of people only see the negative parts, and I’m not talking just about the job-taking part or whatever. I’m talking about the slop, the spam, the scams, what have you. So AI is going to radically increase the budget for proctoring and verification. Proof of humanity is only the beginning of that.

Yeah. Let me clean up one thing you said there because I think it’s really important for the audience. I think we are in this golden era right now where, as intelligent as AI is, it’s far, far better with a human interacting with it.

There’s something missing. There’s clearly something missing. Even though it can solve incredibly hard math and physics problems, it can do entire hour-long tasks for you and come back with near-perfect answers. But there’s still something missing in the creativity system, too. Something ain’t right. But that creates this golden moment of synergy between people and their AI agents.

The beautiful thing about AI is it can do 10,000 things for you concurrently. If you can evaluate the answers and cherry-pick the best one out of 10,000, you get incredibly good results. As an analogy, Midjourney—you mentioned you’re making a movie, and you’re trying to create the next scene and need it to be funny. If you prompt Grok 3 or something like that in Veo 3 and say, “Hey, make me the next scene,” you probably won’t like what you get back. If you prompt it 1,000 times, you get something incredible back.

It’s very easy to set up processes of self-evaluation and self-improvement, use another AI to evaluate the last AI, and create these loops. That’s where an enormous amount of compute is an incredible competitive advantage. That process is working really, really well for code generation, video game generation, movie generation, discovering new physics, and discovering new math. As long as you have evals that can test the results that come back, you can make ridiculous progress ahead of the curve. It just turns out that AI self-improvement is one of those types of tasks.

It’s just software. I disagree with you for the following two reasons. The first is that I think what’ll often happen is the evals will basically just get gamed, in the sense that it becomes like Goodhart’s law: when a measure becomes a target, and so on.

For example, Grok—I mean, look, I like it. Obviously, I love Elon. It’s amazing that he’s in the top 5 among models, given how many other things he’s doing at the same time. It’s honestly superhuman, how good xAI is, given how many other things he’s doing. However, when you actually try to use it on a daily basis, not everything necessarily shows up in the numbers. It’s fine, it’s good, and it’s really good for a Twitter search and so on, but it doesn’t obviously smash relative to Claude or what have you. They’re basically comparable.

One of the issues is that the space is so multidimensional that evals are like hairs on a very large head. It’s like a spike of functionality: this number is up, and this number is up, and this…. But the surface is so high-dimensional that the kinds of things that you want to do with it, the evals don’t necessarily capture. That’s one point.

The second point is that I think what I would argue you’re underrating is that those verification steps, those evaluation steps, are not cheap, and they’re not really that good for many things when you do a team of experts and stuff like that and throw a lot of GPUs at it. Grok’s super heavy, and so on and so forth. Yes, you can flip the coin. I mean, again, empirically, if I use Midjourney, I have to flip the coin 1,000 times before I finally find the image that I actually want. You get in the ballpark very quickly, but then the one that you actually want takes an enormous amount of time to get there.

It is not obvious to me that—maybe that’ll get improved—but I definitely don’t consider that to be solved right now. The verification step in the last few years, I don’t consider that to have radically improved because you still have to prompt the thing a lot.

Another part of it that’s an issue is the way these work: they don’t actually give a pseudorandom-number seed at the beginning to get a reproducible result. So you can give the same input and get different output every single time. What that means is you can only include it in those parts of your process that can tolerate that level of variance, which actually does constrain what—

Which can be a feature and not a bug in many circumstances. Just one narrow point on that: that is true for us as users of the APIs, but it’s not true for the researchers inside OpenAI, Meta, and elsewhere. They get perfect reproducibility. They give it the same random seed, turn the temperature to 0, and run the test 5,000 times. They can debug with exact fidelity. They just don’t make that available to us as outside users.

I know. So exactly, that’s right. But basically, for whatever reason, right now in the UXs, they have not exposed t = 0, and they’ve not exposed perfect reproducibility. I mean, you can share chats and so on. I guess my point is, there’s probably some model that you can call with that to get reproducibility, but when you’ve got stochastic output, let me give you an example.

Somebody had an AI web-browser-type thing. It seemed cool at first. They were like, “Wow, it can automate your posting for you and build a whole following.” I would not do that. The reason is, first of all, if it makes even one misstep, that’s on you, right? It’s like your dog biting somebody. You’re the owner. You’re responsible for this thing. It’s not actually a human being; you’re the human being responsible for it.

Number one, if it likes the wrong thing, retweets the wrong thing, or posts the wrong thing, which it could easily do. The second thing is that what it emits often sounds like the parody version of you, right? People who use AI chatbots—people have tried this, like replicas of people, right? It’s fine for voice cloning if you supervise it. It can help you scale yourself in that way at times.

I have an AI version of The Network State at thenetworkstate.com, but I say it’s AI, obviously, so that if there are errors, people know it’s AI, right? So it’s like an MP3 audiobook that updates with the text. It’s okay. However, that’s something where you can tolerate some glitches in the audio output, right? It’s an analog output. If it’s a digital output, you often can’t tolerate that error, like on the balance that’s being sent to somebody or something like that.

AI web browsing, for example—where is that helpful? It’s helpful for automated testing. It’s like a better Selenium, right? Internally within your company, you run AI web browsing, and you have to click a bunch of things like a monkey hitting buttons, like a human. Fine, you can tolerate the stochasticity there. Also, it won’t take down your live site. But people who are trying to say, “Oh, my AI agent will go and make me a million bucks, like Renaissance or whatever,” just get wrecked because it’s not good for that kind of thing.

I want to hit another one of your points here, and then we’re going to shift to crypto, which is, “AI doesn’t take your job; it lets you do any job.” I think that’s a really important element here that people need to understand and start to internalize. Why don’t you give your favorite examples?

Yeah. So, for example, you can get to a 6 or a 7 in art, in graphic design. People who couldn’t code can do prototypes of things. You can now do a prototype version in almost any area. It’s amazing for founders. It’s amazing for self-directed people. It’s amazing for people who have something in their mind’s eye but just couldn’t get it out, right?

It’s also amazing for people who have time on their hands but not money, because they can prompt and prompt and prompt and prompt and get a good image or a drawing for something if they lack the budget. But to polish it and actually get it to ship and production level, currently you need somebody who’s an expert in the area to actually be able to debug the symbols, usually—again, unless you’re smart enough that you can teach yourself with it.

So realizing that it actually more—you know, the Heinlein quote about a human being should be able to—

Chop lumber. I’ll get it wrong, but you know what I’m talking about, right?

Yeah, exactly. He’s like, a human being should be able to do basic things. You should be able to—

Yeah, exactly.

Go.

No.

No. Go ahead.

I'm paraphrasing: chop lumber, write a sonnet, make wine, sing a song, et cetera, et cetera. Specialization is for insects.

So, in an interesting sense, what I think AI takes us toward is a greater degree of robotarchy. Smaller communities can be more self-sufficient. That's what I mean by the internet versus China, going back to that: you can have a thousand million-person network states where you have robots.

It's effectively decentralizing across the board. Large companies used to gather all the talent and parse it out a little bit at a time. Today, a team of 1 or a small group can do whatever is needed.

That's right, so it reverses the extreme specialization of the 20th century. Your wingspan increases, right? There are lots of things you can do at an okay level. You can learn what you don't know, and so on. As long as you're self-critical about it—knowing that the AI can be wrong, you can be wrong, and so on and so forth—that concept of versatility is there.

With that said, for some areas that are genuinely just data entry or something like that—many kinds of midwit jobs—if you're just a lawyer using templates, or just a doctor who's reading from the book, AI can do it, maybe with a higher error rate, and actually in some cases with a lower error rate. In many cases, medical diagnosis AI has a lower error rate.

No, the numbers are staggeringly in favor of AI alone, without humans in the loop.

That's right. AI in medicine is much cheaper, orders of magnitude faster, because you don't have to go to an appointment, and so on. It's more personal. You can get not just a second opinion, but a 20th opinion.

For certain areas, it's a complete transformation and a radical, radical, radical improvement. Medicine is one of them. Law is one of them. Basically, you use AI, write all the contracts, and then have the lawyers review and sign off at the end, so you don't have to use the $1,000-an-hour lawyer, right?

Similarly with doctors, you research everything with AI. It's like an ultimate search engine, and then you go and get your prescription or whatever from the doctor after you've done all your research. They're the printout step, the final certification and verification step. That's how people are using it in practice.

But it doesn't mean that you have that MD or JD yet. That state certification is still a dangling thing. There's going to be enormous pressure from the AMA, the ABA, and others to ban or limit the use of AI—kind of like unions, but white-collar unions.

Yeah, I think you made a really important point for the listeners, too, which is that the different areas of endeavor are moving at very, very different rates. The 2 big factors are: can it evaluate itself, and is it regulated in some way?

Medicine is moving a billion miles an hour. Law is able to move a billion miles an hour, but you have regulatory issues. Physics, math, and coding are moving a billion miles an hour with no barriers. There are no regulatory barriers.

Give me a concrete example. Protein folding—what Demis Hassabis won the Nobel Prize for—that's genuinely something where there was a huge… That was a—

It was a well-posed problem: given the primary sequence, infer the 3D structure, where the 3D structure comes from the PDB, and you can verify it with X-ray crystallography. There is a very known objective function where you can map this far, right?

But what other examples are there?

The 4 things I named were medicine and biology, physics, math, and coding. And of those 4, the one where Peter's the world expert is biology and chemistry.

Yeah. Go ahead.

Within biology, chemistry, and medicine, there are areas that are very, very discrete and evaluatable, and others that aren't. That's the one of the 4 that I mentioned that's pretty murky. Within that, there's probably 20 different dimensions.

Specifically, if you're reading scans, the fidelity of a scan can be many, many terabytes of data. No radiologist can read it, but when you find something in the scan, it's crystal clear that you found it. That one self-evaluates very, very easily. It's like, “If I had the time to look at it, I would have found that, but I didn't have the time to look at it.”

Those are the areas that'll move very, very quickly. Then there are these other, murkier areas, like: can you build a whole-cell simulator? That's more of a synthetic-data and simulation problem. The AI will keep up with it if you can build the evaluator or the simulator, but can you actually build that?

Okay, so here's my view. The areas I know fairly well are biology and coding. I know some physics and math, but I haven't kept up with the frontier of them.

In biology, AI is amazing for the protein-folding problem, obviously. Another thing is biomedical text mining: all the papers that were locked in archives and so forth, being able to put together the information about p53 in this paper and that paper, all the contradictions, all the references, and so on and so forth. It's amazing for natural-language search, synthesis, and summarization of large corpora.

I feel it's definitely going to accelerate biomedical innovation in that sense, and obviously, as you just talked about, for diagnosis of both text and images and mixed kinds of things. It's smashingly useful in biomedicine.

But I'd say that a lot of that is not really the creation of net-new information so much as the synthesis and bridging of existing things. You've got a bunch of existing nodes, and AI is really amazing at interpolating. Sometimes that is, in a sense, net new, because that synthesis is itself useful, but that's distinct from points outside the envelope that are new physics, new math, or new biology.

Let me use this as a jumping-off point. Did you see Demis Hassabis's prediction of curing all disease by 2035? And did you see Dario Amodei's prediction of potentially doubling the human lifespan in the next decade as a result of AI? I'm curious if you've seen any of those and what your thoughts are.

I've seen those.

Yeah. So here's what I think about that. First of all, obviously these are smart guys, and I love what they've done. I'm a big fan of both Claude and Gemini and everything they've done.

First is, AI is below the state and crypto is above the state. What I mean by that is crypto, with Bitcoin and smart contracts and so on, is a replacement for the Fed and the SEC. Crypto is actually able to have the political strength to take over the government and deregulate not just crypto, but AI.

AI is below the state.

Break when you say “below the state”—you mean controlled by the state?

Correct. That's right. Like above the state, below the state, right? Crypto is above the state; AI is below the state.

For the moment, there is very much a potential future on the post-state side of—

Potential future.

Yes.

Yes, that's right. There's a potential future where you have autonomous law and so on. But right now, as it currently stands, the reason I say that is that a lot of the AI people, as much as I like them, essentially assume a benevolent or at least functional state.

For example, they think of the United States versus China as a unitary America. It's got a functional government with high state capacity, and it's making rational acts, as if it were the Cold War or something like that, versus a shambolic blue America versus red America versus tech guys versus everybody yelling at each other and doing adversarial things to each other.

In the same way, the cure-all-diseases claim—if it's not engaging with the FDA, if it's not engaging with the regulatory process, with CPT, with the AMA, with all that kind of stuff—I have seen for 20-some years that the primary blocker on genomics and biotech is the regulatory state.

But we're also talking about 8 billion humans across the planet who have the same biology. You can get to geographic arbitrage: if a solution comes up, I can go to my friend President Bukele in El Salvador and have the treatment there.

And, in fact, longevity tourism has been going on for a decade.

Absolutely. Prospera, for example, actually has Minicircle, which Altman and others have supported, and I've put some checks in there and so on and so forth. Minicircle is actually gene therapy on existing humans.

Farbod Nivi has posted about it—how he took Peter's post about it, and he feels stronger. Bezos started following Farbod because of that. He saw Farbod's strength and so on.

I'm still waiting for the science to get published there. It's a lot of people's opinions, right?

It's a lot of people's opinions. But the thing is that we're going back into the era of big-effect-size drugs.

To explain what I mean by this, essentially, the entire FDA era—if you go prior to the FDA, you have Banting and Best, who won the Nobel Prize, right? Why did they win the Nobel Prize? Because they came up with the idea of insulin supplementation to treat diabetes.

They tested on dogs. They tested on themselves. They tested on willing patients. Those patients leapt up out of bed, and then they scaled production with Eli Lilly for the entire North American continent. They started their work in 1921 and scaled by 1923.

That was when pharma moved at the speed of software, before the modern FDA. In fact, many drugs are grandfathered because they were in wide use before the FDA, and you just had trial and error over many years figuring out how they worked. Crucially, that was not phase 1, 2, and 3 clinical trials. That was not case-control studies, right? Case-control studies are like A/B tests. There’s no case-control study on the effect of regulation itself.

The regulatory state.

Go ahead. I want to make a very important point. I just want to make a point.

Yeah.

Okay. You don’t have a jurisdiction with and without some regulation to see whether the regulation is actually holding something back, right?

Or actually saving lives, right? And the whole point is that you value the lives that you don’t kill versus the lives that you don’t save.

And tort law is totally messed up. If you make one little mistake, it’s a $100 million settlement. If 1,000 people die when they could have been cured, nobody pays anything. It’s just totally—

Exactly right. Yeah.

That’s right. So the FDA—the fundamental issue actually gets back even upstream of the FDA—is risk tolerance, right? At any big company, risk budget is always scarcer than budget, right? It’s often easier to spend a huge amount of money than to take a risk because people are just—it’s just one of those things. You have to have founder DNA to have the political capital to take a risk. If you fail, what would happen, right?

This is why Bezos doesn’t believe in bet-the-company bets; he believes in lots of small bets, right? But you have to have the psychology and the legitimacy to do that. The decentralization of the world and the arising of all of these startup societies, network states, new freedom cities, and so on and so forth will mean more jurisdictions where you can actually opt in and have a new system of laws—all the special economic zones and so on and so forth.

That’s actually the bottleneck. Not so much—look, I love AI; it’s amazing. I love biotech; that’s amazing. But it’s really the regulatory state that is the thing holding this back, right?

The issue is that a lot of AI guys—not you—are effectively docile. For example, Leopold Aschenbrenner’s paper writes about the U.S. and China as if it’s the United States rather than the disunited states.

You know, he writes about it as if Gavin Newsom isn’t fighting, Trump isn’t fighting, and The New York Times isn’t publishing “Abolish the Supreme Court and pack the judges,” as if it’s not a food fight—a social war—every single day, right?

If you actually showed the political chaos in the backdrop, he’s like somebody talking on the eve of—

China’s Warring States era—that they’re going to launch a space program.

You talk a lot about Bitcoin. We’re going to get to crypto in a moment. There is a direct correlation between people who say they’re Bitcoin maximalists or Bitcoin advocates and longevity. I’m curious, in your decade-long vision of the future, how much does longevity play into that? How much does expanded health span matter at all?

It’s incredible. I used to have as my Twitter bio: “Immutable money, infinite frontier, eternal life.” All 3 of those things go together, actually, because, for example, we’re not going to be able to explore the stars without life extension and probably radiation-tolerant genes and other kinds of things, right? Or sending robots and plugging in.

Yeah, you can send robots. You can do other things, but to terraform Mars, we’ll probably have to modify ourselves. You’re just not probably—

Or send the robots first and build the environments for us little meat sacks to live in, right?

Yeah, that’s also possible. That’s right.

So maybe we can figure that out. It might be some degree of robots, some degree of self-modification. For example, there are various groups on the face of the Earth that have different altitude adaptations, right? Kenyans, Andeans, and Ethiopians each have different altitude adaptations. Some are in the blood, some are in the lungs, and so forth.

You can imagine stacking those together, getting radiation tolerance and disease tolerance, similar to how AI synthesizes from different nodes. You can take different genetic mutations. There’s a company called Variant.io that looks at that kind of stuff—looking for real-life X-Men, taking their mutations, and adding them.

But coming back, what is the connection? There is a deep connection, by the way, between Bitcoin and life extension.

And actually, it’s gone back a long time. The fundamental similarities are as follows: The traditional financial system says that, yes, hyperinflation is bad, but a little inflation is good and deflation is also bad, right? That’s to say, losing all your money is bad, but the New Zealand target is 2% devaluation of your currency every year, and deflation is bad. Then people would go to money hoarding and so on and so forth, right?

Mhm.

And so what that basically says is every year you lose 2%—a little bit of wealth, right? The traditional medical system says, “Yeah, rapid death is bad, but trying to live forever is also weird, right? Reversing death is also weird.” Instead, you should die a little bit every year.

Mhm. You lose a little bit of health every year. Okay, so that’s the deep similarity: The traditional financial system says you lose a little bit of wealth every year, and the traditional medical system says you lose a bit of health every year. None of them really look to first principles and say, “Why? Again, why do we have to do that? Why do we go—”

Why is there a hole? Why is there a hole in our bucket?

Why is there a hole in the bucket? Right? Because you guys have seen the phenotypes for the mouse longevity stuff. You’ve seen the stuff. We have case studies with the bowhead whale living 200 years and the Greenland shark living 500 years. The issue is how we repair our software and hardware to get there. It’s not a matter of if; it’s only a matter of when.

Yes. I think the big thing we have to work on is making the moral case for it.

And also boosting the risk tolerance. For example, people right now will go bungee jumping, they’ll go skydiving, they’ll serve in the military, and they’ll take risks with their lives. Some of those risks are foolhardy, and some of them are for the collective, in a sense, right? We should have medical heroes.

Similar to how the early airplane pilots were considered heroes for risking their lives: No plane crashes, no planes. No train crashes, no trains. There was somebody—you guys are all of a certain age. You remember the CRC Handbook of Chemistry and Physics?

I do.

Yeah. Oh my God.

That’s what existed before we had the internet.

Walls of them. Walls of them.

Well, exactly. We all had that in high school, right? The CRC Handbook of Chemistry and Physics. If you recall, up to a certain point, the compounds had a smell-and-taste list, and then they stopped listing them. Why? There was some guy whose last thing he did was sniff cyanide. He wrote down “almonds.” He took a hit for the team. Hero of science. That’s why we know what it smells like, or whatever.

The scarce resource that we need, in my view, is risk-tolerant people—

—and risk-tolerant jurisdictions.

Let’s tie it back to Bitcoin. I put out a tweet last night asking my followers to propose questions. I got 225 questions back, so we’re going to start with number 1 and march our way through. But seriously, Bitcoin and the connection to longevity—and then, when we built the internet, we built it without a financial layer. We now have crypto as a financial layer. Do you see Bitcoin as a primary currency for AI and machine learning in the future? There are a few questions stacked there, so let’s start with Bitcoin and longevity as a correlate.

Okay. You know how power generation works: You have very high voltage at the power station, then there are step-down transformers, and then it comes a little trickle out to power your phone over a cable, right? In the same way, Bitcoin is like really high-voltage money, okay? It’s transferred relatively infrequently. It’s like digital gold, right?

Then you have various proxies for Bitcoin that you can just blast on-chain. Roughly speaking, Bitcoin is already at the transaction volume of Fedwire. Think of it as the Fedwire of the free world.

Okay? On top of that, you can have various layers. For example, Bitcoin scales, in a sense, off-chain and on other chains. I know people will say Lightning; let’s leave Lightning aside for the sake of this argument. Many Lightning implementations are, in practice, essentially peering agreements between Lightning apps.

So it’s similar to off-chain Bitcoin transactions on Coinbase or Binance. These apps have tens of millions, in some cases hundreds of millions, of users. It’s like transacting within France or within Germany—between yourselves—very, very quickly.

And then, when you want to transact between Coinbase and Binance, they can do settlement of Bitcoin—the true-ups between them. And then, of course, you can withdraw from the system as well, right? That kind of hub-and-spoke system is not quite the same as a total peer-to-peer system, but you can get extremely far with that.

I think you've deviated from the primary question, though.

Yeah, let's go back.

Because you said, “Will it be the currency of the internet?”

Yes.

The issue is that the transaction volume for pure Bitcoin, BTC, is not sufficient for it to be the currency of AI. But WBTC—wrapped BTC, off-chain BTC—can get there. It's the same analogy you have in the real world: if you want to pay with gold, it's a nightmare. As you get more and more toward lighter transactions, you can pay with a debit card, but you don't have great security, so there's a trade-off there. You can do many, many more transactions with a debit card. They tend to be small transactions, and you can do them all you want. When you get to the heavier stuff, you're dealing with bigger transactions that are more heavily regulated and more secure.

You wrap BTC on Solana, you can blast it on-chain as much as you want. It can be the currency.

Exactly.

One question from the community is: if you're willing to say, what percentage of your net wealth is in Bitcoin right now?

Not willing to say.

Okay. But would it be a major—would it be a majority?

Well, that would be willing to say. Let's just put it like this: only hold in USD what you can afford to lose.

Okay, fair enough. Where else besides Bitcoin would you hold?

In terms of USD, is it your hot wallet?

Nobody asked me this. Huh?

USD is your hot wallet.

Well, in terms of fiat currency, by process of elimination, I would say SGD or maybe AED, meaning Singapore dollars or the UAE dirham. Those are the two jurisdictions. I would not say Switzerland anymore. The CHF—the Swiss franc—basically, after the Russia sanctions, the Credit Suisse debacle, and Obama kind of kicking in the doors of Swiss bank accounts, Switzerland ain't Switzerland anymore, right? So the Swiss franc is not what it was.

I wouldn't hold Western currencies in general, nor would I hold the JPY, and probably not the Korean won. But if you must hold fiat, I would say the AED is still pegged to the dollar. The SGD is soft-pegged, but they're independently managed on their own ledger, to my knowledge. They're also global currencies, and they have open capital accounts. I think Western countries are going to have closed capital accounts soon. Even in 2026, it's being proposed, like a remittance tax for the USD. So I'd say my advice on fiat would be SGD.

But if the dollar—you're essentially holding the dollar.

Well, for now. But a lot of the East is moving to gold. One of the things that's happening is that BRICS states are stacking gold. Have you seen the Korean DMZ, where North Korea and South Korea are?

Yeah.

Imagine—this is not exactly how it is—but imagine that on the Russia-China border there was something like that, and they would just hand each other gold bricks each day, right?

Yeah. They'd actually do physical settlements. It doesn't actually work like that, but conceptually it works like that, where states can settle back and forth in gold bricks, right? They can actually afford the security, the trucks, the magnetometers, and all that kind of stuff to verify that the gold is real, that no one's coming to steal it, and that it can back fiat currencies because it used to back fiat currencies.

I do think that gold—both gold and digital gold—is a currency of states, and digital gold is the reserve currency of the network. Those will both rise, but Western fiat won't. Those currencies are currently dollar-backed.

Now, with that said, this could happen. First, all newspapers went online, and then all local papers died because their geographical monopoly went away. The New York Times and The Washington Post survived, and then they basically died against X and internet-native social media. That exact kind of thing is going to happen to currencies. First, all currencies go on-chain—fiat currencies—and then most of them die because the only reason they exist is their geographical monopoly. Once you can hold the best currency in the world in your wallet, why do you have to hold the currency of your geography again? You don't.

Currencies start competing on ideology and features rather than geography, just like online news competes on the basis of ideology and features as opposed to geography. If the Miami Herald, the San Francisco Chronicle, and the Poughkeepsie whatever are all reprinting the same story, that only worked when they had trucks and a geographic advantage. All local currencies compete with each other. Most of them die. USD, RMB, and maybe a few others make it, and then all of those get devalued against BTC.

All right, let me take this back to AI. In a world where AI agents are negotiating, contracting, and paying each other autonomously, what do you see as the role of crypto in that?

It so clearly is the currency of AI, and the reason for that is that crypto is the currency of the internet.

For example, when you set up a program, you can have that program access private keys. One way to think about it is that crypto is a 10×, 100×, or infinity× improvement over the existing system on many axes that people don't normally think about.

For example, you can set up a bank-account equivalent—a crypto wallet—in fractions of a second. You can set up as many as you want, and you can bind those to whatever you want, right? None of that can be done with a traditional bank account. You have to go to the branch and present your ID. So, to simply set up a program or a robot and give it a wallet takes fractions of a second. Just that practical thing alone, forget everything else I'm going to say, means obviously crypto is the currency of the robot and of AI, right?

Then, going further, legacy fiat is locked down in so many different ways. You don't notice this when you're just buying Starbucks or whatever as a mezzanine-level transaction, but if you want to do transactions that are very large, very small, very fast, very automatic, very international, very transparent, or very complex, you need crypto.

Will these banks make it, or will they all be displaced by crypto-native banks?

They're already getting displaced. It's a good question. They're all becoming crypto exchanges. Crypto exchanges are the new banks because they actually have peering relationships across borders.

How many years before we see the current major banks no longer around? When do they become Macy's?

I think you have to distinguish between retail banking and the other types of banking. Retail banking really gets—

Yeah, I totally agree. I was at The Washington Post working for Don Graham for a year, right when the internet and Google were just taking over the world, and there was no turning it around. That was kind of tragic, and all investigative reporting died in the process.

I don't see it the same way, though, with banks, because you're talking about M1 here. Everything you said is exactly right. It's going to happen, Balaji, just the way you said. But banks don't just sit there as a place to hold M1. They have a lot more: they make loans to consumers, they evaluate their credit. There are many types of banks, but very little of what they do is impacted by the transition that you just outlined.

And so it's going to happen the way you phrased it. But, yeah, what's the subset that's impacted?

But I still want to hear your predictions on when we're going to see the current banks we all know and love—

I won't name them, you know, disappear.

I mean, disappear—like Macy's is, I think, still around.

Yeah, it's still around.

And become minor players.

No, irrelevant.

I would argue that, among power users, it's already happened.

Sure.

Right. In the sense that you can have crypto. Crypto is being used in Thailand, for example. Thailand just switched over to accepting crypto for various things.

Everybody, Fountain Life is now accepting crypto for payments on your longevity treatments.

So, by the way, China is the exception to everything I'm saying, right? Because China will maintain control, like the Great Firewall. China's system, the CCP, and BTC will be like the USA and the USSR of this century, because they're centralized-party total control over everything versus the decentralized currency.

2030? 2035? I'm still looking.

2035, I think. But let me explain why I think it comes faster than you might think.

I'm going to keep his hand on the stove.

Very short answer: I wrote another article on Balaji.com called “The Billionaire Flippening.” A long time ago, Olaf Carlson-Wee, a founder of Polychain, and I—he's been a friend of mine for a long time—calculated that somewhere between $100,000 and $1,000,000 per Bitcoin, half the world's billionaires become crypto.

Yes.

Okay. And it basically eats everything else of value in their portfolio.

That's right. The next 10x dilutes down all fiat billionaires.

Yep. For sure.

By a lot.

And including—

That's what people don't get. Like—huh? Go ahead. Including states.

Including states. Including banks.

Yes.

Right. The next—

Slow this down for everybody. I want people to understand what this means. So, if someone's got a million dollars of crypto in the bank and they have a million dollars of other assets, crypto is 50% of the portfolio. All of a sudden, if it goes 10x, crypto is 90% or 95% of the portfolio.

That's exactly right. And the purchasing power of it—at a certain point, what happens is it destabilizes, and it goes from 1 million to 10 million because people flip. That's already happened in the crypto space, where all investments are denominated in Bitcoin terms. Do you see any future in which crypto fails in any fashion, or Bitcoin fails in any fashion? Are we way past that right now?

There are definitely ways it could fail. Let me say how it could fail. First is quantum. Let me talk about quantum cryptography. That's an area to watch. The main issue is that it's not quantum-safe—it’s a very technical issue—but quantum-safe encryption can be deployed in Bitcoin. Moving all of the funds would take a long time, moving them on-chain, so there would have to be some sort of hard fork or something like that.

This is the current state of development, as I understand it. I'm not in the bits and bytes, so I may be misspeaking, but there would have to be some significant protocol change to enable that—like a giant airlift out, a deployment of something that significantly changed the code. Miners and everybody would have to consent to that. They probably would, but that would be a thing.

With that said, crypto as a class will not disappear from this Earth because there are now enough different currencies, with enough different protocols and enough different mechanisms, that there's a degree of herd immunity. There's enough different immune systems that the kinds of vulnerabilities proof of work has with Bitcoin are distinct from, and actually complementary to, the vulnerabilities that proof of stake has.

With proof of stake, you can have a bunch of humans distributed globally. In theory, if they colluded, they could rewrite the blockchain. Proof of work requires physical data centers, but they can't rewrite the blockchain. So, there are different failure modes. You can also do things where you hash Ethereum's state to Bitcoin.

I think crypto as a class of things will not disappear from the Earth. Any individual chain could be attacked, but you can snapshot that chain and issue it on another chain.

Right, so you can have lifeboats.

I think the core here is that once you understand the Byzantine generals problem, that innovation can't be undone. It's here.

Yeah.

That's right. Look, it's also, I think, very possible to see crypto, up until this point, as something like social media. Do you guys remember The Social Network movie?

Sure.

Okay, of course. It's worth rewatching for one reason: It came out in 2010, and what's remarkable about the movie is its complete absence of politics. There is absolutely no mention whatsoever of politics. Facebook is treated as if it were Universal Studios, like a theme park and amusement park for people to poke and like each other.

This was despite the fact that the Arab Spring had already happened, or was happening, and there was destabilization of governments and so on and so forth. It was known that Twitter and Facebook's role in the Middle East had at least partially destabilized things. I may be getting the timeline wrong, but there was definitely stuff in the late 2000s, before—I think it was Iran or something—the tweets must flow.

There was stuff before The Social Network movie came out, and obviously all politics became social media in the years after that. Then there was a counterattack on social media, and every possible attack vector you could imagine on social media that people thought was unthinkable due to social conventions came out. People were deplatformed, censored, and shadowbanned. The president was deplatformed, and then he was replatformed. The whole thing became a battleground, as we all know.

Yet in 2010, if you had told me, even with 500 million people on Facebook, that in 10 years the single most important political issue in the world would be whether or not the president of the United States could tweet, that would still seem like, “Oh my God, Balaji, you're taking the internet too seriously. You're just exaggerating.”

It’s similar to you guys at Singularity University. People will always say, “Oh my God, you're taking this stuff too seriously,” or whatever. With Bitcoin, by 2035 or so—maybe sooner, maybe later—just like you're not the president if you don't have a social media account, you're not a country if you don't have cryptocurrency. You're not solvent, and you're not able to operate in the accelerating exponential future.

Yeah.

Bitcoin becomes the government of governments.

Yeah. I want to hit on networks. I want to hit on the Network State and the Network School. You have basically gone out and built a global talent fellowship in the Network School. You've launched this, and I'm curious about your thoughts around dark talent—the idea that there's this incredible bidding war going on right now in the U.S. for AI talent. There are million-dollar salaries, $10 million, and hypothetical billion-dollar offers. There's so much talent out there, and it's become so easy for this talent to gain access to capabilities. Talk to me about what you've called dark talent, if you would.

Totally. I believe in the America that I grew up in, and I think that we're forking: The internet is taking American values, and the land is becoming something very different. It's going anti-immigrant, anti-capitalist, and so on and so forth.

I don't think that, as an immigrant capitalist, you should build your life in the U.S. I think it's going to get very, very nasty, unfortunately. Moreover, a lot of people can't get visas. They can't get in there and so forth. So, I moved as far away as I could—10,000 miles away.

When did you move?

2020.

Okay. Yes, go on.

I'm just giving the motivation. I believe in America in the same way that America was Britain 2.0. I think of the internet as Britain 3.0 or America 2.0. It's like the next version, version 3.0.

Why? It's 10x bigger than America with crypto. It has freedom of contract and freedom of speech. An Indian may unfortunately be a second-class citizen nowadays in the U.S., but they're a first-class citizen on the internet. They have the same monetary policy, contract rights, smart contracts, and all that kind of stuff.

I think that lots of Westerners, lots of Americans, also believe in that. So, in my own way, I'm trying to stick up for what I think of as American values and kind of keep the torch aloft. Think of me as a minority report. Maybe—hopefully I'm wrong. Maybe all you guys can fix it, and that's great. Salute, and good job.

But I'm like the parachute. I'm the minority report, the backup plan, the hedge. I want to keep the flame of freedom, democracy, capitalism, and so on alive. Why do I even say democracy, by the way? I think lots of Western states are effectively becoming one-party states.

In California, you can hold elections, but the party always wins. The Democratic Party always wins. It's being gerrymandered even further. In response, Republicans are doing the same in Florida. And, of course, communism in China has a one-party state. So, Democrats, Republicans, and communists have all built one-party states. So, democracy is going to—

Bring me back. Bring me back to—

I know. I'm coming back to it. One second. Yes. So, they built one-party states. Democracy is between those states, right? Meaning you're going to go from the two-party system to the thousand-community system.

To go to the thousand-community system, you need to build one community and then scale that around the world. So, we're a startup society that's a template for other startup societies. We are bringing in all the dark talent that has been overlooked from Latin America, from the Middle East, and from Eastern Europe.

Also, by the way, white kids from the Midwest, Asian kids who get crowded out by Harvard or something like that—anybody, anywhere. If they're working hard, they're online, and they're doing well, I will bring them in, fund them, and level them up.

So, discuss how many people you bring in, how are you funding them, and what’s your mission? Are they building companies there?

Well, right now it’s out of the Bank of Balaji, and that’s fairly healthy, right? Eventually, we may take outside capital and so on. Paul Graham did Y Combinator out of his back pocket for a long time, right?

I think that’s always the right way of doing things. If you can bootstrap it, things have worked out well for me, so I want to help others up and do that for them as well. You asked where the people are coming from, and the answer is all over—basically anywhere English is spoken.

My feeling is that we hear about everybody in San Francisco and all the talent wars that are going on there, but I have to believe that throughout China—which, by the way, has 50% of Meta’s AI researchers, who are Chinese—and with this crazy amount of intelligence capability centered in math in most of Southeast Asia and India, not to mention South and Central America, there’s a huge amount of talent that is not being tapped into that could easily create the next unicorns.

Not could—but they will.

They will.

Yes.

They will. Now, the thing about that is, absolutely, and one thing I’d also say—and we’re going to be rolling this out in the near future—is that maybe some fraction of people, a small fraction of people, will be founders and they’ll create unicorns. I agree with that. They’re very important, and we’re going to find them.

However, the number of people who are unicorn founders globally is less than the number of pro athletes, right? There are roughly 1,500 to 2,000 unicorns. Say 2x, with 2 co-founders on average per company, so a few thousand unicorn co-founders. There are easily 10,000-plus pro athletes, depending on how you count.

Having said that, something fascinating is the Thiel Fellowship. I think 5.6% to 6% is the number of Thiel fellows who have become unicorn founders, which is insane.

Yeah.

It’s amazing.

It is possible.

So, my view is that, certainly, I do fund companies.

Yeah.

And I do fund currencies, but what I’m most interested in is the third thing: communities.

I think 1994, with Netscape, was really the first true internet startup, if you remember that moment, because it had all the modern features. You could download the software on the web, it went viral, and it had a billion-dollar IPO that transformed the world—all those features of the modern internet startup. A multibillion-dollar IPO in the ’90s, when that was really real money.

Fifteen years later, the first internet currency: Bitcoin. That was a template for things that came after that. I think what we’re doing—the way I think of it—is creating the first real internet community.

So, internet company, internet currency, internet community. Why are internet communities going to be even more valuable than internet companies or currencies? The reason is that they’re actually the platform on which those things are built.

Just like China is the platform upon which all their companies were built.

Exactly. That’s right.

Exactly. Very good point. I did a spreadsheet, and it was the first time I’d ever written one quadrillion unironically.

Okay, here’s why. If you go—obviously, Google’s a trillion, Facebook’s a trillion, on that order of magnitude, and so on. That’s thought of as the most valuable thing you can do, right?

But I’m pretty sure cryptocurrency—Bitcoin—is going to get bigger than that, because it’s actually still in its early phase. Those tech companies can still grow beyond that, but if you think of a country and what the country’s valuation is, how would you put a valuation on that?

If you take, let’s say, Singapore, and you take its number of square miles and the price per square foot, you do get a valuation. Or you take the number of people and their average net worth, and you get a valuation on the order of 10 trillion.

Okay, that actually seems reasonable, right?

Then, if you say, okay, what’s 10x Singapore? Maybe South Korea-ish is about 10x the size of Singapore. It’s a little less than 10x, but let’s say it’s on the order of 100 trillion.

Then what’s China? It’s like 10x South Korea, or will be, depending. So that’s like a quadrillion, right?

I actually do think that all of the startup societies together—when you have 1,000 million-person network states, and they may not all be 1 million people; some might be 20 million, and some might be 500,000—that is actually the next big thing. That is actually what the startup societies are going to be. We’re going to have a lot of those in 10 or 20 years.

I want to bring this down to your advice for entrepreneurs right now and people who are employees in America. They’re listening to this, they’re excited about technology, but this hasn’t been their lives so far.

Between the entrepreneurs out there and people just listening who are tech fans, what are your top points of advice as we wrap here? Great. Before you answer that question, Dave, I know you’ve got a hard out, buddy, so thank you for your time. I’ll call you on the phone in just a little bit when we wrap here and continue our conversation.

Balaji, a lot of fans out there are going to listen to this podcast probably 2 or 3 times to be able to absorb it, just like I am. If you could, what’s your most heartfelt advice to both entrepreneurs and the average individual listening to this?

For entrepreneurs, first I would say: go direct. Build your own following on social media, get good at AI, create content, and don’t give interviews to legacy media. Only do tech podcasts and tech shows, but even more than that, basically build your own influential individual voice as a creator, as the authentic voice of your company.

Don’t hire public relations people. Only hire creators if you’re hiring them. You don’t have to hire too many, but basically, don’t outsource your content creation. Content creation is as important as code, and you should be doing that in-house. You should be using all the modern tools, and you don’t have to use full AI, by the way. You can use half AI.

So, go direct. That’s number 1. Number 2: move to Florida or Texas. Get there as soon as you can, if you can. Get to a place.

Do not be where I am, sitting here in the People’s Republic of Santa Monica.

You asked me before.

I know. I love it. Yeah.

Move to Starbase, maybe, where Elon is. Essentially, take it from a company town to a portfolio town and a tech town. If I were in the U.S., I’d be in Texas or Florida—specifically Miami or Austin, and near Starbase—and cluster with other folks there.

I think property rights have a better chance of getting protected there. Number 3: if you have cryptocurrency, get your keys off exchanges. Get your coins off exchanges. Do that now. Figure out your cold storage. Figure out all that stuff, and don’t talk to anybody about the details of it, right?

Okay, number 4. You can—I’ll hit that one. You’re concerned about that? You think exchanges are not going to make it?

I’m just saying that you don’t want to—I don’t have any imminent thing, but I’m just saying, do that now.

There are 2 things: exchanges can be hacked, and governments can take them over.

Yes. Both. Exactly. If Bitcoin really, really runs—if it really runs—then at a certain point people basically start getting scared, and all kinds of crazy things are going to happen, right?

So be in a state where property rights are respected, where Bitcoin is respected. For example, Texas—I think the state legislature had something like, “The right to hold Bitcoin shall not be infringed,” right? That’s a good place to be with your cryptocurrency, and so on and so forth.

So, A: go direct. B: move to Florida or Texas. C: crypto and cold storage. D: get involved with your local tech community.

Tech is a community. Thinking of it as a community and doing as much offline social activity as possible—I almost don’t have to say this because people kind of already do it organically in tech, right? But tech is a community.

For example, that might mean not simply homeschooling, but crowdschooling and microschooling, where you’re educating your kids together, right? Start doing collective action, things with tech people in your spare time, because you’re going to want community in the years to follow, right?

You’re going to want people you can rely on in the physical world, not simply the digital world. That’s why I’d go also to those places.

Finally, I would say, read as much as you can of BRICS content in translation, okay? Because unless you’ve actually read it and traveled—if you have not been to Riyadh, if you haven’t been to Dubai, if you haven’t been to Shanghai, if you haven’t been to Bangalore, if you haven’t seen how much better the rest of the world is getting and has gotten—

At a meteoric rate over the last 5 to 10 years.

Yes. You aren’t calibrated unless you’ve been to Shenzhen, Bangalore, Dubai, and also Eastern Europe.

Yeah. Right. Go ahead.

Yeah. Hopefully Russia gets to peace, and the Russia-Ukraine thing gets to peace. I think Moscow has also improved a lot since the Soviet era, right?

But Eastern Europe, Russia, India, China, Dubai, and Riyadh have radically—as you know from FII, right? You and I have both been at FII—really calibrated what the world is, because movies will not show you that.

The only movie that's shown rich, non-Westerners in recent times is probably Crazy Rich Asians, or maybe Squid Game, right? They're not really showing it, right? So you have to get the on-the-ground experience. Otherwise, you're not calibrated.

All right. For the general public, you hit that for entrepreneurs. Much of that probably still holds for the general public as well. Let's go to the individual who's in a job in L.A., in San Francisco, or in middle America. What's the most important thing that they need to be thinking about?

Okay, so first, are they good at math, or are they not good at math? If they're good at math, then they should get offline for a month and totally lock in, get Schaum's Outlines, self-study, maybe use Brilliant.org, and actually learn the details of computer science and statistics. Do whatever LeetCode problems they need to do. Learn everything that you need to learn to actually understand how to implement gradient descent. Maybe do Karpathy's stuff. Do fast.ai, and then you can also learn on the crypto side of things as well, if you're interested in that. There are various courses on this. Maybe I'll teach another MOOC on this. But basically, one month of really all-in self-study.

Focus, focus, focus.

Focus, focus, focus. Get offline. It'll be there when you get back. You won't miss it. Use pencil and paper in a totally offline, Faraday-cage environment.

This is a big thing I do, by the way. I try to spend as much time offline as I am online. You know how, with cars, I've mentioned this before, we advance cars because we have electric cars and self-driving cars and even flying cars, but we also have walkable cities. One of the great things San Francisco did over the last 30 years was take the highway away from the waterfront, right? And so cities have been made more walkable and more green—the Big Dig in Boston, right? So you're both pro-car, but you're also, in some cases, for the minimum necessary car. You take the car away in certain cases, right?

So I'm pro-Internet. I'm pro-AI. I'm pro-crypto. All that stuff, but I'm also pro-offline, right? Pencil and paper, coffee, quiet, focus. It just can't be overstated how important that is. I think we're going to build buildings that are Faraday-cage buildings. You know how you go into an elevator and the phone turns off, right? We'll build cage buildings that have phone lockers at the entrances, where everybody gives the speaker their full attention.

We're actually doing some of that stuff at Network School now, right, just to prototype it. It's great. It's almost like freeing—that people can just be free to focus, right? They're free to be offline because the Internet is so addictive. I would say offline time is important, focusing is important, and the right balance is important. It's kind of like getting a great night's sleep so you're active during the day, right? The better you can focus offline, the stronger you'll be online, and vice versa, right?

If you don't own crypto right now, begin.

I mean, look, I never give anybody financial advice in that sense, but as a technology and as a platform, yes, it's taking over, right? Every bank is adopting it, and so you should definitely understand it.

It's kind of one of those things: Sometimes people will see something in the news, they'll see that it says it was a scam or whatever, and they tune out of it. Then they come back years later like, “Wait a second, it actually developed a lot.” The dot-com crash and then Web 2.0 were like that. People didn't pay attention to the Internet over the 2000s, and the whole thing got built back up, right?

So that's a related concept. When you see something in the news, especially if it's an asset—and again, I'm not telling anybody to buy any asset—a useful thing to do is set a calendar reminder for it. Then, 30, 60, 90, 120, 150 days or whatever later, you can make a purchase or whatever decision, a research decision, when it's not in the news.

Because when it's in the news, it's usually overpriced or underpriced because it's making a big move, right? When it's outside the news, you're not reacting on emotion, and you can reallocate your position accordingly, right? So that's just something I think is useful to do. You can be hyper-rational about it, especially if you write down what your long-term thesis on it is.

Related to that, don't gamble, right? Unfortunately, there's a lot of day-trading-type stuff, and I know people like to do that, but I think it's just too hard to do. Just buy and hold for the long run if you're putting something into something.

Let's see, other things. I would say realize that social media is like watching pro sports. For example, when you watch tech founders or something like that on X, it's like watching really, really successful athletes on TV. If you've ever dribbled a ball, it's actually really hard to dribble a ball and run and dunk, and they make it look effortless. They make it look so easy to do that, and it's really, really, really hard, right?

So anybody who is building something—first, you yourself should be a builder, but second, you should also, in general, not attack other builders, right? It's always easy to do that on social media. It's a cheap way to make oneself feel superior. “Look how dumb Elon is. I can't believe he missed a spot on that rocket.” There's a lot of people who do this kind of thing. You know what I'm talking about, right? So if Elon isn't successful, nobody's successful. If Elon isn't smart, really, there are very few people who are smart. Okay, and so—

That's a great point.

You know, so you have to be able to give respect to others to get respect for oneself, right? You don't have to flip to the other extreme, but game recognizes game, right? Respect people online and try to build. I know X is just a player-versus-player environment. Everybody yells at each other all the time. There are sites like Farcaster and so on that have a healthier and more civil atmosphere.

So try to seek out online communities as well that model the civility that you want to communicate to your friends, your family members, your loved ones, and community members. Don't behave badly; at least try to be a relatively good example. I recognize, by the way, that sometimes you have to defend yourself online and that it's a dangerous environment and all that kind of stuff. But try not to be the person who's doing the first strike. Try to make it a better world for people in that sense.

Yeah, I think that's a good list. Maybe it's not 5 for each, but it's roughly a pretty good list.

Yeah, I appreciate that.

A couple of comments here. I think it's awesome that you're building the network state movement, just because it creates a whole new potential at the edge, and as the existing system fails over, it fails over more elegantly, right?

That's exactly it. Have you read Foundation?

Yeah, of course. Isaac Asimov's Foundation series.

Yeah, exactly. So, like Terminus, right? This is Terminus, where the goal is to reduce the duration of imperial collapse.

Yeah. There's something you said in that podcast that really blew my mind, which was that less than 4% of marriages in the U.S. are between Democrats and Republicans. That really shook me because it shows the depth of the tribalism. How do you come back from that? I think we have to—I mean, one of my fears, and I think it's already happening, is that in the 2010s it was blue and tech against red; in the 2020s it was red and tech against blue; and I think by the 2030s, if not sooner, it's going to be red and blue against tech.

With socialists and nationalists against internationalist capitalists, and some of the backlash to AI, where there's slop and scams, and some of the backlash to crypto. There's now backlash to bio and IVF. There are people who are getting mad at data centers. There are people who are mad at Apple for making phones abroad. Elon, unfortunately, people are mad.

When we start prioritizing data-center electricity over residential electricity—

That kind of stuff, you know, all of this is due to regulations and other kinds of things. But the tech guys—I mean, for example, doing AI in the bluest city, in the bluest state in the union, and making billions of dollars while saying you're going to put a lot of blues out of work—is not a good medium-term to long-term strategy.

And so my view is that tech as a group has to actually realize this, and it has to kind of realize that the bad version is the tech scapegoating, where tech is scapegoated for the decline of the fortunes of red and blue, because tech guys are very publicly being rich.

They’re rich online. Oh my God, they’re so rich, right? And a lot of people aren’t, and they’re getting hurt by inflation and so on and so forth, right? So, Balaji, I would love to have you come back on and continue this conversation. I’ve got a hard out for a board meeting.

Last thing to say is, I think there’s hope. I think the internet can lift the boats for everybody. I think we can get to an age of abundance and all of that stuff. I’m fundamentally a positive person in that way, and I just think we have to get through some of the whitewater that’s coming up.

It’s going to be, if nothing else, an exciting decade ahead.

It needs 3 listenings. I didn’t give any links; I just gave verbal citations, but you can look them all up.

All right, buddy. Listen, thank you for your time. Thank you for your brilliance.

Are you going to be at FII this year in Riyadh?

I will.

Okay. Well, good. I’ll see you there, then.

I’m actually giving a talk as well, so we’ll see each other there. Great.

Awesome. All right, pal. Be well.

Thanks, guys. All right. Thanks, everybody.

Good stuff.