[Music] Hey everyone, welcome to the session. So myself Gatin Deell from ZeroG and I'm super excited to be here to just give you a quick brief introduction about what Zerog is and how you can build onchain AI dabs with zero g storage and compute network. So before we jump into what zero g offers like what is the issue that you currently face with closed AI closed AI currently acts as a big black box with like the agents model alignment data and compute and you aren't able to get much awareness about like what is the being data that is being used for training like how is the alignment that is being used like what are the agents how are the agents are being deployed the inferencing like which are the server providers that are used so this is a big thing without like trustlessness aspect and also there is like no much uh clarity on the consensus that is being used but when it comes to decentralized AI you will be able to completely decentralize the stack of you from uploading your custom data like using them for training using like the data and labeling to having like verifiable inferencing. So you will be able to completely decentralize this whole stack with using zerog's uh chain. So just giving you now a quick introduction about what zerog is.
So zero is the uh largest layer one chain and dedicated for like onchain AI with powered by the decentralized AI operating system which provides like a complete infrastructure to build AI powered applications on chain. So our stack comprises of these different layers. So we do have like there are like four core pillars. So we do have like a layer 1 EVM compatible chain. On top of which we have the storage network running which is like much faster and much cheaper than the existing storage providers whether it may be decentralized or it can be much other centralized providers.
So this is what offers and provides a platform for you to upload the huge AI models and all the training data that is need to be used for training it on chain. Next comes the data availability layer which act which is like infinitely scalable and it has like high speeded data availability to provide support for like the verifiable AI workloads. So that brings in like a good off-chain storage for your large scale data blurs and also like it gives you like an offchain storage which for which you could use it for later verifiability purposes when it comes to uh when it comes to like verifying how the data is being trained or like how the data was used before the model was trained. So next comes the compute network. Here you will be able to get verifiable inferencing.
For example, we use TML to verify the data which the model gives as output. So basically the TML for example, you just given the input to your compute network and the compute network process it with the models that are in it and the final output is being processed in a trustless execution environment and that is given as a verified output to the user. So on top of which we have the alignment nodes. Think of it in a future scale where like lot of your like service supply chain market supply chain is completely run onchain and you need some control over your data and like whether like the zerog nodes whether it can be chain or the storage or like the compute is performing correctly. So we did a node sale with like which is called as alignment nodes which gives the power to the users if they can just slash the nodes if they if they find some malicious activity in it and that gives the power of decentralization in the hands of users.
On top of that we have the service marketplace which acts as a decentralized hub for AI models agents and services and uh this acts like in the future we'll be adding like rag support. We have the fine-tuning available and also we will be able to get uh different GPU providers for you to upload your own models as providers on top of zerog's inferencing. Just a quick brief about what zerog chain is. Zerog chain is advanced. It is like a model layer 1 which is EVM compatible.
So this architecture provides flexibility, scalability and a good innovation in terms of like how this whole transaction execution take place. And we have separated the consensus layer and the execution layer. So consensus layer is where the network agreement and the finality and the block protection takes place and the execution layer is where the state of the transactions, the VMs, the smart contracts are deployed and the logs are being sold. So we have separated these both so you can scale them individually and there will be much more better throughput in terms of like interaction between these layers and this is like a pretty like this is was what was available for the test net and currently we have launched a mainet like four days ago and you are expecting around like 11,000 plus uh TPS and like sub uh sub like 100 milliseconds finality and the consistent throughput as well. Moving on to zerog storage.
So this uh what is the issues with the traditional cloud storage? It is fast but expensive and you need to be logged into to the provider. For example, you're using AWS. But moving on to decentralized options, we have like IPFS which is comparatively slow. Filecoin has like much more limited architecture and RVR is is like much more better but it's highly expensive.
Whereas zeroG storage we set a benchmark by providing 95% much lower cost instant retrieval rates and it we have the support for both structured and unstructured data and most importantly we will be able to handle the workload that you need for uploading large scale AI models. The technical architecture behind this enablement for the zerog storage is like we do have two different parts. One is the data publishing lane. The other is the data storage lane. So the data publishing lane is where uh so I'll just come with data storage lane.
So basically the data that you want to upload to zero g's network is split into multiple smaller chunks which is then uh split up and sent through the zero storage network and this data chunks are randomized in terms of storage using verifiable random function and simultaneously. So as the different storage node receives the data chunks, it will sign the transaction on the data that is uploaded and it will send it back to the zerog smart contract and that a merkel tree is created which gives you the complete structure of data. So if the data is stored in number four, number seven and number 10 of the storage nodes. The Merkel tree gives you a clear clarity on how this data must be aggregated to get the final total throughput uh final uh whole data when the user wants to retrieve it. So this is completely verified by zerog's consensus.
So basically the eraser coding and the distribution among storage networks takes place in the one takes place on one side and simultaneously we have the smart contract creation and the Merkel tree with the Merkel proof of different data chunks uploaded to the zero g's contract which is then signed and approved by the zerog's consensus. Then moving on to ZeroG's data availability layer. Do we have what is the issues that we face with the existing data availability layers that brought in a requirement for zero to bring this up as well to the stack to enable it for like the complete onchain AI. So the limited horizontal scalability so uh many uh system needs all the data to be sent to the network nodes simultaneously. So there is no no like parallelization of data which creates like bottlenecks and it limits the speed and the slowest node will not be able to cope up with what the fastest node achieves on and most of the data availability layers lack just lack the usage of built-in storage.
They either have like a different third party storage that they use and it brings in connectivity issues throughput that is like latency impact you need to have like additional cost if you are using a third party provider as well. Moving on, they have like weaker securities whereas in zero we use Ethereum security standard but a few other competitors have their own security mechanisms which does not uh as is is not as rob uh robust as Ethereum. Then there is like some uh like kind of like a D have like in insufficient randomizations. So there is a lack of randomization for data committees which potentially allowing uh collisions among different malicious node in the network. So this is the architecture diagram which I explained earlier in terms of like the data storage lean and in terms of like the data root and the to data route and the commitment that is like adding in new data to the network.
What are the different use cases when it comes to zerog's DA? It is similar to what other DA does like you can use it for like layer ones and layer twos. uh you could use for shaded sequences, bridges, RAS, DeFi, onchain gaming and data markets but we specify exclusively for AI and machine learning. So this is suitable for you to upload vast data sets which can be stored on chain and it can be accessed for use utilizing for like AI uh DAPs which deployed on zerog chain or you could deploy uh elsewhere and you can just utilize zerog's DA since we are a modular architecture. The last uh part of the pillar of zerog's network is the zerog's compute network.
So this brings uh this kind of like cream on top of the cookies. So we have all the necessary infrastructure that is needed for you to upload AI models, train them on chain like provide necessary storage for it. But then now what we need is like a final endpoint in which users will be able to utilize the models that are deployed on zerog chain. So that comes into the zerog's compute network. The existing uh the existing uh inabilities of the other networks when it comes to like computation providers is that they are high cost.
They have like complex uh setup because you need like to configure like cloud GPUs and you are locked into the vendors. But with zerog's compute this is a decentralized framework which will offer AI computing capabilities to the community as part of like the DAI operating system. So this is a brief architecture on how it works. Think of it like there is a beam meme uh meme generating model. So basically the there are like multiple models and each model has multiple inferencing.
So basically the user needs to just deposit token directly to the zerog's contract and he'll be able to use any of the different model providers and he'll be able to select the inferencing that he needs to use. It's not like specific to like what chbd offers that you need to definitely pay $20 to get this throughput. There are like different uh for example one person has fine-tuned it so accurately that you get the specific thing. So he charges the inferencing a bit higher and the other one just have like a generic fine-tuning over the meme generating model. So he charges a bit lower charge.
So there is a flexibility for you to choose the cost the requirements that you need for different types of inferencing on top of the AI models. So with that the user will be able to interact with the inferencing and when he interacts and gives in the inputs to the model the all the data that is being uh through processed happens in a trusted execution environment. So for example you get for example you ask a centralized AI model provider and you ask uh zerog's compute network uh the output that for example you're asking like a panda image the image that you get out of the centralized provider gives you with a Nike logo for example and with zero you just get an accurate data for that implies that that data is being tampered for example they would have got billions of dollars from any centralized organiza organization providers to just pay for uh having that you can just think of it like ads. If you ask like which is the best shoe in Chad GPT if you want the result to show up to this particular X brand you could just pay up charge GPT like a centralized provider but where whereas in zerog the you it completely relies on the model data. There is no tampering because it happens in the trustless execution environment.
And the final ZK proof of this complete transaction is stored on chain in the smart contracts. So for the AI users, this is the workflow like they deposit funds, they request the user service and they get the result from the best available GPU and uh the automatic payment of the necessary inferencing usage happens. But for the GPU owners, it'll be like for you to register your own GPU hardware, determine the price for the computing power, and the people who wants to utilize your own GPU can use it for their own models. And just just giving a quick conclusion, I know we are like almost on time. So we have this Zerog's builder hub.
In this we have all the resources that you need to get started building on top of zerog and just to give a quick context we are facing a slight downtime in our test net right now. So we encourage you to deploy with local testing like you can use hard foundry and you can ganache for testing and we provide necessary mainet tokens for you to deploy and test things out. Uh so also like we are encouraging people to build in this hackathon for like onchain dev tooling. So here we have all the toolings, SDKs, documentations, tutorials. Uh we have like a showcase of all the community projects which previously built on zero g.
And uh for this hackathon, sorry. And for this hackathon, we are offering like a $10,000 bounty and we are encouraging the builders to build what an early ecosystem or like a layer 1 chain needs and currently we are asking them to use our partner inferencing providers and later they can migrate to once like when zero g launches its compute on zero mainet. So we did our mainet launch on 22nd just four days since our mainet NTG and we are looking forward to see what builders are planning to build in terms of like dev toolings explorers analytic platforms like D5 protocols like memecoin launchpads NFT marketplaces and so on. So if you want to like uh look learn more about zerog here is a link tree for zerog and if you have anything that you want to ask or connect that is my telegram and you can find me at the zero g's booth and if you have any questions or if you want to just come say over hi if you have any doubts just please come over and uh thank you for being a wonderful audience. [Music]
Automatic transcript — names and jargon may be misspelled.