Haseeb Qureshi：Agents 不需要 Crypto，人類才需要！

Thanks for having me. How are you doing?

So Token 2049 at this point is the largest retail conference in the world. It's, I think it was 25,000 people who made it out this year. And of course this is the mega event in Asia that takes place every year. That said, I think probably the best barometer for Token 2049 is Marquee. You know Marquee?

That's right. That's right. So it's like in the lower floors of Marina Bay Sands. Every year, Marquee is booked out pretty much every single night during Token week by some shitcoin project. And this was the first year that Marquee was just a normal nightclub the entire week. No crypto projects booked out Marquee. And I think that more than anything is probably the sign of the times of the ways in which crypto has changed.

Is that crypto has become, even though there is still retail activation, there's still volatility. You know, the stuff that I saw was doing well during the course of Token. You know, you saw Near, you saw Pearl, you saw Quantis. There was a lot of attention on different projects that were in the limelight during the event. And of course, there was stuff about Ethereum and Bunker Bode and cryptography that was a big conversation during the event. But the main thing was that the crazy speculative energy of crypto, like you said, it's not here in the same way as it was in previous years. And even though markets are starting to come back and Bitcoin is finding its footing, probably one of the biggest events that we saw during the week was this closed-door institutional event at Token 2049 where a lot of hedge funds and asset managers and founders and regulators came to Singapore just for that event.

So, you know, you go out the booths, you see the energy, you see the lights, you see the, you know, there was a few places where you have like cartoon characters wearing outfits like it's a sports mascot walking around the floor. So you've got the Asian retail marketing mentality. There's a DJ on every floor, there's nice food, there's a lot of talks about different projects, but the craziness of prior crypto cycles is the lowest that I've seen in a long time. And I think that's a sign of where things are going even for the future.

I mean, clearly 25,000 people thought it was. So I don't know. It might be a little premature to say there's no value in going. The value is different. I think in previous years you would go because it's a giant party and you know, you come out there, you go to the nightclubs, you go to the events, you go to the after parties. There were still events and after parties, but everything was much more tame this year than in previous years. Part of it is also that the industry is getting older. You know, people are more mature. You know, people maybe who entered into this industry when they were in their early 20s, you know, now they're entering into their 30s, maybe they're getting married, settling down, having kids. So there is some sense in which even the cohort that I came into this industry with, they're just in a different stage in their lives where a lot of them aren't going out to the nightlife anymore. So there's not the sense that everybody in this industry is in party mode all the time every time they see each other. So I think that's also an element of not just compositional effects that more of the people here are institutional rather than retail, but also just even the retail that's here has gotten older and has a different reason why they're coming to a conference like this.

All that being said, every time there is a conference like this, there's some alpha. Everybody who I was talking to at this conference was talking about Pearl, was talking about Quantis, talking about what's happening with Near. So there's always some edge that you get from being in the flow of things. And there's a lot of flow at a place like Token. But yeah, it's not as crazy from a cultural perspective as it's been in years past.

I did not see that. I saw Ilia, but did not see him dancing.

Yeah, around the beginning of the year, many people, myself included, were saying that crypto is going to be necessary for AI. And once AI agents really start becoming fully autonomous, that you're going to see stablecoins be the dominant way in which they make payments. That argument, actually, it's a very old argument. I think I first was talking about this in 2023 after GPT-3 came to four. And this year, I mean, so many people were telling the same story. You saw Stripe talking about this, CV was talking about it, Brian Armstrong was talking about it. And the reality is that at the beginning of this year, we really didn't have AI agents. Beginning of this year was when cloud code first burst on the scene. We saw Opus 4.5 first start to show people what agentic coding can do. But agents really were theoretical at that time. Today, they're not theoretical. Today, if you see what's happening with Muse, with Instinct, with ChatBT's dot, with Grokbot, agents have now completely disseminated and gone vertical with respect to their enterprise adoption and even adoption among consumers. And it's time for us to wake up and pay attention and say, hey, the agents are here now and they're not using crypto.

They're mostly using cards. And you saw both the card companies like Visa and Mastercard as well as companies built on top of them like Rain, like Ramp, like Stripe offer agentic card products so that AI agents would be able to use cards, so that you can have spending limits, you can have risk limits. All this stuff that you would want when provisioning money to an AI agent is already live with cards. So the reality is that this story we were telling of like, oh, you know, the card networks will never adapt to AI agents, they, you know, they don't have a driver's license, they don't have a social security number, how are they going to be able to pass KYC, how are they going to be able to use these traditional payment rails, the answer was that no, no, they beat us to it and they knew that this market's going to be so big that they have to figure something out to make sure that card networks are going to be able to work for agents. Now that's not to say that all agent payments will run through cards. Obviously this is mostly for consumer payments. If you are going to do the kind of thing that you do with crypto, probably agents will use crypto for that the same way you use crypto for that. So if you want to send international remittances, if it's most efficient for you to use crypto to do that, it's probably also going to be true for your AI that your AI is going to realize, oh, I should use crypto to do that. But it's going to be able to flexibly use crypto when it's the right tool for the job as opposed to use crypto all the time, the same way that you don't use crypto all the time. You only use crypto when it's the right tool for the job. So we should all admit that we got that wrong.

And okay, we can all admit that. So now we can move on and answer the deeper question, which is, okay, now that agents are here, what do they change about crypto? And the answer is quite a lot. Of course, they change the interface of crypto, is that increasingly you are not going to be pointing and clicking on your desktop to do things on a blockchain, right? You're not going to be manually bridging and checking, oh, you know, I need to pay gas over here, I need to rebalance my, oh, I need to find yield. How do, let me go check what it is on a, go check on Morpho, do this and that. That's going to go away. That's going to go away in exchange for you telling your agent to go do those things for you. Your agent is going to become like your crypto guy. You know, like when people have a crypto guy and the crypto guy manages all the crypto for them, that's going to be your agent. Your agent's going to be your crypto guy. And in exchange for that, you are going to get much more efficiency and much more optimality when you are dealing with, hey, how do I get the best yield? How do I get the lowest trading fees? How do I find the best way to accumulate this asset or send this money to this person? And if that's the way that we're all going to use crypto, it means a few things downstream. One, it means that we are all going to end up having much better security than we do today.

People today are kind of scared when they use crypto. You know, I know certainly I'm scared when I'm sending a large transaction. That fear is going to go away because your AI agent is not going to make the same mistakes that you as a human being make. So your AI agent, you know, maybe today obviously the agents might make mistakes, but today they're the worst they're ever going to be. They're only going to get better with every passing model release. And when the AI agents get better and better, it's much like self-driving cars. Self-driving cars now are better than humans. Five years ago they weren't, but today they are. In the same way, your AI agent is going to be better than you at not getting scammed, at not getting fished, at not putting its money into a honeypot, at not, you know, backing something that ends up being a scam or a Ponzi or a rug pull. It's going to be way better due diligence. It's never going to get lazy. It's never going to FOMO. It's always going to go do the work, read the smart contracts, look up the LinkedIn, do all the background research, see what people are saying on X. It's going to do all those things every time it interacts with a smart contract. And as a result, crypto is going to become more trustworthy. The feeling that most people have like, man, you know, crypto, it sounds good in theory, but in practice, it's full of scams and hacks, and it's so easy to lose your money. It's so easy to lose your keys. That is going to go away. And our experience with crypto will actually be that it's just as reliable, if not more reliable, than traditional finance or traditional banking rails.

The last point that I make in this talk is that the other thing that AI changes is privacy. AI is very bad for privacy because it means that, you know, one, whenever you're sharing your information with an AI, you're sharing it with a third party. And because you're sharing with a third party, being, you know, the lab that runs the AI that you're using, it means at least in the US that you have now lost your Fourth Amendment protection against unreasonable search and seizure. This is known as a third-party doctrine. It means once you share your information with a third party, it means, well, you must not really value your privacy if you're willing to share this with somebody else. And so that means in practice, one, if you share something with an AI lab, not only can they train on it, but also you lose your Fourth Amendment protection. That means that now anybody, any law enforcement can pull that information, they can pull all your chat logs without a warrant because the third party has it. You gave it a third party, therefore you must not value its secrecy. It means that if you're using an AI for legal advice, let's say, that means that if you get sued, then you have to show that chat in discovery. It's not attorney-client privilege. It's not your personal private affairs. So all of that means your opposing counsel gets to see every conversation you've had with your chatbot. All of this shows us that the way that we think about our AIs, we think about our AI as though they're extremely private, as though they are our innermost chambers of our minds. They're like an extension of our own minds. But that's not legally how they're treated.

And so as AI strengthens the state, it gives the state more powers over individuals. It also, of course, increases the ability for the state to do surveillance on a massive scale. So, you know, this is one of the things that in Anthropic's showdown with the Department of War, they were saying that we do not want to let the government use our AIs to deanonymize people en masse, which turns out is very easy to do without breaking any laws because there's so much data out there on the dark web from data vendors and data brokers. All you have to do is buy some data and look it up instantaneously. Crypto fundamentally is going to help you, the human, not the AI, but the human, to protect yourself against the increasing encroachment of privacy from the state. And that's going to happen both through, you know, using privacy-preserving cryptography in the protocols themselves. So things like Zcash and Monero and so on, but also it's going to come from using private AI. So we're investors in Venice. We're also investors in Zcash, but we're investors in Venice as well. And Venice, of course, uses blockchain rails to give people privacy-preserving AI inference. And I think this whole private stack is going to be a counterweight to the way in which AI is strengthening the state at the expense of the individual. Crypto allows individuals to push back. And that's why I say crypto is not for your AI agent. It's for you. And your AI agent will use it on your behalf, not for its behalf.

So we're investors in Venice. For those of you who don't know, and Sapium as well. That's right. So Sapium is less crypto. They are an AI agent orchestration platform. They allow AI agents to pay anyone as well as consume services entirely using stablecoins on-chain if necessary, or they can also do payments via cards and APIs. Venice, I think, strikes more at the core of the intersection of crypto and AI, where Venice is a privacy-preserving AI inference platform. You can think of it as like private ChatBT. Now for Venice, so actually Venice is not itself on-chain. They use actually Neo protocol and Fala to source their GPU cloud. So if you want to do fully end-to-end encrypted communications with an AI model on Venice, they source the TEEs that they use for that from Fala and Mir. But the main kernel of why you actually need crypto for privacy-preserving inference, so if you use Venice today, you actually don't need to do it in a totally private way. So you can go on Venice, sign up with your G Suite or Google login, and you can pay with a credit card. And if you sign in with Google and you pay with a credit card, then obviously Venice is going to know who you are because they can see your name. But even though Venice is zero data retention and they allow you to use these models fully encrypted if you so choose, you don't have to use Venice that way. If you want to be fully private, you can use Venice almost like VPN mode where you log in using your crypto address. So you log in using an Ethereum wallet and you pay for your AI services just using stablecoins, or better yet, just using DM, which is their tokenized inference token, which allows you to gain a certain amount of inference spend every single day if you own DM. So that's the sense in which actually this privacy story only really works with crypto.

If you don't have crypto, there is no way to guarantee privacy. Like I cannot promise you that I'm not going to be able to know who you are if you pay for this service using a credit card. Actually, if anything, I can guarantee that I will know who you are. And if somebody comes knocking, I of course am going to have to have a Stripe balance or, you know, an open Stripe account with your balance there. Maybe I don't know what you're doing with your GPUs, but I know who you are on Stripe. But if you pay with crypto, then I literally know nothing. I don't know what model you're using. If you go end-to-end encrypted with the model, I don't know how much inference you're using, except just, you know, to the extent that I'm charging you. And I don't know who you are. I don't know what your name is. Don't know where you live. Don't know anything. And so that I think is the apotheosis of how crypto can assist AI on the privacy side for consumers.

Correct.

No, the answer is very little agent-first services are being consumed today. What you see now that agents have really started to go mass market is that the vast majority of usage of agents is concentrating in consumer-type activities. So, well, of course there's enterprise coding-type use cases and that's where most agents are being used today. But if you look on the consumer side, most of it is being used for things like travel. So I believe I saw some stats recently that you look at Muse. Muse today is by far the largest consumer AI agent that's launched by Meta. And Muse right now, I think it's something like 40-50% of all the spend going through Muse is happening on travel. So people booking flights, booking tickets, booking hotels, this kind of thing. The kind of thing that a travel agent might otherwise do for you, and certainly did in times past. So it's very simple automation that people are using Muse for today. But I think it's something like 35-40% of people who use Muse have attached their credit cards, and for those who do, the retention is actually quite high. I think their retention is pretty competitive with ChatBT's retention, which is of course a very, very retentive product. So it's very clear that at least for now, most of the focus is on these relatively self-contained consumer tasks, but there's a lot that's going to be changing about the way that agents interact with finance and money.

So just to give you an example, last week there was a viral story from Apollo about the idea of an agentic bank run. And an agentic bank run, it's great marketing for a term. But what does the term mean? What it means is that let's imagine that you have an AI agent like Muse and it starts to optimize your finances. So it's going to go and look at, you know, oh, I have this money in my checking account and I have this money in my 401k and blah, blah, blah. And it's kind of like a financial adviser. Well, the first thing it's going to look and see is that, oh, you have all this money sitting in your checking account and it's earning nothing. Especially if you're hosting it at a big bank, if you're hosting at JP Morgan, then it's like really earning nothing. Even if you put it in a savings account, it's earning nothing, you know? So right now, the Fed funds rate is almost 4%. And if you hold money in a JP Morgan checking account, you're making something like 25 basis points. You're making less than a tenth of the Fed funds rate. And so your AI agent is going to quickly tell you, 'Hey, what are you doing? Why are you keeping money here? Why don't we sweep that money into a money market or into a bank that pays more yield, and then whenever you need it, we'll move it back and then we'll pay out of it.' And basically, like, why aren't you already doing that? Why isn't everyone already doing that? And the answer is because we're lazy. Of course, we should. There's no reason why we shouldn't. It's just that we're lazy. It just takes time.

Well, agents are not lazy. Agents are very cheap to run 24/7. You can have them wake up every hour and go check what's going on with your money and make sure it's earning the most money, and they won't be lazy. And so any business whose business model rests on a tax on laziness, like that's why they make money is because their users are lazy and they can just sit there and rake in the dollars. That is a huge part of how these big banks are earning their net interest margin. They have very cheap deposits because consumers are just having their money parked there and they're too lazy to move it, and they can lend out against it at super attractive rates and they collect the spread. And so most of the profit, most of the return on equity that banks get comes from the fact they have such cheap deposits. The same thing is true of a lot of brokers. So if you look at Schwab or Fidelity, Robinhood less so. Robinhood makes most of their money from trading fees. But a lot of the brokerages don't make that much from trading fees, especially now that the fees have become much more competitive because of Robinhood. Most of them actually make the majority of their money from people having their cash sitting there in a sweep account and not moving it. That's how they make most of their money, like most of their net profit. And so that business model is going away.

And now you can imagine one way it goes away is a financial crisis because people move all their money out of the big banks. That's probably not what's going to happen. More likely what's going to happen is that as they see AI agents are pulling their money out of the big banks into other banks, then what they naturally should do is to just start raising rates. They should just start paying more to keep the deposits at JP Morgan or at Bank of America or whatever. Now that's path one, is through competition, right? But actually that's not the only way they can respond. There's a second way they can respond. And that second way is to just ban you from using an AI. And now this, we've already started to see this. So if you look at Muse, Muse can book you a DoorDash, it can book you an Uber, it can, you know, use all these different services, it can buy stuff on Walmart, but it cannot buy things on Amazon. Why can't it buy things on Amazon? Because Amazon decided to block Muse a couple weeks ago. Now, why would Amazon block Muse? Why don't they want people buying stuff on Amazon with their Muse? The answer is no different from the banks, because Amazon makes most of their money through advertising, not through actually selling goods. Amazon has very low margin on their goods, but when you go to amazon.com and you look at the sponsored links, those sponsored links cost a lot of money. That's how Amazon is making almost all of its net revenue, is coming from those sponsored links. So your AI agent doesn't look at sponsored links. Your agent doesn't care about sponsored links. Your AI agent goes to Amazon, they'll say, 'Great, I'm going to ignore all these and just go see what has the best reviews.' Like, it'll do exactly what you want it to. But that's terrible for their business model. And so that's why Amazon decided to block them.

Well, if Amazon is blocking them, then you can bet that your bank is going to [__] block these agents. And your bank is going to say, you know, not, 'Oh, you know, I don't want you pulling your money out of the bank.' What the bank is going to say is, 'Oh, you know, it's not safe. It's not compliant. You know what? If there's a risk, we don't want your AI agent to mess up. It might hallucinate and you might lose all your money. So we want to make sure there's always a human.' You know, they'll come up with some stupid story, but we know obviously the reality is that your bank does not want you to be able to automate them and pull your money out and bring it somewhere else where it's going to earn a higher yield. So that dynamic is also part of the reason why, you know, for these companies, a lot of these fintech that we're investors into, over time, I think you will see actually that they get more and more agentic demand because that agentic demand is going to come when the old guard of companies start saying, 'No, you are not allowed to use an AI agent with our product.' So you cannot use an AI agent with Amazon unless you use Amazon's AI, which sucks, right? So they have the...

Rufus, which is their Amazon AI, and it's a piece of [__]. But if you want to use your AI, when that's actually representing your interests, Rufus doesn't represent your interest. Rufus will sell you whatever Amazon thinks is going to sell you. But if you have your agent that's trying to represent you, Amazon like, 'Nope, can't enter the room. We're not going to allow it in the building.' You know, it's like only principal to principal. You know what I mean? That kind of business model is going to die.

Yeah, I mean, look, at the end of the day, it's probably never going to completely collapse. There will always be some businesses that can afford to keep you on a string. And you know, if you want to use JP Morgan, they're the biggest bank in the world. JP Morgan can afford to tell you, screw you. We're not going to let you use an AI agent with JP Morgan. And it sucks, but they can do that. They have the power to do it the same way Amazon can tell you, look, you're not allowed to buy things on Amazon. So, if you want to buy things on Amazon, go to the [__] website and do it yourself.

Um, that, you know, now that being said, the other flip side of that is what happens when you have your own agent that's not owned by a third party. So if you're using Muse or you're using Dots, these are not your agent in a strict sense. They're actually, they're somebody else's agent that they are leasing to you. You know, if it's Grokbot or if it's Dots or Instinct or whatever, this is owned by somebody else, and because it's owned by somebody else, Amazon can decide to block that person. They can decide to block that vendor, right?

But if you have your own AI, if you are running, for example, you've got your own local Hermes agent that's running on your machine and it's being run, let's say, with some private inference using something like Venice in the cloud, then now you have your own agent running just for you in your own environment and nobody can force you to be blocked. So, if you're on your own machine and your machine is opening your browser and the agent is moving around and clicking on your browser, then you can buy stuff on Amazon, even if Muse isn't allowed to because they have an agreement with Facebook, where Instinct isn't allowed to because they have an agreement with Instinct. Um, they can force them to stop, but they cannot force you to stop. And so, that's a big part of the reason why I'm long-term bullish as well on local agents.

Um, and so, you know, we're investors in Nous Research, the company that launched Hermes agent, and Hermes agent is the prototypical example of this. Having that local-first deployment allows you to do things that you will not be able to do at steady state with an external party running your agents for you.

I think so far my predictions have been pretty prescient. I've gotten a lot of things right. Got a few things wrong. I'd say first, uh, so obviously Bitcoin unfortunately pulled back quite a bit, um, in 2026, but I think that you do probably still see BTC over 100K by the end of the year. Uh, 150K is probably going to be really hard given where Bitcoin's sitting at today of, yeah, you know, almost 2x-ing.

Um, I think I got it right on the fintech chains. So I think Tempo, Arc have underdelivered. Uh, Robinhood chain still TVL, obviously it's quite early and there's a lot of memecoins. No, I think, I think it is TVL. I think there's a lot of memecoin activity on Robinhood chain. But you know, Robinhood chain is not trying to be a memecoin chain. It's trying to be this RWA chain. And the RWA activity on Robinhood chain is still quite small. So it's like a few percent of all the RWAs.

And they haven't really been growing that much in terms of base. So I mean, every, um, cycle there's some memecoin activity. Maybe Robinhood chain will sustain it for a while. Maybe it'll, you know, there was a while that Base was super memecoin heavy. There was a while the BNB chain was super memecoin heavy. So it feels like right now it's rotated to Robinhood chain, but I don't know that that's going to last a year, two years. Um, so I, and I think when you're talking about developers, it's very clear there's no developers on Robinhood chain. Um, like nobody's building applications on there. Nobody's building like new ideas. It's really just kind of memecoin farming or memecoin cycles, you know.