Technology, Connected - What Is Decentralized AI?

Episode Date: April 7, 2025

Jihao Sun, co-founder of Flock.io, joins Mark and Jeremy to Think on Paper about decentralized AI. LLMs that trains locally, protect privacy, and reward users. They get into data ownership, blockchain-based training, and whether federated learning can really compete with centralized giants.If your data is building the next generation of AI, who’s it building it for? And why aren’t you in control?Please enjoy the show. And share with a curious friend. --CHAPTERS(00:08) Introduction to Decentralized AI(01:33) The Importance of Decentralization in AI(04:40) Accessibility and Cost of AI Technology(07:53) Understanding Federated Learning(11:38) Challenges of Centralized AI Systems(14:30) Exploring Federated Learning in Banking(17:10) Blockchain's Role in Decentralized AI(20:45) Democratizing AI Development with Flock(26:49) Real-World Use Cases of Flock(32:56) The Evolution of Parenting and Technology(35:28) Decentralization vs Centralization in AI(36:31) Future of Decentralized AI Models(37:29) Understanding AI Hallucinations(41:23) The Role of Synthetic Data(44:59) Getting Involved with Decentralized AI(45:53) Staying Out of Your Own Way(47:03) Final Thoughts on Decentralization and AI--Learn more:www.flock.iowww.thinkingonpaper.xyz

Disruptors and curious minds, welcome to another episode of Thinking on Paper. Today's big question, how do we build AI that is decentralized, privacy preserving, and fair without sacrificing performance? Imagine a world where anybody, not just big tech, can train AI models, control their own data, and get rewarded for their contributions. That is what flock IO is building. Mark, talk about our guests, and kick off the episode. sir. Yeah, thank you, Jamie. It's going to be a fun episode today. We've got all the hot

buttons coming later. We've got our backstage breakdown of the show. And I don't think that the world has ever needed flock at IOs as it does today. It's the 27th of March 2025, and the world is reeling from the announcement from Open AI on Chat GPD and their image production. And I've just spent far too long looking at everybody in the world posting pictures of their face. I'm all fortized into a Ghibli animation production. And I think it just highlights where we are with the power that these centralized AI companies wield. Jihal Sun is the founder and CEO of Flock. And they are hopefully going to solve some of these AI conundrums for us and create a decentralized AI that we can all use as cheap, affordable, available.

And mine, my stuff is mine. Jih, welcome to the show. Welcome to the show. Hey, hey, hey, thanks for having me. Hey, Jeremy Mark. Thank you for being here. Thank you for thinking on paper with us. So we're talking about this intersection of AI, of blockchain, of decentralization. Let's start from a high level. Why does any of this matter?

Yeah, just go back to the example. You just mentioned about recently. Everyone's so hyped about the JPLI version of their face and posting on Twitter, right? But then have you ever thought about the thing behind this is you're actually sending your personal photo over to an API and sitting somewhere in S. for somewhere, right? And then they render the photo out and then send back to you. During the process already, of course, you have a beautiful photo, but then you've already sacrificed to your own personal photos. If that's sensitive in some ways, then yeah, that'll be risky. That's where initially, you know, why people were thinking about, oh, can we make models

or can we make AI training more decentralized? And we don't have to connect to a server that I don't even know where it is. And I can put everything, put all my data securely, locally, in my own environment. That's where the start of the conversation, the start of the thinking comes through. Yeah, and Flok is of the earliest actually in the world to doing this. I was so angry about everybody using one of my favorite movie producers, writers, creators, who creates some of the most beautiful worlds that I've ever known in my life and just using it for this frivolous nonsense. But I hadn't actually thought of them sending their face via an API to some data center somewhere. My brain hadn't even worked.

worked to that level next. And I don't think that many people using it even think about that. Well, here's, here's, here's the thing too. Mark and I talk about this all the time. It's, it's use Waze as the example. So you're, you know, I can get from A to B very simply. I can avoid traffic. I can do all this stuff. But guess what? Waze knows exactly where I am at every single point of the day. But I'm willing to give up that security for this convenience for this ability to make my face into a Ghibli. Is the world ever going to be ever going to understand that piece of the puzzle? Yeah. That's a cool part of it.

Yeah, at least from the representation of the final result it is, right? So that's also a cool part of AI. You don't actually know what AI is actually thinking because they might not think in the same way as we do. But as long as the output of the production is what you need, then we can assume, okay, they understand it. Okay, because we are essentially thinking, just like thinking on paper, right? We're actually thinking on behalf of the totally, a totally different, you know, silicon-based creature, which is totally different from us, carbon-based. So yeah, maybe there's something that we can never actually understand.

Silicon-based creatures and carbon-based creatures. We're actually reading a wonderful book called Irreducible. In our book club right now, we're two chapters in, and we're actually getting to the bottom of that. So stay tuned, listeners, listeners for that. So one last piece of why this matters is an accessibility to the technology. So AI is pretty expensive. DeepSeek brought it down significantly, you know, over the last few months or weeks. But what you guys are doing provides a bit of accessibility to building models too.

Correct. Talk us through that a little bit. We actually just announced and released the version that can actually run on a MacBook M-CHIP. So it's like the reason why I like AI, actually I'm so much in love with the whole progress recently years about AI and technology development is something that you thought doesn't really make sense maybe two years ago. It actually becomes very doable maybe in a year or two time. Several years ago, when we started to do flock, right, we thought maybe we can only just run fine tuning locally.

By fine tuning, we mean that, okay, we don't actually train the whole model from scratch. We just do a bit of inferencing and the final several layers. And then that can happen still, not on your personal device, but on a decentralized data center. That's maybe to the most can be the most efficient way of decentralized training. Then, you know, by time we just realize the M-chips are actually super powerful now. that first of all. And secondly, the algorithms, the advancement of the research that we do and also the whole industry do. There are tons of ways that people are trying to tackle the problem of we have to have an expensive GPU. So people are trying to compress the model, basically making

the model smaller, and compress the gradient, basically meaning that making the communication between models smaller. And even like optimizing the whole training process. So he can actually host the whole training in a smaller devices and today run some chips maybe in the future we can imagine something like your phone or a Raspberry Pi can actually run not only one but a cluster of them right can run a large-range model and similarly for the dips deep deep case as well that's also very very inspiring as well like like nobody have ever thought that any any you know let's say second or third tier company nobody heard about before that can use a very limited resources of the the GPUs that they have to actually have a similar quality model, right, compatible to the

open AI ones. Yeah, I mean, before Dipsic, I thought it's a, it's hard to, it's basically impossible to happen. But now like, yeah, it's it happened. And actually, uh, that gives every other company, you know, alongside open air a good signal that actually we don't have to brute force to just put all the resources altogether. There are tricks that we can do. There are ways that we can actually reduce the cost and still make it a high quality AI. So it's amazing, it's amazing Asia that we're living right now. So it's like this intersection of the chip and the hardware cost coming down just a touch and being able to do more with less, which what technology always does, but also compressing the models and like you said, compressing the gradients, which I want to,

I want to kind of click on that for our listeners too. So we can think of the model as a like, like a small baby, you want to ship it to the way that you want it to be. We need to teach it. You need to teach the baby different either languages, stories, and then it will repeat to you in the same way. So whenever there's wrongdoing, you need to reinforce the to tell them, okay, you need to change it, make it better. So the way that you communicate with this baby is basically the gradients that are passing through to the model to make sure, okay, you are going to the wrong direction. I'll pull you back a bit. You are going to the correct direction. Yes, I'll give you more incentives.

Go straight in. So gradient, gradient is parenting. Exactly. Yes. But then for the large language model, that means our huge efforts on parenting, let's say. So that will result in the data to be huge. And then traditionally, if we think about decentralized model training, if you have to send over this huge parenting order book over the internet, that's super, super heavy. It's not efficient.

So that's why people are thinking of different ways. To make it lighter, for example, I don't have to tell you everything. I just give you some very light comments. And then in your end, you decoded well. to make sure you understand the full command of it, something like that. And there are also people trying to make it like what we call the collaborative layer where the instructions just like a textbook, and everyone has that textbook. So that I don't have to send you everything, and I just send you some very light codes.

Turn to page 127, that's the order. So you read the order yourself, and then that becomes your gradient. So, yeah, tons of different ways people are trying. In order to lower down the cost that we have to communicate between, different nodes or different models so that makes things even more efficient yeah how efficient can it get in the next six months in the next six months uh that um i would say uh six months maybe a bit early but i you never know but in my theory so today let's say an invidia 4090 card is great great for games of course also great for model training and inferring i think that one in a

Some frame of about two to three years will be equally powerful, maybe to the phone that you're having today. So if you think about this, some models that you can run on your graphic cards today may just be as accessible and as easy as your phone in about two years time. Then, yeah, that's the direction or the startups are trying to build up for. And especially within that two years, there might be some interesting miracle happening in the technology side itself, not only as well, Jeremy just mentioned not only just the hardware is getting powerful, but also the algorithms getting lightweighted, getting more compressed and getting more event. So yeah, we'll see this actually coming maybe even sooner than two years time. I want to get into flock, but before we just get into flock, what does all of what you've just said mean for these giants of AI,

for open AI, for GROC, for Gemini, like this shrinking down of the power and the compute that you need to create these language models? What does it mean? mean for these big megalists? Are they going to be victims or successes from this? I think still a general model that's been a foundation model that's been trained by such superpowers are still important. If we are talking about open source models, because it's great foundation for all the decentralized scenarios to trend from. Because if you imagine everyone had to do something or everything from scratch themselves, then that's a bit of pain. Yes, for superpowers, Google, Apple, sorry, not Apple, Google, Oman, AI.

Myrle, sorry. Yeah, it's great still to have them, especially for those who are contributing to the open source world like Lamar, right? It's great. It inspires lots of the challengers in the industry, which is way better than having their own just close source model. So, yes, it's still very helpful. And I see in the future the decentralized players still have to rely very much on the open source contributions, but then they will have their own alter or, or. or specified models onto each verticals in different industries, in different scenarios, to make it more relevant to the use cases. Makes sense. So that's, you've been pretty vocal about AI being very centralized, being too centralized. What is broken about the current system and the current things that are happening? So currently, there's actually a very live case. I used to work in RBC, Rob and Canada as their global head of AI for the past 10 years.

The first time when GPD came out, right, all the banking industry banned the access to it because they can't risk their own financial data or internal data to be exposed to those centralized players. Even if they signed some legal agreement, say, okay, we're using you as a vendor for it. But because many of those data, they can create such insight for derivatives and it's hard to trace. It's very risky for those banks to risk it. And similarly for healthcare industry, this turns out to be two very interesting industries for Flok to be Flux clients. So, yeah, that's generally the case, but I think maybe it's too far away, not so related to everyone. I can bring up some very related examples. For example, you want to have your personal assistant, then that will require you to upload all of your chat histories with everyone so that you can understand you better because otherwise it won't be relevant.

But that will include your ex and some secret friends, so, which you need. know it's going to be problematic if one day someone actually revealed to the public. So, right? So there's always that risk that you are actually giving out that you don't want to. So here's a very applicable thing. Mark, we talked about this at the very end of Nexus when we were finishing Harare's book about AI and humankind. The idea that, you know, future leaders, like there's so much information on the people that are going to be the future leaders right now are like 10 or 12 or 14 or 15. Like they have this tremendous chat history.

So if that chat history goes somewhere and someone owns that chat history, they kind of can own that person in a way, which is like really scary to think about. Does that include signal chat history? Yeah. Fair enough. There's actually a quite live case right now about 23 and me because of their financing issues themselves, right?

They have to sell the company to whoever to give the price. But then that includes that data set of almost everyone who tested, including me. I was like, no. Yeah. Yeah. So your gene is, of course, very important. And your private history, private data, photos, chat history is maybe equally important because your gene from bio perspective can maybe clone you.

And your personal data from AI perspective can clone you as well. So, yeah, it's the same thing. Let's talk about federated learning. Because I know this is one of the pieces of the puzzle. And you mentioned banking. You have a history in banking. I've got a banking analogy I want to throw your way and tell me if I'm right or wrong on this. I saw this as I was doing my research.

So the idea of federated learning can be broken down by a bank who does loans in three different sectors, say auto, home, and business, right? You have these three different sectors and three different data sets. So it's basically to prevent the data from the auto, the specific data, the specific information from the auto side, interacting with the home side, but by pushing them all in together. this federated learning process, you get the benefits of the data working together without seeing the actual information. Is that close at all? Exactly. Yes. It primarily being used in many of such use cases, as you just mentioned, where different desks, they can't mess their data up due to regulations or due to many of the

reasons. For example, you know, trading desk, I'm trading the MEA equities, right, while the other decks is trading the Asia ones, which, you know, are not supposed to actually share your information across or I'm treating one client you're treating another and we shouldn't cross our information but across a firm let's say across different firms it's beneficial for them to have a macro large-length model let's say that can give them great insights into the maybe daily news of today or what what should we expect from the news from the market change by doing so that you have to input some of your internal data and yeah so that's exactly what federal learning is doing. Or we can put it even easily, it's not something new. It's actually

very, it's been developed by Google in 2000, where in your type input software on Android, all of the, your typing habits have been trend by traditional federal learning ways, meaning that your data, your role data is not sending back. It's, it's only the gradient, the gradient of those changes are actually sending back and aggregated between different users. so you don't need to worry that your data has been exposed to another Android user or Google. But then still, it's centralized, meaning that if Google wants, they can actually get your data. Because nobody knows that they're not doing it. And there are actually cases of a software company called SOGO,

input software company called SOGO, who claim to be doing federal learning, but then actually being exposed that they're sending back to the raw data. Because it's cheaper. It's regulating them. Why don't we just send a raw data? Because nobody's actually chasing and say, You don't have to, you have to do this, you have to do that. That's fine.

You take the terms and conditions box when downloading the software, which means that you forfeit everything already. And that's why they make the terms and conditions document like 64 pages in very small fonts. All right. So let's talk about blockchain. We talked about Federated learning a little bit. And then after this, this little bit, we're going to go straight into what flock is doing. So with blockchain, there are, there are tremendous benefits.

But there is additional friction in the U.S. for users. So how do you balance the friction and, you know, in the user experience with the benefits of having, you know, data protected and all of that good stuff? Yeah, for us, it's easily that we just deploy the retail apps. So the users don't have to understand everything around it. They can just use an app. And then they will see clearly from the upload stream and also from the blockchain logs. Because it's open to everyone that their data is never sent it out to any third party. So yeah, so basically federal learning, blockchain, that's the reason why we call us a flock.

So federal learning and blockchain. So the reason we put blockchain here is just to replace the role that Google is to play for their Android users, right? Or Apple is using for their own iOS users. So we have an open nature. So we can see clearly who's actually in playing evil. And we can see clearly, okay, there's no one who's actually sending back my data to a to a sort of party server. So yeah, that's part of it. That's being orchestrating and governing the whole own-chain process when it comes to federal learning in a decentralized fashion. And Floch as a company, we don't have to sit in the middle for every of the clients that we treat with, right? We just leave them onto the protocol. So that's a beauty of blockchain that we think is super helpful into this decentralized AI industry. Okay. So this is a, it sounds like a philosophical choice. You're not using blockchain.

because it's necessarily the most efficient way to create a product for the customer. But this is the philosophy of flock it to use the blockchain to keep third parties out of it. Yeah, kind of, kind of like for all the technologies, right? Centralization will make it super efficient, of course, but then there will be a lot of flaws that you can't control. While in a democracy type of way, then things will be a bit slower, but at least you can make sure there's no wrong doing it in the whole process. So, yeah, that's generally the decentralization.

the centralization type of arguments recently and why people want to do that. What does it make more difficult for you, the blockchain aspect of... Oh yes, the obvious ways are the efficiency of course, because if everything is just trend by sending all your data to a server somewhere and then it will be super quick. But then when it comes to decentralized training, efficiency slows down a bit. But then that depends on different company's strategy. They will have different ways to make it quick. by, as I mentioned, compressing the data, compressing the gradient to make it to make the communication more skipped. So you don't have to communicate every different epoch of the

training. You can do maybe 10 epochs each time or 100 epochs each time, so that the model can still converge, but without needing to, you know, sending too much communications, which save a lot of time. Yeah, lots of tricks have been done by people recently, by researchers, recently, even in media, I think they acquired many companies for the past several years for model compressed because they also believe in that by the time that they develop better graphic cars they also need some techniques to make sure that the models can be compressed enough to to host on their graphic cards right so they are trying on both ways who can get the fittest and the fastest is it the chips or the algorithms right in the software right it's kind of

this balance but hopefully they do it together well let's let's talk about let's talk about flock specifically and um in our research we found that you believe that that flock will democratize AI development. Can you explain that in simple terms to our, to our audience, without any kind of tech jargon and, you know, what you feel that, yeah. As I just mentioned about how flock works, right, it's federal learning just like traditionally, but with blockchain for the orchestration and incentives, of course, naturally come from the Webster industry. So for every task on our platform, it will naturally just be a task raised by anyone. For example, I'm a, I'm a medical company. I want to do a glucose monitoring

application, right? I want a nap, but I can't send my client data to any sort of party, right? So I actually raise this as a task on flock, and then the engineers on the platform and the patient on the platform will join jointly train the model without needing to send their data over. So that will train as a task on our platform. And it's it's been connected, all the nodes being connected by, by, by blockchain. So the ones who contributed, the ones who contributed either the gradients, either the model, the parameters, or the data, they will get their share of the incentives by the time when they, when the model was built.

So that's generally how the model works. So someone raised the problem and then we put it as a task on the platform. And the engineers, patients, or the users were joined. to make the model better. And once this model is better being deployed into the Repel app, then people can just use it and grow it. When you open the app, the app that supports Flok, right? Your data stays local with your iPhone,

so you don't have to worry that it's being uploaded to anywhere else. But then the model, a small piece of the model, will be downloaded onto your phone to train on our data. And it will get that gradient again, we mentioned it. The textbook of how the model should grow, they will send that change of the textbook to the other guy, who has holds an original book and then take your, you know, change of the textbook to change his own

textbooks. So now, okay, now the book's a bit different and you train on my own data. So my book will change, I will send over to the next one. So actually, you are not even sending the textbook at all. Everyone has an original book and then we train our own data and we get a different way, a different, you know, shape of the book. And then we just send over the delta, pass on to the next one, next one, next one, until we all converge on that book. Okay, so you're, instead of me pushing my data out to you guys, you're deploying, deploying someone to learn, not someone, but let's just, let's just personify. Local model, yeah, yeah.

Yeah, local model to jump into my phone and kind of be like, all right, let's see, how does this compare to what we have up here? Okay, here's the new piece of information that fits the gap. I'm going to leave everything else. I'm going to take that new delta, like you said, up and train the model. Is that it? Yes, exactly. And that can be super small because it's only the delta of the change.

I'm just having trouble following that. So is this like a alternative almost to ZK proofs and trying extracting the information that's needed without extracting all of the information? Just. Yes, exactly. Which will give the whole picture without giving all the data. Is that what, okay? You can even think of this as an embedding, but, or you can't really say that. But, yeah, like it's embedding your original data, right, to make it into something that's actually not your original data.

But it can be useful for others to learn from the model. So that maybe the insights of your data, if we name it that. Yeah, exactly. Rather than the ZK proof is sending something out, you're inviting something in. Yeah, yeah, which is a model. Yeah. Yeah. There's the reward.

And then there's a reward mechanism, right? Yeah. Yeah. contributing, you're helping. Yeah, incentives. That's also another big perks of using blockchain, which natively come with incentives and it's fair, because all of them are accountable on-chain, right?

You can see clearly who has been interacted a lot with that model, who's not, and then, yeah, there will be a share by simple calculation that you will know how much you should get. So yeah, that's another piece of the democratization that we're mentioning, because, Transly, you are actually helping open AI to trend their model, but you are paying them 200 books. So it's not even fair.

Yeah, I'm paying them. Okay. So these are the incentives. What role does the flock token play? How much can I make? What's my incentive on flock? Okay. So that really depends on how much incentive the task creator is actually putting down.

So for us, the whole blockchain mechanism is proof of stakes. meaning that, okay, for anyone who want to join to trend the model, who want to join to earn incentives, they have to put down a stake of their tokens to make sure that there's no run-doings. For example, if there are people trying to point in this the whole model, right, because it's open. Then if I'm just a devil, right, I want to join the training,

I just want to put some random data over, I just maybe steal some data from others, right? So I'll be penalized and then I'll be slashed out for the training. But then for the rest of the honest joiner trainers, they will get their incentives share of it. And then the whole pool will come from the one who actually raising up the problem, saying, okay, the one who say that I need a glucose monitoring app. So they will put down the incentive, say, like $100K, $200K worth of the token

to bring it as a bounty on the platform, right? And people who saw that bounty, maybe it's larger or smaller, then they'll be incentivized or interested in, oh, this is a huge bounty. I need to try. I need to join it because I know, can get a higher emission out of it. That's actually it. Here's something I just, when we started talking, you mentioned the idea of paying $200 a month,

but also you're helping that thing that you're paying $200 a month. I just had an image of, you know, I pay $100 bucks to go to the gym. And I go to the gym and I work my ass off. And nothing happens to me, but some other guy is getting bigger and stronger. It feels like when I go to the gym. Fair enough. Fair enough. Let's talk about some real, real world use cases for Flock. We understand how it goes. We also understand the why behind decentralized AI and the benefits of it.

Let's talk about some real world use cases that you're super excited about, some recent ones maybe. Yeah, yeah. So we've been focusing on the business enterprise use cases in the past a lot, like two of the London hospitals, especially for glucose monitoring and the eye clinical funds. And then also the trading desks, as I mentioned, the financial use cases in trading, right? Because different desks, they can't just share information. So yeah, we've been working for such clients already, like top of them ones like GSR and some of the bank Web 3 protocols like request finance. Recently, we're trying to push more retail applications, which we will see that announcements very soon as soon as next week actually.

So for the retail apps, so for the retail user to get more feeling of how FlokWords work on their own, on their own phones or on their own MacBooks, especially now we have the version of Flok application. Just test around with our top five data and then we'll push it very soon to all the Mac users. They can now, they can then use Mac to choose different tasks to join. So yeah, yeah. Can we can we unpack the financial one with a little bit more detail? We don't have to share in answer anything. But what, what specifically are they doing on flock and what are the outcomes? Yeah. So there are different parts of it. There are trading desks of it and there are also the credit scoring part of it which you know

Jeremy actually just mentioned and therefore the credit scoring is simply that we in many cases users don't want to share especially in the web 3 dynamic right people don't want to share their wallet addresses because that means you can trace many of their source of income and also the level of wealth they probably have so we have flock as a as a personal node for each of the users like collaborating with us so you don't have to send your wallet address or everything to a third party, but we can still calculate, calculate a credit score for you so that that score can be used as a, as a verification for a lower collateral or a lower rate

of the interest, maybe when you do lending. So that's one part of it. And the other part for trading desks like GSR, the top trading firms who's actually working with us, the trading algorithms that we're providing with them together is the ones that they don't want their different desks to share the information because for maybe for a certain token that you can't see the information from the other token. It's not even legal, right? So you have to make sure it's well segregated, but still we have the macro of the trading mechanisms that's being provided with insight for information or signals for their traders to make executions or even for their trading boss to make executions. So there's like there's a bit of a layer that you guys are the

layer that kind of lives between the users of said platform and then whoever's performing, whoever's offering that platform to the users and you are this intermediary segregation layer maybe. Yeah, yeah, to prevent any authority to actually holding all the data in one place, which is very dangerous. Yeah, very dangerous. On the, after that, we've looked at the finance side of it, could we, could you impact the healthcare side of it? Because I think a lot of our listeners understand finance and crypto and Web 3, but when they

talk about AI and blockchain and then you bring healthcare into the conversation, perhaps especially this the point of incentives. What the healthcare part of this and how the incentives play a role in that healthcare scenario. Yeah. Yeah. So that's actually where we're going to launch a very interesting retail app. Well, we've been collaborating with university college London hospital and also more size hospital on virus research cases, right, but never really pushed any retail applications. So this will be one interesting one you'll see it very soon. So we call it baby for this. Basically means that you you have so for those moms new coming moms they will have a scan of their ultrasound and then some of them in traditional hospitals they are actually paying 100 pounds to have a very

naive rendering of the of the ultrasound image to a real baby face which is quite fun like I I don't have kids myself, but my other co-founders, they just realized this is actually a huge opportunity, especially, you know, collaborating with the professors there, and they think, yeah, this is very interesting that we should do it. And this is the exact... Ghibli Studio rendered versions. Yeah, yeah, yeah. And this is exactly the case, right?

If it's just your personal photo, maybe, that's fine. Arguably, it's okay if you want to send over to a third party. Under many regulations, even for the US under hip-part, right, your ultrasound data, that's healthcare. You can't send it over like it just randomly to any third party. So yeah, so that's one of the use cases where I think the retail users will get it instantly. So now we have this healthcare data can also be used and then and tune the model and then to render the face for me. It's more in amusement, you know, part of the, instead of the actual use because the face might not exactly look the same,

but it's fun, it's fun for people to understand and I feel how flock works without, you know, sending your data, over but we can still you know train a model together to make it the facing face rendering better i was speaking he works in biotech the day and he was saying that the biggest hurdle or challenge that they face is this protection of IP obviously and they they want to share their research they want to share certain results but they also want to protect what obviously their IP because in in this it's worth of millions hundreds of millions maybe even billions of dollars and is what locked in with blockchain and AI a possible use case. Yeah, exactly.

Exactly. Yeah. Yeah. So by not sending those raw data over, yeah, now you can protect a lot of those IP-ish already. I'll tell him. We also have another budding business opportunity built on humanity's lack of patience. You know, you get back in the day where you just waited so the kid came out to understand

if you were going to have a boy or a girl and then you can do these tests to figure out you have a boy or girl and you have these big reveal videos. Now you can see the kid. Like it's brilliant. I think it's cool. I think it's a cool idea to kind of envision on that. But what you just mentioned, I was thinking about my own parenting journey.

Won't bore you too much with parenting here. But this obsession that people have with the next moment, like maybe next week I'll be happy. He's like, why don't you just wake up and your kid's 10? And you just forget, don't bother with all the difficult years. Because when they're 10, it will be cool. I think that what you just said just reminded me of that, of this obsession with next, next, next. Yeah, yeah.

And they're willing to pay premium to just see it a second earlier. Everyone likes the behind the scenes, the sneak peaks, the exclusive access to baby geometry faces earlier. Well, let's explore the idea. When things become decentralized, centralized authorities tend to freak out a little bit because one of two things, they start to see their potential control over something easing. but there's also on the flip side a decent argument for if something is truly decentralized, how do you make sure people are using it the right way? How do you make sure it's not turning nefarious? How do you guys think about that?

Yeah, there will definitely be some kind of verification layers or verification mechanics being lived on such decentralized AI protocols. Just like nowadays, for building airplanes or large machinery, right? There's always a verification department of it. While that code base has been sitting there for decades, you just need to pass all such verification processes in order to build your own robot or your own machinery. But of course, there's also drawbacks. In some cases, those code might be too outdated.

They need to update or many of the tricks of the hacks that they can maybe trick the results and then to make the verification looks good, but then still they will have their drawbacks. or flows that's been putting there on purpose. Yeah, there's always challenge on such things. Even for centralized models, there are people trying to reverse attack such models to make it say something that's not supposed to be saying. Yeah, and even for centralized ones, we've seen many cases where people are trying to hack it. So I think it's just there for Albert centralized or decentralized models.

It's just there. There's just being some bunch of people always trying to, you know, hack it. Well, full clarity, too, the larger models, there's not a whole lot of governance or accountability happening right now with how they're training their data, how they're doing stuff either, right? So, but what happens is here's the spin is where like, oh, these guys are decentralized. Nobody's regulating. Nobody's watching these guys. But I would argue since they're decentralized, since they're decentralized, they're part of an ecosystem, a community that's all has a vested interest in that. Wouldn't there be more accountability and oversight in that?

Yeah, yeah. That's a good point of it. Thank you. And I think so. Yes, at least those models are being built by a group of, yeah, people share similar interests. And the technology itself is used by them for, yeah, for the sake of technology itself, right? So it's not, yeah. Yeah.

For that, yeah, for that, I'm super supportive. Can they get your opinion on something which may or may not be in your, your ballpark to answer. Maybe for some of our listeners, have experienced AI hallucinations. And I've definitely, definitely, definitely noticed a huge improvement over the last eight,

10 months in hallucinations. But I was reading some interesting Twitter conversations this week. And I just want to read a tweet from somebody, quote, I sense that a lot of people now think knowledge graphs will fix the LLM issue, the issue being hallucinations.

But no, they do not. They do not. Even in the case that knowledge grasp would prevent logical inconsistency 100%. A lot of text constructions that are perfectly logically consistent, but have zero relation to reality. So could you unpack that quote for me? And beyond that, talk about hallucinations for regular people using AI, where we are, how we can remove them completely, if at all. Yeah.

So that's quite an interesting question. Thank you for bringing this up. Actually for my past four five years of being a research fellow at Imperial College London, my original research focus is actually in graph networks and graph neural networks, knowledge graphs. So it was once a very high topic. You're the mad to ask you. Yeah, yeah.

While people were very much interested in how to build a great graph, right, that can do all the culturality. So once you have a question, we can just source back. to see the culturality and give you the answer, right? But then just over the night, GPT came out and nobody did that at all. Like I have two PhD students who just then had to change their thesis to make it more relevant or more up-to-date

with a new larger language model or transformer type of technologies instead of focusing solely on the knowledge graph part of it. So that actually just trying to give you an example about looking back to knowledge graph, might go nowhere because from the experience that we saw, let's say 10 years ago, about deep learning came into power, right? There are still a bunch of people who are trying to do possibility analysis or probability analysis on the statistical level of the models, right? But then where are these bunch of people? You don't see them, except for some research papers, they're still publishing things.

Because in terms of applications, people don't really, not don't really, but people care, But then because of the growth of the compute power and growth of the algorithm by piling up the computer power so quick that the causality analysis of such things can just just cannot catch up. So in the beginning, when deep learning become a thing, right, many of the, let's say the hedge funds do not, cannot use it as a proof for giving the, giving the alpha or explanation of the alpha to their clients, right? But then over the time, they gave up to say, okay, it's fine. Let's do a black box back testing and give you your results. And that's just our AI deep learning alpha finder. And yeah, believe it or not, similar things will happen for the LM stage as well. Many people will claim that, oh, a graph neural network can solve the issue.

But the thing I see is that pace of developing graph neural networks or graph knowledge graphs will not be compatible to the growth of the brute force large language model, basically, because it will just go crazily and go super large while it gets getting more explanationally harder for a large graph to be constructed based on that newer model. So basically, the speed of the growth of two fields just cannot match its other. So you can't really rely on the other to do the verification check for the Yes. So theoretically it works and theoretically yes, it's a great solution, but just practically, we won't see the two two sides of the match, you know, be matching each other. You will see this as an example in the finance industry. There's only one company run out with their run out with their knowledge graph. That's Bloomberg. The only one that have enough data to treat the knowledge graph. Rest of it, even the second in the market, I don't want to name the name, they failed. The project failed. I was the one who's with them for that for the project totally fell because not enough data not enough

research not enough even compute or storage for them to grow a similar a similar in a product compared to the other one who's in the top in the world right similarly now same same age of the lms and if you're expecting someone will come up with the knowledge graph that can actually facilitate or verify the large language model of the next day to me i guess it still will fail and people still get used to larger models, even if it's black box. Yeah. So this tweet is correct then, or her assumptions are correct. Where does synthetic data fit into everything you just said?

That might change. But then the question is whether the synthetic data, you know, data distribution in synthetic data is accurate. Because, you know, by the time of 2017, the last moment of human generated data, died because that's where that's where the first version of GPT came out and there's tons of you know synthetic or generate AI generated data on the internet right now right again there's a bunch of people who are just found right on Twitter they create alternative histories by using large language models now who's the 47th

president of the United States they would just put Elon Musk and they'll put tons of resources to back such alternate or altered history so think of one day two years later, right? There's a new AI model to be built. And it's just extracting data from the internet, for example, on Reddit. But they see more data talking about Elon Musk as a president instead of Donald Trump.

Then they'll put that as a grand truth. Damn, there's more information about it. When we think about things like data, yeah, it's been polluted already. So you can't be 100% sure that synthetic data, that you are creating is the actual data that we're presenting the human generated data. So can you can you?

just be ordered already or polluted already by the AI that we created right now. It's a vicious circle that leads to a place. Yeah, yeah. Like it's just no way back. Just like early days when people launching their satellites, we think it's fine. Ilom Musk did a whole bunch of satellites around the whole globe, right? But then we realized it creates tons of garbage around this near space around us. But there's no way that we can clean them now. same thing here. So it's being developed. The area has been developed already and the pollution is already started and it's actually getting larger. And yeah, there's no way back.

It's such a fascinating conversation that in your mind, that example of Elon Musk is president, even if it's by proxy, but the Twitter masses confused the future AI and the future AI believes that Elon Musk was the president in 2025, not Donald Trump. How do we not get there? Oh, how do we not get there? Yeah. Yeah. So back to the question, I think Jeremy mentioned brought up earlier, right, the verification

of protocol of it. So either the real data watermarks that we can put through to all the real generated data or ways that we can preserve the private data, right, not to be exposed to the public domain by using technologies like flock. I got it. So yes, like some of your personal, you know, chat histories, you never actually send it over to iCloud. So there's no way an AI can actually get those data and then generate some alternative history for you because, yeah, you never actually brought out. So there's tons of

there still. So friends and neighbors, instead of just chucking your data into these models, just allow, yeah, just say flock it and just allow one of these one of these little models to jump into your data, figure out the differences, and then just report back on the differences. And it makes it makes it a little bit more palatable. And actually, you could get some money from it too. So who doesn't like that? Let's try to land the plane a little bit here, guys. If someone listening, Jihal wants to get involved, wants to try this,

hasn't really messed with blockchain much, hasn't really messed with AI much. What is a way that they could tiptoe into this and be like, what are these guys doing? And how can I play around with it? Just follow. Well, just go on to our website, flock.io, and everything is there. Twitter, our blog, and also the tools that you can use right now, to train the model to join their training. Yeah, everything's there.

So, yeah, flock to I.O. Thank you. Awesome. Yeah, flock to I love it. We have a question left from our last guest, David Bianchi, who's a Web 3 film producer, actor, writer, director, poet. And poet. This is more on creativity, Ghi,

or actually just kind of getting what you need to get done related to what you love to do. What are you doing? his question was this, what are you doing to stay out of your own way? Wow. Stay out of my own way. Yeah. I think always trying to explore the boundary of your interests, trying to do something that's out of your scope. Like I was traditionally a trend, AI researcher engineer. But then I chose to work in a finance industry, which I have no idea about, you know, before I joined the company. And then I was so long years in a traditional

industry. Then I tried to explore into somewhere crypto, which is also a totally new. new world to me. So for me, it's like stay, stay interested in new things and new interesting things and then to develop some crossovers that can actually bridge your experience from the past to the future. Awesome. Curiosity. Wonderful. Wonderful. Well, Jiha, it was a pleasure talking to you today. A couple of quick recaps. We've learned that AI is a bit too centralized. We also learn to maybe say flock it and try to fix that with decentralized coupling AI and blockchain. We also learned a bit about federated learning, which is an interesting way to put your information into something larger than you without giving up your privacy, which I think is super cool. And real world examples are in place with what you're doing.

And we look forward to track in those. More importantly, listeners, what do you think? Would you contribute to this decentralized AI model if it meant you had more control and privacy and if you were incentivized to it? Can I just say that listeners, every time you use, use AI, whatever model you're using, just when you click enter, just think, just a little bit, just think of what you're sending, where you're sending it, and what that means today and in the future. And when you do that, I think that perhaps you'll start thinking about decentralized AI

and you are important. Your data is important. So think about it. Yeah. See how thanks. Thanks for joining us. Really appreciate it. Keep us posted on the progress.

And hopefully we'll talk again really soon. Yeah, thank you. Thank you, Mark. Thanks for having me. Thank you. Take care. Cheers, take care. Bye. I like, I'm going to do these end of show lessons from thinking on. I love it, Mark. I love that. These reminders to be human, cautionary tales, like... We should write our own book of fables. We could do like a technology fables book. I love that idea. Little short, short stories with some morals. Yeah, a fable, exactly. Like, I think one of the oldest books in existence is the...

Sops Fables. Let's rewrite that for the technological age. That's actually a great idea. Publishers if you're there.