EP201: Why One Breakout Subnet Could Sends TAO to $2,000 with Mark Jeffrey

Bitcoin is to currency what BitTensor is to free markets, not just AI. Once you have the supermodel, the value is not here. The value is in who can provide the cheapest inference, which is a game that BitTensor can play better than anyone. In a race to the bottom in a rapidly commoditizing environment, he who can serve the models, the cheapest, the fastest wins. If shoots by itself can get up to one, two trillion by using a mixture of experts, but you don't need the GPUs or the XPUs in one data center. You can have it spread out for free by the miners. BitTensor on shoots alone is worth 30 billion. That's the value is 30 billion bucks. Mark, welcome to the TAO Pod, the rebirth of the TAO Pod. It's been a few months since I've done a TAO Pod and I guess what the elephant in the room is, JJ, who I used to co-host with, he's not really as into TAO these days. No, he's not. I'm not really sure. And I honestly have no idea why, but I did notice he basically has stopped TAO, you know, basically BitTensor posting and is now fascinated with microtubules and the quantum consciousness and things like that. So I don't know what happened, but it's now different. And, you know, look, to be fair, people have different interests. They move on to different things. I'm sure he's still invested in the ecosystem. I know he is. But I would say I've noticed in general, and this is, let's start at the high macro level, and I want to talk about what's going on in the TAO ecosystem. But there's been so many changes in the TAO rules that Jake Steves, aka Const, has implemented. And we're going to talk about AI and the changes in AI, decentralized AI. But for crypto in general, TAO is a crypto. It's a protocol. And rules govern that protocol. And Jake has been making a lot of changes. I happen to agree with 100% of his changes, but some people get frustrated. They think there's too many changes. Bitcoin doesn't change that much. Why should BitTensor change that much? Do you think Jake is just trying to fix things whenever there's a problem? Or do you think there's like a good, consistent philosophy here? I think what's going on is that Jake is running BitTensor much the same way Elon runs all of his companies. And BitTensor is not a mature protocol yet. So even in the early days of Bitcoin, Satoshi was like fucking around and changing things and fixing things. Like there was a period and it didn't last as long as BitTensor's. But I think what BitTensor is doing is an order of magnitude more complex, right? It's a lot more than just transactions, right? And wallets. Like the whole sort of AI side of it introduces a much larger number of complexities. And I think it's getting closer and closer and closer to sort of cracking the code on reinventing capitalism in general, but also for AI particularly. With all of these changes, and I think there's no way to know in advance what the correct changes are. Until you try something, the market reacts in some way, and you adjust the machine. And I think the correct answer is to adjust the machine and do it rapidly because the only thing that matters here is that one or more BitTensor subnet products make it over the wall and become successful and have a billion dollar market cap or more sometime in the near future. Everything else is a sideshow. So the wrong answer is just to sit there and watch and poke your little stick at it and go, come on, do something. No, you want to fix what's wrong and actively be involved. And I know that's pissing some people off. But as you said, I agree with 100% of the changes also. You bring up a good point about between Bitcoin and BitTensor. Bitcoin was about creating – everybody has an incentive to have a good currency. That's how we measure our value in society. That's how we measure the value of other products. And we make that trade. We'll trade our value for the value of other products. And you need a good solid currency for that. So there's an incentive for a decentralized currency like Bitcoin. There's an incentive to make an incentive mechanism that works so that everybody participates, so that people buy the currency, so that people validate transactions and so on. With BitTensor, it's one step further. You're not just validating transactions. You're validating intelligence. You're like, somebody creates something, an AI, a video, whatever, and somebody else has to validate, hey, this was an intelligent improvement to the system. And it's not just a currency. It's the entire free market system. And in fact, I don't even like the words capitalism because that's a phrase really defined by Karl Marx and other socialists in the 1800s. So I prefer the word incentivism, that we all operate on incentives. And this is kind of the crypto ethos and where AI meets crypto. But BitTensor is really about creating a free markets version of Bitcoin. Bitcoin is to currency what BitTensor is to free markets, not just AI. And I think people always assume it's AI, but it's really more about all things that are a product of intelligence, not just artificial intelligence. And it's difficult. Yes. Well, I think, first of all, I didn't know that the word capitalism was actually a Marxist slur. It's pretty hilarious. It's to basically sneer at the people who are accumulating capital at the expense of everyone else. Like, oh, I don't want Mark to have capital. I'm going to take his capital and just put him down into the proletariat. But we know that free markets is not about the accumulation of capital. It's about creating something of value in the world and being rewarded for it. Yeah, totally agree. And remember that BitTensor, unlike Bitcoin, we have the subnets, right? So we have this added extreme complexity of having emissions, having subnet tokens, having the dynamic of staking Tau to get subnet tokens. That's the only way you can get them, right? And figuring out how to do the emissions. And the most recent thing that Jake has tweaked is what happens with the emissions, right? And before Root Reborn, which is what's got everyone in a twist right now. Well, with Root Reborn, before that was released, the emissions that were produced by the chain were being used to sell subnet tokens with every block. So the chain was dumping roughly a thousand Tau worth of subnet tokens on the market. The chain was working against you. It'd be like if you're in the stock market and the Federal Reserve is printing money to bet against you, right? Like, it's just like, dude, come on. Right? Like, don't just do nothing. Right? And so that is now what in Root Reborn, the chain doesn't do nothing. The chain actually takes the newly printed money and uses it to invest in subnets actively. So it's now assisting with the value creation in the subnets in a smart way. Like, the best ones are getting the most emissions and the most buys. The chain is subsidizing intelligently the entire marketplace based on the market signal. So it's not just sort of making decisions on its own. It's not a politburo in centralized planning. It's actually just taking what the market is already doing and force multiplying it, which I think is the right answer. Right? But this has tended to result in a world where the best performing subnets are getting the most subsidies, which I think is the way it should be. But everybody who's not in that sort of group is like, I don't like this. This chain sucks. Right? So the losers in any change are going to complain. Right. Like, so you bring up two points. And this is really, this is just technical on Tao, the crypto and Tao, the ecosystem. And then there's the larger picture of, okay, what's happening and what's going to be the breakout subnets or products or whatever. Basically, Jake figured out a way to make it more possible for a breakout subnet to succeed, which is what you want. To your point, this is a very just technical thing. We both saw Greg Chavez over at Yuma did a report on this. If even, and then Sami Kasab also did a report is if even one subnet breaks out, Tao's going to a thousand bucks. Like, yes. And there's 128 subnets. And a lot of them feel like they're on the verge of breaking out. Just if they just need that little push, they're going to break out. Whether it's NG or shoots or ridges or Targat or Liam or whatever. Minos. Yeah. I could keep, I could go keep on naming them. Like they're all the Yanez, you know, score. There's all these subnets. Oh, just push them just a little bit and they'll, they'll, they'll break out. We know one is Tagoso a thousand two. I think it was Sami who said this two or three, it's like three to $5,000. So it's like unbelievable. And the other thing is doing in a way that doesn't penalize people who are just buying Tao and staking to root, which is what the average institutional investor is going to do. And so previously there was all this selling pressure every day from the validators on Tao tokens. And now for all these complicated reasons, because of root reward, that doesn't exist anymore. So this is all good. I feel like right now, more than ever, the pathway is cleared. The only thing I would say is quantity in order to get quality. We see this with artists and businessmen over time. Picasso made 117,000 works of art and less than 1% of them were successes actually, you know, made money for him. This is true for every single artist. Quality is a quantity game. So I'm always in favor of more subnets gives more at bats for a successful subnet. So, and by the way, that also increases the demand for Tao because you need to state Tao to register a subnet. So, and the other thing about having a breakout subnet is that it shows entrepreneurs. Oh, wait a second. You mean if I create a subnet, I'm going to be subsidized to the tune of millions, maybe even hundreds of millions of dollars of emissions of Tao to help create my product. If I have a subnet, that's going to increase the demand for subnets. And so they'll make even more subnets and more chances at bat. So all of this is a virtuous cycle. You just need one to kick off that cycle. And that's, I feel the path is cleared for that. Maybe the path needs to now still have 256 subnets or 512 subnets, whatever, but the path has largely been cleared. And then we get back down to what's going to be the breakout subnet. And then there's my most recent paranoid question of the day, because I'm always nervous, nervous. I'm an optimistic, nervous person, but always err on the side of optimism. But what if, and my what if is, and we talked about this last week on Hashtrate, which is a great, it was a great episode to help me think about these things. Because, you know, clause on version five, what if Claude 12 can just do everything that intelligence plus computation can do? So what subnet can beat a Claude 12? Yeah, I mean, so, well, there's two answers to that. One is, one is that for some period of time, you know, you're seeing subnet 44 score, right? They are, they are specialists at vision. And when, you know, at least for the time being, you, you, a general purpose AI is not going to beat a specialist AI, mostly because the specialist AI has all the training data to, you know, for in Scores case, for vid, for computer vision, you know, it knows how to detect fire. It goes, oh, that's fire. And that's a guy with a flashlight. And I know the difference, right? Right. So it knows how to see because it has all the training data, which teaches it how to see. And Claude doesn't, right? Because Claude's got like a wide swath of things. That's a great generalist, but it doesn't know there's this whole vertical of training data that just doesn't have access to, right? And maybe someday it will. So, but I think that day is a couple of years off even now. Can I add to what you just said about score? Here's the great thing about score. So, so again, score, like you said, is a computer vision model. If you know, the state of California could hire it to do early detection of if wildfires are starting. And here's the thing about score. Let's say Claude has a better AI than score. You still need score because, because here's the thing. Let's say I'm the Chicago school system and I want to hire a computer vision company to constantly monitor the cameras at every elementary school and make sure nobody's pulling out a gun. Now I can hire humans to do that, but humans make errors and maybe I can't hire enough humans to do that. There's, you know, 55 schools in Chicago or a hundred schools in Chicago. So I need, I'm going to go to, I'm not going to go to Claude though and say, Hey, can you guys build this for my public school? And then, you know, and do all the code and figure it all out and then take legal responsibility that this works. And Claude's going to say, no, no, no, no. Just hire. We'll, we'll help whoever you hire. We'll provide the AI that someone else could figure it out. But you still need to hire a company to do this. Score is still in that company. The AI vision experts, because they're miners for all we know, they're using Claude. We don't know, but you still need to hire somebody, not the supermodel. You need to hire the specialist to do this. And there's going to be always a role for that. And I also think that eventually the supermodel is a mixture of specialists. It's not like a mainframe, right? It's a network of little, of a million verticalized AIs that are amazing at whatever their specialization is, right? And you basically got this hive mind of experts. And that is the supermodel of the future. And probably Score will be the vision part of that, right? Or might be the primary one or one of several primary ones, right? So I think there's even then there's still a role. It reminds me of like Linux. So Linux versus Red Hat. So Linux was this open source operating system. Anyone could just download and use it. Apple could use it. Microsoft could use it. But instead, they spent hundreds of millions or the equivalent of billions to A, make their own operating system on top of Linux because they still needed to fit Linux to their specific needs. And the Linux open source foundation or whatever was not going to do it for them. Just like Claude is not going to solve Chicago's public school systems issues with kids with guns. And on top of that, you had the Red Hats of the world that said basically, hey, yeah, you could download open source Linux and figure it all out. But we're specialists. You can hire us and we'll write all your Linux code for you and solve all your Linux problems. You still need that layer. Though Red Hat still needed Linux experts to work for them. So whoever, whatever AI vision company you hire still needs a subnet with miners who are specialists who maybe use the frontier models to still solve the very specific problems that customers might have. So this is kind of actually, this conversation is just changing my whole perspective on this one question is that yes, even if Claude is the perfect, there's a need for the specialist layer to be high, to be highly focused and built by either an R&D or a subnet, which has millions and tens of millions of dollars of free R&D provided to it. Yeah. And Scor in particular is distributing their product through PricewaterhouseCoopers. Yeah. Right. So that's, that's the Red Hat in this, in this world, right? So they're already proving your point. Now, the other sort of answer to that question is, you know, as the, as the frontier model, well, first of all, the frontier models are already being commoditized, right? So, you know, Anthropic, OpenAI, they came out with Mythos, they came out with Fable. And then within a week, maybe a week and a half, along comes Kimmy K3, does exactly the same thing. It's a Mythos class AI, basically, right? It's the equivalent or better kind of depending on who you listen to. And so instantly that moat kind of vanishes for those companies. And then, and that's the super AI, right? Like, so, so there is like, even if you get there, there is no moat. You know, I have no, I have no moat and I must scream as Harlan Ellison might say, right? So once you, once you have, once you have the supermodel, the value is not here. The value is in who can provide the cheapest inference, which is a game that BitTensor can play better than anyone. And he raced to the bottom in a rapidly commoditizing environment. He who can serve the models, the cheapest, the fastest wins. And right now we've got several plays on the board, Engie being the latest one, who is serving up Kimmy K3 and GLM 5.2 for half the cost of anyone else on earth. It's literally half off. The next, the next cheapest one, 2x what Engie is, right? It's crazy. And so, and it's only been out for like a day and a half. So this could be one of our big subnets, right? Like there's voracious appetite for inference. BitTensor offers earth's cheapest inference for the best model on earth. That is a true statement right now, which is mind boggling. And I think this is going to keep happening. Corporate America needs that. Like, I will just tell you for me individually, my clawed usage resets every Thursday. And I ran out of my clawed credits on Monday. So I was screwed. But of course, fortunately, I have Blue Tau, which is hooked up to Kimmy 2.5 actually. And actually it's hooked up to Quinn 3.5 I like. But they're great and they're good for everything I need. They're not as good as Fable 5, but I'm not doing enterprise level software. But this is a real big issue. But my worry there is if I was playing with devil's advocate, my worry is just like price per bandwidth from the internet days has gone essentially straight to zero, price per inference is going to go to zero over time. And you know, and already we're seeing Claude and ChatGPT, they're doing deals, you know, with big corporate enterprise, don't worry about it. We'll, we'll, if you have a price from Kimmy, we're going to top that. Like they're already playing that game and they're spending hundreds of billions on infrastructure. They could afford to spend a couple of billion now bringing the price down, you know, because they're losing money anyway. So I don't know. I don't know if that's a long-term superior edge, even though what, what NG is doing to provide cheap Kimmy is, is unbelievable. Yeah. Well, I mean, they're doing it by basically, you know, have recruiting miners to serve that inference from around the world. Right. Now they've done some custom tweaks to, to the model to make it runnable on, on much lower cost hardware than it's supposed to run on. Right. So they got it running on 50 nineties as opposed to, you know, Vera Rubin Blackwells, right. That, that, that kind of thing. Right. So it's order of magnitude cheaper hardware, but also because BitTensor is decentralized. The placement of those machines can be near power sources, which are also earth's cheapest. Right. So this will be profitable. So they're already profitable right from the get go. You're talking about the central AI labs subsidizing the cost. They can do that. And, but they're not profitable and they won't be profitable anytime soon, you know, and then they have never had to compete with this decentralized universe, which is profitable, which can keep making the service better and better and better and advertise it eventually. While still creating the product for the cheapest price on earth. So it remains to be seen who wins, right? Like in that universe, but I don't, I don't think these two forces have yet had a war with one another yet. So not sure where it goes, but it's going to be very interesting. Yeah, you're right. You know, and I wonder, I don't know, I, we're maybe neither of us have the answer to this, but with something like NG, I wonder if they can create, create an AI flywheel of sorts where even though I'm sure they're, they're doing it, you know, privately and cryptic, cryptographically encrypted and so on. So that you can't identify who's asking what and so on. But I wonder if they can take prompts and return answers based on the models they're hosting. And not only serve that to the customer, but also feed that back into an AI that's learning across all the prompts asked on all of their models so that they could cheaply and profitably, unlike all the other AI companies make their own AI model. That's actually the best. So if you're serving Kimmy K3 profitably, and now you can build a flywheel where you get all the prompts and all the output that you can feed back into your model to create the next version, I wonder if NG can create NG 4.0, which is like a better AI model actually than Fable. Yeah, that's actually a really interesting point because the prompts themselves are an interesting training vector, right? Like that's, that's how you create the next generation model. So that's extremely valuable. So if NG becomes very large, which there's a good chance it could be, it's just the prompts themselves will be valuable. But what I would do if I were them is I would then team up with like parallax shoots, right? Or IOTA, right? Somebody who's already working on the training problem and another BitTensor subnet, work out some deal with them, right? Get some subnet tokens from them and basically feed them the prompts to train their model. That is such a genius idea. And actually that reminds me now that you're here, parallax, the biggest subnet on BitTensor right now is shoots. And by the way, you know, we talked about them last week. They are the biggest provider of open source models to open router, which serves open source models to corporate America. And Stripe right now is apparently valuing them at $10 billion open router of which shoots is the biggest part of that. So there's a big question why it shoots a hundred million when open router is 10 billion, but that's another issue. But here is the problem that shoots solves that I don't understand. And you can maybe explain it to me. So the big problem, the hard part of training an AI model like ChatGPT is that you're using a hundred thousand GPUs and the GPUs are constantly having to talk to each other to basically run the back propagation on the neural network to train itself. And this is the big slowdown. Talking, if you have to use more than one chip, talking between chips gets you one tenth the speed as if you're staying in the chip. But but shoots is one step further. You're not just talking two chips within one data center. You're actually talking two chips spread across the Internet around the world, potentially. So or multiple thousands of chips around the world. So how does Parallax even possibly hope to compete with the GPUs trading in a that are sitting right next to each other in a data center? How do they do it technically? Yeah. Yeah. So sorry, I've been studying a lot of these things. I may tend to mix them up. But from my memory of Parallax, Iota solves it in a different way. But what Parallax does is it does a mixture of experts or exactly what I was talking about earlier. So they basically train very small models to do, you know, to do this domain of knowledge the best on Earth. And then later on, they just basically assemble them into a into a network. And then they get like sort of a router, you know, that takes in the prompt that goes, oh, this is a question that you go there. Right. Or I should ask these two and then merge the answers later. But you're not actually merging the neural networks together. So they get rid of that complexity. That's my understanding of how they're doing it. And it's very complicated because like Kimmy does a mixture of experts, but the AI itself figures out what the experts are. Like humans might not understand, well, what's this cluster of GPUs an expert in? Might just be some obscure part of English grammar. So anything with this grammatical form goes into this mixture of experts. And and so it's kind of categorizing using AI. I wonder which, again, requires the GPUs to be close to each other, because that's a whole AI model right there. So I wonder how parallaxes. Maybe they're kind of starting off with a pre prefix, like, you know, collection of experts. So so they're able to categorize almost manually. And then I don't know, it's very this is I don't know. I by the way, I love the science of this. It's like, I almost wish I was a grad student again studying this, but it's fascinating stuff. I think that John Durbin and shoots and parallax are onto something that it really does feel like they've got a tiger by the tail. They're shooting towards a one trillion parameter model, which, you know, at that point, we are pretty much, you know, we're getting into the frontier models at that point. Right. Once you cross the one T line, you know, the frontier is actually a two T, but once you cross one T, like you're within shooting distance. So I feel like they may get there first out of all the subnets. We'll see. So there's two things so far, which I've changed my mind on so far in this conversation. One is my fear that Claude would eventually be better AI than a score, like a computer vision model. I'm now completely over that because no matter what, there's still a need for specialist companies, specialist layers, no matter what. So, so, so that I've completely changed my mind on score. I think it's going to be a huge subnet, but, and, and Yanez and others and, and Minos and so on. But agreed. Shoots, I've been having my doubts on because of this open router question. Why is open router 10 billion and shoots a hundred million? But if, if what you're saying is correct, if they can do a trillion parameter model on parallax, this changes the equation of BitTensor completely. First off, Kimi, which is a mixture of experts on a, basically a 2 trillion parameter, 2.8 trillion parameter model that's being valued right now. They're raising money at a $30 billion valuation. And their whole point is open AI is a trillion. We're just 3% of that. But if shoots by itself can get up to one, 2 trillion by using a mixture of experts, but you don't need the GPUs or the XPUs in one data center, you can have it spread out for free by the miners. BitTensor on shoots alone is worth 30 billion. That's the values. 30 billion bucks. Because there's no, you don't get any value from the brand there. Like if Kimi says they're 30 billion, that is the value of an, of a company serving an open source model of 2 trillion parameters, you know, doing a mixture of experts. So if shoots can replicate that, then shoots alone is worth 30 billion. And this actually is almost making me think there should be other subnets competing with this because there are, this is just math. There's lots of ideas that can maybe get you that mixture of experts on a decentralized network like BitTensor. There are several BitTensor subnets now focused on this, on this exact problem, right? So we talked about IOTA and they're solving it in a different way. Basically there's a, anyone who wants to help train a model like a decentralized Colossus, even down to home max and things like that. They basically solve the problem of what happens when the machines that are in the training rig, the decentralized training rig drop in and drop out, right? Like Bitcoin miners drop in and drop out. Like having unreliable participants in the training network. IOTA has solved that part of the problem quite admirably. And that will, they're also eventually shooting for a 1 trillion parameter model. Then you have Conedo, right? Which we haven't heard much from just yet. I don't know though. Yeah. So we haven't heard much about it. So there's a woman named Isabella and I'm spacing her last name at the moment, but she has been with the OpenTensor Foundation for six years. So she's been with BitTensor since the very beginning. This is her subnet and she is also chasing decentralized training. She realizes this is, this is a big game hunting. And she's got an idea surrounding a mixture of experts architecture as well, which is probably similar in some ways in Parallax. She's probably doing some very different things also, right? So, but she's also another brilliant mind by all accounts. So there's a lot of people, there's several subnets working on this. This is blowing my mind because if all BitTensor is, is a hundred subnets trying to do a decentralized mixture of experts, i.e. a decentralized Kimmy, which is basically a slightly smaller Fable, maybe bigger than Babel, then that's the whole AI problem. You could put a number on it. That's worth 30 billion and then 1 trillion. Yeah. So I just want to switch tracks for just a moment to something else that's going to blow your mind. Cause I don't even know if you know about this. So we were talking about inference earlier. We talked about Angie and you were like, well, what happens when the big companies like subsidize their costs? That's great. What if it was just fucking free? Like your inference is free and you can do as much of it as you want. And there's a new BitTensor subnet called actual computer 95, which lets you and your friends or your company set up your own inference cluster. Right. So it's sort of like a private AI BBS and you hand out keys to your friends. Like I could give you a key maybe, but not other people, you know, so only us and our friends can get on that inference network. And it's accessible from any, anyone it's anywhere on earth. Right. So it is like an AI BBS just for us. Right. And you can have very big machines in there all the way down to home computers also providing inference. Now here's, what's really amazing. Just yesterday, noose research, which makes the Hermes agents just announced that they had, that they, that they're now supporting natively actual 95 inference engines. So the Hermes agents now natively, like you just basically check a box and a BitTensor subnet serves up inference for your Hermes agent. I use these, I use Hermes agent all the time. And so this is extremely valuable. It's earth's most popular agent now. So it is a big deal. So yeah. I mean, so what did the centralized AI inference people do? Well, now they have to pay you for you to use their inference. Like that's the only move they have left once it's free, right? If it's zero, where do you go? You got to start paying people to use your product, right? Are they going to do that? They might, but I don't think they will. I think that's probably a yard too far. Well, what's interesting too with that actual one is let's say, let's say I'm James Alvisher and I have my own idea for how to build a mixture of experts, but I don't know how to make a subnet and I don't know what I'm doing. Okay. I'll just, I'll now get a cluster of computers on actual. And like you say, I can, I can do my own flywheel of, okay, you know, I'm getting lots of prompts and I'm getting some outputs and I'm going to create my own mixture of experts, AI model. And the more, the more inference I host. So let's say I rented out to a bunch of, or let's say I just give it for free to a bunch of companies that do lots of prompting. I can now build my own mixture of experts model. And like basically the tensor could become, there's two, there's two universes we've outlined. One where the tensor becomes this, this specialist layer that the frontier models are never going to do. There's, there's no reason for them to do it. You could say they would do it, but that's like saying Google is not only going to be a search engine, but it's going to be your personal accountant. It's never going to happen. Like, so, so BitTensor, there's this universe where BitTensor is dominates the whole specialist AI layer. And there's strong reason to believe it will because the miners are always going to be better than your local programmers in your hometown. And, and then the other universe is where BitTensor becomes the default laboratory for making the frontier AI models of two years from now. It could be. Yeah. Which that would be insane, but yeah, I mean, it's entirely possible because I mean, you've got the decentralized thing going on, right? Just like Bitcoin, you know, in aggregate is earth's largest supercomputer, 500 times larger than the nearest competitor. And basically the dollars that pay for Bitcoin's existence, i.e. the electricity and the compute for the mining and the transaction processing that costs $15 billion per year. But you know how much Bitcoin pays? Zero dollars and zero cents. Why? Because it incentivizes you to compete for Bitcoins. And whoever competes the best wins the coins, right? So you can't beat those economics. It's such a great point because the big argument that centralized AI companies have is that, you know, like let's take SpaceX as an example. Oh, how is BitTensor ever going to compete with a million data centers floating in space? Like that's their argument. And but at the same time, you just said, look, well, how is it, you know, Bitcoin could have said the same thing. How's anybody going to compete with all of IBM's, Microsoft's, Google's and whoever else's supercomputers? Well, guess what? They just did. They just did it. And yeah, it all it takes is, you know, a lot of people participating and not even like a billion people. Like how many how many how many miners are there on Bitcoin altogether? I don't know the answer to that offhand. I'm going to ask Claude. Go ask Claude. Claude will know. Oh, wait, I have no more usage credits. You just use Grok. Grok will tell you. All right, I'll use Grok. You don't need like the hyper super intelligence for this one. Yeah, right. You're not cracking quantum mechanics here. By the way, Grok 4.5 is pretty good. We discussed this last week. I do think the latest AI models starting with Opus 4.8, really, and then Opus 5, I think they're there. I do agree we've hit some singularity point where this AI now doesn't really need humans as much as older AI. Totally agree. You can feel the difference. And I'm using Kimmy K3 now pretty regularly, and you can absolutely you can feel the fierce intelligence behind that thing for sure. So so mining machines are about 4.8 million mining machines. Yeah, crazy, right? Which which it seems like a lot. But you would figure right now there's probably, I don't know, eight or nine billion computers on the planet. So 4.8 million is not that much. So you imagine BitTensor. If BitTensor was even like a fraction of that, they're going to you don't need the million data centers anymore because you have it on BitTensor. Right. Yeah. I mean, I think, too, I mean, again, BitTensor plays the game of race to the bottom. So even if you have a bunch of computers in space, well, first of all, they're not in space yet. So there's some period of time where you have to launch them. And that costs money. Launching rockets is actually fairly expensive. Right. Putting them in space. Yeah. Eventually they'll be up there, but there'll still be a lot of commodity hardware down here on Earth. And the one thing that we've seen in BitTensor world is, you know, there's a constant drive towards making, you know, making lesser things do more. Right. The Chinese did it when we when we didn't give them chips and they still managed to cough up frontier models like DeepSeq and Kimmy K3, despite the theoretical lack of chips. Right. They just were more clever about how they use what they already had. We also have NG, which has been more clever about the inference side of it. Right. Making 40 making 50 90s work where previously only Blackwell's worked. Right. So and now we have actual computer, which actually got inference running on a Windows 95 machine. Right. So not very good inference, but they did do it, which is it's still it's not that the pig sings well. It's the pig sings at all. It's amazing. So I think that BitTensor incentivizes doing more with less. And that's just only going to continue. And that hasn't been fully mined yet at all. Let's say these two universes play out. And I think both of them are playing out as we see. And we're seeing it on a daily basis, like, you know, you're telling me about Isabella with her subnet and then Parallax and Targon and NG. And then on the flip side, there's, you know, we both met the PricewaterhouseCoopers guy at Proof of Talk in Paris. And I spoke to him how he's bringing score to essentially all of his clients. And he hasn't had one no yet is I don't know if he wants me to quote him, but he told me he has one client say no yet. So there's a real business there. And he can't take Claude to these clients because there's no one at Claude who's going to return the phone calls. So again, it's to my point. But how does this now, let's say Shoots builds the trillion parameter thing. Let's say it's worth 30. But let's say the Shoots company, not the Shoots subnet, but the Shoots company that owns the subnet key. Let's say that's worth 30 billion. How does value now go down to tau as we like, let's play this out. As you know, subnet tokens are, in a certain sense, rehypothecated subnet tokens. You must stake tau in order to get subnet tokens, right? So if the subnet token goes up in value and it's paired in one of the 128 pools against tau, it's going to drag up the value of tau along with its own valuation, right? So owning tau is like owning the S&P 500 of all the subnets. So you're going to do a lot better if you own the subnet, if it's a hot subnet. But even if you just hold tau, you're still going to fractionally, you're still going to go up. You know, if this entire ecosystem goes up and by entire ecosystem, Shoots or one of these big ones might become like 40% of the ecosystem out of nowhere one day, right? So that'll proportionally drag up tau quite a bit just by value of its sort of weight in the tau verse. So, but I think once there's one, there'll be speculation on tau that there's another and another and another. And so a lot of people just come piling into tau and the other subnets just because there was one, right? They'll be looking for the next one. So I think sort of this flywheel, like it just gets crazy at some point. Like we saw something like it happened in Ethereum, but there was no economic relationship between the Ethereum token and the protocols built on top of Ethereum. It was just a gas token. Whereas here, there is a direct economic relationship between tau and the subnet tokens since they are two different versions of the same thing, right? So, so I think it's very significant. I'll take an example you did, an analogy you did a few months ago, which I really like, which is that Shoots subnet is not the same as the Shoots company. And it's useful to think of the subnet almost as like the R&D department of the Shoots company. And so if the Shoots company has a lot of value, now the subnet tokens need to have value to keep incentivizing the miners, but they aren't necessarily 100% correlated with whatever the value of the equity is of the Shoots company. So do we believe that if Shoots itself is worth $30 billion, that that value, that some of that value or all that value goes out into the subnet. So it could be all that value, but what's, how do you think about that? Yeah. So I think of it like the product department, right? So I don't like R&D because that suggests like, they're just like a bunch of professors doing kind of nothing, right? Yeah. So, you know, this is the factory. This is the thing making the product, which is then sold. And the product must, is always in motion. It's always getting better. You always get to make more of it, or it has to keep continuously getting better and improving and improving and improving. So it's on a treadmill. Bitcoin itself, always go back to Bitcoin when you have a question like this, right? What does Bitcoin do? Well, Bitcoin security is always, always improving. The hash rate is always going up, right? As more miners try to compete to get, to get more, more and more expensive Bitcoin, right? So you couldn't just shut the miners off, right? Like if you shut the miners off, then there's nobody confirming the transactions. So your alternative, if you want to shut the miners off, is that you take all the hash rate in house. Let's say you're Google and you acquire Bitcoin. Let's just say you can do that. And you're like, well, you know, fuck the miners. We're going to do it ourselves. Great. You're now spending $15 billion per year to keep the hash rate what it is right now, to keep the quality up. And you must do that. And that price will get higher every year as the hash rate keeps climbing. Right. So you can do that if you want, or you can get it for free. You can not spend $15 billion. And the world will give you that product, the security of Bitcoin for free. That's exactly the same thing with all these BitTents or subnets. I think the valuation of the company and the valuation of the subnet are not the same thing. I think the valuation of the subnet is probably a subset of the valuation of the company. But I think it's maybe 60 or 70% of the value of the company, right? Something like that, my guess. I think that's right, because you and I have had lots of conversations about this. And there's one nuance which I've settled on, which is that, yes, the subnet tokens need to rise to keep the miners incentivized. But they only need to rise enough. They need to be the minimum price that will keep all of your miners incentivized. But the point might be is that, hey, if your company's worth $30 billion and the miners as a group are being paid, let's say, just $1 million a day or $100,000 a day, the miners might say, hey, well, I'm going to go over to Targon instead because they're treating their miners a little better. If there's these two separate universes, one's the specialist layer, the other's the LLM creator, there's a lot of competition among the subnets. So the miners will benefit. It'll be a buyer's universe. It'll be a miner's universe. They'll be able to command the highest token price. And that'll drive token prices up. Yeah, as it should. The miners are the ones that are providing the bulk of the value. Yeah. So I think the company is also providing value by sort of building this infrastructure around it in a red hat-ish kind of way, right? So they're providing the last mile for whoever is consuming the commodities that are being created and consumed. So it's not like the company has no value. The company has a lot of value. But the bulk of the value is in the miners and in the product itself, which is being created by the subnets and the miners. So I'm just going to put a number to this because it's very straightforward. If Shute succeeds, and this is just one attempt, we've listed at least four attempts. If Shute succeeds at doing a decentralized model that's a mixture of experts, that's a trillion, two trillion. If you do a trillion parameters, you can do two trillion parameters. So you're as big as Kimmy K3. Kimmy's valued at $30 billion. You just said subnet token prices could be 60%. So Shute's subnet tokens could be worth $24 billion in this scenario. And again, I know it's a reach. I'm making up these numbers a little bit. But Kimmy is valued at $30 billion. It's a two trillion parameter model. It's a mixture of experts. So we're just plugging these numbers into the model for Shute's. And if that accrues down to Tau, you're talking at least a 10x from right here, which is $2,000 on one subnet success. Yes. I don't think that's unrealistic at all. And that's one subnet of 128. Yeah. Yes. Yeah. I think once the bit flips, right? We're in sort of beat. We're in the zero to one. And we're still in the zero world, right? So there's still a lot of people that just because it's crypto, it's like, that thing, it must not be worth that much. It must be a scam. It must be like, right? So we're sort of, you know, there's prejudice against our AI because it's sort of linked to crypto. But I think once somebody smashes through that ceiling, the world is going to go, hey, wait a minute. You know, like Trimath did when we had the early success in decentralized training. He was like, some crypto thing trained a 72 billion parameter model. That was his words, right? Like some crypto thing. Yeah. So we're going to get a little bit. There's going to be a little bit of that. Right. But then once we do it once, it's going to be like, holy shit, some crypto thing. And then the second time I'm like, oh, yeah, well, somebody proved that it works. So not surprising that there's a second one. And then when there's a third, a fourth, a fifth, a sixth, all of a sudden it's just an AI thing. Right. And then the world at large sees that because it's crypto, they can get in on the early stage valuations of whatever the next one is in a way that they cannot with the equity based companies. Right. In order to get in on early stage AI companies, you have to be accredited and you have to be invited in. You have to have the access. So there's like two gates that almost nobody can get through. Right. But those gates do not exist in crypto. So anybody anywhere on earth can at any time invest in any subnet. There is no barrier. There's no KYC. Just invest. Just use the crypto that's out there. So it's a very unique, you know, wealth creation opportunity like Bitcoin was right. You didn't have to be accredited to buy early Bitcoin when it was like 50 cents. Anybody could have bought it. So I think we're looking at the same thing now for AI. And I think the world will figure that out. And it'll be the crypto people will figure it out first. Right. Like everyone who's like trading meme coins on Solana will be like, wait a minute. There's some like what's this AI thing? What's this Tau? Right. And then all of a sudden they'll start flooding in. Well, because it's legitimately complicated too. Like in the sense that I probably spend a good 10 hours a day now just trying. And by the way, I went to grad school for AI. I've been doing AI, you know, on and off for the better part of my life. But now I even probably spend 10 hours a day studying this stuff, doing things on top of the different AI models and so on. And not only being active in the Tau system, but it's like Parallax from John Durbin at Shoots. It's kind of hard to understand. It's not easy to understand what a mixture of experts is, what these neural networks are. And once you figure out like we're kind of doing here in layman's terms for what's going on at Shoots, what's going on at these models, it's if you figure that out, it's really is like you said, it's like getting in on Kimmy at $100 million before its next round is at $30 billion, which is what it's at now. Or open AI is a trillion and a half. That opportunity exists nowhere else. This is the only way to get in on early valuation AI anywhere on earth. At least as far as I can tell, I'm not aware of anything else where you can get in like this. So it's still pre-zero to one. But I think that bit flip to one is going to happen soon. It feels like to me. It's not financial advice. I just want to clarify all that. That's my opinion. It certainly feels like that's where it's going right now. Like the last week has been insane in bit of time. And we said that last week about the week before. So it's just going to keep happening. But Mark, I know you have to get onto novelty search with Jake Steves. So welcome to the TowelPod and all the listeners. Thanks for putting up with the delay here. We've a couple of months delay, but we'll be back on a regular basis. Hopefully joined again by Mark on a regular basis. And Mark, you've just blown my mind on at least two different layers. And I appreciate it. So thanks for coming on. Always a huge pleasure, James. Thanks for having me.