ETHDenver 2025 AI Summit: Decentralize 1 - Crypto Economic Incentives for Decentralized AI

[Music] the panelists introduce themselves I will start my name is uh James Young U founder of a project called colland and uh awesome hey everybody my name is Sydney Li I am the CTO and co-founder of Gaia hey guys uh my name is Julian Balu I'm the founder and CE of state capital uh we do investment we have a venture Studio and we also participate in build some projects in the defi space in the blockchain space hey I'm Vios uh founder of nura we're a payments and insurance layer for AI uh formerly the first employee at Egan um yeah lovely to be here hello it's kind of hard to hear you guys over there so uh I'm Billy I am the founder and CEO of intuition and our goal is to give people in machines better intuition in all of their interactions by giving them the data they need amazing thank you everyone all right so I'll start off with Sydney uh with Gia how does your platform emphasize decentralized Ai and how does staking mechanisms ensure that the GPU providers and data contributors stay honest so I think there's a lot of conversations about how does token Economics work in the AI web 3 space one of our very bold opin opinions is that we are looking at a staking model for participants on the network to be able to validate their own data and also be able to validate their own Hardware one of the reasons why this is bold is because a lot of AI projects are looking at te and te requires a lot of specific types of hardware and a lot of types of Hardware to be able to scale those types of securitized transactions and because of that te is one of the reasons why we haven't gone down that route yet because we actually believe that it requires a nuanced tokenized model to be able to handle nuanced AI where he really is able to handle binary types of AI applications awesome Julian how do you see token economics evolving to better align interests yeah so my background is actually uh machine learning and I spent quite significant amount of time in uh building some algorithm back in 2010 15 years ago uh we didn't have uh Game Theory back then I I think what's interesting in the U AI in crypto especially in the uh to economics is the fact that you can actually use a staking to either incentivize good weight uh so building LM but you need F or ZK or new type of technology in order to not reveal the weight in order to train new models um and this is the way you can actually incentivize the correctness of the or the vable data in order to get better data or to get better waiting for training more models or to uh get to get more uh better data and this is the way you can actually build um really good economics around that slashing mechanism if your model that you're using is either hallucinating for example you get penalized all the data you providing you also get penalized I think it's not really complicated to the get um good tonics around data verif to verify the data currency um I think it can be theum awesome excellent bias tell us a little bit more about uh nura Ai and what what do you think about uh the economic alignment incentives and what you're doing yeah absolutely um I think like they both touched on like really important aspects where you need these systems to be somewhat um you know operating in like self-functioning mechanisms right we're trying to build like these systems that can handle themselves and so like I look at the agent supply chain as like kind of like three different components right there's the compute layer which like involves gpus FH te Etc there's also like it's the Oracle layer but to me it's like the data layer like Jen just said where it's like you need some sort of actionable information for the agent to execute on that's really important um I think like you can build like really robust you know cryptoeconomic condition around each actor in the supply chain build like proof mechanisms to see who's behaving correctly that's kind of what we're working on um in addition to like a payments and insurance layer to underwrite like some of the actions of these AI models I think the most important thing is like coming down to more like consumer-driven use cases like we're kind of playing this game of like training models further and further where like I'll give you a simple example of dkls or like T like tees being really useful I came home one night to my room in my apartment slightly inebriated it was like three am. and my door was locked to my bedroom right so I started using chat GPT to like understand how to like pick the lock I took apart the door knob I refused to sleep on my couch like I'm trying to get in my bedroom and chat PT refuses to answer because it thinks you're trying to break and Enter but if I could ZK TLS or prove that I live in that apartment I pay utility bills there it should help me right like these are like real consumer driven use cases where like we could Z ktls some of our attributes and like llms could actually start incorporating like useful feedback into our lives but if we just keep playing this game of like oh we're going to train these models more and more efficiently like we're still going to keep missing the plot of like oh every cycle we build more and more infra and like the users will come one day like agents are the way we're going to get users and like if we don't start thinking about it from this Paradigm like yeah I'm I'm not super optimistic about the design space inherently Billy uh what do you think about intuition and how economic incentives through intuition maybe explain intuition briefly and then how do you imagine the economic incentives evolving specifically with intuition and in the context of an AI use case yeah absolutely so where we play is kind of at the the Oracle layer that vs was mentioning kind of the D data layer so if you think about kind of this conscious internet that is emerging like you need kind of the substrate of the conscious internet that allows these things to communicate with one another and in order for things to communicate with one another they need to speak the same language and so we have really interesting cryptoeconomic principles in crypto that are allowing us to come to consensus on the state of a world computer or a ledger um but we're still we're when we use that world computer we're still speaking different languages and so thing even though composability and interoperability is the superpower things aren't really maximally composable unless you have standards and and when and one standard emerges you get you get irc2 and then a whole ecosystem emerges and then you get 721 and an ecosystem emerges so what if we applied cryptoeconomic mechanics to like the State of the State of the systems and we and we incentivize people to come to social consensus on the language that we speak like the semantics of the data so instead of coming together in a standards committee what if we incentivize convergence on data standards through economics where the economically optimal choice is the one that is what the the the social consensus has decided is the way that we should structure the data and I think once we do that then you get intuition because you can just very easily understand what things are because you're all speaking the same language and you have a bunch of data about everything and you just unlock this very seamless flow where agents are interoperating with one another it's like user to agent agent to agent agent to platform user to platform all those things you need data about each thing that is in that kind of supply chain I actually want to talk on Billy's passion about this because I think a lot of times people forget why AI or why AI web 3 even needs tokenomics and the reason is because the next biggest consumer of blockchain is going to be AI applications blockchain is written by code AI is written by code for the last 10 years we have seen humans try to interact and there's a lot of consumer frustration right in the front end and so not only that but also the next biggest consumer of blockchain is also the data itself right so AI is able to be able to process that data Beyond far uh you know much a faster throughput than humans can so I I see your your passion and I just really want to remind everyone why we truly believe the marriage between web 3 and AI is because again AI will be the next biggest consumer of blockchain I think we can even double down on that right like there's also this beautiful idea of potentially you know intuitions providing faster finality from the oracal layer so now like you're dealing with execution at a faster rate and you're reducing latency right because you need the dat data layer whether it's an oracle or like whether it's web 2 Data whatever you're trying to do to go to the model generate an inference and then the agent acts upon set inference right and so it's like if you can provide faster intuitive data then you can provide faster execution it's going to be this like as long as the GPU networks are properly like scheduling their um you know like at a kernel level they're properly scheduling what's happening from a early exit standpoint for determinism but that's a whole different like can worms yeah and it's I I think it's it's efficiency but it's also kind of personalization it's but personalization at every layer so when I go to interact with an agent I want it to have like my intuition so it can personalize kind of its responses to me when it goes to book a restaurant reservation and then when that request goes out you're going to have these agentic swarms that are all coordinating which agent does that request get routed to now you need agent reputation and agent data uh and then that agent goes to interact with a platform which platform does it use to book the restaurant or to do the swap or whatever like now need platform reputation so you kind of need this like interconnected web of data that's very easily traversible so you have like good intuition about all those things so that flow can happen seamlessly it's not inherently clear that we need reputation if there's enough Economic Security at stake true like those two things like they get messy together when you like couple them it might be more El like again we're talking about an open design space so like there's going to be multiple different options of how people execute on this but I think like there's going to be coupled models where you see reputation and stake and then there's going to be very decoupled models where it's just like hey like you're liable for every single action and here's the like underwriting of set action and that's the risk that you're taking on absolutely I I 100% agree and I think they kind of like you can take reputation is one word you could use to describe it it's also kind of like the interface layer it's like how how should I interact with this thing how should I interact with this person how should I interact with this agent how should I interact maybe it's not even the thing that slashes you like maybe instead of your reputation getting slashed you're economically getting slashed but still need data to like be able to communicate it's like the communication layer I I think all of this comes down to the context layer as well which is why when you were talking about your frustration with chat gbt it was missing context if you're venmoing a friend you're putting context as to why you're even venmoing that friend and I think being able to automate a lot of these workflows requires actually quite a bit of context and nuance and so even for us like coming from Gaia our whole framework is about being able to pull up the right context for what your agent business needs we're not necessarily an agent launcher so for example if you build your application with Eliza we'll help you decentralize it contextualize it and then also pull the right verticalized LMS to kind of be the brain of that I'm going to put Julian on the spot a little bit so do you think based off this conversation can economic incentives is it require are economic incentives required to help solve alignment well um I think if you don't have um alignment of incentives you cannot really coordinate uh those AI agents to accomplish task on your behalf and also to be able to communicate between each others but you need I think the key in uh crypto AI or AI it's the the fact that you can verify the data I think that's the uh Holy Grail of training new models and also using um interference for new models so the fact that you can guarantee the data you're providing to models or you're training them can actually be unly verified and this is something that I think the the person going to solve this this problem is going to be the one driving the AI bullan in crypto uh crypto space so verification is key I think verification is everything uh uh especially right now like for example Char PT or um entropic they are reaching um a phase where they are saying that it's extremely difficult to type into accurate uh verifiable data because then you start having models that start hallucinating and and also when you uh pull or feed those uhm with new data you can start actually polluting and those type of training so I think if I was a a big competitor in a space of machine learning like I will try to um interact with bad data into those Alm and it's potentially a business where people are starting to do that they kind of poison the training po yeah yeah yeah and in distributed training with uh distributed computing I think it will be extremely difficult for crypto AI projects to actually come up with um extremely good AI models because of that problem of verifiable data the the incentive alignment that you need to come up with in order to make sure that the data you you you having access to uh as to the cost to have access to extremely good data will be going uh exponential in the coming years I feel like yeah I I think there's a couple of other like interesting aspects there right like the data aspect is like incredibly important and having verified data is like a almost a prerequisite but then there's also like holding each actor accountable like if the agent Builder executes on something incorrectly we need to hold them accountable if the compute network doesn't like provide inference within a 12C latency we need to hold them accountable right so like there there's different layers of accountability that like cryptoeconomics provides I think like there's another thing of like the macroeconomic trend that you know we need to be thinking about which is the whole category of decentralized compute is introducing smaller and smaller units of compute introducing economies of scale and driving compute down right that's the whole purpose of the asset class and so like as we're building out AI agents we're also dealing with like the cost of compute being like a downward trajectory a race to the bottom as compute providers compete right and so like we're dealing with already a ma economic Trend we need to build crypto cryptoeconomic standards to like make sure people aren't cheaping out on their way to the race to the bottom in terms of compute and so like liable parties need to be held accountable yeah we're at time and I feel like we're just warming up one more okay one real quick and then we'll end I I like I think verifiability is like a really like it's it's very important but like okay let's say you know I can verifiably prove that the the data was generated from my device at this point in time like even though the data is you know we we can verify that it's quote unquote legitimate is it good do I want that data from that person and so I think reputation has a big role to play in all of that where it's like who whose data do I actually want even okay everyone's producing all this verifiable data but what do I want to tap into based on like because if if you start to containers in and you say deterministically hey this is good data and this is bad data that gets dangerous cuz then you start censoring we're just now getting to the juicy stuff but here we're going to end thank you so much everyone Round of Applause to the panel