DAPPCON 2024: Towards Decentralised Machine Learning - Jiahao Sun

Transcript

hey hey thanks thanks everyone um yeah um yeah my name is Jan a founder and CEO of flock um I will um yeah so sorry just after a break so I think people want to settle down a bit so I'll start with some questions like how many of you guys here have have used llms like raise your hand yeah everyone did right now like a routine for everyone right even for a simple email we will just run through a grammar check now everyone become native native speaker um yeah so how many of you guys have uh trained llm on your own or even like find tune cool yeah not many people and how many of you have actually trained on

another people's GPU nope or how many of you have Trend or fine tuned llm on other people's St right yeah sounds even even strange right cuz but this is actually what we're trying to achieve a decentralized AI um landscape or actually an endtoend mechanism for people to actually can use any infrastructure any data they can get from the community while we still preserve the Privacy to make sure like nothing's leaked but we can still uh build on top of other let's say that other data sources or other computer infrastructures to make this so that's what what we are talking about as decentralized AI in practice meaning that we have it out now so the website

just trend. flock. iio where actually you can do it right now for today um I'll just run through a bit um how how it uh how it looks like so these are some of our uh uh collaborators um the the whole General web stre AI landscape like many players are now in this in this whole uh Ai and crypto intersection right cuz yeah it's a Hot Topic it's like so popular nowadays we have people who trying to tackle the compute like friend Jensen from from UK as well and also um for decentralized training everyone I think everyone know benser and well some of you maybe know flock already um and also people trying to do the inferencing like rro eare or and

also for applications where David was just here for all us right and morphoses and some of our other brands so yeah this is a big landscape basically that's the whole spectrum of the of the let's say what we call the the the production line right from compute to where you're going to do training to how you're going to influence your data and to how you wrap it into a product where everyone else can can use it so that's that's the whole it um why why actually we need this why do we need uh decentralized training right because centralized model are too centralized of course so if you are going to train a virtual assistant that require you to give all of your data to a third party that'll be that'll

be quite scary right which I think yesterday Microsoft uh launched this this new feature on copilot where if you're using Windows machines right now they look into all your details and they do they can they can help you to do suggestions on all the moves even when you are playing a game you ask it how do I do attack next he'll tell you okay press a or press W so sounds scary right but it's like the the the controls Central so yes they can claim that they don't transfer your data back but there's no way to prove they didn't right there's always this concern about this first of all secondly uh of course there's poor compensation for data um we all know that once gbt become a hype suddenly um

Reddit choose to uh limit their API calls now right and also Twitter as well everyone knows you have to pay 5,000 per month to actually get a decent API call so um yeah just because know gbt in the early days they just take advantage of all this free data they put everything together and then eventually okay El mask figured out oh you are taking advantage of my data now I will I'm going to build my own Gro so um also the risk of bias models which I guess we heard in the news about Gemini Scandal um yeah and also you know s p had an apologize as well because those models are centralized and it's evaluated by a group of centralized the people and they

might have their own um polar of their views right which are actually delivering a model that's aligned with their own values but it's not the same for every corner of the world so um yeah so uh you you probably heard like many of the people on Twitter like recently there's a hype that everyone just post that's me that's me Po like I resigned from orania to build something else right something decentralized so yeah uh yeah so that's basically the the the the um the voices from the community that they want something that they can control they can that's controlled by their whole Community though that's what flock trying to bring um the the way

that we do it is we have well we basically so so flock itself is a p It's a combination of federa learning and blockchain flock so yeah if you're in London I do also do um I also do standup comedian myself that's why I like pounds um yeah so what we do is there's a traceable contribution of course because all the transactions of the training and the model contributions on chain that you can see okay if there's only like a 100 transactions to that model finally being fine-tuned right and I contribute three of them so of course that's 3% of my contribution that's easily to trace and traceable and also there's onch incentivization where well that's the biggest thing of of of crypto right so before if we are contributing

to open source project that it's not clear and how you're going to repay from it get get some pay from it um maybe just uh show off your uh um um status on a GitHub hugging face or Caro and hopefully you can get a job in Google to get higher paid but that that's the only exit for all this open source Engineers before right but now there's some direct Financial impacts and also uh all this gives into our specialized models for the diverse community so um it's not trended by us we can't see any of their data or any of the models it's all is governed and evaluated by each different community so they have their own model they can they can build it um yeah so how the whole framework works like so um

there's four different roles currently well actually three different roles for the TR trainers so there's task creators those are our collaborators who are trying to build up a business case to post up a task on flock then um then uh they can they can they can Define their uh desired models and there are training nodes those are the actually the builders on the in the community who are who who are trying to build models to solve those problems and also validators so that's a important structure or important role in our whole flux system where um they help the whole system to govern um whether the training results

from any training algorithms are accurate or are they faking the result or are they cheating the result and also eventually the end users who actually the users to going to be using those Bots or those AI agents or those algorithms right either it's a it's a chat bot that you're going to chat with or it's like a consultancy medical consultancy application which is actually a real case in our in our uh in our platform right now to be resed to another consultancy company so there's a real value coming in and from from different uh industry collaborators both web 2 and web three um so this is how uh the structure how it works basically the data stays logo and I say so the reason

we use federal learning so feder federative learning is not uh uh it's not something like very very hard like hard to understand or very old it's actually debut maybe 2017 by Google uh everyone nowadays who are using Google phones actually you are using federat learning already so Google is using it to predic typing Habits Like what's my next word based on the words I'm typing down already right so but there's Al always a risk where um yeah it's controlled by a central entity and uh they might if they want they might get your Ro data back so there's no governance over it that's also why if it's so secure why not Apple join this

goog Go's feder learning training so we have a bigger user bases and a bigger databases to train the model right nah they're going to they're going to they are not going to do that that's why we want to blockchain to join this Federation learning training so everyone or every trainer in the community they form this consensus governance so no one in the middle actually see anything so it's all has to go through the whole community base um yeah we published papers on this like our first paper on new ribs 2022 and then several following up papers on uh i e transaction Ai and TBD oh TBD is the title transaction on Big Data it's not TVD right so um that's the base layer so

your data Stace local the only thing changes is the modal gradients basically the changes of the gradients that can pass around and you don't have to have your data send over to any third party it's always local and uh the middle layer is what we call the AI layer AI layer are the ones where we actually collaborate all these roles together the ones who create the tasks the ones who are training those models and the ones who are serving those models and on this whole a layer we will work with our collaborators either those who provide compute or those who provide uh distributed data storage and then eventually the blockchain layer um the uh that's for the whole um security of

the of the whole system so the staking the consensus and an incentive to make sure that okay it's it's it's where we gared and also uh we have the SC staking mechanism not for the financial way but for the uh for for the governance ways that you have to put down something U or some of of the F FML let's say the token name and then to join the training so if you are not playing good you are going to be slashed but if you are always good you'll get those incentives so for good players let's say good as a simple term you'll always get resources to join next tasks or future the new training uh training opportunities or if you are playing bad you'll be slash out then you will run out of your gas basically so you can't join any more training anymore

that's that's basically how it works um um yeah so there there yeah so so there's supposed to be a live demo but we don't have that facility today here so I just got some screenshots so the platform is out already and the um we've already having a great community support and there's going to be a big announcement also with our uh uh biggest node operation uh collaborator which is also the world's first I'm currently still undering Baro it will release 6 p.m. today actually pt1 a.m.

so yeah so yeah you will know the name a bit later and uh yeah um so how the platform works as or we have those tasks basically those tasks will be generated uh it's either you know as I mentioned right either a a bot or an large language model related uh task or even not large large language model related task which we did before with our uh Partners In Medical area where they do glucose monitoring but using flock as a security way so they don't have to send their patient data anywhere else we can use this to do even cross jurisdiction um training because all data stays local so um those tasks will be will be published here and then participants

will join the training they can either choose to become a validator means that they just they don't need to do a lot of work they just need to put down some uh put on some FML tokens and leave it there um waiting for other people's results and then to validate to check the result whether it's getting better for the metrix right that's an easy easy job and also you can choose to be the training notes by which um training notes means those who can join the training and um um and provide their their models either uh change it from a base model or from from some models they created themselves basically to to contribute to this specific task and then those models will compete each other maybe 10 of them 100 of them will

compete whoever get the best will will receive the uh highest incentives on the whole system so yeah there's there's no Financial efforts in terms of the staking things so I know when I talk to people like you have to stake to join the training they feel like oh are are we going to pay in ease or something n it's like there's no Financial staking that you have to do it's uh it's it's just for the whole systems testing and governance it's just for this system itself um yeah so um what's the next steps so basically we currently have our single note trending meaning that you can run it on your own single notes and in the future there will be more uh features rolling out for example the

federat learning close beta meaning that we can have those federat learning agents downloaded or installed on your on your client basis so they can get secure that their data always stays local and then you can use those by use I mean uh States local but you can use those uh data for training purposes and um we will also have our our white paper live we actually have it already but then my team told me like it's too academical let's make it easier to understand so we are just like decomposing now it now to make sure it reads smoothly and uh yeah and moving forward our task creation will be open to almost everyone so um it's it's going to be decentralized so if you have a

task yourself you want to build a model for yourself you can just put up a task leave it on flock and then wait for the community to contribute together and uh yeah that's pretty much it so how to get involved um yeah W list is here um you can well you can scan scan the QR code you can actually um um yeah check all our socials and uh also um yeah join the join the white list you don't um we are currently just running out our second phase so we might open up the wireless very soon but of course early adopters will have higher incentives if you're there there's less competition let's say in the whole ecosystem right now right

so yeah that's us and also for the some of the models which are already live you can check it on beta. flux.io so with our collaborators like those um for the those gbts for those projects and those uh medical use cases that are online here and uh yeah I think that's pretty much it and that's our Motel we are scaling the decentralized AI and which is created by communities and for communities right any questions thank you so much J uh we have a few minutes for questions if not we'll we'll wrap it up there and we'll get started in a few minutes thank you so much J yeah and

also those uh academic papers they are all on this link as well so if you want to check it out um yeah thank you