Dylan Zhang: Crypto Native AI Models Behind the Future of AI Agents

Transcript

You're now plugged in to the Delphi podcast. Everyone thank you for joining and listening on the Delphi podcast today. I'm joined by Dylan Zhang from Pond AI. And we'll start off with just learning a bit about them and love to understand how you got into crypto and particularly crypto AI and why the intersection really excites you. Let's go talk to you again today. I'm Dylan. I'm one of the co founders of Pond. So how did I get into crypto is it's a long story.

So I heard about Bitcoin back in 2017. So that time there was there was a Harvard PhD, was trading tokens and tried to, get me over to trade tokens with him. And at the time, I didn't really believe in it, the story of Bitcoin, because when I looked it up, all the other news that I saw that I read about, scam of Bitcoins, and it's nothing really positive at the time, but once I learned a bit more about Bitcoin, I feel like this is very interesting, but. It was it was two years later. So how did I get in to crypto AI? So we started from developing a user search engine because I've been doing a business for a really long time.

And I feel searching for people is definitely a huge pain from my perspective, because I don't know who knows who, I don't know who knows who. And it seems as a third party, I cannot do such a thing. You need to process Twitter, LinkedIn and Facebook data at the same time because you have identity misalignment, because how can I prove Yen on Facebook is the same Yen on Twitter? And also I, I cannot process the data from them at the same time because they have different data structure, but I feel like blockchain itself as a unified data structure that as a third party, you can definitely process, the data on there.

And also it has different applications on that, which means that you have more data diversities. So we started by developing a user search engine. Apparently different users that have different, on chain exposure. And we do a lot of preparatory algorithm school process the distance between different wallets. So we will be able to know who knows who, and who knows who well. So if today I want to find metallic, I can simply You know, just typing it, Vitalik's EMS, I will be able to know who in my network knows Vitalik and who knows them well, or who knows who that knows Vitalik pretty well. So we started from that way and gradually we have realized that no matter it's social data.

Or financial data, actually they are on the same graph network, which is the the blockchain launching space. So we gradually expanded our use case and our focuses from social to, DeFi to MEV to recommendation, pretty much, I just fully we are just fully using blockchain launching data. Appreciate it. And before going into the specifics of blockchain, I think it'd be useful to, set the stage a bit and just provide a brief elevator pitch of sorts on what the world of DeFi looks like with Pond. Yeah.

Apparently DeFi, I think everything with DeFi now is fixed, no matter, it's going to be the fixed structure. Or it's going to be, the mechanism itself, it doesn't have any dynamic parts. But what machine learning and deep learning can bring to DeFi is definitely a more dynamic structure and a mechanism that allows people especially developers and project owners to adjust some of the fixed parts of within the DeFi to be more and more dynamic. And also I've heard from one of the founders of a very successful DeFi protocol.

So what do you feel about AI and DeFi is more, it can be an analogy in Web2 in a real world economy. It can be such a thing, because the AI is more of the manufacturing business because the generous values it finds it retrieves the values from the data, but DeFi is more like the financial sector, it's trading or, secondary market primary market that help real world economy, like manufacturing business. They definitely feel like there's a relationship there. But it definitely needs a systematic effort from AI agents from AI models and also from the education on the market about, how machine learning can help with DeFi.

Also in their perspective, in the long run, it's definitely, DeFi helps AI. And so this should be a combination there because in real world economy, it's really, the finance helps the manufacturing business to grow. So this is something like what we envision. So for now, because DeFi has a very matured use case and they have a very mature business model. So it's really AI helps DeFi. And later on, it will be, DeFi helps AI and AI helps DeFi. They both should co exist and it should be in an equal position. But what the model, what AI can bring to DeFi is really about some dynamic structure.

We can touch a little bit on the fee structure because lots of liquidity providers on defi, they actually don't make money. All of they provide liquidity to liquidity pools. This is something like we can change with ai. Thanks. Yeah, no, that very helpful. So diving a bit in, you're, you guys are building graph neur neural networks GNS instead of LLMs large language models. So I think it'd be useful to describe the difference between the two and with a bit of a focus on why the former works well for blockchain. Yeah, so AI model, the time when you want to develop an AI model, you need to find the best possible model that's specifically designed for this type of data.

So in Web 2 the majority of the data actually is language data. The model or definition model has to be a language model. But in Web 3, the most crypto native or the most Web 3 native data is 100 percent on chain data. So on chain data, first of all, is not a type of language data. It's actually a graph data. So what is the graph? What is graph data? Actually, it is, as long as it is, the data structure can be represented by nodes and edges in the blockchain context is the wallets, what is on the notes edges are pretty much as interactions between different wallets, for example, by cell transfer, and also knows can be contracts.

Edges are just interactions between what is some contracts like by sale? Transfer means approve this type of interactions. So since it's a type of graph graphical data. The model has to be a graph model and amount other graph models. GN is the most suitable model structure type of model that is designed for processing. Behavioral data and make predictions, and this is especially useful for especially useful for processing blockchain data, because at the end of the day, it's a type of graphical data, and it's very behavioral driven.

Just imagine all the, what is that interacting with each other? What is the next contract interacting with each other? So that's why we chose and that's what is really about. It's about learning behaviors and predicting behaviors. In the future, and just thinking about kind of the journey that you guys are taking from building a single model to building a decentralized model layer. You got to touch on. The rationale there and what that's going to look like over time. Yeah. So initially we started developing a large GNN, but we'll gradually realize actually because the ecosystem is, it's early that comparing with developing a large model, it will make more sense for us to to develop different smaller models and smaller use cases driven models.

Models so this is something I will be working on. Just because, like I said, in the very beginning. All the use cases, no matter is in social in Defi, MEV and and all that, they are actually all in the same graph network, which means that we can develop a bunch of models and these models will be assembled together as a model layer to power different applications. And also we'll help developers to develop something. Wisdom. So that is why we chose our direction from developing a large from from a large graph model to what we are developing now is a, it's a layer of models by opening up our infrastructure of models and also the data to the market.

So that's something that we've been working on because we feel like we definitely need a systematic effort to provide models to developers and to applications. And it seems like it's very time consuming for developers to process data. Data and also for them to, develop come up with different use cases to prefer to have an existing model and the plugin use. So that's how we made a transition from developing a large graph model to a model layer and the model layer itself. It definitely has so many components in it. It has inference.

It has some of the, model and data infrastructure we've been using to develop the cater model. So we have a gradually open it up to the market. And also we have build ecosystem around that from developers, from applications. And also we are working with third parties to have to have their support and also embed some of the models into the SDK, this type of thing. A model is the model itself, but a model layer is really about ecosystem. So we are definitely pushing the ecosystem these days. Got it. Got it. Now that's helpful. Thanks. And so I guess touching on what that, that whole process has been like, I think, you guys had some breakthroughs in terms of the predictive capabilities of these models.

And so it would be great to touch on that. Yeah so for now we have shipped different types of modeling security in DeFi and in recommendation. What we feel about, uh, from what we have developed so far is the AI can make a great impact in crypto applications and in crypto use cases, mostly it's because, the on chain data structure is relatively simple compared with what we have seen in web tool, because if you want to process language data you have to, develop a fairly large language model and you have billions of parameters. And for each time, if you're training a model.

There tends to be very time and a capital consuming and heavy. But just because the simplicity of the on chain data, pretty much, I just like, interactions between different wallets and contracts. So the accuracy is pretty high. Mostly it's because nobody's, first of all, not many people that are doing AI and using AI to processing on chain data, which means that there Are a lot of like behaviors that actually that are very simple to capture. For example, in our security prediction, we had around 92 percent of the accuracy to detect malicious behaviors and malicious addresses.

For example, if this address is a phishing wallet or not, or this token, potentially, is it going to be a phishing token or not? Just because we all have our own inertia as a, as human being, when we are interacting with a token, we tend to have some behavioral partners. We don't change, or we don't, we haven't realized, we haven't really realized it. So this is something that we've been doing also on recommendation. For example, we have built a recommendation system with Zura. So we have had any 32 percent precision. Compare with Amazon's 6% of the recommendation system, which means that this kind of recommendation accuracy can help pretty much all.

The framework itself can help pretty much everything in the market today to build a killer recommendation system. This is something like we didn't really expect by ourselves, but later on when we saw the result. That is pretty amazing to see just because of the simplicity of the the on chain behavior, but gradually because there will be more and more machinery mechanism that has been adapted to this to process on chain data. So we definitely expect that accuracy or precision will drop, but it wouldn't drop significantly, which means that we by using AI techniques, will provide a much cleaner on chain Future and also we can support a lot of locations to have a angle.

Just imagine what we are using today. You want to pretty much every application that using what to they have a machinery angle. For example, the time when you open your instagram time, when you open your twitter time, you open your application to all have machinery angles. In in a theme type app, they are using probably have, different different algorithm to prevent any money laundering. You probably have, some of the, risk control, in on twitter. Do you have recommendations to something better in? But in web 3 applications, we haven't seen. Many of that, but gradually it will be it will be that it will be the case and we definitely believe that, won't be long.

Got it. And so how do you go from this, quality foundational model to a platform? And I know. You don't want to disclose too much. This is part of the secret sauce of where you're going. Just to provide some perspective on how you see this evolving into a platform. It would be useful context. Yeah from our perspective, we feel like model itself is important, but also the ecosystem itself is also equally important. Sometimes it's more important than the technology itself. We definitely feel first of all, the model should be decentralized because a model basically is a business of experiments.

As long as you have a very matured machine learning framework and you have develop another way, you have a killer, infrastructure of model and data, and then later on it's really about. Different types of experiments, we feel like this type of experiment can be achieved by using some decentralized technique. You can leverage the community you can, have some sort of incentive mechanism to allow people to contribute to this experiment. And from this experiment, you can have as many models as possible. You can find the best possible one from selecting or.

From from evaluating the best someone with with some sort of mechanism. So that is something I would feel like the community can help. I can give you 1 example, which is the Wikipedia. So if today you are Google and Google, they do have a knowledge base, but it's definitely. Much smaller than Wikipedia itself, but the reason for that is really because the Wikipedia has leveraged the community. So everyone can contribute it to it, which resulting, is the largest knowledge base in this entire human world. None of the centralized entity can outperform that because no matter how much capital they have, but it's really about.

The community is really about experiment. We are adopting the same methodology here to decentralize the model, your own process. And also we gradually want to open up our infrastructure to the community as well. This is something that we've been doing, and so far, I think we have definitely seen some very good traction among our community about what we are doing. And that's, pretty much it. And also we are giving developers the opportunity to own their model because the on chain data is transparent, which is not something data web2 has. So this is saying that who owns the data, who owns the model?

So web2 developers, they have no sense to own whatever they have developed. They basically just contributed whatever they have developed in the past to BigTag because only BigTag owns the data, but in web3, everyone owns the data, which means they're developers. Can't have a chance to own the models or they can own whatever they have to adopting the past because they own the data technically. So this is something that we're pushing and this is the angle we are using as well to leverage this kind of like decentralization of model development. So in terms of, how this looks in practice yeah, how do you see that this existing is it going to be, existing primitives or outsourcing some of their logic to, to pond?

Is it going to be completely new primitives built from the ground up, or do you see some kind of combination of both? Yeah. So we definitely feel like this is a combination of both because we do have two requirements we gather from different projects. So first of all, just because the the industry is still very early and this entire industry is exploring at a moment. Many people, they actually don't know what to do with machine learning. Also, although they feel like machine learning makes total sense for them to have such an angle because we have the data as long as we have the data, we definitely can process this data through machine learning of deep learning.

But they don't know what to do. They don't know what the use cases are. So we'll have to work with them to find what is the best possible way, to work together. And also what are the use cases they are interested in? For example, in DeFi, dynamic fee structure is definitely, lots of DeFi protocols that are interested. No matter it's going to be a decentralized liquidity pool structure, or it's going to be a lending protocol or sticking protocol. They all always want to have a way to maximize the return or lower the risk of their stakeholders. But also, at the same time, we have seen projects, they do have a sense about what they want to have.

For example some of the different particles that are very interested in having a strategy recommendation, they have a bundle of defy strategy different strategies, and they want to find it out which one works the best for different, users. And apparently they have too many, for example, some of them, they have maybe more than 100 strategies, and it's definitely not human understandable. It's definitely not human readable. So they want to provide a recommendation engine for. Dear users to make sure that they can have the best possible strategy that's specifically tailored to their use case, or it is perfect for them based on the on chain behavioral habits.

So this is something I would have seen as well. So it's definitely the combination of both. Yeah, but gradually, I feel like with the industry gets more and more mature, people will have better understanding about what they want and what they need and what machine learning or what AI can help them to do X, Y, Z. That's helpful in terms of kind of the implementations, right? You have yield optimization. You have better forms of protection whether that's managing leverage or more even Kind of security risks. Do you see some of these I guess what's the lowest hanging fruit you think?

That they can see the see an immediate benefit. Yeah, so we think is that the lowest hanging fruit is definitely trading because to process on China You require a lot of labels because the main reason for that is you want to have explanation of things for example, if I say If I see a contract today they have interactions with different types of what it's, I want to understand what is the contract and what are those what it's, for example, is this a market maker or is it it's a defy. Protocol or such a thing. So you require lots of the on chain data labels.

And this part is extremely immature at the moment in this entire ecosystem because people don't have the incentive to label contracts, to label wallets. And Basically, how it works is is having relay on public goods. So it's a big shout out to all the, open source and the public goods developers. These days, they are actively contributing to data labels to purchase like us, but it's still it's definitely not enough. If you develop a model requires labels, you need explanation of things. But trading itself, it doesn't really need that many on chain labels, like other use cases in DeFi or in security and all that.

So that is the lowest hanging fruit. And trading itself is a big topic. It can be. It can be a training model. It can be a risk management model. It can be a recommendation model. It can be, the combination of the three and it can be much more. So treating itself is the lowest hanging fruit for now on it. From data perspective from the model perspective, I would definitely feel like it is something like we can do, for example, on security and on recommendation, because for A lot of protocols, a lot of projects, they actually have a very mature API and API that can use directly and build a machine learning models around that.

Basically, model business is really about the data. And I think the difficulty of developing models is really about the data as well. So the lowest in food is low or it's not low enough. It's really up to how accessible the data is. It's tough to talk about crypto and AI right now without really mentioning agents and how much they've taken off. And, you see a range from obviously token terminals being the first. And so that's pretty much it. Just a creative good origin story and creative agent, and then you gradually go down the spectrum and you see some with a bit more functionality, like the AI 16 Zs, which is managing a semi AI hedge fund.

And then did there in terminal with a creating a recommendation engine. Do you see this space being affected or enhanced by what pond is building or. Is there a role that pond can play in that ecosystem? Yeah. So AI agent actually is one of the biggest AI model applications these days. So actually there are tremendous amount of synergies between AI models and AI agents, because at the end of the day agents are. The application of AI models. So with more AI models, more diversified AI models, AI agents. And I'm thinking about, AI system Z, or, Dither on terminal, and even we have virtuals these days, which is very popular in this market.

They all are leveraging content creation models, image generation models, and language models, and we definitely feel like this is area that we can tap in and we can support because we will get a more creative models that can literally allow agents to complete, crypto native tasks and provide much better services in crypto AI in general. But yeah, I'm just really curious about your thoughts on AI agents because you are a very visionary investor and you have seen, lots of AI powered applications and agents today.

So what are your thoughts on agents these days? Yeah no it's a very generous description. It's it's definitely something that immediately, caught a lot of interest. I think it touches on a lot of areas, right? You have, these agents that have varying personalities and have the ability to reach a wide audience, right? So that is compelling on its own. And then you strap on functionality and the ability to even manage money, right? So it's one thing to be a useful tool. It's another to be actually, making investment decisions altogether. And I think It's a, there's a really large surface area there.

I think it's also creative because you have these models and these model creators that like you mentioned, otherwise don't really have much they can do in terms of monetization. And so here you have a path, I think tokens are always going to be one of crypto's biggest unlocks. And so the fact that you can now monetize these agents in creative ways and really. Improve the amount of access that people have to them, which, allows them to also improve over time but creating a way for these builders to actually generate an exit.

And so what that ends up leading to, I think is basically a really virtuous cycle of. Of creating this two sided marketplace where on one end you have, speculators, which is also, crypto is very rich in. And on the other hand, you have something that can draw in some of the best players. Model builders, right? Because now they can actually have a path to liquidity. And so I think this kind of marketplace is a really unique and really accelerated way to bring the best agents to market and incentivize people to go out. And figure out unique use cases implementation.

I think, it's still very early and what we've seen now is really just scratching the surface. So it's, it going back to what you mentioned earlier one of the biggest modes is going to be creating this ecosystem, right? Because if you can, Create a platform that simultaneously connects the speculative speculators in the best way and attracts the, as many of them as possible, it'll then bring in the model builder side as well. And so they have the incentive to build their one because. The speculators are there, but I think all of that brings just a lot of attention. And so building their model there would also generate the most kind of use for their model from outsiders that are looking at these platforms to figure out where they can find value, both from an investment perspective

and from a utilization perspective. And I think it's really trying to figure out this platform idea but also gradually. Iterating on what works and what doesn't and understanding where the most compelling and then it's not just, this one off launch a token and it has to be purely memetic, right? There are very direct ways for these agents to capture value from their use via fees. And so I think. Yeah, like naturally you'll have the quality ones rise to the top based on just the interest in using them. And yeah it, I think creating.

a really strong platform that you can bring both sides of the market together. Is really an interesting way for these agents to grow. Cause right now that you see a lot of them exist on a standalone basis, right? And it's hard to understand what they do. And then it's also hard to judge them against each other because there's no kind of standardized methodology to judge their, historic accuracy or, kind of intelligence and there's different scores you can apply. But so I think the platform also helps solve for that and really create discovery for the best ones to really rise to the top. It's definitely very insightful.

I think I feel like I have learned a lot from from the words just now. Thanks for sharing. It was definitely very insightful to me. Yeah. I appreciate that. Yeah, no, I'm learning a bunch as well. Yeah. And so just thinking through, continued implementations, I think one of the areas that's always struggled on the or I think was initially a massive unlock and it's just, we're talking about AMMs, right? They were initially this massive unlock almost like a perpetual motion machine where. We can create liquidity and then the fees are great. But then we started noticing, there are some deficiencies in the sense that you have LVR and IL.

And basically the idea that you're always as a liquidity provider trading you're providing stale quotes. So you're selling low and buying high because someone's usually arbing the other way. And what that meant is that you're There's probably better ways to deploy that capital. I think that there's some holes in that argument in the sense that, for smaller assets, I think it still makes sense because the goal is to really provide liquidity rather than really maximizing. Yield, but over, yeah. On larger assets when there are other kind of ways you can deploy that capital. It's, it, the flaws stand out a bit and one of these issues or one of the ways people have addressed it is through a dynamic fee.

And so this kind of goes back to what you were saying about creating a more dynamic infrastructure or dynamic system. For people to deploy these assets into. Yeah, curious how are you thinking about, the AMM side? Yeah. So it probably would be shocked after I hear that, around like 60 percent of the liquidity providers for some of the taxes, they don't make money, then what's the point for them to provide liquidity is simply provide, providing, liquidity, they are not benefiting anything from this from the market. So the biggest problem for that is. First of all, they have some toxic flow, some of the users, probably they would just use some techniques.

For example, on 1 hand, they hold certain type of tokens on the other hand, and they're providing equity. So this is the equity to pull, and they've done this token, and which they don't will cost a significant amount of impermanent loss. And they make another money, from this another source of money from probably an equity because they cost a permanent loss of another token. And also another thing is, apart from toxic flow, a lot of liquidity providers actually the DEX itself is not, promising a lot of benefits for them which means they didn't do it deliberately is really because they don't have, didn't have a very cool tool for them to support such a thing to have a dynamic fee structure for them to adjust based on, the market performance, for example, let's say when the market is not performing well, You can lower the fee, which

means that it will further incentive people to contribute more liquidity. So users will have a much lower slippage when the market is probably really good and the projected PL of this parallel token. It's going to increase then you increase the fee so you can make sure other liquidity providers, they can benefit from the token price change and also by from providing such a liquidity to this liquidity pool. It requires machine learning because at the day, it needs prediction. You need to predict what's the potential PL of this token pair in this liquidity pool with the market performance in a secondary market of this token pair of this token of this two types of token, but usually one is when it's stable coin.

So this is something like What the market needs, and Uniswap is about to launch their hook mechanism and hook mechanism is further, boost this dynamic fee landscape. So we want to, we need to have more and more dynamic parts in DeFi mostly because, it's a constantly involving market and it needs some technique. To help liquidity providers to safeguard their interests and also such a mechanism also applies to other verticals. For example, yes, as you can see, if today you are, if you mean to ask today you probably will be charged on a pay.

On a fixed price it can be a 2. 1 ether or it can be, 2. 3, but it's a fixed price. What caused that is really we don't have a such a tool to support, based on the market condition, how many people they are maintaining. Yes, you can further, increase the price of the ads and all that. Even like lots of fee structure in. This industry can be changed. It's not only defy, but also it can apply for lots of applications like yes, and all that. So it's definitely something we feel like we can improve onto this landscape. So for the time that we can.

Protect more off the interest of of liquidity providers. There will be more liquidity providers. They are incentivized to provide liquidity, which will further. Further decrease and this leverage for users, because the liquidity. It's going to be there, the depth is going to be there and we'll have a more and more efficient decentralized finance landscape, and we all have more and more efficient decks. Which means that we can further boost the involvement and development of decentralized finance compared with with a centralized finance. So all those like change, we can.

In the machine, they can't break. To the defy space, and would this exist as a vault? So this would be an example of kind of an integration on top of existing defy, or you wouldn't necessarily need to rebuild the amount from the ground up, right? This would just be a managing vault that would otherwise happen to the same kind of, Yeah, I think at this stage, you can have incremental improvement for defy by. Working with them to do all different features.

But gradually with more and more models, with more and more, machine techniques, you can literally just build a DeFi protocol from grown up, you can literally, I think a lot of people, they are talking about I Fi these days, you can literally build an I Fi application from grown up for example, you have a foundation model that's powering your EMM, and you have a recommendation engine based on this foundation model, based on the, this entire on chain behavior, not only, the behaviors in your, Just so many things you can do with with some machinery support. You can literally build a something from ground up. But I feel like the engineering works is also very important.

But they should be coexist. We do have a lot of. Applications, even web tool, they, their entire base is really about machine learning. They have graph database unit, so they can do risk management by default from the very beginning. And they can have, recommendation. They can always recommend you the best yield strategies on the best, portfolio strategy for you. And they will have, entire ecosystems report that, and they give the recommendation for their ecosystem. So we definitely have seen so many things that have been, that has been achieved in web tool through machine learning, and we believe such a thing can be also the case of the future of DeFi.

No, that's exciting. And then just to think about the, going back to the structure of these systems and model ownership. As part of this platform, yeah, how do you think about value capture and both, for you guys as the platform and then all the builders you're going to, you're incentivizing to come on? Yeah, so I think in what we criticize a lot about Web2 is really, It's really about value capturing is like the users on developers contributors. They don't have a share about whatever they have contributed to this company. But from the perspective is really about data and it's really about machine learning models and it's really about infrastructure around that application.

From. On the data perspective, because you own the data, because data itself is accessible for everyone, which means that everyone owns the data, which means that everyone can own whatever they are using, they are contributing. It's really not about how much money you make. It's really about how much money you retain from the money you make. So just because the the accessible data, which means that all the developers, all the model developers, they can own the model they are developing. Once the model itself has been plugged into real time data.

But in Web2, all the model developers contributed basically because they don't own the data. They pretty much are just donated or they. Get whatever, the, all the big tech on the tech company who owns the data, allow them to get but such a thing that this is what three AI can change. So the model owner model ownership is definitely something like developers can own the model. They can own the revenue or the model from people using it because they own the data. So that is a huge difference, and this is the direction we're targeting on to allow developers to own their model because they own the data and we are.

Building this kind of culture mentality for people, for developers, and also for the projects we are working with. So the time when a project uses our model, it's not a model developed by us, it's a developed, it's a model developed by the community. We want to have, we want to give them this kind of mentality that the time when you are, given revenue, you're providing revenue to the model providers. You are not providing to us. You are providing to the infrastructure. You are providing to the community. So that is the.

The value capture in part, so we are trying to maximize the the revenue of our model developers of our community and in the future, even users, they can contribute to this model because you can build a very user friendly interface to allow them to assemble to combine different models together to have a model orchestration so you can have a better agent in different use cases. And you can have more and more diversified use cases as possible. And also there will be a community that's contributing models. And that's going to be very scalable because a lot of use cases in today in this current AI landscape is exploratory, is experimental.

You definitely need a community to help you to contribute to different use cases and to follow you to, for this ecosystem to develop different use cases and to see what works, what doesn't work. Some of the models that we have developed, for example, security and a recommendation, they have been proven that they work really well. Even compared with Web 2, they work like at least, 3x or 5x 5x better. But there are definitely some types of model we thought that we can ship pretty well. But it turns out because the the diversity of the on chain data is not, it's not, big enough.

It's not diversified enough to support different use cases that we thought that makes sense. Which means that some of the use cases that is not supported by the current on chain data landscape. So we need experiments. We need a scalable way to experiment, which is to leverage the, our community to leverage our community. The best possible way is really to provide them a model ownership so they can retain the vet the value they have developed in the past and which will further incentivize them to do something you mentioned that I wanted to double click on was this kind of models interacting with each other on the platform in the ecosystem can you provide some examples of what those interactions would look like and yeah.

Yeah. I think one easy, one easy use cases can be trading, for example, like trading itself is, it's a huge topic. It's not really about a trading model. It can be something like, it's a combination of risk management model a recommendation model, a token recommendation model, a trading model, on chain behavioral model. Prediction model. It can be so many things. Just think about all the hedge funds. We have heard Citadel to Sigma. Et cetera. So they have hundreds of models. They function at the same time to make sure that, traders or, agents, et cetera, they can they can have better understanding about what's going on and to make a best possible decision based on that.

You definitely need an orchestration of models rather than one single model. And one single model is always not enough. And even the model itself requires constant involvement. And this is something like, about the orchestration of the models, and once you combine this model results all together, model outputs together, you can make a better decision still taking the training. For example, let's say you have a risk management to say. Certain types of token, they probably, they have a, another risk profile because of the other holders of this token. They tend to be, some DJ players, or they are fairly new to the market.

So 70 percent of the token holders, they are new to the market and their web profile. Proof that they are new to this market, which means that on this token might be some might be some uptight on the risk in some ways you probably want to, you want to do something around that to manage your risk and this token recommendation model, it can be something, for example back in the days when union swap traded fully on chain, when they pumped the file axiom in 30 days, what happened on chain? What is the behavioral? Two weeks before, and you want to, learn that and then on to apply it to the market to see what is, what, which token has a similar action behavior.

And you can have a security model, for example the malicious address detection model that can help you to detect the potential malicious tokens. So to lower your risk. So all those can function at the same time together. So you can. Have the best have the best outcome for you to make a decision. So the model, one single model is not always not enough for more for a lot of use cases. You definitely need a more complicated model orchestration for you to make a decision better. Yeah.

So that's what a model look like, looks like. Got it. No, that's helpful. And just to think about, what the next couple of months look like for you guys. Is there. Yeah, what you can share, what you're hoping to accomplish, and how you'd like to interact with the broader crypto ecosystem as you guys are building out. Yeah. So in a coming weeks, we are going to launch our point system to help our developer ecosystem to have some guidelines about what they should work on and what they can work on, and also we will do a lot of research on model development.

By providing guidelines to our developers, they will be able to understand a bit more and also catch up the pace quite faster. So they will understand what to do and what are the goals of machine learning tasks we are proposing. And also we will be working with a lot of big players in this market because people gradually understand start people gradually started start understanding. The importance of the models, because it produces values to the market, and we'll be working with them. And also, at the same time, we want to further expand our developer community.

And also, we have launched a series of computations that allow developers to have some hands on experience. And with our guideline, what they can achieve. So we really want to push this entire machine learning boundaries. And also we want to explore use cases with our ecosystem, with our partners fast enough. And also we want to achieve this in a scalable way by opening up our model and data infrastructure to the market. So people can use it. People can use the infrastructure. We used to develop a killer model. If we can develop a killer model, and also we have stretched out our, Model and data infrastructure by ourselves, because we've been, pushing everything to the extreme by developing our own models, which means that all the developers, they will be able to use our infrastructure to develop models by themselves well enough.

Those are the. Those are the tasks that are tasks on our roadmap that we are about to ship in the coming months. Yeah, no, it's very exciting. Seems like the kind of the perfect time to do it as all of the. on chain activity heats up with market volatility and that drives a lot of fees and so I think you know, The surface area for optimization is huge right after that happens and a lot of the stuff tends to look obvious in hindsight, but no, it's great to have that and I think it's timely. So that's exciting. And if people want to learn more about what you're building or yeah, figure out how to work with you guys, what's the best way for them to find you?

Yeah, so the best way is to check out our website, CryptoPond. xyz or follow our Twitter. So we're going to launch a series of exciting announcements. Pretty soon with some big projects and also we're going to, showcase what we have and how we can help the developer community to develop models with us, with our guidelines. And later on, we even can work on the guideline together. Yeah. So we're going to really push this decentralization of model development. Exciting. Thank you so much for joining me today. Appreciate the time. Thank you so much for having me. Yeah.

Thank you so much for the.