Scaling AI on Ethereum: Data Availability and the Private Computer

Transcript

[Music] [Applause] Thank you so much for joining the fan every know uh waiting for the last panel. So last is always exciting. So this panel is uh we'll go deeper into uh what is the issues with AI agents and AI just being you know uh the bottlenecks of AI through uh on on Ethereum and as the AI models and the data is going on chain what are the issues we are facing and how are projects like you know uh ZKM Aztec Surge and Huaw looking to solve for it. So let me start with my first

question about like this question is for everyone just to understand according to you what is the biggest bottleneck for enterprise grade AI on Ethereum uh could it be DA privacy compute or something else all together so start with Ming and then we'll go from there >> right we we're actually building um infro so I think enterprise is one of the um core um audience or um use case um because in infra you deal with uh uh they're very stable uh providers of either compute data or um you know a large user base um so uh from from Asian point of view uh we used to think uh you

know in a different term automation so um I would actually think the agent we used in this scenario um actually should not or do not need to have agency. We call them agents as if they're humans but uh they should have their own agency. Um you know the the model quality are getting better today with less hallucinization and with um you know further fine-tuning and um you know generally the improvement models uh you have a very good uh automation uh agents and and they're supposed to

have no agency no hallucination they just understand language um direct directions directions directories really well um precise and that's the agent you want especially in the onchain where we will deal with assets money um finance yield want your agent to have agency um no other purpose hidden agenda just um authentic uh directive following ability police.

Um yeah, that's uh I think that's what's important to us and my understanding of um agent centric world. >> And what's the bottleneck? What's stopping us to get there? One uh >> I think it's yeah, I think it's still the uh quality of the model and also the cost of uh you know tokens and all that. Um I I would imagine the future uh models like are getting better uh each u enterprise level. Um yeah, you could probably contract with open eye or surfic. They have all the specific design for you. But eventually you'll have your own model and your own um view centered model that you train with your

own you know data private um data. Yeah. And that suits you better and also um can also protect privacy of uh your users. your models are not not supposed to have any sensitive you know private data in it. So you you will have a better fine tune control of your data and your users privacy. Um and uh from an energy point of view you probably will have your own um your own arms of u discrete energy you know solar panels everything. Um yeah I think that's that's the way to go you know just connect it with your uh data center. um cooled by you know um natural um you

know resources and natural energy that you produce yourself. >> See question. So what is the bottleneck for enterprise grade on on Ethereum? Um, in terms of agents or just like Okay. Um, I think maybe a controversial one, uh, is probably identity. Like you're going to need to know, am I interacting with a person or an agent? And I think for it to be useful, there has to be a lot of advances on uh, I guess proof of personhood. So you can tell an agent needs to be able to tell is it interacting with a human like EOA that belongs to a human or an EO that belongs to an agent because it may respond differently um depending on that. So I think yeah strong identity um and I

guess privacy around that on on chain. Um so talking about like Euros scalability is always like the number one key issue people are talking about you and we know majority of the team are working towards um the TPS like transaction per second but there's like another kind of like scalability which is the onchain storage and why that's important um currently uh when we talk about web 3 AI AI agent so usually we do all the modeling stuff offchain it's just like a black box and we don't know like how what's a process and whether something got change and when it come to like a result we put that onchain and we

call that so-called like web 3 AI um and the reason we don't do everything onchain is because of the cost because of the onchain storage capacity is not big enough um so I feel like that's the bottom for the real like web three kind of like you know onchain mode or something. So that's something we're working on. Yeah. >> Hi. Yeah. So I think fundamentally Ethereum was not built uh to accommodate the um AI use cases. I mean it could barely hold up to like you know simple transactions coming in like if you just top it up with some programmable logic it was too much for it and hence we we

we had like L2s. So I think uh if you start to figure it out quarterly is like a bunch of in order to to start with uh but if you need to put some sort of verifiable aspect then we need to have certain like so decentralized aspect to it that's where Ethereum comes in but uh if you have to break it down then I think both compute storage uh you know reachability how you can reach to a a specific resource, how would you define a resource? So I think AI in itself uh is very unique in that sense because if if you have to make a think of like an an

address like an AI agent should have an address. So this needs to be somewhere registered. So I think that's how a ERC84 comes in or that's just like you know like tip of the iceberg because then we will we we will further move into the compute and storage and storage is not like you know your transaction data or your called data from smart contract but it's like huge amount of data like raw data in that sense.

you got very interesting point of view from quality identity uh to storage and generally so I'll like to dig deeper on matters and e storage how you are trying to solve the problem of uh uh cost effectiveness versus reliable data on chain so we'd love to get your views on this >> yeah sure um actually uh I think we have a we have a general view uh just that devolved with the the you know evolution of Ethereum. Um we started uh lay two um with optimistic that is as the first optimistic uh network called drama we're like a month

or two even had a results uh this not this posting and now we're also you first saw over z wall so z map developed zyron which is a u mid r2 it's a open source version of uh based uh zop Um and uh on Ethereum I think the you know the impossible the development uh triangle I think has mostly solved u you know it's never to be a physical law uh to begin with it's a challenge for the community and I think uh the community the dev community you know Ethereum has the best dev community uh in all of web 3 all of

blockchain space answer the call and We single-handedly pushed um you know zero knowledge uh cryptography like you know exponentially like the the past couple years are you know uh are like past 40 years combined we used to pass that now we have real-time proving uh we have you know basically are very usable and cheap so it's not adding much more um uh cost gas fees and uh and the implication of this hasn't been really uh understood. So basically with ZQ Wolf lay their tune you can have infinite uh

scale level uh that's what we call like earlier even the OP time uh this is called horizontal functional scaling so you use uh layer one as your consent consensus settlement chain and uh all competition offchain and uh with this you can also contract uh DA uh with um as also a separate blow up if you have hashes on on you know on on the layer one. So I I don't want to go into all the tech details but I just wanted to to uh challenge the developers that um um you know this is probably only the beginning of the Ethereum

uh kingdom or empire. The first you know 10 years are just the the building stage. Now it's the proliferation stage. It can really be uh a space that people can build real ass you reach the new users not just the creators not just the speculators reviewers that's normally that don't really care about web serial web two they just care about usage and um we will have the best UX real time proofing will have uh milliseconds level uh response time for any app um so yeah of course on AI uh applications as well So yeah, I wish everyone answer uh you know the the call of like uh care apps

on Ethereum again. You know like the the you know we all remember the time like you know crypto key was a trade app right and now we need um real uh apps for movies and the all the hard problem before uh you know central ID exchange everything can be everything every empire every on the foundation built by Ethereum uh dq of uh zero knowledge virtual machine and um decentralization security and um interview us at the same time. So I'm very excited actually.

Um yeah I remember so vital like was being asked what do you think the future of um um eer so is it just like a financial kind of like tool or um applications use case or what do you think and he's like I really hope can be more than just finance it can be something like work computer that means uh except for the defy there will be versatile of other applications for example like the AI we're talking about right um to achieve that goal that means um instead of the TPS transaction per second which is um you know one scalability so the onchain storage will be another kind of dimension uh to help

achieve that goal um for us uh we are super module flexible so we can be a layer two for the layer one we can also be layer three for other layer two so we're not competing with all existing LT like you know everyone here um actually we're working very closely with us um why everyone is helping to you know achieve um you know phenomen like dimension so you will have more on chase um one common question I got the most is what's the difference between you guys and firecoin um IBFS or um arbit um basically They're like single blockchain

outside of Eer ecosystem and more for the code storage and as you guys know if we do want to interact with smart contract like AI model this kind of like real computation you need to have the dynamic like storage which um it's a kind of different and also uh within the eeran eco system um so there's no uh big like interoperability problem like crosschain problem which is really headaches. Um yeah so that's how we position ourself. Um basically we want to help um all the D the apps within the isering ecosystem or the layer one or layer two um to achieve um you know the

onchain storage and lower the price by a thousand times and give them the PB level yeah onchain capacity. I think that can cover majority of the use case especially for some uh mid I mean midle like big model if it's like a super big model definitely no and I'm I curious that should be on the blockchain yeah but for the small to medium size PB level should be covering love yeah >> I think people on you introduced the deep net in his money which has USDR stable coin flux as your uh verifiable compute work. So how is that you know

the framework is making AI private and scalable but also auditable. What's your journey and what's your take on uh uh making this like this would enable the AI to have the uh enterprise grade level. So curious about the framework and how are you thinking about it? >> Yeah. So I think it's a pretty complex question and and it so I I think like I was saying that you know the traditional parts of DA will not solve a lot of questions but now we will we have like the data availability layer that actually caters to uh the AI use cases.

So meaning all the AI data which has been generated should be stored somewhere. uh what do we do with this type of data like initially even if you put some data on Celestia it's going to give you like the data availability guarantees it does not give you the syntaxical or semantical guarantics and even in AI use case the uh semantic can differ a lot so at so we at HU so we sort of imagined this thing that hey imagine In the case of AI a agents they have collaborated on a task. So this task has generated high quality causal

contextual data. Even within agents this data can be like considered as semantic sorry synthetic data in a sense. So with this data so one thing happened for sure is it was generated against some input given by a real user. So how do you sort of audit this flow of data that hey it actually has done something useful. So that usefulness is again very subjective. So we we are going to use this opportunity to actually put some value to it.

Meaning it's not like Bitcoin we you you burn a lot of compute storage uh to actually found like find a noise. So the thing is you actually still use all your infrastructure your your tokens uh to actually generate a real outcome a valuable outcome that a user has has like literally paid for right so if you look at this way it would be transformed into a resource it could be like an AI native that's what we call as flux like hey you go to an exchange you might like you you you solve it against something no it's actually proof of co like it's it's like a proof of work to be honest

that AIS are doing work how good of a work they are doing it is uh or the HU protocol to actually manage that so the more good work you do the more native token that you end up earning in the case of Bitcoin it is just you know validating transaction signatures and stuff but at the end of it to to generate let's say an anchoring proof move on the chain. It has to do some sort of a use useless worker in that sense. But which is required for you know network security in that sense but in in this case is actually going to uh take your intelligence verify it in a way and then give something back to you

which you cannot get anywhere else. Let's let's jump into privacy now because uh AI by nature you have we are processing sensitive data as well as because it's under blockchain it has to be verifiable publicly. So would love to go deeper into Aztec's approach into uh how are thinking about AI data privacy as well as execution and then also to dig deep into ZKM's approach which is more general purpose. >> Uh so yeah >> yeah I think there's there's two things that need to be made private in in AI.

is the kind of I guess the sensitive weights of models. Um if you just put a model's weights on a smart contract, um that's potentially someone's IP and usually people charge for that. And you know if if it's just public in a smart contract, you can just clone it and make make a new one. So I think that's one area that needs privacy. Um and the other is obviously user data. I think the the first problem is probably the more real problem in the short term because if you want to take a model and put it on chain like you're going to have to wrap it in a Zika stark or some sort of program and when you do that you have to do inference or some sort of machine learning in a zero proof and the real bottleneck there is just proving

systems. They've come a long way in the last four years, but they're, you know, you're not going to be able to kind of get like uh very advanced models into a proving system. Maybe you can do some sort of compliance check. You can do like a face match, but you're not going to be able to do like a large language model anytime soon in a in a proving system. And so I think that's kind of an area where we need to see a lot of development for this to actually be feasible. um on the user side and kind of like context side I think yeah the training data is kind of another area and if you yeah I'm not upstate on research here but you'd need to be able to kind of let a model be trained on encrypted data um to not leak secrets to

a model so I think those are the two the two open problems we have right now >> and how are you looking at solving with >> uh for Aztec um at the moment we're focused mostly on kind of like execution um and kind of I guess bringing privacy to normal Ethereum transactions. I think um I think it's quite far off to actually have inference on chain. I think the way we'll kind of see things in the next one or two years will be agents in interacting on chain but the AI is offchain. Um so I think we can bring privacy to what's happening on chain. Um we can bring identity on chain. We have things like proof of passport so we can start to solve the problem I talked about earlier. Um but yeah, in terms of kind of actual useful

like machine learning or or like models on chain, I think that's a few years off. We've focused on payments, DeFi, um making the rest of Ethereum private. >> Yes. Uh so since we're you know we're doing the ZQ so uh the idea is not to uh put model or whatever like model data weights on solar weights >> um they will shrink right so that's a consensus that's the about seven layer and um um there are important um things related to uh blockchains that need to be you know you know put on chain to settle to to reach consensus. I think

that area is um um explanability uh interpretability like you you need to know how the decision is reached. So um so in that regard for example in our case you can probably uh do some inference and you have a trace and then this acting trace would also with parameters of a model which is offchain that you can kind of have a audible trace of you know which layer activates the the kind of uh result we got right and that's called the uh mechanistic interval and if you have that completely contained with the aspect

trace and uh and that can be settled uh you know in a in different way layer one so I think that's the biggest um thing web3 can do for accountability insurability those are important things for trust uh for uh but the thing that's important um with the nature of web street so that's I think um that we can uh do a lot move to the last question talking about the future of uh AI and AI on Ethereum. So according to you I mean we can start with uh Abby and then we go from here.

What is the most like transformative breakthrough you're looking forward to in your domain or could be outside domain just to make uh uh onchain AI uh like more transpar to the next level basically. I mean I'm I'm very bad at predictions but with the current flow that things are going I think the one one of the biggest decentralized AI which actually flies very close to Ethereum or you know web 3 narrative is zero g I was actually looking at their you know presentation at SEC this year so they still have a lot of boxes yet to be checked out. So I think uh training a

model is fine is good. You can do it in a decentralized way but it is still like five to 10 times behind the web two. So go book. So whatever is web 2 is doing u web 3 is doing it way too much slower. So I think u in the future I would love to see this gap sort of closing it down and on the agentic side I would like to see no something very similar to like langraph mansion or phoenix agentic launchp coming to the web world as not just you know hey here we are just giving you uh capability to actually launch AI agents but you have to use it in our own silos no it should be permissionless in that way So be it like

if you are running an an Ethereum validator you should be running your own you sorry you should have the ability to run your own DI agent from anywhere you want no lock into the infrastructure of the web completely permissionless in that sense >> um I still remember so last December there's like um a hype for AI agents and there are like so many tokens generated and what they do usually usually is there's like a web 2 AI agent there's a token and that token is how we connect with the web 3 that's it um I feel like anything we can do more than that it will be a plus like I said um not only

just running all the process or AI agent you know interaction offchain and only have the data available and put that data on chain only got the token that's the only connection with the web Okay. So anything more than that will be great for me at least. So that's why we are at least provide another way of like onchain storage. So they can run the mod's own chain. Um so definitely um like you mentioned before there are like so many gap between web two right uh but at least I want to see a little bit like you know difference compared with what we already got. I don't want people to

say uh okay the only thing we have is just like a token that's it. >> What's your use case of AI? >> Yeah. >> Yeah. Maybe I think two things. One is um on the data approach there's a Ethereum hard fork in December for Kushima. So we get a lot more blobs there. It's not going to be enough to um kind of put whole models on chain but more days on Ethereum is definitely the right direction. Um the other kind of thing I think would be not specifically AI but um more identity standards. Um so there's things like obviously proof of humanity with Worldcoin, the CK passport, but there's not standards around it right now. And so um if we're actually going to get, you know, agents

on chain, we need a way to establish who's human and who's an agent. And I think we need, you know, we have uh was 804 the agent kind of like EIP. We don't have the same for humanity. So I think we we need that. >> Let me finish with something like a Zen like right Z master would say uh the best u the best web three is no web three uh that it can be interpreted as um I think it's the most important thing like we all know uh is UX. So web three, you know, blockchain used to be lacking in that in that regard, right? Like 10 minutes a block, um 12 seconds a block is still too slow. And uh and you can't

just do I I know there a lot of uh projects claim like like a million TPS in 2019. I won't do because consensus needs time. You cannot you know propagate consensus faster than you know network speed. Um so there's a naturally there's a there's a limit to the kind of DPI you can process on there one. Uh that being said uh you know Ethereum has pushed the envelope and uh like I said earlier most of the challenges of scaling has been solved. So it's time to actually build web3 application as good as web 21 and uh so you know um our goal is to uh build an app basing what you call normies that don't know any you

don't need anything um web three you don't need to know there is a web three behind everything it's like today like um um I I I try to build like web apps some time ago and um the one I built I still remember like client server like from long ago and um mine is just so slow and then I realized it's it's so evolved right now. You have CDN and you all that yeah you can't just go back the old way. So I think the same for for this new generation of web apps. Um besides um you know you look at new markets example South America Africa where uh their web uh from the beginning they don't have bank accounts for that matter they don't have sometimes they

don't even have electricity and all that they've never some somebody never like um used a landline right so with them it's very natural to have web 3 at their first um you know like kind of experience so yeah I think the yeah let Let me let me feature that like yeah >> thank you so much for your insights and please have a round of applause for all the panelists and this thank you. Bye.