[Music] All right, next up I want to bring on Ilia from Near to have a conversation about crypto and AI. Please give Ilia a big round of applause.
Welcome back.
Thank you. Um there's a lot of things I want to touch on and um talk about and this kind of jumps in from your knowledge about AI from crypto and before when it was a lot more small and a lot more creative as an industry actually still creative now. Uh but the thing I want to start off by asking you is from your perspective like what's the actual natural overlap of crypto and AI? Yeah. So there are a few different aspects right where it makes sense.
Um I think for crypto AI brings a lot of simplification right you can like right now in a lot of cases you as a user need to figure out all the details of how to get things done where to find things. Is the asset you're trading a real asset or just you know fork of a fork of fork? um what information is going on why the reason right I mean and we see this right now kind of you have you know coin market cap AI that like gives you some information about the thing you're looking at and why uh maybe it moved uh you have some assistance where you can start talking but obviously you know this is going to continue evolving and and uh improving and so this is general pattern across not just crypto but the whole web internet computing We're going to change how we interact as computing to be more AIdriven. Not always natural language. I think this is also like people think, oh, it's going to be chat bots.
No, it's not going to be chatbot. AI can predict what you want and actually have a generated UI with exact information you want that you just need to click on. Right.
So when you say that, are we kind of defining the reliance on AI as a faster way to get answer to a question or is there a different metric here? So I would say there there's effectively two routes. One is you express intent right so and you get to the action not always the answer but to action and the other way actually it predicts which actions you want to take and you just need to authorize or it does it for you even at all. Right. So uh like the the the experience that we will have is not going to be unique.
So specific the prediction and not the execution. Is that is that what we're also measuring or are we counting both?
Both. Yeah. So prediction and action both.
So that makes sense. I agree. We're going to see a lot more of that. How do we account for reliability on the execution of that intent?
That's a great question and there's a lot of uh uh kind of um issues there. So the what what are the potential issues? One is and and this goes to like what are the issues now AI has, right? The reliability is a general AI problem.
Compounded.
Yeah, compounded issues. You have obviously, you know, malicious vulnerabilities that are injected into either AI model training or into like agentic code around that. You have obviously data privacy leakage, you know, people training on data they shouldn't, etc. you have an interesting trust problem where if I'm a model developer I'm don't trust like uh you know generally other compute providers because they may steal my weights and so I generally you know I have a contract with Amazon or you know open as Microsoft and kind of use their compute or build own compute but there's like a trust issue which you know we're kind of in a space
I think we know a little bit about that
yeah we may know about that um there is just gen generally like a sovereignty problem. So if you're sitting outside of let's say US where you know most of the models are built and you're kind of thinking about it you know hey you know our government will use more AI our companies will use more AI and now this other third country can effectively decide on what and how we can use it and what decisions it's going to make based on the whatever training data. So like there's a fundamental sovereign
buy a whole nation.
Yeah. bias the whole nation into some direction or just turn it off. Imagine just like you know let's say the some country decides to use let's say open AI for their government operations and open AAI can just like a we'll turn it off. It's like you know the amount of
found new equivalent of disabling the internet in a in a competition.
Exactly. Yeah. Yeah. So this is kind of and this is you know if you project it obviously if like one company wins kind of the AI race we effectively ending up somewhere between AOL internet and 1984 big brother world right because effectively every single interaction will be facilitated by one company and then it will drive how you make decisions and how you what information you see uh kind of on day-to-day. So those are all things where crypto and block more broadly blockchain kind of decentralized technologies you know have some answers and solutions for but they need to be built for the this kind of AI problems.
So what's the nuance here? Like what do we need to build for and and um is there a difference between building a specialized blockchain that is designed for this or is it that we need to make our crypto networks more compatible with how LLMs are kind of operating so that they can talk to each other a lot more easily. So it's it's more of a kind of again if we're solving AI problems we need to think from AI first perspective and kind of build the stack for that right well AI is about compute hardware um kind of models inference you know then you go into like application stack and then you go into like you know aentic and all the other spaces right so well uh if we want to have kind of no centralization on kind of compute side right we need to figure out how to decentralize compute well the biggest challenge we've had is that blockchains are you know way too expensive in reexecution zk is expensive also we don't have any privacy on the in the current state so we need some way of doing compute that is private and effectively doesn't add much overhead well it wasn't really possible until last year but now with Nvidia you can actually have a confidential computing mode
sorry what changed last year confidential computing.
So yeah, Hopper, hoppers and black holes now have confidential computing. Intel and
the new Nvidia chips. I see.
Yeah, ARM has this new TDX mode which is like at least easier to program and run arbitrary software instead of previous generation. I supposedly have also less issues. But uh obviously the important part the trade dovs here are very different because you are running AI models and not uh uh kind of relying on on purely that for like uh financial transactions but this enables to have now confidential and to some extent verifiable computing that has very little overhead. So we've benchmarked it uh and you know it's 1 to 10% depending on the kind of you know RAM usage etc.
So kind of tying that back in It it provides us slightly better guarantees around people not abusing the information, but does that mean we now need to change up how we're building current smart contract platforms or or how do you actually make that more practical with this blockchain and LM integration or just any AI model agentic integration?
Yeah, so now you can bring these things together, right? You can actually run the I call it the brains in this uh secure enclaves and then the muscle is
you think securely but everything else could be
yeah but you execute on chain and you can verify also that your brain is not going crazy right uh to use analogy and so you can have this kind of coupling where you know the all the certificates everything of the secure enclaves are verified on chain it's really easy to do that and now you can again so we have this mind share agent which uh was given $10,000. And so what it does, it reads Twitter and analyzes for sentiment of different assets and then it prepares effectively through our multi-chain intent uh trading platform a portfolio and rebalances based on the change of sentiment.
And so it made the performance
4,000 so 40% up in two months.
That's pretty good.
Yeah. Not financial advice. Uh so but you you can imagine this like expanding this and and this is going to kind of broad broader sense like this actually enables effectively autonomous agents right right now everything in in kind of both in web two and in web 3 to some extent it just runs on somebody's machine right it's just you know you can turn off that machine and that agent stops running uh well now here you can actually run in this kind of decentralized T cloud And you know you cannot you know whatever machine you turn off there's always another machine that can run exactly the same code that runs off the same state and it can you know holds assets. So now you have autonomous agents which are effectively unstoppable. You can have you know some governance with tokens etc over them.
Then they actually can start doing business. They can potentially go raise money through a token and acquire a gas station or a laundromat and you know run it. Now, you know, we're probably still a bit away from that, but this is kind of the idea where, okay, how do we give agents a way to actually participate in commerce, both onchain, but also offchain. Well, they needed new way of effectively legal contracts, arbitration, like all of this instruments that we traditionally have in commerce. Like normally, you know, I'm sponsoring this event, you send me a contract, I need to sign it, and my lawyers check it, you know, then, you know, wires, etc.
, money, etc. like all of this is decoupled and it's like very complex for I mean it's solvable with AIS but when we go to kind of the
well everything that's not real world compliances in a way because this will represent yourself as an entity if you are doing something that touches real world but payments 100% make sense yeah but but I guess the thing I'm trying to understand is it seems like the bet you're making is that great we can get some better guarantees on computation if it's in a TE or or a more secure computer environment. But we're just hoping that if you just wait long enough, things will be as performant and as
So that's what I'm saying. They're pretty much very similar performance or including
the GPU advantage.
Yeah, that's what I'm saying. One one to 10% and like mostly it's 1% overhead.
Oh, over I thought you meant 1% of performance.
1% to 10% overhead. Yeah. Yeah.
That's actually pretty good.
Yeah, I know.
Okay. Today I learned this is amazing.
Solved. Okay.
Solved. All right. Well, we're done. Um, okay. So, does that mean we have to still change the networks in terms of how we design blockchains or
I think I think we'll we'll start actually changing over time. I don't think we need to like change a lot.
What you described can be replicated as a smart contract or slightly better execution environments on top of existing chains. Is that a good way to think about it? So you can call from existing chain you can call into this let's call it co-rocessor
black box
co-processor
right we'll go back to 2022
intelligent crocessor right I mean but this is what roll-ups are rollups are the same thing
well yes let's this is not this is not our this is being recorded so we'll uh we'll not go back to how we usually um but yeah I mean it's effectively you know intelligent comp processor that you can run and then uh you you can verify that whatever decisions outcomes it did are uh validated and then but yeah for a block from a blockchain side you need to like integrate it tightly so all the verification everything is like done uh very easily now over time I think if we want privacy we probably need to start moving our blockchain itself with into GS and like start like but all this still assumes that we're just giving a secure computer environment we're not actually doing things with like encrypted information being acted on or being modified we do. So, so, so is it to still then solve for like anywhere from rack to just other context setting and having a notion of memory that is more readable. Um, so you're you're kind of feels like if I'm understanding this right, the constraint is still do everything in a box that guarantees no tampering, but the process remains the same.
So, so, so it guarantees no tempering and it guarantees confidentiality, right? uh
confidentially of what compute or something else
confidential compute
but not the data itself
and the well the data is encrypted for that specific box when you send it so you have a TLS connection with that specific box you encrypt it with a TLS certificate only decryptable there and then on top so the next primitive that we on near we have is called chain signatures so chain signatures allow to effectively have a programmable encryption decryption uh and so you can actually attach to a near account or to a specific Docker container that will run in TE a way to recover a decryption key and so have a that's awesome have an encryption key so you can have encrypted data per account or per instance of the of the agent that is only accessible inside TE and so now you can actually have a private memory you can have private state that you can act upon uh and this is just like traditional encryption not
yeah this is NPC based encryption decryption itself runs into
a good context here for size okay
yeah so now you're solving an interesting problem that again in AI right now your context is effectively part of it is an openi part of it is in a tropic part of it is in Gmail like it's all separated um and so you can actually you know take all of those context put it under in encrypted way under your account and then give it access to it to different applications that also run inside CE so you know they not leaking your data right so you're starting to solve actual problems for people
which yes so you're giving everybody like their personal memory box and they can attach
yeah and then you can bring models to it instead of giving you all data to the model
that's going to be awesome so so is something like this possible on year right now in terms of
this is should be like launching in the next uh coming months
and how do you kind of think about just overall intents for payments with uh a lot of the the agents coming into play
yeah So, so intents for payments is effectively this new way of how you can actually do legal contracts, right? So, the way intents work is effectively you say, "Hey, I want this to be done." Right? So, you describe effectively your intent which can be natural language can be in kind of structured format and then you have other participants who say, "Hey, I will do it in this way." So, when you pair these two things up, you effectively create a contract which can say if we don't agree on outcome, we will use this arbitration which can be also AI uh this is like Shirram loves to talk about that.
Um and then if not we can also use this actually court system in you know commonwealth or in in the specific jurisdiction and so when we agree on that this effectively settles and so then it can be executed and kind of adjudicated as needed. Uh the simplest use case of that is you know trading obviously because the adjudication is very easy did the funds arrive or not. Uh but the more interesting use cases so we already had in fiat you can add like other real world assets but more interesting is actually starting to add work right so I want this to be done
like work as a general
asks yeah yeah I want this to be done I want a tree to be planted I want you know website to be built I want this analysis to be done and you can have AI agents now coming in and saying cool I'll do it let's settle on that and then you if you have a guarantee that this AI agent runs inside a secure enclave you can now give it more context text and more data and more more access even to actually act on your behalf potentially.
How far is that from a a pitch to reality?
Uh I mean we have an MVP for like coding to start
actually maybe maybe a higher level question. How much of being able to program something like this is specific to programming on Near versus me taking some Python SDK out of the box and not having to do too much tampering and come up with a new system or a framework to uh play with this.
So the agent itself you can build in Python SDK right blank chain or whatever right that's not the the pieces that you need is the effectively interfaces right one interface is how the agents talk to each other. So in again in normal AI right now when people talk about agents they usually consider agents running on their computer right but that's not when you want something to be done right
it's not always running that's
yeah you usually hire somebody else
push or yeah poll based
right and so so you need a protocol how like a person or an agent will talk to another agent who runs on their own behalf like across trust zones and so kind of that's not really a thing in normal AI space but this is very much a thing in crypto where you know we always work across trust zones um and so we have this agent interaction transaction protocol which effectively AITP as in HTTP for AI and so that's kind of designed for effectively this interface now this it has blockchain components in it to have this intent settlement signatures etc but it's not a blockchain
this is just a communication protocol
it's a communication protocol it works over different uh transport layers and you know you can exchange structured information, you can share things, you can request things, etc. So there it's a lot of just pieces that you know, you use the blockchain and web 3 mindset and then you kind of bring all this together. You know, maybe we're building web four. Let's just let's let's let let's set three settle first. Um
finish the web three.
I think we're still 2.5 from whatever. Um, I think you said something at the beginning which I want to come back to, but you kind of said that ultimately you don't believe we're going to have interfaces in the future.
Yes.
How far of a future is that? And uh, and I guess more concretely, what do you think there will be interfaces for?
Well, sorry. I I don't believe we're going to have websites and applications the
So, what does that mean? So what's a what would be a example experience for some application that I guess people have or things that people need to use?
You just have your assistant AI assistant running on your devices across all your devices and it will you know when it thinks it needs to show you some user interface it will show you that user interface.
So it's just in time versus like not
Yeah. Yeah. I mean it would make have a cache of components right and plug it in like this is actually the the whole I mean back two three years ago we were building this blockchain operating system this was designed for that vision where we just started with humans building components but so we can train AI to generate components um but this is I mean we already see this happening right like you can vibe code your own applications
you're right
and you can imagine if you have all the APIs and services like one of the problem
looks Like time's up. This fireside chat is fire.
Time to exit.
I should have paid more attention in my French class.
Okay. [Music]
Automatic transcript — names and jargon may be misspelled.