Session 2 Web3 开发者如何构建高价值数据集？

Transcript

Hello. Hello everybody. Hey everyone. Hey. Hey. We'll wait for another like two minutes and then No worries. Let me let me get my slide deck out to share folks. Yeah. [Music] This is very exciting. Hello everyone. For those of you who think that I look

tired, it is because it's quite late here. So, I apologize if I feel a little bit I'm speaking a little bit out of it. It's uh it's 11 p.m. where I'm at. We'll power through this. Let's give another one minute art and then we'll kick off this. No worries. This is like this is one of the most important sessions throughout the academy. So I really want like everybody to tune in. This will kind of like help you all to decide like what data you want to build uh throughout the like next nine weeks. It's a very important

session. Uh yeah. Yeah. No worries. I'm I'm very I'm very happy to uh to wait. I think this is also a very important session for everyone. So I expect a lot of folks to have a lot of um a lot of interesting questions to ask as well. Like I'm personally very interested in this session. So I'm just like I'm a listener today. I was thinking that when I was preparing the deck I was like oh you know I feel like I this is stuff that you and I have talked about as well. So I've like packaged it up hopefully in a way for the folks here to understand. Um I could talk about this for hours, but I've tried to condense it into a presentation that you guys can

can hopefully digest within within the time that we have. Yeah. Yeah. Aish, have we started the recording? Yeah. Yeah, it's already started. Great. Let's do it. You can you can airbrush me. All right. Welcome everyone. Hello. Um, it's Art here. Um I am one of the co-founders of Vanna and first of all uh the warmest welcome to all of you to VA Academy. I congratulations on on being part of this wonderful um wonderful um uh community. Um and obviously special thanks to Muskan and Aush for putting together such a wonderful program. I'm

really excited to be here with all of you. Um you are pioneering something that has been a vision of mine and Anna's for several years if if not h the best part of half a decade. Um we we want we we see the value in the data capital that you guys will create as part of this program. Um we're creating new asset classes and you are part you are pioneers and leaders of this this new economy that we're bu building. So um I'm I'm really excited to have you. Um I'm so excited to get to know some of you hopefully on a more personal level over the next nine weeks. Um and I want to I'm excited to talk to you about one of my favorite topics which is the value

of data. So in this session we're going to talk about uh what makes a data set valuable. Um I've developed a bit of a structured framework to think through how to evaluate the value of your ideas at this stage. Um and I've got some hacks to share with you at the end of this session. So make sure you tune in for the entire session to see those hacks um uh for how to spot data opportunities in the market. Um so excited without further ado let's just jump into it. One big disclaimer that I want to share at the start of this presentation is that what I'll be sharing with you are the opinions that I have based on my experience. I have worked in one of the largest data brokers in the world uh prior to coming

to VA. I've also worked on the sourcing side of things doing human subjects research for one of the leading scientific institutions in the world. So um this is my personal view on what value data has. The reason why I say that is because we're creating a new market here. The private data that we're going to be cultivating at VA is traded behind closed doors today. So, um, we're going to see market dynamics emerge as we start to create the public market for this data. Um, and some of the stuff that I say today may not be relevant as we see some of this stuff emerge. So, you are really at the bleeding edge of everything. Consider it kind of like

when none of us are hopefully old enough to remember this, but um, when the first ever stock companies were created, people had no idea what stocks were. um and now we trade them and we have derivatives on them and all these other instruments. So what's happening right now with data capitalization is very new. This is a very new market and I'm here to just give you some offerings and some guidance as to how to think through the value of your data. Uh and hopefully that will lead you in goodstead as we navigate these uncharted waters today. So what I actually wanted to do was start off with a little bit of background. Some of you may not know um about my story. So I wanted to kind of share with you how I became obsessed with human data. Um and it really has

been an obsession. So um in in my younger days um I used to do a lot of policy and population studies. So studies about humans and the impact of certain interventions and um and programs on humans. So I was an adviser to the prime minister of a new country called team. It was the fourth newest country in the world and my job or part of it was to track the national growth of the country using data. And then when I went to Harvard, I also researched for the Harvard Belf Center working on assessing the social impact for Fortune 500 companies. And that sounds pretty tangential to what we're doing today, but trust me, it's very relevant because

all of these types of studies relied on me trying to gather as much data from humans as possible to work out quantitatively how their lives have changed. Um, and so this is a picture of a field collection that um I did when I was in Colombia. Um, I was doing a impact study for Nespresso. They wanted to work out how much impact every dollar they were spending in their supply chain had on the humans that it touched. So I had to go to a bunch of farmers in rural Colombia and I had to go and ask them how had their lives changed since Nespresso started buying their coffee beans. Again, really random and weird.

The data collection itself was quite a challenge. Um I don't know if how many of you are familiar with um agricultural communities but they don't write a lot of stuff down. So going to them and saying oh you know how has your life changed over the last 6 months? How has your spending habits changed? Um how has um you know your your travel and transport changed? They didn't really know. And so, um, part of my obsession with big data actually came from a dinner that I was having when I was with my host family and I was lamenting the fact that I didn't have any good data and I I was struggling to sketch out this model and I saw my host mom and she was constantly on her phone and I said

to her, you know, "Hey, Liz, what are you doing?" And she said, you know, I'm I'm I'm on Facebook. And I said, "You're always on Facebook." She's like, "Yes, yes." And I'm like, "What do you what could you possibly be doing on Facebook?" You know, you're a rural farmer. What What do you like about Facebook? She's like, "Oh, I post all the things that I buy for my kids. I buy a lot of things on Marketplace. I talk about the yield. I talk about the weather." And then boom, it hit me. I was like, "Oh my god, all of the data that I came here to collect that I'm going around trying to get from from farmers is in this lady's Facebook." So, I asked her, I said, "Can I please take a look at your Facebook?" Um, and she kindly showed me and I went through and

I had all of the data that I needed for my model. So rather than doing these, what you see on screen is, you know, a 20 to 30 minute field collection, I then just changed my methodology around and just started walking around asking people to see their Facebooks. I was like, I just want to see a Facebook. I scrolled through and I was like, yep, the year 10 months ago was this. Because she posted about it to her friends. Amazing. Um and so this started an obsession with thinking through well how much data is there on each of us in the world and where does that power reside um fast forward several years and I ended up working at AEN. So AEN is and and you wouldn't be blamed if you don't

know what APE is and you've never heard of them. There is a select few companies in the world that buy, sell, and trade all of our private data. Um I joined AEN as the head of impact sourcing and my job was to try to find data that people that large companies were struggling to get. So uh Appen's clients included Google, we did Google ads, Google search, um Facebook feed, we did GPT, Microsoft, Apple, um uh and Amazon. So all of the big tech companies would source their data through brokers like Appen. There are very few of them in the world, but they control a large amount of the data collection uh and private data business. It is a massive

business. Can I just say I didn't realize it until I started working in the industry that it's a multi-billion dollar business. Some of these large tech companies spend more on data than they do on engineers. That's how large and how big of an opportunity it is. But in these transactions that occur behind closed doors, there was never there was very little participation from the people who actually created the data. You and me and the people that we know, we were outside of the transaction. Um, and this is when I got really obsessed with thinking through how how can we participate in this market. And you know, generative AI picked up and Anna and I were like, look, this is a this is a huge problem, not only for people who

create the data, the rightful users, the the users who are the rightful owners of their data, but also for people who want to build technology with it. Um, the the fact that there was this bottleneck in the industry made this data very very inaccessible for anybody who wanted to innovate and create new AI. Um, and what I mean by that is there is a reason why our Facebook news feed uses such advanced technology and yet we haven't cured cancer with AI. It's because of the economic system that drives the data into different products. And that's part of the mission of what VA is about to try to create a space where this data could be publicly traded so that we can

start to imagine new technologies that can be created with our own data. So that's all a bit of background story about me just to kind of give you a bit of context as to how I came to VA and why I why I'm I guess somewhat qualified to speak about this topic. I know you're all here and you probably have data sets in mind. So I wanted to share with you one rubric that I use to assess the value of data and this is my rubric. There are six points to consider when thinking through the value of data. Now it's tough to value data in a closed market. As I mentioned based on the experience that I've had at

Appen like a lot of these transactions happen behind closed doors. So a lot of it is quite theoretical at this stage because the public data market doesn't exist. We're creating it. But these are the principles that we applied in relation to data sourcing and data value in the industry. And I wanted to share them with you because they're a useful technique for you to think through how to value the data that you're considering. So I'm going to discuss each of these in turn and we'll do a recap. Um as you have questions that come up, feel free to post them and we can answer them at the end. Um but start to think about these as they relate to the potential data set that you're thinking through. So the first principle to think through for your data set value is scarcity.

Just like any supply and demand market, scarce data is valuable data. If it's out there in the world, it's probably not going to be super valuable. And the best illustration of this is private versus public data. So public data is stuff that you can scrape on the internet. It has become a very commoditized industry. What I mean by that is the ability to produce public data efficiently is based on the power of compute, not on any sourcing power and nothing related to the data itself because you just need an efficient comput engine and you can scrape all of that data. Private data or platform data on the other hand is stuff that can be very very valuable because it exists

behind closed doors. So uh at the very extreme if you want to think about scarcity you can think about the dichotomy between public data publicly available versus private data platform data and some you you should be able to fall somewhere in between. Some really important signals of scarcity are and these are kind of the yellow flags to look out for if you're thinking through a data set. Um, if you're thinking through a data set from a platform and they've started to lock down their APIs, it's probably a sign that they see value in that data and they're creating scarcity in the market. They're trying to reduce the amount of data that they have that's public. Uh, we saw this with Reddit. We saw it with Twitter. If

anybody try has used the 20 that used the Twitter API since 2023, you'll know that it's like ridiculously expensive now. And then Spotify started to crack down on its developer API. Why? because they're trying to create these scarce environments where they own all of the data. Um, and it it's a when a when a platform is trying to restrict access, it's usually because they know the data is valuable. The the trick is that data laws actually allow these collectives to bypass these scarcity walls. Uh, individuals in most countries in the world have the right to reclaim the data from the platforms. And your jobs as leaders of these collectives is to help

those individuals come together and structure their data so that they can almost crack through these walls and then come back together and create these fertile data gardens. So the question to ask yourself in the scarcity bucket is how rare is your data and consider that as it relates to your data set. The next thing you have to think about is depth. So when I think about data depth, I think about how much information one has in we call we call it a row in the in the industry but per entry. So like you have one I guess data data point. How much information do you have within that data point? How many uh do you have chats, clicks, preferences,

location, device info, time series logs of the same person? And that refers to the depth of a single data point. The more depth there is, the more optionality there is for training, personalization, and analytics. Um, depth enables multivariate modeling and better signal discovery. So, let me try to try to illustrate that with a practical example. If I only had one of your chat messages, say on like Facebook, I could and I aggregated one message from every single person on this call, I could certainly understand some stuff about you, right? But if I got one message, if I got all your messages actually, if I got your feed, if I got

your device, if I got your location, if I got your age, I have many potential ways that I could model out the relationships between these variables. And so the deeper the data set that you get across several people, the more valuable it could potentially is. So in this category, I'd like you to think through your schema. So that is the structure of your data. and ask yourself how rich is your schema? How many different elements do you have in there about a single row or a single data point? The more elements that you have, the more likely it is that the data is valuable.

The the opposite or I guess the adjacent um problem to look through here is breadth. So breadth is the volume, diversity, and generalizability of your data. So that's how many people you have, how diverse that sample is, and how generalizable it is to that specific use case. And I'll tap into that, double click on that piece in a moment. But when I was at Appen, uh it was a very common word that was thrown around was quotas. So um what would happen is somebody like Facebook would be like, "Hey, I need 10,000 iMessage data points." And you'd be like, "Okay, sweet." And then as part of the contract they'd say but we want a certain percentage quotota amongst this group

amongst this user these types of users in this geography or uh you know for for a lot of visual data collection it was we need these categories of races or this these type of skin color and we need this this number or percentage. For them it was important because they were developing a product that was used by millions across a very diverse societal group. Um narrow coverage is okay if your model is for a specific group. Um so example of that is like the Reddit data model that the Reddit data d created was specifically for Reddit posters. So it was okay that the the the sampling was a bit biased. But if you're looking at something that's supposed to be for society, if your data set is supposed to be broad enough to have a

balanced view, um it it might be very important to make sure that the way that you're sourcing that data has breadth. So you're looking at diversity, you're looking at volume, you're looking at generalizability. It's the fundamental question here is how representative is your sample? And I'd add to that, how representative is it in relation to the use case? Um, so, uh, one of the tips I I have little tips layered in here. So, this is like your little alpha. My little gift to you is if breadth is is an important part of your data set, consider dynamic pricing in your tokconomics so that somebody who is potentially a rarer person or somebody with a unique character trait

might actually get more data, more token uh, to encourage them to participate participate in your data set. um dynamic pricing here can be your friend and I think we have different sessions to talk about tokconomics but I'm just kind of layering in some tips here so that you can start to think through um different elements of this. Um the next thing to think through is structure. Um structure relates for me to your schema. Um for folks who don't who aren't super familiar um when we look at data sets the first thing that I always ask for is your schema. So basically what's the structure of your data? What rows and columns do you have? what kind of operators are there in those rows and columns? So, are they binary? Are they boolean? Are they radial? Are they text?

Are they numeric? Because that structure impacts my ability to be able to generate uh to to be able to use that data for a specific model. Um, usually uh that there is a stronger preference towards a really well- definfined schema that's cleanly formatted and has limited noise. If you have a lot of randomness, um if for example you're just getting a bunch of text data stored in a JSON file, it can be really hard to pass that out. It's unstructured data. Um some people will be okay with using that, but some people don't want to go through the effort of thinking through how do I structure that data to be able to use it in AI. The most common issues that I've seen that have really devalued data sets are where there's misaligned data types.

So for example, uh where instead of using a numeric operator, you've classified a term as like a boolean operator. Um you've got to make sure that the way that you've actually structured the data is consistent with the type of information that's coming in. And the other thing that really devalues the data structure is messy data. When you have a lot of null fields, it makes you question the validity of the data. So your proof of contribution here is very very important to ensure that your data is very well structured. Um the the question to ask around structure for yourself is how clean is your data?

Um I'm I I I would actually say that I'm a little bit OCD. So a lot of my data sets tend to be very clean. Some people love data cleaning. I'm one of those people. Some people hate it. So um this is something that you'll work through. Um I think uh there was a question here about are you learn are you going to learn about this? You you 100% are. This is part of the data refinement process which I believe will be explained in technical workshops later on uh which allows you to define that schema. So just keep this in mind that the structure can be very very important in terms of the value of the data set.

The last piece to think oh sorry second last piece is timeliness. So some data sets age perfectly well. So um there's a fantastic data d that exists on bar right now which is called asterisk data dow. They are a women's non-reproductive health data d and their mission is to try to advance medical um medical intelligence for women. So advance medical treatment for women. Um human bodies um tend not to change as rapidly as for example sentiment on Twitter. So they their data can age fairly okay. If if you have um a person's data point in that data set, it might be valid for several years. And some people might actually like to look at the time series data way back in

relation to that data set. But there are other data sets where the timeliness of the data matters a lot and the the the value of the data decays really really quickly. So um you need to work out what type of timing is relevant to your data set given its potential use case because timing goes to relevance. It goes to pattern visibility and it goes to if there is real time inference that needs to be drawn from your data set the power of that inference. There's an example of a data set that we're looking at right now where uh there's a partner that wants to measure um real time inflation.

So inflation figures are published like monthly at best at some places. They want to do inflation figures you know every hour. So for that the um the updating of that data set needs to be very very uh timely hourly in fact for them to be able to gather any signal from it. So think through your data set as to how how timely it needs to be in order to acrue value because if you get this question wrong, you could have a really powerful data set like Twitter that's just really stale and therefore not very valuable. Um so you need to think through how relevant is your data over time. And my little tip here is your tokconomics. They really matter uh in this one. So if you've if you've

created tokconomics where somebody gets incentivized for doing one um one action and not and there it requires subsequent actions to them for them then you you may not be getting um a great data set here that is that is timely depending on what your need is. So so let me just concretize that for you all. uh if you wanted to get the Twitter data set from people and you give them a huge reward for connecting their Twitter and doing that little bit of a snapshot the first time and no rewards for continuing to update their Twitter data, you might not have the best data set. So think about that in relation to your mechanics and your tokconomics because that's what really matters here. The sixth and final and I I thank you for bearing with me.

There are a lot of them but the sixth and final thing to think thing to think about when it comes to data value is strategic utility. So this is actually a catchall. It is I I put it last but it's actually the most important thing. You can have a rich and structured data set but if you can't squint and tell me how that data set actually solves a real world problem, it's not going to be commercial. It needs to somehow improve a model's performance. you need you need to be able to tell me a story of how it how it improves something, how it makes something more accurate, how it unlocks an insight, how it advances an AI model.

So uh I'll give you a couple of examples. So there is currently not a data set out there that exists that um is human movement like a kinesthetic data set on human movement and there's a lot of humanoid robots that people want to produce. So one can tell a story that if I collect human movement data I might be able to improve humanoid robots. So you need to be able to tell a story kind of like that in relation to your data set. And if you can't um you should workshop it with us. Workshop it with the team so that we can kind of guide that story for you. If that that story the clarity of the utility of your data set is very very important because if you can't tell a story it's highly

likely that that data set may not be important. Um, so the question to ask yourself in number six is what does your data set unlock? You've got to have a very clear idea about what your data set unlocks. So quickly just to recap, scarcity, is your data set hard to access or replicate? Depth, how rich is your data set? Breth, how diverse or representative is your contributor pool? Structure, is your data clean? Is it labeled? Is it schema aligned?

Timeliness, is it current, relevant and refreshed? And strategic utility, what problem or product does it power? That is the rubric through which I I analyze data value when people ask me to. It's the rubric through which the industry also sees um different data value. Um so I encourage you to apply this to some of your thoughts and um and uses to assess your data. Now, um, a presentation from Art wouldn't be complete with some hacks. I have cheat codes for you all because I'm the kind of guy who likes to optimize my life. I hope that the cheat codes will help you to uh either inspire you if you're kind of stuck on a data

set or um will help you to get to your answer faster. So, here are my cheat codes for understanding the value of data. My cheat code number one, draw inferences from the limited public market. So there is a small public market for private data. It exists in niche places. It's not openly traded, but you can start to draw some parallels between your data set and some of these public markets. So look in data marketplaces like AWS Exchange, Snowflace, and Hugging Face. See what's popular or what's selling, and you can use this potentially as a proxy for your data set. I wouldn't feel too bad if your data set isn't listed here because again, these are very infant

marketplaces. But if you see, for example, I don't know, the Twitter data set from 2023 going for $100,000, you could probably bet that a 20 a Twitter data set from 2025 will go for a certain amount of money, right? So, you can draw some parallels and inferences based on what you're seeing selling on these on these platforms. one that I mentioned earlier, see where platforms are locking down on privacy terms or API restrictions. Every single platform that I've more I would say like 50% of the platforms, that was an exaggeration, every single but about 50% of the platforms that I deal with on a day-to-day basis have given me revised um terms of uh use and privacy policies in the last year, probably because

they're locking down uh their their their data. So, uh it's a yellow flag. take a look at some of these restrictions because that's probably a valuable data set. Um there um this is another little hack. Um Appen scale LXT um data strike. There is a small group of these kind of data brokers. They kind of sit adjacent to data labelers that gather human data and you can actually sign up to these platforms as somebody who's going to provide data to them. I'm signed up to all of these platforms.

Why? Because it's a huge, it's interesting market intelligence as to what they're paying people to provide them. That's the kind of data that they're getting paid millions of dollars to provide to large tech companies. So, I would encourage you if you're a little bit lost to sign up for Apple, sorry, Appen, Scale, AI, LXT, um, Data Strike, and any of these kind of provide your data companies just to see what kind of data they're asking from you. um because that could be inspiration for data that then you can also put into the market. And then keep your ear to the ground for licensing deals. A lot of these licensing deals uh they definitely were not public at the time that I was working um in in one of the data shops.

Um they're becoming more and more public now. So, set up a Google alert for data sale, um, data license, um, and get pick up early some of these data licensing deals because often times they'll announce the deals, but the the cycle of data collection takes months, if not years. So, you could potentially frontr run a deal if you're able to crowdsource some of this data and then be able to sell it directly to the person who is um, who's who's on the other side of that deal. So that's my cheat code number one. Draw inferences from the limited public market if you're not quite sure what data you're wanting to aggregate yet.

Cheat code number two for you all is to and this is going to be controversial step away from web 3 builders for a moment and go talk to some AI builders. go and talk to people who are actually building some of the cutting edge technology cuz um so so this this group um this this is from a website it's called Build Club. They're the largest group of AI builders in Asia Pacific. They have 10,000 builders and 11,000 AI projects that have been shipped to date. Um I got involved with these guys pretty early on. Um, and I absolutely love them because when I posted in the in the Slack in this group, I was like, "What do you do if you can't find your data

set on Kaggle?" Uh, Kaggle is an open source data set, um, repository. They all had so many suggestions and ideas. Um, so, uh, AI builders are some of the nicest, friendliest, best people. I mean, just like you guys, people that I've I've, um, I've encountered, and they have plenty of data needs. So, uh, join your local AI hackathons. um stay involved in the AI building space because that will lead you in greatstead as to what's happening in the space and where the data demands are um because these guys usually can't access some of the big data sets that the large tech companies can and they're reliant on open source data right now.

The last thing is cheat code number three use your industry instinct. So um I'm not an industry expert in medical data. I'm not an industry expert when it comes to manufacturing data. I'm not an industry expert in um in even even human movement human subjects data. I'm I'm really not because I I'm the person who gathers and provides the data for the industry, but I think some of the folks on this call, you come from a specific industry. You kind of have an inclination about what sucks in your industry or an industry that you're well aware of. Um I think um you know knowing what AI does not do well in your industry yet is a really

nice entry point into asking why and probably figuring out that in 90% of the the instances it's because they don't have enough data. So, um, think through an industry that you're familiar with, what AI is doing in that industry, and where it's capping off, cuz that's a really nice point of entry, and you probably are familiar with that data set. And I would encourage you for your first data set to to be really familiar with it. So, if it's coming from an industry that you're familiar with, that's even better. You'll be you'll be a great um you'll be a great steward for collecting that data and bringing it to life.

So, a bonus cheat code for everyone here on the call is uh click we'll share this slide link. Go to ki.cc/bana data bounty. Um this is a present to you all if you're stuck. Um we have a list of data sets that we have demand for. Um, and if you're lacking inspiration and you want to use these data sets as the basis for your your your data, feel free to. Um, this is our far data bounding board or we regularly update it. There's a couple of data sets there now that people have picked up and are building. Um, so if you're interested, take a look at this, take a look at this

site. It's just a notion dashboard. um and email Sam at VA Foundation or Muskin at Vana Foundation to discuss your interest in building any or all of these data data dials. Um we give a little bit of context but you know we can we can talk further about the requirements that some of the data clients have surfaced in relation to these data sets. Why I like them is because um usually when there's one person that's interested in a data set, there are a lot of people that are interested in the data set. Um so that that's that's really cool there. So that's my little gift to you. um go to the VA data bounty board and see what other data sets there are. That's that's an additional VA gift. So, just a quick recap for everyone. That was a lot of info and then I'm going to jump right

into the questions. Use the data set value rubric scarcity, depth, breadth, structure, timeliness, and strategic utility to do an assessment of the data set once you have one in mind. If you don't have one in mind, think practically. Look at some of the data marketplaces. look at where APIs are being restricted and then look for some licensing signals. Get that Google alert up to try to work out where these licensing deals are because that'll give you some indication of where data value resides. And lastly, you can find some info on the data bounty board or talking to the ecosystem team. Um, Muscan and Aush are fantastic. Um, I know that they're they're somewhat new to this space, but they're catching up really quickly. So, um, definitely lean on the team for some support as to what data set to build. Lastly, I want to say um I

I don't know if you've gathered, but I'm an absolute data nerd. I could talk about this for hours. Um this is a lot of fun for me and I want it to be a lot of fun for you. Data is super fun because you get to learn and work out things that you never knew were existed in the world. Um when I did that field collection in Colombia and was able to actually work out the impact of coffee on these farmers, it was it wasn't just my job. It was just insightful to see how people change over time. And um I want you to have fun with the data that you're you're collecting. Make it interesting for you. Um and lastly, but not least, welcome to the family. Um thank you so much for applying. Thank you so much for being with us. I'm so

excited to have you on our journey, and I really do mean it. I want to get to know some of you and some of your projects. Um I will open the floor for questions. My preference is that they're more general questions for the group. If you have specific questions about your data set, we can answer them in Discord or in one-on-one sessions, but if you I'll I'll kind of try to offer insights that are more generalizable to the group in the question time. Um, MCAM, should I just jump straight to the the um Yes. Yes, that helps. Okay, let me go let me just quickly skim because I want to make sure that I'm getting some of the top ones and then I can go back. Wow, you guys

are you're going to take a lot of meeting notes. Um hopefully um we'll we'll share some of the meeting notes and the slides in um in Discord and then this this recording will also be available publicly so you can access it. So let me just quickly go through. Okay. Are there evergreen areas or based on what you say, is it better to look at rare categories? Ah, good question. Evergreen areas. Um, it kind of goes in waves. Um, and this is from Asha. Asha, thank you for asking that question. Um, so it goes in waves.

Um, what uh when I first entered the data collection industry, language data was very important. AEN is actually a language data collection specialist because at that time people wanted language data to make their um large language models better. The ones that we now use that we kind of take for granted today. Um I'm seeing different waves come through. So language is still important but not the type of language data that was before. They were really focused in on conversational data before because these lang large language models kind of used to sound like robots. So now they sound a little bit more human. So they wanted like people's messaging data and stuff like that. So now that's in there. What are they looking for in language? Now

they're looking for niche language data sets. They're looking for audio conversation data. They're looking actually one of the areas which is kind of interesting that they're looking at is um uh children's conversational and children's communication data because a lot of that data is not public and it's and and if you try to put an LLM in front of a kid, they wouldn't be able to use it. Um so if uh so so language was one wave. Another kind of and I still think there's a lot of juice to be milked in the language space. Um in another one that's really emerging quite now right now is is robotics. So uh what's what's driving AI will usually also drive data. So a lot of the

robotics research where we don't actually have the foundational data sets for that. Um that would include visual um hum human interaction from a firsterson visual perspective. So imagine like you have a camera here or you're using the the AI glasses. That footage is very important for a robot because that's how their sensors also operate as well as the kinesthetic data that I was talking about. So having sensors on your body, body sensory data, um and human movement data also very very important. Um so uh I would say there's no such thing as a kind of evergreen area, but you can see that these areas kind of kind of start and then they grow depending on what's happening in the industry. So they move kind of in epochs. Uh is there a comprehensive list of data dials on top

of VA collectives for mainet or test net just to avoid overlapping efforts? Um Muscan is that on the website? You can hold keep me honest on that. Yeah, we have one website but that's not updated. We should be actually updating that in like a couple of days. So yeah. Okay, hold on for there are a few actually. Yeah. Yeah. Okay, I hold on hold on for a couple of days. We will we'll add a couple more on there um so that you can um you can see all of them so that you don't double up effort. What I will say though is um if you did see a data doubt on there I wouldn't be dissuaded from also trying to source that data set. Um the way that the market is working is that um or is

designed to work is that the data set that's of higher quality for that specific use case will win. So there might be a Twitter data set, but you might not agree with their methodology. You might think that you can source it better with a better proof of contribution and a better sourcing pattern. Go ahead, vampire attack it. Like I think that that's that's exactly what we what we want folks to do. Um, Build Club. Yep, that is the link. Um, if you guys uh Annie is the is the head of Build Club. She's a friend of mine. If you're having issues trying to get in, just ping me and I'll I'll I'll let her know that you're part of the club. Um it's a very very cool cool organization. Please join it. Um does the utility only have to impact AI

training or a world problem or curiosity in general? That's actually a really good question Asha. So um the way that I think about that utility point is you need to be able to just tell a story about how your data is going to change technology. And I would stress test that with an engineer or a product manager just so somebody who knows AI just so that it's plausible that those two things connect. Um so it could be that your data set is going to change the world directly. That's okay. Like the asterisk data dial, they don't quite yet know what model they're going to build with their data, but they know that if they aggregate a lot of women's health data that they can present that information for the

improvement of medical science. They don't know the exact thing that they're going to build, but that data is very important. There's a fantastic history behind that data dial that I won't go into, but um you should definitely look into them. Great team. Um so your data can have a direct impact on the world. Yes. But you need to be able to connect those dots and somehow connect those dots with the data as the thing that's actually driving the change. So make sure it's not just a correlation, but an actual causation. Uh we yeah you can definitely get the slides. Hopefully it'll be up posting soon. Do we learn schema and how to do this? You should you you will as part of this program learn about data refinement which will lead you to your schema. So

actually if you wanted to play with different schemas um just try downloading your data from any large platform like Twitter um Instagram, Facebook um just you could even just Google like GDPR export and the name of the platform and download it. It will give you data in a specific format. Actually let me see if I can show you my Google takeout data. Um, uh, let me see if that comes up. Yeah. Okay, I'll just quickly share this with you. So, if you use Google Takeout, for example, this is how you get your data off Google. It has all of these different schemas. It actually has so

much data about you, but it also has plenty of these schemas. Um, this is, speaking of depth, this is the deepest, one of the deepest data sets that you can get. Uh Google has Fitbit, Gemini, Google account, everything. Google chat. Um and you you can actually export all of these in the one file. Um and each of these services has a specific schema or a specific structure. So if you wanted to just start even exploring or playing with data schemas, uh that's definitely one way to do it. Um just just download your data and see what structure there exists out there.

Um, Reddit is using Antropic for accessing platform data without permission. It's huge. Yes, that's a it's a big problem. Um, uh, the biggest problem that I saw actually in the data collection space was even when people were compensated, they were compensated a very small amount that didn't actually represent the value that they were creating. So, um, there are lots of AI companies that, um, ethically pay for the data. Uh but I think that the multi-billion dollars they're getting in revenue isn't actually flowing back to the data that they're they're gathering. So uh it's definitely a concern that we're hopefully going to address by creating an open market. Um and I think these are all AI meeting notes. Um

so let's see if there's anything on the bottom. Um uh just want to address all of the questions all here. How does scarcity or strategy translate to utility or incentives for data contributors? Umar, that's a really really good question. How insightful. Um, so your tokconomics one area that we're that I'm really like really really focused in on right now is pairing data tokconomics with the economics of data collection. Um, the dynamics of data collection are a very subtle art. there's an art and there's a science to it. It involves a little bit of UX. It involves a little bit of

incentivization to try to get the perfect data set. Um, and so when you think about the scarcity or strategy piece, you need to think about how to effectively source that data as how to source that data as effectively as possible. And your greatest tool in web 3 is going to be the incentives, the tokconomics. Um uh example, if you wanted a very uh if you didn't care about um the diversity of your data sample, you might put in really strong referral rewards cuz I'm more likely to refer people who are like me because they're in my immediate network. And that could be very good for your data

set because you could grow that quite quickly with people who are very alike. But if you wanted a really balanced data set, you might not actually want a super strong referral program where you award points or referrals because you might want people from who are quite far apart from each other to contribute to your data. So there's subtle art forms that you need to think through as as you design your tokconomics um to make sure that they match the dynamics of that data collection. Um, for data sets that can be quite difficult or quite personal for some folks, I would recommend something akin to something like a bonding curve so that the initial contributors of your community become

your spokespeople and they're the ones who provide you the data first. So, think through rewarding your early people more than your later people. Um, and then think through how you diffuse the the collection um through your community and how you leverage your community in order to to kind of diffuse the collection. So um multi-dimensional problem I I don't Muskan do we have like a session to talk about data tokconomics with folks or the tokconomics piece? Yeah. Yeah. We have two sessions actually. Yes. Yeah. Yeah. Super important guys like I think that like this is the crux of it. We we've been working on this problem for half a decade and um the the the incentives that you provide and the structure of those incentives really matters to get

the the right dynamics. So the data in it should match the data that you're getting. if you can't use the same data tokconomics for Twitter as you would use for a cancer data dial for example two very very different data sets so we'll need to work through that with you um very carefully um Nate thank you um who are you seeing currently buying data from VA gardens mostly AI builders or data brokers active too so we haven't sold any data to data brokers they probably don't like me that much but um I'm actually still good friends with some folks who are working at some of these companies um uh good friends with COO of LXT um and some folks there uh and folks at data strike as well and still folks at API. So I I meet I meet with these folks every day um they

haven't indicated a preference to to buy data through the VA platform. I think the product that we're producing is pretty new. Um we have gotten a lot of traction from research institutions um and uh kind of I would say like early innovator early innovators who have AI products. So, um, folks like Flower AI, um, that's one that I can announce. Um, two of our data dials, um, a very, um, one of them has, uh, presold their data to a hedge fund because they like to get insights. Um that one you can see it's diffusion VFSN and another one has sold their data to a uh uh they they have vehicle telemetry data and they have sold that to a I

believe an insurance company for insights as well as a battery manufacturer for insights. Um one of the subtleties here is because you're using VA you can sell or grant access to your data several times over. It's not a single export of your data to them. So that's why they're able to monetize it time and time again. That's why it turns into capital. Um so uh I think that there's a lot of opportunity for those guys to continue to resell their data as well. Um Raph, my parents own a generatorial service with over 100K employees. How best do you recommend I can collect and structure data on a day-to-day basis?

That's very very interesting. Um, I probably need a little bit more information, but to kind of generalize this to the group. Um, VA is about self-s sovereignty and user own data. That's one of the core ethoses that we have. Um and so um while we can support institutional exports or imports of data um I think that it's really important ideologically that you ask yourself who should have the right to grant access to this data and how can I make it easy for them to gr grant to have the right to give you that access. Um, so maybe in this example, what I would recommend is if you have 100k employees, you can have

them opt into some sort of data sharing system where they could also provide um there's some sort of benefit for them as part of sharing sharing their data. Maybe they get insights um about themselves or something like that in addition to the tokens that they'll get to govern the data set. But I I would be reluctant to uh uh to kind of to to take their data and just say like this is part of the data d without like getting their actual self-s sovereign consent on on that. Um but I'd like to understand the use case a little bit more. Um Ralph, so definitely post in discord or or ask Muscan and the team. Uh Gary, are there processes to request personal data from platforms? How high bureaucratic is it and how soon will people be give up

ask? That's a very very good question. So if you guys haven't tried a GDPR export from a platform yet, you should definitely try it. Some platforms are much easier than others. Um the general requirement is that platforms are required to give you your data within 30 days. Many platforms give you your data much sooner than that and most recently you should search Utah Digital Choice Act. Um VA has a partnership with Project Liberty. The main um drivers of the Utah Digital Choice Act. What that requires is that uh there is immediate programmatic access to user data for any platforms that are serving users in Utah.

That is where the world is going. We're also working with the EU um on GDPR reforms and we're arguing for programmatic immediate access to your data. There is absolutely no reason why a tech platform cannot grant you immediate access to your data when you request it. Uh they say it's for security reasons, but I don't know. Do you have like a bunch of little data fairies reviewing every single line of my data code? That's crazy. You They should be giving you your data at the click of a button and the laws are changing very quickly to um to give you access to that. There are platforms that will just give you access to data quickly like WhatsApp. I believe Chad GPT you can export really quickly. There are some platforms that make it a really big pain in the butt to do it because

they do the minimum compliance. Um, I would suggest uh doing your data flow um testing it out um but also keeping an eye on the regulations because they are changing really quickly and soon we will hopefully have programmatic access to all data um and full data portability. Um Asha again thank you so much. How do we collect data other than our own solo female child? Do we ask others so we know on Facebook? Ah this is cool. This is a sourcing question. Yeah. Yeah.

Yeah. Um, how do you source data? Um, this is a lot of my a lot of the problems that I was encountering at Appen were how how do I actually find this data? There are many channels through which you source. The most powerful way to source data is to create a community around your data set. Find the people that care about it. So, uh, if you want data from solo female travelers, think about an insight or a product or something that you can provide them as part of your data collection flow that could be useful to solo female travelers. Ideas could be uh it could be an app that if you're looking about looking at like um their location. I have friends who are solo who have been

solo female travelers and one of the things that they're really mindful of is safety and security. So could this be an app or some sort of collection flow where you say hey if you share your data with us on a continuous basis uh there is something that the there you know we'll um I don't know we'll will increase your protection by making sure that this data goes to the local authorities or something like that so that they know that they're always protected. I'm just spitballing it. That may not be a great product, but um giving your community something that they want as part of providing the data, providing the data to the collective so that they start to create a community around a purpose. That's going to be the best way to source that data because

then they become your advocates and they can start to um evangelize your data set to other people. Um team Cremont, we're bringing building a credit card data D. Yeah, we're using that. That's a huge one. What are your thoughts on the feasibility and value of this data? Are there specific dimensions of credit card data that you believe would be most impactful? So, I will get specific on this one because this is something that is a very valuable data set um and something which illustrates a very important piece. There are like the credit card companies can actually get they actually buy and sell people's data already. um there are alternative markets uh for that uh which should be created because you know why should they have the full monopoly over buying and

selling people's data so this is a great use case or an example of there is already a market for this data but I don't think that market is efficient and that market is certainly not fair so consider participating in that one great great work you've probably done your research and seen this the other principle here is that um the granularity ity of the data. So this is a depth problem that you get from credit card purchases. So if you look at your credit card statement, it usually only says the the merchant and it has maybe you could probably get down to the granularity of an item, right? So problem with that is that you sorry not problem but opportunity there.

If you can get skew level data, this is always something that's been asked of um a lot of data collection folks. skew level data is like um the the skew is like the size, the color, the specific item that you're actually purchasing. Um if you can get people to report or provide that information to you, that could be very valuable. And I've seen people try to do this with like receipt collection. So like they would try to cross preference credit card data with actual receipts. Um they've tried to scrape receipts off online things. They've got people to forward receipts to different banks. um that is one way to enrich that data set with additional depth. So think through that if there is already a data set that is out there that's bought and sold and traded or

even if it's public can you enrich it to create value on that depth axis because that could be very very valuable. Um Gremont if if you want to chat about this I'm very happy to talk to you offline about it but this is something that um I know a lot of folks in the industry are working on right now. You're you're quite you're quite ahead of the game on that. Um Asha's not taker is taking great notes. Hey, how you doing? Good. How are you? Very well, thanks. Very well, thanks. How's um how's the uh the health data dog going? Yeah, it's going good. I was going to ask if you knew any in your former life which companies I know of a few that work as the that collect health data from uh the providers EHRs and then

take it and resell it to life sciences and etc. Do you know of any on your end? Do you know of any off the top of your head that I I I'll have to tap my networks for that for you, Josh. It's because it's such a heavily health data is heavily regulated. You usually don't find the generalist data collection or data brokerage companies touching that data because it's too much administrative overhead. In fact, one of the screening criteria that we had was is there any HIPPA data collected? Yes. Because then we wouldn't even do the job. there are specialist companies that deal specifically with this type of data. Yes. Um and uh I would say that for for anyone who's doing health data

in the United States uh in the United States specifically um make sure you workshop it with us a little bit because no I know I've been in the business for a long time. I know HIPPA and I know all these different deidentification protocols and stuff. Yes, I'm I'm very versed in that. But uh there are some there I would say that there are some exceptions to the deidentification protocols in GDPR for um bonafide end users as well as users providing their own consent for data sales for example.

That's how I think um that's how I think 23 and me managed to get through a lot of their um regulation. So um I think it's definitely something to tap into and explore. I'll tap my networks to see if there's anybody specifically on the healthcare data. Um, but that's a good point, Josh. Like, um, I speak from a very generalist perspective because I was at like this kind of generalist, um, data shop. Um, but there are specific niche brokers on specific types of data and health is one of those sectors where there's definitely a niche broker. There is one good news is that the Trump administration right now is in the process. They have an RFI out and they're uh asking about things like info blocking and Trump Trump himself and the

RFK are both proponents of freeing the the patients to their data. So yeah, um I'll tell you a quick um anecdote before I jump to Muskan because he's probably like you need to wrap up right now. Yeah, one of my anecdotes is that um there was a I know somebody who I actually invested in a company who's doing longevity data and um they were working out of a a Harvard longevity lab and it took them about 3 years just to get centiggenarian people's data to actually just work out how how can we improve our lives and I'm like oh my god like my great-grandmother was 104 and I was like I could probably convince her estate to provide you that data straight

away because it's such an important problem, right? I didn't realize that health data was so regulated that it was actually stopping research from happening. And that's that's kind of like there is a direct impact if we're able to create those data sets to the world and humanity. So that's like an example of that impact piece. Absolutely. Yeah. Mhan, you're probably telling me to shut up. I'm sorry. I'll I'll shut up now. No, what a great session. Thank you so much, Art. One last question which I really want you to answer is that uh the data we contribute to VA still those platform has the copy of it how that can be solved with VA like and I think that's an interesting question yeah yeah I mean I think that

my my long-term future state on this is that um we have a little bit more choice we vanna uh the VA protocol is a read write protocol so one thing you might not know is that like your your data currently exists on Facebook because you probably use Facebook and it's constantly being written to Facebook. But there is a future world in which tech companies have to comply with user own data where instead of allowing them you can you can allow them to write on their servers if you really trust them. But you should have the choice to be able to say no I want you to write to my personal data device which is on my desktop so that nobody else has access to this data. VAR is designed for that future world where we, you know, we do a

lot of policy work at VAR as well to try to change laws like what we did in Utah where where they're actually going to force them to do that. They're going to force them to put data on an open protocol so that users can move their data around in future. You're not I I pray I hope that you will have the choice to not have to have Facebook store all of your data or your platform store all of your data that you can choose where to put it yourself. Today we're at the tip of the iceberg. So even just breaking the monopoly that these platforms have on these data sets is one step towards saying hey what you're doing is not okay. we're at least going to participate in this

economically and we're going to produce a rival product to what you're putting into the closed market and hopefully other folks will be like I like their product better because it's coming directly from users. It's richer data. It's it's there's something better about that data set and and in most instances there is something better about the data set that you're providing. So um I guess Moscan uh it's watch this space. We're built for the future where we don't have to have our data in these platforms, but for now we should at least participate in the economy.

I think thank you so much Art for joining. Thank you everybody for joining. I think probably our time. So thank you. Bye guys. Happy weekend. Bye. Take care. Bye. Bye. Mask. Thank you.