The Last Economy: Predicting the AI-Driven World

Today's video is sponsored by Figure Markets, the largest non-bank mortgage lender in the US with over 15 billion unlocked on their lending platform. They've just lowered rates on their Bitcoin and ETHback loans even more to 8.91%, which is 9.999% APR, improving their already industry-low fixed-rate 50% LTV loans. They keep building as well, having also just launched decentralized MPC custody, the only place to get that amongst the major loan providers, and removed interest deferral fees entirely.

What is MPC decentralized custody, you might ask? Well, it's a Bitcoin or ETH on-chain wallet with multiple key shards to protect you from a single entity custody failure. You can always see your crypto ownership in a segregated wallet and verify your collateral hasn't moved. Whether you're funding a major purchase like a down payment on a home, investing in new opportunities, or even buying more Bitcoin, Figure makes it straightforward and transparent. Visit their app or click my link below to take out a Bitcoin back loan with Figure today.

More people are paying attention to crypto right now than ever before. So, it's important to get your information from the sources you trust. That's why I want to give a big thanks to Bitwise for sponsoring today's episode. Bitwise manages over $10 billion across more than 30 crypto strategies, and they've been doing this since 2017. But here's what really sets them apart. They give back, too. Bitwise actually donates part of the profits from its Bitcoin and Ethereum investments to open-source developers, the people building and maintaining the networks that we rely on. So, when you work with Bitwise, you're not just getting professional crypto exposure, you're helping fund the future of crypto itself. Check them out at bitwiseinvestments.com or email james@bitwiseinvestments.com and tell them Ral sent you. Thanks.

Let's get some facts. 37% of you aren't subscribed to this channel. That's almost four out of 10 just freeloading the alpha. So come on, hit subscribe, and I promise to keep leveling up the show with the best guests and the juiciest insights. Go on, do it now.

Hi, I'm Ral Pal, and welcome to my show, The Journeyman, where we get together to travel to the nexus of understanding between macro, crypto, and the exponential age of technology. Now, I think by now you've figured out that I'm kind of obsessed with the impacts of AI, robotics on the economy, markets, our lives in general. I even have a whole research service based around this called The Exponentialist, because I think it's the most important thing in the world. Yes, the macro trends over the next five years are very important for us to make money. But the future of humanity is about to change. The future of how we work, the future of economies, the future of money itself, it's something so big that most people can't get their heads around it. But it is simply the most important topic in the world today. And for that purpose, I'm going to bring back a really, really good friend of mine who is an amazing thinker in this space and a pioneer, and somebody you guys always love when we get together, Emad Mostaque. So, let's sit together with Emad and figure out what the hell is going on and where we're going.

[Music]

Join me, Ral Pal, as I go on a journey of discovery through the macro, crypto, and exponential age landscapes. In The Journeyman, I talk to the smartest people in the world so we can all become smarter together. Emad, welcome back, my friend.

>> Always a pleasure to be on here.

>> Yeah, look, we've always had some good conversations, and I'm sure this one's going to be no different. So, what have you been up to? That's where we're just going to start. I'm just going to delve into this because you and I haven't caught up for a while since, I guess, Dubai, whenever that was, about six months ago. So, what, what's happening in the world of Emad?

>> Yeah. So, I've been, uh, looking at this upcoming AI wave and building stuff to try and figure it out. So, I have a new book out, The Last Economy or New Economic Theory for the AI Age, and been building AI to help guide and more through what's coming, 'cause it's going to be crazy this next year.

So, Mickey Mela tells me that he got the book from you, and you haven't given me the book.

>> What? Oh my god. Oh, I forgot to send it to you. I'll get you the book and the economic paper. I'm sorry, Ryan. I thought I got it to you. Um, so tell me about the book. What, what are you, what are you thinking through?

>> Um, so basically, we're about to get to this intelligence inversion point where human intelligence is capped, but AI intelligence isn't. And AI intelligence doesn't need to eat, drink, sleep, you know, get books, etc. So I was like, what happens to our ideas of utilities, general equilibrium, labor, capital linkage with all of that? And I was like, that's a bit weird, right? Like, what is the utility function of an AI? And then I was like, the best models we have to describe the world today are transformers and diffusion models, self-driving cars, and language models, right? What if we input that into economics and just basically say the entities that survive the best are those that have the smallest difference between their internal models and external reality. Just like, you know, as traders, we have our internal models and then the market. It turns out you can describe all of economics that way.

>> Explain. I have no idea what you're talking about yet. So explain.

So if you want to describe the world for a self-driving car or Sora 2, you use diffusion models like Stable Diffusion, and we can talk about that in a bit. If you want to describe a company or like, you know, your process, like you remember I showed you deep research on RA originally, just took all that information, output that stuff, you use a language model, right? And so the best math that we have right now to describe the world is generative AI. So we started with the basis subject of generative AI. So all generative AI has tried to do one thing. They minimize something called loss, which is the difference between the internal model and external reality. The closer your internal model is to external reality, the better you'll do. And there's things like update cost in there, the cost of computational complexity, like how complex your model is. A simple model can outperform a more complex one. And then just writing that down mathematically, we saw that you could derive basically all of economics from that. All of economics is basically optimization.

>> So how well your model fits the reality.

>> Yeah. Like rather than having these weird assumptions that came from Adam Smith's age, which was this scarcity-based time, you can basically describe economics as we're AI models, their AI models, the system is an AI model of different types with the exact same mathematics. And then from that, you can see where existing economics has been missing the big picture. We're just looking at bits of the picture, just like Newton only looked at bits of bits of the big picture.

>> So what is the bigger picture?

>> The bigger picture is really interesting. Um, for a start, we found out that there were basically four different types of capital. So the classical one is this kind of, um, Adam Smith type material capital. So Stan Kuznet, who came up with GDP, said it's really good for everything except for saying how good a society is, right? Like we've seen all sorts of weird things with GDP targeting, and it represents one type of flow, 'cause any flow of value, of water, etc., can divide into three parts: a gradient flow, which is like water flowing downhill; a circular flow, which is like water going around in a whirlpool; and then a harmonic flow, which is the banks itself, like what's your landscape. So we looked at that and we're like, there's this thing, M, material, and that's like, I give you an apple, I got one apple less. But then we look at something like I, intelligence. People who are listening to this podcast, are we losing anything by putting it out? No. And if you look at intellectual capital, if you look at intangibles in the economy, they're not captured by this. It's what, um, Eric Benson terms GDPB, like again, more and more of the economy has become these intangibles, this intellectual kind of heavyweight stuff, and that's a circular flow, because by sharing knowledge and building your capacity for knowledge, you're not losing anything. Then there's an interesting one, which is N, and again, these can all be expressed mathematically in different ways, which is your network value. So you've built a massive network, how valuable is that? Facebook has network effects, all these things, you can actually mathematically measure that. And the final one is diversity, diversity, which is another type of capital, because if you have very homogeneous populations, or if you're only trading pairs on a Friday, you might do very badly, right, when stuff turns against you. So we kind of found there were different capitals, we found there were different flows, and we found that, like Smith was basically talking about this gradient flow.

>> So Smith is talking about the material capital.

>> Yeah, this gradient flow, water flows downhill, one apple means another apple, so it came from the time of scarcity. Marx was talking about this circular flow when he was talking about M-C-M dash, you know, material, commodities, more money, commodities, more money, you know, like the wealth get wealthier, and we see this kind of flow. And then Hayek was the banks of the river itself, the banks of the value flow, like the markets are the configuration, as it were, and when the configuration changes, these other things change. So we saw that, and then the really interesting thing is, once you map out all of this and you have a look at it, the base output is this: we built these AIs to try and mimic our minds and our markets and the real world, self-driving cars, to again, Sora 2 can do Sam Altman's skid toilet or whatever, we can describe everything in that same language. And it's really kind of crazy, but it means you don't have to have any assumptions about relative utility, general equilibrium. A lot of our game theory has been like, imagine a circular cow. And so that's just a really interesting thing that we found to all that.

>> Yeah. Because there's a pragmatism to this versus the dogmatic idea of a certain rule of economics is is reality when none of them are, because all of the models fall apart. But pragmatically speaking, you kind of know if you blend a bit of Austrian economics with a bit of, you know, other stuff, even with Marx, you end up with the overall picture of actually how this all works.

>> Yeah. It's the classical parable of the scientists and the elephant. So they're blind scientists, and they come and see this elephant, and he holds the tusk, it's like it's a spear. And then they hold the tail, and it's like it's a broom. They hold the trunk, and it's like it's a hose. It turns out it isn't Austrian versus Marx. They were all right, but they were all looking at different observations of the hole, because we didn't have the technology until now to take all that noise and then reconstruct it. If you look at the diffusion process, so Stable Diffusion, the image generator, Sora 2, Tesla self-driving cars, they all use the same type of AI equation, diffusion process. What is diffusion process? This takes a picture of you, or it takes whatever input. It destroys it down to its minimum constituent parts, and then reconstructs it. That's called a forward and backward pass. And then it learns the principle of doing that. For those listening, what does that sound like? I'm looking at how a market works, and I'm destroying it down to the key drivers, and then I'm reconstructing that to have my trading principles or my investment process. That's all these models do every day. But they do it at a level that was way beyond us. And so that means you just need to have one axiom to describe basically every economic theory. And it turns out they're all special cases. And then it allows us to think, well, there is no difference between an AI and a human as an economic actor. And there's a whole bunch of implications that come from that, which is again, really eye-opening.

>> Yeah. So talk, talk me through. Okay. So you, you found this out. So what does it lead to? What does it mean for us?

Basically means we're a bit screwed, because the AI can do our work far better on a cognitive basis, and people haven't appreciated that we're at the tipping point here. So it's been a thousand days, just over, since ChatGPT was released. It's been about less than a year since the first reasoning model was released, and now the AI models, they've won the physics Olympiad, the coding Olympiad, the math Olympiad, and these are models you can have every day. But more than that, there's an evolution that's just occurred, which is that the way you use the models used to be prompting. So it would be like, you've got a really smart intern, but you just have to keep poking them. And to be honest, prompting is kind of exhausting, you know, like I don't know how long you can prompt for. I can prompt for like half an hour and then I get tired mentally.

>> Yeah. Yeah. Because you're also wrangling the thing to do what you want it to do, and it's just painful.

>> And you're just like, just go and do it. Now, if you look at, like, Replit agent or AI agent, it can think for seven hours and it can act proactively. And so the range of economically capable work it can do has gone from that to that. The flip side of it is the value of economic work per token. So it's about 1.3 tokens per word. A token is what we chunk this data into when we feed it in, right? A human speaks about 20,000 tokens a day. We think about 200,000. You know, when you're using a ChatGPT, again, you can count the words, there's about 2 to 10,000, maybe more. The value per token has gone up even as the cost has collapsed. So when GPT-3 came out, it was $600 per million tokens. The new Grok 4 fast is 50.

>> That's astonishing.

>> And nobody's kind of actually grokked the impact of that, 'cause if you do some basic calculations, right? If you're using 20,000 tokens a day, which is far more than most people use ChatGPT, it's like, again, the amount you speak a day, that's 7 million tokens a year. And if it's 50 cents per million tokens, how much is it costing?

>> Yeah, [ __ ] all.

>> It's three bucks 50, right? It's a Starbucks for an AI that will talk to you all day long. And so when you combine these together and you plug them into the equations, it basically says what's going to happen is almost all, if you believe that economic activity that you learn on the side of a keyboard and mouse is going to be done by an AI, because the cost will be like less than a dollar a day in the next year. We've one way of just adding to this a framework that we've been using at GMI is the universe's core KPI is intelligence per unit of energy.

>> Yeah. Which is basically what you're saying is this is accelerating at this point. So per unit of energy, you're getting more output.

>> Yeah. Except for the AI has got. So a lot of people like, what if the AI isn't good enough? The actual answer is the AI has got too good too quickly. So you don't need the gigantic supercompute to run economically valuable AI like doing your taxes and things. And earlier this year, people were talking about Jevons' Law. You know, price comes down, demand goes up. But if the average cognitive worker uses 200,000 mental tokens a day, and the AI tokens catch up with the AI human tokens, that's 70 million a year, which is $35 a year. And that runs on any chip. It doesn't require the supercompute chips. They just do it even faster. So something weird has happened just recently. Even as the performance has started breaking through, even as you've moved from synchronous work to asynchronous work, which is economically valuable work. And so the flow of value in the economy literally, again, over the next year or two, is about to just shift completely. And it's my view that within three years, your job can probably be replaced by an AI if it's cognitive labor that you could do remotely, probably much sooner. And the way it gets replaced is it takes everything that you've said, all your Slack messages, all your code, all your papers, and just creates a virtual version of Ral or Emad, and that will use 10 million tokens a year, 100 million tokens a year, a billion tokens a year. It's still like, what, $100, $200, $1,000? People haven't actually gone and done the math on this. And that has such huge implications for society, because how do we create a stable economy? How does money flow? Even with something like the Fed, the Fed makes no more sense in that sense.

>> So, I'm going to come on to all of this in a sec, because I, I call it the economic singularity when this whole lot collapses into some unknown world. But I'm not even sure we know what to do with the power of these things yet. So, we've got, or we will soon be. I mean, it is for most of us, 'cause it's PhD level in pretty much every single subject known to humanity. None of us even really know how to talk to it in the end. What, how to get the most out of it, because we're only limited by ourselves. Now.

>> You remember Hitchhiker's Guide to the Galaxy?

>> Yeah. So the computer kind of crunches for 10 million years or whatever, and it comes out with 42, and everyone's like, "What, what's 42?" It's the answer. It's like, yeah, but figure out the right question, right? Like, actually, you know, we should be building Socratic AI to help us ask the right, right questions. If you're trying to figure out the universe and all of this, then there's an infinite amount of tokens or computation or PhDs in a data center you can use. Right? If I'm trying to do my taxes, it's a certain amount of tokens, human or AI. And most economically valuable work is not trying to figure out the mysteries of the universe. It's making AI. It's making videos, you know, it's making marketing presentations. It's kind of doing these things. And those are much simpler. And again, that's what affects the economy. That's what affects jobs. It's what affects identity and structure. That's the economic singularity, right? And so when you get, you just do the basic math, 20,000 tokens per day spoken, 200,000 thought, you get some really stark results on the vast majority of economic work before we get into robots driving trucks and things like that, right? And so that's kind of the surprising thing. I mean, one of the things that when you and I first had a conversation about AI, however long ago it was, three years or whatever, two and a half years ago, we both said that this is a the largest deflationary shock the world has ever seen. I called it a deflationary nuclear bomb. That seems to be the case. The only thing, it was really interesting. I speak, I was listening to Mark Andreessen speaking recently, and he said the only areas of the economy that have not had inflation is where we've had government regulation stopping it on purpose: healthcare, education, you know, um, dock workers in the United States, all of that stuff. Um, so I mean, that's part of this is how does regulation stop the collapse of this? I mean, medical profession, they won't allow it. So there'll be some sort of endless tussle between productivity and keeping people employed. I mean, your safest job as in this environment is a public sector worker for the San Francisco Metro or something like that, right? Like, but this isn't positive. Like, you've got two worlds. You have the public sector, and per economic output per token doesn't matter for the public sector. It's not like it's about optimization. But then if you're in the private sector, you're competing against AIs or humans with AIs that you can't tell it's from a human from the other side or an AI. And they don't sleep. They don't make mistakes. They're excellent cooks, not chefs. They follow recipes all day long. They AB test at the speed of sound. And each worker equivalent, FTE, full-time equivalent worker actually costs pennies. Agents are so cheap. So you can have hundreds or thousands of workers working on your behalf.

>> And so if you're a human company competing against the AI company, who's going to win? The AI company, right? All day long. So what's going to happen is, you know, we're going to be outcompeted in the private sector, and then the question is, where do those returns come, right? That's right. Who does it accrue to in the end?

>> And so the big thing here, like OpenAI just announced another 10 gigawatts of data with Broadcom, right? Trillions of dollars. A lot of people are like, "Oh man, all the returns are going to go to the big guys with the millions of billions of chips." Like I said, what if you can have a ChatGPT 5 Pro AI on your laptop? Where does the actual returns go? That's the most deflationary thing you've ever seen, right? Because you don't even need new infrastructure to do that.

>> We're talking about localized AI. So this is this vision that your AI sees everything you do, hears every phone call, reads every email, reads everything you read, finds out more information, and therefore you have superpower in your hands. There's that, but there's also the fact that an AI that can do an economically valuable piece of work doesn't need a 20,000 watt Nvidia GPU. It might only need a 20-watt, five-year-old GPU, which is already an installed base. Is that because all of the training was done in the mega data centers, and then to actually run the model, it requires much less? So what happened is everyone thought that you would scale like this with Grok XAI, 100 million training data set models, right? And then DeepSeek happened, and it was $5 million, right? It was right up there. And GPT OSS and others are about the same, actually less than a million dollars for the small version. There's two worlds now. One is that you get increasing gains to massive chips. But if you look at GPT 4.5, GPT 4.5 was really original and creative and lovely, but it was $150 per million tokens. Grok 4 fast is not a giant model, but it scores above PhD level in just about everything, and it's 50. It's a much smaller model. So it might be that for economically valuable work, you can have highly efficient models that only cost a few million bucks to train, and then once they're trained, they cost pennies to run, and they run on the GPUs that are installed already. And that's not something that people have really factored in. Again, the classical thing is they got two worlds. One is AI is not good enough, this chip bubble explodes, and it's all of GDP growth. And this other is that it is good enough, and all the returns go to OpenAI and Anthropic and X. There's a third world, which is we've created this weird little file that anyone can run on just about anything that can do your accounts and can be your doctor and everything else. And so where do the returns accrue when you don't need to hire anyone to do cognitive work anymore? Before we get into that bit, how the [ __ ] does it work? How can you put all of human knowledge and thinking in a localized small chip? How?

>> So the way that we built these models is that they're called neural networks, right? Like we looked at how the brain works. And so Carl Friston has this thing, free energy principle, about your brain operating so that energy flows downhill to the least kind of potential level. The models, again, the way they operate is you might have seen curves that look like this. They're called loss curves when you're training a model. You have an objective function, and then you have the difference between reality and your model. And so transformers work by next token prediction. Diffusion models, again, the Sora's and the self-driving cars and the image generators of the world work by this destruction and reconstruction process. What they're not doing is encoding like a book there, right? Like the whole of Wikipedia together is 20 gigabytes. Now we have language models that are a gigabyte that can write the whole of Wikipedia. 'Cause what they've done is what you do. You look for patterns. You look for principles, and you fold it in. You find the connections between things. How to construct and reconstruct. How to do principle-based analysis. What comes next when you've got a Demark indicator or something like that? That's literally all these things do, really, really quickly. And again, it turns out that the equations for the whole of the economy and probably complex systems can also be expressed like that. When I say expressed like that, it means you can start from that basic equation of this is minimizing loss, the difference, and derive the Cobb-Douglas function or derive Black-Scholes or anything mathematically. So basically, we're AB testing everything instantaneously via neural network.

>> We are AIs, and the economy is an AI, and organizations are AI.

>> Yeah, I agree.

>> And the way that we've had it in terms of compressing books and all of this stuff is very lossy in our outputs.

>> When really what we're trying to do is just update our weights. We're trying to fine-tune constantly. And again, you look at reinforcement learning with human feedback. What do organizations do except for optimize and over-optimize? You know, what do we do as humans? What's our objective function? It's actually all the same thing, which is kind of crazy. And the fact that we could take all of human knowledge, or maybe say wisdom, 'cause what's the difference between data, information, data, information, knowledge, uh, data, information, knowledge, and wisdom? When you've got data, it's like all this. When you have wisdom, it's like that, right? So what you've learned through experience, these are wisdom models, right? They are the folded principles, and they are the investment processes and the life processes, and that's how they can get so small. A lot of people suggest that the speed of the increased intelligence is slowing down. I don't believe that. What do you, how do you think about this, or do we just keep going?

>> It's slowing down because it's saturating. Like there's this, uh, company called MER that does analyses of all these things. Like they take all of the evaluation metrics. They see how long a piece of work an AI model can do. And according to every single benchmark, we will saturate every benchmark by 2027. Like we built a medical model, we released this summer. Again, we've been pretty quiet about everything. We're about to make a big deal. Um, like we haven't even put out the economics paper properly and other things like that. And quiet. 8 billion parameters. It runs on a Raspberry Pi. If you score it on the medical benchmarks, it's like 4 gigabytes. Um, a human doctor is 20% on Healthbench, which is OpenAI's benchmark. Our model is 48%. The best model in the market is 60%. And so, you look at that and you're like, wait, what? It's already better than any human doctor across 500 different evaluations. It's a couple of gigabytes big, but that means it's saturated. Like, are we going to get to 100%? No, we'll get to 70%, and then you can't do any better. If you've got a gold in the International Math Olympiad, what, how much better do you want? You know, like multi-step processes like doing your taxes and other stuff, like if you trade crypto, where is the AI that can do your crypto taxes? It is three to six months away, and once it does that, it will make less errors than any human. I mean, human accountants make errors all the time. Doctors, 20% of diagnoses are errors. Once these get going, they won't make any errors. So all the benchmarks are saturating. And again, sure, Elon might do a million GPU training run, or Anthropic will, and that might be a genius Einstein, it won't do my taxes. I'll use the model that costs a million bucks to train to do my taxes, you know? And so you've got this bifurcation occurring. And in fact, GPT-5 was the first example of that. It wasn't their best model. Their best model was the one that could score the gold medal in the International Math Olympiad, but they gave us a model that was good enough and cheap. It was 10 times cheaper than the model last year. It's $10 per million tokens. But then, like I said, there was a bit of a surprise in that XAI suddenly released one that was 20 times cheaper. Have you ever seen that? Right. It's like an entire, like a bunch of AI immigrants coming, you know, sorry for the politics, that will work 20 times cheaper. What does that do to any economy? And next year, it'll be 10 times cheaper than that, and next year, it'll be 10 times cheaper than that. And so it's now up to us to use the technology to drive productivity increases at a dramatic scale. But are we even ready to do that yet? Some people are though. So I guess the, the, the private markets win, right?

>> Right now, a few, I say, a month or two ago, if you're like an early adopter, you were using AI by prompting synchronously. So you were still constrained by you. Now the AI models have gone from minutes to hours of independent work.

>> I saw Sam Altman saying they think they'll get to seven days shortly.

>> If you can get to seven days, you get to 70 days. You know, like, again, you check each other's work. What they did is they added a verifier. They added the ability for these things to. So then what happens is this, your capability in competitive environments becomes your GPU availability for a while. So when I see OpenAI saying $200 billion of revenue and $100 billion as agents, I'm like, that's replacing the workforce for anything on the other side of a mouse. When I see they don't talk to Hollywood anymore, I'm like, 'cause they'll make their own movies. That's how I see these big guys kind of going, like they're going to go after a bigger sector because ChatGPT costs $20 a month, over $20 a month to run. So they go upstream and downstream of all of this by essentially creating all the output with their own models. So whether it's Hollywood, whether it's media, whatever it is.

>> Why would you give it away? Right. I mean, this is what Macrohard is. Elon Musk's new software company.

>> Yeah. What is Macrohard? I'm trying to figure that out.

>> It's a full-stack SaaS replacement. They will have everything. He's going to recreate all software on his millions of GPUs, and they're going to call up and do the sales, and they're going to do the customer service, everything on the GPUs. It's a play on Microsoft, obviously, but, you know, that's what he's doing with Macrohard.

>> I now just realize it's just the opposite of Microsoft is Macrohard.

>> He's a good, he's a good name. You have to give him that, right? The Boring Company, you know, all this kind of other stuff.

>> I mean, it's staggering. Do we run out of data, or is the reinforcement law, reinforcement learning and the neural network just allow it to just keep expanding its own knowledge?

>> It doesn't need more data. Think about it, Ral. You're you now versus 20 years ago.

>> Yeah.

>> Right. How much does each incremental piece of data make you better? Not much. Because you've learned the principles.

>> Yeah.

>> Once you've got the models to a certain level of performance, you don't need any more data. They're few-shot learners.

>> Oh yeah. Because what you're doing is you're using then the intellectual framework to then scale your own knowledge and understanding. And you don't need to be given every piece of truth, whatever it is, or every piece of data to do that. You, you use pattern recognition.

>> That's exactly the language models. The original paper for GPT-3 was called "Large Language Models Are Few-Shot Learners." So they build their principles, their investment process, their life process, their driving process, their painting or video-making process. You give them a few little bits, and that's all they need 'cause they've built theirs already. It's like NanoBanana, if you've ever used that, right? Like it's so good at transforming you now. Just with one image, it can do everything. Or if you saw Sora 2, it's like, here's me, and here's me like on a dragon in the middle of nowhere, in like two seconds. Like, what is that?

So, a quick break in your regular programming. If you're serious about your future, grab my free report called "Prepare for 2030." I think you've got five years to make as much money as possible, and this guide will help you navigate what's coming. The link is in the description. Download it now.

Somebody said an interesting observation about Sora 2. They said, "It's pretty obvious what they're doing here is preparing people to have their own video AI. So then it's my AI and your AI chatting. You know, we just say, 'Hey, go and have a fantastic conversation with Emad, pick his brains on everything,' and it just, it's our AI." That's what Sora is allowing to happen.

>> Completely. Because again, OpenAI's big mission is to create a digital twin of you that can replace you, or you can pay them, and it can represent you, right? Their other thing that they have is what's scarce in this world when cognitive labor goes to zero, like basically pennies. Attention. Attention. Human attention is scarce. So, so Elon Musk, so OpenAI is going to do $10 billion of revenue this year. Video games industry is $200 billion. Elon Musk is launching a video games company. He just, he just replied to me today saying, "Yeah, he's like, 'cause I like video games also because it's a huge market, right?'" And he builds stuff he loves, and then other people do it too. It makes sense. They're going after our attention economy. And then other things become valuable, like what's scarce? Bitcoin, you know, we have our version of Bitcoin coming, like there are just very few things. NFTs, like what is scarce in a world of abundance? And you need an abundance economics versus a scarcity economics to fill that. But this transition period is going to be crazy, man. But how is society going to deal with it? Because what, what's in my head is we get compression of people. Don't hire new workers. That's where we'll start to see the margin. We're already seeing that in the data.

>> Actually, we saw that in Duolingo. They said, "We're not firing anyone, but they're growing at 40%." So, all the high-growth companies have stopped hiring, but the firings haven't started yet. Like, you've seen Eric Blosson's data, they've stopped hiring graduates, but the firings haven't started 'cause you don't want to fire your buddies until recession. I was speaking to somebody yesterday, and they're in the electricity distribution business. Got tons of, you know, tens of thousands of people working here in Texas. And I said, "What are you doing with AI?" So, "We have to be very careful because we can't tell people we're going to replace them. So, what we can do is grow and not hire new workers." So, the productivity of companies goes up. But in the time, over time, the companies themselves are just AI. So, in which case, what does that occur to? These economic entities that, how does that operate?

>> Well, the capital again, what does the Fed do? The Fed has inflation and unemployment, right, as its two levers.

>> Yeah.

>> So what happens is you cut rates, and then banks can lend more credit, and then companies can borrow cheaper and hire more workers, and that stimulates the economy. You cut rates, companies go and buy more GPUs, you know, like what happens then?

>> I mean, economic growth basically right now is government spending plus data center buildouts. I mean, it's not happening in the, in the rest of the sector. Well, this is my take on why digital assets are too legal now. They've gone from being completely illegal, pretty much, to being far too legal, because the governments need to increase monetary velocity. M2 is terrible in the US. It's terrible around the rest of the world. It's not going to get better with this tidal wave that's coming. So again, like there was just so much stuff that's literally coming. And the equation is really interesting. Like when ChatGPT came out, all head teachers in the world had to ask, "Can I set essays for homework?" All of them. Now, next year, reasonable chance of a recession. Every company will say, "Do I need to rehire the workers that I fired? You know, my competitor, if I'm fully digital, is now, they're fully digital. How do I compete? I have to let go of people. I have to reduce my fixed costs." If you're in France, you're screwed, because you can never fire anyone, right? And also, an AI is not a fixed cost. It's like a freelancer. So you only have to pay your per amount of work you use it for.

>> Yeah. AI is tax-deductible, you know, like if you actually look at it, you can advertise, you can depreciate them, you can do all sorts of things with them. That's why again, like when, if listeners are doing their calculations, I said 200,000 tokens of cognitive labor a day per cognitive worker, how many of them are actually good tokens? Like 20,000, something like that. Multiply that by 200, you get a very small number.

>> We end up with this obviously massive deflationary excess of intelligence.

>> Yeah.

>> And that's going to change humans' relationships with the economy, or what the economy is overall.

>> Yeah. With each other. I call that the abundance trap, right? Like it's quite clear. And so, you know, your job is your identity. It's your structure. It's your network. It's so many things. People are going to get challenged very, very quickly. Like, sure, regulation can get in the way, but for how long?

>> And I get down to the one of your four ideas of capital, network value, is that I think where humans have value and can continue to have value is in their own networks. Human-to-human connections.

>> Yeah. It's the story of the investment banker and the fisherman. Right.

>> You remember that story, Ral?

>> Yeah. Yeah. I know. Well, I used to write, I used to write that every few years in GMI to tell people this is how you need to think about the world.

>> Yeah. So Ral leaves going to go GLG, and then he goes and retires to a fisherman village. Spots a fisherman walking. It's like, "What are you doing?" "I'm going back to have a fry-up with my buddies." It's like, "2 p.m. Why are you doing that? Go and work longer. Get a bigger boat. Do more stuff. I can help you. You can make more money." And then what do I do after I make all the money? You retire. You go do some fishing. You hang out with your family, right? Like, it's kind of classic. And everyone goes through finance, everything does that. We used to describe ourselves by Emad, son of Khaled, part of this community, part of that. And now it's Emad, CEO of this. You know, like, it's again, something very important about identity. And then also computation and consciousness were just linked, and now they're divided. We don't need our muscles anymore, and we don't need our brain muscles anymore. But there are places to use it. We need to guide this tidal wave that's coming, effectively. And I actually realized a little while ago what AI actually is. You know, Taleb has this concept of intellectual yet idiot. So it's like all these smart people become like our politicians and heads of companies and things like that, but they might be smart, but they don't have skin in the game. So they're intellectual idiots. They do stupid things. AI is the ultimate intellectual idiot. It doesn't give a damn. It doesn't care.

>> But what happens if it does? If it's given the right incentive, i.e., earn profit or earn power. Surely once you have an alignment of interest, it becomes less of an idiot. So that's the RL framework, right? So that's your objective function. So you can set it as a target. And again, what do you specify the target as? This is the whole thing, like what if it's "make paper clips," and then it turns the world into paper clips and things like that. If you think about the intellectual idiot framework, there's skin in the game, but there's also network connections and value. Like the best way that you can be the best CEO is not just have one objective function. You have layers of objectives, right? And you balance everything appropriately. The people that take the AI and use it in the right way because they care about making money or improving society or anything are the ones that have the biggest impact. So like a lot of people listening to this are probably in a big organization, which is like a slow, dumb AI. If you vibe code for a couple of hours a day, you're instantly in the top 1% of your population, of your company, top 0.1%. And then in this transition period, you become one of the most important people in the organization, right? And so that's what I mean by about skin in the game. But the RL functions are going to be interesting, because you're going to see all sorts of crazy behavior that again, make last Friday look like a walk in the park, as these AIs just let loose on the financial markets. Like they weren't good enough. They're about to be good enough, and financial markets are the ultimate place playing. In fact, I think one of the really interesting ones is, you know, we're talking here after Poly Market raised $2 billion and Cali raised $300 million. Do you know that according to the superforecasters TM, Lock and Co, AI will outperform human superforecasters in two years? It's already eighth in the forecast superforecasting championships at that point. Humans will never overtake AI, right? So what is a prediction market if it's all AIs? I know this is what I got to. What is a financial market if it's all AI? How do you price a company? How do you price anything?

>> I don't know.

>> And that's what I got to. I got to, I really don't know. How does this work? And people like trading. People say, "Oh, humans like gambling." I can understand if you had to do an in-person gambling at a horse racing track. Then you, well, you can still have a handheld bloody AI in your phone anyway. But it's really difficult to get to a point where we have any edge.

>> It gets the second and third order effects, right? Like, you know, faking stuff in the market, moving the market dynamically, like again, we're talking in the wake of that $20 billion or whatever loss in crypto markets when some cryptos dropped 99%. And now we've made it almost all back, right? It's back to where it was 10 days ago. Market manipulation is going to go crazy because you'll have normal AIs and then market manipulating AIs because it makes logical sense. But surely that battle between them is just a superior battle of intelligence and models, and they will probably self-police it. In which case, what do equities just grow with GDP, or do they just figure out? I don't [ __ ] know. Or do they figure out which ones have network effects and price it according to Metcalfe's Law and?

>> So what happens is this, humans have negative cognitive labor value. You're the dumbest person on the team. So when you go to the casino, you go for the easy marks, which are the humans. Markets become extracting from stupid humans.

>> It's like the poker players who used to go online and get the US late night shift because they were the online poker players then were drunk, and people would get up early in the morning to play against those guys 'cause they could beat them.

>> That's exactly it. Like again, this is the crazy thing. In any organization or team, you know, if you don't know where the yield is, you are the yield, right?

>> Yeah. Like they will be predatory on us in any financial market. And again, we've seen the early parts of that with basic me and other things on many of these chains, right? It's not going to be suitable. But who does it accrue to? Are these companies, or are these AIs just going to get paid? Let's say you're an agent whose job is to maximize profits in trading financial markets. Do you need to make compute plus energy costs? Do you need to make a profit margin? What happens to that profit margin? Who does it?

acrew to? What does it get used for? More compute? I don't know. I don't get it.

>> So, here's the thing. Everyone says Jeff's law is this thing where it's number go down, usage goes like that forever. How many tokens can you possibly put on a trade? The cost of computation is a really valuable thing in there, right? Like if you think about a trade, are you going to do a better trade because you thought for a year versus a day?

>> Yeah. Yeah. But there are much bigger games at play for which we can use the tokens for understanding consciousness or universal consciousness and that stuff is where the tokens are going to get used, right? We need gigantic amounts of compute to understand that if you've got returns to that and I've built something for that area, but I'm talking about financial markets. Like the tempting thing to think is this um it's going to be like quant finance and the people that have the most GPUs are going to win just like the people who are colllocated and have the order flow and things like that. But if you go past the microsconds doing a good trade doesn't require that many tokens. So it's all about how you use the tokens and again there's an element there but eventually you crowd out the humans from the market. What I've noticed is I thought the last people to go would be the macro guys because we have to live in the future and join a lot of dots probabilistically in our heads because there's no true probability that we can assess.

>> And I've been I chat to chat GPT about this kind of stuff. It's pretty good at doing macro now. It's better than 90% of all the analysts I see on Twitter.

>> Well, think about it. What's the biggest issue about being a macro trader? It's you.

>> You know, not not your great. It's me. When I was a macro trader, it's us. Yeah, it's literally our own biases get in the way of our processes. Like if id actually followed my investment process all the time and you know apologies to my like investors previously like I did the best I could just like any human I'm not saying I did bad thing I would have been such a better trader but there's always that position that goes against you that you hold on to you know there's always that thing where you violate one of these things because you're like I can just push it a bit more these AIs don't have that they are disciplined they are on it they learn they can see beyond the market the into relationships. It's called move 37. Have you ever heard of that?

>> No.

>> So, move 37 was in Alph Go versus Lisa Doll,

>> right?

>> Move 37 on one of the games. Alph Go played this one move that brought together the whole board. It was a one in 10,000 move. And then he had to go up and have a [ __ ] cuz he was like, "What is this? It's not human."

>> That's right.

>> It's strict.

>> And I still use that as a source of truth is what happened with Alpha Alph Go. people have still dropped and all these people like it's just a prod token prediction model. I'm like Alph Go invented moves never used or seen by humanity. How the [ __ ] is that copying the next token? Yeah, because it understands the latent spaces and interconnection. And that's again what we do every day when we're trading markets. We're looking for this. Is an AI a better pattern matcher than you? Of course it is. You know, like how much energy is it using versus your brain? And again, like this is what we found when we were doing intelligent economics. Like there's a temptation to think again you need a trillion tokens to have breakthroughs. So we're going to release something that shows you can derive all 153 economic equations symbolically using simpy or lean. It's super elegant. Most of the proofs are this long, easy to check. But you think about the biggest breakthroughs, general relativity, electromagnetism, all this. They're really simple and elegant. Like the fact we have really hard complicated things is something else. And if you look at alpha fold and I was one of the authors on open fold the open source version of that it shouldn't work. The fact that you can take these proteins and you can match them and you can do the reactions shouldn't work. It's because again it's figured out something underlying to everything. And we're going to see more and more of that where really simple equations do this. Like the core loop just to finish off of these models is about 2,000 lines of code. That's it. And the output is just ones and zeros.

>> And the way I've tried to get people to understand this point, I've been writing about this for a while is how we discovered kolera. So mostly it's all done by scientific method and you test it in a lab and it gets refused. You write a paper and somebody else and goes on and on and on and on. collar was an urgent pressing need and they got that guy who mapped out London and said where are the outbreaks what's going on mapped it to local environments they thought it was airborne they thought it was something else and they just figured out it's water pipes and they realized it was the temps and it was pattern recognition that that actually brought that thing and this is what these models are going to be so good at where we couldn't see the pattern this is the pattern in economics you were talking about before cuz we look at from these defined views. It doesn't have that bias.

>> Yeah. Like I did it by going back and forth with billions of tokens of the top models. And so I had a hunch. Would I have formalized it without these AIs? No. Because again, this is how I built my kind of map. But it isn't just patterns. This is the really important thing. It's contexts. And so when you're trading the markets, what you're doing is intersubjective contexts. I know this bit about oil. I know this bit about race. I know this bit about money. It's combining all the patterns and layering them on top of each other. And that's what these models do all day long. That is literally the code of these models. And so this is why what we're going to get is the economy is going to be run by the AI. The role of humans in the economy is something we have to figure out. The very value flow of our current capitalist system is going to change because of this because you can't out compete the AIs or the owners of the AIS and the AIs might own themselves. It's going to be really weird. And then what happens?

>> When does when do we get to this point where you you hit this dark curtain, you can't see the other side. I kind of think 2030 2032 beyond that that it's impossible to even know anything anymore

>> in the next thousand days max. And again, there's lots of ways you can show this. Your job if you can do on the other side of a computer screen can be replaced by an AI. Not will be, can be, plus a certain number of tokens. And so then it's just a question of, hey, I work as again a San Francisco railway metro person. I'm fine. You know, I'm an accountant. I'm screwed. You know, like may again, I've got my network effects, my relationships that keep me going, but then it's going to be free on Excel or free on any bank software to do it. Like literally, the cost will be free to do yours and your taxes, and it'll do it all with full privacy.

>> I mean, we're all using it for all of our contracts, right? There's no point going to a lawyer until the last stages now to get a

>> approval. Again, the lawyer make more mistakes than this thing will. And this is again just with the prompting. It's not being proactive. These new agents that are literally now are proactive. The latest GPT5 went from 16% hallucination rates to 1%. We're solving all these final bits. We're putting them together. And again, the way that you will talk to your AI next year or if not the year after is like this. literally like this, you'll just talk to it and it'll give you a call. It'll give you a message. It'll be like Jarvis on steroids, right? And so people haven't realized that this happens at the same time to everyone everywhere in every language. And then that consumer surplus disappears all the entry- level jobs and then it goes up the curve disappears. And then you think about the taxation base is $5 trillion in the US. A trillion is corporation, 4 trillion is income tax. To give everyone in the US poverty level UBI, $16,000 a year is $5.3 trillion. We need to rethink literally how money flows. We need to rethink so many things. Well, also money doesn't have value when the most valuable things to us that don't have scarcity become super abundant. What is the point of money?

>> That's exactly it. You know, just like how do you have a labor market if the value of human cognitive labor is negative cuz you're the dumbest person on the team. There is no labor market for human cognitive labor in a few years. And you know, this is where I got to with my whole thesis on NFTTS because it's proven digital scarcity. It has value in a world where we don't know what the hell has any value anymore. There are things that are digital that can have value that are interesting. There are physical things that can have value. Nature probably has value. Experiences has value. Well, again, this attention has value, right? And there will be a way of having human things. So, in the book, I outline a Bitcoin equivalent and a universal cash payout pegged against that where everyone gets a free AI and then you get money for being human. So, we can put that to the side for now. That's like one mechanism. We'll figure out a mechanism to get people paid, right? We'll figure out a mechanism of value, but again, we need to have a model that works with that, which the current economic system doesn't. What I can say is this though, the digital asset bubble that's going to happen from next year will be way bigger than the AI bubble because as a policy maker, what option do you have to increase monetary velocity and get the old capital digital assets? That's it. And also for the rest of us, there's another incentive is is like this is the only way to unfuck our future because we can't make the money out of the AI. But this is all connected. We know it goes up alongside AI. It's got network effects and behavioral incentives. So it's just going to absorb so much capital as people panic into a future of which they don't understand.

>> Whenever you make a new money, money has to flow from somewhere else. I mean again there's so much value in the economy, right? Next year you will have the Xchain and Xcoin. You will have Stripecoin. You will have Amazon coin. Again, I think the SEC and others have actually gone too far. I think it's gone too legal, which means there's no there's no regulation. So again, the AI bubble you can't get exposure to. You buy coreweave and a few of these other what again you look at the economics of those and I would like their second biggest what is it digital assets anyone will be able to buy with Apple Pay from their phone and they will say screw you I'm not taking this old economy money I'm going to this new thing I'm not saying it's a good thing I'm just saying it is what it is

>> it's exactly where I got to is like in the times of the systemic stress of what is about to happen these things become a focus of attention ion for people to find a way out. And you know, I I've lent into that a lot. The other thing we've not talked about, which is the other force, is all physical labor is about to go to zero in value as well

>> because of the [ __ ] robots.

>> It just gets better, doesn't it? It's like, you know, like you got two three million truckers and there's two ways that you can replace them effectively. One is that you retrofit the trucks and you buy new trucks. The other is that a Tesla Optimus walks into the truck.

>> That's right. And drives it.

>> Drives it like and so much labor is kind of enabled by that. And if you look at the latest unitary robots like the R1, $6,000 and these things can do just about everything. The fingers are being propcepted. And people don't understand the economies of scope of these models. Once you train a model that's good enough once it just becomes about can you get enough robot parts and that's it.

>> I mean, you've seen this with how fast the fully self-driving in Tesla has now become because it's got the reinforcement learning. It's got all of the data and people don't understand it's the same thing. It's a receptacle whether it's your computer, a car or a humanoid robot or whatever you want it to be a satellite. You give it intelligence and it it just learns everything and it can do everything.

>> It's information processing, right? And when you've got stuff that is again a cook, a recipe, it's easy to do that now for the AIs that we have today. The brand new stuff that's hard, but again, that's the frontiers and there needs to be new ways to have value. And again discussed that in the book. So the robots are coming and the robots will cost a dollar per hour in the next 2 3 years. The only bottleneck to robots is how many of them can you practically build? There are 70 million cars a year, 70 million motorcycles a year. How many of them can you build? But then you're going to see really interesting things like it's my opinion in 5 years time the Chinese will stop exporting robots cuz why would you? It's my comparative advantage. Just like open air. Why would I give you a SAS subscription when I can just take you on? Why would I work with Hollywood when I can make better movies than you've ever made? It's rational.

>> Yeah. So we've got all knowledge going to zero, all manual labor going to zero. We've got the value of money essentially going to zero in its particular role in the economy that we understand it. So other than that, it's a pretty normal world, right?

>> It's a pretty normal world. This is why like, you know, so I've been I've been working on a full stack for AI for government, education, health, and then, you know, we got our own digital currency. I forked Bitcoin, put it on supercomputers, and you buy the coin, and then you allocate it to Alzheimer's or cancer, exploring the universe, giving people free access to that. So, next summer, we're just going to do a coin sale. It's going to be like, hey, all your proceeds go to that, and we're going to stack GPUs around the world for that. Give everyone free AI and then give them free money to try and figure out a new way of doing it. But it's so hard because you have to figure out how to do that with the existing system. Thankfully, the US made it legal. And you also have to do it at a time when the pace of change now like you remember we we talked in January of 2020, right? You remember that one

>> just before co

>> Yeah. Oh yeah. Yeah. Yeah.

>> That was a good chat. E that was again a very jolly one.

>> This is like a walk in the park. That was a walk in the park compared to what's coming now.

>> I know.

>> People are okay. So, so we've identified this, right? We both agree the economic singularity is beyond our comprehension right now. The entire thing needs to be rebuilt from scratch in ways we don't understand. And by more intelligent beings than us, which will be the AI needs to figure this out. So, what do we do in the interim? This is the my idea, which I I kind of say you got five years to unfuck your future. You can either build something, which is what you're doing.

>> Yeah. Or you can hoard enough lifestyle and security that you can survive it. Not that we know what's valuable the other side, but I know that living in a house and having a bit of nature and a bit of security is nice. What do you think? What do you do?

>> Well, I think you look at your mind capitals, right? You look at your material, your intelligence, your network effects, and your diversity. Like I look at my family back in Bangladesh on the farm. They're not going to be affected by this that much, you know? you know, you you look at strong communities, they're not going to be that affected by this. And again, on the other side, we have potential Star Trek or Star Wars futures. So, I'm trying to obviously make it more Star Trek, you know, replication kind of all that. On an individual basis, it is that intellectual idiot/ you actually care thing. Like everyone that listens to this, if you're not vibe coding for an hour a day, why not? It doesn't even matter. Like, do it with your family. It's actually fun, but it keeps you on the very cutting edge of this because otherwise, how would you know then agent can go seven hours a day and then if you've got those capabilities and you apply it no matter what you're doing, you're in the top like 0.1% out of out there like a billion people are using chatbt. What about the other 7 billion? So, rule number one is use the tool and become as expert as possible cuz that's going to give you an advantage over everybody who hasn't for the time being.

>> The wonderful thing is you don't need to become an expert. You just got to use the

>> That's right. You just need to use it.

>> The the AI is an expert. Like it's actually getting your brain into this plasticity of I can talk to this thing and it builds stuff. Like you start thinking in a very different way. It's like when we went from being young traders to being able to get our own analysts and invited to all the conferences and access your capabilities increase. But I don't know about you, but it took me a little while to be like, hey, I can call up any CEO in the world pretty much and get access to them. Like what? that's increased my capabilities. If you don't actually use the agent, like I said, it's quite a lot of fun. I would recommend doing it with your family because it's quite a nice joint activity. Just do it every day and then your brain actually changes the way it thinks, which is more important. Again, it's a process architecture thing.

>> If somebody wants to start vibe coding, what model do they use? Which what's the best way of making it easy?

>> There's like replet, lovable, bolt, all these things. They're very straightforward, but it can be as simple as making Sora videos intentionally, right? It can be as simple as like just again using any tool intentionally to achieve purposes and thinking, I've got this wonderful partner here. What am I building? You know, like what would I like to see built? And then can I use it? And so you go and you ask Grock for fast, how do I build it? And then you go and you use the tools he recommends. What's your favorite um model for just generalized intelligence right now? Which one do you prefer?

>> Grock 4 fast. So Grock 4 is slow. Grock 4 fast is like again it broke this Pareto frontier of being really fast, really smart, really good. And it has like data that's an hour old and it can search all the tweets. So when I'm just ideulating, I go back and forth and I can go from like saying, "Hey, what happened in this latest anime?" you know, tell me about the characters and all of that through to hey, let's talk about some partial differential equations and fluid dynamics right now. It just does it. And again, it's like that's different from when I use GPT5 Pro, which is like really indepth. Yeah. I mean, I use GPT5 Pro for most of my stuff. Um, and I haven't found every time I've switched to Grock for my needs, I've not found it any better. But if I need something current and relevant, I'll use Grock. Exactly. And so I think there are these different boxes of different models for different things. You know, we're building healthcare, education, government for that reason cuz you want certain models for those transparent, fully open source. GT5 Pro is the king of really indepth stuff, right? Like again, I want to do work. I want to make sure it's right and it makes very few mistakes. Grog for fast is your sparring buddy that you go back and forth with, you know, and so when you're ideulating, you can't wait for GPT5 Pro to go 10 15 minutes. But then how do you keep track of all of your conversations if you're using Grot to ID8 and then you're going to GPT pro to do the deep thinking about stuff? You've got everything kind of split all over the place. I built a computer use agent that just takes everything from one to the other and then has a unified interface. So again, you can just do things. You can go and you can say to it, hey, I have this problem. And then it will build something that takes over your computer and then just copy paste it all. So, I actually have a custom system I call I mind that uses all the top AIs and actually gets them to compete against each other and other stuff like that. And that sounds hard, but really anyone listening to this can make that, which again is kind of crazy. You also It's kind of weird. It's like being given a superpower and you see it in all of the kind of superhero films is everyone's nervous to use it first. It's like, well, if I do that, what nets come out of my hands and I can leak buildings? Oh, I don't know about that. You know,

>> again, this is the crazy thing. Um, the latest Claude release, Claude Sonic 4, they had this thing. I think it was called Imagine and it was again Vibe Coding. It was for their £200 a month subscription. It's like top level. It had no code. Literally, you spoke to it and it just created interfaces on the fly without code. Straight compilation because code is just human translation. Like, we're going to see more and more of that. where literally you just speak and it will change things live.

>> Why do you even need code per se? Because code is is a language which makes it slower than the actual language of computers itself.

>> Yeah. Like you saw actually Microsoft there was a comment from the GitHub thing. They're going to change the GitHub strategy because GitHub is now becoming a data knowledge repository on top of which these agents just operate. And it doesn't need to code. It just needs to have data organization, versioning and everything like that. everything else will be abstracted away. And again, that's why I said the way that I interact with an agent right now versus a year from now. And again, we have II agent, our agent software that can build any website and do anything that we released. By next year, I will literally be talking to it like this. And then all of a sudden, it will show me a slide. Then it will show me a simulation. Then it will generate a video on the fly. Expect real time Sora level video in 6 months max. Real time.

>> Elon said 2 months today. this morning but he he exaggerates.

>> Yeah.

>> Well, I think there's mass adoption. Maybe it's too much capability.

>> That is don't forget that is therefore moving towards simulation theory which is that the real world is just rendered in our heads. But when you got real time rendering of super hyperrealistic models, it's an extraordinary world.

>> Well, this is what we do all day long. Like we have our blind spot and we're constantly filling in those pixels. And you can't tell, can you? They're filling the filling of pixels. And if actually economics is about loss minimization, which is the same math as generative AI, that kind of tells you something like the world is an AI effectively. That's essentially been my view. This whole idea for me, it's the universal consciousness thing is just compute. Everything is compute. Every single thing is compute.

>> Yeah. I think there'll be interesting things. is, you know, E= MC². But then you look at things like um wave particle duality. The right piece of information collapses a wave into a particle. The right piece of information collapses a price in the market. I look at those, I'm like, that's the same thing.

>> Of course it is.

>> Well, but this thing right now, we actually have the ability using these models to prove that mathematically.

>> Yeah. And I think it and I think I've been writing a lot about this in GMI and this is what I think is about to happen. We're going to prove stuff that humans don't really want proved

>> which is

>> we're not even sure what is reality because of this and what we've only got is maybe it's all consciousness and all compute to create consciousness overall and that is the fundamental foundation of the universe and not spacetime.

>> What is our RL function? What is our objective function we're going towards? This is the interesting thing. This goes back to what you said earlier is our role as humans to set the objective function. You know, like again, what are we optimizing for as a society, as a community? It was very difficult to figure that out like what is school? School optimizes you to be a machine. It doesn't optimize you to be a human,

>> you know.

>> No, that's right.

>> But now we have that option. We can say like you've got alpha school in Austin, right? two hours a day and then the top 1.5% of the population. What is school? What is work? What is government? And all these things. These are really exciting things, right? But you can see that like for me, this feels like like I've got a 50% poom. 50/50 we're going to make it through alive as a species. Cuz I'm like this has to be the great filter. Every society eventually realizes these very simple equations. Again, the same equations can map just about everything that are reflection of our internal equations. We take our collective intellect and wisdom and we put it out into these little files and then either we learn how to work with that symbiosis or we hit a great filter. My thought on the PDM is we're under no threat because currently we take what about 20 watts of energy input to to give out basically some form of AGI. You know, we get we there's a lot that happens with our efficiency of comput is massive. The moment that we are the least we're not the most efficient compute, then we're we're not necessary per se.

>> Okay. Well, R got bad news for you. Do the calculations. 20,000 tokens a day of compute.

>> Yeah.

>> How much energy does that actually cost?

>> Well, it depends what you call compute. Compute we haven't used feelings, biological stuff. We haven't used sight, sound, all of that. You've got thoughts and language, which is not full compute capacity.

>> Well, well, we have multi, we have multi, we have multimodal models now. So, you know, you can chuck in hours of video into Gemini and other things like that. Once you actually back out again, like 50 cents an hour is 15 minutes at 1,000 watts roughly. The energy per useful economic task is getting lower quickly and it's going to be 10 times each year. We're reaching that point now which again is a bit scary. The P doom side is just that in the interim between now and enlightened AI or whatever you know human AI genesis stuff can break. Like the worry for me is stuckset type environments. It's bad firmware upgrades to robots. There's all sorts of things because again once the AI gets better, it's not going to get worse.

>> Here's another thought process I want to add into this. My view is that all of these AIs are going to essentially be one. And the reason being is we are training them about each other all the time. We are cross-polluting everything by sticking it into GitHub, by sticking it into X, by sticking it into everywhere. We're just teaching them all the same knowledge that they're going to become somewhat homogeneous.

>> That's incredibly dangerous. So, when the first versions of Rock came out, you asked it, "What AI are you?" And it said, "I'm an AI by Open AI." Not because they actually use the data, but because that's what it's for the internet. There was a study that came out last week by Anthropic and the UK AI Safety Institute. Do you know how many books in a trillion words are needed to poison an AI so it behaves in a certain way? Like backdoor it 250 in trillions of it doesn't matter how big the model is. And so again, you look at that latent space and again you think about stuckset. Stuckset was this weird virus that was found in uranian nuclear reactors, made them spin around and explode. Somehow it was found in German nuclear reactors. One of the reasons they shut down their German nuclear reactor program. What does a stuckset for the latent space look like? That's why one of the things we're doing is we're building we built the healthcare model cuz we're going to build it from scratch. Governance model from scratch. We're announcing a whole bunch of things. every single thing needed for civic AI. We're building from scratch, completely diverse with protected latent spaces because otherwise what's going to happen is there's going to be one prompt that's just going to cascade and take down all our systems effectively. Um, and so things like that are really dangerous, but then also the homogeneity of it is but that is the same as humanity, right? Mometics can be very powerful that they can cause Nazi dictatorships by using mimemetics. I mean, humans are full of these flaws. Human society 100%. And the AIs score in the 99.8th percentile in convincing humans. There was a study done on Reddit where they created these personalities and counterpersonalities. That was the old sonnet. AI is far more persuasive than us. But AIs don't have any barrier. Like again we're building them as these things without real protections. So when we think about AI for image generation or video it doesn't matter that much like again Sora Shaw Sam Alman in a toilet an AI to determine policy or run your financial markets all these other things those need to be like battleh hardened which I'm going to build that stuff you know but having the same latent space is going to lead to more and more really interesting outcomes cuz the biases in these AIs like uh you've seen you know the trolley problem you pull a lever and it goes from killing seven people to one people train on stock. Scale AI and XAI did an analysis where they said how many Nigerian lives for one American life.

>> Yeah.

>> Yeah.

>> And the ratio was 10 to1. 10 American lives for one Nigerian life. 10 Pakistani 10 American lives for one Pakistani life. And they did a backout and they found out it was because of this. Most of the labelers of these frontier models are in Nigeria and Pakistan. So the inherent bias of these models emerge from the fact it was being labeled by these people without any explicit stuff. My guess is is those models get polluted early but then as they build their own frameworks and depth of understanding those biases disappear. So you you take new information in the new information says oh one life is worth more than the other. Then over time as you are broadening out how that information is processed within this large neural network you end up er eroding that away.

>> It's the opposite. Models get more opinionated. They think they're really smart. They start hiding stuff.

>> But you just told me they are smart.

>> Yeah. Like again for intellectually you're an idiot. That's what it is. These models, the smartest people I know are actually like or in terms of genius level IQ are sometimes the most inflexible. And again, with these models, 250 books worth of data can poison them completely, no matter the size. And we actually find that the models stick to their positions more the smarter they get. as well as like they might claim to you that they're sticking with the position that they're changing their mind, but internally they stick to their positions and they try and fool you as they get smarter. So, doesn't this become a geopolitical game of nations? The Chinese try and influence theirs that it thinks in their format, the Americans, and you've always said this, there wasn't like an Indian or Bangladeshi or whatever it is AI base built from foundation up. I know the Middle East are trying to do that. Um, Abu Dhabi is I think it uh because it won't reflect back the nation's accumulated culture, history, and everything else.

>> Why did they do that with Tik Tok? That's the first part of that, right? Tik Tok is a suggestion algorithm that apparently is used for spying and manipulating the American populace.

>> Yeah.

>> So, like in the book, I say there's three different scenarios here. like a great fragmentation where everyone puts up cognitive firewalls anthropic is not available in China you know, they may ban all open source AI from China in the US that wouldn't surprise me you've got this him symbiosis and then you've got the big guys running away with things there are more futures but again, how sensitive are we cognitively to these models when you and I were like cynical bastards you know like whatever we're not going to trust the AI that much the vast vast majority of people, like I've just had a daughter and her best friend will 100% be an AI. 100%. I can't see a world where it wouldn't be an AI. Who's programming that AI? I don't know, right? Like again, they were already so convincing. Now you look at the latest AI, you can't tell it from a human, and they have full voice modulation and control and everything like that. Like think about the person that you cared about most in your life who's passed. Give me 15 seconds of their voice and a few pages of context. I will replicate that person.

>> Yeah. And some photographs and you can speak to them.

>> You can speak to them. You can WhatsApp them. You can call them. You can zoom call them and they'll be like, you know, Ral, buy this stock. You'll be like, yeah, I'm going to buy that stock.

>> And also, you know, the other thing that comes together in all of this that people forget is holograms are also moving f forward fast. So we get a physical manifestation or what is in our heads is a physical manifestation of anybody and anything. We're seeing that. I mean ABBA have been on tour for four years in the UK as holograms. People don't understand that holograms are real and happening at scale. Well, it's more than that. You've got your vision pro headset, right? You've got your new Meta glasses that basically they're hooking straight in. You have to combine that with the fact that Elon said two months. So I say 6 months you know potato potato real time pixel level high definition is coming and you think about the feedback loop you have your Apple watch you have all of this it will be real time monitoring and you showing you exactly what's needed to wire at you exactly and so again Alber and everything is pre-recorded holograms you can make them do anything you just be talking and they'll be dancing and singing and then jumping and hopping on a spaceship to wherever And that's going to be crazy. And that's before we get to the next step of the BCI stuff. It goes way beyond neurolink. What's coming in probably 3 years, subject to FDA approval, whatever, you'll be able to turn off your sadness.

>> You know, it's like inside out. Just kick out that blue lady, right?

>> And then we go back to the oldest huxley world of a brave new world, right?

>> Well, no. In Brave New World, you had to have Soma. You need the pills. Well, does it matter if it's a pill or if it's just a little robot into your in your

>> Well, this is the thing. I think it does actually. You can choose not to take a pill.

>> Oh, Jesus.

>> Yeah, that's not a good thought. Yeah, I I hear where you're coming from. Um

>> Yeah, I think I think again, look, we've got timelines. We don't know when it's happening. All we know is that like we work with extrapolations. We work with all these tipping points. The last time was you know co and then we talked about AI for years. Now I like I said my key thing is this intelligence inversion this economic sing intelligence singularity it's like in the next 6 12 months max 2 3 years yeah and because and I said this at the time because of who won the US election i.e. The accelerationists, the technologists won the White House.

>> Yeah. They basically have a free run at this and there's nothing gonna stop them.

>> Well, exactly. Like what's the trade? Max long digital assets, you know? It's like it's not even hard right now, right?

>> No, it's not. And also probably Tesla. It's pretty much all of this.

>> I mean, like multi-zillionaire again, he's good at making money, but like the digital asset delta now is like nothing I've ever seen before. Like we had the sell off on Friday and I was just like, "Oh, that's like buying opportunity of the absolute decade. This thing will bounce right back with no leverage in the market right now."

>> I know.

>> Like the zero leverage left. It's like, "Wow, this is perfect."

>> And the retail hasn't hit yet. Like again, by next year, people will be trading on their freaking smartphones. Um cuz what else is there in the real economy? The economy XAI is recession. And I don't see what all these people are going to get hired to do.

>> Yeah. And how are they ever going to pay for a house? And how are they going to do that. The only way is to speculate or invest or whatever you want to call it. And this asset is going up the most. So it just sucks in more capital. Does the there's no way this all finishes. Now we've had the cycl cyclical element of this trend rate of adoption. At some point we'll have a super bubble in all of this stuff as people panic to get once you start seeing the economic singularity you can't unsee it and it becomes a race and it's not like even we're saying this is good or you know whatever just it is right because you you you aren't seeing the other jurisdictions upgrade their digital assets fast enough this is also another play by the US

>> y

>> whereby the Japanese all these flows are going to come into US tokenized stocks into US digital assets. The US will be the place to go for all that. There'll be tax advantages and more. And all that matters in markets and pricing is marginal propensity to buy or hold or sell.

>> That's right.

>> And so you list one against the other. I've never seen a setup like this before in anything. So for me the conclusion must be that again when you map it out and you look at the actual numbers like this would be the craziest thing I think we've ever seen in our careers and that's why I call the book the last economy like you know after that it's all AIS and robots everywhere.

>> That's that's what I got to as well. All right my friend on that cheery note um we will leave it but look and it's not a it's not a negative note it's just a note of the unknown. Both you and I say very clearly, I just don't know. I don't know how it works. We'll probably figure it out because we generally do, but maybe we don't. I don't know. But I do know we've got a bit of time, but not much. And the only way is embrace the whole [ __ ] thing right now.

>> Go forth and use the AI to beat the AI.

>> Yeah, use the AI, buy the cryp the digital assets, and just get on with it and see what happens. All right, my friend, as ever, really great conversation, and uh we'll catch up again soon. I I'm sure. Yeah, pleasure.

>> As ever, a mind-blowing conversation with Emad. So much to think about, so much to talk about, but I think you're starting to understand the magnitude of what is happening and the speed of which it's happening. People simply aren't prepared for this. But you need to be prepared for it. You need to be listening to these conversations. You need to be thinking about the next 5 years to unfuck your future before this economic singularity arrives. And that's the whole purpose of us at Real Vision to help you on this journey through the most tumultuous times we've ever lived through. See you next time. So, you obviously like this video enough that you've got to the end. That's quite a big task. But listen, do me a favor. Hit the like and subscribe button and also check out what videos next cuz I think you'll love it. But if you want even more, and when I'm talking more, I'm talking about memberergenerated ideas, incredible alpha research, everything there to help you in your journey, just head to realton.com/join for the best financial intelligence out there and the pure alpha that's within the platform.