Unchained - Could Bittensor End Up Being the Only Crypto/AI Project That Matters? - Ep. 712

Episode Date: October 1, 2024

AI and crypto are two of the hottest topics of the decade, but are there any projects truly making waves at the intersection of both? Bittensor, an open-source, decentralized AI network, is positioning itself as a leader in this space, with its TAO token seeing explosive growth and its model challenging traditional centralized AI companies. In this episode, we’re joined by Joseph Jacks, aka JJ, founder of OSS Capital, and Sami Kassab, partner at OSS Capital, to explore why they’ve gone all-in on Bittensor. They discuss how Bittensor works, what makes it different from centralized AI models, and why they believe this project could be transformative for both crypto and AI. Show highlights: OSS Capital’s background and how they got to invest in Bittensor Why Sami and JJ are bullish on TAO What the three roles in the Bittensor ecosystem are How new subnets incentivize miners to develop AI models Why it’s so expensive to launch a subnet Why Bittensor was built on the Polkadot SDK The pros and cons of rolling out EVM compatibility  What Allora and Commune AI are focused on within the ecosystem How Bittensor can compete with the big AI companies  The dangers AI poses to humanity and whether Bittensor can mitigate them Visit our website for breaking news, analysis, op-eds, articles to learn about crypto, and much more: unchainedcrypto.com Thank you to our sponsors! Polkadot Mantle Guests: Joseph Jacks, Founder and General Partner of OSS Capital Sami Kassab, Partner at OSS Capital Links Previous coverage of Unchained on Crypto/AI:  Erik Voorhees' New Venture: Why AI Desperately Needs Privacy and Uncensorability When AI and Blockchain Meet, How Can Each Technology Benefit? The Chopping Block: Why AI Will Change the Course of History in Crypto Learn more: A Beginner's Guide to AI Tokens 5 Use Cases of AI in Blockchain Bittensor: Bittensor’s website The game theory of TAO Nous Research leaving Bittensor  Seth Bloomberg’s tweet on “Bittensor’s Network Effects” Cost of building a subnet Commune.ai Allora Network Timestamps: 00:00 Intro 02:39 Background of OSS Capital and investing in Bittensor 19:06 Why Sami and JJ are bullish on TAO 24:51 The three roles in the Bittensor ecosystem 35:00 How subnets incentivize AI model development 47:50 Why launching a subnet is expensive 50:12 Bittensor’s foundation on the Polkadot SDK 53:00 Pros and cons of EVM compatibility 1:03:03 Focus areas for Allora and Communi 1:06:48 How Bittensor competes with big AI companies 1:09:19 JJ’s take on AI dangers and Bittensor’s role Learn more about your ad choices. Visit megaphone.fm/adchoices

What BitTencer is actually doing is inverting capitalism. It's inverting the traditional business model for AI, AI companies, AI foundation companies. And it's saying, get rewarded by this token, which is increasingly worth more as you have more activity contribution on the network. And it's very liquid. It's a top top 20, top 30 token. And you actually are required to open source all of your work. You're only producing open commodities that can then be used by other people, whether it's a data set or, an image model or video model or coding model or compute infrastructure, what have you, anything

related to sort of the spectrum of things that you do when you're building an AI system. That's what BitTensor is incentivizing. Hi, everyone. Welcome to Unchained, your no hype resource for all things crypto. I'm your host, Laura Shin, author of The Cryptopians. I started covering crypto nine years ago and as a senior editor of Forbes was the first mainstream media reporter to cover cryptocurrency full-time. This is the October 1st, 2020. 24 episode of Unchained.

Mantle's M-Eath is now the fourth largest LST with $1.3 billion in TVL. M-Eath offers folders cumulative incentives and airdrops, in addition to native ETH POS yields. This includes exclusive rewards like Eigen and Cook. Check it out at m-eath.mantle.xy-Z slash campaigns. Pocodot is the original and leading layer zero blockchain with over 2,000-plus developers. And the Pocodot 2.0 upgrade will be a massive accelerator for the ecosystem, making it faster, more secure, and adaptable. Perfect for GameFi and Defi to build, grow, and scale.

Join the community at pocadot.network slash ecosystem slash community. Today's topic is BitTencer. Here to discuss are Joseph Jacks, founder and general partner of OSS Capital and Sammy Kasab, partner at OSS Capital. Welcome, JJ and Sammy. Thanks for having us. So BitTensor, which brings together crypto and AI, is having what I would call its second moment. Earlier this year, the price of the token tau surpassed $700

before drawing down with the rest of the crypto markets to as low as $215. At the time of recording, it's at just about $600. As far as I understand, the reason is due to the upcoming rollout of Ethereum-based smart contracts on the chain. And you both are at OSS Capital, which is an investor in BetTensor. And you actually told me before the call that you are redesigning OSS Capital to go all in on BitTencer, which is interesting and says something about how significant you think this is. But before we get to all of that, why don't we just start with the basics, tell the audience what BitTensor is. And why don't we have you, JJ, start. Well, thank you for having us, Laura.

I really appreciate the opportunity to share our views and perspectives with your audience. So, yeah, I'm JJ. I started a venture capital fund six years ago called OSS Capital. OSS stands for open source software or open source startup, your choice or any other acronym for that matter. I wasn't really that thoughtful in the naming. The thesis has been pretty exciting and really, really rewarding for me. I mean, I kind of probably do the job for free. We invest in open source startups exclusively. We're not really a crypto or Web 3 or like blockchain investor. We've only done a very small handful of a few investments that you could kind of consider

crypto, but we're really known for is investing in. If you're familiar with companies like HashyCorp or Elastic or GitLab, Red Hat is an older example. We invest in companies that look like that, where at the very early stage, there's really just an open source project. And we typically invest in the developers that created that open source technology. And we help them build businesses around those open source projects. So I've been doing that in a lot of areas of the stack, like infrastructure, software, developer tools, data. and analytics and so on, and also lots of applications that we've invested in. So we're the investors behind a company that builds and that sort of stewards the Godot project,

which is a top 10 open source ecosystem and sort of the video gaming space. It's like a Unity competitor. We're investors in Cal.com, which is an application for scheduling time and coordinating people, and that's kind of like a Calendly or Savi-Cal alternative. We've invested in lots of other applications, open-source sort of alternatives to goals like Notion and Loom and Airtable, Bloomberg Terminal, and many others. And so, yeah, that's kind of the history and the lineage, I guess, of OSS Capital. I learned about BitTencer early last year, January, February timeframe.

And initially was very skeptical of BitTenzer, I think, as most kind of non-crypto, non-blockchain native people tend to be historically and really just started following the project and learning more about it over the coming months. And then the network was really substantial. actually upgraded to include this feature called subnets or subnetworks in September, October timeframe last year, so about a year ago, actually a little bit less than a year ago, when they went live. And since then, in the last 12 months, there's just been this big Cambrian explosion almost of developer activity and, you know, contribution to what BitTencer is really incentivizing.

And so I'll let Sammy kind of go into the more sort of practical nuts and bolts of what BitTenzer was all about. But for me, the reason I, I resonate a lot with the mission and vision is Bitsenswer was really created by a very credible team of founders, one from Google Brain previously, Jake, Steve's, and the other from doing computer science and AI, postdoc work actually at University of Waterloo in Canada, which is like a really exceptional university, and also worked at VMware, a la Shabana. So the two co-founders have actually previously no crypto background, no blockchain background. They're really AI researcher ML engineers. And they were really inspired by what Bitcoin has done for kind of creating this solution for permissionless, cryptographically incentivized

and decentralized sort of digital gold and money. And they sort of thought, could we do that with incentivizing AI production, but where the outputs would all be open source and user-owned and user-controlled. And so they were really having these ideas and designs going back 2016, 2017, time frame. Very early on, actually well before Open AI, or I should say closed AI, was really in the spotlight, right, in the industry. And so I think everyone's kind of familiar with sort of the drama and kind of the craziness in the last couple of years. And so BitTencer's just been a really awesome project to follow in terms of, you know, there's really more and more proof and evidence that with decentralized incentives, you know, no single company really has a monopoly on the most capital and energy that you can. apply for producing sort of the most powerful AI models over time. That's really what we believe.

And so we actually think there's going to be a huge shift from the centralized AI, proprietary AI kind of closed companies to BitTensor really incentivizing this large commodity marketplace of contribution. And so that's that's kind of what we're what we're here to talk about. Sammy, can you go into more details on what BitTensor is? Yeah. So BitTensor is essentially trying to decentral. your traditional AI lab. One way of thinking about it is also as a launch pad for decentralized or I guess AI applications in general.

So the way to think about BitTensor is it's not like this AI infused blockchain. It really is just merging crypto and AI together. So BitTensor is made up of a blockchain, obviously. it's a substrate blockchain. And it essentially serves as this launch pad for other builders to come and launch AI services and tasks, which honestly, it actually doesn't have to be AI exclusive. It can be anything that you can write an incentive mechanism for can be a BitTensor subnet. However, the focus has from the beginning been around AI.

And so BitTensor today has around 50 subnets that, encompass almost every part of the AI stack. So you have subnets that are focused on collecting data, structuring that data. You have all aspects of the training. So whether it's pre-training, fine-tuning, then you have inference. And then you also have like subnets that are focused on AI powered applications and sort of the list goes on. Most people have heard about bit tensor, but many do not understand it. And if you've, been on crypto Twitter at all, you have definitely seen some well-known people criticize it as vaporware or a scam or maybe even a meme. But whenever somebody approaches me and asks me to

like break down BitTensor to them, the first aspect I point them towards is Yuma consensus. And this is similar to somebody looking to understand Bitcoin, needing to understand proof of work. And so Yuma consensus. is, and I'll try to be very brief here and kind of speak at a high level, it's a fuzzy consensus mechanism. And so if you think of every other consensus mechanism within crypto, it's deterministic. And these deterministic systems work great when there is an objective ground truth or a binary outcome. And so, for example, is this transaction valid? Is this block valid? Yes or no? For a storage network. Is this storage provider storing this data? This can be proved with a cryptographic

proof and the chain only needs to verify, is this proof valid? Yes or no? When you think about intelligence, like there is no ground truth. It's all relative. And intelligence more broadly, whether human or artificial, is based on like subjective and probabilistic thinking. Now, all these AI systems that we know today function on probabilistic models. And BitTensor or Yuma consensus more specifically is basically speaking the same probabilistic language as these neural networks and as these models. And so you can actually launch a lot more applications and services that are AI focused on BitTensor that you can't have on your traditional L1. And so that's typically what I advise people to spend their time on first because you're you're not going to really get it like

what's the point of this standalone l1 for example for AI if you can just do everything that's being done on bit tensor on say Ethereum or Solana that's that's not the case at all and so yeah today like I mentioned there's all these types of different subnets and there's a reason why bit tensor has been the most successful in attracting these AI devs and and these talented incentive mechanism designers, because the way that human consensus is also designed is where all mining and validating happens off chain. And so this means naturally you can have more like data intensive

and compute intensive services that are built without all the like crypto infrastructure that is very hard to work with, right? Like these co-processors and these cross-chain messaging bridges, like BitTensor simplifies all this. And so, like, it basically has just become this really easy environment for non-crypto-native developers to come and build these AI applications and attract these really talented miners to come bringing that vision to life. Yeah.

I mean, so from my probably not super in-depth understanding, but at least from what I have gleaned from my research, it's sort of like right now we have these huge centralized companies that are pushing these AI models and they have hired lots of people, they've gotten all these chips. I mean, there's like a lot of money that has to go into this. But like if you're more of, yeah, like an open source developer, like, and it's not to say there can't be groups, of course there can, but there's a way of using this in a way where you don't necessarily have to hire. It's very similar to Bitcoin where you have a coin and it gets, it incentivizes people to

do things on a network that, you know, these big companies like opening eyes. and Microsoft and whatever are hiring people to do. And one other thing is when I was learning about it, I kept thinking about Ethereum, how Ethereum was this platform for a bunch of decentralized different projects. And BitTensor is kind of like that, but for AI. So if you think of how, you know, Ethereum led to like defy, where you have, you know, decentralized exchanges,

you have, like, borrowing and lending protocols. There's this whole world of, like, finance that got built, and like developers could have a vision for like, I want to do this piece of it. I want to do that piece of it. And BitTensor is kind of the same way where there's all these different AIs that are, you know, kind of chipping away at different parts of the puzzle. And the models can be like tweaked to specialize in certain things. And, you know, this actually goes to why the coin has kind of taken off again because there's going to be this rollout. And this actually gets further into the weeds than I was intending at this moment in time. But, but we'll

we'll circle back to this later, but now there's going to be what I view is like the equivalent of the ERC 20s on BitTensor, where essentially there will be coins that can be associated with different subnets. But let's just zoom out for a second because I want to, um, to like drill that. So I just gave like a very layman's explanation for this, but can you just explain like, you know, in your vision, like, why is it that you were more interested in this kind of more first of all, open source, but then also like decentralized version of incentivizing AI, then these centralized models. Because, you know, it is very fascinating that at this moment in time, all the dominant ones are these big players. So like, how do you think Betetzer competes?

And like, why do you think that that is the future and why not like what's already happened? I'll let Sammy go deeper here on this. But like, I think the dynamic right now with companies producing state-of-the-art AI systems is very centralized as a dynamic. Like there's really maybe a dozen or so companies that employ most of the smartest people and also have sufficient enough compute and money to like scale that compute and data to like train and deploy these models. It takes a lot of coordination. There's a lot of scarcity and those individuals, those people.

Because the knowledge of how to actually build and train these AI systems, these neural networks is not not evenly distributed in the world. It's just like in these pockets of either academia or big tech or different researchers. And I think if you look at the example of Bitcoin in this massive success that it had being very disruptive to a lot of key parts of the financial services industry and general currency, currency dynamics, it was created by this decentralized community, originally, you know, fairly centralized because no one knew about it. There was, you know, 100 or so people that first learned about Bitcoin and, you know, quite centralized. And then over time, these systems become more and more

decentralized, right? And so as they become more successful, they become more and more decentralized. I think there's a lot of similar things that happen with Linux in the early days of computing where operating systems were very proprietary in the beginning. And then over time, you know, we had this big open source ecosystem. And I think the lessons of open source give us some insights in terms of how and why decentralization matters. whether you have that in an open source ecosystem or on a blockchain, it really matters to kind of enlist and pull together the most capable people towards a common shared vision,

whether it's an incentive that's extrinsic and people care about a monetary reward or whether they care about an intrinsic reward where it's just, I want to work with smart people and learn from them, which is a big reason why most of the smartest engineers work in these open source open source communities. And so I think BitTensor is really, for the first time, kind of like putting a stake in the ground and saying, we should apply these same types of principles,

Bitcoin with fiat currencies, Linux, with closed proprietary operating systems on the server side. We should apply these same principles to AI and require that when you build and participate in producing these AI kind of components, whether it's data or models or training infrastructure or compute, what have you, you should be rewarded permissionlessly. you should not, to your point, Laura, you should not have to go and get a job at one of these big tech companies. You should not have to go and raise capital from a VC to start the next anthropic competitor. You should not have to go and get someone's sign off and okay to be able to actually innovate and participate in this system. And so, Cal, which is this kind of native currency of BitTensor, subsidizes not the right word technically,

but it's sort of incentivized and subsidizes the cost of doing all these today, very expensive things. for anyone in the world. And what we're seeing more and more of is like there's really brilliant people that actually could be working at Anthropic or Google or these other places. And there may be postdocs or maybe have a physics background or they actually have an AI research background. And they go, well, on one hand, I could actually build and participate in this BitTensor project and contribute my insights and my knowledge as open source and sort of solve some puzzles

on BitTensor. Each subnet is its own unique puzzle. Bitcoin has one puzzle. You know, you have to solve shot 256. On BitTensor, each subnet has its own unique puzzle and it's in its own unique incentive mechanism. And so people can kind of pick and choose and see which one speaks to them and which one aligns with their skill.

And they actually get paid and rewarded by the system very, very handsomely, right? Like the inflation on the network is very high. We're still in the first halving. The tokenomics kind of mere Bitcoin in terms of, you know, there's 21 million scarce supply. they get emitted and produced over, you know, I think it's like 140 years or something. So it's a very long fair time frame. There's no VCs. There's no investors. There's no pre-mine of BitTencer. It's not a company. And, you know, it has these really interesting properties that are increasingly more and more like clearly working and showing that you, once you create these kind of decentralized permissionless crypto incentives, it actually gives people another path and another option in terms of how they want to spend their time. And before we delve a little bit more into it, I just wanted to circle back because you said there are no VCs and yet you are VC. So I think you guys are just now like becoming actors on the network or like earlier when you said you're going all in on BitTencer. Like what does that mean? Yeah. So we're a small fund. We manage a handful of like 50 million dollar vehicles. They're seed funds. Almost all of our investing like I said earlier 95 plus percent of it actually has been traditional like Web 2. open source companies. And, you know, you have a cap table and you're, and you're, and you're, and you're, you know, maybe it's a cloud version of the open source or a pro

version, enterprise version, and you sort of have to balance the community needs with customer needs and build a great product at the end of the day. BitTenster is very different. And what it does is it basically says you, you can get rewarded by this token in exchange for actually creating permissionless and open source AI data models, infrastructure. And that's really incredible. And so it really challenges this kind of conventional wisdom around. You have to raise a bunch of VC money, hire the PhDs, get the compute, train the models, release the models.

You know, you're competing with meta and Google and Open AI and Anthropic and everybody else on the benchmarks. And fundamentally, you have to monetize those models. otherwise, otherwise the VC money runs out and your company does, right? That's kind of the story. Whereas with BitTensor, you don't actually technically have to monetize the models. I think people want to see that. And that's a really good sign that what BitTensor is incentivizing is useful for the world and sort of can meet the world where it is in terms of, you know, you can add additional value and charge money for that. But at a fundamental level, what BitTenzer is actually

doing is inverting capitalism. It's inverting the traditional business model for AI, AI companies, AI foundation companies. And it's saying, get rewarded by this token, which is increasingly worth more as you have more activity contribution on the network. And it's very liquid. It's a top top 20, top 30 token. And you actually are required to open source all of your work. You're you're only producing open commodities that can then be used by other people, whether it's a dataset or an image model or video model or coding model or compute infrastructure, what have you. Anything related to source. sort of the spectrum of things that you do when you're building an AI system. That's what BitTencer

is incentivizing. And so as an investor, we looked at BitTensor. And once we kind of like realize there isn't a company behind this, it's a fair launch network. It was mined, you know, pretty conservatively by the founders early on. They could have mined millions of tokens for themselves. Both of the founders, as I mentioned, Jake and Allah, own less than one percent of the network each in terms of the total supply. So that's incredible. It's really remarkable. If you compare that to the early wallet holder base of even Bitcoin for that matter, let alone the other networks that have come out. Actually, it's interesting.

BitTencer is more distributed, more decentralized on an ownership basis, on a wallet holder basis than Bitcoin was in sort of its first three or four years, let alone all the other ERC20s and all the other networks, which are far less, you know, far less sort of decentralized in terms of the ownership. And so what we did was just purchased tokens through OTC like secondary agreements. And we basically purchased some tokens from some of the large holders who were either people that were mining the network on the first subnet, sort of the root network, which is text prompting, which is inference.

You interact with this GPT2 style model, generate tokens, and you would get rewarded on the network. So that was for the first couple of years. And so there are plenty of people who were just early believers and really, really, you know, discovered BitTencer much, much earlier than we did a year, two years earlier than we did, almost from inception. And so, you know, those people now at this point had, you know, millions or tens of millions worth of the token. And so we actually went off in and were able to go negotiate OTC deals with them. And our fund has ended up holding this liquid asset, which is very different from how we've held in them past. And so like our strategy is not as a hedge fund to go

on like trade tau and stile down. We're applying the exact same approach that the OSSS has taken with traditional web two companies, where even though we technically can, we're not going to be trading and selling down our tau for, you know, probably towards the very far end of life of the funds, which are 10 to 12 years. And so that's kind of the time frame that we look at as the sort of like initial frame, frame of time for the tensor to really prove itself out and become really impactful for the world. I think it already has been demonstrating some of that, but it takes a long time for these things

to play out. Yeah, as a fund, we're looking at also participating in the D-Tal ecosystem. That's another really exciting kind of transition or upgrade graduation. That's what I call the ERC-20 versions of Tao. Yes, yes. Dynamic Tao. I would push back a little bit on the characterization of ERC20 because these are tokens that are that are locked within the network. They're not, they're not sort of like swappable and you can't list them on exchanges and this kind of thing. And so tau is kind of the token for the network and the DTal alpha tokens for the subnets are sort of this. They're sort of like, they represent a slice of the overall tau token supply.

And so it's it's quite different from sort of a layer two or, or a smart contract like ERC20 on Ethereum. Right. But some similar things. The last piece I would mention about OSIS Capital's Fund is we're definitely encouraging and seeing a lot of our Web 2 companies and investments actually mine BitTencer and maybe even participate in some of these subnets, maybe even redesign or evolve their conventional business models to, you know, thinking about what would this look like as an incentive

mechanism on BitTensor as a subnet on BitTencer. and sort of like dipping their toe in the water. So that's pretty exciting. All right. So Sammy, can you kind of go into this explanation of, you know, the role of the different subnets, what mining is versus validating? As far as I understand, there's even like a couple of, it's either that there's two different types of validators or they have two different functions.

So can you just explain the nuts and bolts of how everything works? Yeah, I would love to. So there's three different roles within the BitTensor ecosystem. them, you can be a subnet owner who's designing the incentive mechanism. You can also think of them as like the game designer. You then have the minor. The miners are the ones that are actually doing the work. So the subnet owner is specifying the objective function that he wants or that they want the miners to optimize. The miners come in and they will do the optimizing. The validators then are the ones that are validating the work that the miners are doing and making sure that they're

acting honestly and they're actually doing what the incentive mechanism is asking them to do. And so right now, I mean, one thing that I should point out is that the chain isn't actually fully decentralized. And so there aren't any like chain validators at this point in time. It's just the foundation that's running it. So right now the network's proof of authority. So there's a few nodes that are controlled by a handful of entities, but part of the roadmap is to eventually decentralize that, the entire chain. Right now, what the foundation has deemed more important is like economic decentralization, which is decentralizing the way in which the tokens get distributed to the miners and the

holders, which is what these validators within the subnets are doing. So to zoom back out a little bit, each subnet, as its own set of miners and validators. So there could be a specific subnet. I'll take Subnet 12 as an example. Its job is to, or what it's essentially incentivizing is minors to contribute GPUs. So it's like a GPU network. So the miners are contributing the GPUs and then the validators come in and they're

verifying that the miners are actually running the GPUs that they're claiming to have. Right. And so each subnet can choose their own validation mechanism and designs their own validation mechanism. So whether a subnet wants to take a zero knowledge proof approach, an optimistic approach, a trusted approach, a random sampling approach, it's up to the subnet owner to decide or the community within that subnet to decide which direction they want to go down. Now, zooming back out, the BitTencer blockchain emits 7,200 tau per day. And so the question here is, like, how do these emissions flow down to the specific subnets? Like, who decides how much emissions each subnet gets? And so right now we're in this phase, which you could call like the root network phase,

where it resembles more of a, it's like a Dow design, where there's, a handful of like basically the top 64 largest validators on the network who hold the most tau get to vote in how admissions flow to each subnet. So for subnet 12, for example, each validator within the top 64 subset votes, I think subnet 12 should get 2%. Another one says, I think subnet 12 should get 3%. And basically it becomes an average of all these votes. and that becomes the consent that the chain weight essentially of what that subnet receives.

And to be completely transparent, this mechanism has gotten a lot of pushback just because you can see how there's some centralization forces where the largest token holders essentially have a lot of influence in how these emissions get distributed, which is. why there's this dynamic tau proposal that is intended to be implemented either by the end of this year, Q1 of 2025. And so what dynamic tau does, and you and JJ spoke about this briefly, but each subnet now gets its own token. And this token is also basically mimics the tokenomics of Tao. So it's fair launch, 21 million, same emission schedule. But, And at the same time, like, it technically really isn't similar to an ERC20 because it's not even a token.

It's a staking operation. So like you would stake tau into a subnet and receive, you know, this, I guess for simplicity, let's just call it a token. You receive this token back. But so dynamic tau is transitioning the emission distribution process to a market-based approach. So the subnet that attracts the most buyers of its token will then receive a larger share of the 7200 tau that's being emitted daily. And so you can see how that decentralizes this process a little bit more. And just to zoom out a little bit as well, if you look at the decentralized AI landscape, there are a ton of projects that are all going after like a specific stack or layer of the,

AI stack, right? So like Akash, for example, focusing on inferencing, Jensen, focusing on training, grass, focusing on collecting data. However, there really is, well, BitTensor was the first one to take almost this like integrated approach or this approach where like they wanted to control the entire AI stack. And I would argue this was necessary to actually get to producing like decentralized AI, right? Because you're, you're. creating this coordination environment where you can fund and allocate the capital. You can collectively decide on specific research directions to go down. Like there's so much more that goes into like producing decentralized intelligence

where you need to do this in like one ecosystem, right? And it almost goes back to like the integrated versus modular approach. There are all these projects just like building in their own silos, right? And I love all these projects. by the way, like, I'm a supporter of all of them. But there is no one, to your point, piecing everything together. And so BitTensor, while this emission distribution process that we currently have gets a lot of, like, criticism, it is the first experimentation of, like, trying to decentralize

the actual, like, process of building AI, right? Where, like, Open AI has a group of, like, five to ten people. Or maybe just one. Maybe it's just Sam Altman now deciding which direction and go down. bit tensor is like, okay, let's take this collective approach where everyone gets to say. We tried that. We realize, okay, the people with the most tokens are just going to have like too much influence. Now let's run a new experiment. And this experiment is going to be a market-based approach where like the market will dictate which subnet deserves to get the most emissions. Even if it's like two training subnets, right? The market will figure out which one has a better chance of monetization, which one has better technology, which algorithm.

rhythms are better. So yeah, I feel like I said a lot there. Yeah, yeah. So in a moment, we'll talk more about all these details of how BitTensor works. But first, a quick word from the sponsors who make this show possible. Mantle LSP is a permissionless and non-custodial ether liquid staking protocol deployed on Ethereum and governed by Mantle. M-Eath serves as the value accumulating receipt token of Mantle LSP and is now the fourth largest ETH LST with $1.3 billion in TVL. In addition to native ETH-POS staking yields, M-Eath holders can access various yield opportunities across DAPS on Mantle Network L2 integrations and more. M-Eath holders have previously received over 1 million in Icon token airdrops. With the upcoming October 24 launch of Cook, the new governance token of Mantle LSP,

M-Eath holders can start accruing powder rewards under Season 1, Methamorphosis, which will be convertible to Cook. Visit mith.mantle.x slash campaigns to learn more. PocaDot is the original and largest layer zero blockchain with over 2000 plus developers, and the anticipated Pocodot 2.0 upgrade will be a massive accelerator for the ecosystem, upgrading the infrastructure with eight times higher transaction throughput and twice as fast block times, perfectly tailored core time for the needs of every protocol, trustless bridges internally and into Ethereum, Cosmos near Binance smart chain, and revised tokenomics and the implementation of a token burn to reduce inflation.

Perfect for GameFi and Defi to build, grow, and scale with one of the most active crypto communities in this space. Pocodot recently announced a partnership with mythical games, bringing top games like NFT rivals with over 650,000 players and 43 million transactions, to pave the way for GameFi and the Pocod ecosystem. Get your Web3 ideas to market fast with economics that work for you. Think big, bills bigger with Pocod. Join the community at Pocodot.com. system slash community. Back to my conversation with JJ and Sammy. So, yeah, I remember when I was first learning about this, it sort of felt to me like it was

like, Bet Hensor is trying to create almost like a shoot, what was that? The library in Alexandria, Egypt or something like, but in AI where, or just like pretty much anything that you can think of in human, or like a university where, you know, there's all these different specializations. And, and, you know, if you try to bring it together in one place, like, then, then that creates, like, a smarter AI. And I guess, like, so one thing that I was curious about is, let's say that, you know, Bent TensorFlow, so, so I actually don't know how the, the focuses or, like, how the different subnet, the, the topics of the subnets are, are decided. But, but let's say there's something like one for health. And, like, like,

let's say there's multiple different models that get submitted. Like how does it determine either, like, which model or which parts of the different models are correct? Because especially that is something that is very tricky and changing a lot. Like if I think that people who listen to the show regularly will know, I've been dealing with a lot of health stuff. And if I just think about the years that I was going through this whole health journey, there are things now that people talk about that at the beginning of my health journey, you know, people didn't. And so just in the last few years, like if I just think about the way people talk about things like microplastics or PFAS or PFS, they call it, which are like forever chemicals or like the gut microbiome or Lyme disease, like there's so many

things where like the understanding has just changed, like even in the last like year, I would say, or at least the way it's been disseminated, at least has changed. I don't know how quickly, you know, the knowledge or how far back the knowledge got developed. But the point is, that how does how do these models know because like if they're using old knowledge like then they're checking against something that's outdated so anyway yeah if you could explain that yeah happy to take this one so there's actually this new subnet that just emerged on bit tensor called safe scan and it's essentially incentivizing minors to develop like skin care skin cancer detection models right So that is the core focus of the subnet.

Now, the subnet owners are responsible for developing this evaluation rubric. So it's their job to determine, like, what should the validators be looking for within the models that the miners are creating? And how should they rank them? So validators are ranking miners outputs from a scale to zero to one on every subnet. But going back to this SafeScan subnet. So it's the subnet owner's job to continuously be improving this incentive mechanism to achieve the outcome that it wants. So like going back to health care and how there continues to be this advancement in understanding of how to approach a specific illness and so on. So it would fall on this subnet owner right now for them to keep this incentive mechanism.

you could say modern or outputting what the market wants, right? And so the validators and miners are purely really doing whatever this subnet owner is intending them. It's kind of like, you know, the founders of a smart contract application, right? Like they're writing the code and then the validators are going out and like running this code. Same thing within BitTensor. And so going back to like any of these training subnets, the subnet. the subnet owner will choose or they'll dictate how these validators to what data set they benchmark the models on or you know what specifically they're looking for in terms of outcome and there's this process where how a lot of these like training subnets work is the miners upload their model to hugging face this open source kind of repo the validators will download it they'll run the validation according to how this subnet owner has has like specified. So maybe they're benchmarking against a well-known skin care detection

benchmarking system, right? And then they're basically evaluating all these models relative to one another. And whoever, whichever minor performs the best according to what this incentive mechanism is specifying gets the large share of rewards, the larger share of rewards. Right. I guess like then, you know, how do you know that the incentive mechanism is like specifying the right thing. But I guess, like, is there any kind of feedback where people can say like this didn't help me or this helped me? Or is that, I don't know how that part works.

So at the end of the day, like the market should be that feedback mechanism. So with detail, the dynamic token will be needed to access the whatever commodity this subnet's producing. So going back to this safe scan subnet, if there is no demand from like actual users to utilize these models, like if these users identified that these models, like say it's doctors or, you know, individuals who have like explored or like tried utilizing these models and I found them not to be helpful, then you should have this feedback mechanism that says, okay, the outputs here aren't actually valuable. Yeah, but like how do you know that? Because they're, they're just requesting information, but then, like, is there a way to say, like, this helps me

or it didn't help me? It doesn't have that direct. It's kind of the same way where, like, if you were using, say, like, Katjibb or Lama 3, right? You ask it something, you don't have any certainty that what it's told you is correct, right? Like, you're relying on people who are advocating for the model, right? So, like, most likely you aren't going to go directly to a model for health care advice, right? You were going to go to a doctor who might be using some specific AI systems, right? And so that doctor would hopefully be doing that diligence onto those, like into those models.

But in general, like, and so take like an image generation model, right? Evaluating which model is producing the best images is subjective, right? Like all these, all these subnets, all these models, like there isn't this objective ground truth in like, like which model's the best, right? Like even if you came back to like skincare detection, maybe like that one does have a little bit of like a ground truth. Like either yes, that person does have skincare or no,

that doesn't. But when you think back to like text generation, model image generation, music generation, like there is no clear thing that you can point at to be like this model is the best, right? So that's why you rely on these subnet owners to do their best. So they're putting their own subjective kind of like reasoning

into what they think is the best. And then it goes to the market to decide whether the subjective inputs that they put into that incentive mechanism are producing a model that is actually usable, yes or no. Huh. Okay. Okay. So it, all right.

So it basically, at a certain point, there is human subjectivity. Yeah. Well, so one other thing that I wanted to ask was, and a hat tip to Seth Bloomberg at Masari for this question, for models that get big enough on bit, BitTensor, like, what is the incentive to stay there? Because they could just go off and make their own chain and then they can, you know, have basically, I guess, a bigger payout or more of a reward

in terms of the tokens because, you know, then they can kind of create their own little universe. So what would keep them on bit tensor? Yes. Seth is a good friend of mine, or he was a good friend of mine until he posted that tweet. We no longer talk anymore. I'm just kidding. I'm just playing. But, no, he made a really thoughtful post there. And this kind of gets into the complexity of BitTencer. But the question that he posed was like, okay, if there's a specific minor that's a subnet that's gone really large and this subnet owner wants to spin out into their

own crypto network outside of BitTensor, like what's stopping them from doing this? There's two things. One, they don't own any of the intelligence that's actually being produced, right? The miners are the ones who are providing the intelligence. And then the validators are the ones that are providing who are distributing the models. So within BitTensor, if you want to utilize the actual model or output that the miners are producing, you have to go through a validator. Like the validators are the gateway of the networks.

They're the ones who are able to build the front ends. They're the ones who can relay a user request. to the miner. And so if you think about it, if a subnet owner develops this really successful subnet and now they want to go spin out, they can definitely do that. But if the miners don't follow them to their own crypto network, then they've lost out on essentially, you know, all the intelligence or all the actual value. And if the validators don't follow them, which have, which is essentially the demand side, then they're also like starting from scratch. And then the the other critique that he made, which was pretty valid, was, you know, what's stopping a miner

who gets very successful from also going and just like launching their own crypto network? And this one to me is pretty easy to push back on. It's the same way that like a really good gamer doesn't, if someone's a really good gamer, it doesn't mean they're a good game designer, right? So like, take a minor who is extremely proficient in machine learning. That doesn't mean that they have the skill set to go design a protocol, which involves designing incentive mechanisms, dealing with cryptography, having to manage a team, interface with exchanges, lawyers. And so, like, BitTensor offers somebody with a very niche skill set who just wants to, like, focus on AI, the opportunity to do so.

So it's not saying that some miners will spin out and go do their own thing, but just the way that BitTensor is structured, it's almost like this unlock where like JJ was saying, like say you are this developer, Open AI or Anthropic. Like I've talked to some subnet owners who know miners that have day jobs at Anthropic and Google DeepMind who mine at night. And like, that's just a way for them to make some extra money or it's just a hobby of them to go compete on these networks. Nothing. It's like they don't want to go start their own crypto network. Like being a founder is not easy. And so that's, I guess, how I would push back. Okay. And then just to understand when you were talking about how it's harder to go off and start your own thing because you'd have to take the

miners, which are the people that, I guess, create the models. And then you would have to take the validators. It reminded me that I was also wondering, so I guess the validators are providing the front ends. So does that mean for like when people use bit tensor, it's sort of like in the background, but they're interfacing with all these different URLs and like different what to them they think are like different AIs. And they are different AIs. It's just that they're all connected in the background. And the user doesn't have an awareness of that, that they're connected by this incentive mechanism in the background, which is bit tensor. Exactly. So ideally these validators should be developing businesses or applications that give the actual end user no idea that they're using

a crypto network. So it's the same thing. There's really, there's a lot of similarities to deepen, right? With deepen, if these networks are just providing a better service at a cheaper cost, and you don't want your end user to have to care or know that they're using a crypto network, right? So I just take like helium, for example, or a geonet. The idea is that there's these front ends that the actual crypto aspect is all just like in the in the background but they're able to like you know get a cheaper service through utilizing these crypto networks that's kind of the same idea here with bit tensors that these validators would build these businesses and provide these services without anybody knowing they're using you know a subnet

or plugged into a subnet. Okay so one other thing that I was wondering about is if I'm understanding this correctly, then it's quite expensive to start a subnet on BitTensor. It looks like the current cost is about $1.6 million in Tao at the time of recording. In March, it was about $6.7 million. And I wondered why it was so expensive and how you think that affects how BitTensor will grow or what it will become. Yeah. So I'd say that's more of a bug than a feature at this point in time. So the reason, well, it's also like a good signal, right? There's high demand for these subnets and so people are willing to bid up the cost. When detail passes, the current plan is just to open the floodgates, like remove the limit to or the subnet cap, right?

So essentially you should see the price of launching a subnet go to close close to zero unless there's a, you know, flat rate that gets implemented at the protocol level. And wait, why was there a cap in the first place? It was just part of kind of like scaling out the network. They wanted to kind of gradually increase the number of slots because, I mean, this whole subnet aspect is less than a year old. And there's already right now. So initially the first cap was 32 subnets and they just increased it to 52 and we're at 50 subnets.

But there's also like a more detailed reason for why there needs to be a cap at this time. and I won't go too much into the details, but going back to how I mentioned that there's this DAO that has to evaluate each subnet and determine what percent they think that the subnet should receive of the emissions, you can quickly see if there's hundreds to thousands of subnets, no one's going to be able to do an effective job

at diligence in each subnet and voting on what percent they think it should get. So once this switches to a market-based system, you kind of just leave it up to the market to go, you know, there's incentive now to go into the deep corners of BitTencer and find a subnet that could have potential value in the future. You buy that subnet token. And so it's like these capitalistic forces that then force people to go and like explore subnets and let that do the whole emission distribution process. Okay. And the last question about the blockchain aspect.

As far as I understand, this is built on Pocodot. Why was that particular technology chosen. JJ, you might have a better answer for this, but initially, yeah, I know that the foundation wanted to go down the like Pocodot relay chain route. And the substrate SDK was just very accessible and easy for them to work with. Later down the line, they realized that, you know, being a real, like a chain within this whole relay ecosystem was slowing them down. down too much. And so they decided just to stay with the foundation of substrate, but essentially

just create their own L1 using it. JJ, I don't know if you have more insight into that decision, but yeah, that's a pretty deep technical question that I think the founders of the project would probably be best situated to answer. Pocod has what's called the substrate framework, or I think it was recently renamed to be called the Pocoda SDK. So that's basically the code base that allows developers to create their own L1s. Pocodot also supports creating parachines and sort of like offshoots of the Pocodot network, which Bitsencer is not. BitSensor literally just forked the Pocodot SDK and built their own L1 on the same code base

with lots of customizations. And so the codebase is actually quite different and not held like within the Pocodot ecosystem or community as like one of these side chains or parochains. So it's its own unique L1. I think the founders evaluated a few other. frameworks that offered different trade-offs and they considered Cosmos and a few other blockchain development frameworks and ultimately went with with Pocodot for first like I think what Sammy was mentioning like developer ergonomics and extensibility and and you know different

reasons that maybe yeah the founders would probably be best best to answer I don't think it's entirely out of the question over time that the network gets completely upgraded again to support sort of a very customized code base that is not based on the pocadite SDK. But like all things, you know, you build on the work of others. Open source gives you a lot of this leverage. And so I think in terms of speed and launching the initial network, three, three and a half years ago at this point, they decided to build on the poker. com, which, you know, saved them a lot of time.

And they didn't need to write the whole underlying blockchain code from scratch. And so we've alluded throughout this episode to the rollout of EVM compatible. which I guess is happening sometime around either end of this year or early next. Update me if that's not the case. But in addition to the launch of detail, are there any other changes that you expect would come to BitTencer with that launch? And yeah, I just mean with the smart contracts, I didn't mean the detail. Yeah, I think the addition of smart contracts and making it possible to add the

sort of like defy programmability that Ethereum has to Bitensor. We'll come with some, probably some pros and cons, you know, some cons in terms of maybe manipulating and creating these kind of derivative instruments that wrap and extend the tokens for the subnet, the DTal alpha tokens, in ways maybe that the network didn't initially intend. I mean, the sort of the smart contracts that Ethereum introduced with solidity are very, it's computationally expressive and you can kind of do anything, but it's very oriented around the idea of an address and a contract for some type of financial transaction. Whereas on BitTencer,

there's actually a pretty clear reason why smart contracts were not designed from the beginning of the chain because this notion of an incentive mechanism, which is kind of like the BitTensor version of a smart contract, is really just Python code that implements what you want the miners to solve for on the subnet. And it's it's sort of guided in different ways of, you know, ideally not being gamable, focusing on producing a digital commodity of some kind that's very additive and supportive of creating sort of like cutting edge, continuously improving AI building blocks of some kind. Could be data generation, synthetic or organic. It could be models that are trained from scratch. It could be models that are fine-tuned, for example,

like you find to in the Lama family of models, that's another type of subnet could be, you know, solving some particular puzzle in terms of scaling out models to be more parameter efficient or to be more accurate on some dimension. And so I think right now, the way subnets are designed is very coarse. Over time, adding Ethereum smart contracts, maybe it could potentially make them more expressive in terms of adding additional sort of financial incentives. But I think in general, like you mentioned, Laura. It's a big moment for the project because there's quite a lot of people in the Ethereum ecosystem that want to be able to sort of bring their language to the Betencer network. And adding support for that is obviously exciting to a lot of people.

So, yep, those are my thoughts on it. So there was a bit of a controversy when someone named News Research, I guess they were working on one of the top subnets. And then they left and are working on a competitor. What happened there? So I guess I can make a couple comments. It's a really talented group of people at New Research. OSS Capital is actually one of their, you know, we were one of their first investors, and they're collective of developers and researchers building open source models,

principally known for building fine-tuned versions of Lama, of the Lama models from META. And we actually introduced them to BitTenzer late last year, and they previously did some work with the network. And I think Technia knew about BitTenzer. some months before. And so the founders actually helped new build a subnet for fine-tuning the Lama models. And that was actually launched early this year in January. So that was, I think, subnet six, if I'm not mistaken. They started mining the network. They were actually one of the top

five or six subnets for a while. So we're earning quite a lot of the emission somewhere in the neighborhood of maybe five to ten million dollars worth of tau per year. And I think for different reasons and there's no, there was a bit of a Twitter flare up on this. And I'm probably to blame. I probably maybe overreacted a bit. And there was some inside baseball that we don't have to go over. But ultimately, I think the technical difficulty of kind of mining and maintaining their infrastructure for doing that on the Tensor just wasn't palatable for them as a team as a small startup at the time. I think they're also still pretty small today.

And so they couldn't find a full-time developer to kind of help them maintain and update the development and the code necessary. to continue mining the network. And so it was like for that kind of principal reason that they stopped mining the network. And I'm actually not certain that they're launching a competitor. I think that there's maybe some interest in exploring other approaches. But for now, they're very focused on continuing to produce open source research and things like Distro, which is a pretty exciting project that a few people in the BitTenser community

found very exciting and interesting as well. and this has actually already been a version and implementation of some of the ideas from distro, maybe not the whole vision and the whole sort of like plan around what the distro technology will look like. Actually hasn't been fully released by news as far as I understand. But yeah, there was just that sort of lack of engineering capacity, continue mining subnet six.

And like Sammy was mentioning earlier, I think this will continue to happen, I think, for a lot of teams that don't quite have the tolerance for the learning curve and the sort of the developer complexity at some point in time, depending on where they're at with their skill set and their team. And so I think I expect that to probably happen less and less over time. There's certainly other reasons why people would want to build their own blockchain or, or maybe explore other competitors, too bit tenser. There's quite a lot of systems that are kind of like trying, trying to build against a similar mission, which is like decentralized crypto incentives for,

you know, offering rewards that that, that, that, that, that, that, that, that, that, that, that, that, that, that, that, that, that, that, that, that's just a lot of validation for, for, for the original, you know, kind of mission and, and scope of BitTenser. We're, we're very committed to BitTencer in terms of the long, long term development of the project. We're continuing to kind of evangelize and share all the benefits with our, our, portfolio companies. And I think maybe there's some chance that news research actually comes back to the network. There's plenty of examples in history where, you know, something was just too difficult or complex for whatever reason. And, you know, you kind of have to re-evaluate, circle back and then, you know, maybe conditions change in the future. And they can kind of

reevaluate the complexity of mining in the next year. So I think we have actually seen some people leave the community and the network and actually come back for different reasons. Bitensor is really, like, last thing I'll say is like, BitTencer is really, you know, substantially far ahead of the other alternative networks that are sort of aspirational working on similar sort of projects and technologies. And, you know, it's, it's really impressive to see the rate of growth. Like a year ago when subnets didn't exist, there was, there was one kind of text prompting network on BitTencer.

Now there's dozens with many, many thousands of miners across the, across the subnets, you know, and that's, I think, only continuing to grow pretty, pretty rapidly. There were just a few things I wanted to build off of that JJ stated. One is like the competition angle. A lot of people within the BitTensor community and outside, like to point at everything as a competitor where I don't think that's very fair at all because, you know, BitTensor, for example, is not a competitor with Jensen or a Koch, right? There might be specific subnets within that are doing something similar to where it could be framed as like a competition. Like, I don't know exactly, I don't think JJ and I know exactly what Noose is building.

From what they've been publicizing, it does seem like some distributed training network. So again, not competitive with BitTensor, but might be competitive with a specific subnet. What was most interesting about kind of the, that distro release was that, you know, news research simply just published this formula on a slide for Descenternet. centralized training. And that was enough for a subnet developer to take that formula, which was made public, and build out an entire subnet around it for distributed training. And this goes back to like something that I've emphasized a lot in which BitTensor being this black hole that will just absorb every open source development outside of it. And so, you know, anything that is

made open source, whether it is another, buy another crypto team or just like a traditional open source AI team, it's fair game to be absorbed into BitTensor, right? So I really do see this like kind of black hole dynamic playing out. And then one other thing just to build off of is just how far ahead BitTensor is compared to other decentralized AI projects. And like, I challenge anybody listening to this to, you know, point towards other crypto AI projects that have have achieved something substantial. And I don't mean this in any way to like disparage anything. these teams because I think there's a lot of like honest and very talented teams that are building in this space. But in terms of like tangible real outputs, I was a research at Masari for two

years. I continue to study the AI landscape within crypto. The only ecosystem I see continue to produce state of the art AI, push the boundaries and provide something that might be even beneficial to an end user is bit tensor. So like I maintain this website where it documents all these accomplishments of each subnet and just to chill my own website, it's called taupil. AI. But within there, you can see all these different accomplishments on subnets that are model-focused, beating out, you know, the Lama models or even like close source AI. It might be on very niche or like specific benchmarks.

However, there is progress being made. So I thought I'd just jump, chime in there with a little putting my bull hat on. and, you know, I was trying to sell why I'm so, why we are so passionate about this project. Yeah, well, you might have to keep it on because I did want to ask you also about a commune. Dot A.I, which is a bit tensor fork. As far as I understand, I guess it's a little bit more focused on modularity and like flexibility, I guess. I won't pretend that I've done a deep dive on it. And there's another one called Allura Network, which I guess is a,

like a little bit more focused on defy. But yeah, could you just talk a little bit about those two and how BitTensor is differentiated from those? Yeah, I've dove into both. So Commune AI is a fork of BitTensor. The founder used to work at the foundation. And they've made slightly different design decisions. One, there's this concept called modules, which basically means you can set up a subnet around, say, like a human being, right? So like I could be a subnet. and have emissions flow to me. You know, I can, I can set up a subnet around an API or an endpoint. And so there's like slight differences there.

And there's also slight experimentation that the commune team is doing in terms of like emission distribution. So it's very similar to bit tensor. However, they've kind of like, they continue to stray in like different directions where bit tensor is going down this dynamic tau path. commune is staying with the existing emission distribution process. So there's just like slight differences there. And then with Alora, Alora is focused more on prediction models.

And so to your point, they are kind of focused more on defy where there's, I forget exactly what Alora calls it, but they're essentially subnets where there's each subnet focused on predicting the value of a something specific. It could be Bitcoin. It can be Ethereum. It can be the price of homes. And so these miners come in and they contribute to something very specific. And how I would compare that back to bit tensor is like predictive models is just a subset of the applications that are subnets that can be launched on bit tensor.

So bit tensor being this more generalized platform where Alora is focused on one particular use case. And so you could frame the argument as more like specialists. versus generalists, right? BitTensor being the generalist, Alora being more of a special tists. But again, like, to your point, I think, so BitTensor was the first project that inspired a lot of others to now explore, you know, developing AI in a decentralized manner. And putting out my bull hat again, like time and time again, you see that backing the category leader is the best move that you can do within crypto. there's a lot of people who feel like they have missed out on the missed boat and they'll go kind of in

these crevices to find alpha or or or tau beta there's this really interesting chart that Ryan walkins from synchristy capital put put out that basically showed kind of like the category leaders within every sector and how again there's been if you study history just picking that winner like you know instead of backing bitcoin there are probably a lot of people that went and like you know started buying light coin and all these other privacy coins and so on and you can make that same argument for Ethereum, right? A lot of people who started trying to get the the eth-beta or even the sole beta. And so while there are competitors, I think I respect what every project is doing, but again, being very objective and pointing to an ecosystem that's

actually producing something of value, I think the only one is bit tensor. All right. Last couple questions. I did also just want to ask, you know, we had talked a little bit about how it is that BitTencer could compete with the decentralized, sorry, with the centralized models. But, you know, when it comes to the chips that are used to train and power AI, obviously, there's a lot of competition for those. And I wondered, like, how does BitTensor, when it's, you know, these more kind of startup-be smaller players, how can it compete with these big companies that are currently buying these up? I can take this one too. So there's two aspects. There's one, there's like this crowdsource compute aspect where you go after idle compute.

And that's where you see a lot of like crypto GPU networks focused on. So looking at, you know, hardware within data centers that's underutilized, right? And offering an opportunity for those data centers to monetize chips during their downtime. So like that's one option. The second option is that like if the economics are right, like it could make sense for, for a data center to point their chips to a subnet rather than towards a user who might pay it for a specific kind of task, right? And I guess it goes back to maybe this is, maybe this isn't or this could be the right

framework, but like with Bitcoin, right? It just comes down to like, is spending the energy on mining worth more than what I can get utilizing that energy for something else? Yes or no? And I think you can make that same argument for BitTensor, right? is pointing these GPUs at this subnet more kind of economically appealing to the owner than pointing at towards something else, which could be training this model, yes or no. And then the last thing is that distributed training is how open source AI bites back against

these centralized AI labs, right? These centralized AI labs can own tens of thousands of A100s. But if you take this distributed approach where, you know, you tap into, you know, one university that might have 10 or 100, A100s, another university that has some one lab, like you could envision potentially, you know, rival. It's the same way that Bitcoin is the largest supercomputer in the world because it has the right incentives. And so there's a possibility that you can create that same kind of world with a training subnet. All right. And last thing, I just had to ask a philosophical question because people are

always talking about the dangers of AI to humanity. And, you know, these are very wide-ranging and vague, you know, like existential threats like the AI could literally try to harm people. There's, you know, the well-known, you know, incidents of bias in AI so it can take kind of like flawed human thinking and just amplify that. There's, you know, risk to like privacy, manipulation, job loss, surveillance. There's so many. So I was just curious, like, do you think? those dangers are still just as present with BitTencer, or do you think that there's anything about this model that could mitigate those dangers that people fear? I personally believe that all of these kind of fear-mongering concerns that kind of emanate

primarily from the large centralized providers, as well as some French groups in academia and maybe folks that are very worried about a Terminator sci-fi scenario happening. are completely baseless and nonsensical. I do believe that technology is amoral and apolitical. It's neither good or bad. It's neither left or right. And technology itself does not pose the risk. Where the risk comes from is how humans apply the technology.

And so the moment in which humans apply technology is the moment we have the legal system enforced in almost all cases. There is unfortunately precedence in the United States making cryptography or mathematical encryption actually illegal. This happened in the early 90s. A lot of people might forget this. It was actually driven by Joe Biden, the outgoing current president of the United States. And so I think that there's reoccurrences of these concerns and fears that a new technology that's really powerful and proving really rapidly on a lot of dimensions can represent

national security threats and societal threats and issues like that. And I personally don't subscribe to any of those views. I think where the threats really lie is in how humans apply those technologies. And so I definitely believe that there should be regulation on certain applications of this technology. And so if you go and build a system that is then used to harm people or hurt people that the application of those systems and the sort of the widespread production should be regulated. But in terms of the raw fundamentals of this technology, what it is at the end of the day is linear algebra and math and doing some computation on some computers. And so I think they're really concerning narrative, both from regulators as well as

the large frontier foundation model companies like Anthropic and Open AI, is that they really want to do what's called regulatory capture. They want to kind of take this technology, create laws that really prioritize the interests of just a handful of corporations and give them this kind of like authority over how the laws should actually be written. And I think that that's just extremely concerning and worrisome because this is the most powerful and interesting kind of development that humans have ever kind of discovered and are sort of refining and improving on, creating kind of this synthetic artificial intelligence, whatever, however you want to call it. And I think it's never been more important to create.

a platform that allows this technology to be built very democratically. It requires everything to be built in the open. And ultimately, you also maybe have a mechanism to give people the ability to participate in the upside and the shared ownership and the shared governance of that kind of system. And so I think BitTensor is really the answer to that. And there really is nothing else that has shown that it has the potential to really present that kind of path. In contrast, we really have a sort of current situation in the industry. where there's kind of two approaches. One is a sort of open source AI, which I'm doing air quotes with my fingers on the video, the podcast here, but the reason I said that is people think of Lama as

being open, open, open source AI. And it's really not. We do not have the training data for a lot of the Lama models. We do not have full permissionless freedom to do what we want with the Lama models. there's a commercial requirement to sign a contract with META after a certain threshold is reached that competes with their business. So many people don't actually know this, that when you reach a certain threshold of, you know, some number of hundreds of millions of users in your application that you're using with Lama, which is a very broad definition, you're legally required to get permission from META to actually use and run the Lama-based models that they release.

And, you know, there's a number of ways of getting around that. And there's also a question of like how much of that is enforceable and, you know, how do they go about doing that and prosecuting people and so on. But I think that there's a lot of nuance to the freedom and the permissionless nature that people can have with these new AI technologies. And it's very multidimensional. It's not just about the code. It's also about the training data and the weights, the infrastructure you used, and also the accessibility. Some of these models require huge compute clusters to actually run to do what's called inference, to run the weights of the model to execute this kind of initialization function that then

allows you to sort of talk to the model and prompt the model. And so the barriers to entry for that are so high today because the costs both computationally as well as with people and skill are not evenly distributed. They're very scarce. And so that's why we've seen this intense concentration of power in a handful of companies because capital is sort of the thing that dictates that access. And so I think what BitTenzer is doing, but the first time is saying, you know, we have crypto economic incentives that can be applied to offering a permissionless

platform that gives anyone the ability to access these things in a way that's actually, you know, funded. It's really interesting, kind of like quote unquote funded, but really incentivized through a naturally inflationary currency, this cryptocurrency that's naturally inflationary, as opposed to the Fiat ecosystem or the sort of the capitalist model where you have to buy the GPUs and fund a company is highly, highly deflationary. You know, there's a lot of people who say this and I'll just mention both AI models and AI compute, which are GPUs, are the fastest depreciating assets in the technology industry. When you go and spend 10 or 20 or 50 or 100 million or billions of dollars on a data center

and, you know, rack and stack and connect all the GPUs in there, you fast forward a year, it's kind of like buying a Ferrari brand new, and then you go and put 10,000 miles on it or something. I don't know in a Ferrari. I'm just giving the example of fancy sports cars, right? I mean, you're not going to get the same sale price in a year. So the same kind of dynamic is true with these GPU data centers. And then from what you actually produce with the GPU compute data centers is the models.

The models are actually even faster depreciating. By the time a new state-of-the-art model is produced, and you fast forward six months or nine months. months. People are already expecting quantum leaps and performance improvement accuracy, reasoning, multimodal data, what have you, new techniques for, you know, the model basically being more convincingly, you know, rational or reasoning more more accurately, what have you. And so what BitTenzer is saying is like you don't, you don't need to think about constructing some sustainability mechanism that monetizes some snapshot of models at some given point in time. Instead, what incentive

mechanisms allow you to do on BitTensor. BitTensor incentive mechanisms, which are these kind of contracts that you write in the subnet, is they say, let's incentivize people to produce continuously improving. You have to continuously improve this commodity coming off of the network. And in order to earn the token, you have to actually beat these benchmarks and continuously improve your output. Could be a model, could be data, could be compute, what have you. And the really beautiful thing about cryptocurrencies that succeed, as we know, know is that they keep going up. They're inflationary. Their value increases. Now, the question that we sort of have unanswered in the world in terms of going from

Fiat, traditional conventional currencies to cryptocurrencies, is the repricing of everything. We don't yet have enough evidence to say, will current industries, whether it's traditional tech, the AI industry, agriculture, like all these places where you can apply a currency to exchange value and coordinate humans, will the world effectively reprice away from centralized fiat-based government-controlled currencies to cryptocurrencies? Will and when will that happen?

Those are two still really big questions. We believe that decentralized open cryptographic incentives are the future, and that's why we're very excited about BitTencer. But just if you think from first principles about the amount of capital and talent that you could aggregate in a decentralized permissionless system, it vastly vastly outstrips all of the big internet companies combined and then some.

And so that's why we're really excited about the potential for BitTensor. All right. Well, this has been a very fascinating discussion. Thank you both. Where can people learn about each of you and your work? I'm on Twitter at Old underscore Samster. Yeah, I'm on Twitter. My name, Joseph Jacks, underscore, and OSSyscapital at O'SKAPT.

as this capital. Thank you so much for having us, Laura. It's a lot of fun. This was love this. Thank you, Laura. It's been a pleasure having you both on Unchained. Thanks so much for joining us today to learn more about JJ, Sammy and Bittencer. Check out the show notes for this episode. Unchained is produced by me, Laura Shin, with a friend Matt Pilcherid Juana Ranovitch, Beck and Gave us Pam Tamadharamdar and Market Korea. Thanks for listening. Unchained is now a part of the CoinDesk Podcast Network. For the latest in digital assets, check out markets daily five days a week with host Noel Atresen. Follow the CoinDesk podcast network for some of the best shows in crypto.