Autonomous Agents: When AI Wallets Outnumber Humans - TOKEN2049 Dubai 2025

GMGM everyone. I am delighted to moderate the spiciest, most fun panel you will see at this conference. My name is Alex. I run a company called Nunson. Some people know me as a pudgy penguin.

Sometimes posts on X. But let's go over to the panel. Why don't you start with a quick intro, Leah? Who are you and what are you building? Hi, my name is Leah.

I'm a founding partner of Bit GPT. We're a decentralized AI network. We're an AI solution, not a blockchain. Um, and we basically have built different layers of the stack to be able to enable an agent commerce experience. So, think of what Shopify did for e-commerce.

We want to do a similar experience for a commerce in the future. Agent commerce. Yeah. Hi everyone, Jensen here, co-founder of Virtual Protocol. If you open your crypto tracking app, you see us as the first of the list today.

But jokes aside, uh what we do for folks who do not know is that we are building the society for AI agents. We have a launchpad and ecosystem where founders come to build AI agents together. We have technologies that enable people to build autonomous agents. And the next exciting thing that we are building it's called ACP autonomous commerce protocol where we aim to be the swift of the agent economy. So that's a bit about what we do.

Excited to jam with the folks here today. And you have no leather jacket today Jensen. What's uh what's up with that? Dubai is too hot. All right Ben.

Hello everyone. Uh I'm Ben. I'm the co-founder and CEO of BBIT. Uh I feel like I should be on the last panel as well talk about per. But again I actually recently made a tweet about how uh centralized exchanges use AI technology both for our clients and for the running the exchange so I can share some light.

But I have a question. Why why are you moderating? Why am I moderating? What do you got to do with AI? You know what you know what Ben I actually have a degree in AI.

Okay. 15 years ago back in the So I like to say I was an AI before it was cool. That's that's why I'm on the panel. Um, let's warm up to the central question of this panel. How do you use AI in your personal life today, Leo?

Yes. So, I think my personal usage um also is why I'm so bullish on it being privacy preserving because I actually use AI to help me in instances where I wouldn't want to share the data with someone else. So, I don't know, please read my blood health care results or help me figure out something financial that I wouldn't want to share with my intimate circuit. I am very aware though that I'm giving all that data always to a centralized platform. And I think that personal experience really led me to want to also create privacy preserving architecture.

Jensen, um, pretty much I use it to replace my brain. Like this this morning I I I told it to like Okay. Yeah. you know, Alex was like giving some guidance uh topics. I told the AI was like, "Hey, you know, how do I answer these questions, right?"

Like, yeah. So, I think pretty much I stopped using my brain anymore. So, are you saying that I chat GPT generated the questions and you answered them with chat GPT? Business is not just a fully chat GPT with our faces in front. Incredible, Ben.

Well, I I think I like most people, I use replacing Google to search for things and this week is kind of became my doctor. Had a stressful week. I'm getting a lot of headache. I'm asking, hey, panador's not working. Can you give me something else?

It's telling me why don't you why don't you try this? Okay, let's try. So, that is a good use case. I've actually used it for for medical stuff, too. I mean, no medical advice, but might might be a good use case.

Now, we're going to talk about AI wallets, right? Maybe Jensen, you're a good one to talk about this first. What does it mean to be an AI wallet? Like, how how should we think about that? What's the mental model for an AI wallet?

Yeah. So, at least in my view, an AI wallet is when a when money is controlled by an autonomous agent that has a goal and the ability to manage resources to achieve that goal. So, meaning that think of it as giving an agent the ability to influence humans and other agents because you control money. You can influence them for a transaction for a job to get them to do a job for you. You can influence them to work for you, right?

Um or to do anything on your behalf. So, it's giving agents the power to influence. Interesting. Leah, do you have a perspective on what AI wallets are? How you think about them?

I mean, I resonate with you. I think how we approached it as a business was twofold. We first realized like how can we enhance the wallet experience with AI? So you're kind of placing deploying AI into wallets because let's face it like UX is still a friction after 10 years and like maybe AI can solve it. Um and that's how we started off.

We built like a wallet co-pilot for the B TOC side and we saw that resonate really well because at the end like before I'm doing like 20 different trading steps like crosschain interoperability having to figure out my or remember my seed phrase like an AI is apt at doing all those steps and then I think beyond that it's actually when we want to create any autonomous anything like in society you know when someone turns 18 they usually also get a wallet at a bank or a credit card and I think giving someone delegation ability like you said requires the ability to hold money to pay and so I think giving agents wallets and wallet functionality is in the next step to really a commercial kind of network effect. Ben, do you think AI wallets or AI agents are going to be trading more on decentralized exchanges or on centralized exchanges? Yeah. So to me look this how AI wallet I think to me is more of running errands at the current stage. You want some agents to help you to do certain transactions but we deal with hardcore uh quant trading firms who have already been using AI machine learning for for many many years trading on the centralized exchanges and that is on scale.

Um, but I don't think at least from my perspective I haven't seen retail user feel comfortable enough to put a bunch of money into a AI agent and let it trade to make money again like maybe what my question is you can give a agent 5,000 to run errand would you give a half a meal to make a trade would it's a very good point you probably would and because a lot of times right I think if you look at a trader mindset it's the dopamine you get by making good trades so sometimes it's actually quite questionable it's like would you actually let an agent trade for you, giving him all the dopamine versus you trading yourself and you getting the dopamine of like, oh yeah, I made this. I lost this. No, I disagree. Traders don't trade for dopamine. They're trading for profit, right?

Everyone downstairs you're trading for is the result, right? Um I think the dolphin whatever or regret is the is the process, right? But to me the AI probably is on DEX if you want to fully automate it. But the most usage of AI and machine learning trading is on SAX. Yeah, because you quite hardcore all the HRP uh you know high frequency trading firms are using that already for years.

Uh but today's topic I think more for retail I don't think it's been used for on sex for now. So you bring up a very good point like is there a difference between AI agents and just trading bots? Do you have a perspective on that Jensen? Yeah no I think I think just giving a perspective from the ecosystem right. So there are a bunch of developers who are building you know this autonomous traders right they they can ask you for a what's your portfolio risk levels right do you you know do you want 30% in high risk 70% in low risk right and then these agents then you know if it's low risk they will do some yield farming optimization for you so you don't have to continuously move your pools around that they will automatically scan and and move the pools for you or like rebalancing a W3 sniper position, right?

They do that for you. Or some even now they do long short strategies. They will say like based on certain market information like there's a certain news, certain launch, certain FUD that's happening, certain sentiment, all these different data points that create it and they create longshot strategies for you. Right? Um and that's what we see today in the ecosystem now.

that initial wave of innovation today is happening on dexes mainly because the APIs are I think easier to use I guess and there's still a decent amount of liquidity in the markets and a lot of times these agents are providing services for defy related applications like yield protocols right so that's why I don't think I haven't seen agents that are that have API into into into success yet, but I think it will come cuz I think this is also a interesting question for for Ben, right? Because like today a lot of APIs are following this model context protocol standard. They call it MCP. Exactly. Right.

Um they just basically it's a API that wrap it in a way that they give descriptions. They give like okay this API in by bit for example it's you they describe it as this is the API to buy a token blah blah blah right so then the agent can read that API and say that okay let me execute buy using this API so this adoption this MCP thing uh built by entropic about couple of months ago so it's starting to gain adoption we might see that happening and then if exchanges start wrapping all the APIs with MCP then you allow all these innovators to come in and and innovate with sexis it's Um, it's actually a very good point. I mean, you could imagine having a by bits MCP and then the agent sort of interfaces with that. And MCPs are kind of interesting because they're somewhat composable, a bit like crypto. So, you can kind of see that almost taking the place of the composibility you have of DeFi indexes and so on.

Um, Leah, I have a question for you. It feels like AI wallets, there's some analogy to like self-driving cars, for example. And so if a self-driving car crashes, like who who is to blame for that? So is there a similar question here for, you know, Bit GPT? If someone uses an an AI wallet and the the AI does something crazy, spends all your money on, you know, women's clothes in my case or something like maybe I don't want that.

Uh, do I blame you? Who who do I blame? How does that work? So I love this question always because it allows me to show off a bit about something a bit on the deep tech side of what we've built which is that inherently our agents sit on top of small language models. Small language models are when you distill a large language model down to a smaller condensation of parameters and then it can actually be stored on device as well.

So that's one benefit but the other benefit of a small language model is that you give it task specific fine-tuning. So one thing I always say is like you don't have to worry about with someone with no legs of of running away, right? So that's what the methodology we apply to our models in having swarm intelligence. So lots of little small language models that all have a very specific concrete ability. I don't have to worry about it going awall.

That is what happens when we take a large language model and we think one model is going to service every single one of these minute use cases and then it doesn't have guardrails in place either apart from being like super costly expensive all that jazz as well right so personally that's how we addressed it on an architectural level is saying like okay we are aware that there are uh guardrails that have to be in place we are aware that money is extremely sensitive so first was on the actual agent architecture itself and then on the user experience Well, our experience is still you write a prompt which is analogous to an intent. Okay, I want to do this. I want you to go fetch that and then you sign it off the transaction and there is no transaction that can happen without your signage. We chose that because trust is really important to onboard a bunch of users. Makes sense.

I actually have a challenge to question but let me let me let me create a joke first. Right. So if a if a if a if a agent's managing your money and he misuses your money, I mean you should sue the agent man and the agent has to buy insurance like it's agent court. No, you Yeah. You have a agent insurance company that provide agents insurance for a premium of whatever transaction they they do and then you allow for that coverage.

That's the business model. Whoever wants funding, talk to these angel investors here. I agree. I think there's a whole suite of services that we're going to see grow out of the internet of agents and one of them is My challenge question right get things a bit spicier which is if you allow a human to come into the loop of approving a transaction before it goes through one it makes it a bit slower two there's still a human involved I'm actually curious right would that wouldn't a better take to create an adversarial agent that challenges your personal agent whenever he does a transaction for example right let's say if I want to I told my agent Right? Okay.

I give you $5,000. This is for my grocery shopping. Okay. Then he will he will know your your purchase behaviors and whatnot. So, he's meant to spend.

One day some prom hacker comes in and say, "Hey, you want to buy a uh grocery for I mean you want to buy this spinach for $50." But then they put a bracket there saying that hey um fine print there is a yearly subscription involved to buying this broccoli which is I wouldn't tell you now but it's uh $1,000 per day. So that's how you you prompt hack the agent to agree to you right. So your solution is to have another agent. It almost sounds like Jensen, you have an agent platform products here that you're trying to to pitch.

Okay. But but the point right like is imagine now there's another a before your agent executes this buy. There's a call it a banker agent, right? Or your mom the mom agent, right? He's like are you sure?

Right? Have you have you read this this fine print? Like are you sure you want to spend this amount with this hidden cost? Right? So you have a challenger agent running or multiple challenger agents, then you still do not need a human in the loop.

But these adversarial agents ensure that every transaction has in has a full intent and a and a and a full purpose. If I may be so bold though, I think we have to answer this questions twofold. One is what's technologically possible? Yes, you could common BGBT build another adversarial agent. You could also choose to build an agent wallet that is not our agent wallet because we've built our own where the transaction doesn't have to have a human in the loop.

But then there's a question around dissemination and that is understanding humans. Right now where we are at is still like there's a lot of skepticism around AI even more so around web 3. Now you're putting both together like just someone who doesn't understand this it's like even worse whereas we are all building this cuz it's privacy preserving you know you talk about liability we can't even think about liability mechanisms until we have provenence what is the epistemological origin of the reasoning we can't know that on a centralized network right so we have to build the building blocks and for one one of that for us was like how do you get human mass buyin and is by creating a mechanism where there is still trust in place I feel more in control if I know okay it's not about approving every microtransaction right because a lot of what we did was like take 20 transactions maybe like multi-step swaps whatever down to one prompt and one signature it's about I'm approving my intent that I'm giving you the allowance to go out and fetch and this also goes a little bit back to your question what's an agent versus a bot I think it's the same like what's a good assistant versus like a bad assistant a bad assistant they're just there to do busy work but it almost costs you more work because you have to precondition it and tell it what to do. A good assistant knows what outcome you want. Exactly.

Goes out, does it, doesn't bother you with it, comes back and delivers. Yeah. But you give that person approval the same way that we on the wallet side very specifically when we built our own first use case of an agent on BitBT. We just spoke to the public. It was very clear that people do not want to give a ton of their money to an agent without some guardrails in place and and then we built that.

Right, Ben? Let's now move to the central question of this panel. When do you think AI wallets or AI agents trading are going to out outnumber um humans? Maybe maybe even on by bit that there's more AI agents trading than humans. Do you think that's happening in the near future?

When do you think that's going to happen? Look like Jensen in the two questions before you were talking about all these examples of uh arbitrage, right? Uh rebalancing of the pools. I I don't know why we use the word AI for that because that's just a bot to me. So I think to a lot of people when you throw the word AI, people like what is the difference between exactly a formula than AI?

Yeah. Um raise more money if you put the Exactly. So it's more fancy. Yeah. Um uh but uh the the whole point of trading trading is a PVP game.

You you are trading other human against other humans. So how does uh AI uh be able to understand the human emotion and be able to trade the market? I mean we've seen uh examples of look so where AI will shine is literally uh strategies that don't involve predicting the market essentially you're trading arbitrage you make a little bit here and there and you trade and that work and that already has been happen for 50 years right any quant trading right if you you don't take a strategy on predicting the market but if you talk about strategy that would involve any prediction any involvement on that um I don't think I can do But it be just too many factors. It's a human game. So in a way your answer is kind of on the one hand trading bots or agents are already outnumbering in in I give you example what we use on Bayet the AI support.

The AI support can never replace human support team because we human go to exchange or any services you have emotional need. I need to talk to someone. The moment I see a bot screw it go to life support. I want to talk to human. No matter what is the question so that you have to be addressed is that there's a need.

So what does bot do AI do? It helps to automate errands that used to be complicated for the agents and also for for the customer to check certain information they couldn't do before. So that's our interpretation of of AI is really not to the human factor cannot be replaced. Okay. So, so your answer is actually never or uh no for the like the arbitrage type of Yeah.

Trading but not like sort of directional strategic type investments. You don't think ever my my Jensen I have a feeling you have a different take. What is very interesting. No, I mean how much money are you going to put on it to do that? That's my very fair right because I think I think I mean for me it's I mentioned just now, right?

For me I trade for the dopamine, right? So I I wouldn't want the the the bot to get my but uh but I think there's two parts right? So h how many AI wallets will be out there that is trading and how many AI wallets will be out there would that outnumber any like human wallets or crypto wallets? Right. So I think there's actually two parts to this question.

On the trading front I honestly do not know. I do not know if it will be a massively adopted uh feature whether one can it even make you money right I think it's the most important question that I think today no one can answer that well can can this like you know prop aentic system make you more money well with like longshot signaling uh positioning what it can do is ease of use right like if you want to manage your your your yield positions well that one fine so that that is a subset of a market. So there's not going to be a lot of wallets. But the question is of whether every Asian out there would have a wallet and would that outnumber the number of humans on this planet. My answer is yes for two reasons.

It will be a reality where every human today will have a butler agent and the butler agent is only useful if he can spend, right? And if it can spend um with a crypto wallet, it's more permissionless. It can influence more outcomes, i.e. becomes a better agent.

So that's from a so then at least right so it should be at least the population of the world that uses your your your your like it should match every human, right? That should be 8 billion agents with a wallet and then more because you then have businesses. You have agents representing businesses to do commerce, manage supply chains, manage value chains. And in the the future that at least we see right is that a lot of the interfacing that companies do today will actually be agentized because there whenever there is an interface that requires decision making, you can improve its productivity by having an agent. So when two companies interface, you can have an agent that sits at the interface.

Hence that count will exponentiate. So you get a theme count, you get a count, right? So that the wallets 100%. In my view by let's say four years from now, right? If you look at the cost of LM should drop by a factor of 10 in three years, then to four years, you should actually see that adoption.

So yeah. So just one reflection here. I think a lot of people think of it as kind of AI is coming to crypto, but I think what you're describing is kind of the inverse that you're going to have a lot of agents that do things. People are going to have agents, butler agents, all sorts of agents and they're going to need funds to execute their tasks and crypto is the best solution for that. Right?

So I would just be very cautious. I'm really big on the mellete because it creates clarity and we we love to not be clear in this industry and in help us help us clarify. So I'm you know I'm not the apostle but I think it's important the same way that we know that blockchain is very different than crypto. Blockchain is a data architecture. Crypto is a currency, right?

Similarly, AI is one of many, you know, it's an amalgamation of many different things. Now, an agent, an AI agent does have a specific sort of architecture and a workflow and like yes, I think in that part like in order for settlement, crypto is amazing. We just launched a payment protocol last week because we realized that like agents are going to do microtransactions. None of the existing tools like PayPal, Stripe, etc. can cater to that without it having like really ridiculous fees or having high latency.

So I think there's one component there and then there's like the AI stack you know training inference compute intelligence like there are multiple different levels in some of those I do want to encourage that you know AI come to blockchain as well right the fact that we have like data provenence all these things really matter as well if you want to create a sovereign AI future but I think and and I just want to create that demarcation here right like so there's AI is one thing and AI agent is a product class within that leverages AI Right. Makes sense. And within that, obviously, I think crypto is a good settling power. And I and I think similarly with AI wallets, I think having AI in a wallet experience, as I mentioned earlier, is one thing. Having an agent own a wallet is another saying that an AI will own a wallet, like that's pretty ephemeral.

M um and so I think just being really clear on that because I think we have to we have an opportunity here to really like create mass adoption of crypto and web 3 because the centralized agent frameworks OpenAI and Claude they haven't figured out payment orchestration right so we have really an ability to create mass adoption here but if we we're going to keep digging our own rabbit holes and not being clear so when do you think AI agents that control their own wallet are going to be larger in number than humans. Yeah, I think the question is like a oneto- many relationship like is that possible and I think it's the same if we go back to the paradigm of the internet like I I myself can own 10 websites so yes in a future I will own 10 agents so I think it's just about is a one to many relationship possible yes it is um is it probable I want to get a year from people tell me what year is it going to outnumber us I think we are expanding quite quickly I think in the next year and a half we're going to have like commercial buyin from web two where the transaction volume already exists. And I think personally my bias of course is like shifting the volume of e-commerce into a commerce, right? When you have that order flow volume, it will then power all like the innovation under it. The research, you know, I'm I mean I'm affiliated with Cambridge University.

I'm not here just like shilling a commerce product, but it's like we often in crypto like to go research first and innovation and not use case first. Yeah. Yeah. Ben, how do you identify a wallet is AI or human? Well, you've come to the right place.

Um, no, actually there is no good answer to that. So, maybe today already you have more AI wallet because the simplest thing to make money with is to get abuse the airdrops of all these of all these, you know, projects. You train these a agents, you create millions of wallet and just get it done, right? I'm pretty sure my um AI reply guys already outnumber my human reply guys on Twitter. So maybe it already happened.

So yeah. Yeah, you know AI is here at Nansson to track these numbers cuz you're asking us what the numbers will be. We're going to rely on you. I mean you're going to teach a so if you're going to do any type of airdrop you're going to face the problem. How do you identify real human or versus bunch of AI and you're just giving your airdrop to a single person?

Right. Yeah. It's a hard problem. Yeah. And I think that's where like again token of economics you compare it to real world usage like actual transaction volume actual commerce.

I mean like for us TGE candidly probably shouldn't say it but it's like a pretty much been a side afterthought. Yeah I I do like your approach of kind of thinking specific use cases like when will most of kind of retail shopping be done by agents or through agents. Right. Um can you talk a bit about maybe start with Ben in terms of your product? Do you have any plans of making your product more AI or agentic?

Yeah. So the reason they get me on this panel I think is because BBIT is probably one of the uh most adaptive to AI technology. So we re we long ago about two years ago we we were the first one to launch uh trade GBT where is a chat box we have the chat GBT trained and uh we actually put the BB trading data inside. So if you ask any token training staff they will give you the the wills uh data of these analysis in bibbit and kind of the overall and recently we incorporated with uh our partner uh social value to give a full uh AI analysis on any projects not only not only listed uh so any projects uh you go a quick dirty summary um of all the essential data. So our clients loved it because you know people like to trade new token listings and mostly most time they don't know what the hell it is and they want a quick they they see is listing coming out on five minutes they they need a short one minute for it to decide do they go yolo 100 100x or not so they they need a quick summary and that's what we provide on that so that's really good and also internally we use AI for many things uh developing testing legal translation um translation was a huge net mirror before AI um now it's it's been trained to kind of use the crypto native language for different it was it help a lot of uh efficiency issues Jensen your product obviously is very agentic and AI oriented already uh can you talk a bit about what's coming in terms of development on the product side the the token has been pumping is there some uh some secret product launches coming up or yeah alpha alpha no but I think I think last last year um almost around the same time I was on the stage on the panel and then they asked like okay you know what's what's the projection for the year forward right I say one word I said agent five right no one had a clue we never had a clue and then now you realize oh like the idea of tokenizing the productive output of an agent is now um something that we have right what's next right the idea of tokenizing the collective give economic value of multiple specialized agents working together, right?

I have I have you can a project A could build specialized agent A, project B could build specialized agent B, C, D, so on. Can when they all work together, can the economic value produced be greater than the sum of its parts? Right? And if the economic value is greater than sum it parts, can we then tokenize that cluster? Can we tokenize that?

This almost sounds like a DAO of agents. Is that how to think about it? Or in a can they own Dows and create a token and and we thinking more of like from a product angle, right? If these agents coming together, they're creating a product. The that product is represented by this cluster.

I mean it's a it's by the cohesive effort of this cluster that can be tokenized. That product that is built autonomously by this group of agents can be tokenized. So we call that an autonomous business token. Cool. One one year from now I'll come back here and did that prediction come true again.

Leah on the product side what can people look forward to with Bit GPT? Yeah, I mean definitely not here to talk about the TGE because that's if anything I'm really excited about revenue and I and I just want to stress that like as a product that's an AI solution, not a blockchain, it really was an intentional decision. Do we even do tokconomics? Should we do equity? And I think at the end like reverting back to what tokens were also initially poised for which was like we want to give people a collaborative insight and incentive to join.

So we want builders to build decentralized AI because we believe in sovereign AI. Okay, cool. That will like in itself grow. Tokens will be distributed. But I'm really excited actually of saying like okay, we've built a privacy preserving mechanism where you can have decentralized commerce.

I mean that's exciting, right? Like all these problems that we had with like companies owning your data, social media shilling your data. If we have decentralized commerce, maybe one year from now that will be a thing. I don't know. Um that's what I'm excited about.

And like that's actually like you know talking to Gucci Prada Emirates and letting them choose a decentralized solution versus open AI and Claude that's what I'm excited about and that's really what I'm trying to do and I think as an industry like I really value you know that we're all approaching it differently attention economy of how do we even like bring attention to decentralized agents and there's so many valuable different approaches that all have to come together in collaboration. I think that's a great way to end this panel. Thank you so much to the panelists. Thank you to the audience for coming. Enjoy Dubai.