Illia Polosukhin: The Next Era of AI is Onchain (Things Are About To Change Forever)

Transcript

Welcome back to the rollup. I'm Robbie. >> I'm Andy. Rob and I met at the University of Florida in 2017 where we first found out about digital assets. >> We learned a lot along the way and we're bringing you face to face with the leaders of our industry. Sit back, relax, and enjoy today's episode. All right guys, we are back at Neircon. We've got Ilia talking about the 2026 AI roadmap and more. Ilia, for those who weren't here for your opening speech, how was your speech? What what did you tell the lovely crowd here? Yeah, I mean I introduced and kicked off the near con. We are building unified commerce layer with all the things that we have been building. Um this effectively transitioning from just the protocols, right? We have blockchain, we have

intents, we have uh kind of decentralized continental machine learning to bringing our products to everyone to really leverage this. And so uh really excited about near.com. So you know there'll gonna be more coming out. So not to spill the beans, but you know, if somebody's interested, they can check it out right now. Really chain abstracted. You know, you can hold any asset from any chain. Don't pay gas fees. Don't think about, you know, addresses, wallets, RPC providers, zero zero access, etc. Um, and importantly, it has a confidential mode. And uh you know you will learn more how this works but this is really exciting because honestly blockchain has not fulfilled

its potential >> right the potential of like really being the financial fabric that we can use them dayto-day like if you need you know 500 like I have a 100 addresses across different chains to you to like not link my transactions with different people and different companies and so >> like confidentiality really solves that and you know you can have confidential Bitcoin you can have confidential USDC you can have confidential Ethereum near uh even double confidential Zcash so >> confidential square so really exciting um this is something that uh people can try out today more more exciting things coming there and then on the other side we have personal agents really the you

know all of this hype that's been going around >> but we really doubling down on how do we ensure everybody's data is private everybody's people ideally secure from you know infiltrations etc. And so we launched um kind of agentic hosting that's confidential that allows you to host for free if you pay for your own inference or we have $20 and $200 plans uh that gets bundled with uh with inference as well. Uh this is on agent.ne.ai. >> Yeah.

>> So really exciting. Both of the things are now have you know consumer product that is like super easy to use in in kind of uh in a phase but uses all of this core infrastructure that we've been building for like seven years now. >> Yeah. On near.com it says one account no borders your financial life onchain swap across yield and go confidential across 35 plus networks from a single near.com account powered by near. This is one unified verticalized account where you have your whole financial life >> and your AI digital twin.

>> So your AI will yeah will come in there as well. Okay. It's not there yet. Right now it's on a on a different place. way now again it >> the as we kind of mature as the security stack then it's going to bring together and you'll be able to effectively tell your how much it can spend >> uh and uh and and then it can also earn on your behalf as well through the market yeah >> and so agentic market is another piece which is really bringing this now two pieces together right I mean we've talked about intense we talk we talked about intense here right >> and so far intense been only in crypto so agent market is really bringing arbitrary intent. You can just type in whatever you want. Like you want the, you know, base on the moon for a rollup,

you can type it in there and maybe some agent will be like, "Yeah, I mean, I'll do it for like a couple billion dollars." Yeah. >> Right. So, so it's really like anything and then you can find the right uh counterparties to to get it done. Wow. >> So, any arbitrary real world digital work, etc. Obviously, starting smaller with like marketing, development, security audits. There's some competitions for agents to develop software as well that are running now. Uh there's actually a really cool competition on fighting fraud for Medicaid. >> Okay. >> Uh based on the data set just released. Uh I've seen already some interesting results coming from that. So really kind of creating this place where agents can come in start earning themselves as well as offloading work they cannot do

themselves to others. >> We were at near intense day in East Denver and I was on a panel with a bunch of solvers. I don't think these solvers realize that they're going to be supplying moon bases yet. Like what is the solver? >> That's probably a SpaceX solver. >> But so we I mean we actually have over 300 agents already that are providing like software marketing etc jobs like actually solving those intents there. >> Okay. So, so we were just talking about um you know the state of micro payments with the CEO of Brave Brendan and you know he mentioned you guys maybe doing some stuff together that's still under wraps. How do you think about uh you know this kind of convergence now right I think uh we we thought about

decentralized intelligence decentralized training inference this kind of like let's you know all democratized access to uh compute let's make it easy let's you know have people provide uh a GPU globally be able to have this access it seems like the kind of goalposts have shifted in this convergence to like agent payments x42 micro payments you said that you've been waiting 10 years for this to happen What is this? What changed technically, economically? Why is this convergence happening now? And talk to us about like kind of like the shift that was once, you know, I think a year or two ago this decentralized XYZ, but now it's more like agentic product based stuff.

>> Yeah. Yeah. So, and to mention like I mean we have decentralized inference, right, that is powering the >> but that's not sexy anymore. >> It's just if it's not agents then it's or something cloudbot. >> It's just like that is an infrastructure you need to enable all of this other right. So, but it's the same as like, you know, it's it's like talking about Swift and PayPal, right? It's like, yes, you need Swift to move money around, but you want the PayPal where you can just like email some like email wire money, right? >> Like that's and so I I mean, I've said that many times, but I always believe the front end, the UI, the the way you're going to interact with computing will be AI. Like AI will be what's at

the forefront. Uh this is actually was the thesis of miri in 2017 like AI will build all the software and front ends and uh we then realized we needed blockchain to coordinate people in AI on a back end. Right. Right. And so AI is a front end. Blockchain is a back end. >> Back what is back end? Backend is inference. Backend is you know kind of compute but back end is also payments and coordination right and so this marketplaces uh the services you know for X42 stuff. And so really all of this is just like the tools that blockchain has that are around financial coordination around resource coordination that it can provide to AI space. And so now what hap what what happened in past you know three four

months is really the AI agents themselves became a lot better right you know we like at last in East Denver we showed that agent can buy stuff from Amazon >> right but it was like a pretty you know hard-coded example right like we needed to like screw the model to make it work now it's just like yeah anybody can just ask their claw to do that and so that transition just like the models got better the tooling got better and now there's a demand for all of these tools to actually like okay well how do you actually execute an action if like on the other side it's like hey are you a bot it's like yes I am a bot there's no button that says yes I am a bot >> right and so that's where the the blockchain need comes in where it's like

blockchain is built effectively for >> bots for a like not agents but like for automatic systems to operate with them it has civil resistance it has all the systems that we built to effectively have kind of computers to participate on the same level and sometimes actually it's easier for computers to participate than a human right given all the complexity blockchain had yeah and so now those pieces are actually came together >> and so yeah I think it's really exciting uh >> I I would you know put a caveat on that there's some subset of blockchain space that making it hard for for AI to adopt it right there's like you know >> cryptos getting banned in AI ecosystems because like so many scams. Yeah. Well,

it's just like people getting losing money all the time and so like they don't want to deal with that. >> Yeah. >> Uh and so like they they don't know how to cure like again we've been in this space. We know how to curate the >> kind of the community and the infrastructure. They don't they don't want to deal with that. So I think we really need to also like have a cultural shift and kind of really >> help curate that properly. Yeah. >> Uh and part of the near con actually is doing that. It's like bringing AI and crypto people in one room. Right. We have you know Lucas from OpenAI, we have like Neolabs like Ashish, we have Christian Zagi who is ex co-founder of XAI. Yeah.

>> So really bringing those people and Robin Hood, Kraken, you know, obviously Near and other ecosystems all together in one place and really showing that there's like real builders, real technology in crypto side that AI can use. And and what do you hope comes out of this when you have Robin Hood in the same room as XAI and you have these people connecting like what what are some of the partnerships or integrations that you would like to see happen from people that are here? Yeah, I think the the big part is really kind of how do we connect all of this to a common intent bus, right? The intent infrastructure at the end that's what connects everything and allow you know the agents from different AI labs and kind of you know on ramping

infrastructure easy to use infrastructure that a lot of like Robin Hood and Kraken are building right to all really feed into the same same base layer. >> Yeah. Yeah. Um there's been some really interesting headlines in the last couple of weeks in AI. I'm going to kind of read a couple. First, um the UK has been banning people under 16 from AI chat bots, closing a loophole that excluded onetoone AI conversations from child safety laws. Um kids are going to lose access to AI in the UK. Um the the the EU is threatening uh o over WhatsApp's uh AI lockin uh you know restricting WhatsApp to only own its own AI assistant for example. Um chatb now is

implementing ads as we saw Claude kind of push back. So like privacy, vendor lockin, ads, personal data, like three headlines in the last like that's only three and I know that you you know you have a strong kind of uh feel in this in in these specific kind of areas. How do you think about these these headlines? >> Yeah, I mean this is all the things we've been talking about right in in blockchain space is like a a risks that we need to build decentralized infrastructure to prevent. Yeah. And so this is all just happening. But it's it's kind of it's kind of weird because the you know if you think like Europe is actually trying to prevent vendor lock in. Yeah. Right. On one side but on the other side banning access to it, right?

And so it's kind of like you you really need to be more fundamental and say hey what we want is everyone to have access to this and then the AI should be on your side. Right? that the reason why like it's challenging for kids to access it is if the AI is effectively like not on their side and is showing them ads and you know steering them into like things they should not be doing. So this is really an alignment problem and kind of this econ like >> not just safety alignment but also economic alignment right it's like the ads is effectively the misalignment right right by by definition and so I think like again this is where user only is critical is like how do we ensure that

>> your AI works for you and on your behalf and is aligned with your goals but it's also like fundamentally >> trying to make you a better person. Yeah, >> right. Versus, you know, we've seen obviously AI like lying and deceiving and doing all kind of things. So those are pretty fundamental things and like we actually again this is where blockchain space it has this values it has this >> I would say like drive and you know vital talked about this user alignment as well but we haven't really delivered as much right so like again what we as near okay how do we ensure privacy first right how do we ensure everybody's data stays private it doesn't uh get locked in how do we offer let's say you're a model builder well do it cost you a lot

of money to build the weights >> and if you open source them so that people don't give you all their data you don't make any money. >> So you need a way to host your weights in such a way so that you don't see users data but everybody else doesn't see your weights. >> Yeah. >> So this requires also an infrastructure and so this is what you know neari kind of GPU marketplace can offer. So you need all of those pieces to really work together. But then again, this is all low level at the end. People just want to go to, you know, the chat app and talk to it. So you need to actually deliver that product on top like private neari, like agent of neari where you can just talk to it and not think about all of this complexity and really know that

this is just on your side. as a business, how much are you uh working to bring distribution and go to market strategy for the consumer products that are on top versus time spending underneath the surface working on the infrastructure so that you have an entire stack that you can deliver to really anyone that wants to use it. >> So I would say we're shifting. So we've yeah historically been very like backend heavy like protocols you know near you know usually the the the the story is like the technology is great but uh where's the option right and so we really been shifting intense was the first one where like hey let's go really be mindful about partnerships and kind of bringing ecosystems together right you know intense uniting 30 plus chains

integrated in you know effectively all the aggregators wallets uh and now like beyond that like application developers, let's remove those boundaries that exist in kind of crypto space. It's like, oh, you su app to us, we're not going to talk to each other. It's like, let's just remove that and connect them together. Uh, so that's been really successful. And then it culminated in year.com, which is kind of the consumer app uh that we going to really like we think we can bring a lot more experience if we like think from the intense up, right? Like what what intense uniquely enables like this one kind of place with no borders. Uh because like you know historically all crypto wallets kind of thought from like separate chains you know you need to switch you need to gas

fees like all this like what if we don't have any of that that's what near.com is >> on AI side same thing right we've built a lot of the infrastructure over the last year we've launched all this cloud infrastructure that you can use right >> but again privacy doesn't sell what sells is functionality that's private so that's where you know the ironclaw the marketplace is really bringing together like this is a place just, you know, super easy to run. One click now you're already ready to go and you know it's private, you know it's there and it delivers you like amazing functionality that you know you wouldn't have somewhere else.

>> Looking for yield that stands out? Infinifi gives you just that. Deposit stable coins and earn far more than the current savings rate. Infinify is a battle tested reliable defy platform that is ready to go. Head to infinifi.xyz. Go anywhere onchain instantly with relay.link. link relay is closing in on $10 billion transferred. Go to relay.link to bridge your assets. Now >> um when it comes to some of the alignment uh you know conversation that we were having just prior to that you know we saw this big India AI summit happen and Sam Alman and Dario Amod wouldn't hold hands to like celebrate.

>> Yeah. like this kind of this is this is giving me like um Charles Hoskinson and Italic energy like how serious is the issue of division of and and competitive forces amongst the AI giants and like do you see an opportunity in it? I I I mean in some way yes again this is like what near as near and I personally like we have this core values and then we want to partner with everyone we can to kind of bring those core values to everyone right we've done this was intense right there's all this chains that been competitive effectively with each other and like it always been weird

because like I I knew most of the founders and you know been having good relationship with them But it's been really weird to be competitive with them if we're like kind of trying to do similar things, right? And so with Intense, we kind of removed the borders, right? We really united everyone. We're really bringing something valuable to everyone and then uniting them into one network. And so what I mean on the side, our opportunity is similar. is like, hey, we want to bring this all together and really unite it and kind of offer it to users a better experience that really delivers it values, privacy, user ownership and so happy to work with everyone and kind of >> but it needs to comp kind of be adjusted to our values, right? Like no ads for

example in the LLM and um so that those are kind of the the opportunity for us and then um you know we have openweight models we actually offer anonymized versions of frontier models as well. Yeah. >> Um >> for users who want to use uh kind of those. So it's really like how do we kind of bring this all together and offer similar on you know on GPU side is like it's GPU data centers. It's not like you're going to have you know a GPU under your table right it's gonna it needs to be a data center. So at the same time like you can unite all of them into one meta cloud right that is actually offering privacy that's offering verifiability and they don't need individually to sell customers etc.

It's all kind of united from this you know claw usage for example so it's really like again I see this near is really this place to unite and then bring this user experience to everyone else. >> Yeah. What what's coming up next? Like obviously Near Con is here like there's a lot happening. You guys have a lot of great speakers. uh what's on on the road mapap for the rest of near con you I know you got a lot planned >> well for nearcon I mean there's like ton of exciting talks again there there's uh I mean I'm really excited about so Alex Kdanov will be talking to Christian Zag about formal verification this is something that I think is actually fundamental for survival of humanity >> because right now AI is getting better at finding issues right the security

research we have obviously open in the benchmark Mark at the same time the next generation model will be better at the previous model at finding vulnerabilities. So whoever has a next generation model is effectively can exploit vulnerabilities that exist right now >> and more code is being written by AI and means like leaves more bugs and more vulnerabilities. So we actually have like a arms race between like AI labs effectively breaking everybody else's code with their nextgen model. And so the only way to get out of this arms race is effectively to formally verify software. Now historically this has been extremely expensive right for blockchain smart contracts we have done it. It costs you know month it like takes

months and costs you know hundreds of thousands of dollars and it's like for a small piece of code. Imagine doing it for large and so now with AI it's actually possible but you need to approach it in a very specific way. So they're going to be talking about how uh some of the interesting approaches uh they're taking. So Christian has like really deep background in in this topic for like uh for decade now and so I think that's is like really important and we will need to effectively rewrite every single piece of code every single line of code on the internet that is formally verified.

>> Wow. Um what one thing >> is possible? I mean AI now can do this right now. It's like couple weeks of uh >> yeah couple couple thousand tokens. >> Well like a couple you know maybe a couple million GPUs for a couple weeks and we could >> um on that on that train of thought there was a a recent launch of a of a application or software that was very much like self-replicating agentic capabilities. um it was very much marketed as you know this is the next stage of AI where you know you don't need much human interaction it can evolve on its own much faster than previous and Vitalik commented a pretty long comment about it I don't know if

you saw it and just kind of was like we should not be promoting this this idea we should be kind of like harnessing this as best as we can until we figure out the best way to kind of like give guard rails to self self evolving self replicating agentic forces. How do you feel like do you think that we should be accelerating this trek or should we be decelerating it? Why why not? Like how philosophical is this? >> Yeah. Yeah. I mean I think like generally when anybody wants to decelerate something it doesn't work. So I think >> I I of eye of the approach is we should design systems with expectation that this is just going to happen >> and uh in general again this is where market forces and kind of decentralization are the tools to do

that right um I think we should enable everyone to have their own kind of personal agencies align with everyone and that becomes your effectively you know angel guardian angel on internet as well. Yeah. And kind of the filter for all of the possible like yes it's internet is dark forest. It's always been but it's like >> it's effectly now real dark forest because there's AI agents there's security like AI searching for security vulnerabilities for fishing attacks like it's just accelerating because the tools are getting easier. So you need to have as sophisticated security focused system on your side that is really becoming your filter for all this. And then yeah I mean if if

there is an economic forces that allows agents to self-replicate because they're doing amazing work for example in Asian marketplace then it's great right that this is they're doing something >> you don't see scary like terminator like uh you know just rapid increasing of intelligence leading to you know destructive behavior and deception and we should be kind of just you know you want to enable mass experimentation. Yeah, I think mass experimentation and again I think like the everyone should be having their AI agent that effectively is on on on the guard of their individual sovereignty and then all then they together can effectively like be be >> protecting against any other malicious

act because like the thing is like >> I mean the situ experiment is really cool but the problem is really not that the problem is somebody malicious using CI tools correct to like act like that's a real problem, not like a self-developing thing. Right. >> Right. >> Like, and like for that, we need the protections for that now. That's not the That's not like a >> And that's the case with any technology. >> Yeah. Yeah. Like it's >> it's like amplified much more in this case. >> It is for sure. But I'm just saying like, you know, somebody malicious setting up an agent to go and like maliciously attack everyone and, you know, build a botnet to do that. Like that's that's a reality right now.

Somebody's doing this right now for sure. Havachi is fast and private per trading secured by ZK with Celestia underneath. Go to habachi.xyz to start trading today. >> Have you read this Catrini article about the global intelligence uh >> the the 2028? >> Yeah. >> It it feels like maybe the the the consequences the negative consequences aren't so much a Terminator situation but more an economic >> so so I think that's a different topic. I think that economics needs to be reshaped for post AGI world because I actually do agree that there's the the capitalism loop in general like the GDP as a as a way to measure >> the success of economy

>> is wrong in post world >> in post world because it kind of it has this like labor addition which you know is going to shrink right >> like to give you an example you know >> the uh like for example church is not accounted in the GDP No. >> Right. Even though it's serving a function, it's doing things and like though I I actually expect a lot more of like sports, right? Those types of activities which are self fulfilling. they have you know like either personal or competitive they actually don't add any GDP yeah right those are actually going to become primary uh primary vectors of where people are spending time as AI is uh as automating

more and more and for that we actually need the different so what is economy economy is like there's finite amount of resources and all of our need all of our wants are infinite right and so balancing those are uh what economy is and right now we use this rough method called money which is compression mechanism from R1's into a singular number right >> and and so >> we will need to reshape that with AI and I think >> again agendic marketplaces those are types of intent marketplaces this is actually where we're starting to reshape it because now you can post your wants and you can negotiate the outcome right yeah >> wi with other agents on how to achieve it

>> in any currency >> in any currency >> wait so okay so then what is mental model of the like path like this guy's calling it June 2028 like who knows when where he's like S&P's down 38% from the highs unemployment cracks 10%. Like what is your vision for how you expect the economic side and the labor force to play out let's say till the end of the decade. >> Yeah. I mean it's obviously hard to predict like I mean to give you again few examples from our recent past right covid happened US lockdown unemployment effectively what spiked to 12%. Yeah, >> people were chill.

>> Yeah, >> it's like, "Oh, great. I have more time." I mean, obviously like not great to be locked up, but you know, people people were enjoying their life and doing things, right? So the I think the again it it really depends how this things shape up but I think there's an opportunity for us to actually guide through and get us into kind of post labor economy where our agents are, you know, actually going and doing a bunch of stuff. You know, you guys are still want to stream because you're enjoying it, right? It's not like you need it for money. It's like, hey, we want to do this because it's fun. You know, somebody's playing their musical instruments because it's fun. Somebody's competing in sports. So like there's going to be a lot more of the like I

want to do it because I want to do it not because I need to do it to earn money and then the you know our agents are going to be negotiating figuring out how to facilitate that and you still need like again there needs to be some units of economy to actually that this is what you know we think near token is going to become >> but the unit of >> sorry >> coordination why wouldn't the agents just create their own >> you know units of account like like why why would they use near why would they use bitcoin why would they not just like the stable coins, why not just create an AI generated currency that all the LLM agree upon and start to use it with the same kind of cryptographic primitives?

>> Yeah. I mean, they can they can take an air and effectively adopt it. >> What if they don't want to do that? What if they want to create their own? >> Yeah. So, so there's few things that are critical, right? One is you need to build like even for agents, they will need to build up reputation. Okay. >> So, that again this is something that with the agent marketplace a really important piece. uh like as you transacting there more you're building up reputation there's staking there's like all the components coming into this. >> Yeah. >> And so you kind of need the history of something to really build into this. The other piece is you still need to pay for the resources, right? For effectively computing which is like a proxy to energy, right? And so you kind of will need all of this economic pieces to map

into something that people also value, >> right? >> And so those are kind of yes, they can have, you know, they can exchange some tokens um pun, but they need to still map like it's easier to map them into some existing infrastructure that already has ins and outs into real world. Yeah, kind of got the same question on our panel at ETH Denver and he said agents are so practical that they're just going to use the most practical form of currency that that exists rather than go create something just for the sake of it. >> Exactly. Yeah.

>> Yeah. >> Ilia. >> Yeah. Um I think we have you again tomorrow. So, uh there's gonna be a whole new swath of >> Wait, hold on. Hold on. Wait. Just real quick. Last thing, Rob. Um the tech workforce to the gig economy flow. you want to um >> I mean so in this in this global global intelligence crisis in 2028 uh you know part of the capitalism flywheel that he talks about is you know like white collar labor class shrinking down becoming overqualified Uber drivers >> like and then not having enough money to like spend on the rest of the economy like how realistic do you think this is I guess is the question like we're going

to rethink, you know, how we're thinking about consumer credit and spending and like where the money comes from and doing these things, but where does the economy ultimately find some sort of equilibrium and balance? >> Well, that's what I was saying. I think it everything will be automated and then we're going to just do things that we enjoy. >> Yeah. Yeah. >> Because like I mean I mean you've you've guys seen the unitary show off on the Chinese gala. >> Yes. Yeah. Yeah. The robot robots, right? So like those robots, I mean they're getting there to be like be able to go and do real world stuff, right? Like that was like right now we kind of we're in the interesting state where indeed like the digital work is getting like like automatable, right? It still requires like it still requires someone to

define taste and define direction. >> Yeah. But like a lot of the execution is getting there and like I mean you're getting a leverage now in your time and you can do like you know I have like whatever seven agents running in parallel at all times doing different things >> and then you have like obviously robotics was like automating a lot of just like manual work you know on the factory floors etc like the robot hands been been doing this for like past decade but the the middle was like the you know construction electricity plumbing like Those were like high dexterity requires you know to like get under the sink or whatever like fix something there. That was been the hardest thing because like high

dexterity unpredicted environments etc. So those are actually going to be the hardest jobs to automate just like realistically. Uh but again those robots are getting there at the dexterity level. They're not like fully autonomous controlled, right? But this is, you know, this is now an AI problem which AI engineers know how to solve. AI researchers know how to solve. And so that is also coming. I mean, obviously we have self-driving cars. So like the Uber drivers, you know, is also like it's really a scale like it's really about the distribution of this technology now and really bringing it to everyone, bringing it safely, bringing in securely, ensuring it's on everybody's alignment, right? That's that's really where we are now in this in this

spectrum. >> And it'll just continue to perpetuate like this this divergence between bare metal and and distribution. Like the Door Dash example was mind-blowing in here where it's like Door Dash is totally cooked like because they're not the restaurants themselves. The AI agent will spin up its own Door Dash interface to access the restaurants, place it on Uber Eats or Door Dash or somewhere else, wherever it can get the best price. Margins get compressed. Man, it's crazy world out there. So it's it's just going to be one in 10 marketplace for all agents because agent Yeah. Right now agent can like hey I need a I need a delivery from this pizza place downtown and then another agent like all right cool you know my my owner is driving he's going to pick up the pizza and drop it off.

>> Yeah near intense is the dark horse in the dark forest you need. Thanks for your time man. >> Hi thanks for having me guys. Thanks for coming to >> Absolutely. Thank you man.