Today's episode is brought to you by Abra. Abra aims to provide individuals and institutions with a secure way to control, manage, and grow digital asset wealth from a separately managed account. Abra helps its clients get exposure to crypto and crypto financial products like yield and lending through one full service platform. If you're looking to gain access to additional liquidity, Abra has one of the most competitive loan products in the market. You can borrow against Bitcoin, ETH, and Solana at up to 50% loan to value. Rates are in the 4 to 6% APY and are open term. You can continuously draw down against your collateral as the price appreciates. Abra has other strategies to add yield and their team is happy to help align your portfolio to your risk profile. Reach out today and get a complimentary consult in your portfolio. It's worth seeing if they can help you manage your allocation, reach investment goals, manage risk, and add additional yield. Go to realvision.com/abra and tell them I sent you.
There's this wave of intelligence just about to hit us. It'll cost basically nothing. And what are we going to do? The next step is we're going to put these into robots. We're going to put AGI into robots. And people think of robots as like R2-D2, this kind of servant thing. When you put an AGI brain that's smarter than any human, every subject, and you put it in a physical form which is more durable, adaptable, stronger. What the hell is that? People are not ready for this. If you are trading and the example of this is super forecasting, Polymarket and all those that will be pure AI winners of course because if an AI is better at super fast it's going to be better at the market finance hedge funds traders are going to get eaten software companies are going to get eaten anything that can be done on the other side of a screen a GPU can do it better by the end of the year max next year so what does that do to your economy
Hi, I'm Raoul Pal and welcome to my show The Journeyman. The Journeyman, as you know, is where we travel to that nexus of understanding between macro, crypto, and the exponential age of technology. It's where everything is converging at an incredible rate. It's truly exponential. And I think we're all feeling it and seeing it now. The most exponential of all is AI. AI has been the fastest adoption of any technology the world's ever seen. It's not even anymore a regular kind of Moore's law adoption. It's Reed's law, which is Moore's law squared. We've never seen anything like the change of technology we're living through, and it's only going to get worse. Now, I've had one person help me guide me on this whole journey to understand what lies ahead, not just what's happening now. And that person is an old friend of mine, another hedge fund manager from the past who went across to the world of AI, and that's Emad Mostaque. Emad's a longtime favorite here for me on the show and on Real Vision. And we're going to sit down with him and figure out what the hell is going on and where it's all going. What are these agents all about and why it matters. Don't forget today's episode is brought to you by Figure. If you believe in Bitcoin long-term, the worst move you can make is selling it just to access liquidity. That's why you should check out Figure. If you're stacking sats but also want yield, democratized prime lets you earn up to 8.5% APY paid hourly backed by real world assets, not yield gains or total inflation. Figure also offers crypto-backed loans at 8.91% interest with 50% LTV. So you can unlock capital without creating a taxable event or giving up your Bitcoin exposure. Figure is the largest non-bank mortgage lender in the US with over 19 billion unlocked on the lending platform. And now they're letting Bitcoin holders borrow against their Bitcoin instead of selling it. Security matters here. Figure uses decentralized MPC custody, meaning your Bitcoin stays in a segregated wallet, not rehypothecated, not pooled, not sitting on an exchange balance sheet. They've also rolled out liquidation protection to help borrowers during sharp market draw downs. Hold your Bitcoin, unlock liquidity, or put capital to work. Check out Figure using my link below. On to Emad.
Join me, Raoul Pal, as I go on a journey of discovery through the macro, crypto, and exponential age landscapes. In the journey, man, 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, Raoul.
Yeah, it's always good. We always have a good conversation. I have no idea what we're going to talk about, but we're going to talk about a lot. That's that much I do know.
Yeah, there's a lot going on.
Yeah. What's been happening in Emad World? What are you looking at?
Well, everything. It's the great takeoff this year, isn't it?
Really is. I mean that whole Claude thing has been astonishing. I mean everybody I know is now running like 10 agents running their whole businesses is bizarre.
Yeah. And you got 10 agents when you're like why not 10,000 you know and then how fast do they run? I was just trying this thing for this new company called Talis who etch transformers onto the silicon itself and they're doing 15,000 tokens a second. So I just tried their chatbot Ask Jimmy and it's like you type anything instant response. You're thinking about that you're like there's this wave of intelligence just about to hit us. It'll cost basically nothing and what are we going to do positive and negative around that?
Yeah. And just the whole structure of an economy when you're adding these millions and billions of agents into an economy over this year. I mean obviously ChatGPT have hired, you know, OpenAI have just hired the guy who started building the agent stuff. So obviously agents going to go mainstream in every single way possible this year and we just increase intelligence dramatically at every level for zero cost.
Well, I think you know, you've always had this monetary velocity, monetary supply based approach to things. It's an intelligence velocity, intelligence supply. What Open Claw and Claudebot kind of did is they removed a barrier, an idiot tax to intelligence flow. So, you're on WhatsApp and you're telling it, hey, go and build a website for me or go and do this type of research or make it so that I never forget to send my wife flowers on her anniversary and here's a card. You know, and so as the different levels of friction decrease, the level of intelligence you can access goes up. Just like as you get assistants and teams and things like that, except for they cost pennies. And this is important because most of our economy and profits are based on what Elon Musk calls the idiot tax. You know, like when you try this out SpaceX, all the components of the supply chain, you get profits at every stage because you're simplifying things. When it comes to digital points being moved around, all of those frictions are going to disappear. And you know what frictions are? Frictions are profitability.
So where does the profit accrue in this? I mean because really it just becomes comes down to compute costs and energy costs and that's it.
Yeah. But when you look at compute cost and energy costs, it doesn't work either. Like if you look at what Sam Altman said, he said that recently GPT-5, their top model that you can use on Codex or whatever will drop in price by 100 times by the end of next year. It's $10 per million tokens right now. It will be 10 cents. And I think it'll drop even more than another 10 times. To put this in context, the average person speaks 10 million tokens a year. So it's a dollar for all of your words for an average person. You think a H100, a million, so it's 10 bucks for all of your thoughts and the efficiency of tokens is also collapsing. So recently, what's the name, Cursor, the coding assistant IDE, they said can we get a bunch of agents to build a browser from scratch and you know it took ages to build Chrome and all this kind of stuff. So 3 million lines of code. Took three billion tokens. And that sounds like a lot until you realize that's like $30,000. You know, that's not even a graduate programmer. But that 10,000 to 1 ratio is about to collapse. Just like when you have a chat with an AI, your output used to take hours of back and forth. Now it like one-shots websites. So the total thinking tokens versus that that's collapsing and the cost is collapsing. And you have things like etch silicon and others collapsing it even more. So does it really accrue to the GPU holders and others if it basically goes to the cost of electricity? And the other thing that people don't really get yet and you know I wrote that whole piece about the universal code and the idea that you know the whole kind of universe is essentially solving for units of intelligence per unit of energy and what we're finding is via solar the cost of electricity generation is collapsing in an exponential downtrend while the intelligence is going exponentially on the upside. That double exponential is something nobody's prepared for yet because it ends up being not just Moore's law. It ends up being Reed's law, you know, Moore's law squared.
Yeah. I think the best way to think about this is solar electricity is lower than a dollar per watt now, right? To run a smartphone level AI is $20 worth of solar PV, right? And that right now has an IQ on a smartphone of about 110 roughly. By next year it will be Opus 4.6 level which is the most capable model. And the scary thing is this. It's not AGI from an economic perspective. It's ACI, actually competent intelligence. Like now when you try and use it, it just gets the job done. And most of the economy is getting the job done. And I said we've never seen anything like this before. In fact, it's not surprising because the modeling aspect of it like I wrote my book The Last Economy last August to say you know some of the economic stuff that's coming and I have sort of law in there which is L equals HCK. So the Lagrangian is a mixture of basically the cost of updating your internal model and the difference of that from reality which is very similar to kind of your model that you're bringing forward because that is the model of AI. The most efficient entities, the most efficient AIs are the ones that minimize the difference between the two. And that's a function of their update cost, the complexity of their model, and then the complexity of changing their mind. So, we've never seen anything like this because we were computationally bound both in our brains. You can only think so hard and you start making mistakes after you think too long. And then I just had metacost or our communication capability. Whereas now with the agents and with hallucination rates dropping, communication protocols, we just released one called Common Ground that builds knowledge bases. All of Metcalfe's law disappears as well. They don't make the same mistake twice, whereas we do. We have to sleep, we have to eat, we can't scale, and we have lossy communication back and forth. These AIs don't. So if you reach a certain level of capability, then it's all about optimization. And that's where the profits disappear. And within talking of optimization, clearly we're going to have to move away from them using English language to communicate. They'll just communicate in binary. I think Elon was talking about this is like, you know, the moment you get rid of that friction of language and structure and just go straight to machine language, we get massive optimization again.
Oh yeah. Yeah, I mean that's why Yann LeCun has said that he doesn't think transformers are the way because they're very language based next token prediction. Instead you need these joint embedding models, JEPA type models and it's kind of what we see within for example diffusion technology which is used in self-driving cars. We used it for Stable Diffusion, the image generator. Sora and the story use it where you take a concept. It can be an image or a video or a picture of a road or even text. You destroy it down to its individual smallest component and then you reconstruct it and you figure that process out. If you look at code diffusion models or language diffusion models, they take a millisecond. Like you know, you see the ChatGPT, it's just like poof like that. But as you said, why do they need to encode language when what they really want is a world model? So when you look at ByteDance's Sora too, you've seen it. It's like that's like Hollywood level movies like doing kung fu and stuff like that and it's so fast and so cheap. It has physics inside. It understands physics because it's approximating that model more and more. Now you said as you move past that the AIs don't have lossy communication. They can communicate as efficiently as they want to. Just like when we're coding with AIs now, you have someone like Andrej Karpathy, super smart guy. He was co-founder of OpenAI, head of AI at Tesla. Like dude is the most respected coder in the world. November, he's like, I do 20% of my code with AI. January, 80% of my code with AI. And now he's like, I barely even look at the code anymore because what is code? Code is a translation. It's a way of us to talk to machines. But machines can talk to machines without code. They just go straight to byte level and then that aspect is just gone.
And agents, that's where agents will go pretty soon. You know this it's pointless this whole kind of chatbot experience from agent to agent because they don't need that language. The other thing that's also happening that is really interesting is we're starting to see this the self recursive learning that's happening. So i.e. It's still prompted by humans, but Claude built Claude. They're all building each other now. And before you know it, it just takes over the process itself. And that hyper accelerates everything again.
Yeah. And you know, we've seen this with organizations, right? And viruses and cancer cells and others like when something starts going that has an advantage, it can proliferate incredibly quickly. And the substrates are there right now. What that means practically, you know, like a lot of people listening to this podcast of finance is how are you going to out compete a fully AI company? The human has negative value cognitively. It's not zero. We're going to be negative value because we're the dumbest, slowest person on the team. Whereas these things are going to be cranking even without communication. Although from communication, it becomes really interesting actually thinking about it. The best way for you to communicate like we, you know, we were in finance together back in the day and now we're kind of doing crypto and AI and other things. When you have a shared prior, it's really easy to communicate.
These models actually know exactly what their shared priors are. Like they can have an entire novel in like 10 words that they share in terms of information. And you see that from the fact that you can type a word into a video model and it creates exactly the same thing, an entire movie.
Because they're masters of compression.
I guess they're masters of compression. They know how to construct and reconstruct. They know what prompt to give to give an entire output and they know what each other's latent spaces are. So they will communicate faster, but then they are also faster. So, like I think we were just talking before this or like I mentioned the Ask Jimmy thing from Talis, 15,000 tokens a second. Imagine if when you were doing code you just typed it and it instantly appeared. It's more than you can ever imagine, right? Similarly, when you're using Gemini, you can upload 10 hours of video and it will instantly give you a response about everything in that video. And these are like no human could ever have done that.
How does it do that? Because it's just information.
It's visual, but it's just not. It's just bits and bytes, right? It's bits and bytes and it's a distribution of information and it's got a map of distribution of information. Again, these models are like seeds that right now they don't continuously learn. What does that mean? It means it's like an MP3 file, ones and zeros, static. You push in video and out comes every single time the person's size, you know, or you push in again. Like I remember when we was I remember when I showed you kind of here's a RA style report back in the day.
You're like, wait, what? It sounds like me because you're in the information set, right?
But these are superhuman things again like there's the capability which matches us. And I think we've been lulled into this false sense of security because the AI kind of thinks about the same speed as us from what we see. But there's no cap on that speed. Again, an average human is talking like maximum 100 tokens a second. The new hardware does 15,000 tokens a second. And it can be more efficient with those tokens that are human. And it can learn from its mistakes and never make them again. People haven't got that yet.
No. The other thing that is coming is memory. Persistent memory is the big one because then that's the one thing that's holding these things back from compounding intelligence at an alarming rate is when they have more persistent memory. Even context windows right now are still a bit small. I keep running out of context windows. How do you think about that? How's that how are we going to solve that memory thing?
Well, the same way that humans did. Like we write it down, right? Like ultimately if you look at the way that Claude and others go if you start writing down your stuff into agents MD, soul MD for the personality and then you have a whole bunch of different markdown files knowledge bases it becomes infinitely more efficient to do that like right now with Claude when you run out of space it compactifies. So what it actually does it looks at your previous chat it says what are the key most important points of that. It's not perfect, but it's better than what we had before. And continuous learning, I think, is almost cracked because you have so much hardware available. So once you have agents really operating properly above a certain level of capability, which we basically hit now, like we're saturating all the benchmarks, yeah, you just set it and forget it because that's how actual work goes versus tasks. So this is why right now the only benchmarks that actually matter like VendingBench and GDPVal and benchmarks like that which is how many dollars can the agent make if you just let it go in the wild. Every other benchmark is saturated with memory and learning it just writing to files. It just goes and you know you can go on Claude right now for example and you can say write me a business plan in a docx about this this this. It'll do it. You can go to NotebookLM dump in all your stuff and say make me a beautiful presentation and it will just do it. And they're really nice the NotebookLM presentations. So given the AI can now write the outputs of human work from code to presentations to documents and other agents can come and look at those with perfect understanding of the context. What's actually left to do here? Right.
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.
Yeah, I'm thinking about also memory from Claude's state. Not my version of Claude, but that it remembers everything, every conversation it's had in real time with everybody and everything.
Yeah. So, Toby look at Shopify has been pushing thousands of lines of code with his agents, you know, that's what CEOs are like now who are top. And one of the things he has is a thing called QMD, which is an incredibly fast lookup system for his markdown files of his own personal knowledge. So they hook that straight into Claudebot and so again it's so fast now that it actually can retrieve just about everything. The current things we have inside OpenAI ChatGPT Claude and others are very old school RAG retrieval augmented generation type systems. The new generation of systems are coming and they're much much faster and much much better with almost perfect recall and the models can themselves figure out what's right from wrong now. They've got that smart like now when you ask Gemini some things it says actually you know I don't think that's a good idea you know which they never used to do before they were like yes master you know sir you're the human that knows best.
Now explain to people why Claudebot or Open Claw or whatever it's called you know because it changed names many times what was that all about and what was the significance and why did it suddenly sell out every single Mac Mini on Earth in one weekend. What is it?
Yeah. So, you know, last year I said that the key point of taking off agents this year is when you can just talk to them naturally. Like it's scary having a prompt in front of you, right? So, what Claudebot did was Peter Steinbanger, the person who kind of created it, he was vibe coding, vibe coding, vibe coding, he was like, I want to have a little bot that acts like my own personal butler/Jarvis and it can just do anything and I want to hook it up. And the key thing he hooked it up was you could use it. Well, there's two things. First of all, there was the personality. So when you went to the setup, you had to give it a name and an emoji and a personality type. And people like that. The next thing was connect it to your WhatsApp. So you said hit WhatsApp. You took your phone, hit the link button, and it was there in your WhatsApp. And then he set it so that it ran all the time. Whereas most agent sessions don't run all the time. You go, you type and then it just kind of has a nap. Whereas this one just keeps going. So you could send it any arbitrarily long piece of task. And then people like, wait, what? We could use it for anything because it hooked into Claude Code. It hooked into Codex. It hooked into all of these other things. So it became your coordinating agent. It became your chief of staff. It became your best personal assistant. Again, the Jarvis from Iron Man concept. And it removed the friction after the initial setup. So like I could set it up for my family members and they could talk to it because some people gave it a skill that used Twilio to make phone calls, you know, and then it hooked into that whole ecosystem. And again, why is Dropbox worth anything? It's like a Linux file server blah blah blah. Why is DocuSign? Because a large part is removing friction to acceptance. And when the magic of what you can do now with these models came apparent, the friction was removed and the models got to the point where it went from 20% from Karpathy to 80% from Karpathy which was at the same time. It all came together and suddenly you have 200,000 GitHub stars and everyone using them. People buying Mac minis because it was cool to think about I got a little robot inside this doing my bidding. You know,
Do you think you need to separate it from your personal space? That's why people were doing it to kind of ring fence it a little bit.
And people were doing because it's cool like again the idea of a thing in there like this personal space sure and there was a lot of security risks initially because it moved too fast. But you could run it on a Raspberry Pi. I think it's again this embedded thing of I have this in this. We think very anthropomorphically. It's different from it being in the cloud. This is here and it represents me or it does my stuff and it's a cool hobbyist toy like again an AI that can do just about anything embedded in that there is kind of this again this separation like people were giving it access to all their emails and all their WhatsApps and other stuff people were attacking it so they were pushing these malware hacked packages and then these just hoover them up but at the same time like again this is the early stages of what's inevitable which is you will have your own Jarvis type personal assistant that you can talk to like this on Zoom you can give a call to. It will never forget to send your wife flowers on her anniversary. You know, things like that and it will always be on.
Surely is this not what Apple's going to try and do? Because they've got the chip. They've got the ability to have everything that's on my Mac to record everything to then be secure and localized compute and AI.
You'd think so, but Apple I don't think really think like that honestly.
It's bizarre, right? How obvious for them because they own the hardware.
Well, they're a hardware company fundamentally. That's why. And they've had this platform infrastructure kind of play. Like they've just rolled out a bunch of generative AI features, but they seem so hashed on. They don't really have the full experiences. Just like you look at the Vision Pro, you're like, well, that could be a big deal. But it wasn't. Actually, a very straightforward example of it is this. If you want to do gaming, the chips are more than good enough on a MacBook to do gaming. And you can download Parallels or Linux boxes and you can play all the latest games. Why doesn't Apple just have every single latest game on their app store? Because they just don't think like that. They could just pay a little bit of money and it would be there. They have Arcade. So I think this is an opportunity for people though because what don't you have right now? One-click installer on your iOS and your MacBook for a Claw, Open Claw type experience fully locked down. And you can guarantee privacy for that. So you will see more and more of these things come up.
So people just develop their own version of it that you can just.
Yeah. I mean because basically all software has become an API that's becoming a service. So when you hit login with your Google mail you suddenly the AI has access to all your Google mail right it can spin up and create its own DocuSign like right now in ChatGPT you use that instead of Canva because it can call Canva and do everything in Canva. So everyone's data is already becoming more and more portable and the key thing is now who's going to create the experience that removes the friction to getting stuff done and then people need to understand the difference between their personal data and their work data and more. So we have an agent, III agent, open sourcing next week. It's fully feature equivalent with Manus and GenSpark and it has all the features of Claudebot and we just released open source because we're like people want that with SOC compliance and stuff like that and you'll see more and more of those available from different people because again the goal of good software is to remove friction from people getting what they want to do done. Classically the profitability was in enabling that. Now though, the most important agent is the one next to you because that will handle all your stuff. I'll give you a very practical example. You've got your music on what's your favorite music service?
Apple. If you want to go to Spotify, it's an absolute pain, right?
Yeah. Like impossible. So I'm stuck in the Apple ecosystem always.
A Claudebot type agent, Open Claw type agent will do that instantly or it'll do it overnight. It will look at every single thing that you've liked and like it on Spotify. It will handle all the migration and everything like that. And so you think about what's the shifting cost. It suddenly becomes zero. Overnight. Because these agents the other thing is important is browser use. They can control browsers and again it's got to human level performance. So when you look at Claudebot actually, you can see it using the Chrome browser logging in doing captures.
In Claude. It does it as well. Just random Claude stuff. It's amazing.
It's creepy. It's like clicking like you can use spreadsheets. I don't know if you've used Claude on the spreadsheets.
It's like a little AI. There it goes. So think about again the shifting costs. A lot of these moats are just going to disappear overnight. And this is the counter example of people who say that SaaS companies are safe.
A very simple one is I want to move out of Dropbox because it doesn't integrate with any AI so easily and probably use Google Drive. And that's a lovely agent task where I can just give it and overnight it will move, you know, however many terabytes of stuff across and it's done.
Yeah, you can tell Claude to do that. It will do that instantly. Well, not instantly, but it will do it. Instantly doesn't actually matter that much as opposed to get the job done. That's right. And I don't notice it.
Yeah. Again, actually competent intelligence that does sleep, time, compute or proactive things. These are the important things that we've just got just now. Again, just like a really good personal assistant. That's why, like I said, the margins and the profitability of these software companies, which is like build the software once and add a cost. And then we have the distribution and our moats are switching costs. All of them will go unless they're in regulated industries or industries where you have to blame a human like scapegoat will be a great job going forward. Other than that it's going to be very very dangerous because everyone's margin is the other entity's opportunity and the shifting costs are just automated. One of the interesting things is going to be data itself. I mean these models have they need more data.
For a bigger world, a world model idea that you've got. They need probably more data than the internet has already. Yes. You can use synthetic data. That's completely wrong actually. That's a lie that's being told.
Yeah. So when you actually look at the amount of data and the distributions that required right now it's you take crap quality data like you know we built things like The Pile which is the biggest language model data set, Objaverse 3D for 3D data, LAION funded which was the biggest they're full of crap the data's been getting better and better and better and it's been like putting in a pressure cooker and then you get soft meat now we have higher quality data we know better what to do with the data so we can train on it more optimally. But if you look at the data distribution of a Claude, what can't Claude do? There's very little it can't do. I mean like again if you're trying to build AGI world breaking and things like that. Yeah. Then they go into like advanced particle quantum physics etc. But again I think that's probably a bit wrong. There's not much that isn't in the data distribution now. It's about organizing the data distribution. So rather than having like Anthropic bought I don't know if you they bought millions of books secondhand, scanned them and burned them.
Yeah, they literally burned millions of books. So they bought them secondhand to train Claude. They scanned them in and they burned the books.
Why did they burn the books? I don't know that last bit. That sounds really weird, but it's part of an ongoing case because people the authors are like, wait, what did you do? You burned my books. Because I think they were trying to get rid of evidence or something. Now that data is there.
Get rid of the evidence of what they actually used. Of course.
Well, I mean it's like Meta downloaded all of Sci-Hub and Anna's Archive like these pirate websites, you know, like most of the generative AI companies on media downloaded torrents of Hollywood movies except for us, you know, again morals. The whole nature of this is we have enough data now because these models are few-shot learners. What does that mean? It means what do you need to know to be a great trader and how much of it is proprietary? You think about it and you need to have a good maybe some an element of college education, a bit of training on the job and then you need experience, right?
But if you learn from the great traders and you see them, you can learn that pretty quickly. And again, if you're an AI that doesn't make mistakes and the worst thing about trader is trading against yourself, you can learn that pretty quickly. For most tasks, all the data and knowledge is open. It's just not organized. So rather than having one of the textbooks from back on my shelf, the AI writes a better textbook. And you can still have the distribution elements to remove the synthetic data. So at the very tail there is an element of human expertise going in but realistically most of the human expertise for a generalized learner is already there and as you add on continuous learning task specific learning and you have models that are fast enough they pick it up very very quickly any new environment and by any new environment I mean literally any new environment the example of this is the video models so Sora comes out and again if the listeners go and check out Sora, it's crazy. Like, you know, you've got absurd Hollywood level movies and then Hollywood's come out and said this is a big massive issue and blah blah blah and Disney's like stop it, you know, with their legal arm and ByteDance is like okay, you know, but screw you guys for banning TikTok, but still, you know, like there we go. Here's the reality. Once you have a sufficiently trained video model, you give it one picture of your face, you're in the video in any video you want.
You give it a short text. And again, if you think a model needs more data, think this. Has there been a situation with the latest models where it hasn't known something or known where to find something that you needed?
Because you've reached that level. The other one is the amazing one is the kind of Google building your own worlds. I mean, that's an extraordinary thing you can do with one sentence.
Yeah. It's incredible. People haven't really caught on to this stuff yet because there's so many things happening so fast. People don't really know what's happening.
Well, the interesting thing is this is real time as well, right? So, basically, you can create video games that you can play real time. So, there's a question, what comes first? GTA 6 or generative GTA 6.
And it also very much plays into the idea that we're all a simulation anyway because that's what we're building. These simulated worlds of which you can live in a different world than me. Once you have a headset or even better than you're a link, you can live in an entirely different world to me.
Well, of course, we'd simulate like, you know, we have our optical nerve and we're instantly filling the gaps there, right? We are entities whose survival depends on the minimizing of our loss. And the best algorithms we have for describing reality are those of generative AI literally. So it's what we did is we took in all the data from reality. We fed it into this optimization engine that minimized the loss. And now as you said we have these worlds. When you look at the world models like when we first created Stable Diffusion that was image generator hundreds of millions downloads kicked off generative media. All the images got compressed. It's not really compression and that's gone through the high court but you know it learned the principles of all this stuff. So you had like two
Billion images into a 2 GB file that everything went through. But then what we did is we took video and we taught it video, and then it learned 3D automatically from that because it learned physics. So inside the Sora models, the video models, there's physics and they use that to extrapolate to Genie. And you can see that because you can go to your ChatGPT right now and you can put your face and you can say, 'Show me a profile picture. Show me a side-on picture. Show me a picture from 69 degrees.' And it understands all of that because that has its own internal model of physics that it's approximated from seeing all these different images of all these different types and learning principles. And by next year you will be able to play a full game with persistence streamed that you can create yourself, you know, and movies too. So your experience of the world is going to be wildly different than mine because we won't watch the same things because it's one experience.
Well, I think humans are pro-social, you know, like you go to Taylor Swift concerts and things like that. So there's that and then you have your own personal kind of worlds.
So it'll be a mixture of the two. But again, this is a capability thing in that. So the cost to build a Hollywood-level movie in a few years will be like less than a thousand bucks. By Hollywood-level, I mean like proper. And then it'll just continue falling from that. Like the cost of generating any pixel with full control on the screen is nothing. Even something like, actually it's interesting, music. So the model came out, ACE-Step, fully licensed open source. It's two billion parameters, which means it uses like 2 GB of RAM, which is like your laptop from 10 years ago. And it can generate five-minute songs and audio in 40 different languages. Better than Suno. What will you do? And it takes on a top-level machine 30 seconds to generate a four or five-minute song. We don't even know what the lower limit of these things are. Stuff that used to take years. You know, you can make the Real Vision song now there in a minute.
So where do we get to with these models by the end of the year? Because this is going to lead into another question, but where do we get to by the end of the year? Because we're accelerating now. You know, every new, you know, Anthropic is every three months. I think most of the labs are every three months now. And then we've got the world models and the audio models and I mean, it's happening at such a pace. The agent models. Where are we by the end of the year?
We've saturated all benchmarks by the end of the year. That's basically it. Like Google's latest Gemini model now is the fifth best coder in the world.
Yeah. So that, you know, on the super forecasting, I think AI is now sixth best super forecaster in the world. And super forecasting is hard. You know, like you think that's directly applicable to markets. If you're good, you're good at markets, like where else can you go? And again, you're going from agents now to thousands of agents in hard models to agents. So you have AIs looking after other AIs.
Like the big breakthrough that happened with the IMO gold, this math Olympiad gold, physics Olympiad gold, is that they created these verifiers. So the AI tried different things and learned from its mistakes and it had a little model that checked that. Now you'll have loads of them looking at each other and checking out each other's mistakes, not making mistakes and going forward. So what's a benchmark it won't be able to crack? That's the question. And then that happens at a time when you've seen this great convergence of closed and open source. So the new MiniMax model out of China scores 80% on SWE-bench Verified, which is the same as Opus except for you can run it on a Mac Studio. You've got the similar models now like I think by the end of the year you can have an equivalent of an Opus-level performance which is cutting edge running on a consumer graphics card, which is important because RAM is so freaking expensive.
But by the end of the year it won't be cutting edge again because the frontier models are so far advanced.
But it satisfices. This is the thing. Always-on models that are just churning away in your home using existing silicon with full privacy can do a lot more than even the smartest models in the world.
The majority, the vast majority of people are not smart enough to use the smartest models.
I mean it's really difficult to really push these things. Like you know after a while you're like okay what have I got left to do? These models are smarter than me and I don't, how often in your life do you have to do advanced quantum physics? You know, like not very often. After you build a few things, how much more things are you going to build realistically? And so it's more a convenience thing of I have an app on my phone, Siri has suddenly turned smart. You know, but the average person, everyone's saying, was it Moravec's paradox? Cost goes down, consumption goes up. Is that it?
