AI Design Beyond Chatbots

Hey everybody, my name is Sydney Lai, I know this is the last talk of the day, so I made sure to design the talk to be a little weird, to, you know, have you guys have a little bit of fun. So why am I even giving this talk? What is Gaia? What have I been working on? Just for a little bit of context, I have spent most of my career working specifically in infrastructure, anywhere from IPFS storage to DEX infrastructure to key management solutions, so why am I giving a talk at DEF CON, right?

And I think one of the most, how do I, I think one of my biggest motivations or biggest drivers as to understanding how do we provide Web 3.0 Rails to agentic workflows is probably one of my biggest drivers to even bring me to this place today. So again, because we talk a lot about DApps, I was thinking, you know, one of the most important things I'm currently working on in my workflow right now is being able to securitize a lot of these agentic workflows, the data that goes into the agentic workflows and the people who provide that type of data. So when I'm thinking DApps, DApps, DApps, okay, what could that possibly look like? What are we currently building?

I actually wanted to go through this example with Minecraft as a UI. And so I'm going to use Minecraft as an example to showcase how we are currently using AVSs specifically to securitize our network. So this is going to be on a loop. You guys can kind of just visually follow along, and I'm going to narrate what this is actually doing. So in this particular case, we at Gaia, we're building an open source infrastructure for you to be able to build DApps or applications on Web3 Rails.

What that means is that you can build application layers that are agentic workflows. And so in this particular case, we decided to use Minecraft as a demo. And in this particular game, what we are showcasing is that you can have autonomous agents in the game itself. Typically, the game is driven by humans, kids, adults, doesn't matter. And in this particular case, what we are telling the agents to do is to accomplish certain tasks.

We're telling them to get certain resources such as wood, maybe they're going to go gather wood to build specific recipes, build a boat, build a house, weaponry, etc. And what we're seeing is that we are laying the foundations for MCPU or the ability to have this type of integration across different applications so that the autonomous workflow can make these decisions for you as a player because it understands what your intention is. And so in this case, what you're seeing is that because we are running a model on top of this UI, we are also providing the example that we're building an AVS to securitize this application so that it is actually wood or you are actually a good actor or the instructions which is, you know, come follow me, go find this wood, go build me a boat, that they're going to actually build you a boat and they're not delivering you like a house or something like that. And in this particular case, I wanted to further illustrate MCP because I think oftentimes we look at, you know, architecture through the lens of either code or diagrams and in this specific analogy, just think of it visually as, let's say, your MacBook. You have your adapter and within the adapter, you can have different dongles, right?

And in these types of dongles, it's really what kind of application or DAP layer you are plugging in and you're letting your automated workflow make the decision for you if you want to go on Collabland, make a discussion in Discord, maybe Uniswap, you need something traded automatically if a certain condition is met during Discord or otherwise you need to, once you swap it, you're going to sell it on MetaMask to gain an NFT, whatever the case may be. At the end of the day, what we've been working on, on the upper right-hand side, is being able to provide that type of inferencing or context so that your agents can autonomously integrate and interact with the other applications that you so choose to integrate with. So I'll give the example, let's say you are using, let's say, CAT2BT and you say, hey, go pay my electricity bills. It's going to say, hey, I don't have context. What state do you live in?

What is your authentication? How do I go and log in and pay your electricity bills, as an example? So how do we do this, how do we know that this agent is running the AI model that we think it is? So in the case of the example of Minecraft, you say, go build me a boat, but do we know if that boat is actually a boat? Is it actually using wood or whatever the right resources are?

And so one of our works that we've been really focused on at Gaia is being able to build a context network, an index, index context and datasets across both on-chain and off-chain AI marketplaces. In this particular case, you can see, you know, at the last time I took this picture, we have about 700,000 nodes, about 8 trillion in throughput, so you do know that the node or the context network is not static, it is live. And in this particular case, since we are in Germany, we have about roughly one-fourth of our network being supported on a commercial enterprise level through node providers here in Germany as well. And because we're an infrastructure tool and, you know, again, we are at DAPCON, we can look at the application layer on these types of infrastructure in many ways. The most common ones that we are aware of is going to be the chat application, right, the chat interface that most people know, which is like, hey, how do I do this?

What's the information for that? What is the context for this? We can then also stretch our imagination of our AI world from kind of a 2D chat application all the way to kind of a virtual environment. What does an AI environment or an AI world look like in the UI or the skin of Minecraft, as an example? But I think another thing to really understand is that when we are building these virtual worlds, the common question that we're always asked is, what is real and what is reality?

What is true identity of the other person? Is it a good actor? Is it a bad actor? Is, you know, is this AI agent world valid, as an example? So one of the things I really want to walk us through is just a really quick, quick taste of like, what are the current solutions in securitizing AI that we see today?

So one example is ZKproofs. This is, of all the cryptography-oriented types of security, I do like this one. As an example, for those who are new to this concept, ZKproofs, let's just say you are trying to buy alcohol in the U.S. and you need to show your ID.

When you're showing your ID, you're actually giving away a lot of private information because I see your home address, I see your date of birth, your height, we all lie about our heights. We also have the eye colors and all that stuff, right? And this type of information not only becomes a privacy concern, but it also becomes almost too much information, too much context, and it also helps create latency. And when you have a type of latency when you're securitizing these data sets, then you as a user is going to have a bad gap experience when you're interacting with it. You know, another example could be is, using TE, is another cryptography-based or oriented type of security for securitizing AI.

It is not my personal favorite. There are a lot of people who believe in this type of thesis, which is totally fine. And I think one of the main reasons for me is because, there's two reasons. One, it really requires, you know, hardware dependency. And so as we've seen the advances of technology, sometimes hardware can be outdated.

And another reason is because, for me, it requires too much binary decision-making, yes or no. If yes, you know, slash, if no, et cetera, right? And so I think that with the evolution of AI and how we've seen technology impact society and how society impacts technology, it's going to need to really be able to handle nuanced data and nuanced AI. So I'll give an example. You did something bad.

Do you go to jail? Yes or no? Right? That can be nuanced. Is this art?

Nice. Is it a nice piece of art? Yes or no? Right? So then, these types of cryptography-based security solutions, it becomes harder to handle that type of information.

And so in this case, going back to Minecraft, just to stay on theme, is, you know, is the model wrong? So if they're using a recipe book on the upper right-hand side, you'll see, hey, look, we have this guidebook, we have this data set, we've collected the material that we need to make a boat in this particular case. And then you're looking at it, technically, it's a boat. Does the boat work? Does it even look like a boat?

Do we know it's a boat? So one of the things that we've been really focused on in this last year in our engineering is understanding how we've built statistical models into AVSs. So really being able to handle this type of dynamic data and this dynamic modeling for AI to be able to handle validity in this particular case. And so within this public research that you guys can see, one of the most important things that we're looking at is understanding how much the models deviate from each other. How much do they statistically deviate from each other?

So what you'll see here on this chart is that it really shows the distance between the model pairs. And they're about 65 times larger than the highest scattered observed plots within any of these models. And we're running a lot of these like Lama 3 based models. And so this larger separation between the inner model distances really provides a variation that indicates that the guy knows which ones are different, which ones are the same, which ones are true, honest. And at the end of the day, we're really looking at pattern matching as an example.

So going back to the analogy of having a UI layer is what does AVS look like or an agon layer AVS look like with the image of Minecraft, right? So if we're looking at these AVS validation in the lens of Minecraft, what we're seeing is that there's characters in this world, the AVSs, that acts as referees. So as your AI agent is going around getting wood and all that stuff, building a boat, what they're saying is, hey, we are a ref. Did you actually get the materials that you said you would? Is this an actual recipe that I said it is?

And so this can be thought of as almost like a visual representation of staking on an AVS. And then the referees or the node providers being able to put down money and vouch for the agent. And then, of course, the bad referees will then get slashed or lose their stake in that case if they're caught cheating, right? Here's another example of another visual of the AVS because you can also monitor and find bad actors and validate if they are a creeper or if they have bad intention. One example, and this is completely a tangent, and I don't know if you can even see this, which is I was using Gemini a few months ago, and I was just walking around in the capital of the Philippines, and I was just curious, who is the president of the current president of the Philippines?

And it was said, hey, I cannot give you that information. And then I started having an argument with my Gemini, and I was saying, you know, I'm asking for the current president of the Philippines. This is not a controversial topic. This is just a fact, right? And it says, again, I can't give you that information because it cannot validate me as a paying customer, cannot validate if I'm a bad actor or not in this particular case.

And so, you know, this last photo is a little bit larger, it's a little bit broader, but in this particular example, this is the architectural diagram to showcase how the AVS networks works with the indexing infrastructure that we have at Gaia. So really, it's the backend architecture that showcase what the Minecraft example was showcasing. But at the end of the day, what we're really trying to handle is really problematic nodes and being able to have and handle subjective information, and how do you do that both mathematically as well on an engineering scale? Because at the end of the day, our, you know, I think our biggest focus right now is being able to understand how do we scale validation on a AI network? Because if we look at what we're trying to do in the game of Minecraft is that we are playing this game, we humans are building AI.

And what is the infrastructure that we are building so that we can not only create AI, but set the guardrails to be able to validate and really provide the guidelines so that it can be autonomous? And so today, that's really my talk, is just helping everyone understand what is an eigenlayer AVS, how do you use it to securitize an AI network infrastructure, and using it through the lens of Minecraft. Thank you, everybody. Are there questions? Yes.

Is there any questions this time around? Let's see. For Sydney? No? All right.

Thank you so much. Thank you so much, Sydney. All right. Take care.