Anna Kazlauskas I Decentralized Data for AI With Vana I Agentic Ethereum 2025 Summit

[Music] for this one I'd like to bring on Anna to talk about decentralized data and how do we make that way better by leveraging Ai and having AI lever these such high data so without further Ado let's welcome Anna and uh we'll get started and good to have you back hey yeah great to see you all right we'll get you all set up and we should we get to go in a second uh great okay I think I can get started then um so yeah today I'll be talking about um kind of the role of data in Ai and how you can um access better data through kind of decentralized data um for uh training better AI agents and better AI models oh and I'm uh the one of the creators of vanana which is a protocol for data sovereignty that kind of makes your data portable across applications and allows you to pull it into uh data Dows so you can all collectively own your data and kind of uh contribute it towards AI models uh so AI models are kind of ultimately only as good as their training data right if you um post on Reddit you might have already helped to cont train some AI models uh this is a breakdown of training data of gpt3 but represents kind of a lot of the foundation models large training data sets um most of it is from the public internet so common crawl means you kind of just like scrape all the internet that's publicly available kind of put that um in a data set that then you train on um and this is generally how AI models learn today um and then sometimes there's like a fine-tuning step at the end um something to note here though is that this represents a very small percent of the overall data that exists um because such a small portion of the internet is publicly scrapable so people estimate that that's less than point one% because most of the internet requires permissions or sign on to access um so if you want to use that data to train on it you need to broker some kind of deal with a company that has access to the data um in today's kind of centralized web 2 World um one of the problems that AI researchers have kind of uh started to talk about past year or so is called a data wall which basically refers to the fact that we're actually running out of training data to use to make AI models better so here I'm kind of showing um as AI models have gotten better and smarter um they've been trained on more data um often people only think about the compute side of the equation of like scale up compute but if you're scaling compute then you also need to scale data and so that's actually this limitation that we're running into where llama 3 is trained on around 15 trillion tokens of uh tokens as in words in an AI context of data and the whole public internet um if you dup it and just take the high quality stuff it's a data set called Fine web that's around 15 trillion tokens right so we're kind of at Max Capacity of all the data that we can find on the public internet so the question becomes like okay what next like how do we get more data um something i' mentioned too that I think has been like an interesting angle that deep seek has brought in is that the data wall is both an issue of like quality and quantity and deep seek actually managed to kind of like overcome this data wall through quality so they hired like a team of phds to generate really high quality cold start data um people say like uncon people say basically their founder was labeling data himself and that's just like how seriously they took data quality for that starting data set that then they can kind of train on so when you think about what kind of data you want to train AI model there's both both this aspect of like how do you get the most data and how do you get the highest quality data and to date um before deep seek it's been like a pretty Brute Force approach which is kind of like scraped the whole internet you know that if you imagine being like a kid and you can only learn from the public internet and you have I don't know you're like reading forchan or other public forums like it's not the best information but maybe if you consume a lot of it it's okay um but you can also go the approach of okay from a very large subset of data let's narrow in on extremely high quality stuff and use that as uh this training so that's just sort of to set the ground for what makes good data in training AI models um and how things are done today um companies have realized that their data is extremely valuable um in this AI world and they've really l down their data right so reddit's earning $200 million from selling data photo bucket $1 to $2 an image Apple's buying news data for like $50 million and if you are like a company acting rationally that has very valuable data like often the the thing that will allow you to charge the most for your data is to shut down access for other people right so the internet is kind of becoming less and less open like stack Overflow but and Reddit both of change their API terms um and I I feel like now more and more you run into I don't know these sign on walls where it's kind of harder for you to get access um because those Silo platforms are building bigger walls to protect their valuable training data um there's one group that still has access to this data that most people don't realize and that's actually users so um as a user you legally own all of your data in the same way that when you put your car in a parking lot the parking lot doesn't own your car it's still your car you can take it back any time when you put your data in a platform by like Tweeting or writing a message or posting an image that's still entirely your data when you sign the terms of service you're signing like a very permissive license with them to say hey you can use my data for the purposes of like operating the service some of them have ADD is Clause related to like training an AI model as well but uh sort of the default State before the terms of service is it's the irely yours after the terms of service is they have access to it but it's still yours and you can always go and get a copy um so if you use like Alexa in the past you can actually go export your Amazon data and like hear every single audio recording which is kind of like a creepy experience but I think it just shows like they have all this data on you and it's yours you can always go take it back um Instagram is also kind of a fun one because you actually get back both your like posts messages photos those but also the AI labels that they've put onto you um and so you can see kind of like what categories they've put you into um and can be kind of interesting but what's important about this is because each user legally owns all of their data you can actually have users kind of export their data from across different platforms and actually build like bigger and better data sets than any single company can right like if you have say all of apple and all of Facebook and all of Google they're not all going to work together I mean they're also kind of regulations where they can't share your fully private data and kind of connect it across you um but the user is actually able to get more data across those platforms um so that's like this group that often people forget about that really has true legal ownership over the data that's sitting in the walled Gardens um vanana is this self- Sovereign data layer for AI that's really designed to unlock data from those wal Gardens um so two of the core features of vanana one is like non-custodial data what that means is that you can manage access to your data just like your funds while preserving privacy right and this privacy piece is actually really important one like because if you want to put private data and kind of permission it on chain in some way like I don't want my messages or sleep data to all be fully public so kind of just from a user preference perspective second is more from an economic perspective where if your private data becomes public there's not really a reason for someone to buy it from you anymore because it's just publicly available right and so that privacy piece is really important to allowing for um these like economic markets to be able to form around data and then second is proof of contribution and so when you have one person's data it's not that useful for training an AI model because you need a lot of data to actually be able to train you need 10,000 100,000 10 million even users worth of data to be able to create something kind of very substantial so how proof of contribution works is it basically allows a a large group of users to contribute their data to a pool and data is different for everyone right like it's necessarily non non-fungible that's like what makes data data my data is different from your data and so there's kind of this question of like well when we put our data together how do you prevent like a malicious actor from pretending to be 100 people who are contributing their data or how do you weigh like one person's data versus another and so that is the mapping of proof of contribution which is implemented at the data set level so every data set every data out has their own proof of contribution um and if you are a developer on vanana you might already know the term data liquidity pools dlps um that's this more technical term which basically is referring to um the aggregated data set and kind of smart contracts around it um and you might also know the term data which is so from a technical perspective you can have a DLP that's not technically a dated out um so every yeah every dated out is a DLP but a DLP is a slightly more flexible kind of technical term where you can aggregate data that's potentially not managed via a dow um or potentially kind of like operates with a stable coin or something like that so just to give you some examples of dated dows across the ecosystem um here are a few of them that I just kind of pulled up um YK yr is a Chrome extension um and yeah I'll mention this too where when you interact with Vana you will probably interact with an application that's built on Vana that's like the front end that you'll see so yeah yky is a Chrome extension sleep. Fon has you connect your sleep data um Prime insights is having you connect your Amazon data um that verification is actually kind of interesting where you download a a chrome exension that they offer and then you export your Amazon data and it actually generates like a zero knowledge proof client side to prove that the data actually came from Amazon um so there are a bunch of different verification methods that have plugged in like there's ZK email um reclaim has a ZK TLS thing that kind of plugs in um and then there's also a set of tees where you can run General python code to essentially verify data quality um so if you're building um with some kind of like verified tool in mind there could be interesting ways for you to plug into the vonet ecosystem and allow data DS to better kind of verify their data um there's just kind of this uh spec for a proof of how you have to exactly write the proof on chain um but it's uh relatively General um so yeah I think there's some really interesting stuff happening there um maybe one other one I'd call out yeah there's a telegram dated out here um and then nowhere is your your Google maps location data and a lot of this data is very private right and so that kind of like security piece really matters so I can actually talk through how it works when you connect data too so when I connect data to um say nowhere or like the SP there's a Spotify data D called unwrap I'm connecting my data there first thing I do is log in with any evm compatible wallet that's what kind of manages my keys and then I grant access to my data in some way sometimes it's o sometimes it's a Chrome extension or a manual upload and then I sign a transaction to um basically retrieve a an encryption key that I use to encrypt my data client side with a key that I control and so then I write that uh I store that to storage that I control actually so you can connect like your Google Drive your Dropbox um kind of any storage that's self- Sovereign that that you own um there are also some more convenient storage options if if you don't want to host to yourself um then you write that on chain um and essentially have that pointer to your encrypted data then when you're joining the data Dow what you're doing is basically saying hey my data meets your proof of contribution and so then the data is verified um through a te or some by checking the proof um to show hey my data is actually legit it's good to join this and then what you're doing is you're encrypting your data with a key that the data da controls so you've essentially delegated access to the data da to to be able to make decisions about what happens to your data um just to highlight some of the data Dows um the Reddit one has like 140,000 creddit users um a million people tried to join so this shows that kind of proof of contribution is working right there were a million wallets but only 140,000 kind of like real Reddit users who have an account have the karma minimum and now they've trained um a user-owned llm which is primarily just good at chip posting because it's trained on Reddit data but it's kind of like starting point um another one I'd call out is the 23 andme da that actually aims to like buy 23 and me and make it user owned um that is a massive undertaking right but I think that that's actually an interesting example because their goal with a data out is more like negotiation and leverage versus purely like we want to sell our genetic data it's actually we don't want to sell our genetic data um so those are the sorts of things that are possible with data Dows built on this kind of data liquidity pool primitive on vanana um something else I'll call out that I mentioned a little bit already on the Privacy side is that data on Vana is private but it's interoperable right often when you want interoperability you have to trade privacy like if you look at most of the um decentralized social protocols the solution um which makes sense for a sort of like public social protocol is that they make everything public right so like blue sky or forecaster like it's basically like the data is all public um in the case of V vano we are really designed for private data but still making it interoperable because I think a lot of the like information that you want to bring to an application whether it's context for an llm like your past messages photos like even a clone of your voice that you don't want to make fully public right like people want that privacy but they also want that interoperability what that means as a developer is that are two ways to access data on Vana so one of them is individual access what that means is that a user can sign in with their wallet to your application and bring their data that they've already onboarded to different data Dows or different applications into your application so they could bring like their sleep records and um straa records and nutrition records into your like Health AI application and you can then kind of make much better recommendations based on that information um the other way to access data on vanana is through kind of this aggregate access uh basically accessing all of the data from a given data da um and so for example you can train an AI agent on a whole data Dow's data set one example of a data set that's kind of interesting is um fin quarium is one of the financial data Dows that's built on Vana and they have people aggregate their um exports from centralized exchanges and often that data set you don't really have access to right and so like usually you can only see the onchain stuff and then as soon as it's in a centralized exchange you don't really know what's going on um but if you train an AI agent on fin aquarium's data Dows kind of data set then you have access to that um private data that wouldn't otherwise be available so I think there could be opportunities in that case to build like a better AI trading bot because you have access to data that like nobody else has access to um something on the Privacy side I'll mention too is that only the outputs so like the inference or the train model can leave unencrypted which ensures that the underlying data stays private right often whenever you're like letting someone use your data or potentially like working um in a a data sales context um if I give you my data like I it's really hard for me to know that you haven't resold my data we call this like the double spend problem for data and the way that we solve that is that you essentially only work with data in a secure environment and can run General python operations on that data kind of get what you need but you never have that unencrypted data actually leaving the environment um and so yeah if you're interested in accessing any of the data from the data Dows um or building with individual data both of these things I'll say are not that well documented on our docs right now and it's something that we're kind of rolling out this quarter a lot of the sort of fully self-served stuff that's easiest to do right now is building a data Dow more on uh data Ingress side um the other thing I'd mentioned about data on Vana is that it's structured and quality checked right often um when you kind of Imagine This like self- Sovereign like interoperable internet one question becomes like what's the structure of all the data like how do I as an application know even what I'm consuming and what I can build around and so each data liquidity pool on Vana kind of data down more informally is defining their data set structure and their proof of contribution and verifying that all the data that they're writing proofs about is actually of the format that they're expecting so this one uh data schema I've pulled is from sleep. fun which is a sleep data Dow if you have an Ora ring uh you can contribute your data I think they support um a bunch of different data sources too so if you have like a Fitbit or an eight sleep you can also contribute your data um they have a leaderboard it seems like people aren't sleeping that well recently so if you're a good sleeper um it would be great if you could join the leaderboard um and kind of make sleep do fun healthier and sleeping better but yeah this is their format um so then as a developer you know what data you're getting right um and then it's also quality checked so you know that the data that you're accessing has gone through whatever the proof of contribution imple implementation is so you can know okay this is actually good data that comes from users with a real account of at least this age um and sort of have that information because data quality kind of matters so much in in training better AI agents and better AI models um so the best way to get started on bana is to either start a data da or Access Data um something a lot of people don't realize about starting a data out is you can start one for like any data source of you're choosing right because the user has the legal right to go and get their data from any platform you don't need anybody's permission to start one you can just choose a data source and like go and have users export their data and kind of start contributing so in terms of the steps scoping out your data source picking one picking how users will contribute um every data DA has their own data set specific token so there's also a question of kind of how you want to reward users relative to each other Reddit for example does it based on the square root of your karma um the 23 and mi1 I think is treating users all equally um I'm not sure if the spotif One Rewards people if they have like a premium account more um some of the financial ones reward you more if you have more trading volume I think one of them rewards you more if you've like lost more money and I I'm not sure exactly what the rationale is there but it's sort of like yeah basically you're mapping a bunch of users and their data to like how much should it be worth relative to each other um we have a Discord too there's actually a session this afternoon where we'll be going into um some specifics around like data tokenomics um which so feel free to jump into that um then there's a process of building so kind of deploying smart contracts um there's some templates and solidity um proof of contribution which is implemented in Python um and the interface for users to contribute data um on the interface side there are a bunch of different templates you can clone so if you're looking to make like a Chrome extension or a web application um there should be something in the ecosystem that is relatively easy to start from and then there's also um some stuff where people are build build like a unified UI to be able to contribute across different data DS so that's something that's in the works as well um to make it a little bit easier to start uh and then you can launch so you basically deploy your contracts have your application ready um kind of register with the like DLP registry on Vana so then everyone can find your data da know about it know what your data structure is what your proof of contribution is um and then you can also compete um into the top 16 data DS on vanana which are actually earning a share of emissions because they're onboarding data into the network right often these kind of self- sovereign data ecosystem projects which it kind of I think a little bit the original promise of etherium and crypto more broadly like own your data own not just your funds but all of your information across like a self- sovereign internet um really fail to reach like Mass adoption and so I think that it's important to have this way to like bootstrap bringing users in with their data through data Dows so then developers can build amazing applications some of them are data Dows but some of them are also just accessing the data the data Dows have brought in um most of the teams building are like two or three people initial build takes um a few weeks you can probably create a template data da in a few days um and if you're looking for ideas um I've listed some here like I'd love to see an iMessage one Tik Tok medical records um I also posted like a wish list over the holidays uh maybe a month or two ago on dated out I would love to see um so that would be a good starting point to um and so yeah here I've got a link to our docs um the QR code is this specifically the page on how to build a data Dow um as well as our Twitter um the other thing I'd mention is accessing data so if you'll be at eth Denver we're actually hosting um a hackathon where we'll be showing kind of a preview of a lot of the data access and data portability and if you have use case you're already building for um definitely get in touch and we can um kind of like get some early feedback on exactly the kind of like developer endpoints that we set up um for data access um the last thing I mentioned is if you want to learn more about van um a great place to start is actually the white paper um that's Linked In the doc somewhere um and so feel free to dive into that if you want to more understand it at a high level um or if you want to dive into building um these docs are a great starting point awesome uh that was such a good overview Anna uh thank you so much for doing that uh couple things I'm curious about are just understanding um in at least the last six to eight months how have you seen more Ai and agentic experiments on on the the data repositories of Anna and um kind of what's been super interesting to see from your community yeah I think that um the AI agents so one thing that maybe I'd say actually surprise me recently I think today or yesterday someone built a plugin to Eliza where they're letting AI agents actually trade data tokens so like the AIS can buy their own data and that I think is not directly like it's like actually giving AI agents like direct access um so that's been interesting um and then there are other applications um I think more trading applications than I other like otherwise would have guessed or predicted data itself or trading based on understanding and parsing the data from the Dows trading based on understanding and parsing the data from the Dows exactly super cool that have been really interesting to see so does that mean they're taking the data and then modeling and inferring somewhere else and then doing the modeling or is it something different exactly so it's um the two to look at here are data Pig and and fin quarium um and so what they're doing is essentially aggregating the data to teach an AI agent to be a better Trader um and then having that AI agent kind of operate on its own one of the other ones I would mention is y KYR they are they have a Chrome extension where they're kind of collecting a lot of your browsing information better understanding you and they're building this personal agent that's based on how you typically browse the web and a deep understanding and they're doing everything um actually like client side in the Chrome extension so they recently got I think a small version of Deep seek running client side in that Chrome extension so that's been cool to see and kind of preserves the Privacy um how so this might be a dumb question but how ensure the Integrity of the data that's being taken from the Dow to do inference and do fine tuning and then run the models and it kind of does the protocol help with the data Integrity in this way and and if so kind of where does that come into play for these spots yeah uh data Integrity is like almost the most important thing where if you're like working with bad data then you're going to get really bad AI right and so you need to know and have some guarantee around data quality and what you're consuming and where that's enforced at the vono level is through proof of contribution so every data out any data point that comes in needs to go through proof of contribution which is insuring ing a whole bunch of checks on the data so one of them is like proof of ownership does the user actually own it one of them is like proof of authenticity where you want to check did this data point actually come from the source I'm expecting it to come from um and then you have to check kind of uniqueness and quality as well and each of the data Dows implements that themselves with a set of tools that makes it easy gotcha okay um and how does this now interact with other chain here and and is it kind of enforcing that things have to be sort of moved offline or offchain to uh to move forward or does that mean we actually are now also looking into crossgen communication and simplifying this whole agent Universe yeah so there's actually like a tight interlocking between the data da and the offchain compute right where usually the situation you end up in is like okay the da approved something like now what like now how do you like trigger some compute to happen but actually like the um so what you have is like a te that has um the key that the Dow controls and is checking okay has the dowo voted to run code in this environment and for us to decrypt the data if yes data is decrypted and the code actually runs so you end up with a really tight interlocking between that like offchain compute environment that's not um it's not optimistic or ZK based it's actually just purely like okay it's in a te and you have a guarantee that this is going to run without the data leaving right often that's the problem where uh ZK or um other proofs they can tell you something about the data but they can't prove to you that it didn't leak right and so that's this really unique like compute environment that makes the DAT more like interoperable composable while preserving privacy I think that's a really uh I mean this is a really key feature for you to actually guarantee actual safety around what's the output and um make sense of it so I'm glad that he ended up actually covering that because I didn't think of it um in terms of like pointing that out um that's awesome well I want to thank you again for uh this amazing presentation and uh for everybody on the live stream uh still continuing to build on your uh hacks for this weekend uh feel free to uh check out Vanna and uh using data awesome thanks so much thanks an [Music]