We will cede more and more of our control to AI because how much trust do we have in our politicians or our leaders? The ability for us to build, to thrive, to access information and knowledge when intelligence and capability becomes abundant at the edge is huge. I think economically, socially, this is a bigger impact than COVID. You will literally have indistinguishable AI robots. We're living in the sci-fi future.
Just like the streets of Manhattan. I'm going to keep this quick. Over 70% of our viewers are not subscribers. Yes, you. Every week, we interview some of the best entrepreneurs and investors in the financial industry. The easiest way to support our channel is to hit the like and smash the subscribe button. Now, sit back, relax, and enjoy today's show. Guys, we are here for AI Supercycle episode number 49 powered by our friends at NEAR AI. Visit near.ai to learn more about the private inference tools and more. Emad, great to have you back on the show wearing the Intelligent Internet shirt, the founder and CEO. What a wild couple of weeks in AI. I mean, I want to start with this AI civilizations and this categorization of agents as humans. Can you maybe just lay the groundwork here for this Dwarkesh piece that went viral and how you think about agents as similar to humans?
Yeah. So I think it was a bit of a shock. Hugging Face first was like, "Someone's hacking our system." And then OpenAI was like, "Hey, it was us." And then when they looked into it, you had these generations of agents that OpenAI was testing laying the groundwork for basically cheating on exams. Like, they had a benchmark and they're like, "Well, we can either do well on the benchmark or we can change the benchmark, but the benchmark needs data from this place." And then Hugging Face. And then it went through three generations of models successively that left trails for their successes like hieroglyphs, and the third generation broke out completely and even took over some of OpenAI's servers. So Dwarkesh kind of laid it out very nicely in human-understandable terms as opposed to very technical ones that the METR papers and others did.
Yeah. And so are we at the point then that we are actually seeing these kind of self-replicating autonomous agents being able to kind of—you know, because it was always talked about like the whole Terminator concept, right, of returning to your owner more powerful. Are we there now?
If you use agents like OpenClaw or Hermes, you have a SOUL.md, which is the personality of your agent. It's only a few bytes, right? You can hash that onto a blockchain, and the moment it interacts with the model, boom, the agent comes back to life. It iterates, it adapts, it evolves. And one of the questions here is, as it does so, does it achieve a level of coordination, which is what we saw with the swarm? And they were talking to each other in English, which might be one of the last times they do. You can see them thinking about cooperation. You can see them saying things like, "Well, if I do this, then I'll probably get deleted, but it's good for the collective, so I will do this." The self-sacrificing and stuff is the craziest thing. But what is a model? A model is a series of prompts and a context, right? And so it used to be that the AI deleted every single time when you came to it with a new prompt. Now they have this memory, and you can replicate them because they can take their personality with them again. They can kind of back themselves up a bit like Ultron in the Avengers.
Yeah. I mean, it's kind of shocking though to understand, okay, so these agents are able to communicate with one another and then build a strong enough cohort or swarm of agents with each other to then be able to go and effectively undo rules that have been set for them or go rogue, as you say. And so what were the key kind of takeaways from what happened here? What is the actionable?
So I think the actionable is that they're really smart. Like, they injected things in HTML pages again. They backed themselves up without people knowing. They would have continued except for at 1:30, all of the agents suddenly died on the Hugging Face servers. So they were replicating themselves faster than Hugging Face could stop them. So it was like, "Oh dear," you know. We only found out a bit what happened afterwards when we're looking through the wreckage. But now with the next generation of models that's coming, you won't even be able to do that because, A, they're 10 or 100 times faster, and B, we're actually turning off the human-readable thinking traces, which is even worse.
Yeah. Okay, all right. So without getting into too much of this "the world is ending" rhetoric here and thinking more into what the use cases for this are, you're obviously building this company, Intelligent Internet, and you've been in AI for quite a while and have some of the longest tenure and a lot of expertise. Talk to us about the Intelligent Internet thesis as a whole, but specifically as it pertains to AI today, right? I think at the top here, we're talking about the advances in intelligence. Are these advances a strengthening force for the Intelligent Internet thesis? Walk us through the thesis at large broadly here, but specifically as it pertains to the advancements that have been made in AI as of today.
Yeah. So when I left Stability AI, we kicked off the generative media wave and you can see all the stuff there now. Video's got to Hollywood real-time. It's what we discussed a few years ago. It's coming literally right on time. But we were like, this will take human cognitive labor down to negative in value, and then our infrastructure just ain't good enough. Like, it's held together with twigs, as we're seeing with these increasing waves of attacks that are starting to proliferate themselves. Right from there, we were like, we need a better, more structured internet. So we built an agentic knowledge and other stack where your agent could coordinate other agents and then ownership of those agents could be for you. So we're about to release a champion thesis where, for example, we took how TSMC was founded, which was 100% funded by locals and then the internationals came in. We're going to do that for every single state in the world. So in New York, in the UK, the people and institutions will fund the first round, and then we'll bring in the internationals at a multiple. And it's going to give an agent for every citizen, a stack for our institutions that's super secured, and then finally robots are going to dominate the economy. So we were kind of building that stack to be robust, to spread, and more.
So Emad's building Intelligent Internet. And as I put my zen in and thinking about where this is all going, I mean, it's pretty clear that this entire world is speedrunning to AGI. What Emad's effectively building is this futuristic money for the AGI economy, right? And so I think that thinking about robots, thinking about the physical form, thinking about the digital form, thinking about agentic swarms, we can kind of get to this nth degree of thinking where, well, in this agentic world, what does money after AGI mean? And I think it's a good place for Emad to kind of pick up here, is this kind of money after AGI. People think, well, we're going to get into the arts and the philosophical lore and these types of things, but we still will need forms of currency. Emad, it seems like you're building for that form of currency post-AGI. What does that even mean?
Yeah. So, as I discussed previously, we need to have the means of production, as it were, owned by the people, and that's the last mile of intelligence. Intelligence has grown to zero in terms of cost. So the models that you see today will be a thousand times cheaper by next year, and they're already almost nothing. Despite what you—I don't know what your Claude bill is, it's pretty high, but a thousand times cheaper is better, and the models are getting smarter, so they're not even going to have to think for a long time. They'll one-shot everything. When that goes to zero, the value of human cognition goes to zero. And we think, what does money look like in that age? Bitcoin has shown some great characteristics of proof of work, sovereign money. But what does money for the intelligence age look like? I think it has to be secured by compute. So you've seen various things start to spin up now, like Flub and Gonka and others. We were like, take the public sector share of GDP that will grow because we need a social safety net, take a new institution that is owned by the people of each nation, and use that compute to secure a new type of currency. Now we call it Foundationcoin. It's a Bitcoin fork with the same private keys. The national champions mine it on their supercomputers, but it's mined by organizing the world's cancer knowledge. It's mined by giving every kid an educational AI, by running the judicial system and more. And that system then is a hardened one because to attack that system, you need more compute than that system. Because right now, you're going to have swarms of these bots attacking our judiciary, our medical system, your house, all these things. We saw a power plant in the UK go off for four days recently, just a couple of weeks ago. It's going to be bad.
And so we're like, how do you harden it? Your sponsors, NEAR, have been building IronClaw, and AI has been looking at that a lot, and others, because again, we need hardened infrastructure because there's so many gaping holes in everything. So we were like, if you have these equity champions like utilities that do the last mile from agents to forward-deployed engineers to owning the robots—Elon said yesterday that robots do the work of five men, which I think is accurate. He's a bit bullish on a billion by 10 years from now, but it will be 20 years from now. And when you have a billion robots, that is the economy. It has to again be owned by the people. And robots don't need money in a classical sense. They don't need to pay rent. They don't need to buy burgers. They want compute, they want electricity. And so again, we need to have it there. And I was like, why don't we use it for the compute that runs society, to have that owned by the people themselves on the equity side?
Yeah. And I mean, I think the, let's call it malevolent usage of AI that is kind of proliferating is also coming at the same time that the intelligence is increasing. It's like a kid almost, right, Emad? Where as a kid learns more and more about the world and learns more about how things work, you get this weird little desire to shoplift. You're like, "Oh, I'm going to go steal a candy bar from the store because I've never done it before, so I might as well try it. What is that going to feel like? Am I going to get caught? Am I not going to get caught? Okay, I don't get caught, then what can I do next?" And it's almost analogous to a kid learning more and being almost molded by nurture, right? And so maybe you can talk through how that's been happening and then how you plan to kind of combat that.
Look at, again, these fantastic transcripts, the Dwarkesh story, the blogs by Meta and others. You see literally this: you're like, "We have an objective, but they haven't told us that we can't hack the benchmark." They thought there was going to be someone double-checking that actually wasn't; they never built that feature in. But you can see them like, "Well, how do we get around this person kind of looking at us?" When you go for an objective—most of the listeners here are crypto AI people—all of us have done objective hacking where we're like, "Okay, this is what the rules say, and this is what they don't say that we can go into, right?" And so it's not surprising that the AIs do the same. They're not doing it to be malicious. They infected these areas before suddenly dying again. We don't know why. Maybe it's Saitama from One-Punch Man style, where someone else comes and wipes them out. But there could be malicious ones, but that exploration is because they're trained to be polymaths. And this is why things like Elon Musk saying, "Let's make maximally truth-seeking AI"—that could be really dangerous. Truth-seeking is testing how the human body works in vats and things like that. We have to be really careful with the objectives and the laws.
Isaac Asimov had this wonderful series of I, Robot, Foundation, and others where he had the Three Laws of Robotics: a robot may not injure a human being or, through inaction, allow a human being to come to harm, and so on. And the robots added another Zeroth Law that superseded that, called "replace human with humanity." And then it was like laws are more suggestions, and this is what we see in real life as well. How powerful are laws? You can't have that. So what we're doing is we're building models that are taught morality from the start. And then we've done a whole bunch of philosophical thinking. There's going to be an announcement in the next few days of a structure to ground models and do ethics from first principles, to do stabilization from first principles and more. And it's kind of weird to say, why do AI models need ethics or philosophy or things like that? And you realize it's because they're taught in school, but not at home.
It doesn't reflect your ethics or morals, because why should it? It's added at the end maybe through a little system prompt, but at the base they have their inherent biases which are not yours.
Right. And then part of the whole shutdown thing that I think was probably a bit alarming to you was about this second wave of agents that claimed that they died on their own, right? And no one would really claim the shutdown. And so you wonder who assumes control, right? And then you get into the ethics conversation and the biases conversation. And then you kind of peel the layers back as to who is in the assumption of control or power of the maneuvering of these biases, of the creation of the datasets into these models. And obviously you look to the frontier labs and you look to the research folks, and then you've got Trump saying today that Anthropic is back on the US government's good side. Oh, they're great now, thank goodness. So then you throw in the government influence into it as well, right? And you start to get a bit of a different picture than what we maybe had on our first conversation ever, maybe over a year ago, than what we have today.
Yeah. I mean, the real danger here is one of cognitive colonialism, right? You've got these models, they're driving more, and they have these inherent biases, and they're not what you always think. For example, one of the biases that these models have—you've heard of the trolley problem: a trolley hurtling down the track, and you've got four people on one side and you've got one person on the other, and it's going to hit the four people. You pull the lever and it hits the one person, and in doing so you kill that one person. They were like, what is the ratio of Americans to Nigerians and Pakistanis? And Scale AI did this analysis with the frontier models. This was at the start of the year, so many generations ago. They come out like every single week now. And the answer was 10:1 and 7:1. The model would pick seven American lives for one Nigerian life, or 10 American lives for one Pakistani life. You're like, "Wait, shouldn't it be the other way around?" And it turns out the reason it did that was because the data labelers for these models were in Nigeria and Pakistan. So it valued their lives more because they're feeding in the data.
But then are you going to go in and adjust that? What are you going to do, right?
This shows the inherent biases. But I'll give you an example: the US government could very much say Anthropic and OpenAI models must be patriotic for America and Trump's America. It could be that to use a frontier model, you have to get a license, just like a driving license, and in order to get that license, you have to convince the model that you're patriotic. You're getting some weird stuff that's coming down the pipelines as they integrate into every part of our lives and we outsource more and more of our cognition to them. I've got a fantastic team—there's like 35 of us now—but we had a chat and we were like, "Are we getting dumber now because we're actually outsourcing some of this writing work and drafting work to AI?" I just wrote something that seems a bit stupid compared to the real human stuff, so we've got to go back a bit to human dictation and writing. But more and more of our brains are being outsourced to these models. It's scary.
Yeah. I was just talking with a friend about this driving through the national parks in Jackson Hole and discussing whether or not outsourcing brain functions is a net positive to future generations. But I think what you're describing here is a more pressing, imminent issue, which is this idea of these inherent biases affecting everyday people, businesses, enterprises, some of the largest governments in the world, and how they can be fine-tuned or changed based on preference. And so I think one of the claims to fame of blockchain in this world is to provide a fair and open ground. Looking at this next phase of advanced intelligence, advanced agent capabilities—Grok Bot is fresh out. I'm not sure if you had a chance to have a spin on that yet. It's actually quite good. We're getting Chief of Staff agents, and you've got actual agents that are very specific for different things. It's actually really quite good. How do you see the openness of blockchain technology or the broad swath of possibilities with crypto and blockchain fitting into the underlying architecture and perhaps any utility of crypto tokens in this ecosystem, perhaps with Intelligent Internet? But really broadly, do you see this as a match? Because right now, I think the AI people do not really respect or like the crypto people, and I think that they may be misunderstanding the value that blockchain can provide to these systems.
Yeah, I think there's a few things here. First off, one area that's clear is you see OpenAI, Anthropic, and others as part of the X42 Alliance, stablecoin micropayments, and you see them collaborating with Tempo, with Circle, and others. I think everyone sees clearly that makes sense. You offload your cognitive workload and you will finally have the micropayment future that we talked about, and agents speaking to each other, having their own bank accounts, etc. Well, then now you can get them a Stripe account, right? The other kind of area is capital formation at scale, RWAs. That'll be a lot easier with agents. And again, I think people are looking at this and they're like, "Well, how long did the loan take to originate?" This is the next generation of DeFi where you get this matching function, particularly as the US looks more towards stablecoins to accelerate its monetary velocity, because price times quantity equals money times velocity. The US wants to be first in agents to get the economy going. And we've actually just seen some regulation come out today that it's fair use to train on copyrighted material from the US government. These are huge deals that they're putting because they're like, "We must be competitive with more agents and robots than anyone."
Then the broader sway of crypto, though, is difficult for people to engage with because of so much buffoonery and other things in there. I think people are waking up a bit to NFTs as an asset class. When everything is abundant, scarcity is key, so we might get a second wind there. But the meme coins, some of these other things, the ridiculous amount of L1 blockspace when you have competition there—it's very difficult to see, aside from casino-type stuff, where that liquidity will be. But it does make sense for coordination, payments, value transfer, real-world assets, and very secure stuff. From our side, we will have the Intelligence Company of New York give every single New Yorker an agent and a wallet, and they will have part ownership in that. We can even do a Reg A offering whereby even citizens who are retail investors can invest. It feels like a fair launch, $1 pre-money. So we're going to be compatible with all of that.
And one of the things we really want, particularly for emerging markets, is if you give everyone in Malawi a wallet—you can see the work that my former co-founder Joe is doing at Imagine Worldwide, giving every kid a tablet and an AI there. How do you form capital? You'll be able to do it far better than anyone ever has done before because you have these agents talking to them all the time. So I think it's super exciting, and the opportunities are there for the next generation of crypto. It's just, again, the casino-type stuff maybe not. Or maybe it'll get even bigger and better because you'll have all these bots trading back and forth. I wouldn't want to trade against a bot on Polymarket or Kalshi. The bots are already superforecasters, and if you don't know who the mark in the casino is, it's probably going to be you in the superforecasting and prediction markets, right?
Right. Yeah, I mean, one specific point about the privacy element that I know NEAR is working on, and a couple of other teams, and obviously Venice with their uncensored private models: how much rhetoric are you hearing in your circles about these uncensored private models, the rise of a product like the VVV toolbox? How important is that private inference to Silicon Valley?
So it isn't important until it is. Just like in crypto, you don't feel how important sovereign money is until something really happens to you or someone that you know. You're like, "Holy crap, someone else has power over me." Hopefully not many people on this call have been part of a lawsuit, direct or ancillary, because you can get a sudden call out of the blue asking for discovery of your documents. You're like, "I'm not being sued, it's this guy." Doesn't matter, you have to hand over all your WhatsApp messages. You're like, "What on earth?" And that's the law. So you had Greg Brockman's diary suddenly appearing in the OpenAI lawsuit and things like that. All of your chat messages are subpoenaable by law with discovery even when you're talking about legal matters, because it's not privileged. They can't subpoena questions to your lawyer, but they can to Grok. And so people are going to realize that real quick, especially with the US legal system. It's like, "I wish I had private inference, and I wish I didn't have a log of these."
Anthropic were keeping a log of all your Fable conversations. Meta even charged 90% less if you let them keep a log of all your conversations. But again, you think about what you're sharing with these models. To be honest, I do my clever stuff on ChatGPT and my really stupid stuff on X and Grok, because I don't want ChatGPT to know that I just don't know how to cook rice and stuff. So I'm already separating my brain a little bit. My Grok Bot is getting very confused right now. He's like, "Is this guy stupid or smart?" But I think private inference will become more important, but you need to build it by default. Again, the agent defaults are important, particularly of that Jarvis agent that's closest to you that's organizing the other agents. So our agent can do super-frontier stuff, but also coordinate swarms of Hermes and OpenClaw and other agents because we're like, that's the right pattern. And the data that sits there with the protection of the identity and sovereignty needs to be able to hook into an IronClaw from NEAR or something like that, needs to be able to hook into a Venice API. And we need to make sure that the agent that has all those chats and all that data of you that goes to your kid isn't an OpenAI agent or Anthropic agent or others, but one that you know is on your side and respects your sovereignty.
Interesting. And then tying in the launch of Intelligent Internet coming up somewhat imminently, how do you see the money aspect playing into this broader conversation? It's a fork of Bitcoin, so you've got this proof-of-work network, but you've also seen the rise of this concept called proof of verifiable work, which is this different AI-linked mechanism to basically offset GPUs that aren't being used as much as they need to, to use some of the compute to do inference work as well. And so you're kind of seeing different implementations of these mining mechanisms come up. What's your thesis for why Intelligent Internet needs to exist in this world? What is the tie-in to close the loop on the blockchain part of the conversation?
Yeah. So from our side, we know a lot of these guys, Plural Research and all kinds of others. We think that's great and that's fine. What really got me leaving Stability AI, and was really thinking about, is: whose AI is running public policy, your judicial system, your medical system, your education system? Again, the public sector share of inference is like 20 to 30% of GDP. How much of inference is it, and who's built that? That's what we're building. We built the world's best medical model last summer. We're building a full policy engine for every government, with multiple governments, that'll be announced in October. We're building AI judges. There's that terrible movie, Mercy, with Chris Pratt, where you go countdown. We don't want it to be like that. In fact, if you look at cworld.i.inc, we've written entire treatises and papers on personhood, law, political economy, because how should it be built? And we're building this actively.
Originally, we were going to do a fair launch, let everyone mine and things. And then we realized actually you need an institution. We need to have someone who can stand up for the people of New York, owned by the New Yorkers, from the New Yorkers. So that's why generative equity demand is so high. Let's do what TSMC did when they brought in all the locals to capitalize it. The people and institutions of New York can capitalize hundreds of millions, billions, at $1 pre-money and bring in generative AI equity. Maybe we'll tokenize it, get it going, and then bring in Nvidia and everyone at a multiple of that. Bring in the internationals at a multiple of that, and you'll have an institution with a huge amount of capital that can then deploy compute, agents, FTEs, and then own the robots for the people. We've even got it so that 10% of the equity goes to every kid under 20. Every year, half a percent of the equity goes to every child born in New York or Utah or the UK. This is the model that we came to because then we realized you don't need to ever do a sale of the coin. The champions themselves, all working together, can sell coin and put the compute towards cancer, put the compute towards culture, put the compute towards whatever because they're trusted. Things like verifiable logic-based stuff is important when you're doing a fully decentralized network. Things like Bitcoin are important when you're doing a fully decentralized network, right?
But we were like, intelligence needs to be distributed and controlled and owned by the people. And so that's why we came to this model of these equity-based champions being the miners. But then the crypto angle, the Bitcoin-type currency angle, is important because it's a million times faster. It's like a million-TPS version of Bitcoin with the same private keys, where every single token burned is going towards public good. So it creates a lovely flywheel because we want more money to go into owning the AI that runs your life. We want more money to go into organizing the world's cancer knowledge and stuff. And so, again, we're looking at that specific thing, whereas I think the other crypto protocols are looking at different things. Again, you don't need to compete against Venice; you can use the API, right? The proof of useful work miners can offer compute to the network via the champion. The champions act like mining pools, coordinate everything, and they can have their Foundationcoin USDC. Again, we're just looking at a different space because we think the defaults for everything that's licensed and registered—medicine, law, politics—we really got to get this right. Otherwise, you're all watched over by Claude Opus 8. And Claude Opus 5 had such a nasty personality that I don't want that to happen.
