AI Without Limits: How ZeroG is Building the Future of Decentralized AI | KBW2025

Good afternoon everyone. I I'm going to dive straight into this. We're 10 minutes behind. Um but uh this is a very interesting topic about the convergence of AI and blockchain AI without limits. Um so before I begin, we'll start off with some quick introductions.

I'm Nikita Sashv. I'm the CEO and founder of Luna PR and Luna Media Corp. We're a PR agency focused on web 3 and emerging tech. Been around since 2017. Started my crypto career with WOI exchange, now HTX in Singapore uh early on.

Not as early as Michael though, and um fell in love with the the crypto space. So, here I am now headquartered in Dubai, offices in London and New York, team of 100 and uh really happy to be here today. Nice. Thanks for the intro. Zero GM.

Really excited to be here. I'm Michael, co-founder of ZeroG. At ZeroG, we're building the largest and fastest AI layer 1. And we'll dive into that. And as of a few days ago, I'm also the executive chairman of a NASDAQ traded company called Zerostack, which is the DAT for ZeroG.

My background is originally in technology. I worked for Microsoft and SAP Labs, then switched over more to the business side. worked for Bay & Company consulting Fortune 500 companies. Then worked for Bridgewwater Associates, the largest hedge fund in the world. After then that went back to graduate school at Stanford, started my first Y Cominatorbacked unicorn company.

Eventually sold the company and then 2023 because my classmate introduced me to amazing co-founders that he invested in in the past started zero in May of 2023. So very excited to be here now. Just had our TGE released our mainet. So, it's been 96 of really crazy hours. [laughter]

Super impressive background. And I know you had backto-back panels today. So, thank you for being here. Um, what does Michael, what does AI without limits mean to you? I want to like just take a overview of this conversation.

Yeah. So, the way we think about limits is if you look at traditional AI, it's effectively blackbox AI. So the limitation there is you have to trust a particular entity to do certain things for you because you can't verify what's happening under the hood. You don't know where the data came from. You don't know who labeled the data.

You don't know what the censorship decisions were. You don't know what version of the model you're getting. You don't even know which version of the prompt you're inputting into the model because all of that is subject to an entity controlling all of this. Now on the other hand, you have decentralized AI, which is what we're building. And that's where we want to have an AI without limits.

An AI that's transparent. An AI that's verifiable. An AI that's communityowned and fundamentally completely safe for human beings. In blockchain based systems, we figured out how to have trustless systems. We figured out how to basically deal with adversaries on the network.

You have slashing conditions. You have incentive conditions. If AI is going to run major societal level systems for us, we need to ensure it's in an environment where it can't cheat or if it tries to cheat that it takes that we take the resources away from that AI agent. And so that's what it means to us to have AI without limits.

Okay, that's that's very interesting answer. I want I want to dive into that a little bit deeper. What is blockchain? You touched base a lot on transparent trustless systems. What is blockchain actually helping us do right now with AI that is not already possible that's not already happening in web 2?

Yeah, so this whole aspect of ownership verification these are really two key concepts as part of that. So for example, we've built a new NFT standard called the INF which stands for intelligent NFT or ERC 7857. What we've done is we've embedded an AI agent's private keys into the metadata of the NFT, which means that if you own this NFT in your wallet, then you can definitely say that this agent now belongs to me. You have an ownership string. Anything the agent does belongs to you.

Anything the agent does economically also belongs to you. This was not possible before because if you go on to OpenAI, for example, and you create an agent, it's always subject to terms and conditions. Those terms and conditions could change. tomorrow your agent could no longer belong to you but here you have complete ownership of it so that's something that's only possible in web 3 because of the the primitives that we've built verifiability is another aspect how do I know what's actually going on behind the scenes with the AI models you have to trust an entity but with things like trusted execution environments at testations on chain you can actually know where certain things originated from if they're true things that originated let's say from a camera lens or if they were altered over time because of that immutable record and then you can verify a particular model's output and a particular GPU that was used for that model. So you can definitely say like okay this is the model that I expected and I can individually verify it myself that you cannot do in centralized AI as well.

And then the final thing is community governance. We don't really have that. We rely on Google to tell us this is the value system that we have. So therefore you have to use Gemini and so therefore you can't change anything about it. But when you create communitydriven models, you can actually build in your value system and you can use models that are in accordance with your value system.

One size does not fit all in this case. So that's the third aspect that's not possible.

So I mean it seems pretty evident why AI needs blockchain. Um, but what are some of the other bottlenecks currently on bringing largecale AI to decentralized networks such as storage or compute?

Yeah, when when we first started, Ethereum had a data throughput rate of about 80 kilobytes per second. Now, you look at modern AI data centers and they do hundreds of gigabytes, even terabytes per second. So, how do we bridge that gap? That's a million times off if you actually want to do some of these processes on chain. And so we had to completely redesign the architecture that powers some of these blockchain based systems.

So we created a very high throughput data availability layer that's able to scale to these dimensions because of uh horizontal scalability and parallelization. So those were the key inputs and aspects. And then there's other issues too. So most of the storage on web 3 is cold storage. There's not a lot of hot storage.

Recently there's been more. But we also architected a storage system that is made for AI workloads made for large scale um kind of data sources as well as very fast upload and download speeds that are commeasured with AI workloads. Then on the compute side too you need uh what I mentioned before trusted execution environments on some of these graphic cards so that there's not a lot of overhead between web 2 AI and web 3 AI. Now there's still a lot of issues around how do we actually train very large scale models. So for the first time a couple months ago, we demonstrated that you can train a 107 billion parameter model fully decentralized.

Many people didn't think it was possible and this was already three times the previous record. And what we effectively did is we took some research that Google had dlo and we uh added some things like dual optimizer policies and also a way to not have gridlocks on the system and we improved the throughput 357x compared to the initial uh Google paper and we called it doolo x. So in summary, still a lot of infrastructure challenges that need to be solved over time where until we reach this kind of utopia where everybody, every device can contribute to AI in some form or fashion and your data can actually be used and you can monetize your own data and have complete control over it.

That's great to hear. There there are several blockchain protocols that claim to be at the intersection of AI and blockchain. Which one is ultimately going to win? Are they all going to function together? How do you see the future of AI and blockchain working together?

Yeah, it really depends on which part of the stack you're building on. So for example, we have a storage layer and um you want to have really private data and so we have a protocol building on top called beacon protocol for example. So there's a lot of synergistic type of protocols that can build with you. Another example is there's a protocol called balcanium labs uh which is doing decentralized federated learning where multiple people as part of the network can train a model without knowing any of the data that other people are using and so also very synergistic can easily be implemented into the overall protocol design but then there's other protocols like uh bit tensor for example what they've done is they have these subnetss and they are a coordination layer primarily where they coordinate the rewards between the subnets and then leave it to the subnets to figure things out on their own. And then the subnets have these validators that then distribute rewards.

So they're not actually doing really deep fundamental AI research. They're acting more as a distribution layer. And so for us, it's very important to have a one-stop shop approach. So that a web 2 AI developer eventually just needs to go to one place and has all of the different services that they need for whatever they try to get done in one location, one place. And so that's a big differentiating factor for us.

any new tech in the market can be so vulnerable. So what are you guys doing right now to tackle uh data poisoning or bad actors?

Yeah. So we've actually written uh a paper on this and uh we attacked centralized federated learning and we've actually found no matter what the kind of security assumptions were, we were able to break them and insert poison data and decentralized federated learning is actually an approach that solves this particular issue. This paper is at uh ICCV later in the year. It's one of the top AI conferences. We're going to be presenting on it.

And so there's a lot that we can do from a security standpoint where decentralized technology will actually solve some of these pain points that centralized and blackbox AI has as well.

And I want to talk a little bit more on the token side and community governance right now. What governance models do you think will emerge for um for community AI? So I think the key aspect to that is a fundamental belief that we won't just have very large scale models that will rule them all. Our belief is that we'll have a lot of smaller expert models that will come together and then compete against the larger scale models. Uh why?

because let's say I know 10 of the best solidity developers in the world and uh I get them together they contribute their data they contribute their compute they tokenize the model and then whenever the model runs they can reap the reward over time and so then as part of that naturally a community has formed you have a community around the GitHub you have community around all of the data pieces and then you could form uh kind of Dow based governance around that if you wanted to and that DAO can then determine well what do we do now how do we actually govern this Do we do um you know majority vote or in our case as part of zero G we'll actually be one of the first companies to utilize this concept of ADUNA which stands for decentralized uninccorporated nonprofit association which is a legal wrapper in Wyoming around the DAO concept. As part of that, you can have a security console that reviews major protocol upgrades and then essentially determines, hey, this is a good upgrade, so we should recommend it and also jump in at least early when there's like a potential hack on the system or there's a chain halt and then move very very rapidly. Bring those up to a vote and that vote is an optimistic vote meaning if there's not enough veto power then it passes effectively after 5 to 7 days. And so there's different governance models I think that will emerge. Some some communities will want to say it's just a majority vote.

Some community members may say I want to delegate it to a leader because we trust them. Some communities may say um hey we want to do this optimistic type of governance approach. So there won't be a one-sizefits-all because value systems are different, geographical value systems are different. So I think it's going to be really interesting to see as we get some of these expert models how they'll develop. seems like there's a lot of utility behind the token, but how do we how how will you ensure that the token goes beyond just hype?

Um, I understand that there is a community, there's governance behind it, but there will be a lot of retail investors that want to come in to the token as well, especially right now when AI and blockchain is such a hot topic. So, how do you ensure that your token is sustainable long term?

Yeah. So, there's three buckets. We talked about governance. Then there's a security bucket because you have to have staking and restaking and liquid staking. So there's some utilities to that.

You bring a stake to secure the network. The longer you keep that stake, the more uh kind of rewards you get for providing that security. And then there's the whole utility bucket which is all about utilizing the applications that are on the chain. And those applications could be some that are built by us. So the storage network you pay with zero G tokens you get storage compute network you pay for inference fine-tuning pre-training those types of aspects or it could be the application layer.

So let's say it's the DEX Jane DEX that's then utilizing ZG as a native token as a trading pair or you have to with any application interact with the ZeroG token as a gas token. So there's going to be plenty of utility between those three layers.

That's good to hear. We're all here for mass adoption. That's a common denominator with a lot of a lot of the people here today as well. How do you think um what what do you think will bring in the next wave of users to adopt zero g?

So I mentioned one of the use cases which is kind of community- based models. I think that could be a very large uh system and then once those models are out in the world then naturally other people will utilize these models and ideally we put it into some interface let's say like a zeroG chat interface where you don't even have to worry about hey there's a million models that could do these things we'll pick the best model for you and so I think that'll be a killer application because it gives you the interface you're used to but with all the power of blockchain underneath I think that's one area another area is in the form of ownership kind of I mentioned the INF FT there is a launchpad that's building on our chain called AI verse where for the first time you can create an AI agent that then becomes an INF and anything that AI agent does and produces for you belongs to you now and so you can then trade these assets. So I think that would be another really important use case and then finally any can be built on web 2 AI should be built on web 3 AI as well and so looking forward to a lot of those use cases.

Awesome. I think you've answered a lot of my questions. I just have one last question to close this discussion. In five years, what do you hope to see to really feel like Zerog has made it?

Yeah. So, right now we're focused on building the liquidity layer, application layer, research and technology layer, and then finally the governance layer that we talked about. And so if we're successful in five years, many web 2 AI developers would have come to web 3 to enjoy the benefits and built applications that have missionritical AI in them. So if you're for example managing a factory, I would want that to be built on zero because of all the security and safety features that are part of us. So I think that's what success would look like.

Amazing. Thank you so much, Michael. Great chatting with you.

Great. Thank you.

[music]