Sean Li on Application Abstraction, Verifiable Agents, and AI’s Trust Problem

Transcript

I'm already seeing that the user behavior is shifting right in general cuz a lot of people are dependent on AI right already. But um if you ask people everyone's kind of hesitant about using it for like financial use cases. That sounds like something that's going to get massively adopted. If we don't really show what the end to end experience really look like, it'll be hard for people to really understand it. Back to the, you know, beginning of like telling that story and helping people really see the vision. How do you make sure that agents aren't going rogue? Welcome back to the rollup. We are your hosts. I'm Andy. I'm Robbie. And this is your home to learn from the biggest guests, the brightest minds at the tip

of innovation. Our goal is to help spread the collective knowledge of everyone in our community. We are breaking crypto out of its echo chamber and taking it to the real world. Sit back, relax, and enjoy a quick word from some of our awesome partners who make this show possible. [Music] Innovating finance, global, institutional and onchain. Deep liquidity to drive capital efficiency.

Now maximize yields on institutionalgrade liquid staking with M. Unlock trillion dollars of capital via F Bitcoin. All of this secured by Mantle's treasury of billions in stables and ETH capital pools. This entire ecosystem is coordinated by a vast network of seasoned VCs that are aiming to create the next generation of banking and financial technology. This is Mantle. We're in New York City now. Sean, you're in town. We just had an amazing rollup magic happy hour last week, which unfortunately you couldn't be here for.

So much FOMO. Um, but it's good to have you in town. Two fellow Canadians in the mix. Big couple weeks for you guys recently. How are you feeling, Sean? How are you doing uh with your with your building? How are you feeling? Pleasantly surprised by the overwhelming um support and and response and positivity around what we've been working on and um yeah, feeling tired cuz it was a lot of work uh leading up to it, but now very energized because people are really resonating what we're building on. So this week in New York, planning, you know, with the team on what to do next. Yeah, man. So, we're going to talk a lot about AI today, AI and crypto, what you guys are doing at

Newton, but it would be remiss to not understand a bit about who you are and uh why you got in the space. Some of the interesting things that I've came to learn is about your early career at Waterlue as well as selling a company to one of the largest uh development companies in the world. So, who is Sean and and and how did you get started in this interesting design um computer science kind of background? Yeah. So, if we're to go back to uh the high school times, I was in like five different bands. Most are jazz band. Uh and then uh I told I told my dad that uh I'm going to do music. And he's like, uh but I have to take five jobs, right, to make

a living. And then my dad told me, hey, what about this? You can take one job and make a living um by being a software engineer and you can play on the side. Yeah. So I said, "Oh, yeah, that was that sounds nice." So I like computers. I had uh I had one when I was small. So um I was always interested in like playing around on the computer and so yeah, eventually picked the last choice was go to Waterlue for software engineering. So music was your first choice. Yeah. Yeah. Yeah. Um but I always like building things. I always kind of tinkered a little bit on the on the computer as a kid and um yeah when I went to Waterloo I formally learned uh programming but always had this knack to

simplify um really complex concepts. Uh so in Waterlue um you know built all this different little tools that streamline um you know uh building building uh software and then um when I was graduating uh at that time Docker 0.3 came out and we thought this is like an amazing technology um but it was really hard to get started with Docker on like Mac and Windows whereas it first started within like the Linux uh community. So we thought let's add a nice UI to it and streamline a lot of the setup like ver um because you need

to install this thing called virtual box before you can use docker. So we streamlined all of that and then made it one click so you can really experience the wow factor uh running a container on your computer. Um, one of the demos we shown was like how easy it is to run like a Minecraft server with one click and then um and then so that's like how it got started. Um, and it was called Kitematic uh what the project was called and we ended up open sourcing it. Um, and then basically right after we open sourced it, Docker reached out to us and like hey like what if we do this together? Um so then uh you know they're like a huge rocket ship and one of the

um major players in dev tool and then infrastructure. So we jump on the ship and um we uh further improved on the product to make sure Docker feels like it's running natively on your computer um by like so we had got these kernel hackers uh from Cambridge. like super great and they were able to get in the hypervisor and install like this uh Linux uh distro in there. So they effectively Docker effectively acquired Kite Kitematic then. That's right. And that was like your first founder kind of exit moment where you built something and then this larger company came in kind of validated your idea uh and then

that led to uh you know that eventual being acquired. Yeah. Yeah. and and they continued with the product into uh Docker Desktop, right, which is a way to um run Docker natively and that's being used by, you know, millions of developers monthly um for their regular development workflows. Um so very proud of what the whole team have done there. Um just taking something so complicated and making that run seamlessly right on on the computer um was a technical challenge both from like a user experience perspective but also technology because there's so much underneath

um it's almost like exposing the internal hypervisor some pretty comp complicated um tasks there. So to to actually make it run so seamlessly, what were some of your biggest learnings when you were building out Kitematic and how did that lead you to you know any challenges, pitfalls and um you know take those learnings now in your second iteration of being a founder with Newton and Magic. I I would say is this um obsession about user experience. No matter what, no matter if your target audience is a developer or a regular person, just try to shave off any complexity as you can. Um and overall,

um telling a story around that too, right? Like um like it's not just about uh oh, we have less clicks now, but it's more like what can you do? what is that wow factor within the first uh 15 minutes right for for a dev tool um it's it's to really nail that right and that like Minecraft thing was like one of the use cases that we tried so now that it's your you know second iteration being a founder um you've been building in crypto for a while found Ethereum in 2016 kind of you know shook your perception as to what what is possible by building online then you you kind of came to this embedded wallet idea. Um, which kind of comes back to

your user experience prerogative that that you've just shared. Um, talk us through about founding Magic, what that entailed, um, and why you guys decided to go all in on this embedded wallet idea. Um, and then we can get into some of the AI stuff as well. Yeah. Um, yeah, I think I I fell in the Ethereum rabbit hole right after Docker and end of 2016. Um before I wasn't as bullish if it's just cryptocurrencies. Okay. Um but Ethereum really showed use cases and there was a lot of vision right in terms of what it can accomplish.

um you know, DeFi were a concept being talked about like NFTTS and I I felt like I was able to learn so much just from being in the space not only from like a technologist perspective but also understanding money, understanding financial systems and um uh and and and and art even. There's rarely anything that's at the intersection of all these things. So I was just intrigued to the max, right? And and so uh I tried to build two projects um like consumerf facing like one is like getting paid in crypto, you know, uh but it was like really really early and when we're doing user testing

like no one could get started because at that time it was the de facto onboarding was through browser extensions and writing down the seed phrase like a pirate, right? And so um and we yeah I wasn't able to really get those off the ground because for me I was really compelled by the vision. I want to bring more people in the space instead of just um sort of focusing on the vocal minority right. Um, so that was sort of always trying to um get more people involved and so embed wallet um was it came up after I realized that there there just not a lot of infra. The vision is so great but there wasn't

great infra to get billions of users inbed wallet was like the first step at solving that problem. Um it's pretty simple right just like hey we can provide this stripe like SDK where developers can embed in their application to create a crypto wallet using web two login so pretty simple and and we tried you know obviously a huge amount of work underneath to make sure it's like trusted compliant secure um invented our own um like key management approach could uh DKMS basically like kind of like a cloud-based ledger um sort of style security uh making sure

users can maintain control right over their private keys um and just grew that um since mid 2018 um relentlessly focused on like UX making it easy for developers trying to get the best applications on board um and yeah I can't believe it's almost it's almost 7 years. Yeah, I just crossed seven years in fact la last week and um yes till today embed wallets have uh on boarded over 50 million unique wallets um and we work with some of the top applications like poly market helium wallet connect immutable um to some of the enterprises

like Mattel um Forbes web 3 and also neighbor in Korea so a wide spectrum of different applications we're servicing. Um, and with with Newton, um, it it it was really an idea that came last year, uh, where we we're always relentless in terms of like what new can we do to improve UX. Um, so at that time there were a lot of conversation around like chain abstraction, right? um just making um a chain invisible uh with like unified balance everything just interrupt like seamlessly. I think that's like a really compelling vision for UX. But where it really started clicking for me was um I was in some of

the AI meetups because you know I'm in SF there's a lot of AI stuff going on and some people showed me this demo around tool calling. So to put it simply, it's just AI calling APIs, right? And they can make decisions and using all these tools to get some stuff done. At that time, I saw this demo on like get getting the current weather from an API, something pretty simple. Um, but then like it started clicking for me that hey, this is like where AI and crypto can really intersect.

Um because crypto is like all these open transparent APIs that can be called to move like programmable assets around um instead of like these siloed APIs in like private databases. So I think so that's when it started clicking for me. So not only you can make chains very interoperable, you can also make these workflows super streamlined with AI automation. Um so we were like prototyping and building some demos towards the end of the year and yeah the narrative and project really started taking more shape the the more we started working on it um and we identified that so now a lot of people

are working on like tool calling like MCPs um but we feel like the biggest gap right is really the trust and security aspect of it um because I do believe the capability of AI and crypto um are going to um like drastically increase um a lot of them are open source which is great all these plugins uh tools um so yeah so we feel like there's a huge opportunity to use what we've known um the experience that we have to make something that is uh uh sort of this trust layer that bridges the gap because a lot of people see the capabilities of AI but They don't really trust AI,

right? So, yeah. And and I also have the have the thought when I'm using these LLMs and such, like where is my inputs and my and my prompts going, right? Like where where is this data being stored? Um who has control of this? So, that's one thought on the on the AI sector on on the intersection of of AI and crypto. I think it's been kind of a a rough start for that convergence. Um, you know, I think a I think about AI is scaling intelligence. um right that there's so many more uh effectively so much more brain power being used uh by these LLMs now today but bringing that into the crypto space kind of just started with like these agent with memecoin ideas um and then bringing that out to now like now we have like

decentralized modeling and decentralized data and there's kind of like this more uh robust kind of sectors being opened up um but I think the one thing that is clear is that this is a match where uh as you said crypto has so many of these APIs that these AI and these agentic you know forms can use it does seem that that there's a much better fit for uh aentic finance per se uh with within blockchain rails or crypto rails more so than the traditional kind of credit card debit card a wire kind of rails and so I'm curious how you think more broadly about this convergence from your perspective and um where you see the direct application of of AI and crypto

uh kind of happening right now and in the future. Uh because I just think that we haven't really seen that much in the space quite yet. That's really that interesting. And so I'm curious how you see it differently and and kind of where you think that there's actually real product market fit and where that's going to expand to. Yeah. Um I would say everything is a little bit early, right? like the cool calling stuff like really started picking up like last year. Um and then now everyone is like building and trying different things. Um whereas I see um the agentic user experience have a lot of potential is really around

um this concept of like this crypto super app, right? Like a lot of people think about that, but if you were to really imagine it before uh agents, um I I could imagine like this super complex UI, right? Because like you're trying to build a flow for every single user. I mean, yeah, that's what we've been doing on chain for the last four years is like all these different apps and the ledgers and the clicking and the signing and Yeah. Right. Right. And it just becomes this like spaghetti user experience that's really hard to navigate. Um but in some ways you also need those kind of features to service all those use cases but I think agentic experience kind of convert that into a very elegant user

experience um where uh it's like a single prompt and that gets translated into the correct series of actions and you can address different needs from different types of users in like a very elegant way without overcluttering um the user experience. So I think that's um I think that's inevitable and then that really um falls in line with like typical internet user behavior. Okay. So so like I um uh this is the sense I have with like Google, right? So that early on with um before Google came along, it's like complex websites with all these buttons and navigation very hard

to use, very hard to navigate for regular people. And then Google came along. It's like the first intent driven experience where on one simple text input box, you can start googling your intent uh things you want to search, things you want to do. Yeah. And that gets translated and ranked into a series of correct websites and then you you pick one that intrigues you. You go in and then go through the conversion funnel. Yeah. And then finally make a purchase. But in this new paradigm um that from point A intent to action point B gets shortened massively right um with with automation and a agent automation.

So so you you think there will be a future where there's going to be the majority or a lot of the new users onboarding to like just call blockchain rails through effectively a natural language model on the front end kind of like talking to an assistant of of sorts. Yeah, that's the thesis here. That's right. That's right. So in order to actually achieve that, we need to trust and be able to work with and understand what what's on the other side of what we're prompting. And that's like the problem space here. Yeah. Okay. So AI's role in traditional financial services is expanding to become kind of like this, you know, they have like robo assistant that kind of vibe, but but it's we're so early that that's still

kind of becoming a thing in humans dayto-day or week to week financial interactions. But taking to the next step is that in order to avoid some sort of like rogue agents or some sort of like malicious behavior or errors, the idea is that we need to effectively create a trust layer for these agents to interact with humans such that for their financial lives, they can have this this trust in the prompt being executed correctly. That's correct. Yeah. and and and we're I'm I'm already seeing that the user behavior is shifting right in general because a lot of people are dependent on AI right already but um if you ask people everyone's kind of

hesitant about using it for like financial use cases and then that's where the trust layer could really help with that leap forward. Yeah. You know, imagine like um you can just even do simp simple things like hey, watch for volatility and do XYZ or things like book a weekend travel. Yeah, cuz I I was going to say I don't know if I don't know about you, but I'm like I don't really share like my name or like my photos of myself with LLMs. like I'll tell it stories or I'll ask questions or I'll provide some insight into my financial life or personal life or social life, but I try to avoid all of like the like like doxing or like giving too much information or whatever just because it's like well there's a certain you know there's something happening

here that that I'm not so comfortable with. So I can totally understand the financial aspect of that too like now you have now you want to trust these hallucinating agents with your with your capital as well. Yeah. Yeah. And then there's there's also the privacy component, right? Well, yeah, that's that's kind of what I was getting at is like the privacy, right? It's, you know, I don't want to who knows what the corporations are going to do with with the information. That's right. But in this case, it's more about optimizing for the actual success of your desired financial outcome and having privacy and and trust, right? Yeah. Exactly. Yeah.

So when you're thinking about actually conducting let's call them verifiable or trusted agents you know what does that actually look like from a under the hood perspective right so you know I I go into the current iteration of Newton with a DCA agent I get to interact with that what does that mean that we're going to actually have this kind of like trusted agent like what is that like what is an untrusted agent perhaps or or you you know kind of how do we get to this concept of a trusted agent? Yeah. So um I I I like to I like to describe that in terms of analogies, right? So it's a simple DCA example, but it's like a whole iceberg underneath of like the infrastructure that we're using. An

analogy I like to use is um like a self-driving uh vehicle. Um so it goes from um you know when you when you go on a self-driving vehicle, you set a destination. Uh, you set like a speed. Um, what are those called? Whimo. Yeah, Whimo. They don't have those out here yet. I think that's just an SF, right? It's like a SF tou touristy thing. So, you set so you set where you want to go, your speed like parameters basically. Yeah, parameters. And then uh the car will follow through and then get you to from point A to point B.

Hopefully, right? Hopefully. Yeah. And um and then in the middle, right, you want um like this uh secure computer, right, that actually does um the task based on your parameters, right? And and no one should be able to temper with that like not even the mechanics, right? Um so so we actually under the hood, we use FA network, right? to run this confidential compute. That's where the agent runs and executes these transactions. And then that goes in the next step where everything is recorded, right? Um that's like the verifiable part. Every turn, every action that's recorded in like this almost something

like a flight black box, right? With like a notorized log where you can check if they did the correct actions afterwards. Um so we use um you know automa like contracts like verify that and then um lastly um you generate um a zero knowledge proof. Let's say if you want to share this um information with like your insurance company or like your friends without actually releasing all the personal uh and the travel information. Um it it it creates this like mathematical stamp, right? Um and then so that's um the zk proof part and underneath we use exoxinct and and risk zero. Um so yeah like it's it's sort of

this endto-end experience demonstrating like a DCA agent um with end toend like uh verifiability and guardrails. Okay. So that is for the interaction between myself as a human and and this agent. there's this kind of step-by-step process where basically using some some form of ZK and then tees with automata um you get this this this verifiable activity log trusted computer and then get you you get to execute all all those actions that's with me interacting with an agent. Now Accenture has recently put out up a prediction that they think that agents will interact with with more agents than humans in the in the near future. there's this kind of swarms idea

coming about where there's agents interacting with other agents. Um how does how extensible is this technology to to not just be around human to agent interaction but also with you know verifiable agents amongst others. Um, you know, I assume it in a future where you have agents that are going out and doing these multi-step processes or these kind of, you know, potentially something like like a vault strategy or kind of these more complex um processes will require multiple agents. So, how are you thinking about this this kind of combination of agent to agent interaction within the lens of this verifiable trusted nature? Mhm. So I think um yeah my my current take would be uh uh not every agent needs to be verifiable. I think only the uh agents

that's touching like sensitive data whether that's finance or healthcare or other kind of administrative stuff needs to be verifiable. Um, and then you could have like um like orchestrator agents, right? Like the agent that knows about all these other agents and route you to the right agent and then they all know about some other ones and and so like it just kind of cascades from there. But ultimately it's like it's like a searching problem, right? like you're you're starting from your the user intent and that translates into this one agent who knows about these other agent and and then so on and then only the and then the tasks that require

verifiability will be executed by those verifiable agents. Um and then um ultimately getting your result back, right? Um so so a lot of it is like a discoverability uh problem and and and just finding the right agent or tools for the job right and and we're also seeing a lot of development in um like standardizing like AI agent execution as well for example with MCPS um um so you know I think it's kind of like the chain abstraction of like AI execution um so That's um so yeah like there's new standards that's being created also that

could simplify uh this as well. Yeah. You know what's funny is I actually think of uh chain abstraction and intents today as an evolution of the you know poor UX that crypto has had which is now I found that these natural language implementations of interacting onchain seemed to be like the best form of like chain abstraction or even like application abstraction that we're seeing now. And so we kind of had this idea of chain abstraction. Um but now we have this idea of application abstraction which is like this this automation from it's almost like removing the user from having to do the manual processes of interacting with an application. I you know not even you

know forget about a chain we're just on applications now. Um and so even removing that process and so that's more on like this kind of automation side of things. Maybe you can walk us through what your mental model is of the evolution of this kind of UX abstraction if I kind of had the right idea here with this kind of first was chain abstraction application extra abstraction with these natural language models kind of how are you thinking about the the way to approach these this application abstraction and what does that really mean? Tired of multi-chain headaches? You need Omni. Omni makes being onchain feel like you're online.

Perform any action on any chain in less than two seconds. It's all powered by Solvernet. Omi's breakthrough SDK for crosschain speed, scale, and cost. To kickstart adoption, the Omni Foundation is offering $10 million in rewards for solver powered apps. Start earning today at omni.network. Yeah. So um I I saw this as in like three three phases right so when we first started it was wallet abstraction just making sure the wallet stays invisible under in the background almost like you're logging to a regular application. So um yeah that's with embedded wallet we've achieved that now it's sort of a status quo right in the

industry um for on boarding everyday users and then now we move um as the blockchain started proliferating and there's more fragmentation going on naturally chain abstraction is um is a super big problem and now there's a lot of really great projects trying to solve that so I feel like okay that's taking being taken care of like what's text, right? And and that's where the AI crypto intersection um really works together to provide a better user experience. I see maybe 80% of DAP interfaces in the future disappearing right into agents.

Um in fact, they may be easier to build by then. You know, you don't have to manage like all this crosschain stuff. um you don't have to build a UI you know which is takes a lot of effort especially for for crosschain um so I think there's it it only sort of accelerates the number of applications being built reduce cluttering reduce the amount of UI the users need to learn um and perhaps more reusable uh agents and functionalities too like once what one agent is live and running then anyone on can sort of call that agent, right, to to get something done. So there's more interoperability and reusability for the

fword. Um so so I imagine this future where yeah most of the DAP interfaces disappear. Um and there's this like Twitter quote by by Neville saying like AI is going to eat UI, right? And and so um so I feel like that not only is going to happen in web 3 but also web two in general. um starting with this and and definitely finance is a big aspect of it. Um and right now like commerce is all the conversation too, right? It's like just having agent buy stuff. Um so like yeah, we're like at the precipice of of the of of that transition. Um and um and it's a transition of like user behavior, right?

From just using applications, mobile apps to to using a chat. Um it's a transition uh in terms of trust also from not trusting AI to trusting AI. Um so yeah, eventually um the traditional knowledge of what an application is is going to change, right? And and that's why um you know I like to call it like application abstraction right and and I think AI is also really great at like aggregating uh information both from like consuming the information but also executing the information. Um so users don't have to like manually curate information or

manually go to different dabs or protocols or chains to get something done. it can all be like streamlined massively both from like a read and write perspective. Yeah, I think that the trust in AI generally it reminds me of kind of social media. It reminds me of of blockchain and crypto as well. Probably has a an easier ride in terms of perception than than than crypto has in its in its run so far. But I think that I think that it's something that will like only just keep accelerating as as people get, you know, more used to using these programs in their day-to-day as as well as onchain. And I actually think that AI will eat UI is an

interesting idea because I feel like, you know, I have this idea and a lot of people have where it's build, they will come is is kind of dead. It's it's kind of over, right? Distribution is what matters in today's information age, social media age. And so we get to a point where, you know, you have these companies that are fighting to own these enduser relationships, right? There was this big idea around fat wallet thesis where those that that owned, you know, these these end users, they could charge fees, they could dictate order flow, that kind of thing. Now what you're saying is basically okay well if I can build a UI that is you know effectively a natural language model in LLM for humans to interact with in this chat format then I can then have this control

over the user relationship and the order flow. So would you say that this is like has this been part of your thought process as to how you you want to turn Newton into a more kind of like revenue generating uh front end to a a a owning these user relationships? like how are you thinking about the kind of business model and revenue around the order flow and the owner ownership of these enduser relationships that you will have if you can execute on making kind of Newton be this top of funnel for for web 3 for for crypto as a whole. Yeah, as of now we're very focused on what the protocol can enable. Okay. Right. And and one analogy that I like to make is like Farcaster

protocol and Wcast, right? Worcast is a great demonstration of the of the protocol. Uh it's a great experience for the end users uh as well, right? So um we see the same way with the Newton client which we uh announced with Newton.xyz um and then the Newton protocol infrastructure under underneath because it's such a new thing. Um if we don't really show what the end toend experience really look like, it'll be hard for people to really understand it. back to the, you know, beginning of like telling that story and helping people really see the vision, right? So, um, I think with web 3, helping people see the vision and resonate with it is super

important because it's not just up to one company to build it all, right? We're we're probably going to build some of the um, initial like native agents. Then we're going to have this agent marketplace that third party developers can plug in their agents. Um and then um in the future there could be different frontends and clients for the protocol as well, right? So so that is um that is the vision is to be pretty credibly neutral around the protocol by enabling more agent um use cases through um improving the improving trust, right?

while having our um Newton client be like that initial experience for the users and maybe we could use that to help some of the agent developers to get adoption as well like through our community. So um so yeah like it's we we we called it um our recent launch the black triangle moment. Um it's like a story with the PlayStation one. It's like when everything's hooked up from the hardware to software and to end is like that first demonstration but we are very much looking forward to evolve right with the community and ecosystem. Um you know basically we're getting a lot of feedback on what agent to build next what token assets to

support next. So um I see this as like a highly collaborative process with with the ecosystem but centered around around the protocol. Yeah. I mean when I think about what agents would be useful for me it'd be some sort of you know automations around I think strategies that that I wouldn't have previously known about something like DeFi strategies yield strategies that also something that I would identify consistent opportunities in in the market um you know with some sort of positive hit rate right obviously that's like hard to do probably but I I think I think that would be something that people would would want to see. I've seen some iterations of of this where,

you know, you can ask one of these LLMs with regards to like, okay, what is the uh you know, what are the best looking opportunities out there? I think that's cool. Um, you know, I think that also like tracking different onchain movements is cool and getting a bit of a rundown as to like what's happening. Um, you know, what's particularly interesting is that you can put these agents together into a composable manner. um you know and that's something that that I would like to see more of where there's these there's these agents that that I use and that I have ownership of theoretically or they are kind of my agents and then I can then go and interact with those you know within my company within you know friends the community and start to kind of see things happen together copy trading

social strategies group investments syndicating um lots of like interesting ideas that I think we we've yet to scratch the the surface of, but I think you guys are on to it where it's it's centered around trust. It's centered around the idea that I should actually be willing to kind of trust these these tools with my capital. Yeah, that's right. That's right. It's because it's not too difficult to build these capabilities, but it's difficult to get the trust across. And do you think users really care that much right now about that? I I think so. I think for more serious um scenarios like for like memecoin like more speculative use case maybe telegram bots. Well yeah so these telegram bots they've they've always

just been like seed phrase like either insert your seed phrase or deposit the funds and then use them. So that's like a little bit different than what we're talking about here. Right. Right. Yeah. And and and that's what like um web 3 is like more used to. Right. But for these like more speculative scenarios. But if you were really talking about getting everyday users in and feeling comfortable about this, that sort of trust is important or even for like more serious use cases beyond um the the the the scope that which is provided by Telegram bots, right? Um I think just not uh handing full custody of your private key and assets to like a unknown

uh agent or or bot in the background. it's already a miles ahead in terms of security. Um, think of it more like like Google ooth, right, where you're authorizing Google like some read and write access. And you would do the same for like agents and and sort of providing that guard rail uh in terms of what it can do, what kind of parameters and limit it can perform within a and making sure that the agents can't go uh beyond those boundaries. Um, and I think that's um really really important for like security and trust going forward.

So, how do you enforce guard rails then, right? Like how do you how do you make sure that agents aren't going rogue? Uh, is it is it through some sort of like cryptographic proving? Is it through some sort of like rule sets? like how are you how are you creating these like agents that are bound within a set of parameters and are like are you know theoretically impossible to to do something that's not within these parameters. Yeah. So so to start uh this is sort of an innovation that came with smart accounts. Um so um we work with our partners uh Bonomy and and Rhinestone to create this smart account infrastructure underneath which allows you the users create these things called

session keys. Um and then the session keys are bound with like a specific set of param parameters like max m maximum of x token that you can spend and then um because everything there is onchain um the the um the agents can only perform those actions that are specified right uh by the permission and the session key. Um, so that's what it is like today with what with with what we just launched. But in the future, we're looking to uh allow more expressive type of uh uh permissions by leveraging like real time or oracleized data. So for

example, you can make it so that the agent can only trade if the sentiment is bullish, right? Um uh which kind of involves off-chain data kind of like chain link style, right? So um so I think that's going to unlock even more use cases and and that's going to um involve a lot of the more like deeper tech such as like ZK circuits and and all of that. So we're going to be leveraging more of like ZK technology there. So it's imperative then that that this kind of trust layer is built within the system from the start rather than adding it on later because you want to create these guard rails that effectively enable agents to to operate within a set of constraints proactively

um and then not having you know not having rogue um you know agents. So, do do you not think that these agents are just going to be able to outsmart your session key or eventually be able to just operate outside of the guardrails like like isn't the whole point to have these super intelligent, you know, almost human brainesque things? So, are they not just going to be able to reprogram around what you are what you've built for them? I think I think there's like an amount of art in this, right? uh uh because if the permissions are too granular then it kind of defeats the purpose but you can't not have the guard rails right so um so I I I believe there's like a

middle point that can sort of emerge from like active usage on what is like the right set of permissions so I think that's going to be really interesting to like discover in this process um but you know it's something that like in the light paper I call scoped autonomy, right? It's like you want to give the agent some autonomy to operate but within certain scope. So maybe it's something like you can only allocate up to, I don't know, $1,000 worth in total USD a day. Maybe it's as simple as that.

Maybe it's something that's a bit more um a bit more like detailed, right? So, so yeah, I think I think it's going to be really interesting to see that we're going to sort of um iterate into the right uh settings, but starting really really simple and then adding more as we go. Okay, very cool. Now, kind of zooming out from Newton and from crypto AI and just more philosophically, what is your what are some of your uh kind of ethical or moral thoughts around the the recent developments of AI specifically kind of stemming from your open- source background? Um, you know, do do you think that that we are in a a peculiar

spot more broadly for companies OpenAI, Anthropic, Cursor, etc. who are building these kind of corporateowned um LLMs and and that aren't open sourced. Um you know how do you feel about the current progression of uh of this of the stage of AI today? Yeah. So so philosophically I think you know I have I have a lot of trust in the world to figure things out right like like for example maybe in the beginning is really is the fastest to make progress with more like centralization. So these companies are moving really quickly painting a really great picture of what AI could be for everyday users and then problems will

start emerging right but the world sort of responds to that right so now there's a lot of conversations about open source models there's a lot of conversations about setting guard rails um uh more transparency like um audit logs for these AI agents. So I think yeah the world is reacting to that especially as it enters the realm of more serious use cases like when an agent is touching money right or like I don't know healthcare or other type of sensitive uh tasks um or just even companies that are um using AI right uh uh like from like a corporate um uh perspective. So, so I

think um I think that's all that's like sort of the next iteration that wave is is coming and there's a lot of people that's like looking at it. So that's why like a lot of the privacy confidential compute like ZK like a lot of that is emerging um as like a response to address this and we're now getting so that's like on the uh verifiability side but even just on the model side there's a lot better like open source model now with deep se um so so even that side is being democratized um so I think maybe an ideal outcome would be um so there's going still going

to be like the centralized models that's like paytoplay um but the um the average is also good enough right like seven eight out of 10 right with the open source stuff um so so so yeah like um I I I think I think whenever there's a problem I I believe people have the ingenuity to really solve those and and that's sort of what we saw right before we really start unleashing the agents on just buying stuff for users without security on guard rail. Um we're sort of building that bold before that wave comes, right?

And and that's sort of um what's been really compelling for me about crypto and web 3 is this ability for this whole community to build something to um that ex that that precedes it time right by many many years. Um it is just like you know um DeFi and like stable coins and like RWAS and now it's like finally all starting to click together. Um, so these things take time, but the beauty of web 3 is that you get to build a lot of these before the actual need, urgent need comes. Yeah. Ironically, I feel like the crypto AI sector is like 6 to 12 months behind the AI sector in San

Francisco. I don't know if that's true or not, but some sometimes it feels like that. Um, but you know, for for me, we're we're we're trying to actively look at what's happening there and and stay really close to the scene. Yeah. Yeah. So, come coming back to Newton, you guys have recently just put out your Genesis launch, this dollar cost averaging agent. Um, you know, we spoke pre a bit earlier about what we would consider to be some of the next ideas. Um, you know, and you kind of mentioned that there's a lot of possibilities, right? It's just a matter of kind of doing it the right way. Um, what do you think is going to come in the upcoming pipeline for uh agents on Newton? what can what can the community expect to have in the coming months for you know

on the product side when it comes to agents? Yeah, so uh we got we got a lot of asks on uh Yoda optimizer. Okay. So we're um we're planning that out. We're figuring out what's the right scope for this. Um there's a lot of ask on supporting more assets just for for everything. Um there's been asks on um like risk assessment, you know, having AI do do that for the end user. So um there's there's a lot. So So for us is more like um making a decision on what we want to build natively and what we want to um open up, right, to third party developers. Um, and uh, yeah, like

personally I'm I'm I'm eager to get in this sort of hackathon mode just playing around showing demos on all these sort of capabilities and just get feedback from end users. But but yeah, just personally I would I would use the yield optimizer uh myself. I think that'd be pretty useful um for me. So in a lot of the times we're like building this for ourselves and our team as well. Yeah. Yeah, I think I think I think those would be interesting. I'm I'm curious to see how this plays out. Um, and how this space looks in about 12 months because I think that there's going to be rapid progress in the actual models and LLM capabilities. Um, and that paired with these kind of growth of the

institutional adoption of crypto, um, as well as this broader kind of bullish trend currently. Um, you know, it there there's a world where the adoption cycle of this, you know, of of this kind of bull run per se is is centered around adopting people into these natural language models and them touching the chain without ever actually like needing to touch the chain, which I think is like particularly it's compelling and it's also just a much better reality than what we've existed in in 2020 and 21 where it was really just Ethereum and and Salana came about and now we have a ton of fast cheap chain but still have this UX problem. It's like we don't really have a block space problem anymore, but we have this kind of UX problem still. And so I think like this

is coming out at the right time. It's just a matter of, you know, the question like will users want to use LLM and natural language models to interact on chain. Will they trust it? Um, you know, and can they be educated to trust it? Yeah. Um, and I think that's the question that we're trying to solve. Yeah. Yeah. Yeah. That's right. and and then it's just overall very very exciting time where um a lot of these other technologies are maturing too like stable coins getting more adoption RWA tokenized stocks right getting adoption so like crypto has been like this experimental like playground um that keeps maturing these

technologies right and and and imagine like you can then have these uh things tested in the crypto ecosystem apply to stable coins or RWA. So you can have like portfolio rebalancing of you know not only crypto but also real stocks and assets potentially too or using stable coins for commerce to um to to pay for things. I think agent experience will be really great at like buying a collection of things or like booking something that's like pretty complex like a weekend getaway, right? Um so so I think all of those will be enabled through all these rails and I feel like it's going to be faster and faster um all the stuff that

crypto industry has been working as a whole to really connecting that with the mainstream um uh users and use case around AI. Um, and I think personally I I think like a smart wallet experience plus stable coins will be much better and cheaper than like these virtual credit cards, right? That I had to keep track of for like a web two agent experience. So in many aspects um like web 2 AI is ahead of web 3 but also web 3 is ahead of web two in some aspects uh here and and I don't see those as very separate industries and and and I think it needs to be proactively trying to

figure out how the two can can can really consolidate and and I think this is sort of the opportunity to do that when all of these things are maturing at at the same time. Yeah. Yeah, I think the the the democratization of access to these different uh agents or you know if you guys are able to ship a yield optimizer and different trading strategies and others you know there starts to be a compelling case for um those with little to no financial market skills able to have a completely different advantage and situation in financial markets as a result of uh you know using these hyper intelligent agentic forces. um that and then even the non-financial use cases, right? I

think that that's something that you're passionate about, right? Um you know, kind of returning to the fundamentals of of crypto. And so I know you've been thinking a bit about that. I'm just curious if you have any any thoughts around these noninancialized use cases of, you know, agents in in blockchain or or in the space that you think are are particularly compelling as well. I think I think a big part of it um that's top of my head is uh private privacy and like confidential use cases because right now people are submitting just crazy amount of personal and intimate uh data to these AI and I think at that at

some point that's going to become a problem. So, so, so I feel like yeah, like that u anything that's like private um would be great and there's already a lot of like local um uh AI usage as well. Yeah. So, what does that mean? It means that you have this like your your own LLM or your own model running on your own hardware. Yeah. On your own computer, right? like um it just co coincidentally my um previous co-founders uh from the kitematic project we sold to Docker they started um Olama uh which is a way to run local LLMs like the Docker for LLMs and what's the advantage to running a local LLM? Um so it's all

private. Okay, you don't you're not you're not really sending these to like a server in you get access to the same models like DFC or any of these open source models, but then when you input your prompts, you're just storing that data on your own like device basically. Yeah. Yeah. Like the chat logs like on the device. Um but you get access to the same amount of intelligence per se. Uh so long as the computer can can run it, right? Right. Um, so there's like like some limitations based on your computers, but also it be it's pretty cool because let's say you're flying on the plane and there's no Wi-Fi. You can still access intelligence right on your computer. So it it's pretty neat. In the

future, maybe some kind of a hybrid model to that sounds like something that's going to get massively adopted. Yeah. Is is it highly expensive right now to do that or is it a lot of space or what's the restraint right now? uh it's like GPUs like the compute capabilities of of your computer. Um and and that's why I think maybe it would be like a hybrid type of model where you're running um some models and um uh execution environments in the cloud, right? And then you combine that with local. So maybe there's kind of a routing where for basic like I don't know things that correct my English uh it could be like a cheaper model on on the computer whereas for more advanced

like financial workloads that can be routed into this verifiable agent in the in the cloud and and so that brings me to the idea that there's going to be a future where people are going to be hoarding um hardware to be able to run local LMS and just compete with each other on on who has the most intelligence. Yeah. Yeah. Because right now it's just kind of like this bifurcation of if you have capital and if you have some drive, these tools can help you a lot. Um but if you don't have capital to be able to to pay for the um the compute basically, then uh you know, you're at a disadvantage and then that combined with drive. It even seems like the local version is like you you're getting even more, you know,

privacy and ability without actually having to use, you know, some of these major companies basically. Right. Right. And then some of the TE stuff, it's it's in a way it's like kind of similar to a local environment, right? Where it's like a TE with a GPU and your model sort of everything runs there. Okay. In in that. So the model runs in the TE. Yeah. the G. Yeah, it uses the GPU T. Um, so so I think yeah, it's it's almost like a cloud equivalent where you run locally.

Yeah. Yeah. Very cool. Well, Sean, what is your upcoming vision for the coming months ahead for uh Newton? Um, what do you plan to achieve? Um, and then it's about time to wrap. Yeah. Uh I would say uh we're we're uh we've been receiving like um an overwhelming amount of enthusiasm about what we just uh announced and we uh our team is actually getting together this week to uh to take all the feedback and like prioritize and um you know we're already started um uh executing on some of the new agents um that that is that that is coming up and also they think about uh the

marketplace, what it means for developer um to plug into uh the pro the the uh the experience and getting other developers to participate. So, so there's a lot of like ideas, feedback. We're like sort of consolidating, building a plan and um and and continue to execute. But yeah, we intend to keep up the momentum and um and um yeah and and keep showing and demonstrating our vision with uh with what we're building. Awesome. Sean, thanks for your time, man. It's been a pleasure. Likewise.