Sapien: Human Powered Axie Infinity For AI - TOKEN2049 Dubai 2024

thank you again I um I'm Trevor and just a quick background on myself so um I I've been in web 3 for for a long time I'm from Toronto originally that's where the ethereum kind of movement came from have my friend sunny here as well we grew up reading the vitalics white paper early on just being at the right place at the right time um I was also involved in the early kind of Bitcoin scene before ethereum as well and these were some of the world's first like altcoins I guess ever that were based on bitcoin like made safe coin and colored coins so that's going way back to like the early 2010s I guess and um eventually in 2017 I launched my own project called poly maath and it was an early tokenization platform um kind of like the real world asset narrative that is today um is is what uh polymath was and I was really proud of it we raised uh $60 million about we uh got up to a billion dollar token Network Supply uh circulating Supply evaluation we had some really good investors and and now we have a layer one that Charles hoskinson and I kind of helped architect that's out in the wild but um after we decentralized it we we had a full layer one that was completely decentralized now out of Switzerland it kind of freed me up to do something new and that was AI I uh about 18 months ago I got really excited about AI I have a friend from waterl University it's kind of a technical school where I'm from that's where vitalic went to University briefly before the teal Fellowship stole him and um and and I wanted to learn about Ai and I kind of hired him to be my tutor and he kind of broke it down to me in a really simple way which was number one data and number two compute that's kind of what AI is it's like two prime factors it's data compute and then some algorithms in the middle and so a lot of people today are focused on the compute side so this is like the gpus you know having the inference running to to make these models magical but I was really more excited about the data side because data to me is more scalable it's more native to what my background is and so I kind of started looking at the how data is currently structured for AI models and data sets and there's really just a few steps it's kind of like a supply chain it starts out as raw data so like crude oil in the ground is like raw materials raw data so it starts out like that and then you have to refine that data you have to structure it and that's step two and then step three is taking that and running a running a training model on that data and then finally there's the fine-tuning step as well which is kind of like making it better by adjusting dials to make it perform better so that's how AI data is very simple and then I started looking at to how it's being done in some of the biggest companies in the world that do data labeling as it's known as uh these are companies like uh scale AI these are companies like open AI who Outsource their data labeling to companies and as it turns out uh there's one thing that surprised me about how it's done and that is that human beings are extremely important in this process um in fact The More Humans you have fine-tuning data the better the model is going to perform and open aai was the first company that actually used industrial scale humans this is known as rhf reinforcement learning with human feedback they were the first company to ever do that at industrial scale so instead of having like dozens of humans fine-tune the data they millions of humans fine-tuning the data and that was kind of the main reason that open AI broke out early on is because they were the first ones to do this at scale and then suddenly their models performed extremely well that's what it looks like this is data labeling it's image classification it's bounding box there's kind of different modalities so there's text based data labeling there's video based data labeling there's audio based there's liar which is like self-driving data and this is kind of examples of what humans do to fine-tune data sets so that's what it is but what happened is because open AI kind of won the early Market here it led to this explosion of demand for human feedback in these models and so open AI they're they're spending I think a billion dollars this year on human finetuning last year it was like 10 million a month so they're like 10x in their budgets on humans labeling data because that's how you get performant models and everybody's fighting each other to have the best performing models and so the best way to get better performance is to put dump More Humans into your training data and get more and more human training data um as well and it's a big Market even today companies like scale AI they're like I think they raise at 13 billion and that's just one data labeling company um so the Market's growing like crazy um but here's a problem implementing like like let's say you're not big Tech and you want your models to perform well what do you do how are you going to like Implement millions of humans in your data set you really can't afford it it's very messy it's very expensive you have to like have these facilities in Southeast Asia that's literally how uh modern data labeling is done it's like these huge call centers with like 30,000 male Filipinos in one area in Southeast Asia labeling data all day so that's literally the the the modern AI magic that we enjoy today is powered by a lot of Southeast Southeast Asians doing data labeling all day and to me it just seemed a little outdated it's not even it's not even like Web Two it's like web one like call centers with like top down bureaucracies I have some more information too they this is what a data labeler looks like this is an avatar Maria they make1 or2 dollar an hour they have to do a lot of crazy boot camps and training sometimes when they get paid they don't get paid because they have to uh cash out in Fiat and there's remittance problems and pay PayPal could get disintermediated and they only make a couple bucks an hour so even it's not quite a living wage in a lot of areas and it's kind of boring work too so data labeling it looks like Excel spreadsheets like data input it's not really fun or anything like that and so hopefully this demo Works cool and so sapen we decided to build kind of from from first principles a new a new way to do data labeling so I'll just explain what's going on here if you're a company that wants human feedback which is like any model that you want to perform well whether it's an Enterprise model mod an open source Model A researcher model all you do is kind of upload your raw data to the network so it's like drag and drop you don't have to call anyone you don't have to um get lawyers or Business Development people involved you pay and then you're done and then on the Fulfillment side on the supply side we built a a web 3 game like a fun interface think of it like axi Infinity for AI where people get paid to label data but they can have fun and we've we've tried to integrate a lot of web 3 native Fe features like airdrops so if you're playing accurately we have algorithms that detect uh to a high degree of certainty how accurately you're playing you can win airdrops you can win uh nfts you can cash out in crypto as well so you can get paid immediately and so we we're thinking this is kind of the new iteration of play to earn we call it train to earn because it's a little different than just playing a game for fun you're actually getting paid from real customers paying real money and that's kind of it it's kind of like uber for data labeling uh Sapient so we're trying to make it so like anyone can get their data labeled by humans number one and number two anyone can label data from their phone they don't have to like sign up um for like a job it's like an Uber driver you just like all you need is a car and you can be a taxi driver but in this case all you need is an Android phone and you can be a data labeler and because like one of the good things about Sapient is because we don't have to pay all the infrastructure that these other centralized facilities have to pay it's digital first so all the savings that we have by not having to pay for these facilities and boot camp instructors and management we can send all that money back to the labelers so the unit economics are great they don't have to make1 or $2 an hour they can make three to $6 an hour and now uh it's substantially more kind of US dollar based pay for someone who's in the Philippines or in Southeast Asia we have one more uh demo here we have different skins for different verticals so we have like the casino skin for more dgen labelers and then we have this is the Starcraft type theme as well there we go and so a lot of data labeling now is getting verticalized this is a real customer by the way they're a retail customer that needs human to label that image to help them do fashion predictions there's a lot of predictiveness with these models so there's like we have a customer who's a Enterprise defi company you guys would all recognize and we're labeling ether scan data to help build a pricing prediction model for the price of crypto uh but for this case it's it's a retail customer we're trying to make it fun we're trying to make it engaging one of our uh investors is Anam Moka and ygg these are like web three gaming Focus funds and our plan is to like have this interface kind of embedded into their homepage so now any one of their Guild members or labelers can just seamlessly go on the ygg homepage and start earning money immediately at the end of the day and you saw there we have like avatars and gilds and you can like challenge other gills and do like label aons instead of hackathons and things like this yeah so now Maria our our customer Avatar before like I said she was like had to drive to this crammed office she didn't get to work from home she had to do all these kind of like frustrating things and now she can make a living wage she can work from home she can play the Sapien game at a bus stop instead of playing Candy Crush she can just earn some money and then she can also have fun too like you can see we have like airdrops we can kind of like do a bit of dgen gaming as well I kind of joke about this because back in uh the early ethereum days people talked about the ethereum scaling trilemma uh we have our own data labeling trilemma which is how can we pay the customer less the customer pays less the labeler the you know user earns more all while accuracy being high at the same time so this is kind of the the hypothesis we started with that we can make it better for everybody like a win-win-win and we can use kind of web 3 incentives like staking to ensure decentralized uh data quality because that's the big challenge right like how do you ensure quality in a distributed model and that's something that we're experimenting with uh one other quick point I like a minute left is um our distribution strategy so um when I was at polymath we created a standard called ER c400 which is now the default standard for tokenization on ethereum and it turns out data labeling also needs a standard because all of these customer pipelines with raw data that needs labeled don't speak with each other they're not interoperable and they um they have proprietary Tech Stacks so we're actually creating a standard called ai2 um in fact the guy who created erc20 is a great guy and he collaborated on us helping do ERC 1400 but for sapen we want to build our own uh AI data standard called AI 20 so now anybody can hook into our Sapien front end um in kind of a seamless way uh quickly on Enterprise clients um so we have uh seven different verticals and 13 pain customers these are like large Enterprises we have a large language model you would recognize as a pen customer we have um a web 3 customer this is John he's building a model to detect if students are cheating using chat GPT so schools are paying him to detect that and he used us to help label the data uh which is pretty cool and um oh we use we have one of the biggest education companies in Asia and we're labeling their math GPT so they're trying to teach kids math homework and algebra and calculus through autonomous agents and they need um they need data labeling to make that work yeah so that's a quick overview of uh kind of our our traction so far real quick and some of our investors and finally the team and how to connect with us so thanks for your time everybody thanks to the organizers as well um I appreciate it