Data: The New Digital Asset Class | Anna Kazlauskas - Open Data Labs

all right thanks everyone for sticking around for our next talk which is going to be an exciting one it's called Data the new digital asset class it's being given by uh Anna Cass who is working on a project named vanana and it's going to be pretty interesting so please welcome to the stage Anna great thank you great uh just waiting for some slides up here but in the meantime uh oh perfect thank you uh so yeah hey I'm Anna um from vanana uh quick intro on me I uh so got interested in crypto more through traditional currency so I was really into central banks as a kid and in high school like I had a picture of Janet Yellen chair of the FED in my high school bedroom and then like worked at the FED in high school got to MIT and learned about decentralized currencies this was like 2015 and was just like so fascinated by the idea of being able to basically run a Central Bank in your dorm room kind of fully distributed so started mining eth was part of the early MIT Bitcoin Club um some of the economics research I was doing led me to uh actually kind of econometric models and then to some kind of AI models so started doing re research at Cale which is mit's AI lab um in the field of natural language processing and I always wanted more and more data to be able to train AI models on um great so uh this presentation is going to start with kind of a view of how data fits into Ai and then go into specifically Van's role uh AI models are ultimately only as good as their training data so this is the training data for gpt3 it's slightly old cuz now uh the AI companies don't say exactly what they train on but it's pretty representative so you have common craw which is basically the whole publicly scraped internet so everything that Google can index that's basically the data set um web Tex 2 is another kind of public data set it's basically all Reddit links with more than three upvotes go and scrape all of that text you can kind of think of an AI model as like a kid that's learning from experiences and its experiences are the data that you put in front of it if you've post on Reddit you might have already helped contribute to AI models since that is included at least the older sort of posts on Reddit in some of the older data sets what a lot of people don't realize is that most data is actually not publicly available so um this slide is going through kind of most public data but it's actually only that tiny little dot right there it's less than .1% of all the data kind of overall so internet data is mostly living behind kind of permissions or signning to access if you think about I don't know like your daily life on the internet like your emails your Google Docs your messages most stuff is not publicly available right most stuff you have to sign into to actually be able to access so the data represents a very small subset what we've taught AI so far one of the issues that we're running into from an AI research perspective is called the data wall and what the data wall refers to is the fact that we're actually running out of data to train AI models on so what this graph is showing is AI models as they've gotten better and better and how much training data they're trained on uh in AI the training data size is measured in a token which is basically just a word it's unrelated to crypto tokens um so llama 3 was trained on about 15 trillion words and if you look at the whole public internet so that's fine web it's a subset of common craw which is just the high quality stuff and it's like duped that is also 15 trillion tokens so one it is very impressive that we have basically taught AI literally everything available on the public internet like that's a lot of information to teach an AI model but if we want to make AI better then we need access to more data one note I'll mention is that the data wall is both an issue of quality and quantity and so if people are familiar how many people have like interacted with deep seek or have switched to deep seek Okay cool so deep seek is like really smart right and it has this Chain of Thought of reason ing um something that they did to overcome the data wall is they actually started with a really high quality subset of data so they hired a team of phds to generate really high quality data of like coding and mathematical reasoning their founder actually was directly generating some of this data in order to be able to build kind of a better AI model um here I actually have a job posting from their parent company of them like hiring people to do a lot of this highquality data related work uh so companies have realized that their data is really important to Ai and extremely valuable at this point in time and so they've kind of locked it down to make sure that they can charge as much money for it right so Reddit is earning $200 million from selling data that came out earlier last year before they were getting ready to IPO photo bucket earning $1 to $2 an image um Apple's paying like $50 million for news data and generally this trend that happening is that platforms that have this really valuable data in their walled Gardens is making it harder and harder to access so stack Overflow change their terms of service Reddit Chang their terms of service and it's becoming increasingly hard to actually access this data set something that a lot of people don't realize is that users actually have full legal ownership of all of their data so in the same way that when you park your car in a parking lot the parking lot doesn't own your car right they can't randomly take your tires they can't destroy your car you can always go and get it back when you put your data into a platform it is the same sort of legal relationship like it is your data it's sitting there and you've signed terms of service that say that they can use it and they can um kind of use it to provide you with services but at the end of the day you can always go and take your data back um as a user I'll also mention that this is not widely known I think the general uh View and kind of business model of big Tech is around treating this data as if it's their own but it uh is actually not their own they just have kind of uh the terms of service that grant them access to it so with every single platform you use you can go and Export a full copy of your data so you can get your Tesla data you can get your Amazon purchase history um Instagram is actually a really fun one because it gives you both your posts and your messages and photos but it also gives you the AI generated labels that they've put onto you for the purposes of advertising um it's like a list of 300 things one of the labels that they had put on me was stressed but not depressed and it's actually just like hilarious to go and do this data so you can Google like Instagram data export and see what they have so the approach that we're taking at vanana is to basically have users cut through these walled Gardens by exporting their own data um some of the four Concepts to understand on vanana are one non-custodial data so you can manage access to your data just like your funds you can use any evm compatible wallet sign a transaction and you're basically granting access to be able to um Grant an application uh ability to use your data um but you're staying in in full control um the other thing that's important is proof of contribution and so when you work with data if like we just have my data it's not that valuable because it's not enough data to be able to train an AI model but if you have a whole group of people that trains their um that kind of combines their data then that data becomes a lot more valuable right the problem is like what if one of us so we're all going to combine our data like our telegram messages or something and one person pretends to be 100 people to try to kind of like Take Over Control of the data set so what proof of contribution does is it ensures that data which is necessarily kind of not fungible uh coming into a pool of data called Data liquidity pools on vanana is high quality and it actually is basically um giving a score to each data point relative to another data point and it's implemented at the data set level so each data out on vanana has a different proof of contribution so um data Dows are one of the core Primitives on vanana that basically allow people to combine their data together you can think of it as kind of like a labor union but for data um so what it does is it really incentivizes Aggregates and cryptographically verifies valuable data data contributors retrieve their data from either like via API or a gdpr request um every data da builds their own interface so there are a bunch of different data D some of them have a Chrome extension some of them have a standalone UI some of them have a scraper and then the aggregated data can be used to train user owned AI models or sold as data sets so a few of the data DS I mentioned there's one that's focused on Mental Health Data um there's one that's focused on financial prediction data and they actually have you export like your coinbase and your binance trades which usually would be stuck within that centralized exchange um yeah so here are some of the data Dows just to give you a sense of kind of the the different directions you can take them so the Reddit data DA has 140,000 Reddit users in it um that one was an interesting one where um so they every dated out on Vana has its own data set specific token kind of like an erc20 token what happened with the Reddit dated out is early on they had mostly crypto native people joining coming in and then people started speculating on the Reddit data set specific token and because of that if you were a user with a lot of karma because if you have you get rewarded based on the square root of your karma if you have more karma you get more ownership of the data set then you could actually make like3 or $400 for your data so then there were a lot lot of people who were totally new to web 3 but had a ton of Reddit data who came in and contributed and built up this really valuable data set um and now there's like an ex Google brain machine learning engineer who has trained an AI model on the data set um because it's just trained on Reddit data it's pretty much as good at like ship hosting but it shows kind of that full end to-end flow of users contributing data and being able to train a better AI model the other data da I would call out is the 23 and me data da their goal is actually to buy 23 and me and make it user owned um I think that 23 and me is now actually at like a $75 million market cap so it's not insane but the reason why this is an exam an interesting example of a data da is their goal isn't to sell the data it's actually to prevent 23 and me from selling the data so you're able to kind of use this collective bargaining power of a group of users they have a few thousand people in it if you have a 23 in me account it could be good to join um to be able to negotiate with platforms um Reddit actually changed some of the terms of service following the Reddit data Dow and I think there is a lot of power that users have around kind of aggregating this data so data on vanana is structured and quality checked so each DLP that is the more technical term for a data da defines their data set structure and proof of contribution and what that means is that all the data that's available on Vana you know what the structure is going to be and you know that it's been quality checked something like one example of something that you could build with a sleep. fund data set um so you can combine data across different data Dows and then you work with it in a te so it's privacy preserving uh you could train a model that basically has people's coinbase and binance data from Fin quarium and then their sleep data and you could look at across all token holders like who is sleeping worse like of the projects like who's getting bad sleep these days um so that's one example of something that's possible to build on top of this data the other thing I would mention about Van's kind of economic design is our goal is really to turn data into a new asset class so there's a lot of market failure around data today because it's really weird to work with it's very IL liquid one of the major bottlenecks there is what we call the double spend problem for data right so if I sell you my data it's very hard for me to know that you haven't then sold my data to somebody else right so the way that we've solved this with vanana is that you can only run operations on the data within a secure environment but the data actually never leaves so what it looks like is it's closer to renting your data than actually selling it so each of the data DS has its own data set specific token um that is created through a data liquidity pool and contributors earn data tokens by adding verified data depending on how valuable your data is to a particular dat out you're rewarded proportionally so to give you some examples like the Reddit one I mentioned it's based on the square root of your data 23 and mi1 is based on I think it's equal for everybody um something else is uh usually there's some kind of a drop off curve where if you're in the first 10,000 users to contribute your data you're rewarded more than if you're like the millionth user right and that's because the marginal value that you're providing by adding your data is higher at the beginning um thinking about like Twitter as an example if there's somebody who tweets very similar things to you and they've already joined the Twitter data da then your data is potentially not as valuable because that information is somewhat already represented um there's some research related to model influence functions that goes into a bunch of this detail where you can basically show an AI model a new data point and then get back um exactly kind of how much the AI model learns from that given data point um yeah so on the data token mechanics yeah contributors earn by contributing data and then companies and researchers basically buy and burn data tokens for Access um there's also a really interesting like data token ecosystem that's popping up so there are some data dexes you can basically trade data tokens there's someone building a data purpose platform uh there's a plugin so that AI agents can actually trade data um the web 2 analogy I look at for this is maybe like oil Futures right so a lot of companies that consume a lot of oil will buy oil Futures in order to drisk on the fact that they're going to have to use a lot of oil we're starting to see a little bit of like compute Futures where companies will buy compute Futures to drisk that and so I think in the future we're going to see these like data Futures and kind of new Financial Primitives built on top of data as that underlying asset uh so there are DLP incentives on vano which are really designed to help onboard data into the network and what it does is the top 16 DLP by stake actually earn Vana emissions for onboarding data which are tied to how many verified data contributions they have and how much their data is accessed um some of the performance metrics that determine rewards so data access volume data quality and value and amount Stak um there's an upcoming upgrade where the incentives will go directly to the Dex pool of the Vana data token pair right now it goes into the data da and then is distributed and so that's one um upcoming sort of thing to call out and then here are a few of the different data Dows so there's uh the DNA data da which is focused on Raw DNA sequences um mind Dow which is a mental health data set um right now on Vana there are 20 live data Dows where you can go and kind of contribute your data or stake or participate today um some of them are Venture back teams so DLP Labs which is the one that created the Tesla data da they've raised about a million dollarss they have their first data buyer who's a battery company buying buing access to Tesla data um some of the other ones are smaller teams and you can kind of uh get started relatively easily um and then so there's the data outside which is basically onboarding data to vanana and bringing data on um it's always pointers to the encrypted data that are stored on chain and then the proofs and the records so you understand what's in the data the other side of it is being able to access data right so there are two ways to access data on Vana one is you can access indiv visual data via the data portability API um what that means is that if I've joined a data da and say I've connected my sleep data I can log into another application sign a transaction and Grant them access to see my sleep data and what's cool about that is I think especially for AI applications you often want more and more data on your users to be able to personalize the application to them so that is at the individual level the other kind of data access that's available is actually accessing a whole data Dows data so that's the example I talked about with like the ex Google brain machine learning engineer training on a Reddit data set there's another AI company flower AI that's actually doing full pre-training that means you're making an AI model from scratch across three of the different data Dows one of our guiding use cases is how do you get to 100 million users who've all contributed data from five different platforms so that you have uh almost 500 trillion tokens worth of data which is 30X leading AI models today and you can train the best foundation model um but most of the um kind of use of data today is more around um fine-tuning models um and and smaller models like that and on the 100 million goal we're about 1% of the way there um it is a massive undertaking um but that's kind of one of the guiding principles uh yeah oh I'll also mention that only the outputs can leave unencrypted ensuring the underlying data stays private so the data Dow actually has to approve the code that runs in the secure environment right so you can't just like be like Oh I'm going to go run code in the secure environment that just sends all the data somewhere else the code has to get approved so you make sure that it never leaks so if you want to get started building on vaana um here are the steps to start a data da you can also Access Data um first is picking out a data source and how users will contribute uh something to mention is that you don't need anybody's permission to start a data di because users own their own data right if you want to start a Pinterest data da you don't have to go ask Pinterest or get API access or anything because every user has their individual data access uh Reddit actually tried to shut down the Reddit data Dow around when it hit 100,000 users but because of the way that it works on vanana where individual users are getting their own data um you don't need anyone's permissions the user has that legal right and then you can um build and deploy the majority of the work is in the proof of contribution which is the python code verifying data and then you can build an interface for users to contribute data there are a bunch of different open source templates um this one links to a web app but there's also a minor if you look at the Twitter data da that's all open source and you could reference some of that and a Chrome extension um and then you can register and launch your data da you don't have to compete for a stop a spot in the top 16 you can also just run a data da um and most teams are two or three people build takes a few weeks um we also have an accelerator program for dated hours if you're interested in building um and you can probably create a template one relatively quickly uh great thank you so much [Applause]