DeFi and AI for Mass Adoption - James Ross

Transcript

[Music] Hey everyone, thanks so much uh for coming out. So yeah, I decided to put this talk together to kind of go through like the key ways we're seeing DeFi and AI uh converge or be used together. Um, and I think there's like a pretty interesting relationship because right now everyone's using AI for DeFi, but soon I think we're going to be using uh DeFi for AIS. Um, so yeah, let's get into it. A bit of background on myself.

I'm the founder of Mode and more recently a project called Synth. Um, we're building at the intersection of DeFi and AI and uh, yeah, tackling a lot of different use cases. So yeah, we can kick off. So yeah, like our kind of vision at mode is that blockchains are great rails for uh AIS to be using, to be interacting, to be transacting with each other. And so when we talk about mass adoption, the next billion users that come on chain, like we've heard that hundreds of times. I think we're at a point now where the next billion users might not be humans. They might be a combination of humans and AIS or they might just be AIS or they might be

humans controlling AIS. Uh the next billion users could be something we don't quite expect. So DeFi adoption challenge, I'm sure you've heard this over and over again like DeFi TVL has peaked um at this $180 billion range. um it's tough to get DeFi through to the mainstream and there's a lot of innovation that's happening at like this interface layer basically making DeFi more easy to use uh things around like onboarding reducing complexity um having like creating products that mean that users don't need to have deep technical expertise you know to bridge between chains understand how to read

block explorers to basically use DeFi so we're here right now and this kind of like innovation in AI can solve some of these problems uh that we're seeing with DeFi adoption and I'm going to get into those now. So yeah, AI can be this kind of catalyst for DeFi adoption. Um and the way we're looking at it is by these like three pillars. So consumer interfaces is one like what new interfaces can uh we create using like LLMs for people to interact on chain. um what are onchain agents doing? What is an onchain agent? Um and how can they help people manage uh their capital on chain or manage transactions or even

manage like deploying contracts. Um and then thinking the third thing we've been thinking about a lot is predictive intelligence and how can uh yeah we create systems using blockchains to create competitions uh which create predictive intelligence which can be used for a number of different functions within crypto and finance. So yeah AI powered consumer interfaces there's a few different projects like tackling this right now. Um, we're doing this at Mode. We have an AI terminal. Um, it allows you to do things like bridge, trade, um, manage liquidity across LP pro, uh, manage LP, uh, liquidity across, uh,

AMMs. Uh, there's also Heon and another project called Wfinder. Hanon's slightly more focused on like memes and Wfinder is more focused on building complex transactions. So, I think we're going to see like a huge amount more um innovation around like these uh yeah like LLM chatgpt style interfaces and they're getting pretty good at helping people create like more complex transactions. Um and they're acting more like an assistant um that you can ask questions to, get responses, ask it to do XYZ, uh put a stop-loss on a certain trade, add liquidity here. Um that's where we kind of are now. We're starting

to see more complexity where you can ask it to do something like smart automation um where okay so when the price of ETH drops to this level can you buy X amount using this amount of USDC that's kind of where we are now this like smart automation and the next step is um yeah being able to essentially through this LLM create transactions which will be triggered by certain things happening out in the world and that's when we have like this yeah more kind of AI style automation where um yeah the the agent is thinking for itself. So this is yeah the AI powered consumer interfaces um and yeah I'm expecting these to

improve a huge amount over the next 2 3 months uh mainly because we've seen the models uh from OpenAI get much much better. We recently moved from um GPT4 to uh four one mini and yeah the results been pretty significant. So yeah definitely try it out. Um we've seen like a lot of onchain agents emerge over the last 6 months. I think two that we've been working closely with Armor from a project called Giza um and also Autonolas. Uh they have a number of agents. They have a modius agent um from OLAS and these two projects I would say are the furthest ahead. They have actual

agents on chain managing capital autonomously across a load of different protocols. Um this image on the right hand side is from Giza. Um and what we've seen so far is mainly like yield generation agents. So um the agent is monitoring yield across like multiple different protocols and then it's moving uh your capital across those protocols to optimize the yield. I think this is like a pretty interesting use case and yeah I've been a user of Armor since day one and it's it's really like it's just a nice experience to use. Um, the way that these agents work or the way that

Armor works is essentially, um, as you deposit into Armor, you're depositing into a smart contract wallet, um, which uses uh ZKML uh to yeah essentially has all the data um, on all the different yield farms. And then the agent, the smart wallet is acting um yeah autonomously based on you holding the private key to that wallet. So as a user you're always in control of those funds but all the actions activity are under this one smart contract wallet which is essentially the agent. We've also seen like a lot of innovation around like different types of vaults which are AI powered um and OLAS uh they actually have a like a consumer like

desktop application um that you can run all your agentic uh activities through. So yeah, onchain agents right now are very much focused on yield generation. Um everyone should definitely test them out. It's a pretty cool experience especially to see that your capital is just like autonomously moving across multiple different protocols. Um and uh yeah uh I think that the next kind of stage of onchain agents is going to be them using more complex data. Um and also we're probably going to see like a massive uh focus right now towards more trading agents or that's at least what we're starting to anticipate. Um with

the popularity of like protocols like Hyperlquid uh we're starting to move from yeah just yield generation agents which are yeah they do transactions uh like the velocity isn't as high as agents which people are using for trading. So we've been working with a number of teams on that and we're excited to see a lot more of that in the future. the kind of yeah this third layer of AI and DeFi that we're really excited about is uh this idea of like intelligence networks. So we've been building a product called synth. Um it's been live for roughly 3 4 months. Synth is essentially a uh forecasting competition with around 200 machine learning models

competing um and the output is price predictions. Um so when we started building out agents we found that there was this huge gap when it comes like agentic intelligence um when it comes to like predicting or trying to anticipate or forecast the future. So in a lot of uh activities within DeFi, you need to like have a model of the future, right? So if you want to supply liquidity in unis swap, you need to uh think about like if you're using a v3 position, you need to think about what range am I going to supply liquidity in. Um so you need to speculate on the range that you should set before you put that liquidity

in. If you use a lending protocol um and you borrow, you put ETH in and you borrow against that ETH, you're going to have a liquidation price. So, you need to speculate on what's the liquidation price going to be? What's a safe um health factor for me to have on this lending protocol? So, these are questions that like we don't even realize we ask ourselves. Um but when you use agents, agents need to have an answer, right? Right. And so a risk averse agent will want to have uh yeah a very low risk threshold. Um so they might think and they need a model on what this yeah lowrisk threshold is. So essentially we went and built synth. Um and synth uh is this like predictive

intelligence layer. So we have these volatility models you can see on the right. Um and essentially yeah we have 200 AI models competing. um they're constantly scored on how accurate their forecasts are. And the top forecasting models uh have this output. Um and essentially what this is kind of saying is uh each one of these lines represents like a 1% chance. So um I think I took this like yeah around 11:00 a.m. this morning. You can see that um we're expecting there's like pretty much a 100% chance that the price of Bitcoin stays between uh 112 and 107 or

106 uh,000 over the next 24 hours. Um and yeah, you can use this data to kind of get like a probabilistic idea of the future. And so we're starting to plug this into lots of different agents. Um we're starting to see like a pretty yeah it being used in really cool ways. So across liquidity provision, across lending protocols, um across per dexes as well. It can help like define liquidation prices. Um and yeah, this is something we've been super super excited about launching and integrating within DeFi. And yeah, we're at this point where with AI and DeFi where we have like a lot of great tools and we have a lot of interesting

integrations um and now we're kind of looking at like how can we bring these new data sources and data layers um and actually like implement them into DeFi products. So we're seeing a lot uh when it comes like risk management like we're working with one uh partner chain risk on integrating this um these type of simulations into yeah different risk protocols um and different lending protocols and uh yeah Allora are also doing some really cool stuff. Um they have these similar like this simil like predictive engine uh but for like much more generalized tasks. So um what we're expecting is that yeah this kind of data

is going to be demanded. We're going to have a huge amount of like diversity coming on chain when it comes to different types of assets um over the next few years. So obviously everyone's expecting like stocks to become fully on chain. This data can be applied to the stock market. We're thinking of like different types of prediction markets, weather models. Um we're looking at like a uh energy markets. We expect these to all come on chain and the expectation is there's going to be like this huge explosion of these new data layers and agents are going to be like the core uh users of this data and trading on blockchain rails. So yeah, uh we feel like we're

very much at the start of this like AI revolution. Um to kind of recap, LLMs for DeFi are getting much much better. Um another protocol I didn't mention is called Codeex. They're doing some awesome stuff too. So we're going from like transaction building, which LM can do very well, to like smart automation, which we're starting to see, to like much more complex transactions, which are fully automated, which is kind of the next stage. The first agents when it comes to like yield generation have been running for approximately 6 months. These are also getting much better. So yeah, definitely go test these out, play around with them if you haven't. Um, and we're starting

to see predictive intelligence uh actually get implemented. And this is where it becomes like okay the AI is actually like these agents actually thinking and actually using AI to do something much bigger um and much better that humans couldn't be doing which is taking all this information about the future and using it to make decisions on an ongoing basis. So yeah this is definitely the best time I think to start building um start playing around uh and yeah we'd be happy to chat to anyone who's interested um wants to yeah discuss further. Thank you.

Hi. Yeah. Yeah. So far we've had around 700 users. So we actually uh we kind of like put it behind like a token I don't know you call like a token block barrier. So you have to stake your tokens to access it. Um we've seen like pretty good adoption. It's it's pretty good at things like bridging. It's pretty good at things like we have a per deck you can put trades on. It's pretty good at doing like LP, but it's not necessarily like 10 times better. It doesn't give like a

massively better experience. What it does is it enables you to do lots of things in one place. So maybe you want to bridge, then you want to do something, you want to do something else. It's a good experience to use um just the terminal to do all those things. But like if you just want to put a trade on, it's kind of quicker just to go to the perex and just put the trade on, right? Um, if you want to then get more research, you want to pull in like different data sources, that's when it becomes really interesting because yeah, it's a terminal. It essentially means you have everything in one place. Um, I think like adoption is going to massively increase. Like we've been working on this for roughly 3 4 months and it was pretty like clunky to begin

with. Now it's a lot more smooth. Uh, so we're going to open it up to more people in the coming weeks. But yeah, I think adoption is really going to increase when yeah, the basically the experience becomes a lot more slick, a lot faster, and we're pretty close to that. Hi. So there is a take that speed Yeah, there is a take that speed of the order execution matters a lot when when trading because when you are late like you can be front run etc.

So what's your take on that? Like don't LLMs uh decrease actually that speed of execution and why why should for example like big uh parties uh use the the LLMs and why users should care about like LLM uh in in the execution pipeline when it like actually decreases the speed of execution and makes you a subject for front. Yeah. So I think that's a good question. Um, I think that like when it comes to like this like chat interfaces, I don't think it's targeting prousers. I think pro users are always going to use like the trading interfaces. But if we're talking about the type of user that wants to quickly,

they see some news or a friend tells them something, they quickly want to buy an asset, open a position, they want to maybe do it from their phone, it's a lot easier experience just to go onto this type of interface, type in what you want to do and click go and then it will just happen. But I think yeah, the most pro traders, I mean, they don't use even use interfaces, right? They just use APIs and build their own systems. I think that will continue. um to be the case. I don't think yeah prot traders will use LLMs because yeah as you said execution slower but for retail use case I think it's yeah that's really the target market.

Um you were talking about the predictive data um um I'm questioning like this predictive data like your public chain this will be public data like who will incentivize that there will be good predictive data on it. Yeah. So I think that a lot of this like predictive data or what we're building is public like there's an API endpoint that people can plug into um we just charge for uh yeah based on how many requests you want to do. So yeah it's kind of we're kind of building this like predictive data as a public good. Uh we think it can improve like everyone across like a lot of activity across DeFi. Um and yeah it's also an

open competition so anyone can compete or anyone can build a model to compete to create this data. Um and yeah it's uh yeah it's open it's available for everyone to use. It's it's on our website like uh the question comes also like uh I guess you know predictor from ocean protocol. Sorry you know predictor from ocean protocol. Yeah. Yeah. Yeah. Exactly. Like what's the what you take like like they kind of sell like the predictive data compared to your public data. You don't think like if someone can sell the data that they will put more effort in it to have better data?

So I think that like there's one thing we're trying to I guess change like the predictive data we're building it's it's not necessarily like there isn't like alpha in this data. So, it's not like we're giving away this data and everyone can go trade on it and make loads of money. Um, it's quite hard to use this data to like make money, but it's easy to use this data to improve like systems. So, like the data we're doing, it's like simulation data and it helps model volatility and this is like a key thing you need when you're building any kind of financial model of the future.

So this data is really easy to integrate uh into like um a trading bot for example, right? So it's a volatility model or trading bot can help understand like how much risk to take on certain trades, but getting to that point takes quite a lot of effort. Um so yeah, basically like we're happy to like give away the data because we want it to be more useful across DeFi. We want to be used right now. In the future, we might change it. Um, if people want to like have like a huge amount of requests to the API, then yeah, we'll probably charge for that. But right now, it's kind of early days. The day is improving. It's like uh roughly

outperforming the benchmark like volatility models by 20%. Which is good. So, our focus is basically keep improving the data and create this like open data layer for all types of agents to plug into. Thank you. All right. Um, hi. Um, nice presentation. So, I have two question. The first one is um, your agent would actually need private keys to carry out onchain transactions. Um, how do you secure those secrets? And the second one is personally I believe in building um AI agents um that actually argument

human decision. So at what point do you think it's necessary to add some sort of checkpointing into the AI decision? Okay. So first question. So yeah so products like Giza use smart contract wallets. Um you could also potentially use like a a vault contract. um where yeah you as the user deposit into the vault and the vault has uh some like ZKML um element which is running the the strategy is like the vault strategist so that's how uh yeah you can have these like agents running or auentic experiences uh where you as the user still hold your private key um and

the second question sorry I didn't quite catch So at what point do you think um the agents actually need uh human checkpointing? So instead of running autonomous I believe that agents actually should augment human decision and not make those decision themselves. So at what point do they need checkpointing? So at what point are they going to like are we going to have like fully agentic experiences with like no human interaction? Right. Okay. So I think that if you like there's certain ways we can think about like what's like a fully AIdriven experience. Um and right now we're at this point where like you can set rules and you can set fairly complex

rules for like systems to follow but that's still essentially just automation. So if the question is like when are we actually going to see like AI based decision-m um we have to ask a question like okay what kind of rules are we setting and what data sources are we using um and if we're giving like a core goal for the system okay so you need to maximize profit while uh ensuring that the draw down on this strategy is not like 20 it's like less les than 20%, right? So you might see uh yeah someone integrate like synth data and look for

opportunities um to go long on Bitcoin in times when there's short volatility uh a low amount of volatility because that's when you have like the least amount of potential draw down. So that's like one strategy that like uh an agent could use. Um and it's just running off this synth data set. Where you need to like um essentially what you need to do as a builder is you need to ensure that uh you can track and understand and have a record of all the actions that like your model is taking, right? And that's where things become or things start to get more complex.

because you might plug your or might create this model that uses this data source but you need to essentially understand like why certain decisions have been made right by the AI um and that's where solutions like ZKML have become important because you have this like track record of how decisions are being made um from via like an off-chain component right so an offchain component offchain data source and the decisions are being made onchain So yeah, there's lots of interesting stuff being done around this. Um, and yeah, I think like probably using ZKML solutions is yeah, the most the most advanced right now.

Okay, thanks. So we talked a lot about DeFi. Um, do you see mode as um like um let's say if you have one workflow of your day-to-day that you would like to see implemented uh that you would like to see replaced by an AI agent, let's say, which one would it be? So, I'd really like to see like trading agents um that help people who want to run a certain strategy manage risk better because most people lose money trading. Uh I think it's like around 90% of me people lose money Um, and that's really bad. And we've spoken to a lot of big like per dexes and everyone thinks that they want people to come on, do a crazy

trade, and then just get liquidated and leave. But it's actually not what they want, right? They want people to come on and be active because uh, Per Dexes need volume, right? Someone who comes on, puts some money in, gets liquidated, and leaves. Um, equally with a lending protocol, they're just going to leave. They're never going to come back. So what you actually want is people who come on who know how to manage risk effectively even if they have a small balance have systems set up so they can essentially make trades in a much more smart way um by managing risk effectively and they'll continue using that protocol. They'll continue contributing volume um and we'll become like a user for the long term. So, I'm super excited about yeah, trading agents

where people can define what they want to do. Like, I want to go long on ETH because I see all these great catalysts and then they can put a stop-loss in at the correct level. Um, and they can like, yeah, essentially manage their strategy in a much more advanced way like a pro trader would, but everyone would have access to this. So, we have few more minutes for one or maybe two questions. So feel free to ask if you have any.

Hey, hi. Uh I'm curious to hear how much of this information we have about the users is actually based in user research of some kind. Um yeah, so yeah, we're doing quite a lot of user research. I mean, I think the good thing about building an LLM is you essentially can evaluate like every question that someone's giving, right? So, you have a very fast feedback loop. Like, there's lots of things that like uh the terminal we've built can't do, but we kind of see like the demand from users because some users will say like I want to do XYZ transaction, right? But we probably might not have the

capability to do that, but we can see that we can take note of it and we can then build out that product, that feature. So, it's quite interesting um that you can get like a very fast feedback loop when you have like this chat interface cuz people just asking just telling you what they want to do all the time. So, that's quite cool. Um I think like when it comes to agents and like this yield agent managing people's money, I think that this is this is going to happen, right? It's just right now there's like this risk level that people really need to I mean people still don't trust them, right? Like when you're when you're putting like funds into these agents, you might put like I don't know five

figures, but I don't think anyone's trying to put seven figures into an agent to manage right now. you know, you still want to make sure you're using the most secure setup possible. Um, and you're just putting into one protocol and you're happy to forego like some yield to make sure you have like the most secure setup possible. So, I think it's going to take a little time before we see like uh yield agents yet grow to be using um yeah or to be managing huge amounts of capital and that's just built on trust, right? They just need to build up the trust. Just like DeFi protocols in the early days, people didn't want to put huge amounts in, but slowly trust is built and now we have like billions of dollars in

protocols like Ave. So, thanks a lot everyone for your questions. We do we do not have more time and if you have any questions feel free to network with them. Thank you. [Music]