Block by Block: A Show on Web3 Growth Marketing
Each week, I sit down with the innovators and builders shaping the future of crypto and web3.
Growth isn’t a sprint; it’s a process—built gradually, step by step, block by block.
Let’s build something incredible, together. All onchain.
Block by Block: A Show on Web3 Growth Marketing
Jihao Sun--Flock.io Decentralized Federated Learning for Blockchain AI
Summary
In this conversation, Jihao Sun, co-founder of Flock, shares his journey from a background in computer science and AI to the creation of Flock, a platform that merges AI and blockchain technology. He discusses the importance of data in AI development, the challenges of building a two-sided market for data providers and engineers, and the strategies for fostering developer adoption. Sun emphasizes the significance of decentralization in the future of AI agents, aiming for a system where agents evolve independently through community input. In this conversation, Jihao Sun from Flock.io discusses the innovative approaches Flock is taking in the AI and blockchain space. He highlights the importance of user experience for validators, the collaboration with data providers, and the utility of the Flock token in governance. The conversation also covers the model store's role in the ecosystem, partnerships with firms like GSR, and the emphasis on privacy through federated learning. Sun shares insights on growth metrics and future aspirations for Flock, including making AI training accessible to a broader audience.
Takeaways
— Jihao Sun has a rich background in AI and finance.
— Flock is a project that merges AI with blockchain technology.
— Data is crucial for effective AI development.
— Flock aims to give users control over their data.
— The platform has a two-sided market for data providers and engineers.
— Decentralization is key to the future of AI agents.
— Flock's Testnet was launched last year, now on Mainnet.
— The platform encourages community participation in AI training.
— Flock addresses data silo issues in traditional industries.
— The goal is to create AI agents that evolve independently. Flock has engaged a significant number of validators, enhancing user experience.
— Data quality is crucial, and Flock collaborates with leading data layer companies.
— The Flock token serves as a POS mechanism for governance and training.
— Flock's model store allows users to launch and monetize their AI models.
— Partnerships with firms like GSR focus on privacy-preserving AI training.
— Federated learning ensures data privacy by keeping data local during training.
— Accessibility is key; Flock aims to lower barriers for AI training.
— Future growth metrics will focus on onboarding more business clients.
— Flock is rebranding its model store to enhance user experience.
— The vision for 2025 includes advancements in AI agents and hardware support.
Timeline
(00:00) Journey into AI and Blockchain
(02:52) The Birth of Flock: Merging AI and Blockchain
(05:58) Explaining Flock: From Mom to Target Customers
(08:48) Building a Two-Sided Market: Data Providers and Engineers
(11:56) Challenges and Strategies in Developer Adoption
(14:59) Decentralization and the Future of AI Agents
(22:09) Validator Engagement and User Experience
(24:05) Data Quality and Collaboration in AI
(25:53) Flock Token Utility and Governance
(29:46) Model Store and Ecosystem Integration
(31:52) Partnerships and Privacy in Trading
(36:52) Growth Metrics and Future Aspirations
Follow me @shmula on X for upcoming episodes and to get in touch with me.
Jihao Sun, co-founder of Flock. Welcome. Hey, thanks for having me. Now, before we get into flock.io, we'd love to hear about your background. You have a very rich background in finance and also specifically artificial intelligence and finance. Tell us how you got into crypto and why you stay. Right, so a little bit about my background. studied computer science back in old days. And then that was a time when deep learning just suddenly became a big topic. And yeah, so right after school, actually, the first full-time job I did is to become a CEO. That's an interesting time of the AI development time. And then gradually, I also... got a fun exit and then joined RBC as their director of AI for quite good years of corporate life as well. then till in 2022, that's where we started Flock. yeah, so backing all my experience from study to business, like it's all about AI. I actually, Flock's my like a first ever project, you know, working. around blockchain technology. But I got exposed to blockchain, I think, in 2017, the ICO summer. That was the summer where everyone goes so crazy about this, this, this things, right? As as a, like retail public, I looked at this crypto industry and then I did also invest some, some tokens who are now like nowhere. Yeah. But that was a very fun experience. was a period of time where everyone was looking at different ICO proposals and then putting their ethereum just down to a contract and waiting for some emission in the future and things like that. And that was quite a wild summer, I would say. of course, because of my background and everything, didn't or I don't really feel like at that time blockchain is something that matching what I'm doing, because at that time, AI is still a track in the traditional industry, especially when that deep learning summer for AI track, right? That's also a jumping industry, and it's getting bigger and bigger. Even today, it's getting bigger and bigger. It was only until 2022 where I actually see the potential of having blockchain in technical term. that can be merged with AI to build this what we call federated learning plus blockchain mechanism to prevent data from being exposed to any third party entities but can still train a model across different devices. So that's where I say, yeah, that actually matched with my expertise and I want to build this because it's like two very much, to me, two very much exciting things, know, merge into one and it's exactly what I want to build. That's exciting. Now you spent quite a bit of time as the Director of Artificial Intelligence at RBC Wealth. What are some things that you take with you that you learn there that you're now applying as an entrepreneur in your own startup? Well, I'll say mostly about the necessity that I can see for data to be a very important part in the AI development, right? You see that many of those companies like like Nebelberg, or Finitiv, they all are establishing their business quite successfully around data. And then you will also see that data silo issue happens between different entities in this banking world, like everyone have their own small pool of data, they never, they don't want to, and they cannot actually share with anyone else, even between different desks, there are walls between different desks, say, this is my trading secret that you can't share with another desk, even within a same entity, right? So that blocks lots of developments for AI modals, because modals are always, or can only be good when the data... are accumulated together. Because otherwise, can only just have a biased data set, let's say, just on the open domain data. It cannot be fine-tuned or well-trained to your own specific business needs. So that's where one of the necessities I feel like, there must be a solution that can actually tackle it. And yeah, for luck's sake, one of it. Yeah. And of course, also, because of the experience I had in industry, especially in both academia and industry, I knew lots of lots of good engineers and researchers from this space. So when we actually first formed up a team to build FLOG, I actually bring all my good old colleagues over to the team. basically, all the small, large companies I stayed before and then I bring my best lead engineer over all two flocks to build this new team, engineering team. So this is also another quite a good perks in coming from a rich experience so that you know who's the best in the industry. And then you can just call them up together for a new journey. And that's one of the advantages of having been in the space for a while, at least in artificial intelligence that you know, you've worked with or know the big names and you just brought them with you to help with your startup. Now, let's get into Flock. And I would love your help to, if you could explain Flock in two ways. Pretend like you're talking to your mother. How would you explain Flock to your mom? And then, how do you explain Flock to the target customer that you are targeting as a user of Flock? Right, right. I used to attend a podcast, which is a very fun one, that they set up scenarios for you to pitch your stuff. There was one talking about, so you're now, you thought you were going to a consensus or token 24 and nine event, But it up you realize you were standing at the stage of a mom, new mother type of conference, right? Would you spend the next 30 seconds to just explain what's flock? Right? Yeah, yeah, yeah. So, yeah, I used a very good example there. like I was, I can think like for all the, you know, ready to be new moms, right? They are, they all care about their new kids, right? Sorry. And the newborn kids, and then they want to take care of them. And there are new technologies, like a baby camera, keeping at home where you can monitor your baby. So. But then always on such kind of technologies, have a risk or you're worried about that data that can be just exposed to the third party. And then someone might just monitoring your home and then they probably want to do something dangerous to your home. So what Flux is trying to build is a way that can actually help you improve your camera by... course, learning all the things happening at home and give you notifications where baby safe or not. And still not to send any data from your local environment, like your home to any third party. So there's no way that a second person can actually eye on your, your, your home. So that's kind of like the explanation, let's say to a mom or flocks building flocks are trying to, yeah, build a, a building AI or build an AI training framework, basically a a way to train AI, right? Without needing to take your data from your home. So, so to the, yeah, to our, yeah, basically to our targeted users, it's more about giving them the statement about how important it is to have your own governance of your own data. And then you are earning from your data, you're earning from your model. Like for our training platform, which Sorry. Yeah, for our training platform, which we launched Testnet last year. And then of course, now we're on ManNet. So we have about several thousand trainers building actively on such training platforms where they actually contribute either their techniques about how to train a model or their, let's say, data to how to accomplish this task with the model we are proposing. or just validate the model to make sure this model is growing towards whatever I think is for the public good or for the general good standard instead of being biased of some certain ideologies, let's say. So we want to make sure that everyone could participate into this process that will get their fair share of the incentives by creating a model that can be used in industry. So we have several models like being used, say in traditional Web2 hospitals, use cases like monitoring your glucose level to give you predictions about that for diabetic patients groups, right? And also with the eye hospital in London as well for doing the eye disease prediction detection. Hmm. those are the heavy cases in traditional industry, but also we have, you know, fun cases in web three as well. Like, like, like with, it was more fails before that we train a language model that can actually understand natural language commands between different, breach and swaps between different chains and then how that can work within their, within their, their, application where people don't have to worry about that their financial data. so eventually the whole crypto world is all about some financial behaviors and people trading and people having their own records of everything. And if you want to have a AI companion or agent that can help you do all the financial processing, you probably have to submit all those data into one place to fine tune it, but then you don't want to. You don't want to tell everyone, like, this is my wallet, that's my wallet. cases people wanted to hide somehow behind their wallet. So that's one of the use cases we had with the Asian world. also, mean, probably your audience also heard about the agent that we built together with other projects like Anymoka brands, right? So to drive the model behind the agent so the agent can perform in different functionalities. For example, they can read your pitch deck and then to tell you whether this is a good project to invest in or not. To read all the streaming news every day and then to make these decisions or actually suggestions in terms of some of the main investments. yeah, such as such use cases in Pure Web 3, environment is also very popular with Floss. No, that sounds really exciting. You shared a use case that I'm familiar with. A long time ago, I worked for a medical device company that created a special camera that took seven fields of the eye, seven pictures of the eye, different fields. And then those pictures were encrypted and then sent to a university where they had eye readers. where doctors would look at the images and then they would give a score on the level of disease for diabetic retinopathy. But this was like 15, 20 years ago. But now, now AI does that and it's pretty amazing. And will does it more accurately too. Tell us about, now it feels, as a, data training platform, I think that's how you're positioning Flock. It feels like there's a two-sided market, right? There's people that provide data. And then on the other side are machine learning software engineers that create some kind of model to train on that data or something like that. Is that correct? Tell us how you're building both sides of the market. and in some maybe an interesting story of kind of maybe something a conclusion that was learned that was maybe surprising. Yeah, I think two developer building is actually very hard in Web3. But thanks for, again, my rich experience in industry before. There's a platform called Kaggle. It's very popular in AI space or data science space before where I'm also in one. mean, one of my financial data sets is the top, ranked top in the financial data set. I'm kind of like the OG there. And then I know also a group of OGs there. Mm-hmm. who are actually what they called cargo masters, like those kind of cargo grandmasters, like those engineers who actually participate into lots of bounties or tasks in cargo and in a win lots of times. So they will have this tiered reputation on their account. And I know many of them. And that's our actually the breaking point to bring the first wave of developer adoption into flock. Like deep down, I understand why people like Kaggle, right? And also deep down, understand why people don't like it because, know, you know, borrowing the, quoting the book Chris Dickson wrote about read, write, own, right? For Kaggle, it's more of like in the read stage where, yes, it's like a bounty platform where people put down bounty and then a group of engineers will just... build together and then to compete together with their own techniques. then yeah, whoever won the thing and won all right. But then the model is not yours and everything's not yours. You're just competing and then that's it. And that's like a one-off business. So there was like a new platform then came up as Huggingface where you can actually host your model there. That's where I see as in a read and write stage. yes, you are still having your model. You're still building your model, but then you can host it there in Huggingface. But end of day, it's just a hosting platform in that sense where you don't actually own your model. So that's where we think about a flock to be the stage where you have really have right and you have the ownership of the model that for anything in the long tail where your model is being used by this agent or that company, then the rev shares always within the transparent audits on blockchain. So that's where we see the difference in growth of it. back to your question. Yes. So the breaking point of developer adoption is, well, mainly because of the experience personally, or my team had before, right? They break in from the most, I would say, most vibrant team of the traditional Web2AI engineers. And then we quickly had the first batch of several thousand of them joining us. And it sounds small in Web3, because in Web3, every project has like several. several hundred thousand users where I don't know where they come from but for ours it's like a very very what we'll call legit users because they all need to have a graphic card like like a 4090 or even higher with them in order to participate into the training because otherwise it doesn't work right you can't just run on your phone or on your laptop you have to have a dedicated machine to plug in and play so yeah that's where That's where we had our first break of this. then after that, because we are doing this crypto business, we understand that adoption from the community is also important, right? Then that goes to the second question. How do we face the general public? Yes, of course, like in the beginning, we thought if we just do developer tools well, that can be a great traction already. But then when you face... you know, lots of the pressure from the market and everything, you realize, okay, so the general public adoption is also a very key thing that you need to build up, because otherwise that wouldn't create enough balance between the two sides, right? So that's where we started up all the campaigns about the delegation campaigns. So meaning that general users, they don't have to understand everything about Flock. or AI or machine learning, they can just look at the leaderboard of all the trainers and then they can delegate their token over to such different trainers that they can share their revenue between them and the trainer. So that can greatly give also incentive for the trainers to perform better because they know that, okay, if I do well, then there'll be more people dedicating to me and then my return will be way higher compared to others. And also for the general users, yeah, it's easy to just plug and play. I'm just kind of like just supporting some of the best AI engineers in the world to competing in different tasks, were those tasks, some of them landed into the traditional industry to return recursive or retainer revenues, or some of them just like launching a meme, right? They will have rev share about those meme tokens. Also good, like there's no harm to have more of the adoption use cases for any tasks trained and ran on Flutter. Yeah, as you're describing Flock, it reminds me of another project that I met with recently, CrunchDow. Is CrunchDow a, are you familiar with CrunchDow? It sounds like a similar dynamic with machine learning scientists on one side and then data providers on the other. And then, They have not had a token generation event yet, so there's no token. one of the mechanics is that they're able to stake their token to specific developers, that developers that can produce, and they get part of the reward when they produce good outcomes, something like that. It sounds like Flock is similar. Um, yeah, maybe, maybe I'm not quite familiar with them, but then yeah, the mechanism, yeah, the mechanism behind, you know, there's some level of similarity maybe, yeah. How many machine learning data, how many data scientists are on the platform now? We currently is about 2000 and yeah, yeah, a good number of it. yeah, yeah, I think like in the whole world, maybe that's about several thousand people who actually have that time, talent and hardware to actually participate into, into such tasks, right? And then it's kind of like crowdsourcing mechanism to, to, to, get the best model out of from the community. So that you don't need to worry about getting the model is being biased or directed by one single authority. It's like, yeah, nice. Like there's tons of them that building it, right? And then when a model is built, then it can be hosted in a decentralized way. Like you can work with decentralized hosting services like a cache hyperbolic, then these guys, right? So to have the whole process and very much decentralized from. building the model, hosting the model, and eventually, if there's an agent behind it, you can even launch the agent in a decentralized way. that's where I like it, right? Data was sourced decentralized. Model was built decentralized. And the agent was hosted and launched decentralized. That's a very important pathway, I would say, from creation of the model, from ground up creation of the model to, sorry, yeah, to launch of the agent. So that's why we always talk about this like something we will call AI agent 2.0, right? We want an agent to evolve into next summer where yeah, there's totally no one behind it. No one is directing it. We want the agent itself to evolve themselves. And then it's just by the public, by the community. That's exciting. So you have 2,000 engineers on one side. Tell us about the data provider. Is that what you call them, data providers? Or? are... Yeah, Yeah, yeah. Or validators or delegators, yes. And how many are there on that side right now? Current number or minute, let me just quickly check. Yeah, right. So actually from our testnet phase, I remember a number that's like 66,000 already. So it's like a good number of them. then yeah, they've been training, participating into training, participating into validation, right? You know, very, very heavily. Yeah, that's a good crowd of people. And the good thing is like, we keep everything very, very streamlined with the UI that we provide. So you don't need to understand too much about interaction behind. But you can, if you want, use your own command lines to interact with a contract directly. also, we also have very user-friendly ways to use Fernand to interact. And also, for our engineers, we made sure that the entry buyer is as low as possible. The only difference maybe, I always have this comparison, right? I want our engineers or developers from the AI industry to join Flock. And the only difference that they can feel is they just need to provide a wallet address and rest of it should be the same, right? Just like a PyTorch library. But then the only difference is whenever you want to run that model, you have to stake with that wallet, but then that's it, right? That's what I think is the best user experience from the developer side and plus what I mentioned about the user side. I see. Now, in terms of data quality, a lot of the data, so if I'm a data provider, the number one question on my mind is, I need to know what to do with this data. What questions does it help me answer? And sometimes I don't know. And so I'm gonna need help labeling the data. How does that step factor into the process? We actually don't have this whole data layer services. So data layer is not one part of the services that we provide. So we are actually very happy to work with the leading companies in the data layer side, The names like Vana, Masa, Sahara. These are the players in the space already where we would plug in with many of them. And then, yeah, some of their users will have their specific data. being curated and then well, well, formatted, let's say, right, for us to make trainings on. So yeah, so that's like, yeah, I believe like in this industry, like every chain of this creation that needs to be a professional player, instead of someone who trying to build everything from front to end, right? So for us, like we specialized in model training, where it's actually the hardest part of the whole chain, I would say. because you have people who provide data, you people who provide compute, you need to have people who provide models, and all the three will form a triangle that can be packed into, an agent or an application that can actually work. So these are the four different fronts of the whole AI crypto space. Now you've had your own main net. You've had a token generation event. Flock is the name of the token. How does the Flock token work in the Flock ecosystem? Right, so it's quite utilised and then we have this, so, right, the story back that we are a bunch of researchers, engineers, right? actually for the first year of us running together at Flock, we focus mostly on writing our papers. So writing our papers to get in a peer review and then be published into the top conferences and journals. Mm-hmm. That's where in the first version of FLOG, it got all in the best paper award and everything into the top journals. And that's where we actually have this confidence. Okay, this is some mechanism that the whole academia is actually approved, right? It's like peer reviewed and it's something that's robust enough. And then we started to build the product within the 2023, the whole year. So when it comes to... The token usage, a token is actually used as a POS mechanism. So people has to stick in to train. The reasoning is if you have a system that is open, it's always hard to govern everything without a super node in the middle, right? Just like the traditional, in traditional world, there are further learning already. Like there's Google who proposed it in 2017. And then everyone who's now using Google Pixel phones that you are typing habits is being predicted by the federal learning mechanism by Google. But then the thing is Google is the one in the middle. So they can actually see everything. Yes, it's kind of decentralized training that you can't see other users' data in a federated learning setting, right? But Google can see it. Or there's no way Google, there's no way that you can prevent Google not see it, right? So that's where we want blockchain to step in. We don't need Google. We don't need that super node in the middle. We have this. own chain consensus that can actually make the governance between all the users instead of a super node. So this group of users, then when you have the system open to this group of users, you need to worry about, okay, there will always then be some scammers or the pointless nodes trying to join and hack it, right? So how do you prevent your consensus from such attackers? So that's where the POS makes the effect. So in every node, every phone, as my example, every phone had to stake in certain amount of flock to join the training. And if you are not performing well, or if you are trying to actually point the whole model, then your token rewards will be slashed versus other honest people, other honest nodes that will get their incentives. So that's a way where we actually to eliminate such malicious behaviors within the network. So that's the utility of the token. But then also, now since the token is kind of like well spread between different users and also some traders, right? So many of them are also using it, as I mentioned, as a delegation intermediate to delegate to those trainers so that they can actually earn higher APRs based on the training nodes that they chose to delegate to. I see. Let's go into the, I think one really unique thing that as I did my research on Flock is you have a model store. Tell us about the model store and how people on the Flock ecosystem can use this. Yeah, we are actually rebranding this model store very soon, but in the beginning, it's just like something like OpenAI provided as OpenAI model store, right? So you have your model, you train your fine tune or maybe just prompt engineered it and you think, this model at this stage is the best that can maybe answer your question in a more, for example, waifu tone that you think is the best than any other waifu in the world. And then you pack it up as an app. and just launched in Flock model space, right? Or in open-air model space, same case with Flock. So you train a model, you compete with each other, and then you won, and you feel like, is the best model for this task. I need to, you know, even more utilize on the model so you can launch the own Flock model space and people then can plug and play and even pay for the... for the usage of such models, right? And because, you know, now the new agent summer is kind of like super hot. So we will see lots of people not only just launch their model there, but also having their agents just alongside such models. So we are rebranding the marketplace to include such agent launch as well, just to make the whole experience very seamlessly with all the models that built on OnFlux. It's quite exciting. I'm looking at the scroll GPT right now. And this is scroll the layer two, is that correct? Yeah. Okay. No, this is really cool. Yeah, I'm curious to see how this grows as you create more models or participants on the flock ecosystem create these models, how they can be used and how you'll market it. No, it's very interesting. You announced recently a partnership with GSR. GSR is a well-known trading firm, market maker in the crypto ecosystem. Tell us, I guess, the work that Flock is doing with GSR. Yeah, that's exactly the scenario I mentioned back in the days when I was in banks, right? So like the data in a trading firm cannot be shared at all, even between different computers, because they are so sensitive and then you are facing different order book. He is facing another different order book that you shouldn't mess any of those macro up at all. Right. But always such trading firms, they want to have some guidance for trading or some strategy for trading like a firm wise. So that's where that we're actually working with DSR to train a model in a privacy preserving setting within their own internal desks to train the model that don't really expose any sensitive data that's not supposed to be shared between different teams. Yeah, like in that case about data silo issue I mentioned about in financial industry where it's just you just cannot share it. But if you want to have a bigger model or great model within your own firm, right, you have to have enough data to train that model. And then that's like a dilemma. this is unsolvable before FLOC. So now, yeah, there's a solution. then yeah, we're training this with GSR together and GSR gonna onboard many of such strategies and trading agents within their own networks. in some sense, can think of like, yeah, Flock is training the model that's been used by GSR where GSR is a big market maker for many of the top tokens in the world. So kind of like Flock is empowering the whole Web3 world in a way, Yeah. And I'm sure from GSR's perspective, they're looking for alpha and they're looking to flock to help them generate that alpha. Tell us about the privacy preserving way. You mentioned that. Are you using trusted execution environments or how are you ensuring privacy guarantees for the data? Right, for us, it's more about, for us, as we have our own training frameworks, that's federated learning based. By definition, federated learning is a way where you don't have to send data anywhere else. It's more about how you decompose a machine learning training process in a way that you put data, you put model into local environments, instead of you submit data up to a centralized server for model training. So it's like a reverse the way of training. So that's by definition. And in that definition, your data is already safe. That's already been secured. But then recently, not recently, sorry, like last year, we got, we are one, we are actually the only AI infra project being, granted by Eastern Foundation, right? Where Eastern Foundation is very focused on zero knowledge proofs. And we actually co-auth the paper called the DKF flock, meaning that using still this flock mechanism, which I just, a reversed way of training, but adding ZK on EVM, adding ZK as the security layer over the transaction. So A, you already have a very robust training mechanism to ensure your data is local. B, in this mechanism, all the transmissions between different nodes are still secured in a ZK way. So that's like a double lock for the privacy. So for us, like we are quite open for anything that can be compatible with flock mechanism. So for example, TE or even FHE one day when actually become two, right? So that we're happy to explore the collaborations and the possibilities to run such in practice. Yeah, it's almost like a flavor of multi-party computation where rather than the data going up to some cloud service, it's done locally is what it sounds like. And by definition, because it's local, others can't view it. so privacy is preserved. That's very unique. And now I see how the name Flock, I see where that comes from. It's now because exactly because of that. That's really cool. Thank you. Thank you. How do you think about growth as is now Flock is, you know, has a, sounds like a really healthy, you've got, you know, data scientists on one side, you've got data providers on the other side, collaborators. It sounds like you've, it's growing quite healthy in multiple ways. And also training submissions. How do you think about growth and some of the metrics that you follow? and track. Yeah, I think so far that the biggest metrics that we're following is to expand our business developments to onboard more business clients that's actually requesting models to be built. Right. So like I mentioned, GSR and Mocha brands and many others, we want many of such clients to be pipelined into the training platform to populate the platform with many choices for trainers. Cause you don't, you don't want to just, you know, limit people to look at the one task only, right? You want them to have a good window, a window to shop, whatever tasks that they will be interested in. Maybe they are a football fan that they want to model that they want to train or they are more passionate about training a model that can do, I don't know, game betting or whatever, right? Versus there are people who are more on the comic world, who are more interesting, waifu type of agent companions. So, yeah, so that's one of the biggest metrics that we're focusing right now. And also, as I mentioned to you, right? So since now that we have so many training nodes and validators in our network, yeah, we still want to expand it even further, that we still want to collaborate. with many of such traditional, let's say AI platforms or AI engineering groups to expand our adoption from the traditional world. I think for a web-stream project or actually for a successful web-stream project to sustain long-term, you need to make sure there's A, the business needs, right? That's actually really bringing the revenue stream and also the request to make the business case even more stronger and B, you want to make sure that adoption from the web 2 world, that's source of the new commerce into your protocol, into your network and AI engineers, that's actually our play. Now looking at 2025, what are some things you're looking forward to that you can announce to the community? All right, AI agents 2.0, right? Some support for that. So we've seen, I'm not sure if that summer has passed yet, but we've seen the AI agent summer, right? Being super crazy by end of last year. And then, but still, know, majority of such agents are still in their naive period of existence, let's say, right? I'm hoping in 2025, there will be very interesting new upgrade, you know, however they define 2.0, how... 2.0 it should be, right? But there'll be a new level or new stage of agents coming up in 2025 and then hopefully we will have a good proportion of such agents in a trend on flock. So yeah, that's of course one of the very important and interesting things. there's one more thing that we announced earlier about the support of the hardware that we actually now have. Apple's chip hardware support so that to bring the training and validation barrier all the way down from very expensive NVIDIA cards all the way down to Apple devices. And then that's where another thing that I think is very important for the adoption in Web3 or Web2 is that you want to make sure that the accessibility of your platform is as low as possible, right? But of course for AI industries, It's a bit hard. If you say that today that your phone can train a model with 30 billion parameters, it's a bit of joking, right? But then I think the whole industry is trying. They're trying and they're using lots of techniques to making this happen. Like what we did, like we do model sharding, we do model split to make those training and layers of the large-scale model as small as possible that can fit into. an Apple device and that's already one of the interesting techniques that we've developed and actually contributed to the community. It's open source. It's to the community already, right? And that's one of the ways that we are helping the industry to make this barrier lower and lower and to make this, you know, this is also one of the important things because you don't, if you want, if you are thinking about the mass adoption, you don't want to limit users to only come with their high power hardware barriers, right? You wanna make sure that everyone, even with the phone one day, with the Raspberry Pi one day, that they can actually join your training network. And I think that's happening, because think of today, know, the morsel of all of this hardware developments, right? In two years time, maybe your phone can just run as good as, I know, a 49 over today. maybe that's too much, maybe 30, 90 of several years ago, right? So, and then we can see that number going crazy up. And then, several years from now, I think we can see such as edge computing or local device hosted network of decentralized training to be very, efficient. Well, that sounds like a great place to end. And congratulations on all the success that Flock has had. And best of luck in 2025 as you continue to grow Flock and its community. Thank you so much, Jihao Sun from flock.io. Thank you. you. for having me. Cheers.
