{"id":"ba2ixv44eb8","title":"NEAR AI Office Hours #17 with Cerbo AI - Permissionless AI Brain Model","description":"Illia & Alex discuss the latest AI research topics and unpack some NEAR.AI focus areas.\r\n\r\nGuests: Cerbo AI https://www.cerboai.com/main\r\nhttps://x.com/CerboAI\r\n\r\nPioneering the first Proof of Federated Learning and Edge AI Computation. Decentralizing AGI with personalized AI edge models and a powerful computing network.\r\n\r\nNEAR AI: https://near.ai/\r\nIllia Polosukhin https://x.com/ilblackdragon\r\nAlex Skidanov https://x.com/AlexSkidanov\n\nBlockchain for Privacy and Collaboration\n\nIn this episode, the founders of Cerbo AI discuss their innovative approach to combining AI and blockchain to solve data privacy and scalability challenges. They explain their federated learning model that preserves user data privacy while enabling collaborative AI development. Key topics include practical use cases in healthcare and finance, as well as the economic incentives that drive their ecosystem.\n\nTakeaways:\n\n• Federated Learning for Privacy: Cerbo AI leverages a federated learning model, ensuring that user data never leaves the device, with only model updates being shared and validated on the blockchain.\n\n• Decentralized AI Incentives: A dynamic, multi-role governance model incentivizes contributors based on the quality and impact of their model updates, ensuring fair rewards distribution.\n\n• Use Cases Beyond the Norm: Applications in healthcare for secure data sharing and in finance for fraud detection and credit scoring are prime examples of where federated learning can make a significant impact.\n\n• Challenges and Solutions: Addressing challenges such as data validity and preventing malicious contributions is crucial, and Cerbo AI’s unique scoring system aims to tackle these issues effectively.\n\n• User-Controlled Data Contribution: Users can choose how frequently they update and contribute to the model, providing flexibility and control over their data's use in AI training.\n\nTimestamps\n[0:07] Introduction to Cerbo AI: Hosts introduce the guests and discuss the intersection of AI and blockchain.\n[1:39] Founding Story and Team Background: The founders share their journey and motivation for creating Cerbo AI.\n[3:10] The Core Problem: Data Privacy in AI: An exploration of how data privacy limits AI development and how blockchain can help.\n[5:24] How Cerbo AI Works: Explanation of Cerbo AI’s federated learning model and how it differs from traditional methods.\n[8:34] Early Use Cases and Applications: Discussion on potential applications in healthcare and finance.\n[12:02] Technical Deep Dive: Details on model updates, parameter sizes, and the technical challenges of federated learning.\n[18:08] Privacy and Security Concerns: How Cerbo AI addresses security concerns in federated learning.\n[24:07] Incentive Mechanism Explained: Breakdown of the economic model and incentive mechanisms for contributors.\n[30:05] Challenges of Data Validation: Handling invalid data submissions and ensuring model integrity.\n[36:00] Real-World Implementations: Examples of current projects and user feedback.\n[45:10] Future Plans and Improvements: Plans for enhancing the model and potential future applications.\n[1:00:34] Conclusion and Next Steps: Final thoughts and a look ahead at what’s next for Cerbo AI.\n\nJoin NEAR's community:\nWebsite: https://near.org/\nNEAR AI: https://near.ai/\nTwitter:   / nearprotocol  \nBlog: https://near.org/blog/\nMedium:   / nearprotocol  \nGitHub: https://github.com/near \nDev Docs: https://docs.near.org/\n#NEAR #nearprotocol #NEARAI","channel_name":"NEAR Protocol","published_at":"2024-09-10T00:00:00.000Z","thumbnail_url":"https://img.youtube.com/vi/ba2ixv44eb8/hqdefault.jpg","transcript_text":"explore hold the future hand in hand in this Tech driven wonderland hello hello uh Hey again everyone thanks for joining the office hours we're here with Sero I I let them introduce themselves first uh hi you want to go of course yeah glad to be here thank you for the great invitation Elia and Alex finally nice meeting you too um so I'm haha I'm the co-founder of subo AI I came from a background with Masters um in robotics at upen and obviously worked in blockchain for almost two years right now and really excited to see you know how Ai and web spray can intertwine with each other as and provides true user case in the future yeah glad to be here thank you hello awesome yeah it's my turn I'm ccs and I'm in charge of the suppose product um I lead everything from the open source project to the protocols to the like the what the D app that we're integrating on top of the pfl and happy to be here and learn more awesome well uh thanks for joining it's uh always amazing to dive in with Founders uh what you guys are working so maybe tell us more about kind of how you're approaching uh the space what you see and maybe more as well about your founding team of course yeah so me and CCE were were um family friends um know each other since quite quite well and uh yeah so I always like worked in blockchain uh ever since I graduate from master and don't really um you know knowing exactly how artificial intelligence could intertwine with blockchain whereas we started to you know brainstorming about what what is the most um beneficial way for AI model developers and AI application developers in the future whereas we started to notice like data shortage is one of the things this pretty much like the threshold um that um limited not limited but you know in some ways limiting the AI uh Industries development and in another way if you know people we all know human beings and if people don't get enough of or transparent or fairly incentive to you know convince them to you know supports the AI industry so so why you know why are they contributing their private reserve data um and that's sort of the major um questions that we started to see uh I think last year around like around March and that was the first time um whereas I think the scaling Lev was such a big terms that um was being constantly mentioned not only in uh AI but also in uh blockchain uh communities as well so we think hey if we can do anything that leverage the benefits or the advantage that blockchain could bring to AI industry or Community or even to see users that will be the best because ultimately people are the one who is you know contributing and conducting the datas right so um that's the major topic and I can give a deeper understandings about exactly how Sero as leverage that and how do we want to do with it uh but s CC I feel like you know maybe you can give a better understanding from your perspective because you came from a Google um background and also you have been doing AI uh applications research for quite a while as well right right so the problem fundamental problem that we're trying to solve is that we see okay there's a lot of projects on top of the ecosystem that they worry about the privacy and um and what we're trying to do is just a layer that helps every project whether you want to build an advanced application on top of that or you want to build some use cases or you want to build agent or even you want to B to build some models um and because like there's a lot of the open source models in the community now so then like what scenar like whether the scenario is application or use case or models we would love love to um provide the protocol to them so that they don't really have to worry about any data privacy reserved issue um because the way that we're using is not really just relates to data it only relates to parameters so then like in that way nobody has to really worry about okay if we're going to save the data into some you know some decentralized storage and then like to release it afterwards or we have to find different different very expensive um what data consumption like Solutions out there what we do is pretty simple we just compress the data into the parameters so that everybody what they've seen on chain is only the parameters not the data at all so then is we think it's the best way for so far all the project that we've seen to protect the privacy and when it especially when it comes to the Apper layer to see you cases and to be model issuers so give a summarizations of what exactly is subi subi is more like a you know Cutting Edge Ai and blockchain focused protocol whereas we build on the visions of more of a decentralized and privacy preserving machine learning open source Network or protocol um I would say we have a small vision and it could be a big Vision as well um to democratize Ai and also ensure that the user data which is you know the um in our cases the model parameter weights remains private while unlocking the values from decentralized AI computations so that's Sero Ai and so my understanding is like underlying you're pretty much using a Federated learning model right where every kind of users's device is uh updating the model on their own data so indeed the data doesn't go anywhere never leaves consumer devices but instead the updated model is being checkpointed back to blockchain uh kind of as updates being shared and so like that's kind of the general model you're going after where the train happens always on user's device and then the kind of common model is uh then available for everyone I would say that's a good uh understanding perspective um in our words that we would like to say hey this is like more of a a pfl protocol whereas we leverages and a Distributive blockchainbased middle layer infrastructure to facilitate secured and decentralized machine Learning System how to understand this better so through the pfl model or through the pfl protocol note participants in you know using a AI applications and training AI models without requ requiring a centralized data storage to process so yes your understanding for the most part is correct and blockchain serves as more of a a ledger to validate and verify models update and you know give a fairly amount of the um uh incentive back to the communities based on their contributions and also ensuring everything can be you know fully transparent um and fairly rewarded U distributions for all the participants and to also for us to able to ensure this part of the distributions or uh of the uh rewards there has to be you know a really um I would say a multi Rose participations engaged um from the blockchain perspective whereas the pfl creates a more of a a secured environment where multi- Staker stakeholders they collaborate and to train and monetize AI models and AI applications with model parameters weights um in a decentralized manner so I can go deep about essentially like how do we achieve and what exactly we have designed um for for us to make sure that's the entire you know the dynamic uh economic models um and um to you know guides the um guides the U uh incentive mechanisms within the ecosystem so so before before you go into the incentives what what are the uh I'm curious to know what are the early use cases you see like like what sort of models because you know if you're training llama on on you know like on a large Corpus of text that that doesn't have to be Federated right so like I assume Federated learning is more interesting in the situation where uh we want to train like I guess Federated learning came up in the context of training the swipe keyboard right right so like every everybody is typing on the keyboard uh there's a lot of data but obviously nobody wants what they type on the keyboard to be known to Google or especially to other people right so if the model regurgitates uh you know your social security number on another keyboard that's that's substandard right so I'm curious to know what the the use cases you envision or maybe you're already going after yeah that's a that's a really good question um I would answer it from more of e ecosystem perspective whereas CC you can jump in and say Hey you know these are the AI models that we already supplied but to be honest I would say the servo AI target audience is not primarily focused on AI model issuers we're primarily focus on AI powered application developers groups whereas you know some of the killer user case that we already seen that could include private preserving AI applications such as you know Health carees obviously Federated learnings for collaborative AI models training across hospitals without sharing sensitive patients data so that's one and secondly you know Finance um AI or AI for fraud detections and credit scoring Without Really exposing your or anyone's personal financial data we all don't want to expose any of that uh informations out to the public now sometimes even not to the bank and that's why we're having you know blockchains as more of the assets transportations um layers right so in the uh in the other layers or in the other perspective the iot's um projects were um uh Edge based AI for smarts devices which could leverage FL to also improve like device local device intelligence while also keeping the data locally so that's from my perspective CeCe feel free to jump in more yeah yeah um it's it's my answer is pretty much the same we don't really at this stage uh Power like for example 100 different models or even that you can uh open the new models we actually just um so we'll be able to connect with the majority of the fundamental open source models there like Lama opaka and the majority of the models that I think 90% of the application would bu based on that and then um what we uh what we can do is that we can compress the model size or like we can actually compress the data size into 2 KB which um a lot of the isers can or a lot of developers or a lot of the application isers can build based on that no matter what application it is one of the things that we're doing right now is to see what applications they might need a like the 2kb is still like too you know like too big for them or too small for them or just for that use case they need a much more reliable model they need a much more intelligent model so they probably cannot accept the 2 KB's performance they probably want 10 KB or 15 KB so that is the basically all the circumstance circumstances or the case case studies that we want to implement in the open source Next Step just yeah just to double check when you say 2kb do you mean that's uh that's update updates on top of a base model is that what you mean yeah two yeah it means the parameters so the parameters weight is only 2 KB but this is a like low rank uh update right adaptation on top of bigger model yes it Alex I no no that that was more or less my own question I I just I just tried I first tried to understand uh where the number is coming from so it's interesting that you mention uh that they they all measured in in KBS like I never heard of Laura updates of that size so that's that's a novel concept for me but but I'm also curious um so my understanding of Federated learning is relatively limited right so I I was sort of into it back when it came out which was I think seven years ago now uh and uh and so so the way I understand it is that in practice it's been shown to be sufficiently uh my so my understanding right like going back to swipe right my understanding is that when I use Google keyboard uh it does use federative learning right so everything I type does go to Google servers and it's strend and Google TR the weight updates go to the Google servers right the weight updates yeah so it updates locally and then the r r diff is sent and then they all get segregated right so the um but but but I read that paper right back in the day in the Federated learning and the security analysis back then I I assume many things Chang in seven years right but security analysis back then was very handwavy they they they put a very uh hand wavy security model under which it was very secure uh and what I'm interested in is that you mentioned several use cases such as healthare uh or uh like Banks right so so they they have uh security requirements where you cannot just publish a W paper on on archive right so so like I'm curious if it is actually something viable today to use Federated learning for healthare like like if the healthare you know companies would actually be willing to use to share you know weight updates based on the patient data like is it the actual reality today does make sense yeah that makes sense um but to be honest like from my perspective um we have started to uh interacting with some some of the healthcare startup um whereas the majority well uh the developers that we're we're focusing on right now is most of the U Silicon Valley based startup who already started to implement some of the a models into their services right so this is like our target audience and our target audience one of them uh is actually in conversations with uh San Francisco's hospitals whereas they are supplying those hospitals to build a I wouldn't say a AI agent but more like a really comprehensive like more of a supply chain uh uh optimiz optimizations uh systems whereas they could synchronize and track about different like hospitals um and see like what are the for example like a hospitals has the majorities of patients for for one type of disease or for several type of disease so they want to optimize you know how quick and how fast and how accurate they can fulfill the uh the the inventory Ries for most of their medicines or drugs so that's one of the developer groups that we started to engage with and um their their their concern is really straightforward those hospitals they don't want to share their data with any of the third parties including other hospitals that are collaborating with them they could collaborate in the research um and you know case- by casee patient withs but they don't share informations in that levels and one of the things that we uh we started to notice that um you know privacy it's is not only a big con concerns for them but their inventory um tractions is also a big concerns in terms of like despite of the individual informations that could be possibly leaking out I don't know I'm not exactly understanding you know the the concerns behind this reasons but I understand like from different perspective that anything that related with healthc Care status um and even from you know a commercialized the perspective um they wouldn't want to share with any Outsiders including um I would say Googles because they obviously are not using Shopify or are not using other like um inventory optimization sets at the moments whereas they Pro uh they prefer to build their own controlling systems from all you know that links all of the Departments and Laboratories within their within their uh individual hospital um system so that's what we see um currently but again we're we're not directly engaging with the healthcare provider more engaging with the tech uh companies that provides you know consultings or agencies um um for for those Hospital customers yeah yeah yeah because because like something I'm very interested in is whether in like since fed learning came out they they actually managed to if now there is a sufficiently rigorous proof of uh Federated learning not leaking data that uh you know Healthcare or Banks would be fine with that uh because I think that like like if we get to that level of acceptance of Federated learning that enables a lot of interesting use cases yeah um interesting that you mentioned Bank we have not yet catch up with bank but we have catch up with several uh wallets um um developers already so major things that they have concerns is that well they don't really concerns about the Privacy um because everything um the users um transportations of any sort of assets um will be recorded on chain but they do concerns about hey what is the next step for those wallet applications to to engage or to expans their product features to better engage with their communities right now we see wallet as you know um some applications that you receive and you know issue out um the transactions right in the future most of the wallet uh developers they're not thinking about the that they're they're thinking to even intertwine or um engage with some assets management product features in that can be powered by AI or they can either they most of them are thinking hey can they engage with AI uh agents to supply and provide some of the um what do you call it like um investment device within within the wallet product so this is like the next Generations of how I see you know our potential customers will be developing in if that's the case then hey how do you deal with the user data priv privacy it has nothing to do with blockchain in that way it's actually related back with you know most of the users are having the Wallets on their personal devices such as PCS and iPhones like me um but how do I how do I leverage that perspective I don't really use uh usdt and usdt usdc to pay most of my personal Supply expense I use Fiat I use you know I still live a good life uh here on Earth so I have to engage with Fiat payments mostly for my uh for my private life right well if that's the case then how can they understand this person's expense and how can they locally trained uh you know some proportions of that models to be able to power and also provide good informations and good advice within the wallet uh I think that's sort of related back with what we're seeing for our user case and um yeah and do think like Federated learning is a power of that and CC you can feel freee to speak more right before I speak more just wondering if you have any questions so far no not yet okay all right um okay so I think in terms of the application we just talk about you know like different applications like well like Hospital like well even Finance um maybe there's Transportation like a lot of the different scenarios but what we think right now what should be the most application is there's a lot of the agent companies are building some service aggregators or scenarios or when it comes to you know efficiently help customers to leverage the agent or Theos to do something so in that case it is a model issuer or like it is a Tob customer who's building an application but it's directly facing a lot of the 2C users and a lot of 2C users they actually really care about number one how fast this process can be and number two how private like like if they can actually protect their data and number three if they can save some money so our proof of Fe learning actually can be able to enable all of this three we can not only just because you have less parameters out there um for the low training for the low run training so then uh in that case scenario you can we we actually did some calculation we can save the cloud cost by 30% to 40% so it is a low cost approach for a lot of the 2B customers and two when we said that we compress the parameters into 2kb is a way for people to um to have a great user experience out there to even um enable more devices in the future for training and three um when we see privacy reserved we mean that because we don't directly interact with the data itself we only interact with the pro with the the parameters so then we don't really get the user data to upgrade the model um it's very different from the traditional F learning training or the traditional fine tuning or the ways like that so um like on that three perspective we think a lot of the agent applications a lot of the applications that's just directly facing into 2C are the GD applications no matter it is in hospital or it is in finance or it is in transportation so so you mentioned uh you mentioned it it's a cheaper for people to train uh the model so so in this case uh the idea would be that that someone is training the model on their data and and and uh the Federated learning helps them train it cheaper or they or or they still train it on the the data of of someone else what is the use case so so the person who 30% are they training their data or they train on someone else's data they're still training their data but the data is not being trained the parameter is the one that's being updated so you have less parameter and intuitively the cost is less I see so it's not just not just about Federated learning it's generally a different architecture which is somehow smaller than P water yes so maybe maybe Switching gears a little bit um as this is happening right um how is kind of quality insured how do people get rewarded for for it what is the um kind of what is the economic incentives to be kind of doing the right thing as well as disincentive for doing the wrong thing good question so we have already designed really I would say Dynamic economic model whereas we guides users to build our ecosystem build on top of the ecosystem and semi like control the quote unquote evil behaviors of users like you said for example like Elia or Alex you spend it like 10 days playing on one of the AI applications that should have deployed on top of pfl on our pfl versus um you know a soueast Asia Studio that they have like several different devices also spent a few days but doing like you know redundant movement or actions that still accumulated like 32 KB versus your 32 KB like who evaluated more um obviously there has to be a really straightforward and transparent uh scenario and mechanism to guide this but before we dive deep into that I want to emphasize like who is going to be the one who raises hey you know Alex or Elia your 72 KB um earns uh let's say like 0.1% of the entire sub token versus you know the southeast Asia studio uh could earn like 000000 z. one% of the token there has to be a multi-roll on chain governance mechanisms that implement it within the system whereas we design multi different um positions to raise this um the incentive and also issue it right so Ro a is going to be the reviewer and uploader who is going to be linking the results of horizontal comparations of model performance um and role B is going to be the challengers who is going to challenge and also reveals role A's results role C is going to be you know the challengers um admin or super CH Super admin who is going to review on top of Ro bees whereas we see well in the future we hope that you know there will be millions of Toc users um via those AI applications who is constantly engaging with the pfl protocols well if that's the case then row a roow b row C has to be or what we foresee it has to be a primate uh a a prime a paramid structures whereas rule a will be contains the largest participations populations to be able to execute and you know instantly well not instantly but you know we're trying to have them to instantly Rewards or raling the model performance whereas Rob will be more of a prestigious um or you know sely lesser uh um uh popul ation wise of The participations Who on top of Ro a and is checking on any red flags that Ro a raise it up or you know if there are any enormous uh activities that has been perceived or uh you know voted by ro and rosi will be you know a really um smaller group whereas we would like to engage with more of the AI expertise and you know obviously so as teams and you know the other teams that engages and collaborate with us to also overseeing the entire structures in case there are any like enormous um incentive uh executions on top of the protocol so this unchain distributions um systems we will generat different levels of tokens based on frequency um numbers of participants and the importance based on you know role a ro B Ro R C's distributions of different rewards and you know the next questions for us to resolve is that who is going to be the reward receiver obviously there are two um you know stakeholders going to be receiving the rewards developers and communities which means the 2C users who at least have a physical device to be able to engage the AI applications that Deploy on top of the pfl protocol whereas we um we have this metrix we uh we check on the qualities uh and the quantities and whereas the quantities is really straightforward we're quantifying based on the weight over each model updation so Alex and I 32kb versus you know Southeast Asia's 32 KB in this level might be weights the same but then deeper engaged with the qualify the uh qualifying pfl score whereas we combined uh data scores U model contribution scores and um validation accuracy scores together this pfl score is proportionate to the pretty much to the amount of newly mained rewards that will be allocated back to the participants so even though the 32kb is the same but then underneath that the qualities that it brings to the model performance updates are different so at this moment everything is still in testification Period whereas you know we have already Implement one of the universal uh um a model performance um uh formulas or evaluation formulas but in the future there will be you know uh open door for any other formulas that can be raised by the communities and Developers in the future so that's you know the entire um I would say like toomics incentive mechanism that we started to builds up and um for us to be able to better execute on top of it there has to be you know different um participations representing different um representing different um uh scope and stakeholders um to be able to engage so I find it um so I mentioned it every now and then on the office hours right but we've been running a system where people annotate data which is not quite the same as Federated learning um or or or not the same at all uh but it shares an important property that people contribute something uh that is only that can only be subjectively verified uh and they get the reward for that right because uh I don't I don't know very well how proed learning works right but but like from the first principles right because I don't reveal my data I don't think I can provide any that the data I used uh was in any way relevant to the data set right and uh so correspondingly the way I understand uh the fundamental concept of pro refigerated learning is that at the end of the day uh the only way to confirm that what I submitted makes sense is by you know testing it against some held out held out set right uh so uh these a uh uh right so so in the like in case of uh coming back to the swipe example right if I use Google keyboard I have literally no incentive to submit uh incorrect completions to Google right and if there's a 100 million people using Google keyboard uh even if I like try to Doos them with invalid submissions from one person I think they can uh prevent it from happening right in case of uh Sero there's a m massive incentive for people to submit in valid results right because it's uh because I can only get that much useful data to submit to you right but I can fabricate an infinite number of bad updates right uh and so people will be doing that right like the biggest lesson I learned from from our data notation efforts is that once you start paying people money they become extremely creative right like the IQ of people grows of two the moment they can get money for free right so it's so I think what it gets interesting is uh like like the actual concept of of proof of predated learning right that's something that is very interesting to dig into at some point to understand how like let's say let's say like I assume you need updates from you know like let's say thousands of people right so you get thousands of updates you obviously cannot do like every permutation of them to validate and each particular update is relatively small right like when I train a neural network between two consecutive ooks the loads can go up it doesn't mean that the update was useless right it's just a uh uh right so it's interesting if you guys can share like any high level uh of uh uh proof of Federated learning like what like at which level does it operate how how much interaction does it require from the community uh how how reli also who runs this Val like this evaluation right uh like the who who who evaluates the validation accuracy that's a really good question so that goes back to like the mechanism that I was describing earlier um there will be like Ro a roow b r C's right so whereas Ro a who is going to be the reviewer or the uploader who is like making comparations or or horizontal comparations of different model uh performance to answer uh your questions Alex yes each of the updates has to be um has to be verified otherwise you know like you said people get extremely creative to earn um you know to earn money or to earn incentive if they don't know the criteria even they know the criteria they still would think hey maybe I can squee something in or I I can s squiz you know the Redundant data in into the entire boad um board update and you wouldn't be able to find it you will still evaluate me um as the 32kb as the good 32kb no it's not going to be true in in our scenario the pfl score that we designed here actually check based on top of the qualities of each different updates um three things um like I mentioned before has been combined right here uh the data score um the model contribution score and the validation accuracy score so that's the high level like giving just giving you a framework of how exactly the PFO scores has been made up so those are the three perspective whereas we weighted them differently uh with a formula right now um and we issue uh the incentive based on top of the pfl score although the quantity um is really important because that's like one of the first thresholds but we treat it as more of a normal parameters within this entire pfl score formula whereas it's like a like a small things but then you have like a big formulas up up or right after that whereas this is like the true part of how much incentive as a normal to see users or any of the users you will be able to earn so that is the um that is the pfl score that we embedded within the system and I understand what you're saying hey if there are millions of users how can you execute or how can you reveal each individual updates so this is one of the things that we want to engage with the developers and also try to understand you know how the developer communities uh designed uh their applications what we are doing is that we are engaging with one groups of uh developers whereas designing a more of a a application like um a decentralized AI Ser whereas we call them Ferry um so fery is applications that's um is a proof of concept within the ecosystem um you can engage with it um I'm sure like everyone if you guys have time already check on the uh Apple product U uh launch um presentations yesterday or last night um you engage with the this applications and then you you know the AI agents behind this or the AI model behind um that applications would able to help you to execute say hey give you a Uber or give you a LIF lift or you know um buy you something on Amazon uh simultaneously by already knowing you enough um so this is like one of the applications that we're we're seeing and we're supporting them um from the pfl infrastructure U or you know the middle layer um perspective um and this application or this application developer groups they're really interesting they don't update each actions of the models um or of the user interactions um on chain what they do is that they gave the users a freedom for them to choose hey do you want to update all of the actions that you already interacting with the applications on a weekly basis or on a daily basis so actually um surprisingly the the 2C users or the individual users they have the control of how frequent they want to locally even train the models so it's not like you're interacting with the uh applications and simultaneously it has been already trained yes it gathers um that's data from a local devices perspective but it it doesn't really necessarily means it will be trained locally simultaneously so user has 100% controls at least for this AI application group whereas we started to see they gave the users the choice to when will be the best time to train your to train on top of your phone right they are right now supplying iPhone 14 and plus and so this this is the freedom whereas um it doesn't really cost for the entire protocols to constantly have to update or have to review each single actions or each single actions based on top of that what is the model parameter weight and then evaluated that it would been it wouldn't be that efficient um we would actually encourage the developers to well designed um their applications to be able to maintain the best um efficiency for the entire network or protocol uh ecosystem um and then after the users hey let's say weekly uh updates and accumulated enough of actions or interactions with the AI model powered applications locally um they trained um their his historical tracks or um um their their personal datas and then from there um uh we will be able to you know help them to train on on local device scenario and that parameters which will be recorded on chain with individuals um users notification saying hey you already update this part of the things on chain and then from there the AI reviewers or the uploaders will be you know in interacting with um that informations and verified you know what is the what is the value of it so uh overall like there will be or we foresee there might be some traffic gem um for the row a since um we're our our mission is to power millions of users there might be some gems traffic gems for RO a to uh review um because we cannot you know say hey um all the AI application developers you have to do weekly or monthly we give them the ches but we you know nicely advise them to have you know wisely designed of how often they want to update um the model parameters weight and then from there row a is going to be a really big po ation since they are you know not as much as the to see individual developers uh or to see individual users um but they they needs to be at least uh you know oneir or one quarters of that populations to be able to maintain such a big um traffic um so that's what we're seeing but at at the beginning everyone is welcoming to become the reviewer and uploader as long as you have the physical device to run on top of it uh and you know submit your results on chain so that's what we're seeing um yeah I hope that answer your questions so so the idea of uh of doing it infrequently is primarily to save the bandwidth right it's not like security consideration or no I'm assuming if you if you're evaluating every single like update on some evaluation metric as well right you kind of need to wait until there's enough updates for that to matter I think I mean there's multiple Dimensions here that I think worse you know considering because uh I mean obviously these models are highly non-interpretable at the current stage and so one of the challenges that people already seeing is so-called Sleeper Agent problem where we're able to introduce pretty much a triggering behavior that doesn't actually show up at the evaluation stage uh by pretty much tra you know creating gradient updates on specific like you know if it's 2025 you know do this and otherwise don't trigger that and so I think there's an interesting question of the evaluation set is always going to be you know just the subset of the distribution and especially anything that's out of distribution on the data is going to be really hard to evaluate so I think I I I think you you're trying to pretty much say well we're going to have smaller updates right in general which is kind of harder to influence but I think for a broader kind of decentralized training space this is probably one of the major issues in general is like how do you like if you're just getting parameter updates and that's all you don't really know if like this is really malicious update that like looks okay or this is actually uh legit update or this is just a random update that actually doesn't improve the model and somebody just spamming with you know random numbers um as well right so I think kind of a more I would say probably more studying like a full kind of research needed on like the dependency and detection of those things uh depending as well the size of that like of the for example low rank matricies right again I'm assuming the smaller the smaller low rank metrics it is the easier it is to the harder it would be to like create some kind of malicious attacks but uh it potentially will be also harder to see Improvement and validation set right because you can imagine that like let's say your validation set is how good it does on math and uh like it's just not going to get better in math if you just have two kilobytes of uh updates to parameters uh because it requires like a systematic Improvement to the capabilities of the model yeah I do agree CeCe by the way you're muting I know you're speaking no I was taking notes actually this is well said I think I mean the other kind of just thought I've had was broader um like if if you're giving this to developers are there also ways for developers to design model architectures that are trying to extract user information because I think that's that's another dimension of kind of and and I mean we also thinking about it on the agent side right as you're designing agents like if the devel developers militia they're gonna try to actually extract user data from the through this interaction uh so there's kind of another dimension of this as well um which you know you potentially can like train a model to generate model architectures that are uh in updates will actually like when you update it will encode the information of the user that you actually want yeah exactly actually I remembered another question I had so so you said that you want users to send an update once a month right so so in case of these models I I guess it's highly dependent on the use case right but it's one thing when the moduel is already well tramed and you're just updating it right how does the model how does the model boot how do you bootstrap the model do do you like uh do you reduce the update time to like once per day or or is the model initially already somewhat good and Federated learning is only used for maintenance Yes actually depends depends on the project um if the project like you can just set up any frequency you want um and we don't have like um it's not like when when it comes to iPhone you only can update in a certain period because everybody's using it but then we give the what different projects the flexibility to update it like whenever they think uh this is a great time to update it we've got enough parameters uh the model's intelligence has to be updated and and uh for the I'm also curious what is the so the update I assume the like whatever multiple kilobyte update this is not the whole model right this is a uh this is the lur right so so how big are the base models do do you have any insight into what is the because I presume they should be small enough for the for the iPhone to to uh to unroll them without draining the battery right yeah so we have like we're supporting four models at this moment and this just pretty much all the open source model size um like you know there's like 7B there's like uh were just like regular size that we support so that is a big model that we use yeah I definitely need to read up more on the on the actual approach I think 32 kilobyte update it's not L something even smaller I mean what kind what kind of yeah improvements do you see actually through this um I guess that that's probably like an interesting question around like for how many users you know sending updates like that like how like how much more how much improvement you can actually get on on what kind of problems and what kind of valuation sets the improvements in terms of the next uh well like next step what in in terms of the evaluation St that you're using like if it's you know I mean if it's like a 7B llama and you're testing it on I don't know doing math or doing uh next toing prediction for keyboard or whatever right yeah so now we're like doing some improvements on the agent we're doing our own D app and it's a service aggregator and um it basically we it's like we're calling like the whole end to end agent to well like finish for example Uber deliveries and door Dash deliveries and even Amazon ERS and this is the Improvement that we're using in the evaluation data set that we're testing on pfl what does agent do in case of uh what is the agent in this case of like let's say uh door Das deliveries so by the end of the day we do want um this this to be very PR like to be like personalized agent that can do everything for you and personalized assistant that can just one stop respond to all your requirement but now um we only have been pushed like two service aggregator directions when is the Uber when is the door Dash and when is the W like when is when is the the Amazon delivery so those are the three aggregators and we think that the data can be really valuable because it's essentially the human or like the user Behavior data and yeah and we what fundamentally what we want to achieve is uh two Direction one is that okay now we have an agent that can help people to do a lot of the well like the services or the real time applications the other is we do want to prove that the proof of Federate learning in this evaluation data set actually works by um having like by having people seeing their data compressed into a parameter and you can actually access access all the data from the wallet and see the data or parameters on chain what is the model predicting predicting in case of the doorage delivery like is it is it like the actual process of ordering stuff or is it uh at the delivery time what does the agent do like as as a user what is what is my uh what is my interaction with the agent your interaction is the agent is that so for example when you are interacting with Lui right like language user interface and you are telling the well the model that okay you want to eat something and then like it just automatically locates your like your position and it just like you know ERS whatever you want uh based on your the databases you have and based on the prediction of like what should be the taste of the food you like and you just like they just like automatically execute and um retrieval the correct app um that just retrieve all the services based on that so I guess with Uber is the same it's like trying to predict where I'm going yes Uber is the same too yeah I I remember interesting and I were doing that in 2017 obviously less the technology wasn't as sophisticated but turns out it's not that frequent at least in case of an Uber unless it's literally your next uh event in the calendar it's not that easy to predict where the Uber is going uh so yeah interest would be interesting to see if the you know the time has changed and now it's a moreable the user experience that you will be able to get right now is um as a user you can talk to we have our own open source model called Ferry you talk to our fery model and then you can start to engage our faery model is really you know like emotional and it's kind of just like a like a assistant that type of the model in in terms of the DNA and then uh when you tell the fairy that okay when it triggers about Amazon or Uber it just automatically redirects you to there and finishes service for you and you can always go back and you can be able to select if you want to you know you want to earn some tokens by contributing this data and then you can actually access the parameters that's we put on chain of this data and you you can see from the wallet you can access from the wallet you can see on chain and you can see how much token you get so so so I presume you guys are uh battle testing it locally right so like how frequently how like what percentage of of uber right you don't have to intervene and uh and it gets a trade what percentage of the oh it depends on user so like if the user wants 100% then they get more token if the like if the user wants like 50 7 then depends on them yeah yeah no no my my question is when you personally use it what percentage of uber rights that fairy orders are correct and what percentage you still have to intervene oh now we have like around 90% accuracy out there on the Ubers and there's only 10% of the edge cases that you will never know and and like in terms of uh so that's that's uh a question more of like more in terms of U user research right so hypothetical City I I don't use any Apple product right so so let's say Google Google whatever it's called Google Now Google sirity uh she can she can guess more than 90% of things I want correctly right like like if I say Okay Google and ask her for something uh uh it's a he gets of more than 90 % of the time right and yet I never use her because uh the the 5% it doesn't get it they extremely right so so like the the the sheer expectation my head that I can get into those 5% makes me you know like I would set an alarm manually even though it's going to take me 10 seconds longer right so I'm curious if this is something uh if in your experience you you're sort of crossing the threshold where people because ordering an Uber is not a like you're not saving a person an hour of work right like it's a different thing like if it was ordering flights right or hotels right but Uber is a 20 seconds job right so saving them 10 seconds so so do you already see users preferring uh the uh how do you call Lui Lui over uh manual entry or is it is it still in the future um yes this is really still in the future so far like user only interact with the Lui and for the wuber um yeah I mean by the end of the day the service is different from the the the the the the way you interact with gbt or entropic or this models right because the service is more of experience and would you really be able to experience this this good you know agent execution um workflow for you so I would say that uh I think the accuracy in this type of the scenario is not really important but what's more important is if the model is triggering the correct well thing for you or if you can actually even provide a user feedback loop to that and tell you model what do you want so then like it can be more and more personalized yeah I think here's the analogy is like that interesting to compare something like Instagram right where it actually creates demand instead of just um predicting what you want and so and over time you know potentially get you excited about doing something you know Uber maybe not the right example but broad speaking door Dash would be a good example a door Dash yeah like oh you know you saw something that would be exciting for you and you actually will order it uh I mean there's still you know there's still metrics around like click St Etc but yeah I think I think there's an interesting opportunity of like what is experience Beyond where you're balancing and that and this is where like for me an important question like how do we balance Beyond kind of optimizing for profit of the company that presents you this options to actually optimizing for individuals wellbeing right uh in kind of broader sense uh because kind of Instagram right now either way it is it is optimized for you to spend more time on Instagram but that's not because they are care about you but more because it generates some more money and so how do we kind of transition here to more uh more like hey this is actually good for you you know you're learning more you you know eating healthier food for example Etc uh through his predictions as well that's that's a good question but I want to throw a a question for for you as well do you would you like to have like a like a mom or dad or or teacher in your AI or in your you know physical device that's alert you saying hey you know eat healthy um no no snacks um so even though you've been browsing through McDonald's page for like five minutes and then the recommendation is still I check this pokeo U or check this salad I think that's um yeah that's something that I I think the AI um should be balancing or the application should um have a mechanism or a Twitter line where where that um they don't encourage the the evil um um human needs U more but they they don't stop it I think that's that's a clear line for us to see um but that's that's not um that's not the pfl to say um we we has to be honest uh and give back the um the data to to the model right um but yeah it's a it's an interesting question so is AI supporting you and encouraging you uh and you know just fulfilling you uh on whatever you already are or it's has to be acting you know uh more of a Godfather or God mom um to guide you into the right way and what is the right way right um but then uh again uh another interesting um interesting spark uh idea that you just mentioned uh what is the you know what is the what is the incentive that we could potentially um embedded within this uh incentive infrastructures um to give users a little bit more um a little bit more uh feedback or constant uh for them for us to engage with them or for us to use guiltily guiltily use AI agents to promote anything um from from you know the advertisement perspective whereas you know Instagram is obviously on top of that right um they don't give back uh any users any incentive but well Tik Tok might be a you know even worst case uh case study right here that uh they use a great algorithm to know you better and you know track you or trap you into into their application spending more time but for us like um you know the the applications that we see or that we hope to see in the ecosystem that utilizing pfl you have tremendous like private um datas right um how can you use it better um can you potentially also give back the uh individual users a little bit more incentive based on how much advertisements they have already seen or how many like um things that you help you to generate it uh your revenue and you gave us more portion back to the community I think I I think that's really important um um maybe the developers should design the mechanism within the D app or maybe the the infrastructures of ours we should have another layers of incentive in incentive mechanism to distribute it that proportion out um it's open questions for not only us but also for the dev com yeah yeah I mean I think there's there's a lot to figure out and research here so but uh thanks for coming we we are out of time so exciting to uh to hear kind of what you guys are building and uh yeah let's let's keep building uh thanks everyone for tuning in and see you next week thank you","transcript_synced_at":"2026-11-03T05:34:14.742Z","segments":[{"start_seconds":7,"text":"explore hold the future hand in hand in this Tech driven","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=7s"},{"start_seconds":16,"text":"wonderland hello hello uh Hey again everyone thanks for joining the office hours we're here with","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=16s"},{"start_seconds":25,"text":"Sero I I let them introduce themselves first uh hi you want to go of course","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=25s"},{"start_seconds":32,"text":"yeah glad to be here thank you for the great invitation Elia and Alex finally nice meeting you too um so I'm haha I'm","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=32s"},{"start_seconds":40,"text":"the co-founder of subo AI I came from a background with Masters um in robotics at upen and obviously worked in","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=40s"},{"start_seconds":48,"text":"blockchain for almost two years right now and really excited to see you know how Ai and web spray can intertwine with each other as and provides true user","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=48s"},{"start_seconds":56,"text":"case in the future yeah glad to be here thank you hello awesome yeah it's my turn I'm ccs","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=56s"},{"start_seconds":66,"text":"and I'm in charge of the suppose product um I lead everything from the open source project to the protocols to the like the what the D app that we're","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=66s"},{"start_seconds":74,"text":"integrating on top of the pfl and happy to be here and learn more awesome well uh thanks for joining","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=74s"},{"start_seconds":83,"text":"it's uh always amazing to dive in with Founders uh what you guys are working so","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=83s"},{"start_seconds":90,"text":"maybe tell us more about kind of how you're approaching uh the space what you see and maybe more as well about your","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=90s"},{"start_seconds":99,"text":"founding team of course yeah so me and CCE were were um family friends um know","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=99s"},{"start_seconds":105,"text":"each other since quite quite well and uh yeah so I always like worked in blockchain uh ever since I graduate from","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=105s"},{"start_seconds":113,"text":"master and don't really um you know knowing exactly how artificial intelligence could intertwine with blockchain whereas we started to you","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=113s"},{"start_seconds":122,"text":"know brainstorming about what what is the most um beneficial way for AI model developers and AI application developers","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=122s"},{"start_seconds":130,"text":"in the future whereas we started to notice like data shortage is one of the things this pretty much like the threshold um that um limited not limited","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=130s"},{"start_seconds":140,"text":"but you know in some ways limiting the AI uh Industries development and in another way if you know people we all","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=140s"},{"start_seconds":148,"text":"know human beings and if people don't get enough of or transparent or fairly incentive to you know convince them to","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=148s"},{"start_seconds":157,"text":"you know supports the AI industry so so why you know why are they contributing their private reserve data um and that's","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=157s"},{"start_seconds":165,"text":"sort of the major um questions that we started to see uh I think last year around like around March and that was","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=165s"},{"start_seconds":173,"text":"the first time um whereas I think the scaling Lev was such a big terms that um was being constantly mentioned not only","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=173s"},{"start_seconds":181,"text":"in uh AI but also in uh blockchain uh communities as well so we think hey if we can do anything that leverage the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=181s"},{"start_seconds":190,"text":"benefits or the advantage that blockchain could bring to AI industry or Community or even to see users that will be the best because ultimately people","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=190s"},{"start_seconds":199,"text":"are the one who is you know contributing and conducting the datas right so um that's the major topic and I can give a","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=199s"},{"start_seconds":207,"text":"deeper understandings about exactly how Sero as leverage that and how do we want to do with it uh but s CC I feel like","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=207s"},{"start_seconds":215,"text":"you know maybe you can give a better understanding from your perspective because you came from a Google um background and also you have been doing","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=215s"},{"start_seconds":223,"text":"AI uh applications research for quite a while as well right right so the problem fundamental problem that we're trying to","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=223s"},{"start_seconds":230,"text":"solve is that we see okay there's a lot of projects on top of the ecosystem that they worry about the privacy and um and","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=230s"},{"start_seconds":239,"text":"what we're trying to do is just a layer that helps every project whether you want to build an advanced application on","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=239s"},{"start_seconds":247,"text":"top of that or you want to build some use cases or you want to build agent or even you want to B to build some models","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=247s"},{"start_seconds":254,"text":"um and because like there's a lot of the open source models in the community now so then like what scenar like whether the scenario is application or use case","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=254s"},{"start_seconds":263,"text":"or models we would love love to um provide the protocol to them so that they don't really have to worry about","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=263s"},{"start_seconds":269,"text":"any data privacy reserved issue um because the way that we're using is not really just relates to data it only","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=269s"},{"start_seconds":277,"text":"relates to parameters so then like in that way nobody has to really worry about okay if we're going to save the data into some you know some","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=277s"},{"start_seconds":285,"text":"decentralized storage and then like to release it afterwards or we have to find different different very expensive um what data","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=285s"},{"start_seconds":293,"text":"consumption like Solutions out there what we do is pretty simple we just compress the data into the parameters so","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=293s"},{"start_seconds":301,"text":"that everybody what they've seen on chain is only the parameters not the data at all so then is we think it's the best way for so far all the project that","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=301s"},{"start_seconds":309,"text":"we've seen to protect the privacy and when it especially when it comes to the Apper layer to see you cases and to be model","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=309s"},{"start_seconds":318,"text":"issuers so give a summarizations of what exactly is subi subi is more like a you know Cutting Edge Ai and blockchain","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=318s"},{"start_seconds":326,"text":"focused protocol whereas we build on the visions of more of a decentralized and privacy preserving machine learning open","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=326s"},{"start_seconds":334,"text":"source Network or protocol um I would say we have a small vision and it could be a big Vision as well um to","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=334s"},{"start_seconds":341,"text":"democratize Ai and also ensure that the user data which is you know the um in our cases the model parameter weights","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=341s"},{"start_seconds":349,"text":"remains private while unlocking the values from decentralized AI computations so that's Sero","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=349s"},{"start_seconds":357,"text":"Ai and so my understanding is like underlying you're pretty much using a Federated learning model right where","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=357s"},{"start_seconds":365,"text":"every kind of users's device is uh updating the model on their own data so","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=365s"},{"start_seconds":372,"text":"indeed the data doesn't go anywhere never leaves consumer devices but instead the updated model is being","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=372s"},{"start_seconds":380,"text":"checkpointed back to blockchain uh kind of as updates being shared and so like that's kind of the general model you're","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=380s"},{"start_seconds":387,"text":"going after where the train happens always on user's device and then the kind of common model is uh then","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=387s"},{"start_seconds":397,"text":"available for everyone I would say that's a good uh understanding perspective um in our","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=397s"},{"start_seconds":403,"text":"words that we would like to say hey this is like more of a a pfl protocol whereas we leverages and a Distributive","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=403s"},{"start_seconds":412,"text":"blockchainbased middle layer infrastructure to facilitate secured and decentralized machine Learning System how to understand this better so through","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=412s"},{"start_seconds":420,"text":"the pfl model or through the pfl protocol note participants in you know using a AI applications and training AI","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=420s"},{"start_seconds":429,"text":"models without requ requiring a centralized data storage to process so yes your understanding for the most part is correct and blockchain serves as more","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=429s"},{"start_seconds":438,"text":"of a a ledger to validate and verify models update and you know give a fairly amount of the um uh incentive back to","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=438s"},{"start_seconds":447,"text":"the communities based on their contributions and also ensuring everything can be you know fully transparent um and fairly rewarded U","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=447s"},{"start_seconds":455,"text":"distributions for all the participants and to also for us to able to ensure this part of the distributions or uh of","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=455s"},{"start_seconds":463,"text":"the uh rewards there has to be you know a really um I would say a multi Rose participations engaged um from the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=463s"},{"start_seconds":471,"text":"blockchain perspective whereas the pfl creates a more of a a secured environment where multi- Staker","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=471s"},{"start_seconds":478,"text":"stakeholders they collaborate and to train and monetize AI models and AI applications with model parameters","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=478s"},{"start_seconds":485,"text":"weights um in a decentralized manner so I can go deep about essentially like how do we achieve and what exactly we have","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=485s"},{"start_seconds":493,"text":"designed um for for us to make sure that's the entire you know the dynamic uh economic models um and um to you know","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=493s"},{"start_seconds":502,"text":"guides the um guides the U uh incentive mechanisms within the ecosystem so so before before you go","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=502s"},{"start_seconds":510,"text":"into the incentives what what are the uh I'm curious to know what are the early use cases you see like like what sort of models because you know if you're","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=510s"},{"start_seconds":517,"text":"training llama on on you know like on a large Corpus of text that that doesn't have to be","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=517s"},{"start_seconds":525,"text":"Federated right so like I assume Federated learning is more interesting in the situation where uh we want to train like I guess Federated learning","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=525s"},{"start_seconds":533,"text":"came up in the context of training the swipe keyboard right right so like every everybody is typing on the keyboard uh there's a lot of data but obviously","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=533s"},{"start_seconds":541,"text":"nobody wants what they type on the keyboard to be known to Google or especially to other people right so if the model regurgitates uh you know your social","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=541s"},{"start_seconds":548,"text":"security number on another keyboard that's that's substandard right so I'm curious to know what the the use cases you envision or maybe you're already","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=548s"},{"start_seconds":557,"text":"going after yeah that's a that's a really good question um I would answer it from more of e ecosystem perspective whereas CC","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=557s"},{"start_seconds":566,"text":"you can jump in and say Hey you know these are the AI models that we already supplied but to be honest I would say the servo AI target audience is not","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=566s"},{"start_seconds":574,"text":"primarily focused on AI model issuers we're primarily focus on AI powered application developers groups whereas","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=574s"},{"start_seconds":582,"text":"you know some of the killer user case that we already seen that could include private preserving AI applications such as you know Health carees obviously","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=582s"},{"start_seconds":589,"text":"Federated learnings for collaborative AI models training across hospitals without sharing sensitive patients data so","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=589s"},{"start_seconds":596,"text":"that's one and secondly you know Finance um AI or AI for fraud detections and credit scoring Without Really exposing","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=596s"},{"start_seconds":605,"text":"your or anyone's personal financial data we all don't want to expose any of that uh informations out to the public now","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=605s"},{"start_seconds":613,"text":"sometimes even not to the bank and that's why we're having you know blockchains as more of the assets transportations um layers right so in","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=613s"},{"start_seconds":621,"text":"the uh in the other layers or in the other perspective the iot's um projects were um uh Edge based AI for smarts","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=621s"},{"start_seconds":630,"text":"devices which could leverage FL to also improve like device local device intelligence while also keeping the data","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=630s"},{"start_seconds":638,"text":"locally so that's from my perspective CeCe feel free to jump in more yeah yeah um it's it's my answer is pretty much","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=638s"},{"start_seconds":646,"text":"the same we don't really at this stage uh Power like for example 100 different models or even that you can uh open the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=646s"},{"start_seconds":653,"text":"new models we actually just um so we'll be able to connect with the majority of the fundamental open source models there like Lama opaka and the majority of the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=653s"},{"start_seconds":662,"text":"models that I think 90% of the application would bu based on that and then um what we uh what we can do is","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=662s"},{"start_seconds":669,"text":"that we can compress the model size or like we can actually compress the data size into 2 KB which um a lot of the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=669s"},{"start_seconds":676,"text":"isers can or a lot of developers or a lot of the application isers can build based on that no matter what application","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=676s"},{"start_seconds":684,"text":"it is one of the things that we're doing right now is to see what applications they might need a like the 2kb is still","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=684s"},{"start_seconds":691,"text":"like too you know like too big for them or too small for them or just for that use case they need a much more reliable","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=691s"},{"start_seconds":699,"text":"model they need a much more intelligent model so they probably cannot accept the 2 KB's performance they probably want 10","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=699s"},{"start_seconds":706,"text":"KB or 15 KB so that is the basically all the circumstance circumstances or the case case studies that we want to","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=706s"},{"start_seconds":714,"text":"implement in the open source Next Step just yeah just to double check when you","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=714s"},{"start_seconds":724,"text":"say 2kb do you mean that's uh that's update updates on top of a base model is that what you","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=724s"},{"start_seconds":733,"text":"mean yeah two yeah it means the parameters so the parameters weight is only 2","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=733s"},{"start_seconds":739,"text":"KB but this is a like low rank uh update right adaptation on top of bigger model","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=739s"},{"start_seconds":747,"text":"yes it Alex I no no that that was more or less","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=747s"},{"start_seconds":756,"text":"my own question I I just I just tried I first tried to understand uh where the number is coming from so it's interesting that you","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=756s"},{"start_seconds":764,"text":"mention uh that they they all measured in in KBS like I never heard of Laura","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=764s"},{"start_seconds":772,"text":"updates of that size so that's that's a novel concept for me","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=772s"},{"start_seconds":779,"text":"but but I'm also curious um so my understanding of Federated learning is relatively limited right so I I was sort","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=779s"},{"start_seconds":786,"text":"of into it back when it came out which was I think seven years ago now uh","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=786s"},{"start_seconds":793,"text":"and uh and so so the way I understand it is that in practice it's been shown to be","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=793s"},{"start_seconds":802,"text":"sufficiently uh my so my understanding right like going back to swipe right my understanding is that when I use Google keyboard uh it does use federative","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=802s"},{"start_seconds":810,"text":"learning right so everything I type does go to Google servers and it's strend and Google","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=810s"},{"start_seconds":818,"text":"TR the weight updates go to the Google servers right the weight updates yeah so it updates locally and then the r r diff is","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=818s"},{"start_seconds":828,"text":"sent and then they all get segregated right so","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=828s"},{"start_seconds":835,"text":"the um but but but I read that paper right back in the day in the Federated learning and the security analysis back then I I assume many things Chang in","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=835s"},{"start_seconds":844,"text":"seven years right but security analysis back then was very handwavy they they they put a very uh hand wavy security model under which it was very","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=844s"},{"start_seconds":853,"text":"secure uh and what I'm interested in is that you mentioned several use cases such as healthare uh or uh like Banks right so","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=853s"},{"start_seconds":861,"text":"so they they have uh security requirements where you cannot just publish a W paper on on","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=861s"},{"start_seconds":870,"text":"archive right so so like I'm curious if it is actually something viable today to use Federated learning for healthare like like if the healthare you know","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=870s"},{"start_seconds":878,"text":"companies would actually be willing to use to share you know weight updates based on the patient","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=878s"},{"start_seconds":885,"text":"data like is it the actual reality today does make sense yeah that makes","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=885s"},{"start_seconds":892,"text":"sense um but to be honest like from my perspective um we have started to uh interacting with some some of the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=892s"},{"start_seconds":899,"text":"healthcare startup um whereas the majority well uh the developers that we're we're focusing on right now is","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=899s"},{"start_seconds":907,"text":"most of the U Silicon Valley based startup who already started to implement some of the a models into their services","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=907s"},{"start_seconds":914,"text":"right so this is like our target audience and our target audience one of them uh is actually in conversations with uh San Francisco's hospitals","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=914s"},{"start_seconds":923,"text":"whereas they are supplying those hospitals to build a I wouldn't say a AI agent but more like a really comprehensive like more of a supply","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=923s"},{"start_seconds":932,"text":"chain uh uh optimiz optimizations uh systems whereas they could synchronize","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=932s"},{"start_seconds":939,"text":"and track about different like hospitals um and see like what are the for example like a hospitals has the majorities of","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=939s"},{"start_seconds":947,"text":"patients for for one type of disease or for several type of disease so they want to optimize you know how quick and how fast and how accurate they can fulfill","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=947s"},{"start_seconds":956,"text":"the uh the the inventory Ries for most of their medicines or drugs so that's one of the developer groups that we started to engage with and um their","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=956s"},{"start_seconds":965,"text":"their their concern is really straightforward those hospitals they don't want to share their data with any of the third parties including other hospitals that are collaborating with","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=965s"},{"start_seconds":974,"text":"them they could collaborate in the research um and you know case- by casee patient withs but they don't share informations in that levels and one of","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=974s"},{"start_seconds":982,"text":"the things that we uh we started to notice that um you know privacy it's is not only a big con concerns for them but","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=982s"},{"start_seconds":990,"text":"their inventory um tractions is also a big concerns in terms of like despite of the individual informations that could","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=990s"},{"start_seconds":998,"text":"be possibly leaking out I don't know I'm not exactly understanding you know the the concerns behind this reasons but I","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=998s"},{"start_seconds":1005,"text":"understand like from different perspective that anything that related with healthc Care status um and even from you know a commercialized the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1005s"},{"start_seconds":1013,"text":"perspective um they wouldn't want to share with any Outsiders including um I would say Googles because they obviously","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1013s"},{"start_seconds":1020,"text":"are not using Shopify or are not using other like um inventory optimization sets at the moments whereas they Pro uh","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1020s"},{"start_seconds":1029,"text":"they prefer to build their own controlling systems from all you know that links all of the Departments and","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1029s"},{"start_seconds":1036,"text":"Laboratories within their within their uh individual hospital um system so that's what we see um currently but","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1036s"},{"start_seconds":1044,"text":"again we're we're not directly engaging with the healthcare provider more engaging with the tech uh companies that","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1044s"},{"start_seconds":1052,"text":"provides you know consultings or agencies um um for for those Hospital customers","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1052s"},{"start_seconds":1059,"text":"yeah yeah yeah because because like something I'm very interested in is whether in like since fed learning came","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1059s"},{"start_seconds":1067,"text":"out they they actually managed to if now there is a sufficiently rigorous proof of uh Federated learning not leaking data that uh you know","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1067s"},{"start_seconds":1076,"text":"Healthcare or Banks would be fine with that uh because I think that like like if we get to that level of acceptance of","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1076s"},{"start_seconds":1083,"text":"Federated learning that enables a lot of interesting use cases yeah um interesting that you mentioned Bank we have not yet catch up with bank but we","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1083s"},{"start_seconds":1092,"text":"have catch up with several uh wallets um um developers already so major things that they have concerns is that well","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1092s"},{"start_seconds":1099,"text":"they don't really concerns about the Privacy um because everything um the users um transportations of any sort of","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1099s"},{"start_seconds":1107,"text":"assets um will be recorded on chain but they do concerns about hey what is the next step for those wallet applications","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1107s"},{"start_seconds":1115,"text":"to to engage or to expans their product features to better engage with their communities right now we see wallet as","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1115s"},{"start_seconds":1122,"text":"you know um some applications that you receive and you know issue out um the transactions right in the future most of","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1122s"},{"start_seconds":1130,"text":"the wallet uh developers they're not thinking about the that they're they're thinking to even intertwine or um engage","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1130s"},{"start_seconds":1137,"text":"with some assets management product features in that can be powered by AI or they can either they most of them are","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1137s"},{"start_seconds":1145,"text":"thinking hey can they engage with AI uh agents to supply and provide some of the um what do you call it like um","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1145s"},{"start_seconds":1153,"text":"investment device within within the wallet product so this is like the next Generations of how I see you know our","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1153s"},{"start_seconds":1161,"text":"potential customers will be developing in if that's the case then hey how do you deal with the user data priv privacy","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1161s"},{"start_seconds":1169,"text":"it has nothing to do with blockchain in that way it's actually related back with you know most of the users are having the Wallets on their personal devices","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1169s"},{"start_seconds":1177,"text":"such as PCS and iPhones like me um but how do I how do I leverage that perspective I don't really use uh usdt","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1177s"},{"start_seconds":1186,"text":"and usdt usdc to pay most of my personal Supply expense I use Fiat I use you know","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1186s"},{"start_seconds":1193,"text":"I still live a good life uh here on Earth so I have to engage with Fiat payments mostly for my uh for my private","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1193s"},{"start_seconds":1200,"text":"life right well if that's the case then how can they understand this person's expense and how can they locally trained","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1200s"},{"start_seconds":1210,"text":"uh you know some proportions of that models to be able to power and also provide good informations and good advice within the wallet uh I think","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1210s"},{"start_seconds":1218,"text":"that's sort of related back with what we're seeing for our user case and um yeah and do think like Federated","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1218s"},{"start_seconds":1225,"text":"learning is a power of that and CC you can feel freee to speak more right before I speak more just","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1225s"},{"start_seconds":1233,"text":"wondering if you have any questions so far no not","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1233s"},{"start_seconds":1243,"text":"yet okay all right um okay so I think in terms of the application we just talk about you know like different","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1243s"},{"start_seconds":1250,"text":"applications like well like Hospital like well even Finance um maybe there's Transportation like a lot of the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1250s"},{"start_seconds":1257,"text":"different scenarios but what we think right now what should be the most application is there's a lot of the agent companies are building some","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1257s"},{"start_seconds":1266,"text":"service aggregators or scenarios or when it comes to you know efficiently help customers to leverage the agent or Theos","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1266s"},{"start_seconds":1275,"text":"to do something so in that case it is a model issuer or like it is a Tob customer who's building an application","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1275s"},{"start_seconds":1283,"text":"but it's directly facing a lot of the 2C users and a lot of 2C users they actually really care about number one how fast this process can be and number","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1283s"},{"start_seconds":1292,"text":"two how private like like if they can actually protect their data and number three if they can save some money so our","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1292s"},{"start_seconds":1301,"text":"proof of Fe learning actually can be able to enable all of this three we can not only just because you have less","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1301s"},{"start_seconds":1309,"text":"parameters out there um for the low training for the low run training so then uh in that case scenario you can we","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1309s"},{"start_seconds":1316,"text":"we actually did some calculation we can save the cloud cost by 30% to 40% so it is a low cost approach for a lot of the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1316s"},{"start_seconds":1324,"text":"2B customers and two when we said that we compress the parameters into 2kb is a","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1324s"},{"start_seconds":1331,"text":"way for people to um to have a great user experience out there to even um enable more devices in the future for","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1331s"},{"start_seconds":1339,"text":"training and three um when we see privacy reserved we mean that because we don't directly interact with the data itself we only interact with the pro","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1339s"},{"start_seconds":1348,"text":"with the the parameters so then we don't really get the user data to upgrade the model um it's very different from the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1348s"},{"start_seconds":1355,"text":"traditional F learning training or the traditional fine tuning or the ways like that so um like on that three","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1355s"},{"start_seconds":1364,"text":"perspective we think a lot of the agent applications a lot of the applications that's just directly facing into 2C are","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1364s"},{"start_seconds":1372,"text":"the GD applications no matter it is in hospital or it is in finance or it is in transportation","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1372s"},{"start_seconds":1380,"text":"so so you mentioned uh you mentioned it it's a cheaper for people to train uh the model so so in","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1380s"},{"start_seconds":1389,"text":"this case uh the idea would be that that someone is training the model on their data and","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1389s"},{"start_seconds":1396,"text":"and and uh the Federated learning helps them train it cheaper or they or or they still train it on the the data of of someone else what is","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1396s"},{"start_seconds":1406,"text":"the use case so so the person who 30% are they training their data or they train on someone else's data they're","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1406s"},{"start_seconds":1414,"text":"still training their data but the data is not being trained the parameter is the one that's being updated so you have less parameter and intuitively the cost","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1414s"},{"start_seconds":1423,"text":"is less I see so it's not just not just about Federated learning it's generally a different architecture which is","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1423s"},{"start_seconds":1430,"text":"somehow smaller than P water yes","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1430s"},{"start_seconds":1439,"text":"so maybe maybe Switching gears a little bit um as this is happening right um how","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1439s"},{"start_seconds":1447,"text":"is kind of quality insured how do people get rewarded for for it what is the um","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1447s"},{"start_seconds":1455,"text":"kind of what is the economic incentives to be kind of doing the right thing as well as disincentive for doing the wrong","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1455s"},{"start_seconds":1464,"text":"thing good question so we have already designed really I would say Dynamic economic model whereas we guides users","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1464s"},{"start_seconds":1471,"text":"to build our ecosystem build on top of the ecosystem and semi like control the quote unquote evil behaviors of users","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1471s"},{"start_seconds":1480,"text":"like you said for example like Elia or Alex you spend it like 10 days playing on one of the AI applications that should have deployed on top of pfl on","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1480s"},{"start_seconds":1489,"text":"our pfl versus um you know a soueast Asia Studio that they have like several different devices also spent a few days","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1489s"},{"start_seconds":1497,"text":"but doing like you know redundant movement or actions that still accumulated like 32 KB versus your 32 KB","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1497s"},{"start_seconds":1505,"text":"like who evaluated more um obviously there has to be a really straightforward and transparent uh scenario and mechanism to guide this but before we","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1505s"},{"start_seconds":1514,"text":"dive deep into that I want to emphasize like who is going to be the one who raises hey you know Alex or Elia your 72","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1514s"},{"start_seconds":1522,"text":"KB um earns uh let's say like 0.1% of the entire sub token versus you know the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1522s"},{"start_seconds":1529,"text":"southeast Asia studio uh could earn like 000000 z. one% of the token there has to","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1529s"},{"start_seconds":1538,"text":"be a multi-roll on chain governance mechanisms that implement it within the system whereas we design multi different","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1538s"},{"start_seconds":1545,"text":"um positions to raise this um the incentive and also issue it right so Ro a is going to be the reviewer and","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1545s"},{"start_seconds":1552,"text":"uploader who is going to be linking the results of horizontal comparations of model performance um and role B is going to be the challengers who is going to","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1552s"},{"start_seconds":1561,"text":"challenge and also reveals role A's results role C is going to be you know the challengers um admin or super CH","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1561s"},{"start_seconds":1568,"text":"Super admin who is going to review on top of Ro bees whereas we see well in the future we hope that you know there","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1568s"},{"start_seconds":1575,"text":"will be millions of Toc users um via those AI applications who is constantly engaging with the pfl protocols well if","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1575s"},{"start_seconds":1584,"text":"that's the case then row a roow b row C has to be or what we foresee it has to be a primate uh a a prime a paramid","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1584s"},{"start_seconds":1594,"text":"structures whereas rule a will be contains the largest participations populations to be able to execute and","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1594s"},{"start_seconds":1601,"text":"you know instantly well not instantly but you know we're trying to have them to instantly Rewards or raling the model","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1601s"},{"start_seconds":1609,"text":"performance whereas Rob will be more of a prestigious um or you know sely lesser uh um uh popul ation wise of","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1609s"},{"start_seconds":1619,"text":"The participations Who on top of Ro a and is checking on any red flags that Ro a raise it up or you know if there are","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1619s"},{"start_seconds":1628,"text":"any enormous uh activities that has been perceived or uh you know voted by ro and rosi will be you know a really um","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1628s"},{"start_seconds":1637,"text":"smaller group whereas we would like to engage with more of the AI expertise and you know obviously so as teams and you","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1637s"},{"start_seconds":1644,"text":"know the other teams that engages and collaborate with us to also overseeing the entire structures in case","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1644s"},{"start_seconds":1651,"text":"there are any like enormous um incentive uh executions on top of the protocol so","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1651s"},{"start_seconds":1658,"text":"this unchain distributions um systems we will generat different levels of tokens based on frequency um numbers of","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1658s"},{"start_seconds":1666,"text":"participants and the importance based on you know role a ro B Ro R C's distributions of different rewards and you know the next questions for us to","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1666s"},{"start_seconds":1675,"text":"resolve is that who is going to be the reward receiver obviously there are two um you know stakeholders going to be","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1675s"},{"start_seconds":1682,"text":"receiving the rewards developers and communities which means the 2C users who at least have a physical device to be able to engage the AI applications that","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1682s"},{"start_seconds":1691,"text":"Deploy on top of the pfl protocol whereas we um we have this metrix we uh we check on the qualities uh and the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1691s"},{"start_seconds":1699,"text":"quantities and whereas the quantities is really straightforward we're quantifying based on the weight over each model","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1699s"},{"start_seconds":1705,"text":"updation so Alex and I 32kb versus you know Southeast Asia's 32 KB in this","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1705s"},{"start_seconds":1712,"text":"level might be weights the same but then deeper engaged with the qualify the uh qualifying pfl score whereas we combined","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1712s"},{"start_seconds":1722,"text":"uh data scores U model contribution scores and um validation accuracy scores together this pfl score is proportionate","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1722s"},{"start_seconds":1730,"text":"to the pretty much to the amount of newly mained rewards that will be allocated back to the participants so even though the 32kb is the same but","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1730s"},{"start_seconds":1739,"text":"then underneath that the qualities that it brings to the model performance updates are different so at this moment","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1739s"},{"start_seconds":1748,"text":"everything is still in testification Period whereas you know we have already Implement one of the universal uh um a","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1748s"},{"start_seconds":1755,"text":"model performance um uh formulas or evaluation formulas but in the future there will be you know uh open door for","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1755s"},{"start_seconds":1764,"text":"any other formulas that can be raised by the communities and Developers in the future so that's you know the entire um","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1764s"},{"start_seconds":1771,"text":"I would say like toomics incentive mechanism that we started to builds up and um for us to be able to better","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1771s"},{"start_seconds":1779,"text":"execute on top of it there has to be you know different um participations representing different um representing","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1779s"},{"start_seconds":1787,"text":"different um uh scope and stakeholders um to be able to","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1787s"},{"start_seconds":1792,"text":"engage so I find it um so I mentioned it every now and then on the office hours right but we've been running a system","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1792s"},{"start_seconds":1800,"text":"where people annotate data which is not quite the same as Federated learning um or or or not the same at all uh but it","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1800s"},{"start_seconds":1808,"text":"shares an important property that people contribute something uh that is only that can only be subjectively","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1808s"},{"start_seconds":1816,"text":"verified uh and they get the reward for that right because uh I don't I don't know very well how proed learning works right but but like from the first","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1816s"},{"start_seconds":1823,"text":"principles right because I don't reveal my data I don't think I can provide any that the data I used uh was in any way","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1823s"},{"start_seconds":1832,"text":"relevant to the data set right and uh so correspondingly the way I understand uh the fundamental concept of pro","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1832s"},{"start_seconds":1839,"text":"refigerated learning is that at the end of the day uh the only way to confirm that what I submitted makes sense is by","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1839s"},{"start_seconds":1848,"text":"you know testing it against some held out held out set right uh so uh these a","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1848s"},{"start_seconds":1856,"text":"uh uh right so so in the like in case of uh coming back to","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1856s"},{"start_seconds":1864,"text":"the swipe example right if I use Google keyboard I have literally no incentive to submit uh incorrect completions to","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1864s"},{"start_seconds":1872,"text":"Google right and if there's a 100 million people using Google keyboard uh even if I like try to Doos them with invalid submissions from one person I","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1872s"},{"start_seconds":1880,"text":"think they can uh prevent it from happening right in case of uh Sero there's a m massive","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1880s"},{"start_seconds":1888,"text":"incentive for people to submit in valid results right because it's uh because I can only get that much useful data to","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1888s"},{"start_seconds":1896,"text":"submit to you right but I can fabricate an infinite number of bad updates right uh and so people will be doing that","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1896s"},{"start_seconds":1903,"text":"right like the biggest lesson I learned from from our data notation efforts is that once you start paying people money","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1903s"},{"start_seconds":1911,"text":"they become extremely creative right like the IQ of people grows of two the moment they can get","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1911s"},{"start_seconds":1919,"text":"money for free right so it's so I think what it gets interesting","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1919s"},{"start_seconds":1926,"text":"is uh like like the actual concept of of proof of predated learning right that's something that is very interesting to","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1926s"},{"start_seconds":1932,"text":"dig into at some point to understand how like let's say let's say like I assume","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1932s"},{"start_seconds":1940,"text":"you need updates from you know like let's say thousands of people right so you get thousands of updates you obviously cannot do like every permutation of them to validate and each","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1940s"},{"start_seconds":1948,"text":"particular update is relatively small right like when I train a neural network between two consecutive ooks the loads can go up it doesn't mean that the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1948s"},{"start_seconds":1955,"text":"update was useless right it's just a uh uh right so it's interesting if you","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1955s"},{"start_seconds":1963,"text":"guys can share like any high level uh of uh uh proof of Federated learning like","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1963s"},{"start_seconds":1971,"text":"what like at which level does it operate how how much interaction does it require from the community","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1971s"},{"start_seconds":1979,"text":"uh how how reli also who runs this Val like this evaluation right uh like the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1979s"},{"start_seconds":1987,"text":"who who who evaluates the validation accuracy that's a really good question so that goes back to like the mechanism","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1987s"},{"start_seconds":1994,"text":"that I was describing earlier um there will be like Ro a roow b r C's right so whereas Ro a who is going to be the reviewer or the uploader who is like","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=1994s"},{"start_seconds":2003,"text":"making comparations or or horizontal comparations of different model uh performance to answer uh your questions","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2003s"},{"start_seconds":2010,"text":"Alex yes each of the updates has to be um has to be verified otherwise you know like you said people get extremely","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2010s"},{"start_seconds":2019,"text":"creative to earn um you know to earn money or to earn incentive if they don't know the criteria even they know the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2019s"},{"start_seconds":2026,"text":"criteria they still would think hey maybe I can squee something in or I I can s squiz you know the Redundant data","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2026s"},{"start_seconds":2033,"text":"in into the entire boad um board update and you wouldn't be able to find it you will still evaluate me um as the 32kb as","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2033s"},{"start_seconds":2043,"text":"the good 32kb no it's not going to be true in in our scenario the pfl score that we designed here actually check","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2043s"},{"start_seconds":2051,"text":"based on top of the qualities of each different updates um three things um like I mentioned before has been","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2051s"},{"start_seconds":2058,"text":"combined right here uh the data score um the model contribution score and the validation accuracy score so that's the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2058s"},{"start_seconds":2066,"text":"high level like giving just giving you a framework of how exactly the PFO scores has been made up so those are the three","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2066s"},{"start_seconds":2073,"text":"perspective whereas we weighted them differently uh with a formula right now um and we issue uh the incentive based","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2073s"},{"start_seconds":2082,"text":"on top of the pfl score although the quantity um is really important because","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2082s"},{"start_seconds":2090,"text":"that's like one of the first thresholds but we treat it as more of a normal parameters within this entire pfl score","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2090s"},{"start_seconds":2097,"text":"formula whereas it's like a like a small things but then you have like a big formulas up up or right after that","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2097s"},{"start_seconds":2105,"text":"whereas this is like the true part of how much incentive as a normal to see users or any of the users you will be","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2105s"},{"start_seconds":2113,"text":"able to earn so that is the um that is the pfl score that we embedded within the system and I understand what you're","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2113s"},{"start_seconds":2121,"text":"saying hey if there are millions of users how can you execute or how can you reveal each","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2121s"},{"start_seconds":2128,"text":"individual updates so this is one of the things that we want to engage with the developers and also try to understand you know how the developer communities","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2128s"},{"start_seconds":2136,"text":"uh designed uh their applications what we are doing is that we are engaging with one groups of uh developers whereas","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2136s"},{"start_seconds":2144,"text":"designing a more of a a application like um a decentralized AI","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2144s"},{"start_seconds":2151,"text":"Ser whereas we call them Ferry um so fery is applications that's um is a proof of concept within the ecosystem um","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2151s"},{"start_seconds":2160,"text":"you can engage with it um I'm sure like everyone if you guys have time already check on the uh Apple product U uh","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2160s"},{"start_seconds":2167,"text":"launch um presentations yesterday or last night um you engage with the this applications and then you you know the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2167s"},{"start_seconds":2175,"text":"AI agents behind this or the AI model behind um that applications would able to help you to execute say hey give you","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2175s"},{"start_seconds":2183,"text":"a Uber or give you a LIF lift or you know um buy you something on Amazon uh simultaneously by already knowing you","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2183s"},{"start_seconds":2192,"text":"enough um so this is like one of the applications that we're we're seeing and we're supporting them um from the pfl","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2192s"},{"start_seconds":2199,"text":"infrastructure U or you know the middle layer um perspective um and this application or this application","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2199s"},{"start_seconds":2207,"text":"developer groups they're really interesting they don't update each actions of the models um or of the user","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2207s"},{"start_seconds":2215,"text":"interactions um on chain what they do is that they gave the users a freedom for them to choose hey do you want to update","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2215s"},{"start_seconds":2223,"text":"all of the actions that you already interacting with the applications on a weekly basis or on a daily basis so","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2223s"},{"start_seconds":2230,"text":"actually um surprisingly the the 2C users or the individual users they have the control of how frequent they want to","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2230s"},{"start_seconds":2238,"text":"locally even train the models so it's not like you're interacting with the uh applications and simultaneously it has been already trained yes it gathers um","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2238s"},{"start_seconds":2247,"text":"that's data from a local devices perspective but it it doesn't really necessarily means it will be trained","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2247s"},{"start_seconds":2254,"text":"locally simultaneously so user has 100% controls at least for this AI application group whereas we started to","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2254s"},{"start_seconds":2262,"text":"see they gave the users the choice to when will be the best time to train your to train on top of your phone right they","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2262s"},{"start_seconds":2270,"text":"are right now supplying iPhone 14 and plus and so this this is the freedom whereas um it doesn't really cost for","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2270s"},{"start_seconds":2279,"text":"the entire protocols to constantly have to update or have to review each single actions or each single actions based on","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2279s"},{"start_seconds":2287,"text":"top of that what is the model parameter weight and then evaluated that it would been it wouldn't be that efficient um we would actually encourage the developers","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2287s"},{"start_seconds":2296,"text":"to well designed um their applications to be able to maintain the best um","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2296s"},{"start_seconds":2302,"text":"efficiency for the entire network or protocol uh ecosystem um and then after the users hey let's","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2302s"},{"start_seconds":2310,"text":"say weekly uh updates and accumulated enough of actions or interactions with the AI model powered applications","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2310s"},{"start_seconds":2319,"text":"locally um they trained um their his historical tracks or um um their their","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2319s"},{"start_seconds":2325,"text":"personal datas and then from there um uh we will be able to you know help them to train on on local device scenario and","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2325s"},{"start_seconds":2334,"text":"that parameters which will be recorded on chain with individuals um users notification saying hey you already update this part","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2334s"},{"start_seconds":2343,"text":"of the things on chain and then from there the AI reviewers or the uploaders will be you know in interacting with um","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2343s"},{"start_seconds":2352,"text":"that informations and verified you know what is the what is the value of it so uh overall like there will be or we","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2352s"},{"start_seconds":2361,"text":"foresee there might be some traffic gem um for the row a since um we're our our","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2361s"},{"start_seconds":2369,"text":"mission is to power millions of users there might be some gems traffic gems for RO a to uh review um because we","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2369s"},{"start_seconds":2376,"text":"cannot you know say hey um all the AI application developers you have to do weekly or monthly we give them the ches","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2376s"},{"start_seconds":2383,"text":"but we you know nicely advise them to have you know wisely designed of how often they want to update um the model","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2383s"},{"start_seconds":2392,"text":"parameters weight and then from there row a is going to be a really big po ation since they are you know not as","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2392s"},{"start_seconds":2400,"text":"much as the to see individual developers uh or to see individual users um but they they needs to be at least uh you","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2400s"},{"start_seconds":2409,"text":"know oneir or one quarters of that populations to be able to maintain such a big um traffic um so that's what we're","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2409s"},{"start_seconds":2417,"text":"seeing but at at the beginning everyone is welcoming to become the reviewer and uploader as long as you have the physical device to run on top of it uh","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2417s"},{"start_seconds":2425,"text":"and you know submit your results on chain so that's what we're seeing um yeah I hope that answer your","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2425s"},{"start_seconds":2434,"text":"questions so so the idea of uh of doing it infrequently is primarily to save the bandwidth right it's not like security","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2434s"},{"start_seconds":2442,"text":"consideration or no I'm assuming if you if you're evaluating every single like update on","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2442s"},{"start_seconds":2452,"text":"some evaluation metric as well right you kind of need to wait until there's enough updates for that to matter I","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2452s"},{"start_seconds":2460,"text":"think I mean there's multiple Dimensions here that I think worse you know considering","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2460s"},{"start_seconds":2467,"text":"because uh I mean obviously these models are highly non-interpretable at the current stage and so one of the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2467s"},{"start_seconds":2477,"text":"challenges that people already seeing is so-called Sleeper Agent problem where we're able to introduce pretty much a","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2477s"},{"start_seconds":2485,"text":"triggering behavior that doesn't actually show up at the evaluation stage uh by pretty much tra you know creating","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2485s"},{"start_seconds":2494,"text":"gradient updates on specific like you know if it's 2025 you know do this and otherwise don't trigger that and so I","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2494s"},{"start_seconds":2502,"text":"think there's an interesting question of the evaluation set is always going to be you know just the subset of the distribution and especially anything","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2502s"},{"start_seconds":2510,"text":"that's out of distribution on the data is going to be really hard to evaluate so I think I I I think you you're trying","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2510s"},{"start_seconds":2518,"text":"to pretty much say well we're going to have smaller updates right in general which is kind of harder to influence but","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2518s"},{"start_seconds":2525,"text":"I think for a broader kind of decentralized training space this is probably one of the major issues in","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2525s"},{"start_seconds":2533,"text":"general is like how do you like if you're just getting parameter updates and that's all you don't really know if","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2533s"},{"start_seconds":2540,"text":"like this is really malicious update that like looks okay or this is actually uh legit update or this is just a random","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2540s"},{"start_seconds":2549,"text":"update that actually doesn't improve the model and somebody just spamming with you know random numbers um as well right","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2549s"},{"start_seconds":2556,"text":"so I think kind of a more I would say probably more studying like a full kind of research needed on","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2556s"},{"start_seconds":2565,"text":"like the dependency and detection of those things uh depending as well the size of that like of the for example low","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2565s"},{"start_seconds":2573,"text":"rank matricies right again I'm assuming the smaller the smaller low rank metrics it is the easier it is","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2573s"},{"start_seconds":2582,"text":"to the harder it would be to like create some kind of malicious attacks but uh it potentially will be also harder to see","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2582s"},{"start_seconds":2590,"text":"Improvement and validation set right because you can imagine that like let's say your validation set is how good it does on","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2590s"},{"start_seconds":2598,"text":"math and uh like it's just not going to get better in math if you just have two kilobytes of uh updates to parameters uh","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2598s"},{"start_seconds":2606,"text":"because it requires like a systematic Improvement to the capabilities of the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2606s"},{"start_seconds":2613,"text":"model yeah I do agree CeCe by the way you're muting I know you're speaking no I was taking notes actually","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2613s"},{"start_seconds":2622,"text":"this is well said I think I mean the other kind of","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2622s"},{"start_seconds":2632,"text":"just thought I've had was broader um like if if you're giving this to","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2632s"},{"start_seconds":2641,"text":"developers are there also ways for developers to design model architectures that are trying to extract user","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2641s"},{"start_seconds":2650,"text":"information because I think that's that's another dimension of kind of and and I mean we also thinking about it on","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2650s"},{"start_seconds":2657,"text":"the agent side right as you're designing agents like if the devel developers militia they're gonna try to actually","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2657s"},{"start_seconds":2665,"text":"extract user data from the through this interaction uh so there's kind of another dimension of this as well um","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2665s"},{"start_seconds":2673,"text":"which you know you potentially can like train a model to generate model architectures that are uh in updates","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2673s"},{"start_seconds":2680,"text":"will actually like when you update it will encode the information of the user that you actually","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2680s"},{"start_seconds":2695,"text":"want yeah exactly actually I remembered another","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2695s"},{"start_seconds":2706,"text":"question I had so so you said that you want users to send an update once a month right so so in case of these","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2706s"},{"start_seconds":2713,"text":"models I I guess it's highly dependent on the use case right but it's one thing when the moduel is already well tramed and you're just updating it right how","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2713s"},{"start_seconds":2721,"text":"does the model how does the model boot how do you bootstrap the model do do you like","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2721s"},{"start_seconds":2727,"text":"uh do you reduce the update time to like once per day or or is the model initially already somewhat good and","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2727s"},{"start_seconds":2735,"text":"Federated learning is only used for maintenance Yes actually depends depends on the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2735s"},{"start_seconds":2743,"text":"project um if the project like you can just set up any frequency you want um and we don't have like um it's not like","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2743s"},{"start_seconds":2750,"text":"when when it comes to iPhone you only can update in a certain period because everybody's using it but then we give the what different projects the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2750s"},{"start_seconds":2758,"text":"flexibility to update it like whenever they think uh this is a great time to update it we've got enough parameters uh","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2758s"},{"start_seconds":2766,"text":"the model's intelligence has to be updated and and uh for the I'm also","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2766s"},{"start_seconds":2777,"text":"curious what is the so the update I assume the like whatever multiple kilobyte update this is not the whole","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2777s"},{"start_seconds":2783,"text":"model right this is a uh this is the lur right so so how big are the base models do do you have any insight into what is","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2783s"},{"start_seconds":2793,"text":"the because I presume they should be small enough for the for the iPhone to to uh to unroll them without draining","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2793s"},{"start_seconds":2800,"text":"the battery right yeah so we have like we're supporting four models at this moment and this just pretty much all the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2800s"},{"start_seconds":2808,"text":"open source model size um like you know there's like 7B there's like uh were just like regular size that we support","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2808s"},{"start_seconds":2816,"text":"so that is a big model that we","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2816s"},{"start_seconds":2826,"text":"use yeah I definitely need to read up more on the on the actual approach I think 32 kilobyte","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2826s"},{"start_seconds":2835,"text":"update it's not L something even","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2835s"},{"start_seconds":2843,"text":"smaller I mean what kind what kind of yeah improvements do you see actually through this","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2843s"},{"start_seconds":2850,"text":"um I guess that that's probably like an interesting question around like for how many users you know sending updates like","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2850s"},{"start_seconds":2858,"text":"that like how like how much more how much improvement you can actually get on on what kind of problems and what kind of valuation","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2858s"},{"start_seconds":2867,"text":"sets the improvements in terms of the next uh well like next step what in in terms of the evaluation","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2867s"},{"start_seconds":2876,"text":"St that you're using like if it's you know I mean if it's like a 7B llama and you're testing","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2876s"},{"start_seconds":2884,"text":"it on I don't know doing math or doing uh next toing prediction for","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2884s"},{"start_seconds":2890,"text":"keyboard or whatever right yeah so now we're like doing some improvements on the agent we're doing our own D app and","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2890s"},{"start_seconds":2899,"text":"it's a service aggregator and um it basically we it's like we're calling like the whole end to end agent to well","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2899s"},{"start_seconds":2908,"text":"like finish for example Uber deliveries and door Dash deliveries and even Amazon ERS and this is the Improvement that","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2908s"},{"start_seconds":2915,"text":"we're using in the evaluation data set that we're testing on pfl what does agent do in case of uh","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2915s"},{"start_seconds":2926,"text":"what is the agent in this case of like let's say uh door Das deliveries so by the end of the day we do want um this","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2926s"},{"start_seconds":2934,"text":"this to be very PR like to be like personalized agent that can do everything for you and personalized assistant that can just one stop respond","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2934s"},{"start_seconds":2942,"text":"to all your requirement but now um we only have been pushed like two service aggregator directions when is the Uber","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2942s"},{"start_seconds":2951,"text":"when is the door Dash and when is the W like when is when is the the Amazon delivery so those are the three aggregators and we think that the data","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2951s"},{"start_seconds":2959,"text":"can be really valuable because it's essentially the human or like the user Behavior data and","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2959s"},{"start_seconds":2967,"text":"yeah and we what fundamentally what we want to achieve is uh two Direction one is that okay now we have an agent that can help people to do a lot of the well","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2967s"},{"start_seconds":2976,"text":"like the services or the real time applications the other is we do want to prove that the proof of Federate learning in this evaluation data set","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2976s"},{"start_seconds":2984,"text":"actually works by um having like by having people seeing their data compressed into a parameter and you can","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2984s"},{"start_seconds":2992,"text":"actually access access all the data from the wallet and see the data or parameters on chain what is the model predicting","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=2992s"},{"start_seconds":3000,"text":"predicting in case of the doorage delivery like is it is it like the actual process of","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3000s"},{"start_seconds":3007,"text":"ordering stuff or is it uh at the delivery time what does the agent do like as as a user what is what is my uh","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3007s"},{"start_seconds":3015,"text":"what is my interaction with the agent your interaction is the agent is that so for example when you are interacting with Lui right like language","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3015s"},{"start_seconds":3023,"text":"user interface and you are telling the well the model that okay you want to eat something and then like it just","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3023s"},{"start_seconds":3032,"text":"automatically locates your like your position and it just like you know ERS whatever you want uh based on your the databases you have and based on the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3032s"},{"start_seconds":3041,"text":"prediction of like what should be the taste of the food you like and you just like they just like automatically","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3041s"},{"start_seconds":3047,"text":"execute and um retrieval the correct app um that just retrieve all the services based on","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3047s"},{"start_seconds":3055,"text":"that so I guess with Uber is the same it's like trying to predict where I'm going yes Uber is the same","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3055s"},{"start_seconds":3066,"text":"too yeah I I remember interesting and I were doing that in 2017 obviously less","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3066s"},{"start_seconds":3074,"text":"the technology wasn't as sophisticated but turns out it's not that frequent at least in case of an Uber unless it's literally your next uh event in the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3074s"},{"start_seconds":3083,"text":"calendar it's not that easy to predict where the Uber is going uh so yeah interest would be interesting","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3083s"},{"start_seconds":3090,"text":"to see if the you know the time has changed and now it's a moreable the user experience that you will be able to get right now is um as a","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3090s"},{"start_seconds":3099,"text":"user you can talk to we have our own open source model called Ferry you talk to our fery model and then you can start","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3099s"},{"start_seconds":3107,"text":"to engage our faery model is really you know like emotional and it's kind of just like a like a assistant that type","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3107s"},{"start_seconds":3113,"text":"of the model in in terms of the DNA and then uh when you tell the fairy that okay when it triggers about Amazon or","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3113s"},{"start_seconds":3121,"text":"Uber it just automatically redirects you to there and finishes service for you and you can always go back and you can be able to select if you want to you","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3121s"},{"start_seconds":3130,"text":"know you want to earn some tokens by contributing this data and then you can actually access the parameters that's we","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3130s"},{"start_seconds":3137,"text":"put on chain of this data and you you can see from the wallet you can access from the wallet you can see on chain and","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3137s"},{"start_seconds":3144,"text":"you can see how much token you get so so so I presume you guys are uh","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3144s"},{"start_seconds":3155,"text":"battle testing it locally right so like how frequently how like what percentage of of uber right you don't have to","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3155s"},{"start_seconds":3163,"text":"intervene and uh and it gets a trade what percentage of the oh it depends on user so like if the user","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3163s"},{"start_seconds":3171,"text":"wants 100% then they get more token if the like if the user wants like 50 7 then depends on them yeah yeah no no my","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3171s"},{"start_seconds":3179,"text":"my question is when you personally use it what percentage of uber rights that fairy orders are correct and what percentage you still have to intervene","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3179s"},{"start_seconds":3187,"text":"oh now we have like around 90% accuracy out there on the Ubers and there's only 10% of the edge cases that you will","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3187s"},{"start_seconds":3194,"text":"never know and and like in terms of uh","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3194s"},{"start_seconds":3207,"text":"so that's that's uh a question more of like more in terms of U user research","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3207s"},{"start_seconds":3214,"text":"right so hypothetical City I I don't use any Apple product right so so let's say Google Google whatever it's called Google Now Google","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3214s"},{"start_seconds":3222,"text":"sirity uh she can she can guess more than 90% of things I want correctly right like like if I say Okay Google and","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3222s"},{"start_seconds":3229,"text":"ask her for something uh uh it's a he gets of more than 90 % of the time","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3229s"},{"start_seconds":3237,"text":"right and yet I never use her because uh the the 5% it doesn't get it they extremely right so so like the the the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3237s"},{"start_seconds":3245,"text":"sheer expectation my head that I can get into those 5% makes me you know like I would set an alarm manually even though it's going to take me 10 seconds longer","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3245s"},{"start_seconds":3254,"text":"right so I'm curious if this is something uh if in your experience you you're sort of crossing the threshold where people because ordering an Uber is","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3254s"},{"start_seconds":3263,"text":"not a like you're not saving a person an hour of work right like it's a different thing like if it was ordering flights right or hotels right but Uber is a 20","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3263s"},{"start_seconds":3272,"text":"seconds job right so saving them 10 seconds so so do you already see users preferring uh the uh how do you call Lui","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3272s"},{"start_seconds":3281,"text":"Lui over uh manual entry or is it is it still in the future um yes this is really still in the future so far like","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3281s"},{"start_seconds":3290,"text":"user only interact with the Lui and for the wuber um yeah I mean by the end of the day the service is","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3290s"},{"start_seconds":3297,"text":"different from the the the the the the way you interact with gbt or entropic or this models right because the service is","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3297s"},{"start_seconds":3304,"text":"more of experience and would you really be able to experience this this good you know agent execution um workflow for you","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3304s"},{"start_seconds":3314,"text":"so I would say that uh I think the accuracy in this type of the scenario is not really important but what's more important is if the model is triggering","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3314s"},{"start_seconds":3323,"text":"the correct well thing for you or if you can actually even provide a user feedback loop to that and tell you model what do you want so then like it can be","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3323s"},{"start_seconds":3331,"text":"more and more personalized yeah I think here's the analogy is","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3331s"},{"start_seconds":3337,"text":"like that interesting to compare something like Instagram right where it actually creates demand instead of just","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3337s"},{"start_seconds":3345,"text":"um predicting what you want and so and over time you know potentially get you excited about doing something you know","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3345s"},{"start_seconds":3354,"text":"Uber maybe not the right example but broad speaking door Dash would be a good example a door Dash yeah like oh you know you saw something that would be","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3354s"},{"start_seconds":3362,"text":"exciting for you and you actually will order it uh I mean there's still you know there's still metrics around like","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3362s"},{"start_seconds":3370,"text":"click St Etc but yeah I think I think there's an interesting opportunity of like what is experience","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3370s"},{"start_seconds":3377,"text":"Beyond where you're balancing and that and this is where like for me an important question like how do we","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3377s"},{"start_seconds":3384,"text":"balance Beyond kind of optimizing for profit of the company that presents you this options","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3384s"},{"start_seconds":3391,"text":"to actually optimizing for individuals wellbeing right uh in kind of broader sense uh because kind of Instagram right","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3391s"},{"start_seconds":3400,"text":"now either way it is it is optimized for you to spend more time on Instagram but that's not because they are care about","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3400s"},{"start_seconds":3407,"text":"you but more because it generates some more money and so how do we kind of transition here to more uh","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3407s"},{"start_seconds":3416,"text":"more like hey this is actually good for you you know you're learning more you you know eating healthier food for example Etc uh through his predictions","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3416s"},{"start_seconds":3424,"text":"as well that's that's a good question but I want to throw a a question for for you as well do you would you like to","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3424s"},{"start_seconds":3433,"text":"have like a like a mom or dad or or teacher in your AI or in your you know","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3433s"},{"start_seconds":3439,"text":"physical device that's alert you saying hey you know eat healthy um no no snacks um so even though you've been browsing","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3439s"},{"start_seconds":3448,"text":"through McDonald's page for like five minutes and then the recommendation is still I check this pokeo U or check this","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3448s"},{"start_seconds":3457,"text":"salad I think that's um yeah that's something that I I think the AI um should be balancing or the application","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3457s"},{"start_seconds":3465,"text":"should um have a mechanism or a Twitter line where where that um they don't","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3465s"},{"start_seconds":3471,"text":"encourage the the evil um um human needs U more but they they don't stop it I","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3471s"},{"start_seconds":3479,"text":"think that's that's a clear line for us to see um but that's that's not um that's not the pfl to say um we we has","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3479s"},{"start_seconds":3487,"text":"to be honest uh and give back the um the data to to the model right um but yeah","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3487s"},{"start_seconds":3495,"text":"it's a it's an interesting question so is AI supporting you and encouraging you uh and you know just fulfilling you uh","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3495s"},{"start_seconds":3502,"text":"on whatever you already are or it's has to be acting you know uh more of a Godfather or God mom um to guide you","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3502s"},{"start_seconds":3511,"text":"into the right way and what is the right way right um but then uh again uh another interesting um interesting spark","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3511s"},{"start_seconds":3518,"text":"uh idea that you just mentioned uh what is the you know what is the what is the incentive that we could potentially um","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3518s"},{"start_seconds":3526,"text":"embedded within this uh incentive infrastructures um to give users a little bit more um a little bit more uh","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3526s"},{"start_seconds":3534,"text":"feedback or constant uh for them for us to engage with them or for us to use guiltily guiltily use AI agents to","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3534s"},{"start_seconds":3543,"text":"promote anything um from from you know the advertisement perspective whereas you know Instagram is obviously on top","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3543s"},{"start_seconds":3552,"text":"of that right um they don't give back uh any users any incentive but well Tik Tok might be a you know even worst case uh","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3552s"},{"start_seconds":3560,"text":"case study right here that uh they use a great algorithm to know you better and you know track you or trap you into into","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3560s"},{"start_seconds":3567,"text":"their application spending more time but for us like um you know the the applications that we see or that we hope","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3567s"},{"start_seconds":3574,"text":"to see in the ecosystem that utilizing pfl you have tremendous like private um datas right um how can you use it better","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3574s"},{"start_seconds":3583,"text":"um can you potentially also give back the uh individual users a little bit more incentive based on how much","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3583s"},{"start_seconds":3591,"text":"advertisements they have already seen or how many like um things that you help you to generate it uh your revenue and","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3591s"},{"start_seconds":3599,"text":"you gave us more portion back to the community I think I I think that's really important um um maybe the","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3599s"},{"start_seconds":3606,"text":"developers should design the mechanism within the D app or maybe the the infrastructures of ours we should have","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3606s"},{"start_seconds":3613,"text":"another layers of incentive in incentive mechanism to distribute it that proportion out um it's open questions","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3613s"},{"start_seconds":3623,"text":"for not only us but also for the dev com yeah yeah I mean I think there's there's","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3623s"},{"start_seconds":3631,"text":"a lot to figure out and research here so but uh thanks for coming we we are out of time so exciting to uh to hear kind","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3631s"},{"start_seconds":3639,"text":"of what you guys are building and uh yeah let's let's keep building uh thanks everyone for tuning in and see you next","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3639s"},{"start_seconds":3647,"text":"week thank you","url":"https://www.youtube.com/watch?v=ba2ixv44eb8&t=3647s"}],"youtube_url":"https://youtube.com/watch?v=ba2ixv44eb8","url":"https://near-pedia.xyz/videos/ba2ixv44eb8"}