他正在为 AI 打造 Bitcoin……（TAO 详解）

I was the first developer to work on Bittensor. I helped write the white paper and spent many years of my life building Bittensor. I've been building it since 2015, which is a really long time.

Bittensor is like moving into a new era where it's the perfect ecosystem for agents. A machine-learning model is now at the point where an AI can mine Bittensor, build subnets, trade, and buy the commodities it needs directly from the chain. It all starts to really integrate as an ecosystem, and crypto is the perfect playground for an agentic world. What I'm excited about is basically doing that myself. Boom.

Mr. K, welcome to the stream, man. How are you?

Very well.

Nice to have you here.

Yeah, dude, I appreciate you jumping on at one-second notice. I know you're in a little bit of a weird time zone, but my whole feed is talking about TAO. I'm streaming, and everyone's like, “What is TAO? Please bring somebody on.” I think we got the best person we could to come on and talk about it with no notice, so I appreciate you being here. Why don't you start? Give us a quick intro—who you are, what you do. Give me a little bit of lore to set the stage.

Well, I was the first developer to work on Bittensor. I helped write the white paper. I spent many years of my life building Bittensor. I've been building it since 2015, which is a really long time.

2015, really?

Yeah, actually, straight out of university. I had the idea of the neural internet and how to mine intelligence way back then because I was doing my degree in computer science and mathematics, studying artificial intelligence. This was before AI was a thing.

Okay.

It was just the beginnings of deep learning.

Why were you doing this? How did that even come across your desk?

Well, it's kind of an interesting story. I graduated and was really interested in something called neuromorphic computing.

Okay.

Neuromorphic computing is, as the name sounds, like the brain—neuromorphic. I ended up working for this absolute genius named Alex Nent [?], who started the DARPA SyNAPSE program. It's a program out of DARPA where they were trying to build the brain chip. You guys have heard of Extropic, right? Extropic.

It's familiar, but I couldn't tell you what it is.

Yeah. Well, he's trying to make a thermodynamic computer—a computer that you compute with electricity.

Wow.

The idea is that, with modern machines, the CPU has this bottleneck between where we store information and where we compute on it. When you want to do, let's say, a multiplication on your computer, you have to take information out of memory, put it into the CPU, do a multiplication, and put it back. That's not natural. It's completely unnatural for there to be a separation between the memory, the state, and the thing that operates on the state.

The brain is the best example of that. Your neurons actually encode the structure of the neurons. The connectivity of your neurons is how you compute, and it's also where the state of the neural network is at the same time. When you're thinking, all of your neurons are firing at the speed at which neurons fire, which is very fast, and multiple trillions of them can fire at the same time, although that would probably kill you. There's no separation, so there's no von Neumann bottleneck.

Basically, I was working on that back in 2015. It's really interesting to see Extropic blow up because it's actually a very old idea to build these neuromorphic chips. In fact, Alan Turing was the first to propose this back in the 1940s. He envisioned a computer that was a neural computer, where there were neurons and all these things. He was one of the first to think in this direction of making computers like the brain for the purpose of artificial intelligence.

I was working on that after I graduated because it was pretty obvious that artificial intelligence was constrained. The bottleneck was compute, and if we could make things faster, that's what NVIDIA does, right? That's why NVIDIA is worth so much money. They can make it faster, and they're the fastest in the world. They have these $200,000 machines—GPUs—that people pay a hell of a lot of money for just to get speed, because if you can run more iterations, you can go faster. You can try more things.

Anyway, I was working for this guy, and he was a Bitcoiner.

He was a really interesting genius who worked with DARPA and made the SyNAPSE program to build these chips. He funded IBM, Intel, and Qualcomm to build them, and he basically saw them build it. He was so depressed by their speed at developing the chip that he went to start his own company and hired me as his first engineer.

That's how I got into Bitcoin, because he was this cypherpunk with, you know, 10,000 Bitcoin buried in his backyard kind of guy. He paid me in Bitcoin, and he was a really awesome guy. He taught me about real Bitcoin—not pump-and-dump Bitcoin. He taught me what it was about and what the ideals behind it were. As a consequence, I got involved with crypto.

I was a machine-learning engineer, and I was fascinated by how we were able to create this supercomputer in cyberspace that was so powerful. I was studying how to compute with electricity, and there I was watching this economic experiment, this computational network that is run by a form of electricity. Basically, that form of electricity is money. It's a mining network.

I put the two and two together and thought, “Okay, we can compute with money. We can compute with electricity. An abstract form of electricity is money. Here's this example playing out in real time in front of us: this supercomputer that is gargantuan.” The idea is that Bitcoin is so big that even governments can't outcompete it in terms of hashing.

How does it achieve such an aim, where even the United States government can't attack it? There aren't that many things that can say that. Most things are underneath the table, right? OpenAI and Anthropic are all funded by the American military. There are very few companies—

Crazy. Yeah, yeah, yeah, yeah.

—that are like, “I'm going to be at the nation-state level of computing.” I thought, “Wow. Okay, we can do computation with electricity. We can do it with money.” What exactly is that principle that allowed Bitcoin to be so efficient?

There are these super-geniuses out there. I've met the guy who made the Antminer. He's this Chinese guy.

I don't know what that is. Am I going to get flamed for this?

Yeah. This is old-school crypto. You had to know about this. There were the hash wars that happened, and it attracted these super-geniuses who basically won in the hashing game. They won by being—you know, they didn't have PhDs. They weren't qualified for the job of being the largest compute providers.

How did they win?

Gosh, I think about this a lot. How do you win in a game that is perfectly competitive? You're a superhuman player.

Perfectly competitive.

You're a superhuman player. You're a psychopath. I don't know. You're like Cristiano Ronaldo of mining Bitcoin.

What is that, though? What is Ronaldo's equivalent? Ronaldo has a 6-foot-4-inch vertical. What is yours?

Aesthetically, yeah, exactly. It's tangible. You can look at Cristiano Ronaldo and think, “What a handsome—”

Well, I meant soccer skills, but that works too.

Soccer skills—physical. But someone like this guy who made the Antminer, who took over Bitcoin, and in many cases the people who have taken over Bittensor in similar fashions, they're just hardcore capitalists. They know how to procure and organize matter really well and apply it to—

Yeah.

Yeah. Bitcoin is run by all of these people. You may know people who mine Bitcoin, or who used to when it was a hobby, but now it's an industrial process. They seek out access to latent energy reserves in Paraguay. If you're capable of moving across a border and getting electricity at a slightly cheaper price, you go over there.

That's all enabled by the permissionlessness and the global, 24/7, hypercompetitive monetary system that is Bitcoin. We thought, “Hey, let's take that to everything.” That's basically Bittensor.

So, first of all, that was incredible. I've never heard you talk, so I don't know. I just saw you tweet and I was like, “Fuck it.”

Let’s get this guy on. I’d never really heard you speak, so that was incredible. Can you give me the fifth-grader explanation of what Bittensor is and why it’s so special? Then we can dive into the nitty-gritty.

Try to explain it in a different way than I’ve explained it before, using modern terminology. You’re following what’s happening with Andrej Karpathy’s AutoResearch, right?

Vaguely, at a high level. You could break it down if you wanted to start there.

Yeah. Machine intelligence—AI—used to be really complex. People would sit down and try to write down the rules of, let’s say, chess. Then they gave up on that idea. There’s a famous essay called “The Bitter Lesson,” where they say, “Let’s just throw compute at it.”

Which means let’s just try everything, force compute, and figure it out. Yeah.

Whenever we find something that kind of works, we just do a little more of that. You know, that’s why OpenAI and Elon Musk are spending billions of dollars on compute, so they can run more experimentation.

All of AI is basically that loop: try, did it work, adapt ourselves, try, did it work, adapt ourselves, and do that 100 trillion times per second. Eventually, you create something. That’s the fundamental learning process of nature and AI.

Why is that a relevant starting place? When you start doing that process, you recognize that the core problem is defining the thing against which you validate performance. You say, “Okay, we did a bunch of things. How good were we?” That’s the validation, right?

And then we adapt, right?

Right. We push more into the ways that we’re good. Bittensor is basically that process: do a bunch of stuff, validate it, and then feed the thing that did well more money.

Instead of using the computational substrate that’s used in artificial intelligence—for example, genetic algorithms, where you mutate the genetic codes that are better at solving your problem—we give them more money. Bittensor is basically a machine for creating this cycle: things try to solve the problem, we evaluate them, we pay them, and then we run the loop again.

That’s what Bitcoin is. People try to mine Bitcoin, and if they can mine it faster, they get paid money. We took that idea and abstracted it. What if we want to do that for everything? What if we want to build a mining network for storing files, procuring compute, and training machine-learning models?

This is where it gets interesting, because those last two have already been done in crypto. Golem, Filecoin, and so on are all basically the same idea: let’s build mining for GPUs, or let’s build mining for storage.

But they only do one thing.

They only do one thing. Because we focused on that recursive incentive loop—you could call it an adaptation loop—that’s in AI, we abstracted the monetary adaptation loop and said, “Go nuts. People can build a bunch of these things.”

We have this abstract version of it, so we can build all of those things at once. We made enemies with the Render community because I randomly tweeted, “Render is just a subnet on Bittensor. Filecoin is just a subnet on Bittensor.” We made a lot of enemies because it was true.

All of those things are just instances of the same thing: a monetary feedback loop where you have people do things, measure their performance, and pay them money. You can design that game in a whole bunch of different ways.

We tried to build a framework for representing these systems based on machine learning. We used weight matrices, softmax, and all these things, but it doesn’t really matter. The point is that we built this abstract tooling called human consensus that allows us to do it in a trustless manner.

How do we build a mining network for anything, in an abstract sense, that can be trustworthy and distributed? That’s really difficult, because Bitcoin is actually very easy. A Bitcoin transaction is either cryptographically verifiable or it isn’t.

Right? So the truth is—

It’s just one or the other: a 0/1 outcome. But when you get into something like the subnet I’m building right now, the question is: how do you make it so that all these really smart people who are mining Bitcoin will actually go out there and try to train machine-learning models?

The performance of a machine-learning model is really difficult to measure. It takes a lot of time, and there’s a lot of randomness in the evaluation of a machine-learning model. This is why we have benchmarks in AI.

Those benchmarks are standard, so we can take a benchmark and turn it into an incentive mechanism. We can have people train machine-learning models that are good at the benchmark. At the core level, that’s what Bittensor is. That’s what a subnet is: a monetary feedback loop that we can apply to different things.

The examples on Bittensor include applying it to training machine-learning models across the web. You run a node, and we measure how many FLOPs of training you’ve done. That was what Templar did.

Okay, okay, okay.

Follow me.

No, I know. Please keep going. Then I’ll try to reframe it. Keep going.

We can also measure the speed at which someone can inference a machine-learning model. That’s very valuable when you talk about OpenRouter or Chutes, which is the subnet, or something like OpenAI.

It’s a really difficult problem: how do you do that quickly? How do you take an input and give an output really quickly? That’s called inferencing. We can say, “We’ll pay you for the speed at which you can inference,” or for the bandwidth at which you can inference.

Then we can turn that into a product, because now we have people out there not just securing the network in Bitcoin—which is useless in some sense—but actually solving this whole monetary problem of mining.

Why is Bitcoin mining useless?

I’m going to get shit for that take.

I wanted the clip. You just gave me the god intro. I had to follow up.

It’s not useless, because obviously it secures the Bitcoin network. It has infinite value, perhaps. But there are other things in life that have more tangible value, like storing files. How many files can you store?

Those are not very interesting in some sense, but they get more interesting when you apply this idea of feedback loops to things that are much more abstract. For example, don’t provide me compute, because that’s obvious. Anybody can provide me compute.

I want you to develop software that works well against a well-defined benchmark. An example of that was Ridges, a platform on Bittensor. They build coding agents, and those coding agents are evaluated based on how well they can solve coding problems.

Got it.

We can not only mine raw commodities, like GPU speed or how many FLOPs you can provide; we can also mine software by taking this primitive and applying it to actual software development.

This is where I started from AutoResearch. I said, “This is the target of the day. Karpathy’s genius: autonomous agents that just change code.” We measure how well the agents did at changing the code, and then we loop on that.

Eventually, you have this process of annealing the software toward a better and better version of itself. We basically do that for all these problems. One of them is coding challenges, and that’s also just the mining of software.

Let me finish the last part. That’s the subnet level, and then TAO is a market of those subnets.

Okay. The way my brain usually works is that I bring it all in for three days, and then at the end I spit out a finished product. But the way I’m understanding Bittensor right now is that TAO emissions are basically a reward mechanism to incentivize improvement or success across defined metrics.

You would say, “I want this thing to be done very well,” or, “I want this parameter to be met,” and you’re rewarding a subnet, as I understand it, to create and manifest that thing in the world well. The better you do, the more emissions you receive. Is that correct?

So you just have 128 subnets that are all being incentivized with monetary TAO to build X, Y, Z, A, B, and C?

More or less. Let’s go through the stages of Bittensor so people can understand it through its history, because it has changed a lot.

We didn’t do a pre-mine, and that was really good because it gave us a lot of time to develop the technology. Back in 2021, when we launched, there was just one subnet on Bittensor.

Wow.

It didn’t work very well, it didn’t produce anything, and we didn’t really understand it yet.

We didn't really understand what we were dealing with. It failed and it got hacked.

What was it, and why did they create subnet number 1?

We created it. Basically, the original version of subnet 1 was actually quite brilliant. Today, maybe we could do it properly; back then, it was far out of reach, but it was a cool idea.

The miners in the network produced what’s called an embedding, which is basically just a tensor. A tensor is a long list of numbers that represent text—or basically anything—but in this case, text in numerical form. That numerical form is essentially at the center of a neural network: you put text in, you get text out, but in the center, it’s all numeric. That’s called an embedding.

It passes embeddings through the layers of the network. We said, “Hey, let’s build a mining network. Let’s build a Bitcoin mining network for information-theoretic value”—which is a mouthful. Basically, how valuable is the vector that you can produce if we’re trying to solve problems in language?

This was before ChatGPT. For the AI people in the crowd, this was BERT and RoBERTa—old-school transformer stuff. It turns out that it’s possible to do, but we were really not prepared to do it properly.

So we switched to something else. We built another network based on querying the miners: they would answer, and then we would use another machine-learning model to evaluate how good the answers were.

Can I ask you a quick intermission? What benefit would subnet number 1 being good contribute to Bittensor?

Yeah.

The top layer.

If I were doing it now—the industry of AI moves so fast, right?—one of the benefits is that, because we had this abstracted layer, we didn’t build exactly what we were doing into the substrate of the chain. We could then switch to something else. Thank God, because if we had built it in Rust on Substrate, we would have been screwed.

Yeah, you can’t.

At the time, we were working on these embeddings. If I were to do it now—

Yeah.

—I would frame it as producing context embeddings. It’s still a useful field right now. If you look at Pinecone and other companies in the industry, OpenAI serves embeddings, right?

If you want to take your whole codebase and put it inside what’s called an embedding database, which you can then search quickly with LLMs, you embed the data. That means you send the data, and then you get these embeddings. They encode dense information about what was given to you, but they’re also a compression. They’re basically a numerical compression.

If I were to take that subnet and rebuild it—which I’m not doing, so I guess I wouldn’t—I would still train it to produce embeddings very quickly and with incredibly dense information, because that would be valuable to things like Pinecone if you’re making an embedding database.

The way that Bittensor subnets work is that, in order to gain access to these subnets, you need to have a fair amount of stake in them, right? That’s what creates the economic game. The token represents access in some sense. The people who have the token are the people who can gatekeep the network, and therefore they can monetize it.

If you were to try to build a subnet around that—which maybe somebody will do—you would build a website and an API service for people to query.

Try to make money. And so now—

—that subnet would have its own token paired with a Uniswap pool.

Got it. And that’s what creates the demand side for TAO. Basically, it’s like a flywheel.

Got it, got it. So the main benefit of the success of subnets is that they’re paired with TAO: more demand for TAO, a higher TAO price, and more value to compete for future emissions. There are 128 subnets—is that true? Did I make that up?

There are 128.

Okay. Every subnet exists in a silo and is doing its own thing, completely unrelated by default to what anyone else is doing, and everyone is basically competing for increasingly valuable TAO emissions, in theory.

Yes.

Is there any pairing process where everyone’s contributing to Bittensor—the top layer—or is it strictly a financial faucet?

This integration aspect is really interesting. We didn’t build the mechanism to require integration. I build this embedding thing—does it help this guy?

It seems to be epiphenomenal that people work together. It’s a consequence of the system itself, like the ecology of Bittensor. Everyone builds offshoots because an offshoot provides a really good service for producing mechanisms. If somebody solved this really hard problem of verifying something, I can now use that as a subunit for my thing that does verification at a higher level.

Then he’s also my friend because he’s backed by TAO, and we’re in the same ecosystem. We’re all bedfellows here. That tends to work out in practice, but there’s no actual mechanism in Bittensor that enforces it.

We had one subnet, then we had two subnets, then we had three subnets. They were all built by the foundation. Then we opened it up so other people could build subnets.

We tried to manage it with an oligarchy and people voting. This didn’t work, and we always knew it wasn’t going to work. About 2 months into doing that, we said, “Let’s switch to dTAO.”

Okay.

Dynamic TAO was basically the idea that we could build an actual market for the tokens and use market dynamics. People are going to buy the subnets that make economic sense, and they’re not going to buy the subnets that don’t make economic sense. They’re going to buy the subnets that are aligned with Bittensor’s future because they hold TAO, and they have to buy them with TAO.

We took all the subnets, gave them a token, and paired them with TAO. Then we created a mechanism to measure how much TAO is going in and out of the pool. That incentivizes the subnet not only to get more buy pressure but also ultimately to constrain TAO.

We’re minting 7,200 TAO per day, but it’s all being put into the subnets that are most constraining of TAO. If your subnet can’t keep it together and is just leaking TAO, ultimately, you don’t get any more emissions.

We have this structure in Bittensor where the system itself directs the emissions—the inflation of TAO itself—to the places where they can constrain it, where they can keep it locked in the pools.

Got it. What happens? We’re still pretty early in the process. We’re like a third of the way through emissions—where are we, a third?

Around 11 million, so we’re just over halfway.

What happens at the end? What happens when we get close to 21 million? How does the network of TAO evolve when we’ve basically reached the final form of emissions?

Yeah, it’s a good question. There’s a similar question about Bitcoin: how does Bitcoin work?

I was going to ask you that one next, honestly. What do you think?

I’ve always believed that you’re basically building a nation-state with a cryptocurrency. Pump-and-dump schemes don’t make any sense for crypto because they don’t have any time. This is why you shouldn’t sell out to VCs with huge supplies. This is why you shouldn’t go and try to get market makers. This is why you don’t need to get on exchanges.

You should be in this for the long game because it’s going to take time to develop the ecosystem, the mechanisms, the liquidity, and the miners. These things are organic, and they need to grow.

I’ll answer the second question first. Bitcoin’s idea is that after 104 years—that’s how long they’re thinking—this technology, which they made and are pretty sure they made perfectly, is going to be foundational to the world economy.

Yes.

It’s going to maintain control over it plus transaction fees. I actually think that control over it is going to be more important than transaction fees in the future. Bitcoin will be worth enough money to sustain itself past the emission schedule.

The idea is that Bitcoin isn’t even there yet. Bitcoin is just being birthed.

It’s only 15 years into being birthed. It needs time. It’s only really truly galloping after 104 years or something like that.

It’s kind of the same thing with Bittensor. Obviously, we want to go a bit faster.

The first question is about TAO. Once TAO emissions go to zero, the subnets should be incubated, because that’s not going to happen for 50 or 60 years—a long time, maybe 104 years as well. At that point, the subnets will have functioning markets.

They’re fully paired with TAO, right?

It’s the United States dollar of these crypto projects, right? Everyone’s largest liquidity pools are with TAO. Basically, TAO survives, but the question is: what happens with the subnet tokens? They also have 21 million, but they need to pay their miners.

The idea is that if you get it to that point—which not many are going to reach, but after decades, or not even decades, really, I could see this happening with some subnets, and much more in the next 10 years—you start to not need the subsidy. You pay the miners directly in the token because, right now, they’re on a subsidy.

Yep.

Is it like—

—printing to pay to keep them alive. But eventually, the system is mature, in the same way that Bitcoin is mature, and you pay them in the token to continue doing what they’re doing. People who want to access the network need to pay, right? There’s no subsidy anymore. That’s basically how I see Bitcoin working, and also how I think Bittensor will work.

You’re a full-blown, full-fledged Bitcoin maxi as well, huh? I love that.

Yeah.

I love that. How can you not? How can you not?

It’s so disrespectful. I don’t know. Maybe I was just raised differently, but it’s like walking into your grandfather’s house and fucking clocking him in the face, saying, “I’m faster than you, Grandpa.” You know what I mean? It’s like, “Yeah, you piece of shit. You loser. I’m World War II for you.” You know what I mean?

I hold the wisdom of the whole industry. Ethereum can be captured. Ethereum has a founder. He’s a guy with ideas. They can take him out, but they’ll never take out Bitcoin. It was built to be immutable. As a consequence—and I think people take this for granted—governments are just forced to deal with it.

There were experiments with Bitcoin-style things, right? D-money or B-money and Bit Gold, right? They were destroyed because they were centralized, and that was what governments usually did with people who tried to create currencies. But Bitcoin, by its nature, because it’s this massive shield and it is impenetrable, is impossible to take out without mining it and overtaking it.

It’s opened up the space for us to explore and work with the power of money—the language of money. No one in history was ever allowed to actually write with money except for the elite and the bankers, right? Now, as an individual, you can use this language that’s like a coding language.

Thank God for Bitcoin. If Bitcoin goes, they’re coming for you. Why would I want to be better than Bitcoin? That’s bullshit.

It’s also kind of like we’re obviously building on top of it, in the same way that I made an analogy in my post. Telegram doesn’t say, “Fuck you, TCP.” They use TCP. They’re built on TCP. It makes no sense.

I’ve always found it really embarrassing when people come out and love that kind of tweet—“ETH is bigger than Bitcoin.” I’m like, humble yourself a little bit.

Okay, wait. I had about 4 things I was going to ask you, but you’ve thrown me into this last question. Jason Calakanis is talking about TAO over BTC, which has to be a weird feeling because, on one hand, it’s probably everything you ever wanted: all these eyeballs, attention, and capital on top of Bittensor. But on the other hand, based on what you just said, he walked into your house and punched your grandfather.

That’s why I responded. But it’s an open network. It’s permissionless. You can make it your own. It’s not just about—

More power to you. That’s what you think.

But it creates drama, and I think he’s doing that on purpose. He’s a guy who likes attention, and that’s a great way to get attention. But it’s not what I believe. I think my opinion matters more than his, probably, in the end.

That was good. I might have just missed this, but can you explain what makes Bittensor inherently AI- and artificial-intelligence-adjacent?

At least in layman’s terms, I would understand Bittensor as something where you could utilize emissions to incentivize anything. The subnets could essentially be anything. Why is it AI, or decentralized AI? What does AI tie into with Bittensor?

I’ve always had a problem with this because it’s like: is it crypto applied to AI, or is it AI applied to crypto? Are they actually one and the same thing?

I like to sit in that space because the truth is that we took ideas from artificial intelligence—the structure of feedback loops and how to think about it like a machine-learning engineer—and then we took the ideas of crypto and put them together.

This new thing in the center is neither crypto nor AI. It’s monetary computing, you could say, or incentive computing.

Computing.

In the technology, it’s AI crypto, and it’s crypto AI. AI itself is just a bunch of tools, but it applies to everything. You use AI for robotics, farming, language, statistics, computer science, and business. You apply this tool to all things.

In the same way, you can apply what we built because it’s abstract, really powerful tooling for solving problems. It has very similar artificial-intelligence qualities, and therefore we can apply it to all these things, including AI. Why? Because it’s the hottest thing in tech right now, and it’s also the fastest-moving thing.

Obviously, we’re going to apply it to AI because there’s so much demand for commodities and things like inference.

You can take it to everything else.

Crypto invented, for the first time in human history, the ability to write with money and to build AI systems with money. Therefore, crypto is part and parcel of what we built.

But, yeah, it’s neither.

That was honestly a beautiful answer. I guess my follow-up is: the way you see it, if this open network played out exactly as you could imagine it in your grand vision, what does that look like? What value would be created from Bittensor that you would see as a success case?

I know it’s a brutal question, but what direction do you want this thing to head?

We’re beginning to see the beginnings of subnet potential that’s like fully self-contained, functioning economic systems, right? You can bring compute, you can provide it in this permissionless market, and you can make a little bit of an edge with the token. That makes sense, and that’s exciting.

We can have these permissionless markets for all sorts of different things, which is valuable and lowers the price. That’s cool.

But that’s not really that compelling—the finance angle, and just—

Oh, let’s make things cheaper and more efficient.

We spend a lot of time doing that, obviously, because it doesn’t matter if you have this grand vision if it doesn’t work. It works, right? Thousands, millions of people use Bittensor for inference, compute, and things like this. That’s fantastic, and I’m really happy that we can provide that.

But there’s another angle to this, too. We’re not just trying to be fast and efficient. We’re not just trying to be Bitcoin mining applied to things for the good of Bitcoin mining.

We also believe that when you build systems like this—when you build a neutral technology like Bitcoin that many people can hold, own, contribute to, and build on top of, and where they can transparently see what it’s built of and critique its structure inside and out, the way you can with decentralized technologies—that is the bosom in which we should be developing artificial intelligence.

That is the womb for properly aligned AI, not a centralized lab where you can’t tell who owns it or what their interests are. You can never work for it, and you’re always just going to be suckling off the teat of this AI all-seeing Borg thing, right?

We don’t know if it’s Epstein aligned or not, but you would never know because it’s not transparent and you can’t own it. It’s a closed realm, and you can never contribute to it. You can’t go there. They don’t want to hire you.

You don’t know how it works. That is a recipe for a really disastrous outcome for this machine god that we’re building. We’re building this machine god. Is this machine god going to be top-down and totalitarian?

I think the real risk is basically Elysium. Have you seen the movie Elysium?

No. I have to go watch it. What is it?

It’s a sci-fi movie with Matt Damon.

Basically, it’s this future world where the whole world looks kind of like Gaza.

Right. It’s just destroyed.

And the elites live in a space station where they have access to better medical technology and robots.

Okay.

This is the vision for the Silicon Valley elite, right? They don’t have the tools for anything else, so they just imagine that they want to be the people who

own everything.

Yeah. Pick the ladder up behind you, right?

Yeah. There’s no agency for the rest of us in their vision for the future. Elon Musk will have a robot army to himself, he’ll be a multitrillionaire, and you will not participate. You will own nothing and be happy, or whatever.

Yeah.

We can just build it differently. There are 2 paths that people want: one is centralized labs, and the other is government oversight. Government oversight is not going to work because it’s really inefficient, and the centralized labs are winning because they’re faster. But there’s no openness, there’s no shared control, and you can never participate.

Anybody who’s watching this thread can participate in Bittensor and become important in this ecosystem. There’s no way you can do that with a centralized lab. Give us your humble masses. You need to be free.

So the unfortunate story with AI, or with crypto at least as it relates to AI, is that in Silicon Valley artificial intelligence, they build this really cool thing and say, “Hey, crypto kids, go play with the [__] light version.” Peter Steinberger builds OpenClaw, and then we’re trading OpenClaw memecoins on [__] Solana, right?

The question is: Bittensor, you could participate in building this open and decentralized AI future, but can you actually compete at a serious, meaningful level? What does that look like? How can you actually compete? You can participate, [__] it, but can you compete?

Right. Well, take Chutes. This is one of the top subnets of Bittensor. They built this perfect market for compute where the miners bring the compute, you inference on it, and you get the outputs. That runs OpenClaw, as an example, or you can build a real agent or code with it, as an example.

They were getting the fastest tokens per second on OpenRouter across a whole bunch of open-source models, at the cheapest price. Not just the cheapest price because they lowered the price, but economically, they were able to get the lowest price when they looked at the cost per H200 hour.

Why was that? Let’s say you want to become an inference provider with somebody who has a bunch of compute, and you play this game. You can’t actually compete against this liquidware economic system where, at any moment in time, if you’re not as good as the next guy, you just get kicked from the network. The better guy who can run infrastructure and do it for cheaper comes in.

You build this inference-providing company, and then you want to rest on your laurels and take a bit for yourself. The market takes a long time to cut that inefficiency away. We do it instantaneously. We do it every hour. Every hour, we’re remeasuring and replacing, remeasuring and replacing.

This hyper-annealing toward efficiency is how we’re able to get compute for way cheaper and inference very fast, if we can measure that properly. We don’t just talk a big game and do nothing. At one point, we were serving the largest number of open-source model tokens on OpenRouter for a whole 6-month period. It was just Chutes.

We were doing that through this competitive game that allowed us to do it really cheaply and efficiently. That’s just one narrative, one narrow band. But then you apply that to everything else, like building coding agents that are really good at solving problems.

I want to stress here because it’s really difficult to do this. It’s a much harder problem to do what we’re doing. But when we get it right, it just goes way faster, and it’s way more efficient. We can compete. We can actually always win if we can define the problem better than them, because we can consume theirs, put it into ours, and beat it. It just goes faster.

Corporations tend to ossify and die because people don’t want to work. But when you’re [__] the miners on Bittensor, they have to work, man. And they do.

We’ve proved it time and time again. There was a good period in 2024 when we were crushing benchmarks everywhere. We were crushing benchmarks for fastest search, fastest embeddings, fastest AI detection, best 3D generation—best, best, best, best, best. Every single one has something that they’re the best at.

You go and become the best at it, and then you try to monetize it. Sometimes it doesn’t work. But we’re probably one of the only cryptos that can say we actually do things that are state-of-the-art.

I think we’ll continue to do that. We have the largest decentralized training run in the world, the largest decentralized model ever run, and the best algorithms for it. The miners are doing it most efficiently and effectively, and I think that will just continue in that direction.

We honestly just need time because nothing’s going to happen fast.

How does that acceleration curve work once it really starts working? How violent is that?

The thing about incentive mechanisms is that they’re very hard because they’re permissionless. You don’t control everything. There are so many ways to exploit them and so many ways to cheat, so you have to get very precise about how you define the metric of success in a way that people can’t just copy the answer, as an example.

But when you do, it’s kind of like aiming this massive laser cannon. You finally get it in line, and then boom, it explodes something. It has this quality where zero-to-one is always zero. It’s not working at all, and then it’s working in an amazing way.

That’s often what happens when you’re looking at subnets. They may have a really good idea, but the incentives might not be aligned properly. If they can get the incentives aligned, they’ll have success very quickly in whichever domain they pick.

What are you the most excited about right now? You’re working on a subnet on Bittensor. It’s not fair to say you’re getting a lot of attention right now because, honestly, you’ve had it for a while. You’re kind of a [__] Chad. I was going to glaze you when you got off, but you could clearly work out whatever you wanted, I’m sure.

What are you most excited about trying to solve with Bittensor and in the world? I don’t know.

Good question. I have a lot of projects going. One of them that I think we need to solve, and that we’re trying to solve in Bittensor—and I think we will solve—is human identity in an open-source way.

Rather than the orb, let’s build open-source software that can do human detection and uniqueness as well. We can build identity systems that are open source. I think that’s really cool.

Why did Worldcoin fail? Not to throw you off.

Because everyone hates them. You can’t get funding from Sam Altman, the guy who’s trying to enslave us, and then pretend to be freedom technology. What are you trying to do? It just seemed odd.

The orb was so dystopian.

Yeah. Brutal.

Why would you do it like that? The technology is actually really good, though. The guys were really smart.

It works, right? It’s a great technology. I always tweet this whenever some weird [__] happens: there’s a parallel universe where Worldcoin is like flip-flop Bitcoin, you know.

Yeah. Isn’t that a testament to this industry? People hate crypto people—

So bad.

—for good reason sometimes. But deep down, we’re freedom fighters, and that’s why Worldcoin didn’t work: because we’re permeated by people who actually believe in freedom.

You know what I mean? That is actually who we are, and that's not AI, by the way. The AI people who think they're so hot and high and mighty—they're all just normies. They're all blue church people. I love them to death. They're all very smart, but they're not freedom people. They don't really grasp geopolitics. They don't really understand it, and they don't really care. They're just going to go work for the Borg, make their money, and live and play their board games. But the crypto people are like, “We're renegades, man.” And so you can't try to do Worldcoin on us. It's not going to work.

The human thing is a big thing. Taking decentralized training to a trillion parameters is really the end goal for Templar. What I'm trying to do is basically stitch all of the subnets together in an agentic way. You've seen what's happening with auto-research and what people can do with their OpenClaws. Bittensor is moving into a new era where it's the perfect ecosystem for agents because you can play every side of it.

A machine-learning model is now at the point where an AI can mine Bittensor, build subnets, trade them, and buy the commodities it needs directly from the chain. It all starts to really integrate as an ecosystem with agents. You say, “Here's a TAO wallet. Go run some compute, get some inference, store some files, go mine Bittensor, and make money to fund the whole thing, and then make a subnet to do what you want.” Crypto is the perfect playground for the agentic world.

What I'm excited about is doing that myself. I can now just run 10 companies with AI agents. That's what I'm excited about—using it to speed up Bittensor and build on Bittensor.

That was beautiful. I know it's late for you. I kind of want to ask you a sign-off question here. I think we hit most of the stuff. I want to ask you a crypto question. I wasn't expecting you to be so cypherpunk and have such great takes on crypto. So you're really optimistic about crypto?

I'm really optimistic. Crypto changed my life. I would not be here without crypto. I came in through NFTs, so I have a different origin story. I came in to trade and stayed because it's awesome here. But you're building and working on the future, and I'm trading, and most of what's here is slop and garbage. When you trade a lot of slop and garbage, you tend to get pretty cynical.

Taking into account where we are right now, where the market is, and where AI has gone, it feels like a lot of talent is fleeing and running away. We get the light version. I think a lot of the trading side—I’m not even sure how much you interface with this cohort—is very cynical, very jaded, and questioning whether it was all just a vehicle for BlackRock tokenization, whether Bitcoin is an Epstein pedocoin, and whether we all just die and lose.

In 2026, you were here in 2015 with Bitcoin and the ethos of cypherpunk freedom fighters. What keeps you so obsessed and passionate about crypto? You're a dying breed, if you will.

I've been in the industry since 2014 or 2015, and I've seen a number of bull runs. People get so exuberant, and then all these frat boys show up and ruin it. Then they lose a bunch of money, become disenchanted, and it's a great breath of fresh air. Then it comes back. Obviously, there are cycles.

What keeps me going is what kept me going from the very beginning: Bitcoin's vision of a fair monetary future. One where governments are tethered to a system they can't manipulate through war. What are we seeing right now in the world? We're seeing people trying to fight for the dollar's reserve-currency status. The number of people who have died because the only way to win that game is through violence is staggering.

Bitcoin comes along and says, “Hey, we can build it so that they can't take that away from you. You'll always be able to have this digital currency. You'll never be able to inflate it away, and they can't steal it from you. You can store it in your mind and transfer it across the globe.” That never died for me.

People always speculate, asking, “Was Epstein involved?” Who cares? If you're smart enough to read the code, you can see how it works. It doesn't say anything about Epstein in there. It doesn't matter if Epstein himself made Bitcoin. Bitcoin is inherently good. It's inherently a freedom technology. It doesn't matter who made it.

Most people can't understand it, so they're like, “A bad person was involved; therefore, it must be evil.” I don't think Jeffrey Epstein was involved in any way, but even if he was, that's what I'm saying. That vision has never really changed for me.

Bitcoin enabled this whole future where we can play with money, and Ethereum invented the ability to build smart contracts—a classic financial idea—in a decentralized and trustless way. That still exists and is still going to be the place where a lot of the digital web, the Web3 world, is living. I'm a big Ethereum fan as well.

I also think that we're eventually going to draw the attention of real researchers who see what we're doing at Bittensor and understand that this is real technology. You can compute with money. None of that has changed for me.

If you got pumped off Pump.fun, then I guess that's not really it. Maybe it is. But if you were just enthralled with it dumping, I don't know. I think you were setting yourself up if you thought it was going to go to a billion or whatever. What do you think?

Dude, I think you're incredible. This is why I love crypto, and I've been spending so much time talking about oil and everything else that this was a breath of fresh air.

You were the first person to contribute to the Bittensor white paper in 2015. How many people actively work on Bittensor versus the subnets that are separate from it? How many people actively work on Bittensor?

Not that many people, actually.

Do you actively work on it?

Yeah.

Okay.

Yeah, totally. I stepped down from the foundation as the CEO because I wanted to just participate in the network. Being this central figure means you can't have any fun. I can't mine because people see it as a conflict of interest and things like that.

I've stepped down, but I'm always contributing. People know my opinion.

Dude, this was awesome, man. I really appreciate your time and you coming on. I know it's late in Europe. Is there anything you wanted to sign off with, or anything I was really supposed to ask you that I didn't ask?

I watched you talk to Sammy, and I got a lot of invites onto podcasts today, but I wanted to speak with you because I really like your show and your vibe. I think you're a dope person.

Yeah. Oh, you saw the Sammy interview. That was awesome. Someone said to me that we did it in November, which was funny. It was a while ago. That's awesome, man. I appreciate you doing it.

Yeah, that was sweet. I just think you have a great vibe, man. Let's keep crypto going because we need to.

Sick. I'm going to DM you some stuff when we get off, just for when questions come up, because I think you're sick. Is there anywhere you want to direct people? Any links you want to share? Your Substack and Twitter are linked, obviously. Is there anywhere you want to send people?

The Bittensor Discord. Go try to mine a subnet. It's really fun.

Sick. By the way, am I pronouncing it right?

Yeah. Yeah, Const.

Const. Dude, absolute pleasure, man. Thank you for coming on. Thank you for your time. I'm sure we'll talk offline. This was a lot of fun, dude.

Awesome, dude.

All right, have a good night. Peace.