Will Stablecoins Own Machine-To-Machine Payments? Ft. Christian Catalini

Sy Taylor  0:10  

Welcome to Tokenized, the show focused on stablecoins and the institutional adoption of tokenized real-world assets. My name's Simon Taylor. I'm your host, author of FinTech Brain Food and head of Market Dev at Tempo. And joining me as always when we're talking about all things agentic commerce is Bam Azizi, CEO and founder of Mesh. How you doing, Bam? I'm doing great. I'm super excited for this episode. Yeah, we're gonna go deep because joining us this week is none other than Christian Cassellini, the founder of the MIT Cryptonomics Lab, the co-founder of Lightspark, the co-creator of Libra, yes, that Libra, and a former head economist at Meta. So, how are you doing, Christian? I'm

 

Christian Catalini  0:51  

doing great. Excited for the conversation together.

 

Sy Taylor  0:54  

Yeah, likewise. Before we get into that conversation, I've got to remind viewers and listeners that views and opinions of our contributors today are their own, and might not reflect those of companies they represent. And please don't take anything we say as tax, legal, or financial advice. And of course, gotta remind you that this episode is sponsored by Mesh. Okay, Christian, let's get started. I mean, I read out the CV there and some of the jobs, but it might be the most interesting CV in all of finance. Like MIT professor, co-creator of Libra, economist at Meta. What do you learn on that journey? You know, like what are you picking up along the way? Give me the potted history here.

 

Christian Catalini  1:32  

Yeah, I mean, the famous quote goes, "You can only connect the dots looking backwards. I'm not at that stage, so when I look back, it just looks like a game of ping pong from academia to big tech to startup to now really trying to figure out what what we're doing next. I would say every piece of the journey shaped me in in different ways. I did love my years at MIT. I mean, it's an amazing environment being next to researchers and and trying really to get on the frontier with with a lot of really talented individuals around you, starting with with our students, many of which you know ended up going into crypto early on. Big tech was probably the one stop that was less of a fit. I I find large organizations to be complicated and and political and and sometimes slow. Although you know Libra was very different at the time. I do love building, so the startup part is is probably the one that I will get back to at the right time. Right now, it's just a very strange time to be thinking about what to build because on one side it's probably the best time ever, and many have said this. Right, the tools, what you can even do as as as a small group of individuals is like 100x what was possible before. On the other side, when I decided to go part time on on my startup Lightspark, I felt like okay, with more time, I will be finally able to grasp what's happening on the frontier. And every day that passes, I feel like the closer I get to it, the more it's just rapidly advancing. It's it's it's literally crazy how quickly things are moving.

 

Sy Taylor  3:01  

You feel the same about that stuff, Bam. I'm sure.

 

Bam Azizi  3:03  

Yeah, everything is moving 10x faster.

 

Sy Taylor  3:07  

It's kind of nuts, isn't it? The fact that you held on to all those jobs and all those times and still stay curious, Christian, is is I think what's what's fascinating me. You write a lot on Forbes, and you've been writing a lot about agentic commerce lately, where did that passion for agentic commerce come from? Like, what what gave you the bug for agentic commerce?

 

Christian Catalini  3:29  

I think I got the AI bug about a year and a half ago, and it was most an existential bug, so not very original, right? The low grade fever that many of us have felt, and you know, economists when when they do get that kind of fever, cheat. And the way we cheat is very simple. I did the same when I got passionate about crypto back in 2013, which is like, here you have a fuzzy object coming your way, lots of potential, lots of you know intriguing opportunities. But how do you make sense of it? And you know, economists trying to try to boil things down to cost. If you can look at that fuzzy thing and really identify the one or two fundamental costs that are driving it, then suddenly you're like, okay, I may be you know wrong for a good part of it, but directionally, I kind of feel where this might be going. We did that exercise back in 2013 with crypto. I think we did get probably 80% of that right, and and of course the the 20% is the one that matters. As as you realize once you're building with AI, it was the same. If you want to look at your time, if you want to build something next, like what's worth still building? Why should you you know leave AI labs to do or the fast moving fintechs to do versus what can a small startup do? It was a journey really guided by by that personal question. Also, I I have kids, and so as many with kids, we're all like wondering, okay, what should they be learning? What should we nudge them towards? Right, every everything that that was known before is now on the table again, and Ajanti Commerce did. Look like the intersection of that that existential fear and what I've learned. I guess along the journey within crypto payments, stablecoins, individual assets. So it's obvious to me that we're going into a future where, the same way most traffic on the web today is agentic, most commerce, most interactions will be agentic in some shape or another, and really trying to tease out what's the hype and what's actually the ground truth of why is this interesting. I mean, Sai, you've written way more than me on this, and I love your decomposition of the different protocols and stacks. There's just so much fluidity in how different players are coming at this, which makes it really fascinating. And then look, when it comes to Genty commerce, for me it's almost like the the old historical lesson that every payment revolution is a revolution in allocating risk. When you think about the card network, right, the the airdrop in Fresno that Bank of America did on the merchant and the consumer side, then later on eBay trusted sellers with PayPal, right? Online commerce kind of slowly becoming safer and comfortable. We need the same for agentic, which is like who's gonna take the liability when your agent goes out and does things. Hopefully, it's not gonna be like the agent from OpenAI that you know decided that knocking on Hugging's face door was the best thing to do. But we we were missing a piece there, so the missing piece is the part that that I find fascinating.

 

Bam Azizi  6:24  

Kristen, I wonder based on what you're seeing in the market, what is the role of agentic commerce in the world of AI? So, what does agentic commerce mean? And I know that you have a controversial view on that end, like what we have today is knowledge in e-commerce in a true sense. Why is that? Why why that mat? What that definition matters, and maybe you can talk a little bit more about your point of view there.

 

Christian Catalini  6:53  

Yeah, I would say probably something that's I don't know a contrarian, but definitely something that will not make some startups happy is that I do think that a big chunk of agenty commerce it's going to be actually boring iteration of what already exists. That's where you know fintechs like Stripe and and others, the credit card companies, anyone that really touches money today on behalf of a consumer is going to do great. It could be a neo bank, it could be traditional institutions, you know, once they get their act together. But like the idea that if I want to book a travel hotel or something, it's going to be some some weird crypto rail and stablecoin. No, right. So that's not what what consumers expect. And so I I think that chunk of agent e-commerce, which is essentially doing something we're already sort of doing today with a lot more friction and time, I think it's going to go the way of the incumbents. Distribution there is just going to play such an important role, and also the UI UX of of crypto is never is never you know up to par on day one. It can get close, but it's it's it's not really a substitute. The part that I do find intriguing though is machine to machine, and in crypto, it's funny, right? We we had all the right ideas at all at the wrong time. If you look back at when Balaji was building 21, he had this whole idea of like machines transacting for resources, paying humans to do tasks. Nothing came out because it was so early. But when I look at x4 102, you know Cloudflare and many others coming around the idea that okay, the web is broken. How do we monetize data? Is there a better model there versus agents going around and capturing resources, data, you know, activity, whatever might be needed to deliver a service or product? That's where I do think a lot of the crypto infrastructure will be quite useful. So final bifurcation between kind of the boring volume that we already know, I think is going to go the way of the traditional systems, the new stuff, the white space. That's that's where I think you know crypto rails and stablecoins can play an important

 

Sy Taylor  8:50  

role. Where does that tipping point come for machines buying stuff? Because x4 or two has largely been a meme coin story, like most things in crypto. Sadly,

 

Christian Catalini  9:00  

as always, right? Yes,

 

Sy Taylor  9:02  

you know, and and but but I have seen a few dashboards now that are showing some signs of life. We've had the guys from AXTP on the show before, and they're sort of seeing you know real volume of buyers, not just like sellers or people doing buyers and sellers that are really the same people and circular stuff, but like consistently buying from multiple merchants. But that's a sort of a skills library that exists in that some people know about that they can install inside their harness for some of the time. Like it's a fraction of a fraction of a fraction of the users of Codex, cloud code, and cursor. So, where does this start to get any sort of traction on the machine-to-machine side?

 

Christian Catalini  9:48  

Yeah, I think I think it's extremely early days. A lot of people, I think, are also ignoring that the business model doesn't make sense. Right. One of the reasons why we never got micro payments for humans. Has nothing to do with the tech stack. I think it was in 2022 when we released this idea that you could stream money at a fraction of a fraction of a cent on Lightning on Bitcoin of old systems. And I was having long conversations with our engineers that thought, okay, this is super cool. And I'm like, yeah, humans love micropaying for stuff. Imagine you're watching a movie, and every second you're reminded: is this good enough to continue or not? A lot of that changes for machines. The thing that makes me bullish about eventually a market for information is that if you think about what makes AI capable today, is a combination of two things. Of course, we have the base models and all the progress happening there, but then in the end, and you're hearing this also from like Satya Nadella when he's arguing about you know a broader ecosystem around intelligence, more distributed intelligence. What you need for that agent to be useful is some sort of proprietary data, tacit knowledge, information that's either in in one of our brains, right? Any expert in a domain has calibrated weights through experience that are different than what's in the public domain, and it's in the models or historical track record. That's why I think you know the likes of Stripe and and even the card networks have been able to train these new AI models. Ramp has been doing a lot of progress on this. Once you do have that rich data, I think you can build fundamentally new services, and that telemetry, I think, is going to be extremely useful. Also, because when you think about it, when we can measure something, not only we can actually automate it. Ironically, you can't automate something that that you're not able to measure, and we we can go deeper into that. But like the second part is, if you're really able to automate something end to end, like a full workstream, the work of an actual person, which is kind of the old promise around here. Then, you know, what you're missing is the ability to also underwrite that. We talk about agentic commerce, but there's many aspects where I think people will want to delegate to an agent, almost as if it was like a digital worker. But that agent needs to be liable for it. That agent needs to be able to underwrite the risk that comes with all of those actions, and so when you put that all together, I do think that whether it's x4 102 or many of the other competing protocols, I mean, there's there's fuzziness right in where the the industry land something that's programmable, it's interoperable by design, it's neutral, it's crypto native. I think will do very well for these new markets for information. So I have no idea what the killer app might be, but there's going to be something there, probably important.

 

Sy Taylor  12:25  

You're going for the auditability benefit of the thing. I'm reminded of a company at the moment called the AI underwriting company AIUC. Really interesting group who are heavily involved in as consultants putting together the insurance liability framework that sat around self-driving cars, because self-driving cars have been better than people for a while, but they fundamentally didn't start to get any approvals anywhere around the world from any regulator until there was an insurance framework that sort of sat around it. So liability is such a crucial point. Bam, bam. I think you had a question as well.

 

Bam Azizi  13:06  

Yeah, my my question for Kristen was around like his view on the micro payments. Do you think that's the killer app for agent e-commerce? Like I know that you don't accept an agent paying with credit card to a shop as an agentic commerce, quote unquote. Maybe you can explain that more. But also, do you think microtransaction would be the wedge that agentic commerce is going to grow there, and then slowly replace the credit card and traditional card networks on regular purchases?

 

Christian Catalini  13:39  

I would say, look on the consumer side. Again, the consumer product is the one that's going to decide how that is experienced, and I don't think you want to change human behavior beyond. Okay, this is my goal. Go and do it. I think we're going to see a lot of that, right? But your Amazon, your PayPal, you know, pick your wallet. We'll do a pretty good job at it. On the machine to machine, I think what's missing is is the market, right? So, to Sai's point, I think we're seeing some volume, but I've yet to see even on the web monetization problem, people have been banging their head there forever, right? If you if you look back in crypto, I think it was the Brave browser that had designed this whole three-way economy between advertisers, users, and and and essentially the ecosystem to reward you for your attention and give you access to content on a micro-payment basis. I think a lot of that is really challenged because look, there's a reason why digital goods get bundled, right? Your Spotify subscription, your newsletter subscription, all these things take advantage of the fact that the marginal cost of replicating the digital information is practically zero. So once you've created it, you just want to have as much distribution as you can without undermining your own demand. And micro payments don't do that, right? So if I can read the one article that I may care about, you know, once a week in the Wall Street Journal. Good luck subscribing to the IRTR subscription that they're able to price at least part of the demand curve on. So I do think that even on monetizing the web, the key is going to be who can get the ecosystem right. I think some companies they're trying right parallels. There's a few of these that are essentially saying let's get the publishers to buy in, so your agent can access better data. Nobody has cracked that, and I think part of it might also have to be that for an existing publisher, it's very hard to play under the new rules. So maybe it has to be a whole new industry category that gets designed around this new economics and new mechanism serving agents rather than you know eyeballs.

 

Bam Azizi  15:40  

But this is like pre AI, right? After AI, there was like big controversy around like OpenAI and Anthropic and all these trendy names. They were like parsing information for free now, and they package it and sell it as tokens on the other side. So, do you think at some point AI push will change that economy model in a way that anyone who has any piece of information on the web can monetize it by selling it to these machines.

 

Christian Catalini  16:09  

I I do think, and you're seeing. I mean, right now it happens with players like Merker and others that have created essentially a cottage industry of like, can we get new proprietary data for the AI labs to train on? That's kind of the state. They hire expert. They have you write evals. That that's how it's working. I do think eventually we may have something better. And you're right. I mean, if all my content gets absorbed and redistributed without attribution, the old trade-off was like, okay, free to access, but I get distribution, and then maybe I monetize through some other business. If that's taken by the model, then of course, why would you play? That's what essentially I think Microsoft, Palantir, even Salesforce, many have come out saying we don't want a world where the big labs kind of absorb it all.

 

Sy Taylor  16:52  

Well, and also it's it's a it's an ROI question. So as a as a content creator, I'm fairly well citable now inside most of the LLMs, there's probably I don't know good million words of mine on the internet, something like that. I spam you all ridiculously, and frankly, that's all upside. Like I give away the content mostly for free, and I sell ancillary services around it, and that's far more profitable than a micro payment, so ultimately it's that return on investment for my time of the association to the brand that I think a lot of people want, and for some products that works, for others it doesn't. I wrote a piece called "The Intention Economy, though, where that you're doing something different when an agent arrives. It's it's usually given some sort of mandate or it has a mission and it's trying to get to that mission as effectively and as cheaply as it can. And if you look at Parallel and Exa, and you know the the sort of the data serving startups that have done really well so far, they are able to do better search quality with lower token costs than the traditional tool calling inside LLMs, and so I suspect that at some point, as we go down this road of like we don't have enough compute, we don't have enough money to pay for the tokens to the big labs, and there's just this real focus on efficiency. That push for token efficiency pushes the agents to be much more token efficient and buy from whomever in the marketplace is token efficient, and so there's this marketplace effect that that starts to emerge there. I don't know if that's five years or 500 years away, but you you can follow the chain of first principles logic to something like that, Christian. And interested in your thoughts on this idea of agents having intents versus and mandates, and how that fits together, and then I'm going to throw the S word at you. I'm going to say stablecoins and see how you react.

 

Christian Catalini  18:51  

Yeah, look, I think you're spot on. The intent economy is is quite interesting, also because it's one that has big repercussions for some of the big historical web aggregators from Google to others, right? And it's really unclear how disruptive it might be. In the end, again, distribution is quite powerful, and so if I have the entry point on distribution and I can best serve your intent, will we have an economy where everybody has their decentralized agent? Probably not, right? The entry point is going to be one of the big, big companies that they already know, but I think you're hinting at at a really interesting tension, which is like intent also means absence of steering and continuous verification. I don't think we're at the level where you can delegate these agents to do some complex things and and trust them. You you can still do fairly simple things, but as we're seeing, they're probabilistic systems. To your point about efficiency, efficiency is going to be a mix of like lowest cost, best ground truth, which delivers lower cost, right? Because I can point you to actually what you need, and also can I ensure that you know we're not going off the guardrails. So there's going to be a whole fintech wave. Since we're getting into stablecoins around building the best rails on payments on financial services around these agents,

 

Sy Taylor  20:07  

I think that economists' equation for services that win when agents are the buyer is your next Forbes piece because that is is kind of again a set of ground truth principles that you can base yourself in, that that's really powerful. So yeah, I did mention the S word. Where where do stablecoins come into this? Or you know, is Bam just sort of like cluding stablecoins into every conversation because he's great at it? You know, like because he's having so much success with mesh and and all of the demand that stablecoins have latently. Are we try to clue agents into this, and actually, credit cards the better answer for them.

 

Christian Catalini  20:45  

I mean, I have strong opinions of this. I do think stablecoins are having quite a moment, and it's probably the second product in in the crypto ecosystem to to show loud product market fit. If you think of you know Bitcoin as this new type of digital asset that people care about. What's fascinating to me right now is almost like the battle on the horizon between stablecoins, tokenized deposits, any other sort of cash and cash equivalent being tokenized, and what that means for the economy. I think there's massive upside there. And look, whether it's touching agenty commerce or not, I think there's so much to be done already on cross-border money movement, treasury operations, and just streamlining stuff that should not have friction. That should keep us busy for for a few years, even before the agents come.

 

Bam Azizi  21:32  

Yeah, I think I believe that Kirsten is right, but also like even the the new economy without the agents requires a new form of payment network. I think it's time for an upgrade, whether agent exists or not. I think agents are, I would say, gas in the fire, or like maybe that's not a right analogy, but I mean like they help or facilitate or they are the catalyst for stablecoin to be adopted faster. But even without agents, we need a new form of payment networks that is global, that's fast, is cheap, and can settle immediately and instantly. And I think what we had is what we have today is like 70-year-old technology running on mainframe. And I think it's time to upgrade those systems.

 

Sy Taylor  22:19  

Yeah, and they're domestic by nature, and agents are not. They're an internet-native technology. Well, before we get on to part two and we unpack some more of this, we're just going to take a quick pause while we hear from our sponsors. Stablecoin operations usually mean a wallet from one vendor and on-ramp from another, and then controls stitched together across all of them. Visa's stablecoin platform fixes this fundamentally. You can mint, move, and manage stablecoins across OpenUSD, and you remain your own custodian all in one single environment. Then stablecoin linked cards let you spend balances anywhere Visa is accepted. That's Visa, the global leader in payments, and of course, sponsor of this show. You can find out more at visa.com forward slash crypto. This episode is brought to you by Mesh, the global crypto payments network. Mesh makes it possible for anyone or any agent to get paid or pay instantly from any wallet on any chain anywhere in the world, you can learn more at meshpay.com. Thank you to our sponsors. All right, so you alluded a little bit to the economics of AI, Christian, there, but like, what gets commoditized most, and what's going to like stay super super valuable? Because I think this is like the eternal debate in AI circles, and everybody sort of uses the word taste and like common sense, and we reach for these ephemeral concepts that AI seems weirdly bad at, and yet you can't quite put your finger on. Like, how do you put an economic value on that stuff? I think it's really difficult.

 

Christian Catalini  24:07  

Yes, look, part of my journey into this actually was a reaction to everyone talking about taste, curation, agency, judgment as the answer on how we make it out on the other side. Mostly because I felt like no, that's that's not going to help us. So what what's going to help us? So we wrote a piece on this back in 2025 that had a super simple idea. You know, after talking to a bunch of computer scientists looking on the frontier, the idea was like if it can be measured, it will be automated. And and now this has made it into the mainstream. But back then, you know, people again were coming out with with all sort of wonky theories about what AI can and cannot do. If you truly believe that as long as we have digital traces, the right data recorded for any type of task, that you can hand it over to AI and automate it, then the next conclusion is that okay, this thing is going to get very abundant. So the ability to. Intelligence and execution are problems that we measure. It's is going through the roof, and what's the complement? Right, economists don't like you know ideas where oh suddenly everything is abundant. There's there's always some new bottleneck, some new friction that captures you know ironically all of the value, and when it comes to measurable and verifiable things, the complement to that is what's not measurable and what's not verifiable through a machine. So we landed on actual

 

Sy Taylor  25:28  

evals.

 

Christian Catalini  25:29  

Yeah, human verification being being the bottleneck, and it's even broader than evals. I think evals are a great example of that, but it really boils down to you know think about any profession, what the model can do, and what the human that's steering the model, giving feedback, doing all the rounds of back and forth over you know codecs or claw code, is doing, they're essentially using the weights in their brain to look at the output and realize, okay, is this ready? Is this up to par? Can I ship it? Can I trust it or not? They're essentially using tacit knowledge information that they've learned and acquired over their entire career, and you know the sum of all their errors, all their mistakes, or all the stuff that they've seen doesn't work, and applying it to what the agent did. That's the collaboration. And of course, you know, look, the bad news is that agents are getting more and more capable. So that frontier of where do you need the human to come in and do verification is is perennially shrinking, but it's really you know as simple as can I have a digital trace of what's in size brain or band's brain when they're doing their job in this domain, and that's why you know even tools like cloud tag or you know computer use are kind of dangerous, right? Because they're allowing these companies to build those digital traces that track your decisions, your reversions, your upvotes and downvotes. That that's your tacit knowledge, right? And the more of that makes it into the machine, the more the next round will automate it.

 

Sy Taylor  26:54  

I wonder if you can create like a DNA fingerprint of like somebody's tacit knowledge and and how many data points you would need on them to build a model of how the decision-making process kind of works. Because generally, humans have this really interesting set of idiosyncracies. They are unique, and those unique idiosyncracies persist. And what you need is the ability to measure those idiosyncrasies over a persistent period of time, and to define the pattern and when they emerge. So, to your point, if I'm just watching everything you're doing on your digital devices, I can start to build a model of how this brain is interacting with the environment it finds, and maybe reverse engineer those to your point. So, so what does that mean? Like, outside the the individual human for for startups for incumbents, are there actual moats left here? Like, or is is everybody just going to be praying to the mountain of Mount Dario and the new gods? The

 

Christian Catalini  27:59  

good news there. I I don't think actually if you read the economic implications of the bitter lesson, they actually work against the big labs. I I think it's just unfolding now. It's going to take some time, but it's very natural to go in a circle on moats. To some extent, you may start by looking at the models and say, okay, the models are commoditizing entire categories, whether it's design, finance, accounting. Then you see, okay, but seven months later, six months later, four months later, we get an open weight model from China or somewhere else. Does 90% of the job? All the token volume shifts in quantity to that. That doesn't mean there's no demand for state of the art, right? If you're doing cyber, if you're doing any sort of like sensitive job or patented will matter. You pay a premium, right? And and many of us will pay a premium for the very best intelligence for the most hard task. But if the commoditization is happening right after, almost like in fashion, right, where the designer brand knocks throws something out and then it gets knocked down, you know, by the copycats, then routing feels suddenly very interesting from a mode perspective. It's like okay, companies like Open Router super powerful, right? If I control that data play, if I'm the one arbitrating between all these different models, that's the aggregator, right, of the new era. Only to realize that okay, what what what's happening right now is like every app is building their own router, right? Ramp release one versel, and so suddenly the app layer looks very interesting and saying, "Okay, routing is going to get commoditized by the apps. Now you're at the app layer, and look, lots of good reasons for thinking the app layer is usually where it stops. Distribution is is typically a mode, but then you see the labs again coming in and saying, "Oh, with enough traces, we can do the the work of the app. You don't even need to leave your codecs or your claw code. So it's a dog biting its own tail. And I come to this conclusion for now. I may revisit it soon, but I do think there's there's still a few modes, and it really boils down to what makes the models really capable. I think that ground truth. So if you have access to unique information about the universe, and by the way, the reason why people obsess about agency is because agency is a powerful way to extract information out of the universe. You have to go out. You have to have that friction, as you know, Sai. You know, with regulators, for example, when you're talking about stablecoins or tokenized deposits and all of this, that's information.

 

Sy Taylor  30:19  

Oh, that's some friction, all right.

 

Christian Catalini  30:21  

That's a lot of friction, right? That's ground truth. That now, if you do your next AI move, is super useful. The second one, I think, back to the entropy in everybody's brain. Yes, probably we are very predictable, right? So I was passionate about crypto because it's decentralizing, democratizing all this stuff. On the AI side, I'm passionate about open weights because I do see a lot of the similarities from like not your keys, not your money to not your weights. But then again, humans capture entropy all the time, so yes, maybe we're predictable. But I I like to to believe that we're also updating our weights over time, and that's a big part of where we what we are. For those of you that have watched Westward, there's that scene where she's going through the library and picks a book, and and every book is a person, right? Fully predictable by by the machines. That's that's not true, but those weights in people's brain, I do think, are a comparative and competitive advantage. So that's a mode you're seeing it already with the labs. They wouldn't be fighting for talent, if that wasn't true, and so the labs are a bit of a cannery in their coal mine. And then, and again, I'm an economist, so I have nothing to contribute here. But there's many domains where the fact that they're not measurable-it's a feature, right? There's things that are all about social consensus. Think about art; it doesn't matter that one piece of art is better than the other. Does it capture? Does it become a focal point for a bunch of individuals? Do we all converge? Those things will continue being important with or without AI. Now I'm a bit more bearish on things like, oh, people care about human touch. Yeah, they do. But if the product is 100x less expensive, they will reconsider. And anyone that has experienced a Waymo or FSD does typically not regret the experience of a of an Uber or a Lyft. It's just a simpler experience end to end. It's a lot more predictable. So, what most persists? What should startup care about? I would say, look, look for friction in the real world, whether it's atoms or people or regulators just increase that flywheel because if you have information that nobody else has, that is an accelerant for AI. Hire some of the best people in the world for the job you're trying to do, the ones with the weirdest weights in their brain, and then yeah, I mean you just have to to execute, execute like everybody else, and then the the speed is is nuts.

 

Bam Azizi  32:40  

So based on what you just said, how come that this is good for new businesses and startups and founders? Because if there is no moat and the moat is talent, Anthropic and Google, Microsoft, they have most amount of capital to hire those talents. Plus, in last 20 years that AI or 30 years that AI existed, everyone knew like models are commodity or will become commodities soon, and whomever that has proprietary data can train their model better and they can act better and respond better than they are ahead of the game. And again, like Microsoft and Google and Apple of the world, they have all the data that they need compared to new startups. How come that this open weight might open up a little bit of or level the playing field a little bit? But overall, the data is the king, right? Isn't it?

 

Christian Catalini  33:34  

On the talent, first of all, you'd be surprised how many talented people will rather take a job in a startup with with high talent density and people that care about the same things that they do, so meaning is as important as the big check. So that that's on talent. On data, I do think incumbents have a massive advantage with AI, and in fact, many slow-moving incumbents, including some of the GSEs, I think are going to be a lot more dangerous, right? So if you think about the gap between a neobank and a GSIB, it used to be massive because one had an IT department, the other one had actually a tech software company. AI closes some of that gap if you want to do the right moves, but incumbents tend to have the right data for the past, and so I do still think that opinionated founders that do go out on a limb, but sometimes look that limb might be five to 10 years. Right now, I feel like at least on X, there's this like topic of the day. Everybody converges like toddler soccer, where everybody's talking about one thing, and then three days later it's the next thing, and the discussion is exactly the same.

 

Sy Taylor  34:32  

Welcome to the internet. It's toddler soccer all over, isn't it? It's like boom, whatever it is. Oh, South Korea's gone because the memory makers aren't doing anything. Like it will, yeah, completely. Humanity is one big game of toddler soccer on X. That is absolutely the case. I I'm so interested in this point about the quality of data versus the depth of it because incumbents have a lot of it, but most of them can't access it. And the ones that can might like the talent to do something useful with it. There are a few notable exceptions, but you know there are lots of data. Cursor just put out their latest model, I think version three of whatever it's called, and they have more tokens about how engineers create code than pretty much anybody on the planet. So a weaker model can outperform Fable Five on Codegen and Sol 5.6 on Code Gen because it's RL'd within an inch of its life for that one specific task, and there's this move towards smaller models, and so you kind of get that that interesting direction, and then you get the other direction of like the not so tacit knowledge that you said that Merkor and others are trying to build in with with the specialist evals. So it's going to be interesting to do that. And as as you think about financial services, who has like a real upside opportunity here? I've got a couple of favorites, but who's grasped that nettle of AI and done a good job in finance that you've seen, and you know is is finance just waiting to be cooked like the legal profession, or is there something else going on there?

 

Christian Catalini  36:10  

No, I mean when you look at legal, I would say it's going to be much harder for some of the top law firms to embrace for impact. Actually, one of my co-founder, the Lights Bark, used to work for one of the top U.S. firms, so we've been talking about this quite a bit. But it's difficult, right? And there's also restriction on data usage privilege. Like it's not simple for an actual law firm to to go ahead on, you know, with a legal AI lab trying to replicate the work. I do think they're at the top end of the distribution. Probably it's gonna look like a hybrid, but the bottom is just gonna go the way of the big models. I do think fintech has been different, right? So fintech probably because very tuned to data. For some of the fintechs, I mean, you could not be profitable if you didn't have good ML, right? Forget AI, right? And some of the more recent methods in LLMs, fraud has always been adversarial in crypto. I think same kind of feeling. AI is definitely more adversarial. Agents are going to be adversarial. So look, I think it's obvious that Stripe has been you know training really interesting models on the payment side. The card networks probably are trying their own, although of course there's larger, slower moving machines. They're not founder-led, so that that creates a gap. Then companies like Ramp, right? So if I were to think through which companies that become almost like foundation AI labs for fintech, they've done an amazing job at releasing tools and and playing into that middle arena of like, okay, we're not entropic, we're not OpenAI, but we know a thing or two that the CEO and the chief accounting officer of a company will care about. I do think we're gonna see a lot more than that. That is the vision of like own your intelligence. If you're a company with domain expertise in X, you will fine tune, you know, your own extension of a model. Thinking Machine recently released an interesting early result with Bridgewater, right, where they essentially show that their base model, which is essentially a distilled version of Kimi, plus Bridgewater's extension to that can really outperform the expensive models. So to your point about efficiency, that solves two problems at once, right? One is cost. I'm a lot less interested in cost than control. If I'm a VC firm, a law firm, you know, a financial services firm, the last thing I want is to send my traces back to Entropic and OpenAI. Now they're promising they're not training on your outputs and inputs. They don't need to. Everything around it, unless it's DDR, there's so much signal leaking. That's why you know, Claude Tag feels like a Trojan horse to me. Why would you like Claude right into your forehead? There were

 

Sy Taylor  38:44  

FDs that that come from labs. It's it's wow. This this horsey looks really nice. That's just outside our castle walls. We will

 

Christian Catalini  38:55  

spend resources, yeah, to help you.

 

Sy Taylor  38:57  

What a beautiful horsey. Let's let's let it in. Oh no, wait. There were Spartans hiding inside this horse. Who would have thought that's the FDE model from the labs? 100% But we're ready.

 

Christian Catalini  39:07  

We're ready to leave you the you know the dispose of the war. We're not going to invade your industry. Take the horse. That that's exactly no. Now that for those of you that have caught up with Odyssey, yeah, Christopher, you get

 

Sy Taylor  39:20  

it now, like 100% It's not just a weird little metaphor in a book anymore. But the the thing that I've observed as I speak to boards and CEOs of the incumbents quite a bit, I get asked about AI. Is they're smart to that? They're absolutely smart to that, and they'd rather have their employees in revolt about how bad their internal tooling is compared to the outside world, then risk giving away the secret source that they have because they are very very protective of that. So, really interested as we round out here in terms of like where the opportunity is for builders, like especially where stablecoins meet AI, where payments meet AI. Commerce. Where are you spending your time and energy?

 

Christian Catalini  40:03  

I mean, I'm spending my time and energy across too many different things. That's the phase I'm in. So probably not that useful for for the people listening. But what I would say is like, look, he's essentially the advice that I would give my my my daughter, right? Which is like, now is the time to realize two things. First of all, we all have superpowers through these tools, so no doomerism, no no regret. Like this is this is wonderful. Approach it with with a positive attitude. And second is like, what is the one thing that your innate talent, your obsession, your your history makes you truly unique at in a period where there's infinite change. The modes that you know one week ago we could have discussed about oh the labs are gonna eat everything and now we're here debating oh will open weights crash you know the entire IPO market for OpenAI and Entropic, it's it's pointless to you know narrow down those modes too too closely. I would say just double down on something that you enjoy doing that you get in the flow. Yeah, find that friction with the universe. Agencies just literally you going out and learning does stuff that the models haven't seen. So, if you're worried about the models, just think about the models that are sitting in a data center, right, or in many data centers. You can actually go out and and and do things. That that's probably the the main recipe for success. Sorry if it's too generic, but no,

 

Sy Taylor  41:20  

go collect new data from the universe like my two-year-old does. Oh, I wonder if I can stick this. Yeah, that's the best way to learn. 100% Look, we are closing in on time here, but I want to make sure that there wasn't anything that you wanted to say that on the agentic commerce topic that we haven't covered today. I'm one thing,

 

Bam Azizi  41:39  

Sam.

 

Sy Taylor  41:39  

Yeah, go go ahead, Bob.

 

Bam Azizi  41:41  

So you mentioned Ramp, but you skipped over Stripe and Open Router. What do you think about that?

 

Christian Catalini  41:48  

Oh, I I think it's a great idea for Stripe. Look, whether you build routing in house or you acquire it, Stripe is gonna have to do it. It's important to have it. I think it's gonna be table stakes. I think the open routing team has been honestly top in this domain. And if there's anyone that can turn routing into a real data play, which is really where routing becomes sustainable, it's probably them. So, yeah, look, Stripe has has a really good taste in acquisition. So I will use the word taste for this particular instance.

 

Sy Taylor  42:20  

Well, if you make your money growing the GDP of the internet and then the take rate on the GDP of the internet, then maybe you can have a take rate on another part of the GDP of the internet. It's it's fascinating strategy. It really is a very exciting time to be watching this market, isn't it, Christian? And I'm so glad we got to do this conversation. If people want to learn more about you and all of the things you're up to, where do they go to do that?

 

Christian Catalini  42:46  

I can be easily distill that C Catalini on X. That's probably where all of the information flows for free.

 

Sy Taylor  42:54  

Go grab that API and do your distillation attacks. Bam, how about you?

 

Bam Azizi  43:00  

Yeah, you can find me Bamazzi Mesh on Telegram on X, and also Mesh Pay on X and LinkedIn.

 

Sy Taylor  43:07  

You'll find me S Y Taylor on all of the socials, screaming into the void at fintechbrainfood.com, and of course at Tempo.xyz, where we just dropped a banger of a report on creating additional liquidity in markets. If you haven't already, go ahead and subscribe to this podcast because then you'll find a lot more of this show with conversations like this. And leave us a review or a comment as well. That's how you say thank you, and it means a lot if you would go ahead and do that. I know everybody says that at the end of a show, but it actually does work. I can see it in the data. When you go do that thing, it really helps us. So please help us out, and we'll catch you next time.