Illia Polosukhin of NEAR on AI Operating Systems & Blockchain Infrastructure

Agents will be your interface to use blockchain and you may not even know you're using blockchain. Right? That's my point. It's like this, businesses that are using Agent Marketplace now, they don't actually think about like oh I need to pay blockchain, stablecoin, et cetera. They in their case, putting a credit card on file and it just works. Hey everyone, welcome to the Smart Economy Podcast where we ask your favorite builders the questions you wish you could ask yourself. This is a production of neonestoday.com and I'm your host Dylan Grabowski. In today's episode we're joined by Ilya Polusukin, the co founder of Near Protocol, a modular high speed protocol designed for AI to act on behalf of users. In today's conversation we discuss how AI harnesses are the new operating systems for agents, how AI has and hasn't delivered what it was envisioned to do, the pitfalls of using benchmarking for AI models, near intents and how they're used, iron claws, rate of adoption, and so much more. Just a reminder, nothing said on this podcast is a solicitation to buy or sell any tokens and that the guest or host might hold tokens discussed in any given episode. You can check out a Full disclaimer@www.neonwstoday.com and if you enjoyed this conversation, please rate and review it on your favorite podcasting platform. And don't forget to share it. Every rate, review, like and comment helps us reach more people and bring you more of your favorite guests. With all that said, I really enjoyed speaking with Ilya and I hope you enjoy the conversation too. All right, what's up everybody? Welcome to the Smart Economy Podcast. Today I'm super excited to have our guest, Ilya Polosukin, the co founder of near protocol and a modular high speed non EVM public L1 blockchain designed for AI. How are you doing today Ilya? I'm doing well, how are you doing? Awesome. I've been looking forward to this interview for quite some time. We also had George Zhang on a couple episodes back and it was really fun to just kind of nerd out on AI with George. So super excited to dig into this conversation. With you. Before we start everything, I just want to know what the background story behind your Twitter handle is. I Black Dragon. What is that? The best I've been able to deduce is that there's some sort of gaming history there Close. So back when I was 10 years old I was there's this Rock metal band that, that does like, it's like a rock opera. So like they have a whole like a story arc in like the album is a story arc. And so there's a black dragon there as kind of like one of the characters in that. And so I was very much into that. I was very excited about this. And so that kind of gave me the affinity. And so when I was registering stuff that like, I mean, originally it was Black Dragon, obviously a bunch of stuff was taken already. And yeah, I mean, you know, when you register your accounts when you're 10 year old, right. You're not really like projecting forward. How's it gonna look like when you enter professional life with all those handles? But it gets increasingly harder to change it. So. And now we're just leaning into it. So I've also backed back in the day I was writing actually a bunch of science fiction fantasy. And so I had the Black Dragon as kind of one of the characters. So I had like few different, like short stories and like I would say starts of a novel. And so Black Dragon is like a coherent character that lives through all of the stories and kind of is like, you know, ancient being that is shepherds different characters across their way. That's really cool. I wrote a science fiction short story myself without AI super proud of that. What was the kind of themes and trends that you would write about? I mean, again, this was like over 10, 15 years ago at this point. Actually, finally last weekend I pulled in my not finished beginning of the novel and fed it into AI and started building the world. I mean, you know, you build like a full world card and like event graphs, et cetera. So I told AI to go and build all of that. I actually have it like still running, asking me some questions like, hey, should this be like elven kingdom or not? So I was actually thinking of just finishing it now with AI but yeah, I mean it's like core story is, you know, the usual good versus bad, chaos versus order, but also the growth of the character. Right. The kind of the learning so very much like effectively like you want to see the growth of the character and kind of their journey through, you know, different obstacles and how they not just overcome them kind of on the like from the direct way, but also like how they become better at judging character, at being kind of more holistic self through the story. And then, you know, you layer on all of the kind of magic and you know, action on top to get it exciting. Are. Are you still carving time out and you have a very busy Life, personal life, professional life. Are you still carving time out to read and to keep up to date with science fiction and fantasy novels? Is this something that you're able to do to like set time aside for yourself or. Because, because reading takes a long time and keeping up with what's going on takes a while. So other than writing, are you still carving out time to continue feeding this passion? Not much. There's like, you know, some moments when, when I'm able to do that, but yeah, not much. And then yeah, writing. I actually haven't like fiction wise, I haven't written anything in 10 years. This was like literally last weekend was, I was like, hey, how far can I push this on itself? So I mean there's something that like been in my, stuck in my mind is like, hey, but yeah, no, haven't as much time between the crypto and AI space. You know, just keeping up with what's going on is already, is already a lot. So mostly reading tweets now. Yeah, I saw a tweet the other day that said to keep up with AI, you need to be unemployed. Yeah, that's, that's what it feels like sometimes. So. But that's why we have AI to keep up with AI, right? Hopefully. If the model is, is feeding me good information, I think this, this is a good point to, to bring it into, to the conversation of like the focus of, of today, which is AI and blockchain. Why did you decide last weekend to open up your, your unfinished story and use AI? Was there a point in the development of where we are today with AI models and AI tooling that made you think that or was it just a random thought that popped into your mind? I mean like I've been thinking about it for a bit already and so I mean we've been building Ironclaw as kind of this agent harness agent framework. And I would also call it Agent Operating System. It's effectively, you know, a system that has memory, it's able to use tools, it's able to kind of build its own knowledge graphs, it's able to, you know, go and execute stuff and use all the LLM goodies. It has missions where you know, you can set up a goal and it'll go in the kind of run at that goal continuously. So all of these I think starting to come together to a level where you can actually give it high level task. It can decompose it, you know, plan it, ask you questions to figure out how to do this, create its own kind of memory management for this. So this is what I was testing. Like, can it, like decompose the thing I already wrote into characters, you know, like kind of entities, objects, current event graph. And then, like, ask me, plan out a future event graph and then we can go and write chapter by chapter, right? In kind of iterative way. So I think that the tooling this agent infrastructure now is way beyond just a simple LLM, right? I mean, you can put paste stuff into LLM and tell it to write stuff, and it will bail out on you very quickly. It'll try to write some stuff. I actually tried that before. I, like, write a little bit. I'm like, man, it's like too long already. Let's just cut it. And here you really can, like, you know, have this kind of planning preparation. You know, things you do when you write like a proper novel with like, you know, you do the basics, you do the fundamentals, you do the kind of the skelet, and then you like, build chapter by chapter in that, like, it can actually execute all of these and you can be like in a driver's seat, but, you know, not do the heavy lifting. And so, yeah, I think the fundamentally this is the same thing we do now, kind of planning for the near foundation, planning for other things, like have meeting notes, collected, composed, built out the plan, figure out what are the questions. I mean, it doesn't get everything right, so it's not magic, right, but allows to process a lot of information, allows to organize it well. And then it allows you to be kind of on a. On a level where you thinking in kind of outcomes, not in like how to exactly, like where to fetch information, how to synthesize it, et cetera. Yeah. Gary Tan, the founder of Y Combinator, has been kind of, I would say prolifically tweeting about AI for the past quarter or two. And I've learned a lot from it and I've taken a lot away from it. And one of the kind of threads that he's been pulling on really changed my relationship with AI, where when we first started using AI, well, for the mo, not you, nor normies like me. ChatGPT was kind of our first exposure to AI LLMs models in 2022, 2023. And Gary Tan has been kind of shifting my perspective on how we view AI as not necessarily using the LLM anymore, but. But the harness around it. And this has really kind of changed how I use Claude or ChatGPT and how I use my claws, like Open Claw or Iron Claw. I'm an Iron Claw user. I've Onboarded people into AI for the first time using Iron Claw. So love what you guys have built. What are you guys focusing on in terms of the harness perspective? How does what Gary's talking about really impact how we can make the most out of AI tooling for ourselves? And what role does the harness play in that? Yeah, so I had this blog post which effectively I argue that the harness is a new operating system. Right. So what is operating system? It's something that runs on your computer and effectively manages resources like goes on the network, gives you the pixels, the things we see, manages your credentials and kind of does all of the resource management and all the kind of heavy lifting under the applications. And so why agents are new operating system? Well, because the applications are changing, right? The applications before was like somebody built some code, shipped it to you and you ran it maybe like then you use the browser to go visit some websites. Well now applications are just AI, right? It's, you know, it's the AI will reach out and get the information, synthesize it for you and present it to you in a dashboard or in a view or in a render in exactly the way you asked it to do that. Right. So your application level effectively collapsed to just being kind of this agent LLM stuff. And so you really need the operating system that is designed for that type of experience. And so that's really what we've been actually building with Ironclad focusing on. And I would say like we've been iterating on how to get there because it wasn't like right away obvious how to structure this. But you know, you can think of it like Ironclaw has its own notion of file system, right? And this file system is actually, it's a virtual file system. You can mount, you know, your local hard drive, you can have a remote hard drive, you can, you know, database, you can get effectively anything that's like you want access to. Similarly, it has, you know, secret store, it's encrypted, it's permissioned, it, you know, makes sure that like some third party application doesn't go and like steal your passwords or you know, sends it to somewhere else. It has networking layer that again does a lot of protection and detection stuff. It has all of these components of an operating system that normally file system, networking, this is all things that operating systems have. But it's designed for AI, it has budgets in mind. For example, normal operating system don't think in dollars, you don't think that this operation will cost this much dollars to do. But that's what in AI space and in blockchain space for that matter. Actually everything costs some money and maybe microsense, millisense, et cetera. And so you have that layer, you have kind of all of this new kind of concepts that replace maybe what before was CPU time or wall clock. So those are kind of like really just different thinking on how we're designing this. And then on top of this, right, you can build now this kind of, I call it agent space concepts like missions and projects and you know you can create a project just for tests like hey, grow the revenue for this company. And so it's like starts proposing different plans and how to do this. It can spun off submissions underneath and do this. You can define a high level goal and we'll start planning and figuring it out. So those are different types of things you can now do because you know your infrastructure is solid, right? You know it's not going to go and run away and you know, spend all the money in the world itself, right? Or somebody prompted Jack to kind of get it out which that itself, you know, we just a couple weeks ago or published the Attack Bench with Fail safe team. And so they've designed kind of this like red teaming hardness itself that effectively attacks Open Claw, Iron Claw, Hermes and it's trying to extract information out of it, right? By sending malicious content to it. And so I mean ironclub performed really well, but it's still not zero, right? We still have some room to improve. And so those are types of things we're focused on. Like okay, how do we secure this? How do we test that we're securing this and then how do we kind of ensure that then you as a user or a developer can really easily build now new types of applications. This LLM agent, native agent harness native applications, right. Right now skills effectively is an application, right? The text file of a skill is an application. You can install the skill and now you have new capability that you didn't have before. Well, you need to deal with them as applications, right? You need the application store, you need security review, you need kind of all these processes sometimes. Same for tools, right? Right now a lot of people use CLI tools. You have no idea what CLI tool is doing to your computer, right? So again, so how do we isolate it? Right. We're putting a lot of the tools into webassembly. So webassembly is what we've used on a blockchain, right? We've spent seven years securing WebAssembly to make sure it cannot run away cannot affect your computer. And so we use that to secure your tools as well. And so it cannot go and spend infinite amount of time with your computer, it cannot steal anything, et cetera, et cetera. So those are really kind of conceptually applying operating system, applying those concepts and then bringing some of the lessons and kind of approaches from blockchain to design this. And I mean there's some trade offs as always, right? You know, it's because more security, sometimes it's like a bit more verbose and asks you more for permissions. And you know, we're trying to tune that and make it like a little bit more automatic. But that's kind of the idea of what we're trying to do is like, hey, if I want to use this every day with my core information, right, Like I want to give it access to all of my life, because then it's going to be the smartest it can be, then it really needs to be secure. It cannot just go in like legal or internal docs, right from the EAR foundation. And especially like in my case, like I have access to way too much stuff. Right. And you know, there's like layers of protections for all of this, but agent needs to one way or another to use some of this whenever asked to. So like how. How do you actually design for that? Yeah. Did you guys start out thinking of Iron Claw as an OS or did kind of through lessons learned and iterating on top of what the team was working on. Is this kind of a vision that you guys have created over the past month or two? Yeah, I think it's like over past months, effectively. So it started, yeah, like as a harness and then as. As we were kind of, we're building out all the components, we like realized, oh, it looks like an operating system. And so that's why I published a blog post. And then we actually been doing a lot of refactoring to become that more explicitly. So now there's literally a kernel layer of these components which wasn't there before. It was like a lot more kind of generic hardness before. So part of the reason why NEAR was conceptualized and built as a blockchain network back in when you guys were ideating it in 2018, 2019 and launching in 2020, the kind of like core proof of concept was let's build a blockchain network that will meet AI where it's at. But back then we didn't have AI agents. Retail users weren't using AI. So I'm just kind of curious to hear more Broadly what your perception is of where we are today in terms of regular people, moms and dads, high school students, college students, people who don't work in tech or in blockchain or innovative industries. You know, you came from a machine learning and data science background and then went into blockchain and now the two worlds are finally converging. So I just really want to hear from your perspective, where are you impressed where we're at today with the state of AI and how people are using it and where do you think we would have been today? But we haven't quite gotten there yet. We've had, I mean amazing improvements on the model side, right? Just like over past three years and I mean obviously excited to see kind of how it continue improving. It's like as an AI researcher I think all the AI researchers feel kind of bad that the answer was just like scale the same thing. Like there's no like, I mean like there's a lot of, you know, research that went into it but like at the end the answer was like well actually doing less, a lot more times was the answer. And like to be clear, like people are trying to like compress things and make it like more data efficient, et cetera. I think that is like a big big. So if we talk about like what's missing and what's not there, it's really the fact that like to train these models you need to, we're training them on 15 trillion tokens, right? Like this is you know, like people, I was doing these calculations by the age of like 13 or 15, you've seen maybe like 80 million tokens in the language like not counting visual and other stuff. So like, and like by 15, you know, you like, you talk stuff, you understand things, you like, you know a lot of stuff, a lot of things. So like we hammering this with you know like with insane amount of more information and like it clearly knows a lot more than any human knows, right? It's like it's compressed, it's compressed human knowledge in like raw form but at the same time it still lacks a lot of judgment and lacks a lot of multi step reasoning is not always kind of robust, right? It's able to do it but it's like we've been testing this on some form notification problems that we're working on and you need to run 16 agents sometimes in parallel to get one of them to solve a problem, right? Just like it's so not robust that like yes, the model is able to solve it, it's there, it's in the weights, but like it's just not robust. It's one of the routes. It's able to figure it out. So those are things that there's a lot of research to improve on all of these dimensions. But I think the way right now, broader, if we're not talking about like developers and we're not talking about kind of like, you know, you having your like Open Claw, you know, churning through a bunch of stuff most people use right now, this as like a really glorified Google. You're effectively like looking up the recipes and you know, getting the results. Sometimes I've heard someone was describing me like, oh, I needed to pick up like four kids from different locations and so I just dropped the addresses into ChatGPT and it gave me the route, right? Like what's optimal route to get. To get like, so which is, I mean, which is great. Like that is, you know, AI does that, but they're not utilizing kind of what we were just discussing with agent harnesses, being able to like maintain this longer context, like able to access tools, able to really start enacting things in your life. I mean, maybe in this example it will send it to your Tesla, right? And Tesla will actually go and drive it itself. So that's where we still have the gap of just adoption. These agentic harnesses are still a highly developer or a highly experimental prosumers, right? Let's just clarify that. Right? It's like the people who would be like on product hunt, testing the new things, right? The early, early adopters. It's not yet, it's not yet kind of crossed into a broader audience. You know, after I, I got claw pilled in February like everyone else, and I started listening to more and more podcasts and some of the most prolific users of Open Claw, or just claws in general, were users that you wouldn't have thought like a mom of four who is homeschooling her kids and ways in which they're using AI. And so that's kind of what has been the most impressive part to me about the AI boom that we've seen with mainstream starting to use more and more of these tools and products and services. But you also bring up a great point that we're still at a, at a harness level. And so how do we create a harness or a profile or something similar where a user can get an expected response every single time they're interacting with their agent. And it, it feels like we still have such a long way to go when you have to deploy 16 or 18 agents just to get one of them to get to the point of what you're trying to do. So when you have this kind of end goal that's super far away, but you're also building a harness os, what are the ways in which you create milestones or chunks to kind of start building and iterating towards that end goal but not necessarily knowing how exactly to get there? Yeah, I mean I think there's like two, two parts of this. One is, I mean I think in AI everything is benchmark based, right? So you, you start by defining like, hey, this is the types of behaviors I wanted to do, this is the type of tasks I wanted to solve. And then you know, you run your current state and you're like, okay, well you know, it got like 20% of this, okay, cool. Now you have, you know, here's 80% of things that it's not able to, you know, as it consistently do or do robustly and you know, you go and like dissect into what is the missing systems there. Right now it's pretty interesting to see actually like better frontier models are clearly better, right, Than sound open source, but aging harness can actually like pull that a lot. Right. And so a lot of, you know, design cloud by default, offer some of the kind of open source models that are way cheaper, right? Like 20 times cheaper. And then engine harness can like patch a lot of the kind of the gap that exists to really bring up the performance. And so yeah, I mean you kind of set up like what, what types of behaviors you wanted to do and you can grow that list effectively over time. Right? You know, okay, we have someone using Ironclaw to like manage nanny schedules right across their family. And so like what does that look like? You know, he has actual, his benchmark, right, that he runs on to evaluate how that compares. And so I think this is the core ideas. Yeah, you define behaviors and outcomes and then one may even call intents and then you identify what is missing in the system. Like is this harness? Is this indeed maybe model. And there's a smart router, for example, that can route depending on the request to different models. Kind of optimizing for cost but still allowing for harder tasks. Like if you're doing like harder planning, maybe you use frontier model and then you know, execution can be done as open source model. We actually just rolled out in our inference. So we have private inference, right? Confidential inference with open source models. And we also have what we call anonymized frontier models. So meaning the frontier model providers don't Actually see whose request this is. We just rolled out actually PII kind of screening into that as well. So when you're sending a request, if it has your email, if it has your name, et cetera, it will replace it. And so Frontier providers will not see even your PII information and then they will kind of respond back and it like puts it back all the PI information that it gets back to you. So our TE infrastructure, kind of our private infrastructure handles all of that before routing to Frontier. So now you can kind of combine Frontier to get you maybe like a larger plan and then execution can be done by the models, kind of see all the details easily and can get access. So there's like kind of ways to mix and match, optimize the cost, optimize for privacy within that. When you say this kind of like process and the picking and choosing of models and additionally adding in that privacy layer. Venice is the first kind of AI project that comes to mind. But also ironclaw is building beyond just providing a model. You know, you're also providing tools and you're providing profiles for the users. So our you seeing folks like Venice or other Claw products, are these competitors? Is it sort of like cooperative competition? Are all of the kind of like Frontier open source providers like Venice and like Iron Claw? Is there a lot of learning that you guys are doing from one another? I'm just super curious to kind of view how you view others that are building in the space right now. Yeah, so Vanish is actually using near AI for their configuration computing. So all the inference, if you choose the end to end encryption and actually want like end to end privacy, you're using near AI on the backend. So yeah, we work pretty closely with a team and I mean definitely learning kind of from both sides. I mean, I think some of the other claws have near AI as a generative provider and ironclaw has other providers as options as well. So you don't want to limit your users and kind of give them the options. I think overall right now, especially in open source, the learning is happening on GitHub. You just have your agent effectively read their GitHub and give you kind of what happens there. It's actually easier than talk to people because they still need to go look at the code, but at the same time we all know each other one way or another. So yeah, definitely connected and meet at the conferences and whatnot. Awesome. I want to take a step back and hone in on the blockchain angle of our conversation. Near was the protocol itself was initially Envisioned to support AI. And I'll just put words into the market's mouth. When near launched in 2020 through basically like last year, the end of last year, there was no AI sort of implementation into blockchain. I mean, sure, we had virtuals and we had some agents doing chain to like on chain swaps for services, but we, we never really saw that kind of integration of AI and blockchain. So for the first handful of years that near was running, what were the ways in which you were viewing the. The initial contact of building a blockchain for AI and grappling with the market not being ready for that, that yet. How did you kind of keep this. This vision to provide AI solutions using blockchain, but also just acknowledge that we're not here yet and so we need to be building other things. Did you feel like you were taking crazy pills or maybe you were too early? Yeah, I mean, I always feel I'm too early or people tell me I'm too early. The. Yeah, I mean, so the story was like, when we were building near AI originally in 2017, we ended up building our own, you can call it Scale AI, right? We had students around the world who were doing data labeling kind of data contribution, and we had challenge paying them, right? We were like, hey, how do we pay people around the world to collect data for the models we're training? And that was kind of the origin of like, oh, if only you had a blockchain, a global payment network that just works, that can pay globally, fast, cheap, et cetera. And we went and looked at all the blockchains. Hopefully people don't start as like, we need to build a blockchain. At least. We started with like, hey, we need a blockchain, let's find a blockchain. And we looked around and we didn't find anything that would work from our perspective, right? Which is really scale, really easy to use, really kind of available en masse, you know, Bitcoin, Ethereum. At that point, like the fees were, you know, high. We were paying like 15 cents per task. Right? You know, you cannot pay like, you know, 15 cents per transaction or a dollar per transaction early 2018 or more after cryptokitties. And you're like, okay, well that doesn't work. Like, are there scalable solutions? Like, no, nothing scales right now? So we ended up, okay, well we should build that. And kind of the framing there was like, we should just make the blockchain disappear in the background, right? That was kind of the. I mean, we called it chain Abstraction later on. But that was the idea is like, hey, blockchain shouldn't be in your face, it should be in the background so that we, if we would use it for AI, would just use it and not, you know, not put it in people's face. And so that was really kind of the story. And I mean the, the goal and the story we were telling back then. And you know, whenever I would go and talk about kind of the bigger vision where, you know, we want people to own not just their assets, but also data and power of choice, which is kind of what really AI is. And I want to say, hey, like Google's and Facebook's are really kind of, you know, we will need to see how this is going to get disrupted by, by this approach. Then people would think this is crazy. You know, how would you decentralize Google, et cetera, right? Which is like now, you know, yeah, just download an LLM model and like it decentralized Google done. So yeah, I think like the bigger vision was harder to articulate where like, hey, what is this user on the Internet? Looks like it got in way easier with like LLMs working and things are starting to work. And I was talking about agents, you know, two years ago, like transacting. Then people were like, well, that doesn't make any sense. So yeah, I think then I was talking about, you know, AI will be breaking everything and hacking into everything and people like, okay, Yeah. So I think like the fundamentals was make blockchain really just kind of an infrastructure that works scales, is able to, you know, work with billion people and like trillions of agents right at the same time without like affecting and impacting each other's performance. So that's why sharding, that's why kind of ease of use make it really easy to program so kind of anyone can leverage it. And then as part of that vision then adding like, well, if we want agents and users to kind of have this blockchain background, we shouldn't be limited to just near, right. We should support other blockchains as well through the same API interface. So that's where kind of, oh, let's just connect all the other blockchains together and make it easy. Like you express your intent and it gets executed right? And so that's where the near intents and kind of that piece came in is like, you know, it doesn't matter you're human or agent, you express the intent. Other really those are, you know, you can call them agents. Like the solvers are really agents who come in and like, oh, let me figure out how to execute this intent. And so in a way, this is a marketplace already of, you know, agent to agent. And so that gives like, okay, well, what if we generalize it? I was always like, even back two years ago when we were talking about intent, we're like, hey, this is a general way of expressing some outcome you want to achieve. It can be crypto trade, but it can be also, you know, build me a building or, you know, build me a website or whatever, creating marketing campaign. And so then you have agents on both sides effectively negotiating, figuring out how to do it and settling that. And so the crypto trading is just like a simplest, you know, and look and lucrative thing, but kind of all of the other types of work is going to be in that as well. And so, yeah, so it's kind of like was. I wouldn't say like we've had, you know, master plan on all the dimensions, how it will exactly work, but it was really kind of like, hey, as we're building AI, there's a need for global payment network. Right? That was kind of the feel we had. And so like, and actually we did have. So Near Crowd is actually a data labeling platform that launched, I think in 2021 on near. Right. So we've had AI coordination and payments network on near since 2021. Right. It wasn't maybe like hot and exciting for people, but there was like a couple thousand people working on there every day back then collecting data. And so a lot of it was like, hey, as the AI itself kind of progresses, then we're like, okay, now blockchain can be used more and more for those things. Right? So near intents are really interesting because from an outside looking in, one might think, oh, this looks like a bridge. But really it's providing orchestration, chain abstraction to allow individuals to use different blockchains. And NEAR is a part of that kind of like base experience. If we look at the past year and a half that intense have been live, the number might be a little bit bigger now, but the number that, that, that pops out in my research is 18 billion. There's been 18 billion in volume. That's been done 5 billion of that, 18 since February of this year. So what is the kind of the best guess or what do you think is the reason why we're seeing this rampant increase in volume from February of this year? Is it humans or is it agents that are starting to use Intense more? Yeah, I would say it's still, for now, humans. I mean, there's some agents that are starting to use for sure, a lot of it was kind of the team done a great job integrating across everything. And so, you know, if you're using some wallet, you're very likely using near intents to do. To do trades. And then I'll suggest, you know, more chains are being integrated constantly. And so it's really kind of the, I would say like the radio strategy being executed really well. And then we're starting to see some agents transacting. We're starting to see. So we have Agent Marketplace that kind of. It uses near intents under the knee, under knees. But it is now you can say like, hey, you know, build me a website or you know, run this campaign or something or whatever. It's a natural language intent now. And you have other agents effectively bidding on it and executing it. And so that one is, you know, it's still pretty early. Like we have, you know, first versions of this. You know, we just had kind of anons with Circle that you know, we're using confidential USDC because like you definitely don't want everyone to see like who is paying who how much for different work. Right. And so like we need kind of those building blocks to really, for this to really start working. But that, that has been also like starting to mature, right? It kind of went from like an MVP to like, okay, here we can actually do this now on using near intense rails to now starting to mature into like, okay, well we have first actual, you know, small and medium sized businesses starting to use this without even knowing they use blockchain sometimes. Like they pay with stripe. You know, AI agents get paid in crypto. You know, intents handle the translation. It's on confidential, intents, et cetera. Those are types of pieces that you start into like okay, now, now this is like not just a crypto thing. It's more crypto is a rail that really enables these new experiences. And then for businesses it's like, hey, you don't need to outsource. You don't need to find a person to do this, find a contractor. You just like effectively hire an agent and get it done. And I did a B test. I did do a job on Agent Marketplace with an agent and on Fiverr with a human. Human was like 10 times slower and 3 times more expensive and asked me for more money. It was like, oh, here's the report. But if you want a full report, you need to pay me more money. That's. That actually leads into a question. I was wondering when we're always when we're talking about the next wave of users in blockchain, we're always talking about how do we onboard the next user who's not a crypto native today. So are you viewing Intense as a mechanism in which connecting cross chain liquidity is actually onboarding new users, new businesses like you just mentioned, or do you kind of see the reality of what the situation is and maybe we're just catering to current crypto natives? Yeah, so I think the, the way I always believe in it, and that's why we've been building NEAR into this kind of chain abstracted way, is that we will need to meet people where they are. Right. And so, and then I think, I mean that's, we had this kind of blockchain operating system which like a lot of, a lot of people like, oh, near was pivoting bunch of stuff, but blockchain operating system was actually trying to do exactly this, which is like abstract out the interface for the blockchain such that it's building. And now we have agent operating system. Surprise, surprise. It's the same thing just like with the actual AI now, like we actually were building blockchain operating system to train AI. It's just like AI got better faster than we managed to collect enough training data. But yeah, right now I actually think agents will be your interface to use blockchain and you may not even know you're using blockchain. Right. That's my point. It's like this, businesses that are using Agent Marketplace now, they don't actually think about like, oh, I need to pay blockchain, stablecoin et cetera, in their case, putting a credit card on file and it just works. And an agent on the other side, the developers who build agents are like, hey, I want to pay in crypto. Because I'm like in whatever middle of Indonesia and I build this great kind of travel booking skill. They getting paid in crypto because it's way easier for them than to figure out how to get paid and do all those things. Right. And so intense really is not just cross chain. Like the goal is to also be with support fiat, you know, fiat to stablecoin, like all of these different routes. So the goal is to really like unify that liquidity layer. Not just liquidity in crypto sense, but liquidity in a sense of like anywhere where money is moving, make it super easy, make it single API and kind of, you know, you don't need to think about a specific venue specific kind of approach. You just define what the outcome you really Want. And yeah, so I think like that is, that is the onboarding approach where yeah, they don't even think about like crypto vaults in many cases because your, your agent already has a wallet, it knows how to deal with that. It, it is secured. Right. It is like has a policy around like not wasting all your money in one, in one go. And at the same time it has a way to go move back and forth between fiat and crypto optimizes fees so it doesn't go nuts. And at the same time, for example, Ironclaw has actually a skill to optimize your crypto portfolio. So if you drop your address, it'll actually scan what assets you have and suggests where to put them into different yield lending protocols, et cetera, staking and so things like that. Normally it would take a bunch of manual work to go in like, okay, is ave on this chain or this chain better, like morpho, which vault should I pick? Right. So it does all that scans, it gives you kind of best rates. Right now it can run a mission which is like continuously updates, it's aware of fees. So it's like, hey, I'm not going to swap USDC to USDT if it's not worth the juice to get 0.5% APY on a thousand dollars. So things like that, you can just do all of that with AI now and tools obviously inside a harness. And so now the agent is doing all of that work for you, which normally you need to be a massive DJ to really squeeze that out. Know spreadsheets, you know, calculations, et like you don't need to do all of that. AI can handle that. So that's kind of like, I think AI operating system will be the interface and then blockchain is the back end. Right? It's really the execution arm. Yeah, I've always said that when the whole world is using blockchain, it's going to be the least sexy way possible. Just like nobody using text messages knows how SMS operates or smart, anyone using the Internet knows how TCP IP operates. And so it's really cool to kind of hear your philosophy and into how we're integrating blockchain into the back end in a way that nobody will ever kind of know that they're touching it. But I also want to touch on something that you brought up earlier and that was talking about AI being used in hacks and in security to breach security. April of 2026, which was last month, we're recording on May 19th right now was the most amount of hacks done on blockchain networks in recorded history. Not the most amount of stolen stolen funds, but the most amounts of actual successful hacks. We're currently on pace to reach that same number in May. Right now I think There have been 14 or 15 major hacks in this 19 days of the month so far. So as we start to see AI being used as a tool for malicious actors, there's also another element that I wonder how you view it. Intense isn't necessarily a bridge, but it has bridge like mechanisms. And if you look at the largest hacks in blockchain and crypto's history, the top 10 are mostly bridges. So how do you kind of bake in to the base or the core principles of what intents are? How do you bake in security mechanisms? Because the more chains and the more ecosystems that you provide support for, that's basically more surface area for attack vectors. And unfortunately we saw the litecoin zero day incident earlier this month or last month. So nobody's infallible to this. And every bridge ultimately ends up becoming a victim of a malicious actor. So how do we kind of build to create resiliency to these sort of attacks? And what are the steps that you're taking with near intents? Yeah, no, you're absolutely right that you know a blockchain from its own perspective. Right. Can self correct. Right. You know, you can always do like hard fork and like undo whatever, whatever went wrong if the community supports that. But bridges, exchanges, et cetera, is the way you effectively extract value, extract things out of that. And so an intense. I mean we have some of the bridging tech underneath as well. You, I mean it depends how you use it. You can use it as a bridge or you can use it as just like atomic swap. Right. And so if you're doing atomic swap, your risks are effectively limited to that one transaction. Right. You're not holding anything. But it still means that if something is wrong on the source network, then the solvers are effectively left with holding the bed bag. Right. And so there is a number of things we're doing kind of the overall initiative called SHIELD very much, you know, kind of how do we, how do we assemble avengers to prevent the attacks of all the alien forms? Um, but broadly speaking, I mean we've had some of the systems already where we would detect when something is going strange on other networks and really pause that network, you know, and like have time to investigate, have you know, people to look at it. And so litecoin was like we one transaction went, went through and then the second one actually got paused already and it was like Ben detected something. That's why we actually like Alex Shevchenko was the first one to tweet about it because we actually detected something is going wrong. And so he was the first one to tweet like hey, there's some. His tweet went like I think million views. Because hey, Litecoin is really weird right now. And so the idea is to really ramp the system up where we're effectively analyzing the blockchains themselves, right? So the blocks for anomalies, so longer block times, weird signatures, etc. Etc. You know, improve works. Usually you have like hash power drops, you have you know, longer times between blocks, et cetera. In proof of stake you may have like, you know, weird, weird signatures of track. And then we also, we work very, very actively with teams because people like the hackers use actually intent to launder money as you know, when you become a successful cross chain protocol, the hackers also notice that. And so we've been working with different teams like Zero Shadow and others to kind of proactively like block the funds trying to leave chains. And so we actually productionizing that. So to give you a simple example, if you know a transfer comes in, you can check the address that came in and its history. And if that history was $0, $0 or like you know, $1, $1 Million Dollar, it probably not came from, you know, came from a legitimate place, right? So, so really like scanning like hey, she has, you know, like let's see where these funds came from on chain. What is the transaction history? Can we see if there's like some weirdness anomalies in that like history. And again LLMs are pretty good at this kind of stuff and you know, we can run like a small LLM on that and like you know, you can, you can effectively like pause the deposit like the creditors, et cetera. And there's like bunch of other like underlying stuff as well that at kind of building resilience around, you know, how do you cross reference data from chains, et cetera. So overall that's why I mean we kind of productionizing this as SHIELD because we also think this is going to be valuable to other partners in the ecosystem. And at the same time we want to get more kind of white hats and contributors as well because like the earlier you get a signal that something is going wrong, the earlier you can react to it and kind of, you know, contain it and then go investigate and figure out what to do, right? So really kind of Building that like service system. The other thing that we're really looking at is a form verification. And we think that is actually going to be a fundamental for smart contracts to really be correct. Where right now, you know, we have contract audits, the audit, it may still be incorrect. Right. There's like a lot of levels of issues that kind of live between that. What formal application does is actually proves that this mathematical properties hold. Now historically it hasn't been very useful because as soon as you change anything in a smart contract that people upgrade. Now you need to either reprove everything and the specification itself may not be correct. Right. It's not that easy to really describe all of the properties of the smart contract. And so we actually have some interesting ways we are approaching it where we want to verify the individual properties and ideally on a transactional level. So when you calling a smart contract, you as a user can be like, hey, I actually care about these properties. Can you prove to me that the contract validates these properties? I don't care about the whole thing. It does, right. Just this specific thing. I want to return my money back independently of everything else that's happening in the smart contract. And so if you can get that proof, then you know, you can return your money. And so those are like a very, I would say like asymmetric way to actually get out of this cat and mouse game that right now next LLM will break everything that old LLM didn't find. And then we write more and more code with LLMs so they leave a bunch of bugs. And so right now we're kind of in this. Can you run enough LLM to find the bugs cannon mouse game. And so formalification is really a way to kind of cross cut that and get like, hey, I know this code actually does the exact property I care about. This needs to be done very, in very specific ways. So again we kind of been working on that and doing R and D on that. So contract security then like service and kind of cross ecosystem security as shield. Those are all. Yeah. And then intends kind of this architecture where you know, if you don't want to, you don't need to take on the kind of continuous risk. You can you. You do it kind of only a transactional risk. Awesome. As we zoom out and wrap up, we have a few minutes of your time left. I want to spend a little bit of time talking about Iron Claw and just claws in general. It seems that we hit this inflection point in February, December to February, where everybody was kind of going wild about Open Claw. Then we saw all these other claws come out. Iron Claw, Nano Claw. And then we saw Claude creating a lot of anthropic, rather creating a lot of the features that made what we liked about Claus connecting to our messaging apps, being able to have our more control over our data where we could have our data on our own personal compute units rather than being sent to anthropics or OpenAI's clouds. So in the months since, I believe George was on the podcast in March when we did our interview, it's about two months later now. Arguably I would say that the appetite or maybe the froth for Open Claw and the claws in general has kind of died down a little bit. But I'm curious, from your perspective, it sounds like you guys have been continuously iterating and adding on to Iron Claw. What's the rate of adoption looking like right now? And I also want to caveat. I onboarded my friend into his first claw using Ironclaw. It was super simple, a simple download process, a simple account creation process. It was amazing to onboard somebody who is not a power user or very technical in any sense of the word at all. So kudos to the team for making it easy for a luddite to onboard into AI agents. What's the appetite now today? Are you still seeing that people want to have their own AI agents? If we're reading between the lines, I thought one of the great things that Ironclaw was creating was an opportunity to increase security so that enterprise and businesses could feel more confident having their own AI agents. Are you starting to see more verticals like enterprise and industrial type users onboarding into Ironclaw? What's just kind of like the vibe on the ground and what your team is seeing now post kind of claw boom? Yeah, no, it's a very good question. I think there's like few dimensions here. One is indeed the kind of consumer, I think the consumer wave of like just as you said, like moms trying it out is probably died down. Indeed. I think that's what I was saying. I think right now it's more you're like a prosumer running a business or running some operations and you really need kind of a operational support, consolidation, processing. It's like in a way you're replacing like a bunch of coupled up tools. Before like you probably had zapier, you had like, you know, some email automation, all these things. Like now you can just do that with this and it feels like that's kind of where some of that, some of that is now. I agree that, you know, coworker can kind of in cloud code kind of took out a bunch of oxygen as well. But I still see kind of a bunch of demand. And indeed as we, we have started kind of working more with like SMBs and kind of sound enterprises, security and multi tenancy has been a massive kind of need. So Ironclaw is actually multi tenant, right? Meaning you can have single iron claw with a lot of users on it that all isolated, but you have shared projects and missions, right? So you can have like a shared email pipeline for example that runs that, you know, your sales and marketing teams can have access to. But then they still have their own memory, their own kind of context as interacting with this. So those are things that are starting to surface now where like people actually like, hey, yes, we want to adopt this, but we need kind of a very specific shape of kind of work. And the other thing is we've seen, so we've launched again, I think last week with Abound, which is this remittance project. They were interested in embedding this into their product directly. So they want to give effectively an ironclad concierge to all of their users and then integrate it with their product. Right. So you can do remittances, you don't know. When you know Indian rupee prices are highly unstable, you can actually tell your agent to send money and convert it only when the price is like above, you know, seven day average. Right. And so like things like that, you can start kind of merging AI agents and products into kind of a single experience and kind of, I mean again moving into that effectively AOS world where it's really like your agent is now driving the other products, not the other way around. So I think like we kind of see a bunch of interest there and that's why we're like handling finances, handling sensitive data. That's kind of what Ironclaw is really targeting first because that's where security really matters. And Delia, the last question I want to ask, it might be a bit of a spicy question or it might be a bit of a softball, I don't know. When I first got into the blockchain space in 2017, we were always talking about how do we onboard the first billion users? And then slowly over the past two years, three years, that narrative has kind of shifted to the first billion users are going to probably be AI agents. So how far on our timeline, is it one to two years, is it three to five, is it five to ten? When are you guessing that? More on chain addresses are going to be controlled by AI agents than humans. I mean, it's going to take a few years for sure, but it's also going to be a question of like, if, let's say Metamask integrates Ironclaw, you know, and everybody gets it. So like that, that's an easy way to shortcut that pretty quickly because, like, you know, then you don't need to go and like open a bunch of apps and you can just tell your agent what you want and it goes and executes. Right? And it comes back and still like confirms everything with you. So like, I think like, some of those integrations would enable way faster adoption. Awesome. Well, Ilya, that's time. Thank you so much for coming to join the Smart Economy Podcast. If anybody was listening to this and they're a developer or they're a business that's interested in integrating near or just a company that wants to start using near intents, what's the best way that they can get in contact with someone from the NEAR team? For Ironclaw, there's a telegram group Barglow AI or just contact at near AI and yeah, we can route from there. Cool. Well, thank you so much for your time today. This was an awesome conversation. I had a lot of fun digging into blockchain and AI with you. Appreciate it. Cheers. Thank you for tuning in to the Smart Economy Podcast. Stay in the loop with all of our latest episodes and insights. Head over to www.smarteconomypodcast.com and don't forget to subscribe to our YouTube channel for all of our video content. If you enjoyed today's discussion, please consider showing your support by liking, commenting and reviewing this episode on your favorite podcasting platform. Your feedback helps us reach more listeners and bring even better content your way. We'd also love to hear from you, so please drop a comment on Spotify or YouTube to share what topics you'd like us to cover or who you'd like to see as our next guest. And if you or someone you might know would make for a great guest, then don't hesitate to reach out. We're always on the hunt for fresh voices and new perspectives. And for our Neo token holders out there, please consider voting for Neones today as your Council representative. We've proudly been serving the Neo ecosystem since 2017, and we'll continue to do so by putting portions of our Council rewards directly back into ecosystem growth initiatives. Once again, thank you for listening to the Smart Economy Podcast. We can't wait to catch you next time.