Hype to Verifiable AI: What It Takes to Run Trust-Minimized AI on Crypto Infrastructure | Will - 0G

Good afternoon everyone. All right, next session we have Will, head of Devril at Zerog, who's going to be taking us through how to build a decentralized AI platform. He told me not to say centralized AI platform.

Thank you. Thanks. Thank you. Um hello everyone. Um the the cool guys that are on the bean bags, they've left now.

So if you want, you can come you can come chill out at the front if you like. Um I am from ZeroG Labs. We um have built a decentralized AI platform, which sounds like a bunch of buzzwords, but I'm going to talk you through uh what that means um and what the advantages are of building on ZeroG Labs. Um are you guys developers or researchers? Put your hand up if you're kind of technical like a developer or researcher.

few. Okay. Um this should be interesting for you hopefully. Um so this is called from hype to verifiable AI. So obviously there's been a lot of hype around around AI.

Um I think over the past year we will probably all agree that it's kind of entered into a lot of our favorite apps and stuff like that. So even though there was a lot of hype, you know, we're seeing real world use cases for AI very very quickly. Um to be honest, this feels more like an AI conference to me than a crypto conference walking around. If you didn't know this was ETH Denver, you probably would think it was something AI related. So, I don't think I have to actually kind of sell you guys on AI, right?

You probably realize AI is a is a big thing. We've probably all heard of this. Um, what you can see on the screen here is that um in the last three years, as you'd probably expect, um the amount of money uh made in the agentic AI market, the the revenue there is increasing year on year. That's not particularly surprising and is estimated to increase to $9 billion this year. Um but you can also see at the same time trust in autonomous AI is decreasing down from 43% to 27% last year.

And at the same time um in the crypto economy uh the total number of uh dollars lost to crypto exploits is $3.4 billion up to 2025. Nearly $2 billion lost to malicious me. So if you combine these three things together it doesn't paint a very good picture, right? uh the use of AI is increasing a lot.

Trust is decreasing in autonomous AI and at the same time in the crypto economy there's huge amount of hacks. So if you combine crypto and AI together in the agentic world or the agentic economy, you know, this sounds pretty bad, right? We're going to be leashing these um autonomous AI systems, giving them access to payments and crypto and you already have huge hacks in crypto. So the combination of these three things together is kind of worrying um in the current centralized systems. So what zero G is what we've built is a decentralized AI operating system or platform or stack whatever you want to call it.

Um you can see here a bunch of words in a chart. Um these are the different components that we that we have built. Um I'll talk you through some of these uh more specifically. What you can see in general is that we have a chain. We have a layer 1 uh EVM compatible chain.

And then we have three other main components, storage, data availability and compute that are all fully decentralized and will allow you to leverage these components when you're building uh decentralized AI applications. There's two other kind of layers that tie these things together. One is an AI alignment uh layer which helps kind of keep the models in check and then there is a service marketplace that lets you actually bid and use these resources. So the four main components that I mentioned there, the chain, the compute, the storage, the DA. I'll talk a little bit about these uh specifically.

Uh the storage layer, the decentralized storage layer is fully modular. So you could use just this component by itself. You could build something on Salana if you like and still maybe I shouldn't say that here today, but you could use something on a non EVM chain, let's say, and still use our decentralized storage. And the storage is specifically set up for very very large amounts of data. So if you want to actually train a model and store the entire data set that you've trained the model on on on the storage then that is that is completely possible.

The costs are very very low less than 10% I think of of S3 um on mainets currently. So you can store large amounts of data at a very low cost on a decentralized system which is going to be very important for decentralized AI. We also have the zerog compute. So this is a a a component that gives you access to clusters of GPUs provided by different uh service providers. So it's an open permissionless marketplace where different providers can come in, they can provide GPU resources using our software um which will basically set up these resources within a TE and then you can um send out tasks to these different clusters of GPUs to train models or to run onchain inference or to do fine fine-tuning of models within those clusters and then the verification of those results from that TE is then posted and verified on chain.

So as you can see there uh in those two green boxes onchain inference and fine-tuning is currently live. So that means you can run prompts directly on existing models um and that is fully on chain. So you're interacting with the smart contracts to make that happen. Um you can also do fine-tuning currently. So that means you take uh existing models uh like GLM for example which we just released GLM 5.

2 too. Um, and you can take that existing model, you can run uh additional data sets on top of that to fine-tune the model even further and then you can publish that model um and then you can run inference on top of that as well. Uh, we will also soon be adding um support for pre-training. So this will enable full uh training of models from scratch on zero. Um, and then there's the the chain and the DA layer as well.

So the chain is what ties all this together. Every time you do onchain inference, you're running a query to a smart contract. Um, and the provenence and the proof of that uh that computation is then is then stored on chain. And we also have a decentralized DA layer as well. Uh, which gives uh 50 Gbits of of uh of speed and gives you the full providence of the data that you're storing on chain as well.

So you combine these four four layers together and it gives you all the kind of the building blocks you need to build decentralized AI applications. Um on top of that there are some other components uh you can use. Um we've uh developed a standard we call I NFTts. Uh there's an ERC and draft for this. Uh which which means intelligence NFTTS.

Uh these are basically a representation of the agents that you deploy on top of models. Um and they represent the ownership of those models. So if you want to deploy a model and then you run an agent on top of that model, you can just simply trade the NFT to sell that agent. So let's say you could use all of these components, right? You could use the storage component.

You can train a model um or you can use an existing model and fine-tune that model using a data set which is stored on the storage. You can then build an agent on top of that model, go unleash that agent to go do some interesting things. Let's say it's a trading agent and it turns out to be a ve very successful agent. Someone might want to buy that agent that you've then built on top of this model that you've trained. All you have to do is just sell the NFT on a marketplace.

Someone can buy that agent and then they own it. Um, and we also have support for X42 and American Fortress as well for private transactions. So, we launched um, it's been a staggered launch. We launched the chain itself in September. Um, I don't know why I put the money there just because it's a large amount of money.

We raised a lot of money. I don't I don't know if that's interesting to you, but uh yeah, that's a thing. Um, and we, yeah, we have a lot of money available to builders. That's probably more interesting. If you're a builder and you want to build something on top of the stack, um, yeah, we have a lot of funding available.

So, if you have an interesting idea, something something that's unique that's um, that you kind of couldn't build on a centralized system or has disadvantages on a centralized system, then we do have funding available for projects. Um and yeah, a lot of integrations. Obviously, um in the agentic economy, integrations are very important, right? If you're going to be unleashing agents, you need to have access to liquidity on other chains. You need to have access to oracles and things like that.

So, a big focus of ours has been adding as many integrations as possible. I also mentioned um an AI ali alignment layer as well. Um so, this is basically what will keep the uh the open models uh in check. Um so these are run by a separate validator set. So we actually have uh three different validator sets.

One is for the chain, one is for the storage nodes for the decentralized storage and then the AI alignment layer also has a separate set of validator nodes. So these validator nodes run our AI alignment software and basically checks up on these models the verifications that coming out of the models and checks that the the the output is basically within an expected format and they're not doing anything that you wouldn't expect. So, this is pretty similar to, you know, like checking transactions um on a chain, right? You're checking transactions, they're not going to do something unexpected. They're not doing double spending or something they shouldn't be doing.

It's a very similar concept to that, but applied to the output of AI models. So, just going to dive a little bit into why this is um important. Um you know, why why does it matter that we we can't trust current providers? What what is the advantage of actually having a trustless system? Um so currently models uh you know from open AAI, Anthropic, XAI um they are built on data sets that are unknown to us right they tell us certain things they tell us they use certain amounts of data and data that they tell us they don't use um which often then turns out to not be true but we don't actually know the data they train these models on.

Um, so this is an issue, right? Because if you have um, let's say you're an enterprise and you're using clawed code or something like that and it gives you some kind of output that you make business decisions on and this ends up going to court or something like that, right? You can't say that, okay, this model was trained on this data. It's all open source data. You have nothing to point to at all in a legal sense.

Um, so this is obviously an issue. Um, zerog fixes that by putting all of the the training data and the references to those data onchain, making them fully referenced onchain and verified onchain. Um, another issue is vendor lockin. So obviously, yeah, only three sort of main providers most people are using for AI, maybe four if you I guess if you include Gemini as well. Um, and from a developer standpoint, it's not great.

If you've built in AI, you'll probably realize that they just deprecate things when they would like. um models obviously aren't deterministic to start with and then when they're just pulling pulling uh features and things like that, it's not a great building uh experience. So if you build on a decentralized system, you can control the model and the data yourself. Nobody's going to remove that model. Nobody's going to make breaking changes to that model.

You've deployed the model. You've deployed the data itself. You've deployed the agents on top of that. So you have full control. Um no ownership or accountability currently.

I think that's kind of similar to what I was saying there about um the data that it's trained on. Um there's no accountability there in terms of which data is used. Um so if agents take actions that are malicious or break laws or things like that, it's not quite clear in the current system who's at fault. Um and how you aortion the blame to the different parties. Um in an open-source open trustless system like this is a lot clearer.

Everything is verified. every step of the the journey from training your model and running agents is is a full provenence and a history there. Um and with INFS as well, the trading of agents, there's some kind of identity that's linked there as well. Obviously, um the identity is on chain, right? It's an NFT, so it's not necessarily linked directly to a person itself, but there is a sort of a pseudonmous identity.

Uh centralized control and censorship. This is a big kind of issue. Um, obviously the main model providers at the moment are kind of competing for market share. So, um, they're probably not going to be redacting information and things like that too much at the moment, but you can probably imagine a future where there are dominant model providers. Um, and there might be political pressure and things like that.

They could be asked to remove things from the model. They could be asked to give certain responses to certain queries. Um, I think that's going to increasingly become a problem. And Zerog solves that by be permissionless infrastructure. Anyone can can deploy a model.

Anyone can train a model and run an agent. Um so yeah, it fixes that issue. Um AI applications, yeah, pretty fragile. Um at the moment, you know that most of these companies, they're losing, right? Even if you're paying $200 a month for Claude, they're losing money on that, right?

So um there's a lot of compute that goes into these uh these models and these from these providers. So it's not really going to be surprising if they start aggressively changing their terms, right? If they start um rate limiting you significantly, maybe in the future they don't like some of the prompts you're asking and things like that and they could start restricting access. So having an open permissionless system where nobody can actually shut it down is clearly going to be an advantage and misaligned incentives as well. um at the moment you know these these three providers are obviously you know losing money on on um subscriptions um but in the future if they end up being you know one or two dominant providers um and the whole kind of world around us becomes reliant on these tools things like claude codes as an as an example um then obviously the the incentives are not going to be really aligned between the people that are actually using this software and the providers themselves they'll be very dominant with these models and able to just increase pricing as much as they like.

So there's really misaligned incentives here. Whereas in zero G, the actual training of the models can be done by anybody at all and you can be using GPUs provided by any resource provider at all. So it's a kind of an open uh an open market and a fair system. Um so running a little on time. This is basically just a summary here of um the trust versus trustlessness.

And I think we would agree trustlessness here is going to be better. Um, this gives you a kind of overview of the stack. So, the thing if you're a builder, you're probably not interested in how all this stuff works and why it works, right? You maybe realize there are advantages to to decentralized AI over centralized AI, but the only part you have to really worry about here as a builder is the top part, the app layer, um, is building AI applications and using the service marketplace. The service marketplace is your gateway to actually uh, compute resources, to training models or to accessing existing models.

um and just building on top of that and building on top of the chain with us with solidity contracts as you might do uh right now uh in web 3. So the other three layers here, the the alignment nodes, um how the actual chain itself works, uh the infra layer, the compute and the storage, you don't actually need to worry about how these things work, right? This is just an entire stack that works together. All you need to do is deploy contracts. You deploy data to the storage if you need it.

Um and use the service marketplace to to bid for models. Um so yeah if you're interested in actually building trying out the stack um whether you're an engineer or not we have uh a zero coding session uh tomorrow uh which is a vibe coding workshop uh where we will kind of teach you how to to vibe code in an efficient uh way. Um it's pretty easy getting started with vibe coding. Um I imagine most of you here have probably given it a try but there are kind of you know better ways to vibe code and uh design patterns that work pretty well. Um, so if this is something you're interested in, then come along to the session uh tomorrow.

Um, we'll teach you about the stack itself and then we'll teach you how to vibe code. Um, and there'll be a little bit of a competition going on for like the best app. Um, and there's I think about $3,000 that's going to be given to um a number of teams uh for the best apps on the day. So yeah, if this is something you're interested in, even if you're not an engineer, if you're more in the product side or the BD side, something like that, this is still a session that's really going to be um accessible to you. Um and in fact to be honest I would much rather those kind of people um uh actually come along because we've we've been building in web 3 for a long time and it's mostly developers and researchers.

Uh we don't really have a lot of product market fit let's be honest. So um if you're a product engineer or someone like that then I think you'll add a lot of value to these sessions. So please do please do come along if you're free. Um there is a QR code here which I hope is big enough for you to scan. Um please do come along to the event if you're interested.

If you can't scan that, please just come up to me after the presentation. I can uh I can give you the link and would love to see you there tomorrow. If you have any questions as well about the stack, anything else you'd like to like to know, anything you'd like to chat about, I'll hang around at the back for uh the next 10 20 minutes or so. Um so feel free to come over if you have any questions. Thanks a lot.