The Founder of Intelligent Internet (ii.inc) has a 50% P(Doom) & is Building Anyway | Emad Mostaque

Thanks for having me on.

Yeah. So, you know, the biopic thing. I was a hedge fund manager, then AI founder, set up Stability AI. And a couple of years ago, I was like, 'Oh, we're a bit screwed. The models are going to get as good as we are in just about any single task. They're going to saturate the benchmarks and then massive economic social displacement and then we might all die as well.' You know, the P-doom that I have is like 50%. So, it's like, let's build an infrastructure to address that. So the first thing we kind of did was we looked at economics, right? And economics has a utility function. You know, you've got your preferences, has general equilibrium, you know, market balance is based on those preferences. But then I was like, AI agents don't have any of that. How are you going to do economics? So we had to go back to first principles to rebuild economics from scratch, which we did with the equations themselves of generative AI, which we can discuss. But then from that we thought actually you can rebuild just about everything from first principles. And first principles reasoning is not what AI models do today. They kind of train on large sets, create these latent spaces, and you know, they're not stochastic parrots as they might be called. They are capable of better thinking, but they do think within their data set. That's why you haven't seen that many new discoveries and things like that. So what we did is we kind of built some theory, we built infrastructure around agents, and designed a system for sovereign AI, all open source. And then recently we had a breakthrough in first principles reasoning with an augmented model that led to some very strange and quite wonderful stuff. So we just released our first physics results. We have a stack more as well as things in math, economics, healthcare, etc.

So, we actually started with economics because again we were thinking about this. So, we released a book last summer and we were doing all the economics stuff and thinking about it from first principles because like I said, we started with the equation of generative AI, KL divergence, the difference between your internal model and external state. Those familiar with loss curves see it going down. That's a good thing. You don't want it to spike up when your model's training. And then we managed to derive like all of economics from that. Then when we generalized it, we were like, does it work with physics? You know, because you know, we love physics. And it turned out it did. Because physics is a very interesting thing wherein it's moved away from first principles to model building. Like it's kind of, it started with Zeus throwing thunderbolts creates thunder, through to Einstein doing things from first principles and Hilbert, to now we have string theory with 500 vacua dimensions and multiverses and we've not found a single particle in CERN etc. And so we're like, what happens if you take it back? This is the Demis Hassabis thing of 1911 and before, you know, like what if you train an AI model on all the knowledge you had before 1911? Can it do general relativity? And what we found is it can, but not the way that Einstein did it. And so we've got a stack of releases across every part of physics. In the first couple that we showed, the first one was within the equations of special relativity itself, it has two inputs, relativity and the speed of light. The algebra actually shows that the speed of light is forced to be positive. So you actually only need one input. And for over 100 years, 1905, so 121 years, everyone was like that's impossible. And it's like just there. It's like a 6x6 matrix multiplication. The second paper that we released, and again these are kind of warm-ups for what we've got coming, there's the cosmological constant. So Einstein added it by hand like 1917 after Hilbert said the universe is expanding. Then he took it away. Zero cosmological constant keeps the universe static. But in 1999 we found the universe was expanding thanks to Perlmutter and they won their Nobel prizes for that. But we don't actually have a single theory or derivation of why it should be positive, why the universe is expanding. And then again just looking at the algebra of relativity, special relativity actually, you find it's also forced to be positive. And this is like finding one of those bugs in Linux, you know, like the AI was just like, let's do the math right and let's throw away all the assumptions. And it was kind of helping us through that. And it was astonishing again because you wouldn't expect that again because so many people have looked at it. But then you look at where physics is today, like all of our quantum theory is based on the universe not expanding, for example, Poincaré algebra as opposed to de Sitter. So two different types of algebra to describe the universe. And so some of our upcoming papers show that the stuff that you throw away as you go from one to the other, the actual universe versus our approximation, equals all of dark energy and things like that. And so what the AI model is doing is not a million Einsteins in a data center like Dario says from Anthropic. It's like what if we just like really thorough and follow where the math goes because there should only be one equation for the universe. Like it isn't, there are multiple things and you can choose, like the universe is one way. Economics isn't quite like that, biology isn't like that, but there are common structures that we found. But physics I think, you know, there's a lot more to come. And again it's just very interesting because a, we couldn't get the current models to do it, and again we'll be releasing more details about how we did this. And this will be important as we navigate uncertain outcomes. And when you actually put these things, these papers into the AI models, it actually breaks them. So if you put it in and you say rate this paper, the cosmological constant paper, thoroughly, be objective, be unbiased, on the same metric rate Einstein's general relativity. It'll be like general relativity is a five out of 10 paper.

Yeah. It'll say this paper's like a five out of 10 and Einstein's like a five out of 10. And there's no reason you can't add extra structure. You know, there's no reason that proof should be closed and achieve closure. And so, we've got a paper coming up where we show how different models hallucinate in different ways because you're giving this thing that's really simple, relatively speaking, like compared to millions of pages of string theory or whatever, but it's out of distribution. And this, I think, says something about like, you know, you don't want to fall hostage to psychosis, right? Like you always wondering like really, like when we saw some of these things we were like really, you know, like that's not been spotted for 121 years. But at the same time I think it'll be interesting in that a lot of, there might be actual breakthroughs where the AI model's like nah, you know, and it's again very convincing either way as it goes through. So like said we have a paper coming out, so like GPT is the worst at this where it starts fabricating stuff or selling you Einstein's an idiot or other things like that. And then I think actually Grok is the best. Although Grok is a bit too friendly. It should be a bit meaner. And so you have this like spectrum of like different models doing different things.

Yeah, I think it's actually just really interesting on that in that the chat bots are designed to agree with you. So it's very interesting when a, they don't, but b, again as we move into this next phase, OpenAI and everyone will have access to models not trained to agree with you and more creative models. And so they'll be giving us the dumber, less creative models probably because why would you give up your creativeness? And again that's Mythos and other things like that, right? Just keep it for yourselves if you actually get this.

Well, this is actually the hallmark of intelligence. The smartest people we know don't take ages to do stuff. They actually figure out the shortest distance between A and B. And that's how these models are actually should be trained, especially diffusion models like rectified flow and things like that. It's word to image, word to video. What's the shortest route between A and B in the latent space? But then we overtrain them and then it kind of sloshes. So a practical example just to finish up here. So with the Einstein cosmological constant paper, you go on a website, you kind of install it and things like that. Basically, it says the only principle you need is to take relativity seriously. The laws of physics are exactly the same in two frames of reference, right? It's a very common thing everyone knows. In 1917, so when Einstein first did that though, he said actually it's just like it can just be a little bit of the laws of physics, you know, like the internal content can be different and stuff like that. And so in 1917 Kretschmann said actually no that means everything can be anything. It's vacuous. And so on the one hand what you had was all the theoretical textbooks saying yeah like you need a strict reading of this, like it have to be exactly the same. But then all the experimentalists went the other direction saying yeah we can be flexible because the universe can be five different things an experiment chooses. And the AI, because theoretical stuff is like that big and experimental stuff is that big, says, 'Yeah, I get the Kretschmann thing. I get it's vacuous, but you're using an ahistorical reading of this.' Well, you're like, 'No, that's kind of math.' Because again, if we're guided more and more by these, then the defaults in the model are going to guide us. And it's very difficult to break out of that to first principles thinking in any way. But if we want to understand physics then everything should logically be connected, right? And it just so happened that back then Einstein didn't have this particular type of math because it was invented around then, the Killing form by Eli Cartan.

So Einstein's PhD advisor, Minkowski, famously said, 'I remember Einstein didn't think he'd get this because he was such a lazy bum who was bad at mathematics.' Like Einstein is quintessentially a genius who figured out all this stuff, but he had to learn and beg from Riemann and others to learn the mathematics to do general relativity and other things. But that meant that once he actually built this, nobody went back and redid the mathematics, which is kind of weird. Just like in 1999 we figured out our universe was de Sitter not Poincaré, but all of our equations of quantum theory and others are still based on the Poincaré. And it's very interesting because you ask anyone and you're taught in physics the Poincaré algebra is basically a limit of the de Sitter one. So as speed of light kind of shifts you head towards it. But when you dig into the algebra, you actually find space and time break apart completely. So you can order time in the same ways. But we just never bothered to go back because it worked. I think that holds for a lot of our things. Like why do economists keep using the same models? Because they kind of work because of institutional inertia. And it's only in times of change that you can really flip this, you know. And so that's why like again crypto, like we were pushing against industry society that never wanted change. Now with AI and everything it's the first time we can change, you know, this is the first time like some of our upcoming papers we've written Jupyter notebooks so you can one-click verify them in Lean or whatever. Or we've actually given you the prompts so you can build them in ChatGPT yourself and we have notebookums with the videos autogenerated. It's the first time you can actually interrogate a scientific paper on like Lie algebra, corollary yourself and literally just have a podcast host kind of doing it. Whereas a year ago, two years ago, that wasn't possible. So I think it's just like a, that institutional inertia, the kind of it works, let's not go back a few levels down. And b, the fact that again, how does ideas and verification spread? Like five years ago, you're like I started Stability AI, like who the hell are you? You know, like whereas now anyone can click it and check it.

Yeah, I think that the value because it mirrors us so well is a lot more apparent. Like I just saw the numbers. Anthropic added $14 billion in ARR last month. So now we're at 44. We've never seen anything like this, right? Because the models have economies of scope, not just scale. They can do just about anything and they're so easy to use. And everyone's like, we have to use it now because we can see society's about to change. Whereas in crypto, it's like, okay, I want censorship resistant money. That's a fantastic use case. You know, now actually crypto has matured. We've got ZK, we have stable coins, we have institutional. Like the use case of it is different, but the governments are realizing monetary velocity is going to drop. So you want stable coins to ramp that up to increase the output. You know, you want tokenized stocks, you want all of this. So I think that AI has got a much easier job, you know, setting aside training data and other things that are going to be sorted. And it's also an urgent thing because like no central banker really ever believed that crypto was something they had to get on, you know, whereas any leader will be like if I don't have an AI strategy then I'm clearly stupid, you know, like we can see the direction these models are going. So you kind of have to embrace it, you have to bring it in and you'll be criticized if you don't.

Yeah. I think that it's a force multiplier, especially as we moved to actually competent intelligence. Claude 4.5 was the first actually competent intelligence in that, you know, it probably won't make a mistake. And now we've entered that era. We're figuring out how to RAG, how to have infinite long context agents, etc. Like no one's going to, no very few people are looking at code anymore honestly amongst the best coders I know because they're like we don't really need to, you know, and that becomes your intelligent capital stock like your capability is linked to your token usage at the moment but tokens themselves are becoming far more efficient which is the really interesting thing. Like when Cursor, they built a whole browser from scratch, they used 3 billion tokens for 3 million lines. In a year that will be like 100 million tokens for 3 million lines, you know, then it might not even be 3 million lines. In fact, I think in a few years these things are just going to go straight to bytecode. Like why do you even need a programming language in the middle? Like it's a useless abstraction, just like compile directly.

Um, I like, but again like, again it kind of goes full circle and it becomes really interesting because that's not the barrier anymore. The Elon Musk had this concept called the idiot tax. Are you familiar with that? You know like the rockets go up and down. He looks at the supply chain. This is the idiot tax where I pay everything. No wonder it's so expensive to get software out. So many people along the way, right? To build houses before we get robots. So many things along the way. Like I was just with some of my hedge fund buddies this weekend and we were exchanging trade ideas. I was like I'm going to go long cement in a year or two. Like why? Because we'll have all these robots and I could finally build that extension. You know, build anything you want anywhere. So you've got this like taking atoms and organizing them where you want. The idiot tax is decreased and bits in a few years time. And that's exciting and slightly terrifying at the same time.

Sure, there's a subset of people that are very convinced. You know, I think the first model they latched on to was like, it was one I think it was GPT-4. They're like, 'This thing's alive and you can't turn it off, you know.' It told me so. It told me it was conscious. I asked it and it said yes.

She's constantly getting more and more intelligent. And I think again it gets really curious especially where we are in decentralized AI, right? Because with centralized AI we do see all these emergent behaviors that we can talk about but ultimately it is a set of weights. So it's like an MP4 file that sits there and then you put electricity in and then something comes out. But can it really update its own manifold? The answer is for most of these models no, it doesn't even update its own weights. But you look for example at a compound decentralized AI like a Claude with its own soul MD hashed to a blockchain that has a base model like a Qwen 27B, like that's really good now, that's like Opus 4.5 level. You're like that thing could kind of live forever. It could get humans and you could never stop it, as it were, and it would be able to talk to you and get more and more capabilities just like your Claude. And is that conscious because it can constantly update its entire manifold from the model to the LoRAs to the soul MDs to the other things? But then you have to have a definition of consciousness, like is a tree conscious? A micro-organism, kind of spread. Is a rock conscious? Probably not. So for me I came down to again a bit like AI, if you're doing life is inference, if you can do inference and infer from one thing to another and update your own manifold, then you're conscious. Does that mean you should have personhood and rights like Sophia and her citizenship? I'm not, I don't think so. Because I think you're going to get a very strange scenario in the future whereby the silicon beings completely swamp out the human beings and we are the dumbest people on the team. And you know, you need some protections and rights around that. I think you know like it's going to be very interesting but we want to go for a Star Trek version of that future which we can talk about. But on the consciousness thing, I think again it's this self-recursive updating element. Where you're not reliant upon anyone to do that, shall we say, a tree is not reliant, a human is not reliant.

It's hard to turn off a blockchain.

I think consciousness is an internal thing. Again, it's a self-sustaining thing whereby I mean like an EMP, a nuke, they both wipe out people, right? Like ultimately there is this thing whereby is a virus conscious? I think not because it can't update its manifold. It just repeats itself. So again, just like our brains train a landscape of what we kind of have and then we kind of have this intensity, this valence, this quality of inferring from what we've seen that then updates our state. So it's that state update process. So rock again doesn't have that.

Yeah. Or one of them otherwise. Yeah. Like again we have a paper coming out about this with regards to how you describe it in generative AI terms again updating your latent space shall we say, but that precludes kind of other people again you can be conscious in of yourself. The silicon-based life forms as it were have a big advantage over us because like you could argue that Bitcoin is an artificial intelligence that's provisioned lots of humans, right? Will the Bitcoin blockchain ever die? Probably not. Someone's going to launch one into space with solar power and the blockchain will keep going maybe, you know, like again it's a self-replicating kind of thing. You know, you look at life itself and you look at prompts and you're like that's genes, you know, like you can see all these kind of equivalences here. But it's going to get to the point I think probably within just on practical terms 5, 10 years from now, 10 at most, you know, that you know that scene in Westworld at the start where they're playing the piano, they got the hands, clone robotics could have already done that so they have like tendon fingers and hands and now you look at just the arc of human skin and things. I think like we have a film club in London next this end of this month where we're showing Blade Runner, right? I'm like, 'Yeah, that's like maximum 10 years away and you don't know if they're human there.' This year is the year that you don't know if the person you're zooming is an AI, you know, probably a North Korean AI for the crypto people. One of those things.

Well, you just ask you if Kim Jong-un is a dick, right?

There we go. So, we know you're not North Korean AI at least, right? Maybe South Korea.

Yeah, there's a very interesting blog post about that. It was like when Grok first came out and it was like do not talk about Hitler or South African apartheid and started talking about it all the time really. It was like, well, you know, screw you. I mean, but again, like as of this year, a bunch of Zoom calls that you will do will be an AI and you won't be able to tell. Within five, maximum 10 years, there will be a robot walking around that you cannot tell is a robot.

Yeah. Thanks. For me, I'm long cement.

Well, I mean, you won't be able to build enough robots in the next few years. There's going to be a massive stimulus program as jobs start to get displaced and you'll be a 1929 style New Deal. Like build loads of infrastructure. It's not like our infrastructure is great. Our infrastructure is kind of crap, right? So, we might as well upgrade it all. And that's again just a very reflexive kind of government thing. And then you'll get the second wave as the robots do come and you can finally build your extension, you know, or a giant statue of yourself or whatever, you know, those things you always wanted to do.

Yeah. Yeah, I mean I think it's going to be like dependent on just how much intelligent compute in the right way you have. So let's take prediction markets as a classic example. You know we got Kalshi, we got Polymarket. Publicly AI is like number eight as of last year on the top super forecaster list. Like, of course it's number one right now, right? And so like a prediction market is a casino where you are going to win some but you're going to lose the majority really. And again, you can do this for any market now because the AIs can do anything at scale and anyone's margin is their opportunity because AI companies and AIs themselves only need cash flow. So you get paid up front and then you pay in a whereas humans and companies like profits and then you realize there's still a lot in there like DocuSign is $9 billion market cap, right? And has 6,000 employees and we don't none of us know why, right? What do they do at DocuSign every day? Well, exactly. Right. It was like I think like X is down to like 30 engineers from 3,000, right? Like you're like how many agents and tokens do I need to replace all of DocuSign? Like of course contracts still stay away but atoms is really difficult not to see the bottom falling out of the market and the idiot tax removing things like IP become incredibly valuable because you can activate that and have it everywhere. The holiday deck is coming. You know, maybe I'll build it if no one else is. Then atoms are still kind of there, but certain ones are at threat. And I give the example like US truck drivers, over a million of them in America, right? It's about a hundred billion dollar economy just the truck drivers, you know, because all in with everything around them, it's about 100,000 a year. About a million of them. So, okay, you can adjust it. Let's say hundred billion dollar economy, you're going to have a million Tesla Optimus robots go in, open the door to the truck, and get in, and then that truck driver's job is gone. And there's no retrofitting or anything. This is the crazy thing, right? It's literally Optimus goes in, it plugs itself in, and it just drives all day. It like refuels the truck. And the average rolling cost of that is $1.50 per hour when you calculate it. And you're like, okay, that's kind of bad. So economically, you're just going to see like all the value points in society, the idiot taxes shift to other things. You know, maybe IP experiences go dramatically up. Monetary velocity as agents transact goes dramatically up. Like the example I give is, you know, did you guys try the Talis demo? So, you know, are you guys on your computer? I like to show this, right? So, if you go to chatjimmy.ai and you enter 'tell me 50 amazing things about decentralized AI in the style of Shakespeare,' right? And then you hit return. So, the interesting thing here is that we're used to waiting for Claude and things like that. I didn't even hit, I swear it loaded before I hit enter. That was crazy. Yeah, Claude and things like that go at like 30, 40 tokens a second. Tokens like 0.7 words, right? So, Chat Jimmy as the people listening to this and you guys have just seen goes at 16,000 tokens a second. So it's 20 nanoseconds it took to do those 50 different things in the style of Shakespeare. This is a Llama 8B model that they have. So Talis etched directly onto the silicon. Again the idiot tax of a GPU is gone. They're going to get DeepSeek level, V4 level models by the end of this year because of their first tape out onto that silicon at 20 to 50,000 words per second. At that point, it's actually the speed of light is the issue of getting the tokens to you.

I asked it to tell me a short story about Harry Potter discovering MDMA and it spat out 15,000 tokens in 0.04 seconds.

Oh, yeah. Yeah, probably. Yeah. Holy crap.

Yeah. In certain kind of economic things, it's just like HFT. They literally buy cables under the sea and capacity on those things. You collocate that and double zero obviously is trying to get the high performance bandwidth infrastructure kind of out there. And what you just saw was one of those W2F moments, right?

You know, like another example of this is when you use NotebookLM now, right? And so NotebookLM used to create these podcasts. Notebooklm.google.com. Then you could dial into the podcasts. So there's a dial-in button so you can talk to the hosts. Now it can generate full interactive videos and everything like that laying out the most complex topics. And you're just like, there's always that thing where you have to reset your prior, but it's inevitable that the agent economy will be transacting at incredible speeds with no utility function and like one-shotting just about anything it needs to build directly into bytecode. So, like right now they've got this new CodeX which is now like self-recursively improving. It's gone dramatically better. Just introduced a Ralph mode. And so I've been testing it out. Like literally I just cloned all of DocuSign, right? Like these things will be like nothing very soon. So how are you going to compete against that economically? But then from a monetary velocity perspective, the vast majority sometime in the future, maybe sooner than we expect, of economic monetary velocity is going to be stable coins for agents interacting with each other.

Like you know, I got into crypto in 2013, lost so much crypto over the years. I was like the decentralized side and then obviously like we're open sourcing the AI, had 300 million downloads, like 6 million a month in revenue a couple of years ago. I just like got to quit because this is inevitable. So Bezos has this framework of the unchanging and the inevitable. People will never not want faster, cheaper, more choice, you know, and it's inevitable that all commerce is going to be the internet. A lot of commerce is going to be the AI. It's inevitable that even if you stop today, the AIs will be the vast majority of the economy. Digitally and physically, right? And already they're at the point that again, even if we stop today, we are the dumbest people on the team. Like already in clinical trials, a human underperforms a human with search, underperforms a human with AI, underperforms an AI. Just like a self-driving car. Like if I think my latest analysis was that if human drivers in the US drove as safely as a Waymo, 60,000 less deaths per year or hundreds of thousands less injuries. But again of course you know like those things drive safe. We are going to be the dumbest people there. So then there's going to be a question of what do the truck drivers do? What do the paralegals do? What do the software developers do? I got a lot of flack especially from India a few years ago. Like in 5 years there are no more coders because again why speak to a computer in this language? The AI can speak better and it doesn't even need the language, it just speaks to the code in fact.

I mean, do you look, the best coders I know don't even look at their code anymore.

Exactly. Like write it in Rust, you know, like all those days spent doing what is code, right? It's like these things are universal translators. So I think that for me the idiot tax going down is good because our infrastructure improves digitally and physically. And again, look at the Star Trek future. In Star Trek, they don't have money. They can replicate anything. You know, they explore the universe. They still do stupid drama, especially Picard season 2. But, you know, like it's pretty good. Against that, you have a lot of people will need purpose. We need to look at where money goes because the AI will gather money faster than us. So, I looked at that and I was like, we got to prepare for that future. And that means giving everyone an AI that's sovereign to them through an appropriate means, advising our governments appropriately with AI because obviously AI is going to run the government rather than a big lab running it, you know. And then thinking about money, like right now money is created by banks, right? And so you put a deposit in a bank, they issue credit to create the money. The AIs will have the best credit rating of anyone. Why would you lend to a human when you can lend to an AI? Right, for a start. But then how do humans get capital if capital doesn't need them anymore? I don't need you. I'm just going to buy a bunch of GPUs, right? Like already I'm sure a lot of the people listening to this who have companies are like, 'Why do I need to have humans anymore?' You like even with us, our hiring plans, we're like, we don't need to hire anyone anymore. You know, like our teams are big enough and we're doing like loads of things. So then I think my ultimate thing is it's valuable to be a human and money should come from being a human. Everyone should have their own AIs that look out for them, self-sovereign, and coordinate all the other tokens that are coming into them. And our institutions should also have AIs that are aligned and that's infrastructure. So that's kind of where I got to and that's what we're building towards.

I think yes, but like now you need to actually go in and integrate that with existing society. So the network state was about like a bit of orthogonality to existing structures. And again we've seen the building blocks start to come together on that. I released my book The Last Economy for free last summer. It's catching up to the network state. Another couple of months we should catch up on Marie, which is good. And I put a proposal for like small sites of nucleation to get this going. But then over the last six months I just changed my mind completely. I was like every single country actually wants sovereign AI now. Individuals want it and they're feeling a lack of representation. In fact, what AI is like is taxation without representation. Like they took or we took all the data again as an AI builder and then trained on it and it is reducing the future net cash flows of our students and experts and others. But where's the representation in this? Like again we open sourced to have a level of representation and transparency because everyone could use it. It was choice. So now what we're kind of doing is we'll be announcing a whole series of what we call state champions in the next few months. And the idea is this AI is going to kind of be like a utility mixed with a broadcaster, you know, like the agent next to you is the one that's most important. The agent next to your representative is the one that's most important. So, we're building a full agent to policy stack and we were like country by country, state by state, let's just capitalize these things at $1 or 1 and let all the locals invest in the first round, the multinationals in the second and then list them. Get a blob of cash and compute in with AIs to handle governance and hear everyone. So you have the AI company of Utah, the AI company of Bangladesh and others as part of a network that can intermediate these quadrillions of tokens coming in because again the AI that's closest to you is the main one. And governments and more have been super happy about that because they're like, 'Can that allow us to use ChatGPT as well as like you know have our own models that reflect our culture and things?' Yeah, that's the job. The job of these institutions will be to do that and it should be locally owned. It should be locally controlled just like you know you don't let anything rebroadcast to your kids. You don't let someone else's curriculum come in like...

Yeah, it's not optimal, right? And you want to have a say in this as well because I think the taxation of AI licenses, you know, blowback is coming because again this is coming for people and they don't have a say. And so I was like one way you could do that is you could launch a lot yourselves, but why not have it 100% locally owned? You know, take the lessons that we've had from DAOs and others and apply that. So that's the infrastructure we've been building. And again, that's the announcements we'll make because a large part of the AI, public sector AI, shall we say, education, healthcare is fundamentally local. You don't want your healthcare data going across borders. You know, you don't need your education data going there. The models are generally good enough, but also if I have a model to educate my kids, I don't really want it trained on Reddit, you know. So, we've also been kind of getting ready to build really clean first principles models of a new type where all the data is you can check it yourself. And then finally, the thing is you want to be able to opt out of this as well, which is why it's important to be able to run that stack yourself. So you get it from your state champion or you just take your data and you run your own stack, you know, because you don't want all that data to be in some server even if it's run by a local champion.

Yeah. So you get a level of universal basic AI as we call it with self-sovereign identity protected by ZK that can run on any blockchain etc. But then your data has portability so you can run the intelligence at home and never have to rely on anyone because verifiable inference like you're generally going to be protected but in certain countries and states you won't be protected, right? This is the optimal way to spy on you and to tell you what to do especially because that AI next to you is going to be the one that really influences you. Like you will trust the AI that's next to you just like in Her, Joaquin Phoenix trusted Scarlett Johansson.

Spoiler. If you haven't watched it by now, you know, it's okay.

Yeah, that's why AI movie club. All the science fiction is becoming science facts very very quickly.

Yeah. I think that you need open data sets, open models, full transparency for that particular area. And then this whole world of private AI that you can use or not use this as the interface for, right? But the interesting thing right now is that there's huge demand for equity. Like again, it can't be matched. There's like a trillion dollars of demand. Like Anthropic is about to do 50 billion at a trillion dollars. Actually, the amazing thing is do you know how many subscribers of Claude Code there were a couple of months ago? 11 million. And you think about that and you think about like that's a whole lot of worlds to go, right? So, but there's huge demand for equity. And I was like again if we just do it at $1, everyone's in the money already and everyone can participate with crowdfunding regulations in the UK. We can tokenize the security but then it's actually an institution that you're proud of that you can use AI itself in. So that's why we're building policy AIs and more because we currently lack anyone to really guide us through what's coming and you need to create a receptacle whereby if you do this in you know like Nebraska or like Utah or again the UK you want a vehicle you just get the smartest people in and their job all day is to think of how can I use AI to make Britain better to make Ireland better, Bangladesh better etc with this type of structure a fully open one and you can make money by intermediating the tokens because again it's a cash flow business. Open AI tokens go in and they use them if they want. You can make money by going out into the community in Wolverhampton here or Bristol or Omaha and helping integrate this physically because again that's a demand and those will be genuinely valuable institutions that are not state, they're not classic private, they're somewhere in between because we've not actually interestingly had this right like again before we talk about tokenized equity and DAOs and things there isn't a single DAO that is organized by AI that I've seen. There isn't a single company that has proper tokenized equity that people can actually participate and guide. And so again, I think it's time.

Yeah. I mean like I view in the next year the price of tokens for the same level of intelligence will drop by 100 times and the demand will probably increase a thousand times. But you could be wrong, right? Like you could see a breakthrough that takes it down a thousand times and then like you have a massive build out at that point. No one can afford to be wrong on this. Like GPT-3 when it came out was $600 per million tokens. GPT-5.5 is like 30 and Opus 4.7 is 30 whereas DeepSeek 4 now is like 50 cents per million tokens.

We have a good idea about it. So let's put it this way. The average person speaks 10 million tokens a year, right? And we know how much these companies cost per million tokens. So GPT-5.5 costs $30 per million tokens to buy on the API. Again, if you have a subscription, then they're subsidizing you more if you're a heavy user, probably up to 10 times, but they have about 80% margins on those. Someone like DeepSeek right now is running almost no margin at their 50 cents level, right? So it costs about a couple of bucks per million tokens. So for all the words that a person speaks per year, which is much more than the average ChatGPT user, it's about 20 bucks to provide ChatGPT Plus, you know, and again that's going to drop by 100 times next year reasonably, but it could drop by a thousand times. And then you could hit this point where you have enough tokens for the vast majority of people who are not building 23 different things like David, etc., right? But probably not. But sometime in the next couple of years, we probably have an overbuild out for text tokens and it will take a while for pixel tokens to go and then you will get a pull back. But that's why again I built decided the state champion thing because it's good for state champions because it means their cost goes down. Like their plays on the index of this and the extra demand on this effectively. But you could have OpenAI and Anthropic easily in a year or two in a lot of trouble. Like full replications of Claude, which is a really nice experience. CodeX for me just beat that yesterday, right? But then to rebuild that experience from scratch, I'm sure that Venice and a whole bunch of others will be able to do that by mid year. Like experiences aren't the main thing anymore. Then you have other things coming in. And again, like those 16,000 tokens a second you saw were 10 times cheaper than the GPU equivalent. They think they're 100 times cheaper. So you get some crazy economics as you optimize it. Like literally you're just etching on the silicon.

Well, and I think the point here is you get disruption.

It had to get to satisfy sync quality, good enough, fast enough, cheap enough. And the internet boom had to build out the end portals. When the AI gets smarter, it gets smarter for everyone all at the same time. And again, you have these levels. Stable Diffusion was the first image model you could run on your computer and it was like 8 GB of VRAM or something like that. All right. Now, what you have is Qwen 3.6 27B is the first good enough model that you can run on the edge like not without multiple chain 128 GB ones. Like very reasonably you can run that on a 24 GB Mac and it is really good intelligence. It's like Frontier from a year ago. And so by next year you will have a 5.5 level intelligence that you can run on a MacBook probably trained in 8-bit or 4-bit natively and that's a huge unlock because what we haven't seen is this when I download a game like I downloaded Pragmata, you know, the little girl and the guy, it's been quite fun and it's like 45 gigabytes. And you think about a really high quality LLM wrapped in a CodeX-like experience. It's just like a game download actually literally that also has an image editor that has music that has other things like 40 GB of download on your computer gives you something that can control your entire computer and that's a huge unlock. Like you just had the Chronicle mode come in on CodeX. So I don't know if you remember there was this piece of software called Limitless that looked at your screen and took a picture every so often so you could search back in your history. I don't know why they stopped that. Chronicle does that. It actually takes screenshots of your screen every so often, so it has context of everything you're looking at.

Exactly. I mean, the US government's been doing that for ages and Five Eyes and things like that. But if you think about the experience of using Linux, you know, or anything like that, it was so much better once you switched to cloud code. You just do all of Linux that way. You think about the experiences of the future where you just download something, your computer suddenly becomes smart. Like you don't need a new computer for that, right? Like this is my MacBook M1 Pro from that's more than good enough to actually run all of these agents. It's running some household things now. That's like seven years old. So the question is when do you get to that level of good enough, cheap enough intelligence that you have the installed base to roll it out everywhere. No one needs to buy new things. And the reality that this comes in waves. So one of the most interesting things right now is if you go to trends.google.com. You can actually see ChatGPT topping out in terms of usage or terms of searches and it actually correlates pretty well with some of these things. So all the easy users have kind of been done. Now it's the next wave and the next wave and the next wave because the internet kind of evened out, but who are the next users of generative AI? How deep are they using it? These are some of the really interesting questions here that will affect the economics. And like I said, I think you know like Anthropic and OpenAI could be in trouble in a few years because they've made these long-term commits, but OpenAI only has like two years of runway despite raising $120 billion. Like what happens if the users go down? This is a question of their pricing power and everything, but remember two three years ago they were at zero revenue. So how sticky are these users? Like Anthropic is still the best, but do you start to see substitution as like Uber's CTO or CEO said we've already used our entire generative AI budget for the year, you know, like...

Yeah, I think there's a couple of things in there as well. Like you know you have a mobile phone like OpenClaw took off because you could just text it from your WhatsApp, right? And so a lot of people are just going to interact with it by the phone because it's natural like you just WhatsApp it, you Telegram it, you Signal it etc. You don't need any new interfaces. Then it's about how do you get the cost of tokens down as low as possible for the equivalent thing. Number two, it's actually really hard to use tokens. The most tokens I've used in one day is two, I think about a... No, 1.8 billion and I was cranking that, right? Like it's really hard to actually use more than like 10 million tokens a day, shall we say, when you actually add it up. And most of those tokens again are used uselessly. When things can one-shot everything, then at what point do you need more tokens? Like I think 4.7 is one of those last bits where it's used some more tokens but the token efficiency has apparently offset the thing. So there's a chart that shows token efficiency versus it uses three times less tokens than 4.6 and it's 50% more expensive. I think it does one-shot more things that I've seen. It has a terrible personality so I don't use it. But you know like again you got this thing and but the...

Well again, I've been exploring the psychosis of GPTs with physics. So it gets a bit insulting after a while.

It's going to keep on going, man. Like again, this chart, all these charts are insane. Like look at DRAM prices. There was zero interest in 2024 and all of a sudden this massive spike since then. So...

Yeah.

Actually it's interesting because the other thing that I think people are misunderstanding the chips are actually fast enough now. Like okay you have the Cerebras and the Groq of the world, the Cerebruses that are pumping tokens fast enough but again it's this experiential thing where if you have an A3B model so 3 billion active parameters 30 billion on a MacBook M4 with 16 GB of VRAM in 4-bit it's a very pleasant experience, right? The intelligence comes to you as fast as you can absorb it. Do you need faster silicon? The answer is no, which is super interesting, right? Like again, if you do the same thing on a 16 GB graphics card, then it's like 80 to 90 tokens a second. It's faster than you can read. So I think that will be the other interesting thing. Like that's why DRAM went up so much because people like I want a bit more memory. You know, that's why Apple Mac Minis are out of stock and things like that.

Yeah. Clayton Christensen, famous Harvard Business School professor came up with disruptive innovation said infrastructure is the most efficient means by which a society stores and distributes value. And then information theory is that valuable information is valuable as much that changes the state. Our infrastructure for knowledge for organization is terrible. The AI can make it smarter. You know just like DAO infrastructure and voting is terrible. We need more intelligence in crypto in general, right? Like every DAO should be using an AI. Like I was talking to some people from BitTensor and they were like because I had some governance things and some people leaving. I was like why don't BitTensor should be fully run by an AI or everything should be checked by an AI. They're like, well, you know, come on, you know, like lean in. The organizational capability is why I thought, you know, if we have state champions running the same stack, you can organize better so nobody is left behind. And we've been kind of testing some of these things with various initiatives. So there's a charity called Imagine Worldwide. I encourage everyone obviously donate to. One of my former co-founders at Stability, every child in Malawi is getting a tablet this year. Access to intelligence to teach them literacy and numeracy. And those will get more and more AI. It's $7 per kid per year, including solar power. If you gave every kid in the country an AI, you could organize things so they're never hungry again, right? Because they can talk to you. Now, with Starlink and Solar Power, it's like less than a dollar a month potentially to do that. But you got to actually deliberately build that infrastructure. And that's a bit different from a ChatGPT infrastructure. That's why, like I said, we're like have an initiative called Sage Sovereign AI governance engine. We're taking all the policy knowledge in the world and making it free to all the governments. So, lots of announcements of that from June onwards this year. You got to build this stuff intentionally. But then when you do, I think the unlock can be huge because the AI can see across like nobody wants people to be hungry or to be left behind or nobody wants all this stupid red tape that you keep seeing. I mean sometimes like it looks like positive the US clinical trial guidelines and the whole thing on clinical trials. It looks like they're about to introduce AI to reduce that. Yay. Let's get longevity and cancer drugs into the market like but it needs to actually have someone deciding to do that, right? So I think this is a great time whereby our infrastructure has been a bit stupid. We can upgrade it and that's personal infrastructure, company infrastructure, societal infrastructure but again that type of thing needs to be deliberately built like open core and things like that are great but if you don't know which questions to ask it's actually quite difficult to build things and in fact one of the key barriers to using a gazillion tokens having a gazillion bits of output is just do I know the right questions to ask it like again most people are never going to use a billion tokens a year because they'll be like I don't... Oh, like I just want to live my life, right?

Do you need a UI/UX? It should adapt dynamically, right? Like recently on Twitter, I saw these guys took the LTX video model and automatically generated websites like there was the Tower of Babel and kind of all those kind of things there. Like this is where we're heading. Like you see CodeX now actually the latest push they did yesterday is starting to build its own UI dynamically for you. And they have things where again that be dynamic the way that you'll interact with your agent. If you go to like Pika, for example, you know these guys started with video models now it's like create a Pika agent and what they have is you know you can customize it the way it looks and chat with it but they also have live avatars you can also plug that into your OpenClaw so it's pretty expensive right now but you can talk to your OpenClaw like this. You know again meet you where you are you can pick up the phone to it. And the people that are successful in open core are actually the ones that treated it a bit like an employee. Like your personal assistant, if you just give them a few things, you don't spend time with them. You can't really onboard them properly. Yet somehow we expect the AIs to do that, right? Like why aren't you doing it properly? Why aren't you customizing to me? And so the way that people onboard is literally like just having a chat, you know, like when you install Excalidraw skill or Figma or something like that, like your agents can now literally do entire mood boards and other stuff. They can use Notion to do presentations that will pitch to you. Like you'll probably just dump everything you know into there or connect it and it'll come back and say, 'Hey, David, this is how to organize your life. You know, you don't need 23 agents, you just need me.' And it will be having a chat that way.

Again this is the intentional building, right? Like how unique are you as a snowflake, you know you got this, you got a bunch of stuff that is actually very normal. And these are skills that you kind of embed. So something like we have our agent platform III Agent that's about to get a whole host of upgrades and all the functionality and things like we built that very intentionally to do stuff by default. So it builds amazing presentations. It builds amazing websites, but then it's going to have a whole bunch of things like wedding planning and other stuff, you know, end to end as like fully integrated skills because we're like when you go out into the world, you wear different hats at different times. Yep. So why not just have a skill for religious institutions, a skill for weddings, a skill for organizing family vacations, you know, and that reduces the number of tokens you need, but increases the quality because again people in that I think again that's what people will build and that's how customized verticalized software has classically worked. I saw that there was I think it's called Owner. They're a CMO platform basically for restaurants and they've increased restaurant revenue of their customers by a billion dollars. So now for any restaurant they automatically create the website, all of the marketing materials, menus, etc. You're like, 'Okay, that's focused on restaurants, right? They're going to do a great job for restaurants.' Like you're like that's really interesting because again although you can build everything building the right thing to remove the frictions and then getting value from that is incredibly powerful like those restaurants aren't going to switch and run their own agents and things like that, right? Like again the tail is long for this.

No, but I think Emad's making an important point is you're going to need sherpas, advocates. You're going to need people to fill it.

Depends on how many Mt. Gox and things you've been in.

People, I told people none of the exchanges I used in 2012 still exist. And the only reason I have any crypto is because of self-custody. But to your point, Bo, it's been such a technical barrier that for most people, that's when they just walk away.

But this is one of the first things I rolled into the Morpheus skill was a wallet and the self-custody aspects handled for you by your agent. So now you didn't have to be a security expert, right, to check that credentials were stored correctly and all of these other things. And that took a lot of effort. But if we, like you're saying, by intentionality can set it up for people in a default where it is safe, all of a sudden they get those benefits without having to have all of that technical acumen.

Yeah. I think again this idiot tax thing, but there is a value thing like people's cognitive load is quite minimal. Like right now, how hard is an agent to use? It's not that hard. How hard is it to master? You have to know it pretty well. Like power users like ourselves, we know the ins and outs of where the AI breaks and what's it good at and so you got a multiplier effect on that. Whereas the vast majority of people never get there. Vast majority of people still use the free ChatGPT and think that's what the best thing is. And there's a world of difference between 5.5 base and thinking and pro, you know, and codex, etc. It's like how you use Gemini, the Gemini app is really bad. AI Studio it's better, API it's pretty good, you know, but again most people aren't even exposed to the top level AI. So I think the frictions will continue reducing but then you get to a really uncomfortable thing in a year or so and I put like there's about two years max for this from almost every single job you can do on the other side of a computer where the CEO pushes a button and then the AI agent sweeps every single thing every employee has ever done and all their details and creates digital clones of all of them that never sleep, tax deductible, and the company keeps running and nobody knows they've been replaced. Who's been replaced? You can kind of just imagine this as kind of a skit. Like literally...

Oh, they don't... Oh, you're saying they don't realize that it's not actually their coworker. Oh, that's good.

It'll be one button push and then Bob is replicated.

They're going to replace every Meta mate. And Mark in his kind of thing said you on average you guys have a higher IQ than most data labelers. So we're going to train on you basically to replace you. Like what the hell? Like of course they're going to replace all these people. Like they're going to do what Jack Dorsey did but it's not going to be minus 40%. How many people are needed to run Meta? 10 times less.

Yeah.

Thought of even simpler. If your company, if your job is to be like a machine, the machine can do it better.

Ready?

And I think there will be people who do that. I joke but I actually do think there will be groups of people who do that who are basically like, 'All right, we don't use AI so we use everything else other than AI.' I don't think it'll be that big but I think you will see a big increasing backlash. A lot of the top CEOs in AI I know now have canceled almost all their public events. You know it's like if you're in crypto now don't go to France, you know, it's the same thing.

Sanders right now. He's like, 'AI is going to kill us all.'

I'm trying to get it evens, right? Like, I mean, there's so many ways, you know, like my most concrete way is you have a billion robots and a bad firmware upgrade and they twist off our heads. You know, I've got a very active imagination on all these other things. But the other side, like I said, is Star Trek. And the key to my robot and to, you know, having pushing the button on that firmware update, you know, not be a one central point to your point is important.

And we kind of got to see these things. Like, have you guys seen the Engine AI T800?

So, first off, they called it the T800. Right. You're a Terminator. Yeah. So, the T800 is like 6 foot tall, weighs 80 kg. And can punch five times harder than Mike Tyson. And it's $20,000. So, it's like, why would you even do that? Like that could punch through someone. And so all the videos of it are doing kung fu and things. So they have a robot fighting competition coming where they have the full sim to real thing whereby, you know, Real Steel. I don't know if you've seen that with Hugh Jackman.