Building an Autonomous Web3 Infrastructure From Secure Data Exchanges to Decentralized Computing

hello eth Denver how are we doing yeah there we go good you are the champions thank you for staying up until the last session or one of the last sessions it's the end of the day so we're going to try to make this fun and we're going to try to bring you pretty good content here so let's get started uh I'm going to introduce the speakers so I've got David Minaj he is the leader of the team that built the first framework for developing multi-agent systems in the blockchain space he has uh he's the CEO and co-founder of Valerie and a founding member of Olas we've got Sydney Li the CTO and co-founder at G uh G is focusing on Distributing access to Ai inferencing and she's focused on developing adoption for infrastructure tools and has built developer funnels for eight tools engaging over 2 million developers nine years on blockchain that's a long time uh and she has contributed to consensus mantle filecoin the graph and Ox protocol and finally another David David Snider the co-founder at lead protocol which is the first distributed Key Management Network currently securing more than 100 million in crypto assets and 25 million uh in encrypted data assets lead protocol is a key enable key enabler [snorts] uh of user controlled AI agents and prior to LIT David was a co-founder at a software as a service company acquired by LinkedIn and uh your moderator me Paul Lansky is the founder and CEO of dab node and we have more than 1.4 billion tvl in ethereum validators validating on our free open source software that allows people to run web three infrastructure from their own homes all right are we ready ready let's set the stage so we're going to be talking about the centralized AI but I'd like to know what is wrong with centralized AI I mean as far as I know it works pretty well right anthropic open AI um what's what's wrong with it why do we need the centralized well I think one one point when we need it is uh when we want to take control of the AI so when we want to fully own it as a user understand what it's doing um maybe bring our own biases to it rather than uh consuming someone else's uh uh product um and um I think it's a bit like you know control uh like in in crypto we're all about controlling our keys and um with this concept of decentralized AI it's about maybe like kind of controlling your own Ai and the data that goes into it and the the models that go into it so I think think that's one way to answer it but I'm sure we have others another uh way like there's a big talk of alignment in the context of AI uh largely it's concerned with the ethics of the machine intelligence itself um but for me one of the main concerns of centralized AI is the the untalked about part of alignment which is to David's Point who controls these models and we can look just as recently as kind of the unfolding around social media in 2016 and the splintering of uh and bifurcation of society into multiple camps because you had uh commercial companies potentially with some government nudges in the background basically driving to an outcome of let's keep people scrolling on this page Doom scrolling by any means necessary because there's a profit incentive to do so and the externalities of that were quite significant and so the fear is in a centralized AI world we have that but with order of magnitude larger negative effects I would say for us one of the most important things about differentiating from centralized AI is that for True autonomous AI applications it truly requires the centralization if you have a centralized AI it is really mimicking artificial intelligence you already see this right now with I'm going to say it with your Alexa your smart home device oh you've ran out of toilet paper let me deliver you some more more toilet paper that is not true autonomous artificial intelligence it just understands your purchase behaviors and not even in a dynamic way because it just sees okay this month it's time for your toilet paper next month it's time for your toilet paper but it doesn't actually understand did you actually deplete all of your resources so for us the obsession is to have true autonomous nuanced artificial applications it requires decentralized knowledge domains yeah all right um so there's there's a um a quote from uh Andre karpathy um that resonates with pretty much what you guys said and in crypto we have this um not your keys not your crypto and Kathy mentioned not your weights not your brain in this in the sense that if you don't control what's in the AI what's what's like the AI itself the llm the model if you don't if you don't know what goes in there um it's not necessarily working for you you don't know for whom it's working and I'd like to talk uh about trust now because um I think the trust is like a point that we we are very familiar with in in web 3 um but in AI what are the trust assumptions that we're getting from the centralized uh the centralized actors what are the trust assumptions that that that we have I mean it's fully trusted when you type something into open a you know you're even to what model is being used you're just trusting what's on the front end there's the the entire stack is trusted [snorts] right so in the central isai how does that change what what what assurances do we have or like what you you Central the centralized nowal in the centralized for Assurance there's there's two types right number one is when you look at how ownership has been done in centralized AI you'll see the different um let's say like news sources that have sued open AI for scraping vague and vast models to take that information one of the things that we're tackling at guia is understanding how do you upload your own data sets we get a lot of requests from let's say law firms we have a lot of requests from universities uh even those who specialize in specific data sets on on training what this looks like like what this looks like on practice is if you use the analogy of YouTube what our framework is doing is you'll see the analogy of YouTube they are building applications white labeling YouTube content videos and monetizing on their own domains what we enable is for different applications projects who have their own specialized knowledge such as law firms universities to be able to build businesses on top of that we can all also talk about the different trusts and currently what teams are currently doing with like te zkl and how we how we tackle those uh but I also want to make sure you have a chance to share your perspective as well yeah um no they were great points and I think like for the audience like it's important to pick apart maybe the different layers of the technology stack so obviously you've got the data at the beginning um you know most of those models are are trained on a corpus of open web scrape data plus like proprietary data and then you have like the training run and which is hugely proprietary in most cases and then you have the model itself like with the weights um and that can sometimes even be open right there's now quite a few uh cases and um and then you have the point of inference right so when you kind of basically calling the model you're prompting it as as we all probably did in the audience um and uh you get a response um and then the question is like where is it running right like and what's the underlying infrastructure there like is it some some centralized provider doing this who I trust that they're running this thing like David was saying or um is there maybe like some like kind of smart technology let's say like in simple words like eli5 right like so ZK technology or some tees or something like that if I may yeah just Just One Last point and I'll hand it back to you and then like after all this you have the actual application on top right like like in most cases the application isn't the model itself it's like something built around the model so then it's like okay is it an AI agent maybe in Social which tweets somewhere right so then it interacts Maybe by Design with some centralized platforms like Twitter um and maybe with some decentralized platforms like a ledger where it maybe has like a wallet right and so then you have the question around like okay how does that interaction happen where does this system even live like is this agent something you as the owner you're the owner of like you have it on your local device it's open source software you can modify it see it whatever or is it something which maybe someone else offers as a service to you and so there's many different layers and it's actually like we have to be very concrete I think to to really pick it apart but yeah back to I think trust in in decentralize AI is a very big conversation because what I've seen for example with lit protocol the question is how do you trust by giving keys to an agent because I think with the last few years in this space Trust of your keys is to another entity or another cooperatively owned entity and so now how do you provide that trust let's say from a key management uh situation we talked briefly about uh te but te is commonly you have to trust your Hardware as an example and it doesn't scale very well if if Hardware is outdated uh then there's a question of on Prem which we were talking about so if you're running these models locally and you have your own access to your own Keys excuse me your own weights then another concept is do you trust yourself and some some some developers or some people like I don't trust myself to build this own I actually rather just offboard it to you and you can just handle it because I don't I don't trust myself with like losing my keys or whatever the case may be and I guess one other comment here in in the uh open vers closed setup we talked a little bit about like what does a malicious centralized AI look like but then there's also it's also worth considering what's the opportunity costs of building the new economy in the context of a bunch of gigantic centralized operators and to me it feels like in that direction we miss out on what is potentially the most groundbreaking of this technology which are these vast societal upgrades things like a decentralized collective intelligence that could functionally be really owned by nobody but everybody can contribute to and those type of applications which are still very much on the periphery like the infrastructure is still being laid to enable those type of things like uh what the team at recall did with lit protocol encryption of encrypting agent memory uh with uh with with Mother Dow those are like starting to point the pathway forward in terms of uh the next level AI applications and to me that's much more of an exciting prospect around we need open AI because we get large scale societal upgrades um rather than we need to make these systems decentralized because of a fear motivated reaction right um so I I want to talk about uh Laton especially like of out of giving giving agency to these agents right giving the the power giving the power to sign things for you um but but I would like to stop for a minute to two things that one that that Sydney said first of all thank you David for laying down the different layers um and then I'd like to dig into one of these layers a little bit with you Sydney because you mentioned like how can we provide our own data to to these to these models right and I think here um is where really trust comes into place when you're giving your data to one of these centralized actors um for their llm to process you really have absolutely no assurance that this data stays yours and yours only but not only these but also you you don't really have the assurance that it's making the best out of this data so how can we deal with this differently how do we create models that truly work for us and not for a an unspecified actor or like or like the owner of the machines yeah I think something that you just talked about which is what does this look like in a practicality standpoint if I do a quick refresher if we look at how centralized AI is currently created it is on the backs of the data that has been created for years the these models are scraping YouTube blogs a lot of these generalized content one of the reasons why the importance of having very specialized data is that if you let's say go to chat gbt Gemini you'll ask very vague questions and it will pull information but you actually don't know is that from a peer research data set or is that from someone who has a blog post that talks like they know that information but they don't have citations that's when you start to see let's say hallucinations as an example now you're talking about how do you build this very verification in the current space in web 3 I can talk about it in one sentence like very generally and I can talk about it in one sentence for Gaia so from a general perspective right now it's a lot about verifying the the the data sets like do are you do you own do you actually own the data sets are the data sets actually relevant one of the approaches that we've taken which is not te which is not zkl is surprisingly still a token stake model behind how do you actually uh verify this is the data that you do own that because this requires a lot of nuance and we believe that uh the current security models cannot handle Nuance very well which is why we didn't take a very concrete security example for securitizing those data sets gotcha okay uh let's move on to agents um so agents it's like everywhere everybody probably everybody here that's planning to hack something will will at some point touch an agent but like what is what are desirable properties of a good agent and what's a bad agent I mean we just we just go on to Twitter and we see a lot of bad agents we see a lot of a lot of chat GPT regurgitated even Bots yeah exactly that's that's difference there's no difference between these and Bots right um but what is a good Agent what are desirable properties yeah I mean I think an agent is just a software product like an uh a good Agent is an agent that is something that somebody wants um it's like no different than what's a good startup what's a good app what's a good software product so uh I know that's kind of like a general answer and then within that um in the context of crypto agents we're kind of seeing two camps we're seeing those of the uh what they would call the internet Cinematic Universe which is a lot of the uh agents that are posting on Twitter and then we're seeing agents that look a lot more like software products which is there are things that people used to do um that were tedious and annoying that are now automated things like um automated yield optimization across various dii protocols uh dollar cost averaging uh across top coins where they used to have a human in the loop to understand data to click a button to make a purchase and the work that we've done at lit protocol with the lit agent wallet is basically like ooth for assets where you can grant an agent runtime some scoped and specific ability to use certain assets with a certain spending limit on a certain contract or blockchain make it as broader specific as you want and you can almost kind of think about this as an app store for agents in many ways where you're saying here I'm giving this agent runtime this set of permissions and then that agent runtime just like an app you install on your phone needs to provide a benefit to the user such that they are going to pay for it deposit their funds in it Etc very curious about David's take here though yeah um I think it's a great question because it gets also back at the decentralization like the answer will be different for everyone right like what's a good and what's a bad agent and it can only really be different if you have like very big diversity in the terms of agents and if we can effectively own our own agents that do things that are very specific to us I really truly believe we we will end up in a world where all of us will have hundreds of agents that in some cases augment our life right so we are there and we're doing something and this agent helps us right there and then and in some cases autonomously do stuff for us in the background so for instance a self-driving car whilst I'm reading something or chatting to someone or like my trading board that does things on behalf of Me Maybe on the ol St with like lit protocol like assurances or not kind of going off the rails right so these are some examples but in all all those cases for them to truly be like good to me they need to be aligned with me and the one big Topic in like AI is always this question of alignment and then people like well how can we align the AIS and the the thing is well we can't even align ourselves like we the four of us won't align on something like let alone the world right so one big I think requirement is that we that we enable everyone to have like really customized really purpose built agents for themselves and decentralized AI in it sort of abstract sense I think can enable that I think this this is what we talk about in terms of going hyper vertical right because how useful is a general model versus like when you go on for example I keep using the analogy of YouTube you don't watch a general YouTuber right you watch a YouTuber that specializes in plants and then maybe you're interested in a humidifier so then you'll go and watch a YouTuber talk and review all these different humidifiers is this same thing you're not looking for a general YouTuber that will talk about plants and humidifiers and you know what cars fixing something like that and I think my easy answer to you which is what makes a good Agent I think it's product Market fit if that agent has product Market fit it's a good Agent it has a user base excellent so I see I see several like threads of what you guys are saying number one is that if we concentrate the brain power of these like agents or these new economies these new AIS if we concentrate all these in like the four main foundational models we're missing out on the benefit of actual decentralized intelligence um secondly agents must be personalized to be good and personalized with our own data uh personalized with our own use cases with our own um preferences weights maybe uh and uh and maybe with uh access to our own assets as well um okay uh We've run out of time but that has been an amazing panel uh thank you very much guys in 20 minutes I hope uh we could teach you something thanks everybody thank you thank you everybody