Builder Workshop with Anna Kazlauskas

So this morning I'm going to dive into a bunch of technical detail on VANA and the different sorts of applications you can build. Yeah, before we dive in, I think probably everyone here is at least somewhat familiar with VANA if you came to the VANA event, but I will, yeah, regardless, introduce myself and some of VANA. So I got into crypto through like mining Ethereum back in 2015. I was really into central banks, actually, so I had like a picture of Janet Yellen in my high school bedroom, like worked at the Fed in high school, was like central banks are kind of controlling currency, which is central to the whole economy, and got to MIT, learned about Ethereum and the fact that you can run kind of like a distributed central bank just from your dorm room. So started going to the dumpsters, getting GPUs, setting them up in like milk crates, and became really fascinated with just kind of decentralization and how you can build systems that are controlled by a whole bunch of people rather than a single entity, especially for pieces of infrastructure like currency or data that are very central to the world.

I got into AI because I wanted bigger and bigger econometrics models, like I was modeling economic data, found what was called natural language processing at the time at CSAIL, which is MIT's AI lab, and then started training some models there, always wanted better and better data, which led to Art and I, my co-founder, starting VANA, yeah, a few years ago, and now we have a whole bunch of different builders in the space, many of them in this room. Something too, I know it's been a long week for everyone at East Denver, so I might be slightly lower energy. I had Twitter spaces with CZ last night, and I'm quite sleep deprived, but yeah, excited to dive in. Okay, so first, oh, actually, how many people are coming more from a crypto background versus AI? Like, crypto, raise your hand, and then AI, raise your hand.

Okay, cool. I'm excited to see the AI hands up, too, even at East Denver, please. So AI models are really ultimately only as good as their training data. This is a breakdown of the training data for GPT-3. GPT-3 is one of the last times where they were super transparent exactly what they're training on, which is why I have kind of the GPT-3 breakdown, but what you'll notice is that the majority of the data is common crawl, so it's just like the whole public internet scraped.

If you post on Reddit, you might have already helped contribute to AI models. Webtex2 is an older Reddit data set you might be familiar with, where they basically take all Reddit data with more than three upvotes, and then they scrape all those links. They can't do that anymore, because Reddit's changed their terms of service, but that was one of the old kind of historic data sets. Something a lot of people don't realize is that most training data is private, not public, so on the last slide, I was showing public internet data, so that is that small dot on the right here. This data that I've broken down here is a very small percent.

It's less than 0.1% of the overall data that's actually available on the internet, and the vast majority of data requires permissions or sign-in to access, so most data is sitting inside of the walled gardens of Web2 today. One of the problems that we're running into from an AI research perspective is called the data wall. What that refers to is the fact that we're actually running out of data to train AI models on. If you look at the data, so here kind of graphed are better and better AI models.

Llama3 is this one on the right. It was trained on 15 trillion tokens, basically a word from an AI perspective. That is equivalent to the 15 trillion words that are available on the whole public internet. FineWeb is a subset of Common Crawl, which is just the high-quality stuff. One, it's pretty impressive that we've actually trained AI on the whole public internet.

That is like a huge amount of information for AI to have learned, but if we want to make AI models better, then we do need to be able to get more data from somewhere. One thing I'll mention, too, is that the data wall is an issue of both quality and quantity. If you guys saw DeepSeek, they actually overcame the data wall through quality, where they hired a team of PhDs to generate really high-quality cold start data, where they were actually having really smart people generate coding data, math problems. I think their founder himself was actually generating some of this initial cold start data. This is one of their job postings, actually.

They're hiring a lot of people to collect data and stuff like this. This is really powering DeepSeek. But because of this data shortage, where people realize, hey, if I can get better data and more data, I can build better AI. Building better AI is a very valuable thing to do. It generates open AI, like $3.

4 billion in revenue. It's kind of becoming this central part of the economy. Companies have realized, hey, my data is super valuable. Reddit's earning $200 million from selling data, PhotoBucket, $1 to $2 an image. Alongside this monetization, they've also made it harder and harder to access data.

Reddit and Stack Overflow have changed their terms of service, changed their API terms, so it's much harder to get access to. I was talking to someone this week who was trying to build something using Pinterest data and got rejected from their API, because they're not allowed to use it for AI. More and more, that's the case. The places where this valuable data is sitting is becoming way less open. One group still has access to their data, which is users.

A lot of people don't realize that you legally own all of your data. In the same way that when you park your car in a parking lot, the parking lot can't just randomly take your car or take your wheels. It's still yours. It's just sitting there. When you put your data in a platform, it's like the same kind of rules.

It is entirely your data. When you check the terms of service box that says you can use my data, you're granting a permissive license for them to do things with your data, like operate the platform. But it's ultimately yours, and you can always go and take it back. So even as these platforms lock down developer APIs, individual users still have access to their data. One of the examples I really like, this is just a fun activity to do, exporting your Instagram data.

They'll give you not just your posts and your photos and kind of all your comments, but also the AI generated labels that they put onto you for marketing purposes. So you'll get a list of like 300 things. One of the labels they had put onto me was stressed but not depressed. And it's actually just like a hilarious activity. So I don't know if they're still using that label, but maybe, yeah, you can go and see.

And then, yeah, 23andMe is another example. Like, yeah, you can just always go and get your data back. The platforms generally don't want you to know this, and they don't always make it super easy. There's some data regulation that makes sure that you can always get it back within 30 days, and it's getting better and better over time. So, for example, in Europe, there's a lot more API access available now.

And then in California, it's quite good. So if you're requesting the export and they ask where you live, sometimes they do, you can put California, and you'll generally get kind of the highest quality export from there. So what Vana really does is it unlocks data from walled gardens and kind of puts it in user's control. This is kind of a classic diagram of like the walled gardens of the Internet, right, where each big tech company has kind of built up their walled gardens to protect their valuable data. And then Vana, kind of the symbol of the black iris, we have some physical flowers around too, is really about like breaking through those gardens through the individual user, right?

Like all of us can go and take our data from every single platform and actually rebuild better datasets than any of these leading platforms have. Okay, so actually, if folks in this room, you might have seen me present these slides to date already. So now we will get into the kind of like builder portion, which is going into what are all the things you can build on Vana, especially beyond data dials. So two of the core concepts on Vana, one is non-custodial data. What that means is you can manage access to your data just like your funds.

So in the same way that you control like your crypto funds with your wallet, with Vana, you control your data with your wallet. So you just sign a transaction, you can grant access to an application or decide what you want to happen with your data. The other core concept to understand is proof of contribution, which I'll go into more later. But the high level of this is basically your data on its own is not that valuable. It's much better if you can pool it with many other people, right?

And so if all of us together, we're like exporting our Telegram messages, what Vlog might do is like pretend to be 100 people and then like try to take control of the dataset, right? So what this does is it makes sure that the data coming into the pool is actually high quality and allows for a group of people to aggregate their data all together. Okay, so I'm going to go through a bunch of different applications you can build on Vana. Most people are only familiar with data DAOs today, which is the most self-serve kind of ready to go thing. It's actually been great now there are data DAOs deploying to Vana.

Like I used to know all of the teams who were deploying and now I just see a data DAO and I'm like, that's crazy. They're live. So yeah, data DAO is fully self-serve. There's like still some friction in setting it up, but that's relatively straightforward. I'll talk about Datify applications, which are still quite early.

So I'd say this is like, you can build them if you're a deep DeFi builder. That's ready to go for you. And then I'll give a preview of some of the data access stuff that's coming. This data access stuff is still in the works, still very much in beta. But if you want to, yeah, try this out, I'll show you some of it.

And then lastly, I'll talk about data portability. I put two construction emojis here because I don't, I think the engineers would be mad if I, I don't know, put one, but it's like, yeah, so this technically works. Like you can use it. There's a lot of friction to use it today, but you actually could build an application like that. And I was talking to someone earlier who's building like a longevity application and kind of wanted to consume some data from different data DAOs.

So I think that, yeah, this data portability piece would be really helpful for that. Okay, so diving into data DAOs, the, yeah, one of the biggest data DAOs today is the Reddit data DAO, which has like 140,000 users. They've now trained like a user-owned AI model. These are kind of examples of what you can build. Some of the data DAOs are venture backed.

I don't know if Ryan from DLP Labs is here, but he runs the car data DAO, where you basically take telemetry data, connect it in and are able to earn for that car data. I also don't know if the Nakamoto guys are here, but they run a 23andMe DAO. They're working to buy 23andMe, make it user-owned, which is obviously a very ambitious goal. They have a few thousand people in the data DAO. They haven't made a bid or anything, but that's kind of their end goal.

One of the important pieces of data DAOs is kind of understanding that data in the data liquidity pool, which is a more technical term for a data DAO. So yeah, basically data liquidity pool is a collection of data that has been verified and is tied to a token. Most of the time that's implemented as a data DAO, where people have governance rights over the data set, but that's not necessarily the case. You could build a DLP that doesn't necessarily have DAO-like properties. So yeah, each DLP defines their data set structure and their proof of contribution.

This is a data schema of sleep.fun, which is one of the sleep data DAOs. And what's nice about this is when the data gets written on-chain, you're able to know what the structure is actually going to be, right? I think often when people talk about building like massive kind of amounts of data or this idea of a data treasury, it's like really only valuable if you know what the structure of it is so you can actually work with it, right? You need to be able to do things with the data.

So that proof of contribution is ensuring structure, but then also ensuring quality. A large portion of the work in building a data DAO is in implementing this proof of contribution, which is a way to validate data on VANA. It's basically just Python code that runs on a set of TEEs to ensure data quality while preserving privacy. A few of the data DAOs use like client-side DK. So if you're thinking of doing something like that, you could look at the Amazon one, which has like a browser extension where they client-side generate a proof of your data quality as they get it.

And then there's also an integration with Reclaim, which uses DKTLS. I'm not sure if any of the data DAOs are using that live today. Yeah. Oh, are you guys? Okay, cool.

There is one data DAO. You can go talk to them if you wanna learn more about how to use DKTLS with VANA. And then all data DAOs make their proof of contribution publicly available. So you know exactly what guarantee you have of the data when you're working with data from a DLP. This is the Python template for running kind of your proof of contribution.

I think Vlog, who's right here, is the best person to talk to if you have any questions about how to implement this or how to get it running. The other thing to mention about a data DAO is that alongside, so proof of contribution is basically mapping data, which is necessarily not fungible, right? My data is different from your data to a fungible asset, which is a data token. So it's kind of deciding how much is my data worth relative to somebody else's. And the data DAOs do this in different ways.

So the, actually, maybe we can pull up some proof of contribution. I think that, there we go. Okay, I have this linked actually for Synquarium, which is one of the financial data DAOs. And so what the financial data DAO does is they have you export data from centralized exchanges that isn't usually publicly available, right? So they have you export like your Coinbase and your Binance data.

And then they're giving you a score based on how much your data is worth relative to other people's. I think in this case, they are, yeah, rewarding you based on trading volume, trading activity. It looks like if you've lost more than a million dollars, they actually give you more points. I don't know what the rationale behind that is, but maybe for some reason, they see that as more valuable data. And so, yeah, this is an example of it.

You can, you wanna structure it. Oh, something else I'll call out is often they have a halving schedule because the first people who join a data DAO are often bringing in kind of more relative value than like the millionth person. So in this case, the first 10,000 people, they reward more aggressively and then they start halving the rewards from 10 to 100,000 people. This is kind of like a whole new design space, which is data tokenomics. It's something I'm actually, like I love spending time on.

So if you're trying to figure out how to design this, like come and talk to me. Or we also have Francis, who's kind of our resident DJ on the team who can talk to you about some of this as well. And really what you're like, the perfect economic model that you're trying to get closest to is you want to be rewarding someone for the marginal value of their data point that they're bringing in, right? And so for example, if you have like an email data now and it has like all your emails or maybe all your writing, like Google Docs, if there's somebody who writes really similar to me that has already contributed their data, then I might like, you might wanna reward them less because they're not adding that much to the data set. One research area I would call out here is actually model influence functions.

Let me just, oh, there we go. And I know, I think, yeah, you guys might be doing some research related to this too, but what a model influence function does is you basically give it a data point and then it tells you how influential that data point is to teaching the AI model. So if people, maybe the like AI data people in the room know this classic data set, which is MNIST, it's just handwritten digits. And it's like, kind of like this toy data set that machine learning people use to like look at examples. And the goal is to take this handwritten digit and then say, this is, yeah, this is a digit.

These are the most influential twos, I guess. And you'll notice it's a lot of the edge cases, right? It's kind of the weird data where it's like, oh, like this is a two, not, I don't know, what someone could think it was maybe like a G. And so, yeah, it's often these edge cases that are really valuable. And that's also true in like Web2 data collection, right?

In Web2 data collection, which Colin and Art will go more into kind of later this morning, often the problem is, how do you get a representative sample of people to all contribute their data, right? So, and often you end up in scenarios where you're like, I need someone, I need like a man in their 60s who's Asian and has this specific, you know, and you're kind of like trying to get that full spread. So yeah, model influence functions are one tool. I'm not sure if any data data, do you guys have it in production or still in research phase? Okay, I'm not sure if any data data have it in production today, but I know that a few are working on it.

And yeah, also if you are, yeah, curious about proof of contribution, a really good place to start is to look at the different data data and what they've implemented for verification. Although if you have a bad proof of contribution, you will probably find out very quickly as soon as you go live and start rewarding people because you will get aggressively civil attacked. So proof of contribution is both about mapping kind of these non fungible, yeah, pieces of data to a fungible token and also like marking them as zero when they should be zero. Let me just pull up the Amazon, oh, it's called Prime Insights. This is the one that has a Chrome extension to verify your data.

And what they do is, actually, let me see. Okay, so they have you upload a proof file and the proof file is a ZK proof that's generated client side when you download your Amazon export, right? And so one question I get sometimes is, and you get this export from a platform, right? Like the file, how do you know someone hasn't edited that file or tweaked it somehow? And the answer is honestly, like it depends on how exactly you get that file.

So in the case of Amazon, because you get it directly from the browser, you can use a Chrome extension to generate a proof. But in other cases, like with ChachiBT, when you export your data, they send you an email and you click that link and download it. So then that's the link that you would need to verify. Our team, like a bunch of our engineering team has thought a lot about proof of contribution, especially in like an adversarial setting where people can get money if they slip through. And so, yeah, if you have a data set you're thinking about, we're kind of happy to chat through that.

In terms of the mechanics of data tokens, so each data DAO has their own kind of data set specific token, which is created through the DLP, right? So you have this data liquidity pool and each time someone brings a piece of data in, they're rewarded with tokens proportional to how much they contributed based on proof of contribution. So contributors earn these data tokens by adding verified data. One thing I'll note on the contributor side is whenever somebody contributes data, it gets tied back to their wallet. And so if you're using the same EVM compatible wallet across different data DAOs, all of your contributions are actually mapped to you.

So I can pull up my wallet and we can look at some of my contributions for data DAOs. This is my public wallet, I haven't been using it as much, so it might not be. Yeah, okay, so you'll see like I've kind of added files with permissions and contributed to data DAOs. And this data is all available on chain. Obviously the data part is private, but the fact that I've contributed is publicly available.

So if I sign a transaction, then I can grant access to that data to somebody else. One sort of technical note, if you're building a data DAO, something we've noticed is some people don't like the friction of a user having to sign a transaction. And so sometimes what a data DAO will do is basically relay the transaction on behalf of the user. And then there's just like a data owner address or something like that that you can specify so it still gets mapped back to the user. And then companies and researchers can basically buy and burn data tokens for access.

That's something that Art and Colin will get into later as well, of kind of like how the data sales process in full works. And there's also an ecosystem company, TensorSource. I don't know if Sam, the founder, is here today, but what they do is kind of help the data DAOs sell their data because most Web2 companies don't even know what a data DAO is or kind of like how to talk to one. Oh, another thing on data DAOs, just at a very high level, I kind of think of them as similar to almost like a labor union for data where it's kind of this data cooperative where people have pooled their data and then have this governance right over the data set. And there's actually a lot of flexibility in what a data DAO can do.

So the earlier example I gave with 23andMe, their goal is actually not to sell genetic data. It's actually to prevent 23andMe from ever selling genetic data to insurance companies, right? And so you get a lot of kind of like collective power when you have a group of people who pooled their data. And you can use that to train an AI model, to negotiate for better terms of service, like following the Reddit data DAO, Reddit actually changed some of their terms of service, right? So you actually can have a fair amount of influence to be able to negotiate how the data lives in the existing platforms or whatever else a data DAO might want to do.

Something else to mention about data DAOs is that VANA is designed with incentives for the top 16 data DAOs. So if you're in the top 16 DLPs, just a more technical term for data DAOs, by stake, you're actually earning for onboarding data. And the earnings are tied directly to verified data contributions and data access. And so the performance metrics are basically what determines the split of rewards. They're related to data access volume, data quality slash value, which can be a very hard thing to measure, and the amount staked.

There are a few changes coming to this. If you're deep in the VANA ecosystem, you might know like DRC-13, which is one of the changes that's coming. And what that is doing is making data access much easier to measure and ensuring that it always happens within the TEEs. And then it's also directing the DLP rewards directly into the data set specific token. And if that sentence meant nothing to you, that is okay.

That is like a very detailed note for folks who are already building DLPs. And so yeah, here are some of the top DLPs. Should we go to Data Hub? Check it out. So here, what we noticed, there was a lot of kind of success to the successor happening, where people were just staking with the top data DAOs.

So you actually can't, you can see who's in the top, but you can't see exactly how much stake they have now. I think I might have zoomed out. Okay, here we go. Yeah, so here's one, sleep.fun.

You're able to contribute. So yeah, you can go and contribute to their data DAO. And you can also kind of see, there is, okay, someone sent me, oh no, it's vanaseek. Someone sent me this UI. I actually don't know who, maybe someone in this room built it.

Oh, maybe, okay. Maybe they took it down. Someone had sent me this UI, which was basically like a way to query stake across all the different data DAOs. I'm not sure if it's kind of in production or even necessarily where it came from, but that's a tool that I saw some of the DLPs were using to track some of the metrics. Maybe I'll pause and actually ask, are there any questions on data DAOs or on these incentives that would be helpful to talk about?

Yeah. Do you consider the time domain, like especially around data being useful for some period of time? Yeah. And you want it to come to scale, and you kind of want to gain it all? Is it sort of like a one-time transaction?

Yeah. Or like you are subscribing, you can actually get a value over time? Good question. Yeah. So it depends on the data DAO and kind of how they set up their proof of contribution and data tokenomics.

I think there are some, like I believe the Spotify one is getting your API key. So it can constantly kind of go and refresh your data and kind of reward you proportionally. I guess the most extreme version would be that you actually, you give people a non-transferable data token, and then whenever the data is accessed, you only distribute rewards to like whoever, like, yeah, basically people who contributed to that data set. There is one element that's like kind of surprising where I feel like some AI companies want historic data because they know it's not AI generated. They're like, can I get the stuff from pre-Chat GPT?

But yeah, it really depends on the data source. One thing actually I call out with Vanna today is that it's not great at super real-time data, which is not really what we've designed it around. I think that'll get better over time. And if there's a super real-time data source, like I'm sure a common question we get is do hedge funds buy the data? And actually, at least as of now, there might be specific data DAOs selling data to hedge funds, I'm not sure, but that's not really a case we're focused on.

And part of the reason for that too is that often the hedge funds require an exclusive license, which is kind of hard for a data DAO to be able to guarantee, and potentially limiting in terms of what they can do. Yeah? I know most of the data currently is unstructured data, but do you see unstructured data having value and having value? Yeah, I think that unstructured data, I guess there are some video and audio data sets that would be kind of interesting. Actually, maybe folks, yeah, so I think that there could be other data sets.

You do need to be able to run proof of contribution on the data. I guess if you could throw just like the unstructured data at an LLM and trust that proof of contribution is gonna run properly, that could work. It might also be challenging then when you're consuming the data if you truly have no idea what the structure is. So I think you would want some level of uniformity, but of the question like does it have to be, does it always have to be? be a specific JSON file, like not necessarily, although maybe you could argue you can like convert any unstructured data eventually to a JSON file, but maybe more of kind of a technical one.

Yeah. Any other questions? Okay. Cool. So that is, yeah, everything about what a DataDAO is.

In terms of how to build one, these are kind of the steps to get started, and this is QR code with the DataDAO template, but I can share these slides after because there are a bunch of links in here. So first, you want to pick your data source, how users will contribute, and kind of start designing proof of contribution with tokenomics. The way you should think about proof of contribution is really like what's the minimum requirement you want to have about a particular data point coming into the system, right? And then how you want that to scale as well, of like, okay, if someone has more data, how do you let them in? And then from a build perspective, then it's a question of implementing your proof of contribution, which is in Python, building an interface for users to contribute data.

So each DataDAO builds their own UI. Some of them are Chrome extension. I think I have YKYR. Oh, I might have. So like YKYR, for example, they have a Chrome extension where you just download a Chrome extension and it's like a web analytics DataDAO.

Some of the other ones have like a template UI like this. There are a few Telegram bots, and I might be forgetting a few other interfaces. But yeah, any way a user can contribute data. Something that I think someone in the ecosystem is working on, and I know Colin has been thinking a lot about, is a unified UI so you can contribute data across DataDAOs so that if you're a DataDAO who doesn't want to build an interface, you don't have to do that. So that would kind of be a more standardized way for users to contribute data.

And then what you can do is deploy template smart contracts. You don't need to know Solidity to deploy these contracts. It's actually relatively straightforward. You're like changing the ticker symbol to your specific data token. If you want to make more specific modifications, you can.

There are some requirements coming out as part of VRC13 so that all the DLPs have kind of unified data tokenomics. For example, a fixed supply. But yeah, it's not too bad from a Solidity perspective. And then you can launch. You don't need to compete to be in the top 16.

You can just launch a DataDAO and kind of build a collection of data in that way and aggregate it and then use it to rent out to train an AI model. Or you can compete in the top 16, which earn a share of emissions. Most of the teams are two to three people, and I think initial build takes a few weeks. If you've built a DataDAO, actually, can you raise your hand and then people can come talk to you if they have questions? Okay, nice.

So talk to those people. And I think all of our engineers, too, can also raise their hand. So those are the people you can go talk to if you have questions about building a DataDAO. You can create a template one probably in a few hours. One of the engineers was like, you can do it in five minutes with an AI, and I'm skeptical of that and have not verified that claim, but maybe one of the engineers, Chris, I don't know.

Someone said that. I was like, I don't believe that, but okay. And so, yeah. Here's a DataDAO template. That's kind of the best place to get started.

If you're wondering which DataDAO to build or kind of thinking about what data source could make sense, for Christmas this year, I wrote a wish list of DataDAOs that is in a blog post that has a bunch of different DataDAOs. I can actually pull up some of them and maybe talk through how I think you could run proof of contribution on some of these. Okay. So there we go. How readable is a dark screen with a text?

Okay, cool. Yeah, so health DataDAOs are something I think are really interesting because there's so much kind of like regulatory red tape around this data. Something that's really special about VANA and kind of the way that we've designed it is the user acts as their own data processor, right? And so it's basically in the same way that with Bitcoin, it's non-custodial because if you control the wallet, you control the funds, Bitcoin isn't your bank, right? In the same way with VANA, like the user is their own data processor, and so they have kind of full control over their data, and you don't have a lot of the regulatory red tape that comes with working with data in a centralized way.

Oh, I see, I think that's the Biohacker DataDAO founder too, who is probably one of the first health DataDAO founders. If you've got an Aura ring, join the Biohacker DAO, and you can kind of, yeah, contribute your sleep and other health metrics in there. I'd love to see like medical records, DEXA scans, fitness tracking, nutrition. Something that's nice about VANA is that the data is interoperable, so if one DataDAO is focused on sleep data, and another DataDAO is focused on health records, that data can actually be combined across users or used in different applications. Let me think of one other one to call out.

Okay, this is actually home mapping. I was recommending a Roomba DataDAO. I think that there's a lot of interesting stuff to do with robotics data these days, and so I was talking to like an AI robotics company where they actually bought a cleaning company just so that they could put like cameras on the person cleaning, so I think that that is a very high in demand kind of robotics data set, so that is like a really good area to be building. Screen recordings, like Rewind AI, they use custom encryption, so I haven't been able Yeah, if you can figure out how to get Rewind AI data, I think that would be a really cool data set because it's basically like a recording of everything somebody has done on their screen. Okay, so that is a bunch of information about DataDAOs.

This is kind of the most in-depth version on, yeah, the four different things you can build on VANA. Now going to DataFI apps. So what this refers to is basically applying DeFi concepts on top of the data tokens, which come from data liquidity pools, right? So the data liquidity pools are really that foundational layer, and on top of that, you can actually just build, yeah, any DeFi app with data tokens, and so there are a few kind of early ones in the ecosystem. There are a couple of DEXs, so there's DataDEX, which is one of them.

It's, I think, a Unity 3 fork to allow you to swap data tokens. There's another one, Flur, which is actually just a different UI on top of the Unity V3 fork, which lets you kind of swap different data tokens and kind of explore what's going on with a particular data token. And something I'll mention is that today, most of the DLP rewards are just going into DLP treasuries rather than burning the data token specifically or kind of like acting directly towards the data token. With VRC 13, which is the update that I called out, VRC is like VANA request for comment. It's where we just put kind of all the major changes to the protocol and kind of seek feedback and stuff like that.

One of the major changes that's coming is essentially having the reward in VANA by the data token and then put that into a DEX pool to really like bootstrap a lot more data token trading and DeFi activity. There's also an integration that Datapig made, which is an Eliza OS integration, just letting the AI agents kind of trade data. Okay, there's someone building a data perps platform that is not live, but I would love to see like basically data perps, right? So like anything you can build in, yeah, kind of typical DeFi, you can apply to data and in interesting ways too. One design space I was talking to Sven from Polychain about is like insurance.

So if you're working in like insurance or data tokens at all, he would love to talk to you. I think that's like an interesting design space he's looking at on VANA right now. Maybe one thing from kind of like the Web2 world is a lot of companies who consume oil, it's like if it's a big part of their spend, they'll usually buy oil futures, right? Because they want to kind of like de-risk the fact that they're going to have to spend a bunch on that. And now people are starting to do that with compute to some extent, like compute futures to basically hedge your price risk.

I was asking someone about egg futures recently because of the egg price spike. Someone was complaining to me, like, can you buy egg futures? But I think, yeah, like data futures, you could see being used in the same way, right? Where you're actually like, hey, I'm going to need this data set, so I'll buy kind of some future, hedge it. And then there are some like trading firms who are really interested in data as an asset class, right?

Like how do you value a given data token? What is the likelihood that this data set will be sold, will be used? What is kind of the lifecycle of it? And so I think there's a lot to build here. It's still pretty early in this space.

But yeah, this QR code is a link to VRC 13. And if you want to talk about building any Datify applications, Francis is your guy, our resident dGen. He always wears flip-flops and shorts, so he should be easy to find. Okay, so next, I wanted to talk about the third category of applications, which is AI applications powered by data access. So on Mainnet, the way that access to data sets works today is a somewhat manual approval process.

So what you do is you put up a proposal to say, I want to access this data, and then you go through kind of the full, like, okay, this is how much I want to pay. This is the code that I want to run. You get that approved. I'll actually show you – maybe I'll use the Reddit data now. Let's see.

I know there's a proposal. Let me see if I can get this. Okay, there we go. So this is a proposal from a while ago about kind of like someone who wanted access to the overall Reddit data set. And you'll notice it's pretty manual, right?

So they're saying, hey, I want access to six months. I want access to 10% of the data set. And this is what I'm going to use it for. This is what I want to kind of do with it. And this ended up going through kind of like a voting process, got approved.

He got access to the data set. So he got access to 10% directly. So he could look at the data and see, hey, does this data work well? What can I actually build with this? He had intended to try to beat some benchmarks for LLMs.

And when he looked at the data, he decided that it might be better to try to train like a shitposting model because it's kind of like more casual Reddit data. So he trained one of the leading shitposting models on then 100% of the data set in a privacy preserving way, right? So he saw 10% directly, looked at some of the data quality, saw what was possible. And then he trained kind of the full model on 100% of the data. And so this like works for data access, but introduces a lot of friction.

And what we found is that often people want to access data from across different data routes too, right? Like FlowerAI is an example of a Web2AI company that's training a model. They're training like a full, maybe people have heard me talk about like a foundation model owned by 100 million people, which I think sometimes people are like, oh, that's a huge undertaking. But they're doing full pre-training. Like they actually have in a distributed way trained a 1 billion parameter model.

And now they're training it across data from three different data DAOs. And the friction we were running into is we're like creating a ton of different telegram groups with every single data DAO. And we're like, oh, here, coordinate with them. And here are all these proposals you have to put up. And what we've noticed is that that is a ton of friction to access data.

So we're working on way better kind of data access where you actually have automatic API access and access to data from across data DAOs, right? So that means that I could access somebody's sleep data and their Coinbase or Binance trades. And then I could understand like, okay, which token holders of a given project are like mostly deprived right now, right? And you actually can start to do that kind of like cross-platform stuff. And then, yeah, there's also a proposal to kind of have a standardized way of burning data tokens and VANA to access data in like an 80-20 ratio.

And if you're curious, I'm going deeper into this. This is, we were, oh, yeah, with a community, we were hosting some workshops on this. There are a bunch of kind of like detailed notes in our developer Discord, which is a small Discord you can get added to if you want to go super deep into VANA. Okay, so these are, and this is not even on testnet yet, kind of very, very early. And I think the engineers actually didn't intend for me to demo this, and I saw it on one of their screens, and I told them I really wanted to demo it.

So if it breaks, yeah, apologies. So this is a query interface, which basically lets you query data. So you need a signature that shows that you're actually able to access the data. When this is on mainnet, you would sign in with a wallet and generate that signature to prove that you're you. And then you can essentially query the data.

You know what the data format is because that is recorded by the DLP. So, yeah, this is kind of what the query looks like. And then let me just reload this. So what you would do as a user is you would say, okay, this is the data that I want to be able to access. You'd write your SQL query to get access to the data.

You'd likely get some kind of synthetic data just so you can, like, test it out and play with it directly. Because when you're actually operating on the real data, that only happens in a TEE, in a secure environment, right? And that's to prevent what we call the double spend problem for data, where we want to make sure that, like, if I sell you my data, you can't resell it, because that's really what protects the economic value of the data. And so then on top of this data access, you can build, like, all sorts of applications. So here it's kind of like, yeah, fetching user stats.

This is, I think this is from the Unwrap Data Dao, the Spotify one. And you can see kind of all the different aggregate data that's available. One thing I, like, used to run into a lot in, like, training AI models is I would always want access to specific data sets and then just not be able to get it because I can't get API access. Or if I can get API access, it's hard to convince, like, 100 of my friends to give me their sleep data, right? And stuff like that.

So because of the Data Daos, like, all this data liquidity is sitting there ready to go, ready to be used. So as a builder, we have a ton of really fun data to be able to build with, and this makes it way more accessible. I think a lot of the applications that we'll see being used for the Data Daos, actually, let me pull up a bunch of the Data Daos and kind of go through, yeah, some of the things they could be used for. So, yeah, I think Prime Insights, which is an Amazon data set, basically shows, like, all of your past shopping history. That is one that people are very interested in for Consumer Insights.

So there's, like, a large, very large retailer who wants to use that to be able to decide, hey, when are my customers shopping at my competitor, right? And, like, what should I stock so that they don't buy there? And where should I open new stores? The auto DLP is actually selling data for car battery companies that want to understand how are people typically using their batteries. And I think that's, like, $3 a month per person per car.

So, yeah, those are some of the examples of data access. I think, I know I mentioned, like, there's one company, Flower AI, that's doing full pre-training. I think most companies are going to be doing more, like, fine-tuning. Like, I don't think that the main use case is this full pre-training. But this is some of their research that allowed for the, like, 1.

3 billion parameter model that was done in a fully federated way. Any questions on data that's available or data access on Zona? Yeah. Yeah. Yeah.

Yeah. Yeah, that's a great question. And I think for a lot of people, like, selling your data for, like, $10, it almost feels dirty. Like, it's, like, I don't know. I think that that's something, and that's often something that happens within data collection today, right?

So, 60% of data sales contracts actually fail to, like, completely collect the data. We have a list, Art might share it later, on, like, recent failed data collection contracts in the industry where someone's, like, I want to pay $50 for audio data from children. And then people are, like, yeah, you're not getting that. You know, it can be quite difficult. I think that the, like, data token aspect is really this change where rather than a one-time payment for your data, which, I don't know, might feel like selling your plasma or blood, you know, feels, like, a little dirty.

It's more about actually having ownership of this collective asset and deciding what happens to it, right? I think that whoever decides kind of what data goes into AI is sort of deciding truth, right? Like, if you look at the example of Google borrowed a while back where they had put a on the AI and if you asked it, like, show me a medieval, a picture of a medieval prince, they would show you, like, an Indian woman. And I think people laughed at it and were, like, oh, it's this crazy wokeness of the 2020s. But if there weren't another source of truth, people would just start believing that, right?

And they would just be, like, oh, that's just true, right? And so I think that for a lot of people, it's less about let me get a one-time payment for my data and more about let me have, like, ownership and a say in the future of AI. There's a lot of fear today about AI, right? Of, like, oh, is it going to take my job? And I think a lot of it is more of, like, an economic kind of market failure problem where if you own the AI that was created from your data, then you would be delighted that it could take your job, right?

Because then it can just do work for you and kind of earn on your behalf. So I think it's kind of that difference between a one-time cash payment and, like, true ownership. One very, like, tactical example of this is around the Reddit data now. And the first group of users were very much kind of, like, crypto-native users just coming in. And then people started, like, speculating on the data token.

And what that led to is that if you had a lot of karma, you could make, like, $300 or $400 for your data. And so then a lot of, like, real Reddit users with very good data came in, like, they were new to crypto, figured out how to set up a wallet, kind of got on board. And so I think that that is a really powerful tool to actually bring new people into crypto, right? I'm just thinking if there's anything else to add. People have different, like, go-to-market strategies.

I know Ryan, who built the Tesla data now, and he used to work at LinkedIn, selling LinkedIn data, was telling me how cheap it is to actually get paid posts from LinkedIn influencers. He's like, you can take over the whole LinkedIn feed for, like, $100,000. In his case, his plan is to just, like, yeah, pay a bunch of LinkedIn influencers. Often it's specific to the platform. Like the Biohacker DAO, for example, they have a bunch of, like, biohacker existing communities of people who contribute their data in.

Maybe I'll actually look at a few others. Okay, so... Synclerium, I don't know what their kind of, like, marketing approach is. Oh, actually, Datapig is one where they are very much kind of, like, having you contribute your trading data. I actually, I mean, I think their go-to-market strategy is basically, like, really good memes, which I realize is not, like, that much of a strategy.

But yeah, it kind of, like, depends on the data DAO. They have, like, a very engaged community and kind of all this lore built up around, yeah, the Datapig kind of character. But that is actually something really good to call out, where starting a data DAO is not just about deploying the contracts and implementing proof of contribution. It actually is a pretty significant undertaking to onboard a large group of users who are willing to contribute their data. And so something that has actually been interesting to see over time for us with the data DAOs is early on, like, maybe last summer when some people were coming and building data DAOs, we were kind of nervous if someone didn't have an engineer on their team because we were like, well, how are you going to build it?

Like, is this going to happen? And we've seen that actually teams that are really go-to-market focused have done really well and been able to kind of build while also really driving forward that, like, go-to-market piece. Any other questions on, yeah, data DAOs or on data access? Yeah. Yeah, so when a user or when a data DAO votes to say, yes, you can get access to my data, then the data is decrypted in the TE and you can run operations within that TE.

For some model training today, there isn't a TE that works really well with GPUs. And so in some cases, it is running on a server, which is kind of as secure as possible, but not directly a TE in cases where you need to train, like, a relatively large model. I think we just, we switched from SGX to TDX. Is that right? Well, yeah.

Any other questions? Yeah. Yeah, that's a good question. So the way it works right now on mainnet is that it's a manual approval process, right? So each data DAO is, like, literally being, like, should I run this code or not?

And obviously, that's a lot of friction, but it also, I mean, it works, right? I think that what we are building towards with automatic API access is being able to verify that more programmatically. So one example that's used, one of my favorite data sets is the Panel Study of Income Dynamics from the University of Michigan. They follow people over, like, a 50-year period, like 100,000 people, everything, like your work, your health care. You can train really interesting economic models around that.

And they have rules where it's, like, you can't take any data out unless it is a summary stat aggregating at least 21 people, right? And so there are kind of other practices like that that can ensure that any of the operations being run aren't actually kind of leaking the data. Any other questions on data access? Yeah. Yeah, so actually storage is a really interesting question, where often people assume that VANA has a storage layer built in.

VANA is storage agnostic. So when you're connecting your data as a user, you actually connect your own storage. So you can connect, like, your Dropbox, your Google Drive, or you can choose IPFS. Are people familiar with Urbit or Solid Project at all? Okay, so those are some of the earliest, more, like, federated architectures, where basically, like, you host your own data and you can use it.

So VANA follows, like, a similar architecture, where it's actually a user-controlled storage environment, where they put a copy of their data that's encrypted with a key that they control. But when they join a data DAO, so, like, if I've connected my LinkedIn data in, and then I put it in my storage environment, if I join the LinkedIn data DAO, I might – and they give me a token, then I might delete that data, right? And so what the data DAOs do is they can hold a copy of the data basically in escrow, which is encrypted with a key that they control. We're actually trying to standardize storage across the VANA network, because we found – like, people have used a bunch of different things for storage. And so, yeah, if there are kind of, like, specific ideas around storage, it's – like, people use all sorts of things, like GCP, IPFS.

But, yeah, that's how storage works. Yeah, so the way that data – and that's, like, really important, right? You want to make sure that it's not fake data. So the way that's ensured within a data DAO is through proof of contribution, which is actually measuring and recording on-chain that you have verified the quality of the data and run this Python code that says, yeah, this is good data. And that gets written on-chain, and it's readable by anyone, right?

So then you can know, hey, this data point living here, if I decrypt it, it meets these properties. Yeah. Okay, okay. Okay, so available data. Okay, and then the last thing you can build on VANA, which is the thing that I put the two construction signs next to so the engineers don't get mad about me telling you guys this is live, you can use data portability.

So when a user has joined a data DAO, they've encrypted a copy of their data with a key that they control. So, for example, I could have connected, like, my sleep data, my mental health data, and my 23andMe data. And then they can log in with their wallet and actually just decrypt their data client-side and pass it to an application that wants to use their individual data. And so if you're building, like, a personalized AI agent or anything, kind of, especially in the health and longevity space, I think there's a lot of opportunity to use some of this data. So, yeah, here are all the different things you can build on VANA.

Again, data DAOs are kind of the most production-ready, self-serve thing. The Datify applications are, I'd say, very early. And I think that that space is going to be supercharged by the VRC-13 upgrade, where the VANA rewards go into data tokens. And then the AI apps, that was a preview of data access I showed, and data portability. If you want to dive deeper into VANA, this is a link to our developer Discord, data.

wishlist. If you have an idea you want to chat about, here are the right people to talk to. So if you want to talk about data access, you can talk to Chris or Katas. If you guys can raise your hands. Wait, where is Chris?

Oh, maybe they're outside. Okay, data portability also could offer VOD, data DAOs. Muskan, who I think just joined us this week, so don't ask her too hard of questions. And you can ask some of us other questions. And then here's also a link to these slides.

There are a bunch of links in here. Feel free to use them. Some of them might not be production-ready. So, yeah, it's kind of in developer mode. And then finally, we have a global hackathon kicking off at the beginning of March.

So if you want to take place in the hackathon, here's a link to be able to apply. Great, thanks so much. It was fun. This is kind of the most...