Bittensor creator Const on Affine, dTAO, "mining reasoning," and more | E2326

24:31But not just faithfully, perhaps we're also going to check to see if you're doing it fast and faster than the last guy. And what happens in all of these networks is that people joining, there's a continuous role of people joining into the network, and they're measured based on what the network is measuring. in this particular design. An example would be, I just described what's called inference speed, right? How fast you can inference a machine learning model. It's measuring how quickly you can answer these questions from the clients. And it's paying you more and more based on how quickly you can answer these questions and if you're meeting some sort of bar, right?

25:06So think of it like a very well-defined, written-to-code description of how we're going to check to see if the computer that you added to the network is faithfully following the rules and doing the thing and an axis along which you can perform better right so perhaps you can combine and add more computers to your cluster or um you can speed them up or you can improve the software yourself and make it faster at serving these inferences and if you can do that you can you can serve a request and over time you can make more money in this network and so each of these networks are usually the the way that we visualize them is is like a curve and so along the x-axis are all of the different participants.

25:46Usually it's about 200, it's 256 of them. There's 256 computers, which is often more than enough and I'll explain why if you're interested later. And then there's the amount that they're getting paid, which tends to go up, well, it goes up to the top. And then there's the ones that are being cycled out at the bottom. And think of it like a league, like the Premier League is a good example. I love it. So each of the networks on BitTensor has this relegation system, this Premier League, of the thing that the miners, the contributors to the network is contributing. So some of them may be inference.

26:21A good example, like a sub-sample of those projects would be you bring GPUs to the network that people can use. You provide computing power that people, you inference machine learning models. So when people are talking to your endpoint that you respond. Or it's an example where you actually train a machine learning model. You produce an AI and contribute the AI itself to the network and get paid if your AI is better than the other AIs in the network. That's what we would call a model competition. So that's a little subset. That would be like Liam, Engie, and Affi with some of the top subnets on Binted.

27:00So that's what they do. The whole premise of one of these projects is that they're able to use this permissionless, hyper-competitive market to produce this digital commodity, inference, compute, or models faster and better than anyone else in the world because we're using the power of Bitcoin. And if that's not their premise, it doesn't make much sense to build on BitTensor. And there then that network has its own token. So think of it like a subtoken to TAL, which is like a derivative token we call a staking token the the name we use is an alpha token um which is a a secondary cryptocurrency which also has 21 million cap that only exists with inside of that network that we just described you need to pay it to enter you get paid in that by doing well um and it's if you sell it you sell it into tau that is what a subnet looks like and that That Komopti, what we call the alpha token, that token that represents that network, it only has value if at the end of the day, people are going to want to buy that token to get access to the computers in this network.

28:09So everyone is evaluating these subnets based on whether or not they actually have some sort of long-term potential for producing value. If it's just, oh, it's a network where you can just join and if you just break random numbers, no one's going to buy your token because there's no way to thread any relationship between that thing having value.

28:28Lon Harris:In startup speak, for me, that would be a product with product market fit. So for me, using energy or ENGY, for me, that was, wait a second, these tokens for Quen, Kimmy, and GLM52 are half the price of other places. And it's like, well, wait, how come those are half the price? And now I'm looking at it, you know, using OpenRouter, Claude, Fable. And hey, maybe I just plug in my API key for energy and I use that in my Hermes agent. Right. Right. And behind that product is this liquid swarm of anybody in the world that can enter and try to reduce that price over time. So right now it's half the price, but it doesn't necessarily need to stop there.

29:18Because whenever there's excess compute, that person can plug themselves into this permissionless network and it becomes quite liquid, right? So, hey, I'm not using these computers anymore. Well, I just run the software and then I just sell the inference to NG. And as a consequence, we can dramatically reduce the cost of that commodity.

29:36Lon Harris:And this dovetails with your previous statement about permissionless and anybody can join to make a little bit of extra tau or whatever the subnets token is. So if I'm sitting there and I was, I don't know, providing servers to startups and other folks, and I happen to have a rack that I haven't provisioned yet, and it's going to be provisioned in 100 days. I have a client who's coming online in 100 days. What do I do with that for 100 days? Well, I could give it to Engie without asking anybody permission and start making money from it. So the downtime would turn into productive time, would turn into revenue generation time.

30:17Lon Harris:Precisely. If you've got a company, you probably have a tool for invoicing and one for running your website. And then finally, you got one for your CRM, of course. And somehow, even with all these time savers, you're still moving data around by yourself at 10 p.m. That's why here at Twist, we recommend Odoo, O-D-O-O, the all-in-one management software that's already serving 16 million users across more than 170 ,000 companies. Think about that. They're bringing everything you need into one platform. That means your CRM, your sales tools, accounting, manufacturing, your website, inventory, and of course, your point of sale, if you have a point of sale, all right where you need them.

30:52Lon Harris:And they're all in constant communication with each other. No more logging every change order or new sale across three different apps and the dreaded spreadsheet as a database. No, you make a sale and your invoice gets created instantly while your inventory gets quickly updated. So if you're still combling together your back office across five different tools and apps, get started today at odoo.com slash twist. and your first app is free. That's O-D-O-O dot com slash twist. It took quite a long time. It's important to put an asterisk here. It took quite a long time for us to figure out how to build these mechanisms.

31:26Just like how the first smart contracts on Ethereum, like the DAO, didn't work. And you had CryptoKitties, and that was cute. But now you have production-level stuff and all of these side chains.

31:39Lon Harris:How long has Tao been around, and how long did it take you to get to So people have been building mechanisms on BitTensor for two years. So this is the second year that it's been open for people to build mechanisms on BitTensor, but actually we're 2021. So the network launched with us learning the art of what we call incentive mechanisms. We call subnets the core design, the structure of what a subnet is, which I can talk about at length. It's actually quite an interesting field of the study, the economic study of how do you build a permissionless adversarial game where even the worst person in the world, if Hitler and the devil were to have a child and that child were to mine on your subnet, they would just produce value.

32:26Lon Harris:you know um so you can take away intent but you need to have some validator that says hey this hitler idi amin putting servers on here is not doing so to kill people uh or to do harm in the world for being a little bit uh uh you know dramatic cheeky here yeah but they are being judged on a specific criteria which is does this server provide inference and provide kimmy or GLM-52. And there's something there with these validators that do this, yeah? In the subnets? Yeah, so we built a mechanism that would allow a distributed set of computers that could take any code that you wrote, Jason, for describing whether or not the computer that joined the network is actually, if Hitler and Satan's love child was actually doing inference properly, you wrote code that would check that.

33:26And these computers, which we call validators, you can elicit them to all run the code at the same time. And if they do so, they will reach consensus if more than 50 % of them are running the right code. It's very similar to the way in which Bitcoin works. If more than 50 % of the network is running the correct code of Bitcoin, then it doesn't matter if 49 % of them are cheating. It's irrelevant. the person, the majority, the honest majority will determine the direction of the incentives in the network. So the core mechanism of BitTensor is that we can build mechanisms like this so that, yes, we can build a network that's permissionless and verifiable and auditable, right?

34:11If you know that this distributed set is all running this code and it's going to take more than 50 % of them and the decentralization of that set, which is measured in proof of stake, It's not like Bitcoin. Bitcoin is proof of warp, which means that the weight of a node is based on how much compute they provide. In Ethereum, it's proof of stake, which means that the weight of a node in consensus is based on how much economic value they have. It's the same thing in BitTensor.

34:34Lon Harris:In this case, if you were, you know, NGL, I'll go to that one because I actually use it. If somebody got on the network and said, yeah, I'm providing GLM-5-2, you know, this open source model from XAI in China. But they were actually using like some old DeepSeq 3 and they were passing it off as GLM-5.2 to use less compute and they were giving wrong inference. The validator would say, uh-uh-uh, we're checking that you're actually using the right GLM-5.2 code, not 5.1, not 4.0, not some other hack. Not quantized. Yeah. Yeah. And we are going to make sure in a way that there's an SLA, a service level agreement here, which you might have with Google or Amazon Web Services.

35:21Lon Harris:In a way, you've smart contracted or built into the system an SLA. The service level is mechanically and architecturally built into the system. Am I understanding it correctly in layman's terms? Yeah. And instead of a contractual SLA, it's a programmatic SLA, right? So you define the way in which you can check to see if they're doing the right thing. So for an inference subnet, it's actually very difficult, but the technology is there. The way you do it is you query the endpoint, the endpoint gets a result, and they have to pair that result with a very cheap, what's called a ZK proof. In case of NG53, it's Topolock, which is an algorithm created by Prime Intellect, which basically hashes the hidden states of the model.

36:20Hidden states is like halfway through the machine learning model. They take those, basically intermediate representation, and then they project them onto a different mathematical space. And then you don't need to know too much more about it. But anyways, that's what - Proving it is what you say it is, is a way to say it. Yeah, I'm planning. Precisely. And on BitTensor, these things get like, it's one thing to write a paper. It's a completely other dimension to write on BitTensor because we have the smartest people in the world and we have the smartest hackers in the world and they're just right next to each other and they're going to break your system.

36:57And if you can build a system that on launch doesn't break, you know, it's like, oh my God. Wow. It's unbelievable because -

37:06Lon Harris:So this is critically important, Const. If it's permissionless, you can have bad actors, and then the bad actors can be forced to act well and be good actors by the architecture, but you're constantly being stress tested. You're constantly being attacked because it's permissionless and global. You have attackers who say, hey, there's something at stake here. I can get tokens that are worth something. So there is a value in hacking the system. Therefore, the system must be architected properly. And that's a big part of your job and the BitTensor Foundation's job is to make sure the rules and the architecture and in a way you're the police officers or something.

37:50Lon Harris:You're the Jedi Knights of this system, making sure there's peace and that it's being done properly. This is the role that we play reluctantly while it's still required. As all subnet owners do themselves. So at the level of the individual subnet, they build a mechanism and it breaks and then they fix it. And then it breaks and then it fix it and then they get it right and then it starts working and then they make a million bucks. But in that time, it often requires a lot of massaging and realignment to fight against effectively nation-state militias. There's a level of the people that are trying to destroy your project.

38:36And for a lot of people starting a subnet on BitTensor, they don't really understand that that's what they're up against, which is, I think, kind of cute. I often have these phone calls with people. They're like, I want to start a subnet. And I'm like, oh, that's so nice. But you're really not coming at it with the right level of intensity because your adversaries on this network are very serious. And so they could literally be a nation state. It could be North Korea, which like North Korea is so desperate for revenue.

39:03Lon Harris:I don't know if you remember this story, which we covered. North Korea was placing developers as remote developers in companies, not because they were hacking the companies. They wanted the six figure jobs. Yeah, yeah. No, we talked about how those companies had to develop like whole protocols to identify when a North Korean was applying for a job, pretending to be a different kind of remote worker. Yeah, I remember. Yes. Right. Building that is difficult. But we have now, I would say, like the top 10 to 20 subnet-subbitters, which are premier and they're experts at this. And they can do things like, you know, subnet 51 figured a GPU attestation, you know, trusted execution without trusted execution, which is quite incredible.

39:44They don't even have TEs, which is what Intel does. They built the algorithms that could go in, SSH into the GPU, and checked everything in the world to make sure that that person has that computer and is not cheating. And they built economic derivatives to make that make sense. And so, you know, now those churn. But it took a while for us to build those things. And so this is all just one level. That's just a level of the mechanisms on BitTensor. But in an absolutely insane, but it turned out well decision by us, we decided to make the actual creation of the mechanisms themselves on the Bitensor blockchain mechanism itself.

40:27So we went, okay, hey, let's apply ourselves to ourselves. We're really good at building these permissionless mechanisms. Let's build a permissionless mechanism which selects permissionless mechanisms. And so that was what we launched just over a year ago, which we call the NanoTao, which is where all of these subnets got their own token. All those tokens got paired to BitTao and all of them are trying to push the metric of success inside our ecosystem, which is increasing their price while staying on basically. And so our thesis at the highest level of BitTensor is that we want to, hey, this network is going to incentivize and try to optimize people for bringing inferences.

41:14Great. Well, we're going to optimize people bringing projects into this ecosystem that produce a lot of value, which we can measure with price, more or less. and with a couple of knobs here, and obviously it's complicated and we can get into it, but that is what we think will pull the most amount of innovation into the ecosystem and also drive the most amount of value into BitTensor, which I believe does work. When you look at the way in which the hand overhead crawling towards performing well in this ecosystem, the more competitive we make it, the more badass the teams have become. And it's quite impressive.

41:58It's really reached a point now where the entrepreneurs in this chain are so much better than I am and many times smarter. And I listen to as many of them as I can because they all often are mentors and guides for us at the ecosystem. How could we improve the core incentives of the BitTensor blockchain?

42:20Lon Harris:So this is a good pausing moment. Again, we'll just stay on ENGY.ai. We've been talking about the person running that is a guy named Ning Ren. N-I-N-G-R-E-N. This person came to BitTensor. How and why? Who are these people that come to you or to the BitTensor Foundation? and we should understand what that is and how that works. Who are these people who are attracted to creating these projects and what's their goal? Are they like freedom-loving individuals? Are they entrepreneurs? Are they hackers? Are they some combination of these? A good number of them are scammers. And you kind of can't avoid that because, you know, you have the doors open.

43:09And that's part of what we're doing is where we're saying, We can open our eyelids the most and that's what is going to make us win. But we're staring right at the sun. So you're going to get a lot of stuff. So there's a lot of bullshit. But then there's also the best stuff and the highest quality and the most intelligent people that if they can get across the stigma, because there's a lot of stigma in crypto. Well earned, honestly. And if they can get through that stigma and they can look at the technology and they can see what they can build here and they understand technically and philosophically, like why this is such a powerful primitive.

43:56Yeah, those people that see it, they come. But you need a level of openness for sure, because it's not the most treaded path, right? Right. And but so someone like Ning, like he he was brought in by Algod, I believe, who built a team and said he knew this guy who'd worked at Google Brain, where I'd also worked. And he didn't know about it. And but he was told about it and he thought it was super interesting as it is. And so he got fascinated and obsessed with it. So there's those types of people that come in. And then there's people that have absolutely no academic experience, but they're just raw entrepreneurs that go, hey, wow, this is novel.

44:54This is the thing that will allow me to build something that can break down that glass ceiling. where, you know, can you really break past these fiat-funded companies by just doing the same thing that they're doing? And I think that the thing that compels a lot of entrepreneurs in our ecosystem, which compels me, is that we need to do something different and better and more powerful than what they're doing. Otherwise, they'll just beat us with more cash, right? And so...

45:28Lon Harris:So this is also mind-blowing. The fact that BitTensor exists and it's relatively stable and providing functionality to the world that some number of consumers and enterprises are dependent on is extraordinary as a technical achievement. But it's also extraordinary, as you're pointing out, hey, there's frontier labs. There's open source projects out there. There's a lot of competition. This is the highest degree. Like, what is the chess rating? ELO or something? Like if this was chess, you're going and playing with the grandmasters, the grandmasters being, you know, Elon Musk, you know, Claude, Sam Altman, OpenAI.

46:11I mean, these Gemini, Sergey Brin and his team, like these are the most elite and you have to beat them in the offering if it was inference in this case.

46:22Lon Harris:So it's an extraordinary achievement on so many levels right now. And yet the largest supercomputer in the world is Bitcoin. So this is the technology that has the example of the only thing that's beat them in SaaS. So there's something to be mined here. And it takes time to build the future. But now we're really starting to see what's so exciting right now in BitTensor is that these primitives, they just make perfect sense now. So, you know, we just we turned up the emissions for a lot of subments. We said, hey, you've got to turn up the amount that you're paying miners. And, you know, some of the mechanisms like like 51, they just the revenue just scales with the more the more money you put in because it's permissionless.

47:12But the companies don't act like that. Oh, here's a bunch of money. Oh, it's not necessarily that you can just distribute that immediately and scale efficiently. There's so much inefficiency when you have that. You try to scale a billion dollars through a human organization. You have to hire people. We were talking about this at the beginning of the call. You have to fire people. You have to give them contracts. You have to get spaces. You have to find all these things, put them all together. It's not easy. But a mechanism that's just described by really just a code base and a permissionless market, you just pump more money through it and it just scales.

47:46So because they've tripled the amount of money they're paying liners, they've tripled their amount of compute in two months. And so these things are really beginning to work and it's really exciting to see how they are coming together and meshing together in an ecosystem. But our goal is not just to do more compute and more inference. That's fantastic. Our goal is actually to truly head to head with the elite centralized labs at intelligence. This is, I think, the pinnacle commodity that we can measure. Inference and compute and storage, these are sort of predicates. But this is where I think that, well, this is where we're going to go.

48:32And this is what we're trying to build right now. And it's much more difficult to measure intelligence. A computer is very difficult. An inference is extremely hard. But measuring intelligence is still possible. It's just very, very, very abstract and hard to get at. So this is what my personal mission is. And I actually run a subnet on the tensor right now. It's where all I was working on right now is called Affine. So we're building the code, the mechanism that can actually measure that pinnacle element. Like what does it mean to actually grasp in your hand intelligence as a commodity? So that's the thing that we're thinking about.

49:09Lon Harris:And how would one measure that? We do have like humanity's last test. We have all kinds of benchmarks from LM Arena. is it as simple as saying, here are intelligence tests, and I want to hear more about Affine. A-F-F-I-N is... Affine. Affine. A-F-I-N? A-F-F-I-N? A-F-F-I-N. A-F-F-I-N. A-F-F-I-N. Okay. I-N-E. I-N-E, Jason. A-F-F-I-N. I-N-E. And that's subnet number... 120. 120. There's 128 of them right now. There's 128. There's been talk about 256. what does it take to start a subnet? How does that work? Do you go to a board of subnets? Do you go to the other ones and say, hey, I want to do this?

49:52Lon Harris:How does it work? Well, if we had a board, then we wouldn't really be permissionless.

50:00And so anyone can register one. And the way that you do that is exactly the same fractal-like design. So it's the same way that you would register into a network of a subnet. you basically burn some token um to pay an entrance fee into the into the uh league in this case the league is 128 in size a normal sub is 256 so we'll probably get to try we're trying to get to 256 at that level as well um so you enter into the network and you start building

50:33Lon Harris:your system and people show up and go hey how many tau do you have to burn or contribute this is like buying a franchise right like buying a team uh in a league yeah let's look it up oh okay and how is that determined yeah it changes um every block um on on on the tensor and and so what we we use is a it's kind of like a dutch auction um so the current rate is 608 tau and we're trying to lower that so it's it's about it costs about 121 000 to have one of these 128 slots, which is a lot of too much. And we want to lower that a lot because, but it's a Dutch auction. So the price decreases until somebody is willing to pay for it.

51:18And then when a subnet registers, we double the price and then we lower it again.

51:22Lon Harris:Well, I mean, it's essentially the cost of joining the money you would get if you joined an accelerator. It's classically been 125K. So it's not, to my mind, like a crazy number, but it's certainly not nothing. So where does that Tau go? It just gets burnt and it lowers the amount of Tau in the network or it gets into the foundation. Where does the Tau go? It goes to you. Who does it go to? The other subnets? It actually goes to create liquidity in the initial pooling between Tau and your alpha, what we call alpha token. So the subnet token, it creates liquidity. So all of these subnets in BitTensor, or Alt 128 have in the liquidity pool, a V3 automatic market maker.

52:07It's basically somewhere you can just buy the token through with a little bit of slippage. It's a smart contract itself. And so each one of them has a pairing with TAU. And so you have to buy TAU in order to buy those tokens. And this is one of the ways in which we build demand for the underlying token, right? It's like the US dollar, right? It has all of these amazing companies inside of the US system, and it has a massive network effect. So we're creating -

52:36Lon Harris:Are people speculating and just buying five of them and sitting on them? If I was like an investor and a speculator, can I just buy four of them for$500 ,000 and sit on them? You'd be like, I'm a speculator? Not financial advice. Yes, you definitely could do that. And people do do that, right? Like, you know, we had this question problem, which was how are we going to, we built the platform so people could build these systems. But we wanted to know how are we going to incentivize the teams? Okay, we still have, you know, 10 million Tau to distribute. And okay, well, what we're really good at is an optimization mechanism.

53:19So let's build an optimization mechanism that these teams can compete in, right? Where they come in, they get relegated, or they go to the top. And so we came up with this idea of using the price of a paired token as a thing that we would measure. And, you know, that allowed us to build, allowed us to distribute the inflation of Tau itself into a network with people who are mining by creating subnets. So people create subnets and they mine Tau by creating subnets. That's the meta system in BitTensor. And the idea is that, you know, as a whole, this network, if we optimize for all those products to produce value, which is measured by their price, we can actually elicit the internal machinery of capitalism, which is thousands of individuals from DGENs to scientists and the like to long-term investors.

54:11We can elicit all of that swarm intelligence to properly order rank these projects. Like, hey, if you were to sit down right now, Jason, and go, hey, how can we order all of the companies in the United States? And you didn't have the NASDAQ. How would you do that? It would be an impossible job. So what we do in society is we use markets to do that ranking order for us in some way. So we did the same thing inside the internal market of BitTensor. We elicited this market mechanism, this trading system to order the submits and push the cream to the top, which is what you see. So, you know, when you're entering into the Bitensor ecosystem, I often tell people, you know, be careful because there's some rotten milk and there's some cream.

55:02and the cream is more expensive and there's raw milk and your job for playing this game is to actually help us contribute to the actual movement of this internal system, right? Like you're actually governing BitTensor. In that way, BitTensor is highly decentralized. They're probably one of the most decentralized networks ever produced. It's not like 10 nodes. It's not a thousand nodes. It's tens of thousands of nodes contributing and making... Making informed bets on the truth. And so that's how we built. Yeah, that's how we built it. And that's holistically, I think, what it looks like. It's extraordinary.

55:44Lon, you had a question. Yeah, we got to talk more about AFI before we were on a full-on chat mutiny. The folks want to hear about this. So I'm curious. This is the idea is you're building models. And if so, how are you setting up the competition or how are you setting up the landscape to ensure that your miners are producing the next generation of models, like of the kind of model that you want to train and design. Last year, we did produce a model that was better than Quinn's best model at the 35 billion range. And just before we launched it, they launched another 35 billion model that was better than ours.

56:22And so, you know, everybody's in the race now. It's an unbelievable race. And like, you know, it's not an understatement to say that we, We are up against nation states because this is truly like China versus the United States. And there's that level of funding. So it's in no ways easy. But we did produce a very good model, but it didn't take us to the level we wanted.

56:49Affine's logo is mining reasoning. and reasoning is the way in which a model thinks to itself so that it can answer a question properly and um so what we have the the miners on the network do is we have them produce machine learning models that can produce reasoning that they can reason in a way that makes other machine learning models answer the right question so it's sort of indirect it's like um imagine if i were like i you would think that i'm smart if whenever you talked to me you felt a lot smarter yourself right um like i know it happens better right that's a really good sign it's like a it's actually reflective right it's like you you you feel you're you're you know that i'm smart if when you talk to me, things make sense to yourself.

57:45And so right now, this is what we're honing in on as the core mechanism for that network. Even if it doesn't necessarily produce models that are stylistically perfect, they can, at the very least, if this is successful, be adapters to other machine learning models. So you could run GLM. And instead of GLM thinking, you would just talk to the smaller model and then be able to go like 30 times faster. Got it. But that's not the end goal. Actually, the end goal is that we think that mining that latent space of thought is going to be the prerequisite for us training incredibly good models. And so that's what we're doing right now.

58:31If you go to Affine.io, you can participate. For people on the call that are interested, we're in this really interesting period of time where you don't need to be a machine learning engineer and you don't necessarily need to be a computer scientist to participate in these networks. You can literally just get your open claw and send them the website and he will tell you or it will tell you whatever you want. The gender you want. They will tell you. They will tell you. They them. They them them. Gap is a he. I don't know what you guys are talking about. He's a he and he is my friend. He's a he.

59:04Okay, great. he will tell you what is needed to participate in these games. And you can just basically send your AIs to participate in these networks, which is a really, really interesting time to be allowed. And this is what I was talking about before the call started with you, Lon, about how we work remotely. So I'm nowhere near the rest of my team right now. But we can all work together through these games. We don't need to be in the same room. We just build these really well-defined computational and innovation games. And then anybody in the world can contribute almost without communicating by just sharing the best work they can.

59:49And if you're not good, if you don't come to work, it doesn't matter because you're just going to get replaced very quickly in these hyper-competitive systems. So, yeah, the call-out I've always said from the very beginning, like, I don't actually really sell Tau token. I sell that you can be part of this, and you should, because it's very exciting and it's very fun. It's like, for some people, Mining BitTensor is like the most fun game they've ever played, ever. Like, most dynamic, most interesting, most rewarding thing they've done. Along those lines, one thing I'm really fascinated about, we've talked to other subnet, or we've talked to subnet owners about this.

1:00:27Like when you're first setting up these competitions, how much are you sort of thinking about the 4D chess of it? Like I have to create a competition or a system that's so tight that no bad actor can come in and like game my system. Like how much of a part of the project is that, that sort of adversarial thinking about? I would say that that is the entirety of the project. Okay, fair enough. um the the all the other stuff is fluff in some sense yeah because you're you're you have to think about your you can start with all the marketing you can start with a website and that might help you and people will maybe invest in your project because they're like oh this guy's got a good sense of sales um but if if you if you can't solve the adversarial problem in the network then the network produces no value right and so you know it's then it's just lipstick on a pig and the you you're you're wasting your time you're wasting my time you're wasting everyone's time and that's the most fun thing and when you get down to you know how do you resist adversaries you often find that there's this very like compressed idea at the very core um a very simple compressed idea that um you're truly building that is like an elemental, right?

1:01:48Okay, oh, interesting. So inference verification is actually information checking, information pairing. You're checking that the information produced by this thing is similar to this thing. And so they're actually, they're producing inferences, but underneath the hood, they're producing information and they're trying to produce the information as fast as they can. And so that is actually the commodity of an inference network where people don't talk about. So anyways, that's like the inside baseball philosophy of the stuff, but I love that the most.

1:02:22Lon Harris:There was an interesting moment when this language model got created. We were on All In and Chamath brought it up with Jensen. You probably saw that clip where he was like, this is so impressive. Talk a little bit about not only that moment and like what you took away from it, but then the controversy with that subnet and it imploding. Right. And what we learned from that, yeah. The reason why Templar is called Templar is because when we were building this subnet, we talked about how one of the holy grails of in the artificial intelligence field and the psyche and neosphere, whatever you call it, is not decentralized inference or decentralized computing or even what I'm talking about with affine, with the model training.

1:03:17It's decentralized training, which is where you have a computer and I have a computer, Lon has a computer, and they are making the same machine learning model at the same time. Because in order to train a trillion-primor machine learning model, you need a hell of a lot of compute. So in order for the rest of us to organically self-organize and to take on OpenAI, we need to come up with a way that we can work together because only together we will have enough compute to compete with the big guys that have the billion dollars investments in infrastructure. So, you know, we need to figure out how we can train together.

1:04:02But in order to do that, we come up against some fundamental limits of physics, which is that the way in which these machine learning models are trained is that they merge their weights every step, more or less. They merge them. So you do some work and I do some work and then we just merge them together. And if one of the parameters in your network is pointing this way and the other one's pointing this way and this one's pointing this way, we find this middle point, which would be where the model is most intelligent. Because you've trained on some data and I've trained on some data, right? It makes sense, right?

1:04:34But in order to move these models across the wire, it's heavily expensive in terms of bandwidth. We're talking like a terabyte of data more, right? And if we want to do hundreds of thousands of steps, well, hundreds of thousands of terabytes that we need to communicate, which is more than any of us have at our home connections. And it's certainly not what an average person has at their home and not on their laptops. And so how do we aggregate together in the first place if we can't even do it with the internet connection we have? And maybe we can do it, but it's going to take 10 years, right? Which then is no longer important, right?

1:05:10So, and then on BitTensor, to put it from another perspective, we have all of these compute aggregators like Liam and Targon and Cube, which people actually bring in, they plug in their GPUs. But these GPUs are all over the world. So for us to use them, we need to come up with a decentralized training mechanism. So in some sense, like training a trillion per hour model is sort of beyond what BitTensor can do until we can come up with an algorithm that stitches together all of the compute in a way that gets around this bandwidth problem. And it's also the holy grail because, you know, a lot of people are thinking about this problem and people haven't thought about this problem for a very long time, including myself.

1:05:50So that's why it's called Templar, because the Templars were trying to find the holy grail or they're protecting the holy grail. And so we came up with that name and we started working on the way in which to do that, which was to take advantage of some of the Tensor's primitives. Like we have Hippias bucket storage and so where all the computers can upload to single places and then download, which is kind of cool. And so we built this out and then we also built the algorithm where we could what would we be measuring that the miners are doing in the first place? so as I as I talked to you about like the the core of the the the problem is okay well how do I know that you you did the inference right so in this it would be how do we know that you did the training how did you how do we know you trained on that particular subset of the data that we needed you to train on in order for you to merge with us and how do we know that you don't just contribute bullshit that destroys our model while we're training because that's really difficult and training machine animals is a very fragile thing if you have one person there that's fucking around specifically language it can destroy the whole thing so this is why like when

1:06:57Lon Harris:yulon talks about hey the next version of grok's coming out he has colossus it's a some number of day training run you throw some you know ranch into that machine you got to start over and so that's perfect so what happened when templar and the rug pull i know this is like the one thing people use as an attack vector on BitTensor and the subnet and the architecture? Just candidly, what happened? Did somebody run off with essentially all the tau in their subnet and just tell everybody to fuck off and it's just part of the system? I mean, effectively, in any company, if somebody works in the company, the CEO starts a company, they can just leave there's actually usually there's no contract you invest in me i just you invest in my company and just be like i don't want to do it right um this happens in early stage startup because some people look at what it's like to be an entrepreneur and they go well shit that's a lot of work not for me yeah and and not only is it a lot of work it's a lot of stress and I just got some investment into my company and maybe I think I'm just going to take that.

1:08:17So what this person did was they just sold their shares, more or less. They sold their shares and while they also sold their shares, they wanted to cover up that they were leaving with a reason for leaving. Like it's not that I'm a bad guy, it's their bad and so I look good. I'm going to cover my ass as I do something that's really shitty, which is take a bunch of people's investment and then just walk away, which is a pretty standard thing in crypto, right? Because it's, you know, one of the things about permissionless markets and us allowing for early stage startups from non-accredited investors and things like this is that basically people get burned.

1:08:56And because most of the time or like a lot of the time, this kind of stuff can happen, right? And so they wrote an article that basically said, hey, we're not bad for leaving. It's the mean network and cons for being an evil dictator because he sold some of our token and that was mean. And then what did they go do? They did another startup or they want to have you with the tensor? No, the project fell apart immediately. And I mean, of course it does because you kind of like, you can't burn your reputation like that. You can't. Yeah. Take money from people and expect people to follow you.

1:09:36Lon Harris:This is a reoccurring issue for Y Combinator, where somebody goes to the Y Combinator program. If you get accepted, not only do they put that$125K in, they will give you a loan$375 ,000 in an uncapped note. Then you get a bunch of people excited, and you might get a bunch of people on demo day that let's just theoretically say put in another million. Well, I've had multiple people contact me that say, Jake, you wrote the book Angel. What should I do here? I'm not getting updates. The project's not active. They raised$1.5 million, and they're just using it to live off of, and we don't see any updates, and they're not shipping any product.

1:10:17Lon Harris:What do we do? And I say, that's the cost of going to the casino, is you could have bad actors. Now, you could file a lawsuit. You could do all these kind of things. you put 50K in to file a lawsuit would be 250K. To then pursue it would be a million dollars. And then the outcome is maybe you get back your 50 in the best case scenario. So talking about trust, the startup ecosystem is largely based on trust. We've literally had people take the money from their bank account when we give them these small checks and just YOLO it and go crazy. and it's like you it's kind of like um credit card bad credit cards you know you get like a two percent or one percent fraud yeah and you're like okay whatever 50 basis points of fraud is what it's going to be i mean this happens in every industry you remember there was that netflix sci-fi show they were creating and the guy just took the budget that they gave him and just went and bought a bunch of mattresses that really happened that really happened i think he's in jail now where he got, yeah, he got sentenced already, but I remember this.

1:11:22Yeah, he bought like watches.

1:11:24Lon Harris:Fraud can happen in Hollywood, startups, and on the BitTensor network, I think. But you did tighten up the Tensor network a bit based on that? Related question from the chat room. Oh, sorry, go ahead. Well, I just want to make a point here, right? So it's a double-edged sword to allow anybody permissionlessly to have access to early stage startups, right? Like this is one thing that crypto does, right? It's like, okay, you can be on the ground floor potentially where you can't for OpenAI. And so like this, but as a consequence, we've learned that there's a lot of issues. And so crypto has become a lot more prickly and has thorns now because people have woken up to the realities of this network now.

1:12:11So it's a double-edged sword, but the benefit is that people get access. And I would say that overall in the business ecosystem, we've made people a lot more money than they've lost. So that's, you know, I hold on to that. No, but the reason why it's particularly bad in crypto is because in early-stage startups, you would invest in, they can't just go and sell their equity into a market automatically. Because there's no market for that. There's no, like, order for -

1:12:43Lon Harris:secondary market of startups. Although people have tried, but even for the nascent ones, there would be no buy side. Exactly. But we, in crypto, build the buy side, and we let these things float. And as a consequence, it's part of our technology, right? I spoke about how this internal market is how BitTensor works. It's actually a functioning aspect of the machine, is that we have individuals that govern the network through market dynamics. It's not something that we're not just creating alpha tokens for nothing. We're using them to move like a computer through space. But so the reason why it's particularly pernicious in crypto in general is because anyone can just go and sell on market, hence rug pulls, right?

1:13:30It doesn't exist in startups because you can't do that. There's not going to be a secondary market. And if you just go to your investors and go, hey, quickly, give me my$10 million, I'm out. They'll go, actually, I don't know if I want to buy that. because you're selling. So what we built into the chain was basically the best that we could do is go, hey, let's build sort of an ability for the subnet teams, not enforced, that allow them to express their conviction, we call it conviction, by locking their tokens effectively. And in a way that if they want to go and sell them in one big sell, that's a public event.

1:14:14And anybody who invested before them could be like, you know what, I think I want to get out of here because you just did that. And so that's actually led to this. It's been very beautiful, actually, to see a lot of the teams, you know, go up to the plate and be like, yeah, you know what? I'm locking this thing perpetually for years.

1:14:32Lon Harris:Which, by the way, we call that vesting in startups where you vest your shares over time and then you can't sell them. And if you do want to sell them, there's a board decision. Hey, we're going to do a secondary offering. Hey, as we wrap here, tell us about the foundation and its role. We had a bunch of questions. Maria123 asked as well. Not Maria678. There's a Maria123 line. You've got to keep your Maria's distinct, James. It's so separate here in the chat. But what's the role of the foundation in all of this? And then what's your stake in all of this? Like as the creator of this, Do you have like a gazillion Tau?

1:15:07Lon Harris:And how do you stay motivated to keep this holding on? Because my gut tells me it's, you know, Tau isn't at the point yet where it would keep going if Kant went away. So how far are you away from making yourself irrelevant in this equation? And then what does the foundation do? I think it would definitely continue if I went away. Most certainly. It would be, I think, I hope that people would miss me. but there's a vibrant community of people that want to really participate in this network and that they do. And there's an open source community and those people contribute. And there's 128 different teams that understand this technology really well and want the system to go along.

1:15:51And there's the foundation in Canada, which is no longer involved with development. And there's now another foundation, which is more shielded, which is purely about upgrading the chain. So we push a lot of changes to tweak and improve the mechanism. And that goes through our internal governance system on the chain, which we call the triumvirate, and which is this year going to be expanded. Basically, because we have a chain that we can build government systems directly into it, we intend to build all of that this year. So that's the two foundations, you might call it. There's the OpenTensor Foundation and there's the RAL Foundation, and they hold different sides.

1:16:39One's more marketing and outreach, et cetera, and one's more development and upgrading the chain. I'm no longer involved with OpenTensor. I work just in the RAL Foundation, and I program. That's my language. Actually, this is very unfamiliar for me. I am a programmer at heart. That's how I speak.

1:16:54Lon Harris:Awesome. Well, listen, continued success with it. We're going to be monitoring all these subnets. As I told everybody, I think buying one Tau, just buy one Tau is what I've been telling folks. Why? Because I think it is a ticket to watch something, a spectacular experiment occur. So if you think of it long, like going to see a basketball game, just your 200 bucks or whatever it's trading at. If it's the finals, you're$15 ,000 or whatever. OK, sure. But who's got it? It's a front row seat to like bet and learn. And it's not financial advice. But it's also like in my in one way, a vote for me. like if you were giving a, you know, GoFundMe, like I almost see it like as a vote of confidence that maybe there could be a decentralized AI intelligence platform out there.

1:17:44Lon Harris:And that's good for humanity. Yeah. Pardon me. So I see it as like, and then, oh, and there's a third. What if it is Bitcoin? And what if it goes from 200 to 60 ,000 or 120 ,000? Hey, that could be like a great bet. Don't sell too early. That's what I learned about Bitcoin. I had a few and then I sold it when it hit like a thousand. Like I didn't think it was going to go. Dummy. You're a big dummy. You got to ride your winners. Anyway, how important is it to get people promoting this and getting involved in it? Or is that like actually a distraction? The fact that like people like me are speculating on it now and interested in it.

1:18:19Lon Harris:And I'm obviously a venture capitalist. It's a very passionate fan base. It's a very talkative, passionate ecosystem. Anytime we talk about TOW, yeah, the video goes crazy. How do you think about the pumping in crypto or, you know, non-builders like myself saying, hey, I want to invest in this because I think there's something here. And I'm fascinated by it. And I like to make a return. And I think this is like risk adjusted for me, like a great, you know, hey, maybe this thing, you know, I look at it and go, hey, maybe this thing can go 100 or 1000X. That's why I make the bet. bet. It's like a real long shot kind of bet in my mind, but it's also fascinating.

1:18:56Lon Harris:So how does speculators like me and then pumping and all that impact these projects on a practical basis, if at all? Well, I mean, I think that it's inevitable and we can't really avoid it. It's the nature of markets, but it's not the goal. The reason why we have Tao in the first places so that everyone can have some and that everyone can join and that it can be split up and and in the first place they can be split up it turned into a whole 21 million of these things so like i love it that there's people that are excited and that's amazing and and the the ones that are you know authentically promoting it they come they go um there's always people they're going to say shit on both sides.

1:19:43I can't stop them. I am way more interested in people that see it the same way that I see it, which is more as a really, really powerful technology and that are super excited about that and the potential for us to build something that's novel and unique and competitive and the third path for what artificial intelligence can be born out of. so the yeah but the market dynamics are also really fun I'm in all the press chats as well

1:20:13Lon Harris:alright there you go we got over an hour with the man himself constantly shipping you can follow him on Twitter constant reborn great follow and I think it's great that you're going out and talking so I appreciate it I know you gotta get back to coding but I think it's important for people to understand it like just from first principles and you did a great job of sharing that with our audience today. All right, let's drop it. Yeah, thanks so much for being here, Conn. That was great. Here's the thing, Lon. You know, these projects are so, they have so much potential and the intention is super important.

1:20:48Lon Harris:And when you spend an hour talking to him, it reminds me of what I saw in the Bitcoin early True Believers. And I had somebody on, I think, in 2011. And I saw it then. I didn't make a big enough bet. And now I see this and it just all the signaling is going off. So I was like, all right, let me put like, I don't know if I put like half a million or 750 into this, something like that. Not a lot of money for me be the equivalent of like, you know, maybe somebody putting in, I don't know, a couple of thousand dollars, right? Or I don't know, $10 ,000, whatever. You know, it's a, it's a smaller bet for me.

1:21:25Lon Harris:It's not out of my funds. It's just a personal, because I think there's something here that's notable. And my signaling goes off. Just like I had the open claw signaling. You remember when that happened? I was like, guys, we got to like pay attention to this because this reminds me of the mobile cloud computing era when like local mobile GPS all started coming together. I got that signaling for Uber and Robinhood. Hey, what would mobile do to trading and getting a car and GPS and all that stuff? And being early on a lot of these things, same thing I saw when I bought the 16 Tesla. Now that doesn't mean you're going to be right.

1:21:58Lon Harris:Right. Yeah. But you do have to place the bet is what I've learned. Yeah. I mean, sometimes you are right, but it's just not the right bet at the right moment. I mean, there's no way, you know, there's, there's no way to know which bet is the correct one, but it, you know, it keeps things interesting. Yes. And, you know, like the interesting thing for Tesla was, you know, I bought the two cars for 300 ,000. I wish I just bought 300 ,000 in shares and just let it sit forever, you know? Uh, and I always go, Oh, you know, I did have shares and whatever, but I, and I did fine. So I'm not complaining.

1:22:28I mean, hindsight on this stuff is always 2020. Like, if I could go back in time and tell myself not to sell my few Bitcoin when they were in the thousand, I was like, I can't believe how much money I made. What a windfall. I'm going to go get myself, you know.

1:22:41Lon Harris:And the mistake was you should you could have sold 10 percent or 20 percent, but you want to keep it. If something's accelerating, you have to ask yourself, is it going to stop accelerating? Is this just. And so when I look at this, yeah, I think I might actually be down right now in my BitTensor bet, which I made this year. And I like to be vocal and clear about this so that nobody thinks I have some nefarious reason. And that's why I told folks just by one, because I realized people were starting to take my tweets online, Lon. And, you know, there's all these Pelosi track or now there's a JCal tracker.

1:23:14Lon Harris:So when I bought Figma or I said Uber was at a bottom last week when it was like 67, when people were panicking. They think you're Michael Saylor-ing. They think you're - I'm not YOLO-ing. You're pumping. I'm not trying to influence anybody. I just like to be honest about it. And when I J-trade something, I just take a screenshot now. Like I bought Figma when it was at 20 because I was like, you know what? Dylan's like a beast. We should have him on the program. I was like, Dylan's a beast, you know? Like he's not going to sit here and lay it down. He understands designers better than anybody.

1:23:42Lon Harris:He understands how to make a great tool, build a great brand, get people to pay for it, all that great stuff. I like to make a bet on him. It's now up to$25 a share, whatever. Now they're doing the same thing. They're retweeting it. J-Cal called the bottom on this. J-Cal called the bottom on this. You know, my promise to you as the people who listen to this specific podcast, when I do it, I'm going to try to be transparent about it to the extent I can. I can't do that with private companies because it's not my choice to make the funding announcement. Right. That's the founder and the board's choice.

1:24:12Lon Harris:So sometimes I will make a private bet. I'm not at liberty to talk about that. Yeah. Your public. But I try to be transparent with the public. Yeah. The J-Trades. Only because, and the J trading is all, I'm not day trading, I'm J trading. J trading is hold for a decade. That's the nature of J trading. Now, you may want to quit a stock if you get new information, but I like to find stocks that I'm comfortable holding for a decade. That's my hold period I'm looking for. Sure. You got to wait. In tech, the strategy is always wait for there to be a bad news cycle. The stock dips, you get it at a little discount, and then you just hold on to it forever.

1:24:48That's the play. That's your This Week in Startups. We'll see you next time. Bye-bye. Bye-bye.