How AI Will Enable the Next Generation of Intent in Web3 | Isabella Yang - dappOS

Sorry, it's not It's not presenting the slide. Oh, it's coming. Okay. Cool. We are ready to go.

Hi, good morning everyone. I'm Isabella. Are you okay? Okay. Uh some chit-chat.

I'm Isabella. I'm from Bay Area. Anyone from California? Yes. So, and I'm from DappOS.

DappOS is an intent execution network, but today this talk is mainly more like a inspiring session that I'm not going to talk about DappOS or any particular projects or promote any specific uh things. But, I'd like to like bring you all to a discussion where the AI web3 intent and how they can interact with each other in the new era. Okay. Uh Thank God I saved a PDF so I can use this. So, the convergence of AI and web3.

Why this comes to our mind? I think on a very interesting phenomena I've seen in Bay Area is that every people in Bay Area is talking about AI right now. Like all the big companies or startups, they're all doing AI innovations, applications, infrastructure, everything. But, today this talk will mainly touch base on the application layer. Like I'm not going to talk about GPU or computational um, uh, resources, uh, sort of things.

Because I think application layer is more like it's more accessible and it's easily more easier to understand for most people here. And before we go into that, I want to ask like how many of the, how many of the audience here like write codes? Python, C++, wow. Almost to 50%, I see. But do you know in this world only 10% of the people knows how to program?

Like 90% of the people don't know how do you write codes. And also I want to know how many of you have interacted with large language models at least once in the past of 3 days. Oh, almost everyone. Cool. I chat GPT.

I would guess so. So I think this I think this this is definitely a very hot topic right now in AI with the advanced, uh, large language models and all the capabilities. And we cannot help but to think how we can bring this power to web3 and how can we utilize these strengths to strengthen our web3 community and industry. So what So that that makes me think about what challenges that we currently have in web3. So for me, I have a, like, 10 years of background in AI, uh, before joining crypto.

I had an internship at Microsoft Research Asia where I did question answering system. I also did some recommendation systems like back in the old days when we used the deep learning like not large language modeling. And now with the web3, I when I first entered web3, I noticed that it's really really hard to use. Like it's powerful, but it's also very complex and it's very counterintuitive for a lot of people like including me uh, back at that time. I spent like 1 hour setting up my first wallet and had to remember write down not screenshotting the private keys.

And also had to understand a lot of concepts such as gas fees, bridgings, and you know, all those sorts of very hard stuff to understand. And then with the current AI, AI is definitely revolutionizing, I think especially in this intent-based interactions. The biggest value it could bring is that it could remove the friction in execution. So, our vision, or my personal vision, is that I want to have a web3 world where the intent is seamlessly understood, optimized, and executed by AI agents. So, before that, let's have relatively clear definition of what is the intent here.

Understanding intent in web3 means that the intent we are talking about today, it's something that is concrete. Like, I want to get rich, that is not my intent, but that's my life goal. But, intent is like I want to do something that can be verified successfully or in failure. Right? This the intent has to be verifiable.

So, that's the that's the main difference between intent and anything else. And the current model is that users have to be very, very specific, like with all the details and explicitly define an execution, just like the example here. But, in the intent-based model, I want this to be rephrased to I need 2,000 USDC and at a best rate within 5 minutes. And however, today, there is a lot of limitations that prevent us to fulfill our intent. The complex UX where you have to understand the gas fees, all the all the all these things I mentioned earlier, and also the execution friction that you sometimes have to manually like manually do the wallet interactions by all yourself.

And also you have to manually search for the best timing. And there's also a lack of personalizations. For For example, every user follows the exact same rigid execution path. And all of these adding together create a lot of barriers. Like if you I don't know how many transactions do like do you do by manually per day, but how many TikTok videos or short videos that you swipe to see every day.

See, that's the difference because the friction is so high that prevents you to use it and go go deep with it. So, how AI can be helpful here? I see this as a I see AI as a missing piece in web three intent. It's that AI can help with at least first, it is able to do the NLP translation like natural language processing where users they can describe their goals in very high describe their intents in very high level plain English or any other languages, of course. And AI can translate them into a smart contract interactions or a sequence of interactions that needs to be take.

And the second is the reinforcement learning. So, the best So, the easiest way to understand reinforcement learning is you train your pet. Train your cat or train your dog. When it does right, you give it a reward. When it doesn't When it does wrong, you don't give it a reward or you have some little punishment for them.

So, over time it will learn how to behave as intended. And with this RL, like reinforcement learning, AI can optimizes the routes, slippage, gas fees, and also timing for the best execution um using my previous example. And the third one is the risk management. So, AI has been adopted in a lot of fintech areas to detect abnormality, right? Like abnormal transactions, abnormal cases.

And this can be definitely adopted in the web3 world as well. Like it can detect market trends, rock pools, and math attacks in real time. And that's the most importantly that you have to do this in real time because in web3, we all know that speed and the fees are the like the fundamental elements that we like develop things here. But AI, sometimes it can be slow and can be expensive to use. So, here I come up with some use cases, but definitely not limited to this and happy to to happy to discuss this after this after this talk.

So, the first is the DeFi trading and asset management. For example, here AI can enable a trustless and AI-driven smart contract automation where users can specify their desired outcomes such as the one I said like maximize my yield with minimal risk. And AI can dynamically rebalance the portfolio across, for example, Aave, Compound, Lido. And AI can predict the slippage and math risks ensuring optimal trade execution. And the second is an AI-powered DAOs and governance.

I know DAOs are so like nobody talking about right now. But I still feel this is so powerful in like governing or managing a community where you can where you can use AI to summarize the governance proposals and recommend optimal voting strategies based on your past behaviors or your history data. And before you make any decisions, the AI can simulate the outcomes of your all your potential decisions and propose and recommends the best one for this community. Right? And the third is a personalized AI wallet.

I think this is relatively easier to understand. Like users can interact with their AI wallet assistant and uh get things done. Like they can set their financial and privacy preferences and uh this AI wallet uh assistant can automatically help them to suggest and to execute uh transactions. And the last one is NFT and Metaverse economy. I think So maybe here I'm more talking about the uh Gen AI like generative AI where you can use AI to curate the digital assets based on users pre- preferences and buying history.

So all of these can be I can I can see the potential of using AI to like be applied on all of these use cases. And there are definitely some challenges that come coming with the AI and Web3 when it when when they combine together. One is the security risk. A security risk is so important. Like probably a lot of you a lot of you have heard about the accident happened just few days ago uh with Bybit and the Safe right?

So safety So security is always the top one um uh thing that we need to monitor. So we have to ensure that AI doesn't introduce vulnerabilities in smart contracts. For example, I remember previously there was a game that you can send any instructions to AI and your goal is to make it transfer a money to you. So, that was a very interesting challenge and that actually shows that AI is not like it it it it's not perfect. Like there's there is ways to hack that, but just like all the other technologies, like none of those are perfect, but we have to be aware of the risks and to provide and provide relatively like uh management on the risks.

And there's also decentralization concerns. I feel like Open AI is not open, right? A lot of models are not open. Uh but Deep Seek is open. But just just to say like I feel decentralization is also the core spirit of Web3 and I hope the AI we use here is decentralized, too.

And also there's bias and fairness uh fairness issues. This is a very like inherited in the AI technology because AI models behaves based on the data you give it. Garbage in, garbage out. Biased data in, biased outcome like you will get. So, this is also very important to keep in mind.

So, the future I see I I feel like AI and Web3 definitely can enable a new digital economy and this I with the democratized access, trustless execution, and the intent-based uh interactions. But also I want to talk about more on the AI limitations in Web3. I was sitting down there and listening to the previous speakers talking about the AI project AI Web3 project. I'm very excited, but I also want to like just bring this to our attention that AI is definitely have many limitations. For example, it For example, AI still struggles with with interpreting nuanced user goals.

I For example, if I if the intent is that I want to swap ETH for stable coins at the best rate. But how do you interpret the best rate, right? This is very ambiguous. And LLM is still sometimes hallucinate as well. And then you have to make sure that it follows the instructions.

And that is also a very hard tech like tech Nicole tackle to deal with. There are some current solutions. For example, you can do interactive prompting. Like or ask it to reason why it takes the steps. And also you can predefine some intent frameworks.

Like adding all these together to make sure the the the AI interprets your intent precisely. And also there is the adversarial environments that it cannot like perfectly anticipate market shifts or detect all the attack vectors. There's always misses. So that means we cannot like 100% like it's not 100% guaranteed to be safe or to be, you know, to be protected. And there's also privacy concerns.

Like a lot of users may not want financial intents publicly visible on chain. And as well as the centralization issues as I mentioned before. Uh but eventually I still see the road ahead. Like all these limitations are more like a reminders or signals for us to push the boundaries forward. And where the limitations are, there are chances are, and there potentials are.

And I see the road ahead is decentralized AI and with some technologies such as zero-knowledge proof and also with the on-chain memory. So that's pretty much the talk. I'll be like around for for some time and happy to discuss after this. Thank you for your attention. All right.