Build an Onchain AI dapp with 0G's Inferencing I Gathin T

Welcome to this workshop and I'm glad to be here. So myself Gatin Deil at ZeroG and today I'll be giving a short workshop at this beautiful venue in Buenos Harris on the topic of build an onchain AI DAP with ZeroG's verifiable inferencing. So let's get started. So before I jump into explaining on what zero G is and what do we do? So the first thing that I would like to just give you a brief is like what the AI that we're using on a day-to-day basis like chat GPT is currently our closed AI.

So which is actually like a black box with agents model alignment data and compute completely being blocked blacked out in from the user's vicinity. Whereas when you use into decentralized AI that's where you could make this complete process transparent verifiable and also bring in few mechanisms with which the users can make sure that the data was not tampered or it was not any there is no censorship resistance that are being impulsed into the data markets. Moving on to before I jump into our architecture I'll just give you a quick brief on what Zerog is. ZeroG is the first intelligent layer one for onchain AI that provides complete infrastructure stack for building AI powered applications uh onchain. So with that our complete stack you could like developers can leverage zero's four core components.

So our stack is built with a EVM compatible infinitely scalable chain on top of which we have the zero's data availability layer. we have the zerog storage and on top of which we have the zerog compute. So the chain lies like a foundation for this complete stack on top of which we have the decentralized storage network which is an AI optimized storage with ultra low cost and verifiable verifiable permanence for both structured and unstructured data. Other than that we have our compute network. This is like a trustless AI inference and compute which is secured by decentralized infrastructure and cryptographic verifiability.

This acts like a foundation for you to bring complete uh AI models on chain. So moving on to just giving you a brief with the architecture that's that there's like a deep in hardware which lies as a base on top of that we have the infinitely scalable chain and with the infinitely scalable chain and us being EVM compatible so we have better a better like block time which is less than 500 milliseconds and we can achieve a throughput of 11 ktps which is extremely high and which can is capable of bringing the AI workloads Moving on to the compute network, storage network and data availability layer. These are other three pillars of our modular stack. So just to mention that we are a modular chain. So you could use individual components separately or you could utilize the complete stack as well.

So with the computer network you could do training rag and you could build deploy agents and we also have like verifiable inferencing that is like I'll just dive deep into it in the upcoming slides but just give you a brief that lies on the service marketplace is that we have uh models like uh quen we have uh GPToss and we have llama. So we have few other model providers. So you can get verifiable output which where the like the complete output of the model is processed in a trusted execution environment. Moving on to like the data availability we have high performance and provenence uh supported data availability layer and this acts and provides like a off-chain verifiable storage for all the data transactions that happens on the chain as I mentioned. So we are able to bring the infinite scalability and also give better throughput as we have separated the execution layer and the consensus layer of the chain.

And on the execution layer we have the state VM and the smart contracts. On the consensus layer we have the network agreement and finality. So you could scale individual components and you could uh bring in much more throughput for the chain and we have achieved with our recent mainet launch of like less than 500 millconds of block time and also a highest throughput of 11k TPS on like moving on to the next component we have the zerog storage. The traditional cloud storage is fast but it is expensive like Amazon AWS that you use on your day-to-day basis and other decentralized options like RF filecoin and uh uh IPFS they like they have like their different uh limitations but with zerog we are trying to achieve this complete uh complete uh solution which with which you could like upload large scale AI models and we have achieved like uploading a 100 billion plus parameter model on chain by using our new framework which was developed by our research team which is called as DLCO X with which you can upload large scale stream uh large scale models on zerog storage with just office network bandwidth which is insanely fast for like a decentralized option. We are claiming that we are 95% much cheaper than other AWS solutions.

You can get instant retrieval. We support both structured and unstructured data. So it is universally compatible. You could you could just think of it like a decentralized Google drive or you could just think of it like an S3 bucket. So you could structure you upload structured and both unstructured data to it.

On the technical side, how we are able to achieve this throughput is by having like two different lanes with which you will be able to upload uh files to the zerog network. So the current when you upload a file to the zerog network, there are two different parts that happen simultaneously. One is the data publishing lane. The other is the data storage lane. So in the data publishing lane uh let's start with the data storage lane.

When the file uh then the file is uploaded to the zerog network using a mechanism called as eraser coding it is split into multiple chunks and it is then distributed into the storage nodes on zerog chain. So when it is distributed to these nodes, what these do nodes does is they verify the data and they provide a signature to the contract and all these aggregated signature from multiple such nodes is cumulatively created into a Merkel tree and the Merkel hash root hash is uh verified and like stored on the contract in ZeroG chain. So which is in turn verified by zero's consensus. So zerog chain acts like a base for uploading uh AI uploading data to the zerog storage. Just giving you a brief about ZeroG's uh data layer and the main challenges that we solve compared to other data availability layers is that there is limited horizontal scalability because uh there are like there are like big requirement when datas are sent simultaneously to the network nodes.

This creates a bottleneck and limits the speed to the slowest node. And also we have an inbuilt storage network which improves the latency and throughput with rather you relying on like a third party storage and we are relying on like Ethereum security. So we are pretty much stronger than the existing data layers and we have better randomization compared to other existing DA layers as well. Here are a few use cases. uh you you could find a few common ones that you could use other DAS as well like using it for offchain uh storage for like L1's and L2s uh having like decentralized shard sequences bridges roll up as services defi onchain gaming and data markets but one thing we specialize in and we are happy to enable is that it is like a great option for AI and machine learning for example when you're training a model which is used after like several months by users and they want to verify on like what data it was trained on, how it was trained on and they want to get access to the historical data.

They will be able to do that with this offchain offchain with this offchain place called as a data availability layer which is affiliated to that chain.

Okay. Moving on to the most important and like the core concept of like a brief uh zerog's compute network or like the service market on top of which so before we jump in like the issues that I mentioned earlier we had the others face like high cost complex setups they are locked in into vendors but with zerog compute it acts like a decentralized marketplace where the GPU owners sell their computing power uh to developers in need and it's like kind of like Uber for AI computing. The only issue is that you get vendor logged in there are like rental providers but we are like 90% cheaper than and we are like payer use uh pricing models that we have. So this is a brief architecture diagram that we have. So if you want we have like a provider setup.

So initially we can uh act like a provider by uploading your own model and like connecting it to the T that we have you uh the requirement that we have is that you need to have H100 or H200 GPUs and for the settlement it makes like a ZK proof and it settles back on the contract on zeroG chain from the user side how they interact is that there are different inference providers. Users can select one of this. Users don't need to pay for each transactions or sign them. Instead, they could make a deposit to a smart contract which is called as like a ledger on ZeroG uh computer uh service marketplace and that will autodebit based on like the user's request for like multiple models. This is the same brief that I mentioned earlier.

So users deposits funds, they request services and which can be like send your AI task, they get the results uh which they get like output from the best available compute GPUs and automatic payment takes place for GPU owners. They can register their hardware, they can list their GPU specifications, set their prices, process requests and yet instant earnings. That's with like the complete architecture and for this hackathon we have this special bounty that it's like an open track and we have like a total bounty pool of $10,000 and our topic is to build launching adapts with zerog's intelligent layer 1. So we are looking for projects that can uh build projects that can be only possible to build using zerog by utilizing this complete stack for onchain AI. So it can be AI agents with verifiable computation and permanent memory.

It can be decentralized ML training platform with proof of contributions. High performance storage solutions which has like an AI analytics part embedded into it and also like having like offchain storage like a DA layer connected to it as well. So we have a EVM chain, we have a DA layer, we have a storage and also we have like inferencing that is available for you to utilize. And one bonus that we have is that the projects who complete deployments on testn net if you can come over to our booth and just uh show us a brief on like your deployment we can help you migrate to zero g's mainet by providing necessary tokens and even like for test net if you need additional faucet support we can help you that with that as well. So we are looking for projects that builds things that are impossible elsewhere with other chains.

So that is what we would like to achieve with this hackathon and with you fellow builders. We have built this beautiful builder hub which acts like a one-stop resource platform for you to build and uh acts like build with zerog and you could have like look at the documentations we have some tutorials you have SDKs and also the necessary toolings that you would require to build with zero g you just give begin journey you have the list of all the tools faucets scans like chain scan we have like a scan for like storage uploading as well. We have a validator dashboard, RPC if you need RPCs. We have our own RPCs or you could use our partners as well like quick node, third web or like anchor. We have SDKs and we have our starter kits.

So with this you could get quick access by just uh dep like just pitching it into your CLI and you will be able to start building immediately. So we have like the template ready for you guys and we have the code links for the documentations and we have like a brief tutorials and office hours with our partners as well. And this is going to be like one of your most favorite places or it's like the most favorite one of mine as well because we have the list of all the projects or like community build projects. uh it's like a big collection that we have and we have it in a repository format which is called like a awesome zerog repository and this acts like a template and base for you to see what projects have built in the past how they have utilized zero. This can solve like 50 to 90% of all your doubts that you have or to see like what they have implemented and it's been a year and a half and we have like pretty much all projects from multiple domains as well.

So you could find the base template or like a small ideas from these as well and we have our documentation which is you can head over to docs.zogg.ai AI and you could head over to like inference SDK and the comput SDK that I mentioned over here you could fetch and start uh spinning it up uh you could fetch and spin it up locally and you will be able to access all the models uh that are available on zerog's uh inferencing network. Uh so this is like a link to all the zerog resources this is like a link tree for zero g and if you want to connect with me you could scan the other QR as well and if you have any doubts for the next 3 days myself we are uh Kumar uh we have yon we are like de at zerog and we our team will be here to support you throughout this journey if you have any questions please head over to our booth and we will be there at your disposal. Thank you so much.

Have a good day hacking and uh it's been great so far and looking forward to see what you guys build over here. Thank you.

Oh yeah. If you have if you guys have any questions, let me know. Yes.

Awesome. So if you have any questions, please do like visit us at the booth and we'll be there to help you with that. Thank you so much.