Decentralized AI: Crypto’s Killer Apps? | ETHSofia 2024

[Music] Hello everyone. Uh today uh the topic of our panel is decentralized AI. So we're going to try to demystify the AI concept. Um see uh how it works and why would we need it. And today I'm pleased to be joined by uh great panelists here.

and I'm going to let them uh quickly introduce them themselves and um tell things about what they're currently working on. Okie please. Perfect. Well, thank you so much, Alex. Why why don't you introduce yourself first?

Oh, okay. Thank you very much. So, I'm uh I'm working at Aura Protocol and I'm a Deel. So, what we're trying to do is we're trying to bring AI on chain. We have this product AI Oracle and um if you want to build some AI powered apps, you can use our products.

Um yeah, thank you. Okay. Well, you're welcome. Hello everybody. My name is Luki.

I am the head of ecosystem and strategy at Chainbase. What Chenbase does is we provide the new data stack for uh datadriven DAPs and also AI initiatives. So what does that mean? Right? We have all those data on different blockchains.

Each blockchain is its own island. You have different way to index. You have different way to decode those data. And now we have all those AI projects coming up. We need those data to feed their models.

So what Chinbase does is we provide this data processing standards and we open sourced it. We invite developers and also data scientists to create those framework so that you can fetch the data you need from all different blockchains and backstage we just talk about chain abstraction etc. We want to make sure the user experience developer experience is up to the maximum level. So that's what we're here to do is to help developers to get those onchain data that's specific specific for the parameters they set for the models and projects. Y uh I'm Yosi Blah.

I'm the co-founder of ICP Hub Bulgaria. We are regional we are a regional hub here in Bulgaria as well as the Balkan region when it comes to ICP which is an L1 blockchain developed by the Definitity Foundation. Uh and on top of that I'm also a head of innovation and creative at Resolute which is a web3 marketing agency working with a lot of projects within the AI deepen uh space. Hey everyone, I'm Lawrence from Extropy. Uh we're a consultancy.

We uh provide auditing uh and education. So I spend a lot of my time uh explaining concepts such as zero knowledge proofs uh and ZKML to people. Uh so yeah, that's my my going to be my contribution today. Thank you very much everyone. Um so first let's start by um seeing what does decentralized AI mean to all of you.

Um we can start from bookie. Okay. First of all I just saw myself on a monitor. I didn't realize my big blanket. It's that big on my neck.

It kind of distract me from everything here. But it's really cold here in Sophia for some reason. Everyone was telling me October is actually really beautiful weather and I believe so and I agree but it's just a little cold in the morning and at night anyways. So that's a yeah in Balcons it's a little bit cooler. Yeah it's a it's a pivot.

I came here from Greece so I got a little spoiled by the sunshine over there. uh anyways decentralized AI um I will just briefly touch base a little bit of what I know and I will pass on to the rest of the panelist. So when we're talking about AI, it's artificial intelligence, right? But what really are the components for AI? We have compute, we have model, we have data.

And when we're talking about decentralized AI, that means we decentralize compute, model, and data. And for all each three of them, it has development, it has training, and it has deployment. So combine all together six elements or nine elements however you calculate it we have to decentralize all of them is that feasible but this is what we're aiming to do is to at least from one piece to another computation now we have decentralized computation projects right and we have models we have all those marketplaces for different models hugging face if you get on hugging face there's a million of open-source models. Of course, they all are um optimized or forked from another model that's um that has been trained from those big giant companies. Um and then we have a decentralized data.

So decentralized data we have those data dows they get voluntarily um provided like uh people will provide data voluntarily uh like the Twitter data or uh reddit data those are web two datas and also we have the web three data which is onchren data so all those things we have to put them together to achieve the concept of decentralized AI so that's in my in my opinion it's something that we're uh we're driving towards to um but to think about decentralize every single component of what I just mentioned um it probably is going to be a very interesting journey for all of us. Okay, thank you for that. So you think that we should decentralize every layer of this whole process, right? If we're talking about decentralized AI, we have three components as I mentioned. That means every single layer needs to be decentralized and there's also a reason behind every layer being decentralized.

So um like the data layer decentralized why because then we have everyone who has a slice of the pie that this is my data data self sovereignty um I govern this data I get incentivized for providing it computation I provide this resource I need to get I need to get incentivized. So there are different reasons behind the decentralization of the AI. Thank you very much Luki. We'll touch more on that later. U Yseph, what's your take on that?

Yes, I must say I agree I agree with what Luka was saying as well in terms of the sort of the currently from sort of if we're looking at it from the centralized standpoint. So let's take for example open air or something like that. It's currently a black box of uh how the data is being collected, how is it being trained and it's taking the data from the internet. It's taking the data that we've posted on the internet all across the space and then it's basically training the model based on that data. And when I say in terms of the decentralized AI, what is what do I mean by it is I like to know where my data is going for, who has it, uh if if that data is private, if my sort of uh information will not get leaked and then also on top of that uh enabling users to have not not users enabling essentially that there's no single point of failure, meaning we decentralize the computing around the AI and enable them to get rewarded for those actions as Liby stated afterwards.

Okay, thank you very much Lawrence. Yeah, similar uh ideas. I guess I as a uh web 3G I guess I've been really uh excited about decentralization for a long time. I want all my systems to be to be decentralized. So I would want you know all the areas of uh AI to be decentralized if possible.

I think the way that's achieved is going to be different for the different areas and it mean will mean different things but one of the the most important things for me for decentralization is the idea that I can have control over my data or over my assets and so uh particularly with AI and the fact that you know there's going to be this huge amount of data very widely used I want to be able to control my part of that as much as possible. Thank you. So we just mentioned that we want to decentralize different components of AI and um also we mentioned um identity and privacy of the data and um ownership of of that data right so but um I really want to clarify why do we want it to be decentralized so we we just don't want to decentralize just to be decentralized we really want to to have some good use cases so um can we think about some problems or maybe some um um consumerf facing use cases that you that you can think of. Um what do you guys think? Where can we use this technology?

Yeah, there are a lot of places that we can use the technology but um healthcare is one of them but there's a lot of regulation around it right. So there's a report showing on Google Google search 40% of a Google search is a health related people asking about what happened to my headache or things like that. So 40% of a search is about health that means we need people to understand that this is amount of data that it needs to be trained. So if I have something going going on with my arm, I want to do this search and because of all the data that's generated that can can give me some sort of information regarding what's going on. But when we talk about decentralization, it's user data and it's the data that I need specifically for myself.

So the things that they detected from you probably does not apply to me. And that's why we're talking about decentralization because when we're talking about data control as you guys mentioned this amount of data it's mine and we use this model to train my specific data AI agents. So if I have something going on that actually related to the case that only applies to me. So that's what the decentralization is about is it goes to specific individual is goes to specific use case. All the data that's been trained it's not biased.

all the data that's been trained that's used and necessary for the individual who need that data who need a model to be trained. Okay. So if I understood um each user can have its own uh specialized agents trained on their specific data right and uh also in some use cases we don't want to expose that data publicly in in healthcare or some other um confidential areas especially when we're talking about now we use CHBT to get answers but the thing is you don't really know where the data comes from um the data that used to train and to be honest with you I was reading some reports um regarding the data the open source data when the data is open most of them have already been used so there's a convergence of a models having the same result because they all use the same data to train the model so then there's like another layer of a synthetic data in etc actually just announced um their I think the video um generation generating video um kind of kind of um product and they're saying that they did not use any of the user data use user data to train that that specific AI model or um the the the product. So they only they say they only used midjourneys results to train their data. So if that's the truth then that means now companies are thinking about creative ways to use the synthetic data to train the models.

So is that is that something that wouldn't be biased in the future? Is that something that actually will uh help our human beings because it's a synthetic data it's not the real user data. So by thinking about using the real data, the user data, we can control the data to train the models, then the result will be much better and more applicable for our daily use. Okay, thank you very much. And and what about um uh can we somehow incentivize that data for for these processes like for training for for other components of of uh AI?

What do you think, Yosip? Yes, of course. I mean there's there's one very good use case which we've also touched before especially when when it comes to the healthcare. Let's say that you have so there's a couple of you know diseases which are unccurable and there hasn't been a cure for them currently which was found but let's say that there's a big research center somewhere happening whatever the country there is and they you all the people who have that particular disease can voluntarily give that data to that research get of course rewarded for it but also they have the possibility for that data to cure the disease in the long run and I think That's one of the things where we're slowly getting there in terms of the decentralization especially when you think about it and that data which they provide is the privacy. It's sort of very much private because it's their own uh sort of information that they're putting out towards that research.

Right. So uh I'm not sure if I answered the question. I believe you mentioned something else in terms of the incentivization. Incentivization. Yes.

So currently as I said before the models which are currently being used uh to you know the chatbd everything there's no incentivization for from the users p perspective they've taken the data from the users from all around the world and they haven't given them any incentivization similar similar to be said like if you were a beta tester at Facebook which is a sort of a back in 2008 you haven't gotten any sort of a reward for beating being the beta tester in that whereas here when we're talking about decentralized AI you can get rewarded for allocating your computing power allocating the resources on that sort of a front but then also allocating the data uh the your own data and getting rewarded for that be that if you're an artist whatever it is it applies to everything right thank you very much so um yeah another thing that that uh at Tora specifically we we focus on is not on that data but more on the on the execution layer. So if you want to use this AI inference in your smart contracts and you want to use it in um verifiable and trustless way um you have few choices like you can execute this um AI algorithms uh completely on chain and uh but that is really expensive you you will spend a lot of gas and it's usually not feasible so the other option is to to execute it offchain and then you need somehow to prove to the onchain context that that execution is actually correct and that specific models executed that right so that that's the problem we're trying to tackle at Aura and there's many teams in the industry have uh different approaches to that they try to use different cryptographic methods um and uh I would love to to touch on these different approaches and maybe Lawrence you can you can touch on some of them um what are what are some of the ways that we can actually verify this offchain AI inference to to the to the smart contracts. Yeah, I think that's a very important aspect. I mean the I I really like the privacy and and I think that's important but also we want to be able to verify that these models are actually working correctly uh particularly because we do find that there's bias sometimes and uh we need to yeah so we need to make sure that the the inference stage is working correctly. So how do we how can we do that?

as well. Uh some technologies that have been used uh we have uh FHE fully homework encryption which uh allows us this is very good uh for for data uh processing generally. This allows us to do operations on encrypted data. So you then have the uh confidence uh that your your data is uh not going to be understood. You can outsource your data uh to be processed and you know it's encrypted.

So that that gives you the confidence there. Uh other other techniques the uh the one I I know best I guess is ZKML. So this is taking zero knowledge proofs uh and then using those with machine learning. So zero knowledge proofs provide uh verifiable computation and so they can show that a particular computation was done correctly and in this case some a machine learning inference for example. they also it also allows uh private inputs to computation.

So that that's an extra advantage. But really I think we're focusing on the the verifiable aspect. And then a nice sort of not exactly a twist on that but the uh another alternative to that is the the OPML side the the optimistic machine learning uh where they uh they're assuming that the machine learning has been done correctly but then they provide economic guarantees via uh blockchain to uh to give uh the the opportunity to uh do resolution if something is found to be incorrect. Uh so those are some of the the sort of cryptographic sides. We also have trusted execution environments as a hardware solution.

I'm personally perhaps a little bit skeptical, more skeptical about those. I mean, they're they're good, I suppose, in that, you know, they're they're a hardware solution. So, um they, you know, they they can be available, but uh we do then have to trust the the hardware provider that there's no backd dooror. And I think you probably know there have been exceptions to that. uh and also they only work uh properly the you know in that extra act actual environment they're only saying there's security within that environment so they're not making guarantees about anything outside of that.

So a quick a quick rundown there of some of the uh the techniques. Um yeah, there you go. Yeah, thank you very much. Um yeah, definitely there's a lot of a lot of um u talk about trusted execution environments these days. Um it obviously has some some pros and cons uh each of these approaches and yeah I'm excited to see um where will it end up.

I just want to mention about the incentivization part. I think uh Aura actually has done a really good job regarding the IMO which I'm like I'm a big fan of the initial model initial model offering. Yeah. So that actually incentivize the models like the incentivize the model providers. If you train the model and you list it on Aura, you actually get incentivized.

Is that Yes, that's correct. So we have this um offering called initial model offering and u um uh what it allows you to do is if you're developing open-source um AI model you can kind of um do some incentivization around it uh to reward contributors um community around your AI model and then if you integrate it in AI Oracle um and many decentralized applications actually use your model you can have the the full loop like uh you you have a constant development and then from from the usage of of this from the onchain usage all of these token holders um get rewarded. So that's kind of we believe that um the future of AI should be open sourced. Um as we as we've seen rapid development in the space um AI is shaping the the human future and we want to make sure that we do it correct correctly. Right.

So we can touch also on the reg regulation side of things. Yeah, that's also interesting that because the model that you can get incentivized then what about the people who provide the data to train the model how are they going to get incentivized you know I think there are so many different layers when we talk about incentivization even though uh Lawrence was touching based on like privacy and stuff like that all those technology that we can integrate with the model training in integrate into the technology but like coming to incentivization There are so many different layers of contributors that we have to figure out how the incentivization really goes, right? And what what about user experience like um what do you think uh about these u big tech solutions and um opposed to some open-source decentralized solutions that that we're proposing? Like do you think uh regarding um first of all um who who will be the users? Will it be more user centric or do you think it's more on infrastructure level and what do you think are some user experience um uh drawbacks there and do you have any so you can okay I will just I just be really fast about um couple things that I noticed um when I was working with projects or training models or even chain base we train our own model which is called a C world model Um I mentioned earlier about three components of AI model, compute and data.

So we talked about data decentralized data which is easy to understand but what about computation? I would just touch base um on that lightly is we were we tried a different decentralized computation um projects out there and there's some really good ones if you guys search Twitter there's some really good ones out there. uh raised a lot of money. Um and by saying that it's decentralized, we we give it a try. But when you're training a model, that decentralization, it does not work well.

It works very very poorly compared to centralized um computing system. So if one node is not communicating with another node, the whole model is failing. like it needs to be in the same place the neural network needs to be together in order to train model just to I can't think of the uh the power resource that you need to train chatbt but um by 2026 I think I saw it was like up to 5 to6 billion dollars just to train you to train the model so that means if open AI they just raise $6 billion for the finetuning, for the training and that's a centralized that's closed source and for like open source and decentralized think about all those nodes are all scattered into different places. Then we have to think about the the regulation that we'll be touching base later on is that how feasible it is for us to decentralize the compute power to train model. Um so that's one thing that I wanted to talk about too is the user experience.

And so when we talk about AWS, it's yeah it's a centralized it's closed source but they give you all those plug and play. You click click click click and this is the parameter you need in order to do something. But with open source right now hugging face you get on the GitHub hugging face and you try to understand the code you try to understand how to integrate. You have to have the knowledge in order to use what they what they have offered. So I think as I mentioned it's interesting it's a long journey for all of us but at least this is a good initiative for us to think about how to improve the user experience.

Yep. I mean I must say that I I agree on this uh entirely as well because you know when you have sort of graphics cards running in SLI bridges on big big farms it's definitely going to be much more performant when you on on that front rather than when you decentralize them. But also on one other thing, I mean crypto in general, crypto started from the trenches, right? And it was not built with UX in mind. Uh it was built by us who understand it and then when the new sort of people get on boarded, it's very much hard for them.

I'm talking this from sort of a user perspective. It's very much hard for them to even realize what's happening where in which chain they are on. How do they need to bridge? I'm deviating a bit but there's uh there's new narratives like chain abstraction coming out uh which essentially abstracts all that uh in the back end and user doesn't need to necessarily know it can he can know but it doesn't need to be aware of what's fully happening in the background as the same case is as the end user doesn't know what's happening in the background of open AI difference being here that this is decentralized and you're able to see it if you're very interested into how does it work right so I think that when we push this sort of a chains rational narrative which is coming out I mean it's not something which is here now it will be in the future and then when we sort of sort out some of the bottlenecks when it comes to computation side of things I believe that we can get in we can do a very nice user experience in the next few years as the technology progresses because I mean if you even real if you know it like first the graphical user interface like which was in the 80s it needed to get adopted like seven eight years it's not something that comes out uh immediately so but I'm very bullish in terms of this so that we're going to see it and the user flow will be good and the data will be private and sort of a DI will shape sort of the future of it awesome um Lawrence do you have anything to add yeah I think that echoes those points so what One thing I would say for for the end users, I think the there's always going to be an attraction towards the the large centralized uh projects. The way that we can do something about that I think is if we do have incentivization.

So the types of things you were saying around the medical data that that kind of thing will help move people away. Um but again and also what you said about the the user experience. We've been web 3 has been fairly bad at user experience. Um but uh I think things are improving and perhaps this could be something that can then develop further and become uh something that can be used with AI as well with with uh decentralized AI. Um I also wonder if the the efforts that are going into deepin could also help here that the decentralized infrastructure will also feed into this and could could help.

And then uh and finally one uh good point I think with reason for for the decentralized approach I think is around the standardization. I think if we have very centralized projects that are leading things and uh pushing the pushing the debate in a certain direction, we run the risk of uh moving away from open standards. So if we can keep things decentralized, we can get good standards. I think we're going to be in a much better place. Totally agree.

Um yeah. Um I mean uh so far what we what we mentioned is on that data layer, we want to to have ownership of the data as users. We want to be able to monetize that data. We also in some cases want to might keep it confidential while also contributing to training of some AI agents. And then in my view, we can also use blockchain as if we have many specialized this AI agents, we can use blockchain as um communication channel between these different AI agents.

They're highly specialized in in specific tasks. And um I heard some people talking about the whole uh uh commerce network um with within these uh between these AI agents uh and they're utilizing blockchain to uh verifiably um transfer messages and communicate and uh yeah, I'm definitely excited about about the future of um of DAI. Um, I think there's going to be some use cases, but I wanted to ask each of you um do you have uh one project in particular that that's not your project obviously, but uh that you're really um excited about like the the approach they're taking and yeah, I I actually wanted to go back to what you said about um just AI and and blockchain um the use case user experience. One thing I would love to invite you guys to think about is now we have a chatb, we have a mid journey, we have all those generative AI uh products out there, right? So now we have those um like faking someone's voice, those photos, you generate like Donald Trump doing something, it look like it's real, but actually it's not.

And you generate me saying something, it's not real, but it look like it's real. So I think that's also what blockchain is really good at is the transparency tracing back to where the data is tracing back to where the model is. So for that specific auditing of the social platform and uh what's what's like what's out there is that authentic is that actually from that that data source. Um I think that's what blockchain also has a big role in the future too. Yeah.

um couldn't agree more on that. And then for the the the project that I'm really um bullish on, I think it's something to do with the deep projects for sure. I think with all those user data federated learning about academic research and hospitals about cars there's like uh from 2024 to 2049 29 just those five years the data from just car companies will go up like four four to five um fold. So the thing like from what I was reading the information car companies has is way more than what Chad GBPT has used for training. So in the future there are a lot of information those edge nodes the deeping projects health healthc care projects they can be integrated with those blockchain like user data.

I think I'm super bullish on that and with federated learning and incentivization of using data and also combining with the different kind of a models um and I I think like a different layers all together it's going to make um the whole whole landscape of AI and crypto big. Thank you very much. Yseph, is there one thing that you're really interested in? I I mean in terms of the deep deep in side of things I'm also very bullish on that particular behalf. Uh especially I I was I was very big into the node sales and I'm trying to figure the time will tell how how will this sort of progress in the future but I think that it's going very good.

Uh but on some of the projects I'm very much when it comes to like fully web 3 side of space there's AI agents which enable you to do your own meme points. Uh you provide a liquidity to it. they created Twitter profile. They have their own the entire the entire pipeline of it. Uh and you're able to do certain things.

I've tested a couple of them. And this is sort of the approach that I'm currently playing around with. It's not nothing like revolutionary, but it's uh very fun and entertaining way to see how it works and what are potential possibilities later down the line. Uh, so yeah, those and I believe that there was I'm not I just followed it a bit, but I saw that there's a rogue AI on Twitter and that recent Horowits put up I believe it was 50K in Bitcoin to that AI and now it's creating its own memecoin and potentially it's going to rug. So there's a lot of things happening over there.

Yep. Um, awesome. Lawrence, your thoughts? Yeah. Uh I I also I think one of the most exciting areas is the the AI agent area.

I I think there's uh there's a lot of possibilities there. So I don't have a particular project. I'll just say that uh that area in general on the DeFi side. Well, that's sort of taking arbitrage bots to the next level. Um and there's lots going on there.

There's all sorts of marketplaces that you can uh you can sell your your bots or your your agents, etc. Um but one area that I I like uh and this came from web 3 and particularly on the game side is the idea of autonomous worlds and this this really does excite me. Uh I like the idea that you can create these worlds and they can uh evolve independently. And I think uh if you're doing that then you need to know that the the agents involved in these um are operating correctly. They're working according to the rules that they have they claim to be following.

uh and that's where we can uh do some verifiable uh we can check the verifiable aspect of the the computation. Um so yeah that's that's the area I think that excites me most. Yeah. Okay. So um yeah we're getting close to the end of the panel.

So I definitely want to to engage with um with the audience as well. So if if anybody has any questions for the panelists uh we can take few of them and uh continue this discussion. Um so yeah do we have any questions there is one so hi uh it's great conversation what do you think is one thing that the only decentralized AI will do and centralized AI will not be able to do amazing question um do you have any thoughts can you repeat the question yeah the question was um what would be one thing that only decentralized AI can achieve and centralized solution cannot. So basically the the core question of of the whole um discussion I think um do you have like clear view in your mind maybe or some direction um so say on on a general level the the the uh if we if it becomes too centralized then they're going to dictate the you know the large players are going to dictate where this area is going. uh and uh potentially that you know we as as the end users we could potentially lose out there.

So I think that's one of the the main reasons we want decentralized AI is so that we can have uh more of a say and more impact in the the direction things are going. So yeah I can maybe add on this as well. Um so from from Aura perspective um what we're trying to achieve is this um you you can have these use cases where you need to keep decentralization um like build fully trustless applications on blockchain and um you don't want to rely on this centralized points of trust like if if you don't want to to trust that the treptt provided you the correct answer but instead you want to have verifiable flow to verify that um um that the computation is correct on the provided data. And also um touching back on this uh data part of of our talk um also you want to to have um data integrity and as as Luki mentioned I think uh in the future as as as we will have more and more data available on the internet specifically uh blockchain will be like a source of truth of the information. So I believe that it will be very important to to keep that truth on on this fully replicated databases and use that uh as a as a starting point.

I think just to visualize this we're now training a model right and we need every one of you in the room provide specific data to train this model. It's very transparent. You know what data we use, you know what model we use and you know what the result is. But this is decentralized way. It's very transparent.

The uh the result that's trained by the model because of the data that's provided by everyone here is uh it's not biased because everyone has the consensus that we're providing this this data for for this model training. But if it's a closed source is if it's a centralized the data you don't even know if what data I used to train this model. I maybe just use this group or that group. So the the result it could be biased even though it's not intentional but the result could be biased. So that's the difference.

Well in my opinion one of the differences of the AI decentralized versus the centralized is decentralized. First of all, you get that slice of a pie. You provide your data, you get incentivized. Second, you there's a transparency and the result of the model, the result of um yeah of the model is not biased. Everyone knows where the data comes and everyone knows what the resources is.

Um so that it does not give you the answer that specifically they want you to know. It's actually the answer that you um out of the all the different ways to achieve verifiable computing, I think that ZKML has the higher trust assumptions. Correct me if I if I'm wrong, but when do you think we it will be fast enough and cheap enough to implement it? Yeah, more massively. Thank you for that.

Yeah. So ZKML uh I think along with many of the other approaches uh does suffer from the fact that uh the performance uh is not uh as uh as fast as uh if we're not using verifiable computation but there have been a number of breakthroughs and the cryptography is advancing all the time and people are finding ways to fit that cryptography to machine learning in better ways all the time and so there was a breakthrough uh about 6 months ago from modulus clubs where they managed to uh verify uh the chat GPT2 uh which was something which was seen as a breakthrough because it had a a billion ways a billion parameters in the model um and so that that was a sort of a nice threshold to to pass uh and they managed to do that it okay it took 90 hours to create the proof of the the computation but still it it was something that was achievable uh and so I'm quite bullish about that I think we're we're moving forward and yeah uh in the next years, we're going to see the everyday sort of child GPT applications being verifiable as well. Thank you. [Music]