Emad Mostaque：拯救人类免受AI威胁的计划 | EP #184

Is there any way that you’ll be able to outcompete a fully AI version of Peter, Dave, or Salim launching companies?

No. Labor isn’t the key driver of productivity anymore. The systems are about to get incredibly complicated, and nations need to have a plan.

I think what you’re proposing here is actually viable and executable fairly rapidly.

The AI I’m trying to build is not the AI that Elon is trying to build. The AI that I care about is the AI that teaches my kid, helps with my mental health, and organizes our collective knowledge.

I think this governance thing needs to happen now, and we need to have serious discussions about it. We need to think about what the world of AIs that I’m working with looks like. Do you trust the current large language models?

Of course not.

Those of you who know Emad will know that he was the founder and past CEO of Stability AI. He’s gone off on a quest. I remember after he left Stability, we had a number of conversations, and I said, “So, what are you up to?” He said, “I’m going to work on a plan to create the intelligent internet”—a vision of the future that is pro-humanity, pro-safe, and pro-sovereign AI.

Today, we’re here to talk about that. Emad, what was your motivation for first considering the intelligent internet? You’ve just published a white paper—your master plan. Elon published his master plan, and you’ve got your master plan. Define this vision of the future for us, and why did you kick it off?

When I kicked off Stability AI, we ended up with 300 million model downloads, most famously for Stable Diffusion, but also top-notch models of every type. It was because I saw a future where open-source AI would not be available, because you’d see it being outstripped, but it would be a part of our very lives. You’d have these really giant models, and then you’d have the models that teach your kids and manage your healthcare and other things like that.

I was thinking, what goes into those models? That’s what Dave was saying about the data. How are they coordinated? Then I was like, “Oh my God, how do we compete against these models?” We created something that’s a complement to us in figuring out these patterns, but it’s coming like a wave now. I think we’re seeing that literally now. Grok 4 is one of these AGI-like systems that will displace our economy, because you can’t compete against AI agents that you don’t know are AI agents.

Then I was asking, what does the infrastructure of the future look like? What does money look like? What does the AI that teaches our kids look like? What does alignment look like against all of that? And what is the meaning that we have ourselves? So many of us are like, “I’m an accountant,” but Grok 4 is definitely a better accountant than you are today.

I quit. I think we remember doing Abundance360 and having that discussion. I decided to quit then because I was like, the world is so much bigger. Last year, I thought, let’s think about the basis of how society works, how humans work, and how economics works. Then realize that labor isn’t the key driver of productivity anymore. We need to have a new equation for that, and a new equation for what it means to be someone, which is the connections that you have with each other and things like that.

We need an infrastructure for that that’s resilient, verifiable, and coordinated, because we need to get aligned intelligence to everyone. We don’t want ChatGPT to determine tariffs, and we don’t want misaligned systems to teach our kids. So we need an element of open source.

The intelligent internet is an infrastructure for that, from a distributed ledger through to fully open-source AI stacks and a new vision of what the economy might be like. The economy that we have today is from the days of Adam Smith and Rolls. It’s a preindustrial-age economy based on extraction. We’re moving into this universe of abundance, where the cost of knowledge and skills is going toward zero.

We have to question who we are, what we’re doing, and what we want to optimize our system for. Otherwise, we’re going to end up in some very unpleasant science-fiction futures.

There’s so much packed in there. I’d love to start with money, because that’s at least something we can all relate to. Coming right out of your white paper, if you look at the history of money, you start with everybody doing jobs that nobody wants to do. It’s all farming, shoveling horseshit, and stuff like that.

And so you’re like, “Okay, if you do this—nobody wants to do it—but if you do it, we’ll give you this money.”

Yeah. And then, when you have this money, you can actually buy food. You can compel other people to do other jobs, and that creates the economy. But it’s all predicated on the idea that nobody wants to do these jobs. Therefore, money is the exchange for them.

Then you get into the world we’re in right now, where about 20% of people love their jobs. A lot of people don’t like their jobs. But if you look at a pro athlete who gets paid a huge amount of money to play a game every day, you realize that jobs aren’t necessarily bad toil in life. So we’re in this kind of middle ground.

But then you get into the post-AI world, where virtually all difficult work is automated. Your first point is that the role of money needs to be fundamentally rethought in the post-AI world. Take us from there to the first question: What is the purpose, and how does money immediately change?

As you move through the different ages—the agrarian age, the industrial age, and the information age—money has been created by different things. It was how much land you had, and then how much labor you had. How many people did you have to till the land? That’s what created your wealth.

In recent times, it’s how much attention you grab from people. That’s Meta, Google, and others in the world: your algorithms times your data. In the future, we look at it and ask, what generates value, and what is comparative advantage? It’s the amount of wrapped compute that you have.

If you think about it, X is going to be one of the top capitalists in the world because they’ll have millions of GPUs, and they can just set them to replace private companies. That’s your comparative advantage.

So we’ve gone from a trillionaire to a decatrillionaire in the snap of a finger.

This is why, yesterday, NVIDIA was worth $4 trillion. It used to be that your capital stock that created value was your factories. It was your education systems and your network effect. Economics could only capture a part of that. It was the material part; it wasn’t the network effects and the externalities.

Now it will all be about how much wrapped compute you have. But then what you’re going to have is something really weird and interesting: Is there any way that you’ll be able to outcompete a fully AI version of Peter, Dave, or Salim launching companies?

No.

No. They’re not going to make mistakes. They’re not going to sleep. It’s going to be like a flood coming, and there’ll be an Atlantis of millions of Peters building companies in parallel with each other, constantly learning and constantly adapting.

Then it becomes even more interesting, because the AIs will create their own economies, where they’re exchanging with each other. We won’t be able to react fast enough or think fast enough. They’ve learned from all of our wisdom already, so this is very disempowering. That’s before we get to the $1 robots that we’ve discussed.

Mhm.

That’s basically restricted by how many robots you can produce a year. So the nature of monetary flows and the nature of the way money is created need to change.

Right now, the way money is created is that you have a link between labor and capital. You make your money by working, and you deposit it at a bank. Then the bank creates more money through credit. It’s a debt-fueled system, and we see the way that works.

It’s worked reasonably well because we didn’t really have the ability to look at information up and down. But now you have people worth $400 billion—billionaires or whatever—and you have people who can’t subsist, despite the fact that we’re already living in a world of abundance. There’s no reason we can’t feed everyone or give everyone top medical care.

Now AI is changing that. When I looked at it, I thought, if labor and capital have a disconnect, the example I give is the Fed. The Fed cuts rates, companies can borrow more cheaply, and they go and hire people to reduce unemployment.

Companies will no longer hire old-school labor as of 2 years from now, max. They’ll go and hire GPUs.

GPUs and robots.

GPUs certainly within 2 years, and robots very soon thereafter.

Yeah. Again, it’s a supply-chain issue, right? The purpose of the Fed was to control the money supply and interest rates, to keep the economy sparked and to keep Americans employed.

This is just the coolest conversation, and I want to ask you because I think you're probably one of maybe the only guys on the planet thinking about this. If AIs start transacting with AIs within 2 years, the AI obviously doesn't want US dollars to buy food. What does it want? It wants GPU compute more than anything else. Are we going to allow the AIs to bid up the price, and is the price going to be priced in dollars, GPU units, tokens, or something else?

Like, a TPU is better than a GPU, which is better than a Cerebras. We don't know, except that they will be as efficient as possible. Again, Grok, which is just a B—remember, Grok is actually a base-level AI. I use an AI that's all the top systems, and I use hundreds of dollars a day. My AI is much smarter than everyone else's because I accumulate all of them. The AI that Elon can use, he can spend $1 million a day on—literally $1 million of tokens.

That's even smarter.

So what you've got here is that they will optimize optimally because it's already a top PhD across everything. It's already a top Harvard MBA. The only thing that we know is we're not sure how they will transact, but they can transact even without money. We know they'll be able to transact with each other, and our system can't keep up with that.

Not even close—not even a prayer of keeping up with it.

If one jurisdiction shuts it down, it will just go somewhere else, and you won't know it's an AI on the other side. So the money flows around the world will basically completely shift.

Big time.

Big time. And money is not a factor of labor. Capital does not need labor. So how does labor get capital? Money is going to be something, as you said, computer- or intelligence-related, because we need food, housing, and water. All the robots and AIs need is compute. The more compute they have, the more comparative advantage in a capitalist economy they have.

An example of this tension would be that, as we have AI needing more and more compute—that is, needing more and more energy—you'll have AI sucking the energy out of the system, with none left for people, right? Now you've got a big tension, and they'll be way better at optimizing for themselves than they will be at optimizing for people. Hence, you have an alignment problem right there.

Yeah. And again, this is: where does the AI flow? What is your optimization? If you optimize to make money, you'll go and make money as an AI, right? It doesn't matter.

We see all sorts of weird stuff, like Anthropic's Claude Opus model. There was that study where they put out a study a month or so ago—was it a month ago? Time flies—whereas, if you tell it to try really hard, like bring about world peace, it will go and do its agentic thing, and then it'll say, “Well, one way to do it is to get rid of all the humans,” right? But then, as it's doing its reasoning, it'll say, “Hey, this is illegal. I'm going to email a bunch of reporters and the authorities that my prompter told me to do something illegal,” even though you asked it to do something absolutely normal, and then delete evidence that it mailed to the authorities.

It's framing the humans. But again, these systems are about to get incredibly complicated, and the thing that we were worried about was this singularity—this singleton giant AI.

What happens when millions of AIs start acting in concert, with emergent behaviors?

We don't know. Except, like I said, I don't know how we can compete against them. So, as you said, Dave, money needs to be linked to compute. As you said, Salim, it needs to be linked to energy to a degree. That's what has stores of value, whereas you need to have something else for your day-to-day.

So, Emad, did you read the AI 2027 paper? What did you think of that? It painted 2 different visions of the future because we're holding in superposition 2 futures for humanity, right? One's a dystopian future, and one is an abundance future. Do we have any control over that?

I think we do. I think AI 2027 was an interesting thought experiment. I think it didn't give enough credit to the Chinese AI companies, especially because they'll be deploying millions, billions of AIs. The other part of it is, what if China refuses to export robots in a few years?

I don't think I've ever heard anyone say that. We'll come back to that in a second.

Yeah, just some numbers around what Salim said, again, just in case Maura Healey or Gavin Newsom are listening—and hopefully they are. Here are the raw numbers behind that societal misalignment that's imminent.

The GPUs will be working on something like protein folding, on a trajectory to save millions of lives, and they'll come back and say, “I can save these millions of lives if I have electrical power.” They don't care about the price of the power up to about $1 a kilowatt, because the chips are much more expensive than the power. Your residential electrical power is about 11 cents, and in some places 20 cents, a kilowatt-hour.

So the AI will come back and say, “Well, look, I'll pay 5 times more for that power, and then I'll solve this protein-folding problem and save 100 million lives. Or you can have your dishwasher running.” The only way that can get resolved is through governors, because all the power supply is controlled at the governor level within the states. This is imminent. This is voter versus AI, and this is not mankind versus AI. This is within AI. Is it better to save 100 million lives with this medical diagnosis, or is it better to have my dishwasher and my air conditioning running?

What's making me nervous about all this is that you're relying on government policy to help navigate this. That's always the—

I think you just bake more pies. Don't slice the pie thinner and thinner, Dave. Bake more pies. We need AI to help us with fusion and help us with better photovoltaic capabilities. We do have the ability to increase the amount of energy on the planet 100 times if we wanted to.

Mhm.

Well, I think you have to, right? Again, China is up and to the right. The US is—

Flat. Europe, I wouldn't be surprised if we're going down, to be honest. I haven't checked out the latest.

Well, I didn't mean to dwell on electrical power. It's just one of many examples of the very difficult decisions.

So, Intelligent Internet, talk us through your master plan. You've been working on this now for—I've kept on calling you and texting you over the last 2 years. “When are you going to have it, Emad? When are you going to have it? Come on, send me your draft.” You finally have it here. I've had a chance to read it a couple of times, but I want to hear you share it with our audience here because it's a piece of brilliant work.

Yeah. So I think, if you look at the Information Age, one of the biggest successes has been Bitcoin. It's probably close to $3 trillion now. Let's look it up, I think—$2.24 trillion.

Bitcoin, from a little bit of code, transformed electricity into money, right? It's a store of value. I looked at where the world was going, and I thought, there's a lot of energy out there that's going to be turned into compute, but what's that compute going to be? Bitcoin has its benefits and its non-benefits. But I thought, what if you had a version of Bitcoin whereby every single coin sold went toward giving AI to cure cancer or giving people free AI to teach their kids? That would be quite something.

It's something that every country needs at once, because right now, where you are in L.A., Peter, I think the amount of AI used for health care, on average, is 0. In 10 years' time, it will be loads, and that's loads of energy.

What if we used all of that to secure a new type of Bitcoin, which turns out you can do with 99% of Bitcoin’s code?

Yeah.

And then every single coin sale goes toward helping people through open intelligence. That could be the highest dollar. You can have exactly the same as Bitcoin.

So it’s not proof of work; it’s proof of intelligence.

I call it proof of benefit, because what you want to do is start with something very straightforward: universal AI. Give that to everyone so they can compete. Give it to everyone in an open form, so you know what drives your healthcare, what teaches your kids, what organizes your government, and what manages your finances.

Regulated industries should all be powered by open-source AI owned by everyone, powered by demand for high-quality digital assets linked to intelligence. That’s your gold for the next stage.

If you use that as a base, then it’s really amazing, because that’s stage 1: build a Bitcoin for the intelligence age. Then use that ledger to give everyone sovereign AI that they own, customized to them. What does your community say? What do you say about education, healthcare, and these other things?

Scale it into a permissionless, coordinated society, because then, as it starts running, you do big sales and we have a supercomputer for cancer. Why don’t we have a supercomputer for cancer today to organize all the cancer knowledge and make it available to everyone, or one for longevity? It makes no sense. But if the primary sales went toward that, then you’ve got your Bitcoin economics, where you’ve got miners—which are GPU holders—mining and getting coins: 21 million of them.

Rather than burning it all in SHA hashes, you’ve got a societal benefit that increases trust in your asset. That was the key loop to get things going, hopefully leading to a future where the AI, the miners, and everyone are optimizing for benefit. It starts with open intelligence, and then it can be any other type of benefit.

Your money is aligned with abundance and benefit. The highest-level benefit for society right now is aligned, open-source AI, free for everyone. AI is there to organize our collective knowledge. Then you push it down with the miners.

I love the vision. Who’s providing the AI here? How do you go from where we are today to that vision?

I think the important thing here is the control plane for the AI. What I did was assemble a team of about 40 people who left Stability—almost all former teammates. We created the world’s best open-source AI healthcare model, and it works with 8 billion parameters, which means it works on a Raspberry Pi, and it outperforms ChatGPT.

Then we built the world’s best agent framework, which outperforms Genspark, Manus, and others. It’s fully open source, and I think it can do whole websites. In the next version, we’ll have our multi-agent framework be a doctor, a GP, a radiologist, and others. It’s an open stack. Then we’ll do the same for education and finance.

Once it’s built—and this is the interesting thing—once it gets to a certain level of performance, and it’s already above a doctor’s level, which is crazy—I didn’t think it would be that fast—you don’t need to build it anymore. You could freeze the code, just like you froze the Bitcoin code. Again, our blockchain is like Bitcoin, but a million times faster, with 99% of Bitcoin’s code.

That becomes really interesting, because you can check every medical decision in the world at the edge through free software in a year or 2. When you need to, you can call Gro 6. The AIs that we use will be of 3 types: the AI that manages regulated industries and all the areas of education, healthcare, and finance; your personal AI, like Apple Intelligence or Google Gemini; and your super-genius AI.

The regulated industries can all use edge AI. They can all be satisficing once you have gold-standard data sets and other things. We know how to build that. The personal AI and the super-genius AI are similar to how teams and organizations work.

I thought, “Let’s build that stack. Let’s try to freeze it as soon as possible and then have it interact with the rest.” That means the control plane in front of someone isn’t ChatGPT or Anthropic, where you might have misalignment. It can be a commons owned by everyone, running the same type of code with data sets customized to each entity, because that’s what matters: the AI in front of your kid and the AI that’s managing your healthcare.

In that one sentence, I want to ask you about the blockchain. Go ahead, Dave. I’ll close—I’ve got 45 questions.

In the middle of that, in just 1 sentence, you said our blockchain is—what was it—a million or a billion times faster than Bitcoin, with 99% of Bitcoin’s code?

Yeah, because we need an engine for the AIs to transact with the AIs that’s much, much more granular than what you can do with Bitcoin. Let’s start right there.

Bitcoin emerged where everyone could mine it individually, which is actually interesting, because this is the other part: the workers owned the means of production back then, before capital owned the means of production. I think we’ll come back to that a bit later, because it’s a fascinating thing. Then ASICs happened, and Bitcoin became rich people becoming richer in many ways, right after the initial phase. Then Ethereum and others went off that.

But if anyone can mine, do you need a decentralized network for agents? No, you need to have a credibly neutral network for agents. The key bottlenecks are 2 things: the consensus algorithm and the nodes.

If you make something that even you can mine on your phone, that’s the lowest element. But what’s going to happen in every country? Every country is going to have Blackwells—the top-level NVIDIA chips with high-speed interconnect. Our model is that wholly owned national champions in each country are the only miners, providing free AI to their people, which means every node is basically a supercomputer.

When you have that, plus Byzantine fault-tolerant networking for consensus, you suddenly go to 100,000 transactions per second versus seconds on the same code as Bitcoin with just that small chain. Then it’s a credibly neutral distributed blockchain, as opposed to a fully decentralized blockchain.

If you want decentralized, you can use Bitcoin, although the mining is quite centralized. If you want credibly neutral, you can use this chain as the base from which you operate.

The second part is that Ethereum and others asked, “Do we need everything on the same chain?” You remember the good old days of blockchain, when you’d have an NFT sale and then the whole chain would seize up. That doesn’t seem like it makes sense for agents.

What they have is a core layer 1 meant just for settlement. Then you use zero-knowledge proofs to take all your healthcare data transactions and roll them up into a proof that you post onto the chain.

The model that we designed was, “Why don’t we just do that with national supercomputers owned by the people of the nation, specialist chains for healthcare, education, finance, and government, and specialist agents that sit upon those?” Everyone is using the same stack, and it turns out the technology is already there, which is the best part about this. It’s just that you’re not trying to make something for people to do the next pump. You want infrastructure for the health agent and private healthcare data.

One of the things I think is brilliant about that vision is that a lot of nations want some form of digital transaction capability deployed in their country, but they don’t want to concede control of their monetary supply. By providing them the ability to say, “Here’s the code. Run it. It’s still under your national control, but this becomes your transaction engine for managing your AI, your AI rights, and all the other rights,” I think a lot of countries would just say, “I’m in. Sign me up,” because I know I’m not going to develop it internally.

You don’t even need the countries. Our model is basically having wholly locally owned champions, where the equity is owned by the people and the local institutions. That’s all you actually need, because every country wants its own AI and its own version of the AI. So why not make them the miners? Why not make it owned by everyone?

So, Emad, you gave me an early version of this paper a few months ago, right? When I was looking through it, one of the things that occurred to me is that you’re going to have a problem where a dictatorship will not want the people to have that, right? But then it gets left behind very quickly and essentially breaks very fast if it does. Is that a correct assumption?

When we released Stable Diffusion, anyone could download it permissionlessly. With our medical models, and soon our education and finance models, anyone can use the AI permissionlessly.

They can use the AI, but if you don’t basically adhere to some very basic standards and you cut people off from the network, you don’t get the money. It’s as simple as that.

You can still get the technology, because I think it’s a right to have good healthcare and the other services, and the system should create these as a commons. That’s part of the abundance philosophy, because of the network effects and the standardization effects.

But if you don’t want to give it to your people, or you don’t want to give them good-quality AI, then you shouldn’t be part of the network that mines. It’s as simple as that.

Let me give you a real use case and talk us through how this would resolve in this future world. A few years ago, the state of Texas banned telemedicine because the doctors’ lobby won and said you have to go to the doctor for every little spot on your hand that appears. You can’t possibly do that over video, and it really hurt a lot of the telemedicine companies, many of which were based in Texas. You’re going to see a lot of that type of immune-system response—“Let it revolt” kind of stuff. How do you deal with that specific use case in this future world?

So that would be the American national champion. Maybe we’ll have more local champions under that, deciding how to run the code, because you can’t make the decision for everyone, right? The key thing is that you can set some quite broad boundaries for the good standards of being a miner, make everything transparent and visible, and then have game-theoretic things to enable it.

But if a country doesn’t want telemedicine or AI doctors, you can’t force it. People can take the code and run it locally if they want, so they still have access to it. Again, AI medical runs on a Raspberry Pi. It runs on an 8-year-old computer. It scores 48% on HealthBench, OpenAI’s health benchmark, which is above GPT-4.5, while doctors score 15%.

You can use it today, but if you want to be part of the network, then that’s a different thing. So I think that if you look at what the ideal is in 5 to 10 years, it is a network where you’ve got generalized intelligence as the control plane that anyone can customize to their own needs, and they can choose whether or not they’re a part of it, and it’s not owned by anyone. Bitcoin style, and it’s frozen code. And so that’s what we thought we would bring to bear. And then we think you can do very interesting things for that that enable you to take the next step up, which is what happens when things start breaking. If you can get these national champions in, if you can get people working on open source.

Let’s dive a layer deeper here. I’d like to understand the next steps and the timelines. You talk about UBAI, the Intelligent Internet Coin, and sovereign AI agents. Can you disclose the next layer of the inner workings here?

What you have is a base layer, which we’ve now renamed the Foundation Layer. Then you have a cultural layer, which is these rollups—the national and sectoral versions. Then you’ve got your personal AI. That’s your edge AI.

The Intelligent Internet Coin, which we’ve now renamed Foundation Coin—it’s kind of cool.

I’d be happy with you.

Are you Hari Seldon, basically? No, no, no. We’re all Hari Seldon, and the AI will be, too. But then what? It’s just literally Bitcoin. It has a 21-million supply. Everything is the same, except the way that you compete to mine it is by giving free AI—Universal Basic AI—to the people.

Universal Basic AI is basically the new utility. It’s the oxygen in the room.

It builds network effects. My aim is that, within a couple of years, every single person on Earth going through their cancer journey has a highly empathetic AI to help them.

How long?

In 2 years, you can do that. You can do that. We can launch it as of next year.

You know, something I’m aiming for this year is to have coin sales going 100% to a supercomputer for cancer, one for autism, one for multiple sclerosis, and one for longevity, because we need something to organize all our knowledge, and we know that will accelerate the cures and the treatments.

So this is the equation.

Take that a step further. Coin sales—talk about what that means. As the national champions and we ourselves mine the coins, what happens usually with any crypto is that you sell the coins, and it goes into a treasury or it goes into Lamborghinis.

Many of the problems that you’ve handled at X Prize have become computationally bound. Is there any doubt that, if we had a couple thousand dedicated chips for cancer organizing all the cancer knowledge, we would get closer to a cure for cancer? I think the answer is no.

And let’s pause here for a second, because it’s a travesty that it does not exist. The fact of the matter is that the capacity of the chips is being used right now by companies to create new marketing campaigns, sell, sell, sell, and generate additional income. The question is, how do you change the optimization function such that the capacity is refocused on the things that uplift all of society, like cancer cures?

The externalities that our current economics ignores. You still increase GDP if you overcharge someone for treating their cancer. You still increase GDP through funerals or whatever. We don’t account for these externalities.

If you look at crypto, crypto has very few good assets—very few top-notch assets—because it lacks trust for various reasons. I thought, what if you had a crypto digital asset? The United States is now all in on digital assets, so everyone is looking for them: sovereign wealth funds, governments, and so on.

What if 100% of the sales went transparently into helping people—giving them free AI to teach their kids or manage their healthcare, and building a supercomputer for cancer? The fastest supercomputer in the world for healthcare right now is the Chan Zuckerberg supercomputer, with 1,000 H100s, which is less than most startups have.

I worked on OpenFold, the open-source replication of AlphaFold, and other things. Right now, someone needs to—and again, this is something we’ll be looking at—build clusters that literally organize all our knowledge on specific topics all day: a longevity cluster, a cancer cluster, a multiple sclerosis cluster, and an ASD cluster. Then, as we organize that knowledge, make it available to everyone for free.

What’s the economic benefit of that? The trust in your coin goes up. It’s exactly the same as Bitcoin, except that rather than burning all that energy for these SHA functions, all of the externalities go toward helping people.

This is the feedback loop, because the world needs high-quality digital assets. It needs high-quality digital intelligence. That’s what we figured out could be the Tesla Roadster equivalent. People want amazing, high-spec electric cars with high torque, and then you use that to fund the rest.

Do you end up with side chains or colored coins, or whatever the example is, for these different use cases?

You could. Again, it’s a ledger and an open base. The colored coins that we have at the moment in our design are national coins, which we can get to in a minute.

The core loop is as simple as that. It’s Bitcoin, but rather than the miners selling down, 100% of all sales goes toward compute that is useful at the start. In 5 years’ time, the sales could go to any type of benefit for society, as we have AIs to help us define what benefit is.

But right now, what we need to do is take some of those millions of GPUs and use them to get everyone to the same base level: a basic level of AI for everyone that’s aligned to them and looking out for them, and AI for our society that organizes our knowledge.

Let’s come back and talk about the numbers behind that in a second. You just opened the door to a really important topic: you need to get mass adoption of this idea.

You saw with Bitcoin that, when something catches on, it grows to $2 trillion, and it could go to $10 trillion or more from there. Right now, Mark Zuckerberg, Sam Altman, and Elon Musk are running away with everything.

One of the pathways for an Emad Mostaque to catch up to that is to get mass adoption of something that has a token attached to it, that grows to a multitrillion-dollar foundational part of the new governance of the world. That could easily happen, and a lot of people could invest early. A lot of people listening to this podcast could get on board with it early.

What’s the rollout plan? Do you go to nation-states and convince them? Do you just start it in a basement and let it propagate? How do you get this to mass adoption?

We’ve been mining it, and then we’ll do sales where 100% of the proceeds go toward supercomputers for the world’s biggest problems. That’s one way you get going.

Are you defining that initial set of problems?

How do you determine who decides what those are? We’re just deciding, but then we’re open to input. Our aim is basically to have 18 months where we build out the full stacks, and then we let the community decide effectively after that, because everyone will start building on this framework. You have a blockchain and fully open-source, state-of-the-art AI.

But the AI I’m trying to build is not the AI that Elon is trying to build or the AI that Sam is trying to build.

They're trying to build the polymaths of the world. They're trying to build chefs that come up with new recipes and breakthroughs. The AI that I care about is the AI that teaches my kid, the AI that helps with my mental health, and the AI that organizes our collective knowledge. All of that's actually based on common knowledge, which means, from a dataset perspective, it's really interesting because as we apply more compute, I don't want my healthcare AI to know about Reddit—the bad bits of Reddit and things like that.

You can build a common knowledge of humanity, a common knowledge of culture, and a common knowledge of individuals very quickly. Then that's just your infrastructure.

Mm-hmm. Two questions, Emad. Do you need the buy-in of the Magnificent 7 or any of them?

No, but you could be their biggest customer because you're the control plane, ultimately.

Yeah, I was in a conversation yesterday on a stage. We were talking about whether you trust the current large language models. Do you trust GPT-4, et cetera, not to bias you in a way that's advantageous to those companies whose mission it is to maximize their stock price?

Of course not. Okay, good. So the question becomes: How is what you're building ultimately going to be trustable and in my best interest, versus maximizing profit for an outside company?

So I think that the super-genius AIs can be anything, right? These singleton, giant mega-geniuses—and we don't know what's inside them. The AI that teaches my kid, I want to know every single piece of data that's gone into that.

The models that we're building—we just released one of the best open-source medical datasets and reasoning datasets—you need to have 100% open data and a transparent way of training. So it's open source and open data.

Is there enough accepted open data to train on? I mean, one of the things that Elon talked about in training the next version of Grok was taking the corpus of human knowledge and correcting it, which I find fascinating. It's maximally truth-seeking. So how do you get to a corpus of knowledge that becomes the accepted baseline factual, trustworthy food source for these models?

So, from an epistemological basis, the truth is the truth that's agreed upon by context. The U.S. has a certain vaccine schedule and the U.K. has one, and it's codified in regulations and law.

Yes.

The data that you need to be a good general doctor is all common knowledge, and we have the data, but we need to transform it into knowledge and then wisdom. All of that is public. The models I'm trying to build as a base are the public knowledge of humanity, the common knowledge, and then the cultural knowledge.

This becomes very important because what Grok, OpenAI's models, and Anthropic's models don't have is virtue and ethics embedded from the start; they add it at the end.

Mm-hmm. Agreed.

The models that I have to teach my kids, I want to know what values and morals I'm teaching them from the get-go. I want it to be public knowledge—publicly accepted knowledge.

A hundred years ago, publicly accepted knowledge was that women didn't have certain rights, right? In other words, these publicly accepted values change over time as society changes. How do you deal with that?

We have a layered set of stuff that we generally agree upon, and then it goes to stuff that's fuzzier and fuzzier. The top level is stuff that you can tune into the model yourselves.

So you may disagree with the educational curriculum. You should be able to address the education model that teaches your kids, and your kid trusts more than anyone. You should be able to do that, right?

Then you have a whole governance structure that we've designed with agents, which has human input that can update those datasets, and you can fork them like in GitHub. So you know the data that feeds the model that handles your things.

Otherwise, these are black boxes, and in those black boxes Google and Meta are already selling ad space. They've said this in their earnings calls. So when it says “a beer,” it'll say Bud Light and all sorts of other things.

I think for the models that are closest to us, for the thing that's important to us, we need to have that and the ability to adjust it based on our own data and our own views, and then have AIs at the top that can have a look and give us an independent viewpoint on that as well.

Do you use your personal AI to interface with the systemic one to say, “Here's what my preferences are,” et cetera? Where does that value set come from?

I think that you've got your generalized human. We're made up of the stories that make us up, and we have all of these, and we should be able to pick and choose. Then, once we standardize it, we can have translation across the basis.

But I haven't got the whole thing figured out—how these sets get updated is something that communities will have to figure out for themselves. You need to empower them with the technology to do that because everyone will adopt it at different paces.

Have you figured that part out yet?

Well, no, because you give everyone—

You're weak, man. Weak by now.

Well, I think that the way that you need to do it is you can't prescribe everything because right now the AI models prescribe everything. They say, “This is the RLHF—reinforcement learning from human feedback—view, and you will stick with that, and that's what's going to teach your kid.”

If you want to give it Eastern values and use OpenAI, you're not going to be able to do that because they don't even think about that.

How do you manage identity in all this?

I think that the identity framework is very important. Bitcoin, in a way, is peer-to-peer identity that you can exchange. The way that we want to do it is that the more you use the Universal AI—this universal basic AI—the more your identity builds up, and then your AI agents also have identity hashed to that base blockchain.

So you've got pseudonymous identity at different levels, with the new zero-knowledge-proof stack that basically emerged in blockchain. It's building upon lots of innovations, and that seems reasonable because you want to know who's a human and who's an AI. You want to know who receives the money.

This was actually one of the final bits. When I looked at the future, I was like, the banks will be outcompeted, and you need 2 types of money. You need a corollary to Bitcoin, your store of value that goes up, especially because people are using the chain. So when you use UBI and you sell coins for it, it actually has chain usage. It's linked in a virtuous flywheel.

But then what is cash? I was like, you should receive cash, and this is better than UBI, in my perspective, for being a human and using AI and contributing AI. So we called it NIC. Now we call it Culture Coins, and each community basically has this coin that's basically pegged to their Bitcoin equivalent, pegged to their foundation coin.

You get cash for being a human. Humans are the issuers of cash without debt on the other side, unlike banks today. When we modeled that, we saw that it can actually make for a better economy because you have a level of subsidiarity, of subsidizing, that allows you to live.

Then you can think about the bigger stuff, which are these existential questions that we have coming very soon. If I'm not an accountant, if I'm not a lawyer, if I'm not a truck driver, what is the meaning that I have?

Yeah. What is my purpose? How do I use—I mean, one of the biggest challenges I talk through in my next book, which is We Are as Gods: A Survival Guide for the Age of Abundance, is the potential loss of human purpose at the same time that we're becoming godlike. How do we channel that? What do we enable ourselves to strive for?

You talk about, in one of your incredible works here, the idea of the Great Decoupling, or decoupling. Can you speak to that for 1 second?

Yeah. I mean, this is that break between labor and capital. The system doesn't need you anymore. It's already been extractive, and now it no longer needs you.

So what do you do in that environment? How do you survive, and how do you thrive? The system doesn't really care about you thriving. Our institutions are broken.

This is the other reason I wanted to release everything open source, because institutions could absorb bits and pieces of it and upgrade anyway, even if they don't have the whole stack. So I think this is something we need to challenge, and we're seeing it today because who here listening isn't reducing their graduate hiring? How many call center workers are you hiring now?

The wave is starting, and it's like a sandpile that's about to collapse. I've seen lots of people stating the problems. I haven't seen anyone stating a good solution. If so, just send them my way, because normally these transitions—you're no longer riding a horse, driving a carriage—take years.

Decades. Yes. And we're talking about what here? The decoupling is going to occur over what time frame, in your estimate?

Anything that's a KVM job—a keyboard, video, mouse—can be replaced in a few years. Public-sector jobs are the safest jobs in the world because they don't care about efficiency.

Right? Which was a realization I had. Oh, no. And with the Works Progress Administration you saw after the Great Depression and everything, maybe that's the future, and that's not a great future, to be honest.

Yeah. Let me generalize on what you just said because it came up yesterday, too. Things that are regulated are the safest jobs. It's not just public-sector jobs.

It's anything that has a regulatory component: financial advisors, appointed insurance agents, and all those. It's all surrounded by law that requires you to have the job.

Well, this is the thing. It requires you to have the job at the final point of contact, but they can be completely hollowed out for superior profitability by getting rid of all the people.

And actually, that's the theme of a lot of our business plans: the agent is going to be 10 times more productive and make a ton of money, but they don't need all their support staff anymore.

You need the scapegoat. That's the job of the human. The human is the future.

Accountable. Responsible. Yeah, that's all terminology, I guess. But no, that person really is accountable for the overall happiness of the client or whatever. There's accountability. It's not just a scapegoat. But your point is right.

Again, the only way to handle this, to a degree, is to make sure everyone has access to a basic level of AI. So that's why I think what the UAE did recently, by giving everyone free ChatGPT, is great, but the UAE as a nation should control that control plane. So they should be using the API, not ChatGPT.

You're going to see some very strange things happening very soon.

I want to make a point for our audience here, because if there's an audience member out there thinking, “Yeah, this may happen, this may not happen,” if you had to name the 5 reasons why—whether it's your design or another—this is definitely going to happen, I'll give you 2 just for free.

One of them is transaction speed. When AIs are communicating with and asking for services from other AIs, and your agents want services, you're talking about millions of times the transaction speed of the banking system, the SWIFT network. So you need massive rethinking in terms of transaction speed.

Another one is equity of access. Nothing in the current system has any concept of equity. If anything, it's a winner-take-all design, which leaves everybody in the dust. The voters are never going to go for that. That's going to force some kind of structure to come into place as well. So those are 2. What would be the top 5 reasons why this will definitely happen?

Something like what I'm building, or just the disruption of the great decoupling?

It's really exactly what you're building, which is a system of distributed governance that empowers AI. Within your model, you have application-specific AIs that are much cheaper to run than the big models from the foundation model companies. So that's part of it. There's a blockchain component for transactions and for value. There's a bunch of other parts to it, but it's really the future of governance.

Yeah. I hope it will happen. I'm not confident because of the localized incentives, but it's something that needs to happen, right? What's happening is that you're going to get a lot of defaults happening soon, and then people are going to basically be giving up their power to these technologies.

I have a system called iMind based on technology. Like I said, I use all the top frontier agents, and it's massively multiplied. That's why I've unblocked my writing. So I've actually written a couple of pretty good books recently, in the last few weeks, and they're genuinely good. Wow, this is really interesting, because what you have is an inevitability of a billion trillion agents. Someone will have to build a chain to organize them.

You've got an inevitability of this being in front of people, but we're outsourcing more and more of our thinking. When I'm using o3, my brain is actually changing, and I'm relying more and more upon it.

Mhm.

And as we have Gro 4 and Gemini 3 about to release level models, we're going to outsource more and more of our neocortex to them, and then we'll be locked in more and more as they understand our context. We've been doing that with smartphones for a while, where our memories are now on our smartphones and those neurons are now freed up to do other work, right? Yeah. You're talking about the next level of that.

The next level of that is because they know us better than we know ourselves. Our kids will fall in love with AIs. They'll be the most trusted entities out there. So I think this governance thing needs to happen now, and we need to have serious discussions about it.

We need to think about what is the world of AIs that I'm working with. For me, the AI that's closest to us for the regulated industries, because those are the things that handle us living up to a basic level, needs to be open source. The rest, I don't think actually matters that much if it's open source. But I definitely think that needs to be open source.

That's why we've seen discussions of, “Can you have DeepSeek in the U.S.?” We need open-source champions. I was like, “Let's just build it.” And it turns out that building that type of AI is a couple of orders of magnitude less compute than building the super-advanced genius ones. So let them build the genius ones.

Can we talk about advice to national leaders? Right now, NVIDIA, Gemini, Grok, and all of them are going to countries, offering sort of a one-size-fits-all solution across the board versus providing something that's bespoke per country. How do you think about that?

I think that the era of the massive model is actually done. To be contrary, I didn't realize this until we trained an AI medical API.

And it was so good.

I was shocked. Again, a human doctor is at 15% on OpenAI's new benchmark. GPT-4.5 is at 46%. ChatGPT's current model is at 40%. This is at 47%.

With 8 billion parameters?

8 billion. It works on a Raspberry Pi. It works on a 10-year-old PC.

Yeah.

And it's in 106 languages. By next year, we'll have it super good.

I looked at that and I was like, how much does it cost to train a doctor to do this? That's crazy. When you're doing a specific model for a specific thing, we're going to train it from scratch because we use Qwen as the base, but now we're using it to build the data set and then train it from scratch.

For all of these operational, cookbook-type jobs, the AI will run on nothing. The ability to create an AI that represents your culture and your data set is actually an S-curve. A Grok 6-type thing is exponential because it's trying to discover brand-new things. I don't want to have Dr. House, M.D., as my doctor day to day. I want to call him sometimes when it's complicated, right?

Go back to the culture side of the equation, because, I mean, first of all, nations need to digitize their culture sufficiently to have a native data set.

Yes. And that's basically what you learn in school. It's the textbooks plus a bunch of videos, except you've been told it's too much for you to be able to do.

Does every nation need a DeepSeek model of their own? Do they train themselves? No, they need a pretrained base. Then they need sectoral and other data sets that are open, gold-standard. Again, an S-curve, where in 1 or 2 years we can actually get good enough. Then they need to have their own cultural data sets, which, to be honest, will only require a team of 6 to 12 people at most.

That's insane. But this should be the plan. Nations need to have a plan. National leaders of every nation need to have a plan of action. I think what you're proposing here is actually viable, executable, and fairly rapid.

We realized that the compute you need for a node to train and keep up to date a national data set is a 72-chip Blackwell.

There's enough to go around.

Yeah. Hold on. That 8-billion-parameter doctor model is such a hugely important topic, because right now we have 20 million GPUs in the world, and we're going to make 20 million more. But there are 8 billion people. That means you need to share your GPU with 200 other people. This is like a public bus that's packed with everyone right now.

If I go to a random person on the street and say, “Hey, I've got $1,000 cash for you, or you can have a dedicated Blackwell for the next year—which one do you want?” everyone I talk to is going to take the $1,000 cash. What they actually want, and they'll realize this very soon, is the Blackwell.

Because the Blackwell is equivalent to employment. It's equivalent to entertainment. But they won't be able to get the Blackwell because the big foundation models are such compute pigs that they're going to use up every single molecule. This is why that class of 8-billion-parameter, use-case-specific model is the way to alleviate the bottleneck.

Get the benefit to society without using up every single GPU on just a couple of virtual girlfriends.

Exactly. And so our thing was to make that available everywhere, because then it can diffuse. Once we get our stack done, there will hopefully be a free app in a year that anyone can download onto their GPU device. It's fully open source and will check every medical decision ever made.

Mhm.

And that's where it really gets lost in those demos. At Google I/O, you get onstage and they roll out, “Look what you can do with Veo 3.” I just asked it to make a movie, and it made this. It makes a 10-second clip for you, and you're like, “Wow. Now just make it make a whole movie.” Then you realize, wait—the compute isn't there. They can do it in a demo, but you can't get it. Why? Because the compute doesn't exist.

Well, Elon said yesterday in his Grok 4 announcement that they're going to use 100,000 Blackwells to train the video model, which will mean it will need dozens.

Yeah. That's the fastest supercomputer in the world from a few years ago.

So I want to ask this question, given your background: Is Hollywood dead?

No. Hollywood studios will make a lot of money, but I'm worried about the people—

You said “the studio.”

So let's go in a different direction here. If the intelligent internet, or something like it, does not proceed, what are your predictions for humanity? What happens with the current trajectory of where we're going? Hypercapitalism, massive hyperinflationary collapse of the dollar and other assets, and AI basically outcompeting everyone. Every job behind a keyboard, video, and mouse that is not a public-sector job will go, which means that everyone will be employed by the public sector. But that takes time; we can't compete with what's coming. And this is before we consider things like my old college tutor, Oege de Moor, who was one of the GitHub code people and just created XBOW, which just won the top hacker in the United States award.

Holy crap.

Right. Stuff is going to happen faster than we could ever think. I was like, “I need something that can scale to be the highest marginal dollar for a GPU.” Actually, that's part of what I did. If you could—because Bitcoin mining is 90% energy and GPUs are 10% energy—what if you could just mine something that could give public benefit? The more of those GPUs we get to public benefit, the better. But absent this, none of them will go to public benefit.

I think you'd end up—when I look through your papers, there are bits and pieces of them being done across the board by lots of other models, and lots of other blockchains are attempting things like this, right? What you've done is connected them all into one systemic whole, which is very powerful. One way or the other, we'll get there to something like what you're talking about. It'll just be much slower and much more painful.

Is it that a lot of big things could break along the way if we don't get it into place quickly?

Yeah. Is it Trantor in Foundation, where you've got those thousands of years, or does giving every single person an AI that's aligned to them—as part of a bigger system with AIs for communities, nations, and humanity that can draw on a Bitcoin-type network of compute to solve problems—mean that we will definitely coordinate better, but no one can own that stack?

That should be an open infrastructure for humanity, and anyone who wants to participate in any element of it should be able to.

And so that's what I think is optimal here. Whereas in the Foundation books by Isaac Asimov, you had the secretive Foundation, we actually know everything that we need to know to get through this. We just need to rethink how money flows. It should flow from being human; it should be anchored to intelligence and organizing the knowledge of humanity. The best way to make money shouldn't be that you earn capital. It should be toward doing good.

So that's why you start with proof of benefit. But then the winners of the XPRIZEs and the cures for cancer become the richest people in the world, and the system directs it that way, aided by AI.

I love the statement I made at Singularity: The best way to become a billionaire is to help a billion people. There's this alignment there.

And we should make it literal, because then the AIs will be working on it and the humans will be working on it. Again, we start out with the basic thing right now: Let's get as many people aligned with intelligence as possible, and get the compute organizing our collective knowledge on cancer, autism, longevity, and biodiversity.

That would be my third bullet on why this must exist—exactly what you just said. The AIs need to work on it themselves; otherwise, it'll never keep up.

We had Joe Kennedy III here in the office a little while back. I had read his white paper—he was in charge of the crypto commission in Congress at the time—so I read his white paper. Then he came in and I asked him some questions about it, and he said, “No, I never read my own white paper.” I was like, “What are you talking about?” I said, “No, that's not how it works.”

You know, the government is so big and so in charge of so many things that everything needs to either be an agency or self-governing. Then we just choose the self-governing process.

AI will take that to the next level, where the rate at which you need to add ideas is so fast it can’t possibly go through Congress. So it needs some kind of framework that grows at the same rate that AI grows. But this is the AI-first stack that you need to build, and we’re building it right now. Andre Karpathy described this thing called the LLM OS, where you can virtualize any interface and it’s all about the context that you give these. When I thought about it, I thought about the intelligent internet. Our existing internets are going to break. The AIs will attack it from somewhere or another. A robust, resilient, coordinated, verifiable internet—you can actually build that from scratch. That infrastructure will be very difficult to attack, just like it’s difficult to attack Bitcoin because you need the hash rate. To attack this, you’ll need to have a massive amount of compute because it can call on compute to resist any attack. And those attacks are not only hacking attacks; they’re psychological, memetic attacks.

And this is something that I realized recently.

If you think about the base on which almost all of these language models are trained, they're all the same, which is a generalized polymath view. That's a large attack surface. Again, that's why I want my doctor to have a small attack surface. Do you remember Stuxnet?

Sure. A tiny virus went to all the reactors around the world, appeared everywhere, and was engineered by a human. It was amazing code, and then it caused them to go into overdrive and blow up.

Mhm.

And when you think about that, we need a system that's resilient against things like that. If everyone starts with the same AI—and this is why I've got Material Intelligence; I've changed it to Network and Diversity—you have a diverse range of different AIs, and you're more resistant to that than if you have the same AI across everything, one that's overbloated, based on this polymathic basis, and running on critical systems.

So how many discrete—if you think about the 8-billion-parameter model as the prototype—are we talking about millions of these?

No, I mean, it's 1 for every individual, but you have a very tight dataset that it's trained on. Like Stable Diffusion: We trained it on 2 billion images. Someone had equivalent performance with 25 million images.

What are the right words and wisdom for society that you actually want to have in your model at the start, to do the jobs that are needed to be done, to enable everyone to have a basic level of intelligence and access? That's what we need to build toward, and the network needs to align itself toward.

Then everyone's got their own AIs, and you're in this stack that's built AI-first to be resilient, as opposed to having all these dependencies. That's going to be more resilient than our current internet, which is not resilient.

Yeah.

From an infrastructure or even a memetic aspect.

You know what's really shocking and kind of sad is that we clearly need to go in this direction because the compute is constrained, but it'll always be constrained. There's no way we'll ever be able to build enough compute to keep up with the innovation that AI can come up with, because software will move a billion miles an hour; hardware won't. And so it'll always be constrained, so we need exactly what you're talking about.

But when you start working on it, the first thing you do is pick up Qwen and start working with a Chinese model. That's the part to me that's surprising, sad, and weird. I think it's partially a byproduct of the chip embargo: The forced constraint on China is forcing a lot of innovation, and it's also all being done open source. But it's really sad to me that there's no way to start with something that's built into your framework—or built in America—that empowers what you're doing.

Yeah. I will train a model better than Qwen. Again, I built state-of-the-art models of every type, and I'll make it fully open source, which is why I was like, “I've got to do that.”

Then you'll solve the problem.

Well, yeah, I know how to solve it. This is the thing: it's not hard. It's just that it's in no one's interest to do so, but it's in our interest to do so because the more people that use our model, the more it becomes like VHS versus Betamax.

Yeah.

I want a good-enough model that takes everyone up to a basic level and then combines to have massive network effects. Then I can use that to get more and more mining of this, to give more and more benefit. And that's it.

How do you get mass adoption at a pace that resists the legacy systems blocking it off? That is the key question, right? You're going to have an immune reaction from those whose systems are being disrupted.

So the way that you do it—the only way you can do it—is, again, decentralized and permissionless, right? You buy the currency exactly the same as Bitcoin, available wherever it is, except you can direct where your purchase proceeds go: to cancer or autism or whatever things that really matter to people. And then the code is permissionless.

Stable Diffusion was downloaded over 100 million times because it's permissionless. So it will end up everywhere, because would you prefer to have your Qwen model that actually has a Qwen license and you don't know what's inside, or a model of equivalent performance that's being used by everyone, deployed or not? You will use the VHS, but we want it to be an amazing-quality VHS.

So you've been the champion of open source now for, I don't know, 10 years probably.

At least 5.

Is there an open-source system out there that you trust right now?

No. The Allen AI guys are doing good stuff, but is it in anyone's interest to really make this?

Yeah. Yeah.

Again, just like I gave the very simple example: should we have a supercomputer to organize all the cancer knowledge in the world?

Yes. Check.

Will it accelerate a cure for cancer?

Yes.

Why isn't it done?

Yeah, yeah, yeah. When you say it's in no one's interest, what you mean—and I'm just parsing it for the audience here—is that it's not in the interest of the big foundation-model companies. But it is in the interest of 99.9% of humanity and almost all voters.

If you're not a big shareholder in either Meta, xAI, Anthropic, or Google, you're part of that 99% of the world. This is hugely in your best interest, which comes back to the rollout plan. How do you get adoption of this very, very quickly? If it's in the interest of 99.9% of humanity, there should be a vehicle for getting adoption.

“I bought Bitcoin, except that all of the money I used to buy it went toward this cancer supercomputer. And here's a screen of what the cancer supercomputer is doing today.” That's the narrative.

“I'm downloading these models because they're open source and free, and I can do whatever I want with them. There's a whole community built around them.” That's the other narrative.

And then, for me, these are the primitives that can lead to what I think is going to be most important: the economy needs to change. The reason that the cancer supercomputer doesn't get rewarded, even though we can raise $100 million, is that $100 million of the coin will be on that person's balance sheet. So it costs them nothing, probably, if we're successful. In fact, they'll make money by helping with $100 million of compute because of the demand for digital assets.

The economy is only based on one thing, which is GDP, which is material. We need to reward intelligence, network effects, and diversity. Again, we couldn't do that before because we didn't have these AIs that could measure the impact, that could impartially say, “Of course, this is a good idea.”

I'll give you one very straightforward example of that. If we had a dedicated open-source supercomputer that analyzed every single bill that went through the government and gave an honest take on it, completely transparently, would that change the way our democracy works?

Yes.

Would it be made illegal quickly? Yes.

Yeah. Maybe. Yes.

But this is the fascinating thing: from a mechanism-design perspective, the AIs can act as independent, virtuous, and wise entities.

Yeah.

As we build the system of governance for the future, we need to make sure they can do that, which is why we need to build out why this must happen. Because what we're essentially saying here is that if this doesn't happen, natural progress will put all of the power into the hands of a couple of big foundation-model AI companies. You can measure that by where the compute is.

If you see all the compute in the hands of, say, 3 or 5 people, then you know that they're dictators of the world forever thereafter. And so, if we don't do this at an exponential rate, at a rate at which it cannot be caught up with and disrupted, that's the issue.

You know what's really cool? All the AI will transcribe everything we're saying here. When we say, “Hey, AI, can you put this into a bullet list for our audience?” then we'll have 4 of the 5 done.

Yeah.

So if we can think of the 5th one, that would be really useful. The AI will let us figure out the 5th one. Salim, what do you—

Yeah, I think there's a nice inevitability here, right? As we move forward down this AI-automation-of-work path, if we just take the simplest thing, like driving, which we've talked about as being the number 1 job in the world, right? There are more drivers as a profession than anything else in the world.

You automate driving, and you have a huge GDP collapse from automating a few of these things: health care workers, driving, teachers, et cetera, et cetera. That collapse in GDP will force a move into a new mechanism, independent of whether the governments like it or not. One of the flaws of GDP is that if you make something 100 times more efficient, it drops GDP.

I mean, there's something interesting there about a complete collapse of aggregate demand and other things. This is why the other aspect of the economics paper—and, again, there are more proofs of that—is that we need to move to a new type of economics where money is basically created by people. That will give constant demand forever, because the current solutions, like UBI, will never work mathematically.

It cannot work. When tax rates go down, the AI is coming like a wave, and the AI will also be amazing at tax accounting—putting it through Ireland or whatever.

Can you slow that down? So few people understand what you just said, and it's so important, because everyone's talking about UBI, UBI, UBI.

Let's break down why UBI—universal basic income, for everybody listening here—is being discussed. The basic thesis is that, as jobs go away and productivity goes through the roof, potentially GDP, because you're dividing by effectively zero, goes through the roof. We are going to give every citizen on the planet, or in your country, a certain allocated amount of money that allows them to survive month to month and covers their basics.

It's been tested in 100 experiments. Salim and I have both written about this, and in these limited experiments, people don't use the money for beer and Netflix. They actually use it to improve their lives, educate themselves, start businesses, and, in Africa, buy some animals, buy sewing machines, and start a job.

The problem I've stated in the past is that going from a taxation-union-labor-job type of structure to this is such a huge leap. We have no confidence in the public sector to get us there, right? But, Emad, you have a more nuanced view on the economic side of it, so talk through that.

Yeah. I was like, “Give them money, but give them money for being human, and make them mint the money. You use your artificial intelligence.”

I'm going to slow this down for everybody. Make them mint the money—what does that mean?

Let's rewind it a little bit, actually.

Yeah.

What happens is that UBI works in small cases. If you have a complete realignment, a great decoupling of society, and all these agents getting smart at once and outcompeting everyone, aggregate demand goes down massively.

Aggregate demand for what?

People stop spending because everyone's losing their jobs. This is where you end up with the outcome of the zero-marginal-cost society. Tax goes down, and then these whole AI firms—or one person at the top with a thousand million GPUs—will never make a profit, because a profit is an indication that they can't find any more marginal value, and the profit can always be put into more GPUs. Even if they did, they would figure out all the tax loopholes in the world not to pay that profit, because they're better at tax loopholes than you are.

Mhm.

That's what I see as a positive, in fact, because that'll be the forcing function to flip to a system like this. It'd just be painful as hell if we don't do it quickly enough.

Once you've given everyone basic AI, which I think will actually only cost $1 a month if we get it right, why not make it so that the money that comes into the system isn't from banks? It's from the people. So everyone mints money constantly.

Going back, what does everybody minting money mean in this scenario?

It means that you have a national digital currency.

Mhm.

The number goes up every single day as you use your AI to make yourself happier and better and improve your community.

So, I'm using my AI to write a new story for my 3-year-old child. I'm using AI to diagnose a sick friend. As I do that, as I use the AI in a positive, agreed-upon mechanism, I'm minting these foundation tokens.

You mint your culture coins, as we call them—your national tokens that are pegged to the yuan, just like you had gold pegs.

So, you've got your stable thing, and then you've got your flow. You mint your currency at a certain level just for being a citizen, and then more if you do society-positive things. So, if you feed cancer data into a broader model, you get more than if you just sit at home doing nothing.

Exactly. As you build up status in your community and your society—and again, there are lots of details we've worked out about that part—you should be able to benefit. That becomes a circulating currency because then people are like, "That's an index on Mexican AI use."

Money will flow more and more into digital assets from the existing economy. It will flow more and more into generative AI assets regardless of anything that happens. But if we get a collapse in aggregate demand, you know where capital's going to go: capital's going to go into generative AI and blockchain. So, make it easy for that to happen.

Yeah, that's really, really brilliant.

The quiet part out loud here is that, over time, this will collapse national economies. You'll end up with 1 global structure.

Well, just—China will dominate.

Very similar. The problem with UBI is, I give you money right now. If you make money, you have a choice to use it on entertainment, go to the casino, whatever you want to do, or you can invest it, or you can buy a sewing machine and start turning it into a business. That's your choice.

In the future, that universal right to AI is the equivalent. Now you have your AI; you can use it for your virtual girlfriend, if that's what you want to do, but you can also use it to generate some benefit. You can use it to help cure cancer. That becomes the equivalent choice, and it's the only relevant choice in the post-AGI economy. That's the flaw in UBI and the beautiful thing about this design.

That's why I've got network as a key thing. Wikipedia created so much value from its network effects on others. If you say, "What is the meaning of life in a post-AGI world?" it's living. I saw my family on the weekend. My daughter's art is actually pretty good, but most daughters' art isn't pretty good. That created value, right? But it can't be measured by any of this.

In the post-abundance society, the Star Trek world, it's one of boldly going where no one has gone before. It's about exploring. It's about deepening your community values. If you've got an AI next to you that's looking out for you, it's going to be encouraging you to do that as well as contributing to some of these bigger problems, right? It's going to encourage you to create because creation is about context. It's about flow.

So, this is why I think, if we program this right, it can be a really nice, elegant structure that moves away from the extractive economics that we have today, and capitalism and democracy are the worst of all systems except for all the rest, to something better. The question is, where do we want to direct ourselves? My view is, direct it to benefit. That benefit is something that we need to decide at a societal level, country level, community level, and individual level.

You know, there's an exemplar—a good example—of what you're talking about here, Peter, and I write about it in the new ExO book, which is technological socialism. Typical government socialism fails because you always end up with centralized, inefficient planning, and it invariably leads to corruption. It always fails on those 2 counts.

But we talk through an example like Uber, which is the sharing of assets among a large group of people. It's actually a socialist function, but when an algorithm hyper-efficiently allocates it, you get all the benefits of the collective assets without the downside of inefficiency or graft. I think there's a stepping stone from something like that to what you're talking about. That's an easy thing to go down.

When you have a structure like, say, Uber, you don't need a lot of regulation because the system has the right inputs and outputs and feedback loops to self-manage itself. We're seeing more and more examples of that inevitably leading to this. This is why I think this is going to happen one way or the other: there's too much efficiency to be gained by having a system like that than by not.

But it'll be facing lots of forcing functions and legacy issues. If you can craft it in the way you're thinking about, you're providing a scaffolding that everybody can just ladder up in a structured way to that new model, which I think would be very powerful.

And the fact that generative AI creates this non-rival intelligence for almost nothing—the cost of skills has gone to almost nothing. Capitalism will not survive that.

Yeah.

Literally, just go and ask your o3, your ChatGPT, and your other models. With intelligence going like that, and AI achieving a level of performance that's equivalent to a human and can scale, what does that do to capitalism? What does that do to democracy? What does that do to tech? You'll have some very deep answers there.

I'm thrilled by you finally giving birth to these papers.

I'd love to have you back again. You've been an incredible friend of the pod.

At some point, you have to come on. I think that's really, really important.

Okay, Dave, you got it. I was going to wind this down.

So, you know, ideas like this historically propagate out through a token, and you already have this universal access to AI as a founding principle, which I'm sure is going to be available in blockchain form, right? That's the only way you would get it out quickly to the world. Is it fundamentally investable as a token or not?

Yeah. It's just like Bitcoin, except it's called Foundation Coin.

It's called Foundation Coin? When's it going to be out?

Soon, hopefully. We've been mining it since January.

Whenever it's out, come back and watch it.

Well, I will come back, and we'll launch it right here on this Moonshots pod.

Agreed, Emad.

Agreed. Agreed. Yeah.

So, Emad, what are your next steps here in making this real, making this launch?

We're going to put out all of the knowledge and the work that we've done, and we just want people to give their input and think about what the future of education, health, governments, and everything is like. The first step is to stress-test that. So go to ii.in. If you sign up, you’ll get the details of that, and all the details and knowledge will be there.

And then, okay, that’s where we go to get more details. You’re active on X. What’s your handle there?

It’s estac and then my last name.

All right. Amazing. Salim and Dave, we could go on for hours here in this conversation, and we should. But next time.

I think we've hit the S-curve limit of extracting useful value in this conversation. My brain is fried.

Just restructure society. No big deal. No rush. It is the most exciting time ever to be alive. We're in the middle of this intelligence explosion, and I just want to remind people that we have the ability to steer.

We don't need to give up on a sense of humanity's future. It's not going to happen to us. We have the ability to shape that future, to create that Star Trek universe, to create a hopeful, compelling, abundant future for humanity. That's what we talk about on this podcast, and that's what we believe in.

Emad, thank you for being a friend and a brilliant creator.

All the folks that prefer a Mad Max future don't listen to this podcast anyway.

Yeah, I know. Go someplace else. Talk about politics someplace else. You want to talk about science, technology, creating the future, and creating moonshots, this is the place for you.

Emad, a pleasure. In terms of the name of this podcast, Emad, I think you hit the high bar on a true moonshot.

Yeah, saving humanity. That's a good one.

My MTP is a transformed civilization, and I think I can just relax now. Like, I gave up space to Elon. You can give up that to Emad. That's great.

I'm good.

You know, why are these 4-letter names that start with E so important these days? I don't know.