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

how's it going I'm Trevor if you can't read and yeah I'm a web3 guy I'm from Toronto I was part of the early ethereum days uh when vitalic was handing out the white paper back in 2014 or something and um my last project that I I uh launched in web 3 that was like a token based project was called polymath it was an early um security token what's now known as real world asset token uh we evolved that into an L1 called poly mesh Charles hoskinson and I built that back in 20121 and now as a category leader but now I label data for a living that's my job I am a lowly data minor person and the reason let me just back up AI is really three things it's compute algorithms and data I think compute is not a sustainable competitive Advantage because the next guy can just buy more chips I don't think algorithms are competitive Advantage because right now they're all at parody anyways um but it's data data labeling and um proprietary slf Frontier data that I think is a competitive Advantage so humans are kind of paradoxically critical for AI um and and you you hear sometimes you know machines and synthetic data is going to replace humans it's actually the opposite if you look at um any category right now ai data is the fastest growing by Revenue I'll just back up for a second a lot of folks are saying that the reason AI hasn't reached AGI yet is because a lot of the data we have on the Internet isn't um the result isn't how humans reason it's the result of human reasoning so we've kind of hit a plateau because when you write something on the Internet you don't write how you got to that answer you just write the answer if that makes sense um so a lot of people think you know data is the new oil it's a new competitive Advantage it's going to be just as geopolitically strategic as chips are today if the US government has a lot of proprietary data it's not going to want that data to get out in the same way it doesn't want Nvidia chips to get out either um and and just follow the money you know on human data labeling now the big llms pay about a billion per model per year on human data labeling and that's up like 10x from the previous year which is up 10x from the year before that so it's growing at like quadruple Moors law right now the the demand for data is insatiable and it's only growing and accelerating uh especially human feedback or rhf data all right oops so what is data labeling it's a lot of things it's basically taking raw data kind of like crude oil in the ground and refining it before you put it in your car as gasoline and so this is kind of what it looks like you take these bounding box modalities that's like an image and you say this is a stop sign this is a baby or it's like a tech based thing where it's like hey Mr human do you think this is funny or not and you're training the the model how to be funny but have you ever wondered until right now how uh data is labeled I'm I'm saying it's this huge Market there's millions of data labelers today around the world you don't hear about them and I'll tell you in a minute why um but this is how it's done it's done this is a real picture it's from centralized labeling facilities in the Philippines mostly and other places in Southeast Asia um we're just thousands and tens of thousands and hundreds of thousands of people in a physical location label data you do not hear about this because big Tech doesn't want you to know how the sausage is made and I'll tell you how the sausage is made it's made in these huge facilities um it's not ideal working conditions they make like one or two dollar per hour um but I don't really want to focus on the negative either and there are a lot of negatives I'll say one more actually the the value is clearly not acing to the to the humans that's what we're what people the people that are doing the labeling are the unsung heroes we're the people that helped open ai's model work because open AI wasn't the first model it was the first Model that worked because they put a lot of humans in their data sets uh so anyways these humans don't really get the credit that that they deserve and that's why we built Sapien sapen is a decentralized version of a big company like scale AI who we actually have a lot of respect for um but just quite simply sapen is a two-sided Market place for the Enterprise it's the fastest way to high quality labels on demand so every single model whether it's a huge big Tech model or it's a mid-market model needs to structure their data so we're making it it easy for even the smaller models to compete with the big models to close that performance gap which is like insanely big right now and for labelers the supply side of the marketplace you can think of that like the drivers in Uber uh sapen is just very simple a new form of gig work where you can earn money anywhere anytime from your phone and more importantly we think this could be the first time ever that anyone anywhere can earn a living wage without needing any skills without needing any assets like they don't need a car to drive an Uber they just need an Android phone and you can label data and make a living wage and that's kind of powerful because you can get paid the minimum wage AG to to survive just by being like a no skill low wage human or you could be like a highly paid specialized human and you can make a lot you can make you know let's say you're a radiologist and you want to label MRI data on the weekends you can make hundreds of dollars per hour on demand instead of playing Candy Crush at a bus stop you can like make a 100 bucks labeling with your human skill from your brain uh on your phone I want to do a quick demo this one's a little more uh active this is a real customer that paid us um this was a fashion model so it's a fashion company like an aritzia type company that's trying to um build a model to predict fashion trends that's the beauty of working in a business like sapen we get to see all kinds of fun use cases like we had an Enterprise client uh an Enterprise defi customer where we were labeling solidity data like ether scan data we've had a pancreatic cancer researcher like I said before this is an example of a fun one which is a uh a verticalized fashion company that wants us to find humans to label data to build a model that predicts fashion trends so this is we think this is the future of human work it's like using our skills and our intuition to teach machines how to do that job even better um and so what we're trying to do here is like gamify the experience that's why we have uh some of our cap tables like ygg and Anoka we wanted that gamified layer so it's not just like a boring Excel spreadsheet it's like a fun interactive way to play and earn kind of like a new version of play to earn but we actually call it train to earn or label to earn um so we think that's the future of of work and especially for data labeling um a few different strategies we really want this to be a simple interface if you guys pull out your phones now we have a live working product you can label data and like start earning rewards uh you got to go to game. Sapient .io on your browser so just game g. sapen sap pen.io doio I'll say it again at the end of the the presentation but when you see it on your phone you'll see what we mean we're trying to take this little embeddable front end and put it everywhere it'll be a telegram app it'll be a Chrome extension right now it's a responsive web app on your phone and I think more importantly for web 3 this is what we need to bring adoption to these chains we don't have a problem with block space right now we have a problem with empty block space we need like real use cases I still consider us an infrastructure company but we're also an application as well that's like bringing we hope millions of new users to web 3 um so yeah we want to go AC crossplatform and you can see our first iteration of that on your phone right now for labeling and by the way on your phone that's real customer data so you're actually helping us make money too because they paid us uh to get you guys to label the data and then we'll support um rewards for you guys to earn a percentage of that income as well we've only been around about a year we've already got uh labelers from 91 countries we won a big contract recently from uh a big labeling model you would recognize that CHS us because of our diversity of labelers another problem with this centralized labeling model is that they're all in one physical location and the people labeling in them are all of the same nationality and culture and gender in a lot of cases and I'm not talking about like a Dei argument or anything I'm talking about quality of data if you have more diverse labelers you have higher quality training data if you have homogeneous labelers you get slanted and biased results and suboptimal data so um diversity is strength in this case and we want to have the the biggest largest most diverse network of labelers in the world and quickly on the web 3 side you know nowadays I think it's important to get a lot of traction in users before launching tokens around and that's what we've done um but we have a lot of interesting web 3 use cases that I want to touch on quickly so connecting a wallet is important because that's how you get paid you can get paid on chain we would much prefer you and you should also want to get paid prefer to get paid on chain because you don't have to pay fees and remittances and delays you can just get paid instantly that's an easy one uh we have another concept called stake to play where we give users the option to stake a token while they play and while they're staking they earn more rewards but if we detect that you're kind of cheating or cibil attacking the network meaning you're trying to cheat or game the system or if you're not human or if you're a bot uh you could get slashed so this kind of if you play well and do a good job you get more Rewards or if you misbehave you get less Rewards or you even lose your rewards that's part of our kind of mechanism design or incentive design to get people to play genuinely and accurately and it's actually a big unlock because the biggest reason this doesn't exist today is because it's hard to ensure quality in a distributed model so that's why everything's centralized right now it's like Banks right before Bitcoin everything was centralized to ensure um settlement um but now you don't need to do that centrally because Bitcoin is a new protocol so we've kind of done the same thing with our incentive design uh just to go down the the stack a little bit we're kind of placing ourselves within the decentralized AI ecosystem um we're actually working on a new standard right now at my last project polymath we wrote ERC 1400 which became the default tokenization standard for ethereum and uh the the guy we co-wrote it with was Fabian Vogal Stell who created erc20 so he's kind of the father of of modern ethereum without erc20 we wouldn't be where we are today uh so we want to do the same thing for decentralized AI data where we're solving real problems because there's no standardization we want to make it so data can flow across the Dai stack so it can be collected by a project like grass they can send it to us for structuring we can send it to uh Jensen and bit tenser for training they can send it to narage at ritual for inference and he can send it back to us for fine-tuning so we're kind of seeing the need for standardization it also helps us on the business development side because every time we work with a new Enterprise customer that needs humans to lab their data we have to do custom Integrations and build custom modules and there's so many different data formats and modalities it's very inefficient but if we can create this standard and get everybody on the same playing field we can um update and onboard customers a lot faster um yeah so we we we've been around for about a year we've we've raised from really good investors that focus on kind of the decentralized marketplace ecosystem as as you can see um I want to talk a little bit about a couple of our customers we have one of the biggest Chinese education companies where we had to bootstrap Mandarin speaking mathematicians which is crazy I didn't even want to do that because it's a big pain in the ass to get that specific amount of specialized labelers but it was worth it because they're willing to pay us a lot more for specialized human knowledge that's what I'm excited about everyone in this room has knowledge that they can monetize to the machines so the machines need us more than we need them in a lot of ways um and you can see some of our other big Enterprise customers as well uh the last little slide here is what I said earlier you know to get to AGI we need human knowledge and we need human reasoning AI is not good at reasoning yet primarily because data on the open web is the result of human reasoning it's not how humans reason and if you look under the iceberg the open web is just the tip of the iceberg there's Enterprise data like 100x bigger than open web data JP Morgan has 120 pedabytes of private data for example and Chachi CH Chachi pt4 was trained on one petabyte so it's a whole like under the iceberg and then even bigger underneath that iceberg is human knowledge data that's Frontier data that's knowledge that we have in our head that needs to be trained um so yeah we're around check us out everyone with the sapan T-shirt to chat we just uh brought on Rowan Stone he's a co-creator of Bas like literally at coinbase he co-created base so he joined us as head of web 3 and um if you have any questions or anything we'd love to chat thanks for your time