ETH Denver 2025 AI Summit Bitcoin for AI: The Poster Children Catalyzing Decentralized Intelligence

[Music] coin for AI uh or at least analyzing how AI is going to be breaking out of data centers and how we can use like technology much like Bitcoin and like how we could use like early Primitives of blockchain to maybe unlock more benefits from AI so I'd love to kind of kick off this panel by giving everybody a chance to quickly introduce themselves before we like jump into questions I'm Trevor from SA and we give Enterprise companies access to the world's largest network of human data producers I'm Julio from cluster AI we're a decentralized inference and fine-tuning engine and what we are set is to make AI available to everyone in the world hey I'm Paul with render and render is used to uh generate Graphics we are used heavily across Hollywood a lot of the movies you see some component of the movie was rendered on the render Network um things like the sphere the graphics for the sphere Nasa uses the render Network to generate Graphics so widely widely used and uh you know the token for rendering hey I'm Vlad I'm head of BD at Cosmo Capital where liquid token fund based out of Chicago we invest heavily in the bit tensor ecosystem we have a subsidiary called Nakamoto mining that runs a validator on the bit tensor Network and we also mine and do fine-tuning on the network so that's why I'm here all right so I'd like I'd like to kick this off by asking if there is a path for AI to truly follow like a decentralization um similar to how we did for financial technology um like with Bitcoin um or is this going to be harder for like autonomous intelligence unlike for finance or is there like a a bridge between the two um in your opinions yeah I can start so if you look at a grand scheme of things like H AI is a problem similar to Bitcoin in the sense that it is a mathematical problem and it can be decomposed so multiple nodes can participate and and solve the same mathematical problem uh but it's way different in another way which is in Bitcoin you don't have sensitive data traveling to these machines H you don't have an end user who really doesn't care that things are decentralized and for me it is all about value so Bitcoin has a really clear value of why it has to be decentralized and there's value both for people using Bitcoin and for the miners themselves El so on on AI um there's no value really for the demand for it to be decentralized unless there is some advantages of some trade-off that a decentralized network provides H and then for the supply side we can cover that later but there's also about incentives and value for for the suppliers right for a Bitcoin Min Miner it's very clear what value they get if they put a mining rig they they can like create Excel sheet and see like what the value is very quickly and if you're a user of Bitcoin you can also see the value of Bitcoin you can transfer money there's no borders there's no like nothing but for AI it's all about value right and I think we're still figuring that out it's one of the missions of for cluster yeah for the more um philosophical thing about decentralization and and centralization of AI versus crypto I don't have as strong of a comment but for for us at a more kind of granular like tactical level uh We've noticed like we're AI data production company and uh the way that a lot of these these data foundaries as they're called it's like Invidia is a chip Foundry now there's also data foundaries that are producing like new Cutting Edge custom data for these models because all the General open data has already been used so the next Frontier is like new interesting data that hasn't been trained on and so it turns out that being distributed in data production is better so a lot of data is produced in labeling facilities by these big kind of hyperscaler data companies uh that have thousands of people in these facilities producing data but it turns out diversity and decentralization produces better data so if if you have 28,000 dudes from one country in a data center producing data or structuring data you get suboptimal biased results so that's one thing we've learned in terms of why decentralize AI well at least on the data side for our particular use case to build custom data for Enterprise um it does matter to to have the broadest range of of humans the the most diversity produces better data so I have a specific if you want to go no you can go on all right I have a specific question for Paul actually from render now you guys are decentralizing compute for training like what do you see the biggest value of decentralized and biggest risks are in like your field and why did you choose to decentralize it rather than using like a centralized data center or like a compy form yeah absolutely I mean you know the last panel was all about agents and decentralized agents and running those and you know decentralized compute matches that architecture right so if you're running an agent that's on the blockchain that you know is looking it's not being ran in a hyperscaler data center uh you know matching where your computes done running that on the render Network on a blockchain uh that's distributed matches the whole architecture so from a risk perspective from a benefits perspective you know they all match up and um you know the the power of consumer gpus which is the core of the render network is going up day by day right I mean week by week we get more powerful image generation models language models that fit on you know 4090 or something that we see a lot on the render Network and so you know these models uh are going to be state-of-the-art and we're going to be able to run state-of-the-art uh models both for image gen and language generation on the render Network and it's going to be very exciting to have uh the power of decentralization uh be you know accessible to a wide range of people so there are like when when you decentralize AI there it kind of like opens up a a larger kind of area for like attack vectors right like when you're going to like decentralize agents or when you're trying to um have ai transact on chain um how do you guys like think about the complexities that decentralization kind of exposes AI to and like why put yourself through that when you have centralized Solutions in place yeah I mean you know there's a lot of risks in centralized Solutions as well I mean uh you don't if you don't control uh the data center on AWS or gcp I mean there's risk to that as well and um you know by bringing down the cost dramatically we're able to like do security checks do inference checks uh you know run the model multiple times uh diversity of generation there uh is a really powerful security Paradigm that uh by leveraging decentralization you don't have to trust Amazon you don't have to trust a hyperscaler uh the trust is built into the network and can be verified and so I think that's an important uh benefit I mean yes you have to do the architecture well you have to do security well you have to like you know the things we've learned in the crypto Community about how do you build a secure blockchain decentralized project all apply to you know AI on the Chain uh but it also provides some real advantages as well yeah I would say as well that not only in the compute part there is attack vectors it's also like how the model is created like what what is the training set what is the process to train and Clos models uh you don't know you don't really know what's happening behind doors right so by having a decentralized and completely transparent uh process end to end on how to create a model it ensures also that things are more secure and and better to use in production yeah you could be paying for Chach pt7 or whatever the most expensive one is and it's not actually running that you'd have no idea right I mean within the context of bit tensor I think a lot of the security stuff in terms of whether the models are reliable and are producing useful data it comes down to this validator ecosystem so so benser is unique in that it as what they Define as validators is validating quality of work so one of the things that um a term that's starting to be used now something I call proof of useful work so kind of a play on bitcoin proof of work bitcoin's proof of work is kind of arbitrary math solving and within the context of like AI it's actually solving like useful problems to like that people need an answer to and so the validator ecosystem within bit tensor the whole job here is to do a good job of validating the output of the people mining the subnets yeah so that kind of like brings us to the question around like incentives right most people are doing this in order to get incentivized in some way how do you make sure that the people that are coming in aren't really here to farm that that's a huge problem across like most web three companies that you introduce incentives and that becomes like malaligned very quickly so how do you kind of make sure that whether it's incentivizing data or labeling or compute or agent production whatever it might be across the AI stack how do you make sure that those in incentives are truly aligned and like you're incentivizing good behavior not bad behavior yeah for for us that's the question so the question of can you produce uh quality data in a distributed model was the hypothesis that we started out it while maintaining quality and the problem is exactly what you said is people will Cil attack the network and they'll if if if our customers are paying for humans to do rhf on data sets and we have Bots doing it they're not going to be very happy so we have to have a good authentication but B uh we don't have like hierarchical QA we have to figure out a way to do it in a distributed way that's automated and so that was like the problem set that we started with and we started kind of experimenting with different techniques some kind of more traditional stuff like mean reversion and anomaly detection and we did some stuff with like probabilistic modeling where we gave the same question to thousands of different labelers and we could kind of build a distribution over who we think is being accurate with the ultimate goal of giving them real-time feedback on how well their labels are um but then it actually turns out we were looking too far back like right in front of us is the best consensus crypto incentives and and consensus mechanisms like stake like we have staking now so people can they have the option to stake while they're playing to earn more rewards but they have the threat of getting slashed just like that's how how Bitcoin and and proof of stake and everything else gets secured is there's there's penalties if you don't follow the rules so it's it's a it's a non-s solved problem at least in data production where we are but it's it it is the the hypothesis we started with how do you ensure quality in a distributed model um so we're working on that yeah and in cluster for example that they we have an incentive for people to bring their own gpus right so if you're like data Center imagine you have H2 200s and you know that you can put them on contract and get $2 an hour for for those things so what is the incentive for data centers to say I'm going to participate in this decentralized network right like sometimes doesn't make sense for them so we have a program where um every week we we talk to all of our suppliers and we tell them like this is the number of cars we need and these are the price that we're willing to pay for and they look at their excess capacity and then they assign that access capacity to our decentralized Network and that way we we you know it's a win-win situation because we pass the cost savings into the demand and there's an incentive for the demand to use cluster instead of other providers because they pay less for the service right so everyone is incentivized and that's how we fuel the marketplace yeah I think in some way it's it's a it's almost like a venture capital feedback loop problem because you know Venture capitalists have the same problem it's like okay how's my startup doing you got to look at like you know how much product Market fit we're getting eventually it's like revenues things like that and so I think like within the context of something like bit dener the network is just upgraded this model called Dynamic to which gives each of the subnet on the network its own tokens so now you have like kind of to be the money of the network and every subnet token be kind of like an investment a comp a startup within the network and so that model basically allows the investors and the token to decide okay how are we they get to vote on how they direct future emissions of the token and so that provides a financial incentive to like say okay like I'm going to reward the stuff that's actually being quality work Beyond just the validator ecosystem itself self doing the same thing the other example I'll give is is another uh is a data startup called vanana they've had this problem ever since they launched they launch these things called Data Dows and of course it's being botted and farmed uh I think there there's an interesting idea of like rewarding data Dows at the point of the data being used instead of it being contributed I think there's probably some technical work to be done around that but I think it makes a lot more sense to reward the contributors at the point of their stuff actually being used to train AI models things like that just one super quick Point like the points programs and these kind of incentives get a bad WP I've kind of come around though that um they're actually really powerful I think having a token having a way to give your users uh ownership of the network early on and incentivize them to use it so you can kind of get out of the the chicken and egg problem is one of the most legitimate things you can do in a in a crypto network is that you have an advantage where you can do what PayPal did and just bribe people to sign up uh by paying them but you can do it in a more ownership incentivized way and um but that's only part one part two is you got to keep them yeah but but but that's okay like if if you can get a lot of people in the door and they're leaving you're halfway there but it's it's better than nothing and it's up to us as as the founders and and engineers in the company to actually like figure out how to build products that they stay but it's a great mechanism to get people to come I am I actually kind of really believe in what you were saying before about like how do you only pay people out once something is used and we're building a an agent project where what we're trying to get contributors to contribute towards like specialized vertical agents and the only way that their contribution has any value is if it get if the agent gets hired and that contribution is used and we have to Value attribute back to that contribution and then do a retrospective payment right but the current system of pre-incentive ising folks with like incentives ahead of time gives them the bad like it gives it a bad rap because at the end of the day they're just giving it for usage and then they can kind you know put up Bots or you can sible attack it or you just spin up as much garbage as possible and that's a huge problem for AI because garbage in garbage out and that's like a structure that we're all seeing right now and I think that's like a big pain point that needs to be solved I know Anna from Vana and that's sort of like a a problem that they're kind currently facing as well so definitely um now when you think about incentives one of the like questions that I've been having in the space are like we've got established systems right that we've been using staking retr funding what are some ideas that you guys have around like new innovative ways that we can approach incentive design um in the space so I'm I'm big on um Distributing incentives uh distribution to the community so again like I think bit tenser is doing this cool with detto the other one that's like weirdly relevant here is if you guys really like the dgen token on farcast So What DJ did that was interesting was they said like we're going to do this token we're going to give a bunch of community members the initial allocation and then we're going to do seasons where we release more tokens but those tokens are going to be released based on people tipping posts right so you you post on farcaster and you get a d gen tip and everybody a dgen holder gets a certain allocation based on how relevant they are to the ecosystem how many tokens they have but it's essentially like a proof of contribution of proof again proof of useful work that's entirely reli on the community deciding on what's a good what's what good output is what useful contribution is so I like that approach a lot I think that's the the key is like thinking through what metric you're trying to optimize for the the old school example was when uh blur designed their points program every other nft Marketplace was optimizing for volume but then you had people react to that and just do kind of wash trades and then it wasn't real traction and so Pacman's like well let's do something that's more interesting and harder to game like liquidity provisioning so it changes and then that gets gamed and stuff but you kind of always got to be thinking about what what metric you're trying to optimize for and what's the most sustainable um growth channel for your protocol awesome I'm we're at time I did have one last question and just a quick thought on like outside of trading and agents kind of like transacting on chain where do you think web 3 is going to actually start adopting AI I think uh one of the cluster missions is specialized models like if we not talking about agents but if we see AI like taking over the world uh running every single AI operation on a huge model that consumes power and resources is not sustainable in the future so smaller models are really specialized at good or perfect what they do I think that's the future and we're going to see that being adopted a lot in web Tre yeah I mean we we' seen it already very much in on the render Network where you know our bread and butter before was use blender or something you use distributed gpus you get the output of that it's very easy for that to become you know standard diffusion or another Leading Edge image generation model it's just a super natural movement and so yeah no it's it's a very natural next step for our Network and we're seeing big adoption of that already so yeah that's that's here today um I like the idea of like uh AI being trained on blockchain data like with that's the easiest source of public data there is and the easiest way to attribute data to a certain person and reward them for it so I I'd love to see more projects just leveraging their AI models just Based on blockchain data thanks everyone