Ed Felton & Illia Polosukhin I Where AI and Blockchain Intersect I Agentic Ethereum 2025 Summit

[Music] and for this panel I'd like to bring on Ed from offchain labs and Ilia from near I'm going to talk about how a and blockchain intersect so without further Ado let's welcome them both on stage hello hello thank you cik hey um so we've done a lot of uh convers ations together and uh one kind of cool thing here is that we always end up learning super interesting things that are just on top of your minds and uh this one is a very interesting one because both of you have had experience and uh and direct work in in in these kind of fields and now we get to intersect Ai and blockchain so my kind of first question to you uh both is kind of looking at your past background from kind of Google research in Princeton and then kind of transition into focusing a lot of your time on blockchain scalability I'm kind of curious to understand how has research influenced how you see um practical applications being built and bringing those research ideas into uh into the real world and uh where do you kind of see uh the world of as shaping in for uh for your personal interest and maybe it'll be good to start with that sure yeah the way I think about research is um and impact as a researcher um was always to try to look for what are kind of the fundamental Al hard problems what are the barriers or roadblocks that are stopping the progress that everyone sees is possible and then figure out what are those barriers and then how to knock them down so you know you're really focused on fundamentals um and so you know we I got into this field through uh looking at fundamental issues around um I got into blockchain by looking at fundamental issues around how to scale um but um one of the things I think a research background is useful for is that it helps you spot the big picture and to sort of find The Leverage points what are the technical leverage points you need to attack um and then look at the big picture I mean for me the the the big um the exciting thing about both blockchain and AI is that they have the potential to change the relationship that people have with technology right with blockchain it's about by putting things on chain they become more transparent more accountable um you they can be governed more by Community instead of you know behind behind the walls of some data center um and AI is the same kind of thing it's a way of empowering people especially people who are less have less facility with programming to uh to actually be empowered and to be able to do more with technology than they thought they could before um so like looking but for me it's really exciting to be able to look and say what are the hard interesting technical problems at the core of making this Vision happen and then just to go after those yeah so uh for me actually we got into blockchain because we were doing AI startup in 2018 and uh we were doing a lot of data labeling um at the time to effectively train our model um and the challenge was paying people like coordinating people effectively with incentives uh you know we had people around the world uh some of them didn't have Banks a lot of countries they were in had some restrictions on transfering money and so we started looking at blockchain just to solve our own problem and kind of obviously second ads position that like you know when you start looking at some problems you start looking at like okay wh why is this why is this the state like what is the fundamental blockers right that's kind of how we got to build near but I agree that kind of what inss to me is like how do we use this tools to coordinate because one of the challenges I think in general kind of research um has a lot of broken components right the grant systems the corporate research there's all of the incentives that are quite misaligned and they don't have the same kind of outcome like how do we Advance the world because of all of this kind of misalign incentives and I think kind of blockchain off like cryp crypto industry offers obviously it has its own vline incentives but offers a way to structure incentives in a constructive way where you can actually advanc technology bring something to the world and uh reward everybody who participated in this and I think again that that's kind of what attracted us back in the day to do data labeling and kind of those things and I think this is where very attractive right now for AI where we see big labs are closing up right not open sourcing not allowing other peop to participate effective open source now is used as like a weapon kind of than than than actually to advance the science and so yeah I think I think of blockchain always is like Hey how do we coordinate people in more effective ways how do we incentivize people to really create common goods even if People's individual incentives are uh you know self-interested right I mean that's kind of that's kind of the story of Bitcoin right everybody's self-interested but the result is you know public good that everybody can use um both of you kind of ended up bringing in something um pretty interesting and and we have a few more questions want to cover on that topic but it might just be easy to surface this now which is um the difference in definition of what the the AI Community things open source and and like non- centralization means which may not be the same as decentralization um is kind of a very uh very different uh interpretation that the crypto Community does so what I'm curious about is knowing kind of how do you each Define what does it mean to uh to kind of see crypto and AI look like as a as an overlap and um and I guess from your past experiences and even just kind of now your refresh view of this world um how do you kind of feel the the AI Community right now interprets uh decentralization and and the crypto Community um in terms of actually bringing both these things together you want to go first uh sure yeah I think I mean I don't think AI Community uses world decentralization uh that much I think there is a there's set of values that are there in open source a community which are obviously like you know permissionless access ability to you know keep your data private uh ability to kind of build on top of it easily and those are very much the same as what we have in crypto and uh I think one that a specific property that like we we summarize all this decentralization and I I I really kind of like I've I've I've talked about this before that decentralization is not the point it's actually a way we get the properties we want right um and so if you can get them somewh other way that's great uh but one of the properties in decentralization that I think AI is very much missing is effectively you know some property of livess right some guarantee that this thing will actually be there um and so right now that uh especially on the like EV from evolutionary perspective doesn't exist right like yes meta maybe published the model but we have no idea if they going to publish an next one right so like you have no guarantees that they're going to be continuity of some service or property even if it has other properties you know out there and I think that's that's what where again crypto and blockchain has kind of a opportunity to come in and provide that right hey we have the incentives and and kind of coordination to make sure there's continuity of this thing and then we align on you know permissionless and uh kind of ability to build on double with Etc so that's kind of how I see this both convergence and and kind of where some of the Divergence right now is maybe as a as a quick clarification what does it mean to to have liveness for for AI are we purely talking about availability of kind of the expectation that I'm going to get a response and and not having a central point of failure or or something else so I mean I think that is important so I I think of this as more like a general value right like when you have like a general livess of things right you can build on top of it permissions really easily and you don't need to worry that you know you're G to get red pulled effectively and it's accessible at all yeah so so Services you know libraries new foundational models like that that's like I I I talk about it in the more broader sense I mean definitely for any like inference service Etc right right now all of them are still you know like you don't have any guarantees we actually have something coming out we have a paper coming out on this exact thing soon um but uh but even beyond that yeah like how do you guarantee that the next you know model foundational model the next advancement will come and not you know get kind of consumed so like those kind of things are are not there yeah so for me in in terms of decentralization um I think it's really about who has power or who has control right it's really about so one thing about blockchains are that they are shared resources the whole point is that everyone can come there even though we may have misaligned interests we may have different goals um and and that it is a shared environment with sort of a single rule set about how things interact with each other right and so then the key question is okay how does that get decided what are the rules how do the rules change how does this evolve over time who's got the power uh in order to do that right and really effective successful blockchains do an amazing job of creating something that seems almost magical right that you have this space where parties can interact in a fair and transparent way um even though though they're in even though they're they have very different goals um and um and things work out and decisions get made collectively about how all that is going to be going to be managed at the technical level um it's an incredible thing right and it's something it's it's both a big challenge for the blockchain space that we have to do this but also it's the main reason why that the things that we're doing have value right and so that's the thing that blockchain can provide to the AI world is when you move AI things onchain or chain adjacent um you do change kind of the Rules of Engagement um and that can be really powerful uh right because in the AI world people talk about uh there's a lot of talk about power and a lot of talk about who has access to what and so on a lot of push and pull between different players but there's not really a central space where they all have to work together like we have with blockchain but blockchain can provide that um I I think one piece that I feel like I I have noticed um especially over the last few weeks and we kind of looked at everybody's application was trying toh hack is that they kind of compartmentalize where crypto fits in um to AI for them it's kind of like the the equivalent of I really want to do something with AI but I should see if I can swab one or two things and use blockchains for it and that's kind of what my matchup is U and obviously um there's nothing wrong with that because it's how we started in in our own ecosystem too you kind of look at how do you sort of Swap and and make a better platform um over time but the thing I'm kind of curious about is um from what you've seen and in the interesting P you both represent two interesting ecosystems here um that are all kind of still working together so maybe starting with that again like how do you kind of see rollups in general kind of the evm uh giving people a leverage for how they can understand where crypto fits in and just overall yeah how that better I it starts partly with the community and what people are doing how they work together and and so on um our team um you know working in the arbitrum community we've been really focusing on trying to sort of upgrade the capabilities of of a chain starting with evm you know we shipped arbitrum stock which allows people to write smart contracts in other languages rusts the C++ and run them composedly on the same chain with evm right so um you know I think letting people letting people program in familiar ways and getting the the speed advantage that you get from being able to go at the speed speed of native code rather than the speed of evm um of an evm interpreter uh that makes a big difference um and so you know there's been a focus on trying to sort of upgrade the developer experience if you will without taking away everything that people know and love about evm it's all there you can call it you can use it you can compose with it but um but also you can uh you can program in a more familiar way so I think that's one piece of it um for sure uh on the technical side the other one has to do with data right AI applications can be data hungry if you're going to be doing stuff on chain especially um you need to have a really strong story about how applications can store access use data um that is accessible on chain um and that also requires a lot of careful thinking because I think one of the great strengths of evm and ethereum is the really careful thinking about uh gas and cost accounting and the ability to have a system that very robustly allows home us users to run nodes um and yet still is sort of provides a very general programming platform so you know I sort of think of within that tradition and within that approach um of kind of sustainability and the ability of Home users to run nodes uh how do we get rid of the parts of the or improve the parts of the evm experience that uh people do complain about um and so you can have the best of both that's really what we've been focusing on doing um I'm going to get into like what you've seen so far as well but uh it's interesting because IIA you kind of have a similar view and and you've kind of taken the wasm approach um how do you kind of think about making that experience better as you uh now see a big increase in AI applications yeah so I mean we we've taken uh kind of approach to use vasum from the day one because we believed uh that that we want to give developers the flexibility to build with whatever language but also because we wanted them to be able to use any library right one of the challenges withm always been like you're kind of relying on cry on whatever 2,000 crypto developers to kind of build everything uh and I mean near actually is a home for probably the most complex smart contracts in web 3 right we actually have evm full evm implementation running a smart contract right I don't think any other ch is even possibly can can like fit that into into limits um but the other thing we've done is the rvm is a synchronous and so uh when you're interacting with contracts uh you are effec sending a synchronous messages and every account and contract is effectively um you know you can consider them different roll up a chain or whatever and we just like package and make it really a nice experience that's why we're also doing a lot of ch abstraction because it's very natural for us because we already abstracted effectively all the near chains under the hood but the interesting thing for AI is and and and broadly for verifiable compute is that we can now directly in the same transaction uh yield the transaction execution and call a method inside a secure Enclave uh which can run yeah inference which can run some complex Computing which can do you know pull information from internet and apis make a decision and push it back on chain and restore the execution of the same transaction right so to give you an example you can effectively call an uh amm contract this amm instead of deciding that you know x y equal K it actually goes runs an AI model that does inference over whatever the market state is and decides you know that you should get this much tokens and send that reply back on chain and you receive those tokens in one transaction from a user perspective so that's kind of like you can effectively overlay AI execution and calling other services directly into execution of your smart contracts and you get very similar properties from executing inside secure enclaves because they effec give you verifiability and and privacy and other properties from there now I mean obviously there's as there but um that that kind of how we think about it is like really expanding capabilities of what you can do on chain with this kind of ke compute how um might be helpful for especially the developers here who who are who are listening how does um what's a good way to think about intuition around doing something like that in in this example and then understanding there's also a notion of a block time where I gotta make sure everything fits in that same limit and um and what's like a way we can simplify how to think about these things because some of these are not going to be 100 millisecond actions yeah so that's what my point like on near it doesn't matter like it can take seconds and then you can resume the transaction so near transactions can take multiple block times and uh effectively effec every cross contract colum here is a message passing to another entity right be that another contract or in this case we added ability to send it to verifiable compute engine and so you now or you can send it to eum or you can send it to salana right so to bitcoin even so like near itself is you can think of it as like a multi-chain multi execution environment that kind of orchestrating message passing around around the whole web 3 and uh you know that includes all the inference capabilities and everything else I also add because of vasm we actually I mean we've had kind of JavaScript and typescript for a long time but we're just adding python as well and so you can write smart contracts in Python now and uh so now you effectively can write python smart contract that calls into python agent code that runs you know pytorch model and gives you back result and you can interpret it in your python code to act something on chain or send a transaction or send a message call to another chain so through near you can effectively have a inference actually posting stuff on salana or ton or ethereum or what Bitcoin and and is similar what I think about this for stylus as well Ed sure yeah I mean the whole point is to sort of unlock the ability of program to express what they want to express right and then you get to the scaling problem right which is um you know there's an unlimited number of programmers who want to do stuff that is as complicated as the chain and you know and sophisticated as the chain will let them do um but there but there are resource limits right that the nodes of your chain the proving system of your chain needs to be able to handle the resources you have right and because demand is so high um resources will always there there is always a need to manage resources resources will always kind of be scarce right the point of blockchain is that different people's applications different smart contracts and so on can interact with each other right it's that kind of uh that sort of um a Gathering Place if you will where they can uh where they can they can talk to each other um and so you know we're really focused on trying to drive the efficiency of that exec as far as we can um and then present to users the tools they need to manage complexity and resources and so on because ultimately you know just like if you're when you're doing Cutting Edge programming that you're going to be using as you you're you'd like to use a lot of resources and so you're concerned about cost and scaling considerations everybody has to be programmers have to be and the way we think about this is first of all to provide as much Resource as we can and then second to give people the tools and the feedback that they need in order to um manage the resources they're using in their application as as effectively as they can um a good piece of what I want to use um this panel for is actually helping um either understand things a bit better in terms of how developers should make the best use of leveraging blockchains and um and kind of what we can help them unlearn so kind of the two follow-up question I have from both of your replies are one um does this significantly change the definition of what it means to have H inference now uh because we've kind of at least my interpretation for both these sers that like we've now abstracted things to be just message passing and what matters now is like as long as we can communicate things whether it's over blocks or just to the end place whether it's onchain or offchain uh We've effectively just use a blockchain as as the or for orchestration uh but when you have things like okay I want to have a complex prompt and I'm going to have it do reasoning and even if I pass it on to um an open source model here like llama and it takes 20 seconds to get a reply back or maybe to a minute um do I what's a good way to think about okay how should that be built as an onchain interaction is it just I just wait it out and have somebody just deal with the the data and then send it back to me and all I care about is the Integrity of the encrypted message or am I should I be introduced to A New Concept um how can we simplify llm specifically here and um do do both of your ecosystems uh have a different interpretation for this um yeah I I don't know about whe whether there's different solutions I think um you know time and asynchrony are actually really important to think about um you know understanding what it is that you need what assumptions you need to make um in a lot of the applications that people are focused on I'm thinking things like Defi and gaming timing is really important right not only how long does it take you to actually take an action when the world changes because fundamentally the world that your you or your application are in is always changing and you need to be able to decide and act before the world changes um again um and so that that becomes important just thinking about this timing the other thing that you see is uh in some applications is what's sometimes called Meb which is basically about somebody seeing what you're trying to do and then sort of jumping in front of you right so if somebody sees you sent a query off to a model um that is that may be a hint for them to uh if you know if you're in a market or you're in a game or you're in some situation where you're interacting strategically with them you need to think about like what it is that they're seeing you doing and um and how they might react to that um and you know so one of the reasons why original blockchain models were focused so much on synchronous execution synchronous transactions is they make it a lot easier to reason about that stuff um but I think you know one of the pieces of advice is just to think really carefully about timing think really carefully about latency think really carefully about if you're in a strategic environment what it is that will be visible to your to the Strategic Rivals of you or your user in terms of what it is that you're doing um and that could be fine right depending on which application you're in but I think it's it requires application specific thinking and these These are going to be some of the hardest issues right there's work going on I think in a lot of different uh communities to try to develop um programming models programming tools that are going to help people cope with all of this uh so that balls don't get dropped in the same way that um in you know programing programming of Internet services there's a set of tools that arose for writing asynchronous programs that interacted with other say apis across the uh the net and the different ways of managing that and dealing with errors and so on and I think the same kind of facilities are are going to fall out here um especially when you get into more complex and cross-chain worlds because um you know there are complexities that come up with asynchron there are errors or uh other things that can happen and um you know getting to a point where programmers don't have to think as much about that stuff is I think a big picture goal that all of us are going to be working on for some time um IIA you are muted I'll give you from from our from our side I think first it's important yeah to define something things so uh we Define effectively an assistant agent right so this is an a something that drowns on the user side and it's effectively designed to benefit at you and kind of be on your side right so you know think of it as your co-pilot or your interface to the internet and I mean a longer term I think the assistant agents is actually how we gonna use all of the Computing and internet and communicate to each other and uh all the apps are going to disappear and websites are going to disappear so so that is like one just assuming that is that we can we can go down that down that line I mean that's that's effectively just just you know I think it's selfeed at this point so the internet we well and I mean the interesting thing is actually it's gonna drive everything to be more peer-to-peer because right now youly have checkpoints to go through like Google and aggregators and if your system can inspect the whole world you know at any time and evaluate it uh you don't you don't need to go to checkpoints but anyway putting that aside so you have ai assistant uh that's effectively you know from a crypto perspective just to give you use cases this game anything from like hey you know go DCA into this thing uh you know make sure my portfolio is you know not get liquidated Etc right like you can give it you could give it instruction to do things it executes it on your behalf now the second type we Define is effectively we call them verifiable agents uh where you have an agent that is running inside secure Enclave you know that you know this code and this um kind of model is running so you know you know what what to expect uh it may it may still be non-deterministic because it's fetching external data right so like the exact state of the world is changing um and uh so this verifiable agents can you know either be pieces of functionality or they can be like orchestrator so you know hey run a marketing campaign it's like cool I'll do it you know let me find other agents and people to do it and coordinate them uh and so they can themselves actually act on chain right they're like oh okay I'm verifiable agent that knows how to rebalance portfolio tetr back end finally we call them onchain agents although uh we actually explained the name called based agents and so this is where the smart contract itself has a logic and execution uh and it uses effectively external compute right inside secure Enclave as a kind of you know offload to it and so the difference here between verifiable agent is like verifiable AG has its own key it manages all that and it's effectively like constrained by security of that te and and on chain agent uses to but like kind of doesn't fully trust it right it can verify whatever conditions and can apply any restrictions on top of it uh it can you know ensure liveness Etc through that uh and again we have some work coming on on on how to ensure liveness and response across decentralized Cloud um and so those are effectively like I think when people say onchain agents like they people imagine those things where you're effectively writing something smart contract code that can call things but actually nobody built those like so far because it's actually very hard and so we actually we have the whole framework to do this you know you can do this we have like Oracle example we have this amm example I mentioned and near is very well designed for that because of this as synchronicity at the core so you literally are in the same transaction right so you have all the same context when the response comes back uh which uh which normally you wouldn't have uh but verifiable agent I think is like extremely important concept that because they also serve as private right any session that users have is private uh you now can actually have them do you know read your email and act up upon that you can have them you know act with the credentials of your we website to optimize the COO and and and all those other use cases in real world that you know people right now are actually paying money for somebody else to do um I think a both of you have mentioned a pretty consistent theme around sort of leveraging um secure enclaves and and tees for for a lot of this and it kind of is very obvious because it fits it nicely with how we interact with these services but um mostly curious to see are there any obvious things that Zer knowledge proofs SK can simplify or enable when we lever these things and I'm curious what that looks like sure um yeah one of the places where zero knowledge proofs uh in principle can help is if you have uh if you have a situation where there is a model that is uh agreed upon or someone has committed to the model and they want to prove that applying the model to some data yields some particular results um right and there's different ways of doing that right so in in the cryptos in in the blockchain space when people talk about zero knowledge right if you were to talk to a cryptographer them zero knowledge means something about privacy or confidentiality right um and there are applications where that's what you want right where maybe you know I say I have a rule for making decisions about people are situations and I'm cryptographically committing to the rule but I'm not telling you what it is and then later I can apply I can prove that I applied my committed rule to the particular data you gave right so like I have a I don't know I have a model that can predict who's going to win the uh you know who's going to win the next Champions League game or something and I'm not telling you uh you know based on whatever data about the players or the teams I'm not telling you what my model is but I'm committing to it now and then like later I can prove to you that yeah applied to this data here's the result so that is using the zero knowledge aspect of zero knowledge proofs right but you can also do the other thing that people in blockchain do which is to quote quote unquote ZK which is that you're using these sort of um concise cryptographic proofs but not for secrecy but just like for some notion of efficiency and there again maybe you have a model that is too expensive to evaluate on chain but you can publish some cryptographic commitment to the model and now somebody can prove to an onchain contract um that yes this is the result of executing that complicated model on on particular data um and you know that also is very empowering because it lets you move sort of offchain into a kind of co-processor the the evaluation of the model but in order to do that you need to be able to generate efficiently that concise easily checkable proof of model evaluation and you know that's that's a real technical challenge to be able to do that on models that are large and sort of Cutting Edge um a lot of interesting not deterministic right sorry and also non deterministic so well it' better be deterministic if you're gonna prove it right um yeah exactly it's hard to back apply it not without but you can take you can take models and make them deterministic in a sense where there is a deterministic answer right um and then um or I there's different ways to work around that but fundamentally you one as a pre yeah one of the Annoying problems you need to deal with if you're if you're working on this model proving stuff is how to deal with the fact that um you know a typical model evaluations are mildly non-deterministic right there's a whole set technical challenges there right that and that's but people are going to be pushing that envelope and I think we're going to see more and more sophisticated model evaluations being provable on chain sorry I mean to be clear the reason why they not deterministic is Hardware like the model itself computation is deterministic I mean given a seed like for sampling yeah the reason why they not deterministic is because Nvidia gpus are effectively optimizing in parallel when they do matrix multiplication and rounding errors like accumulate exactly right yes that's right floating Point arithmetic is not actually commutative and associative yeah so so the the I mean I agree that this is like a very nice use case but indeed it's like very non-practical right now right this is not really applicable right now like any model that anybody wants to use is effectively not GK provable in like reasonable amount of time meaning like even I'm not even talking minutes it's probably hours um and I mean obviously we all hear H for you know ZK continues to speed up and like there a ton of research put in into implementing this but indeed like from our perspective te's right now solve all of the problems and they solve privacy as well so like to me to right now are effec solving privacy and uh verifiability and yes they have kind of you know there is ways to um attack them in in in from a few different directions but given we are piping data like large amounts of data through them instead of you know what normally people use them for like private key material and stuff it's like so much harder to to get anything out in those cases because all the attacks are like side channeling and the two nanometer uh Intel CPUs now like from a even Hardware perspective trying to like effectively uncover the hardware private key on the chip uh I mean at least like an analysis of security researchers is that it's uh kind of Highly technically um like the the cost of that is extremely high right so again wouldn't store like on a singlet te you know all all of everybody's money like that's not but like for any kind of data piping in know influence AI I think it's actually a very good solution and then uh yeah we should continue investing in speeding up zuk for sure um and then we can actually have both oh that's it's it's obviously um it's a moving Target too and we kind of get to understand what the limits are and improve them and that's uh quite kind of frankly what the goal of research is and you kind of Define the new constraints and see if you can exceed or reach um two more questions before we end this a quick one would be um I guess from what you've seen in your ecosystems or generally in the space what are kind of cool and interesting crypto AI mashups and applications that you've been excited about and uh I mean I think I think the the first things the first things we saw for sure were people uh just putting agents in especially in defi to act on their behalf um nothing too sophisticated is just really the ability to have an agent and to have it um engage in some kind of more sophisticated Behavior than just like a simple if then rule set um there's a lot of lwh hanging fruit you can get without going to the most advanced AI as AI um I think that's really what we've seen more of than anything else there are people pushing the boundaries trying to understand what are the limits but um you know as in any kind of engineering making very practical decisions and not getting too fancy or complicated for the problem you need to solve um I think I've been impressed by just sort of how practical people have been in thinking about what to do that's awesome U IIA so yeah I mean from our side like last year we we invested a lot in kind of building out like fundamental infrastructure so this is you know uh kind of tokenization of compute to make sure that you know there's a kind of ability like liquidity in compute markets um the some of the kind of inference distributed to decentralized inference um some of the foundational models that the train specialized for for example on chain data like so you are able to have like a lot richer information uh you know AG indic ments um and kind of synthetic data data Crow sourcing so kind of the the more fundamental blocks the way we've saw this is you know if we want to have a alternative to one of those closed Labs we kind of need to have that stack available and and uh people can address this in parallel obviously as Ed said there's a kind of a bunch of people attacking this from like okay well how do I give people tools to use the blockchain in a more natural way uh again in in this mindset of like chain abstraction like the final form of chain abstraction talking to AI right you don't really need to think that there is blockchain even existing so bit is really good example of like a fully chain abstracted wallet it it can use dozens of Chains It's all AI there's ton of agents that can actually like plug in and execute actions for you so uh so that's kind of One Direction I think what uh what we see kind of as the next wave now coming is actually starting to replace different functions where right now you either have a people or some kind of very strict game design like Game Theory design which is not uh very sophisticated right so the examples are obviously the governance right so starting with like your governance co-pilot right something that can summarize all your governance Forum threads and give you like actually what you should know about this topic but also going into what do AI delegate looks like what is like replacing the person like especially now there's like professional delegate companies which are fact just like doing that as a company like what do that replacement looks like it can summarize all this information it can make decisions and can propose things um it can query it uh you know its delegates and then the thing that I'm really excited I mean this is more longterm is AI president so like actually managing work right in a dow or in a in a in a company full end to end right removing hierarchy the other examples are you know AI like oracles right now there's humans effective resolving like for for prediction markets or for like more complex questions there's humans resolving effectly markets uh based on you know reading information on the website and deciding you know if something happened like did Trump win or not did you know this team one or not like that is replaceable by I now obviously there is still gaps in its understanding so you need a backdrop with people that you can start automating you know 80 90% of the work so effectively now it's moving into this direction where like you can you find places where right now you need a human in the loop and reduce that time and requirement 80 90% And you still have a human behind to kind of make sure it's going and yeah so I think that's GNA be like the theme for this year and that's as I was saying kind of some of the mechanisms we built is exactly for that oh that's awesome um and this is probably a really good note to end on because this kind of gives us a good uh look into what I guess from your prediction just this looks like so I can't wait to uh see for the next year AI hackathon just one of the AI Bots being the the other panelist we'll see how that goes but uh I you don't know you find a or not well I do know because you were muted for a while you were talking so any can make that mistake unless you program that of course yeah obviously right that should be built in you can go all the way back to turing's original paper on the Turing test and in it you need Tock them but uh and Ilia thank you so much for this chat and uh we'll keep you posted on the cool projects to come out of this next week really appreciate your time thank you cardig yeah take care [Music]