Decoding AGI, Specialized Intelligence, and Agent Infrastructure with Ram Kumar

Transcript

How humans need the brain to go ahead and tell you to do actions, right? Agents also need a brain to go ahead and do some actions. If agents are humans, then what what is open ledger? What value does open ledger provide? Let's say tomorrow you go ahead and launch an agent and uses a particular model, something goes wrong, the agent goes rogue and it does something bad. Welcome back to the rollup, your home for technical talks and open conversations that shape the modular future. Nothing on today's show is to be construed as financial advice. You can see our full disclosures list at therollup.co/ventures. Sit back, relax, and enjoy today's episode after a quick word from some of our awesome partners.

Ladies, gentlemen, anons, and dgens, it is stable coin season brought to you by your favorite, the best in stable coins, Fracks USD dollar stable coin, backed by the best in the business, the Bidd Fund by Black Rockck and Superstate. This is stable coin season. It is the yieldbearing stable coin by Fracks, coming soon to a chain near you. All right, Ram, we're in Hong Kong. Yeah. My last podcast of the day has been a marathon. It's been a fun week. Awesome.

Um, lot going on here. Yeah. Yeah. There's a lot of stuff. Hong Kong is always busy. Yeah. How's how's life for you here? You've been traveling, right? Yeah. It's my second week or my second time here. Oh, wow. Uh, so it's it's it's just very interesting to me the difference between you go to like East Denver and conferences in the west and the the like the culture, the types of projects, the conversations are just different than they are here. Yeah. I think that this is more like go to market like you know what's your product going to be how the end consumer is going to use your product yes how would retail kind of relate to your product right yes it's less like technical research like west is more research yeah longterm partnerships yeah 100% I

totally agree well the the AI and crypto space has one one cycle behind it I think it's Yeah, it's kind of it it but I think what's next is where we're going to see some of the more legitimate and kind of venturebacked and longer term projects emerge one of which being open ledger that you guys are are building. So I want to talk about that today. I want to talk about the kind of state of AI and crypto. What's what happened the last couple months. So just to kind of lay the groundwork, uh we had that we had the initial launch of truth terminal on Twitter which kind of came out as like this whole new meta for uh AI or Aentic crypto interactions, Fartcoin and Eliza

A16Z and all these kind of different uh projects and and we kind of we had that cycle. So may maybe you just kind of speedrun us to where we are today. Yeah. uh following the last 3 or 4 months kind of what happened from your perspective um and what was the core infrastructure and core components that were built and that will be left standing following kind of the the downfall of price action which is not always related to fundamentals. Yeah. So if you kind of like take a look at it right this happened in the traditional space as well when open AI was launching chat GPT there's a huge rage we've never seen something like this you could chat to so to an AI and then get answers and people like going crazy they were asking like random questions and then you saw all of

these chatyp rappers that came out of that like bunch of random stuff where you can generate an image for a particular thought like a topic and stuff like that you could probably ask questions about legal you can ask questions about your person stuff and then AI would answer. All these were like simple rappers on chat GPT, right? I I very similarly see this uh you know AI agents to be like that to be like chat GPT rappers because it is what it is, right? It's it's an open AI model in the back end with some system prompts and a theme around a particular meme and it talks about that. Um most of them were speculative but some of them like really good have some kind of use cases.

Um and unfortunately or fortunately in crypto people like to play around with tokens, right? Since all of these agents came with tokens, they kind of like took retail to like in a frenzy kind of because the tech was cool. It was interesting. You could talk to it. You would see a Twitter agent like interacting with you on your comments. So people like going crazy about it and they wanted to like speculate this token because they thought okay this is going to become big because it has like a lot of eyeballs. there's a lot of attention to this and that kind of like made people launch random agent tokens and there was too much speculations that was going on. We've all seen this. We had I think from right about like August to like December we saw random kind of AI agents that came along. 15,000 launch on

virtuals 15,000. Wow. Which is tokens as well. But how many actually are like really standing out, right? Like very few of them. Virtuals as a platform is great, right? Eliza as a framework is like really great but like people could not like I would say like take it to the better front like I we've saw we've seen this happen in the traditional space but people evolved they built much better apps like perplexity I would say is a great app which is still uses open AI in the back end and probably use other models as they kind of evolved from there but it's much better version of a chat GPD wrapper right that should have happened with agents but again in crypto So a lot of stuff is initially

speculative. We've seen this happen with ICOs. We've seen this happen with DeFi. Uh I think that's what happened with AI agencies. Yeah. And a lot of fluff kind of like really went down now after you know Trump token came in and all the things that happened. Uh but again this will come back. I think the meta will come back again. AI is definitely here to stay because there's a lot of stuff happening in the traditional speak in mutitional space that's going to kind of creep in back to web 3. uh I'm sure that there will be agents that will be coming back which has much better utility uh much more purpose than the current agents that is there and I think the infra way will also come back for AI right because better infra obviously means better agents better products and

that that's what I expect so that's what I see will evolve from here so maybe you could outline what the the sector really is so we have agents we have frameworks and there's also um some sort of middle ground here where it's uh like so I know we have frameworks to build which are kind of like akin to an L1 at least that's how like people have been kind of making a mental model around it and then we have agents which are like almost like applications like what is the what is the like what is going on in this sector here or we we have like what is what are the key categories of this like AI sector currently in crypto? Yeah. So like if you take a look at AI like there are three major components in terms of

infrastructure it's compute there is algorithm which is like which is basically the model and then there is data right so this is the core infrastructure what's built on top of that is the app layer which is the agents right and the frameworks are kind of like in the middle which support these agents to be launched right you need a no code platform people to go ahead and launch these agents right I would say frameworks are those no code platform that enables this to happen. What like kind of people were experimenting with the frameworks is more on how can we use the existing models that are there the existing solutions that have been already provided like open AI lama and all that and like how can we tweak that to build

our own use case right it was more of experiments right like AI XPD is a good experiment to showcase to talk about what's happening in the industry to talk about what's happening like onchain and offchain sends and all that we don't know what happens behind the terminal. I don't know I don't know if anyone actually access the terminal but um like I think if you get to access the terminal and the terminal actually lets you do something that's the real utility that is where I would like people to go and explore. Um so that's you know that's what we want to see happening right we want to see agents which can let you do something like agents are basically a bloated version of apps right when you interact with an app you

get to an endpoint you get something out of that I'm able to order e-commerce I'm able to you know go ahead and book my ticket and do all that agents could do that for you agents are just smarter versions of apps right but the infrastructure currently in Web 3 at least doesn't support that right need to have better models. You need to have better data flows for these agents to be smarter. And uh there are very few firms that are working on that like we are one of them where we want to support by providing models which are specialized for a particular need. And it's like think of agents to be like humans. Okay.

Like how humans need the brain to go ahead and tell you to do actions. Right? agents also need a vein to go ahead and do some actions and usually the the current agents uses a brain which is very generic which is an open am it's very generic this tells you it has information about your brain you know it's like they just stamp boom stamp boom stamp boom stamp boom right yeah it's not customized to your need it's no you can't use like your butter knife to have your stick right you need to have something customized for a particular need right if you want to build an agent for trading then you probably use a model which is better off for trading right which is specialized for that and that's what we want to enable to happen

like we want to make people build models for their own need like let's say I want to build a model around you know to find the best taking options that is there right to find the better yields for myself and I want to build an agent around that people can go ahead and use opener collect the data sets around that build a model for that particular need and then launch that so that any agent can use this model and make better decisions with that. So that's the ultimate goal, right? How can we build like verticalized AI models on top of us? An AI for trading, an AI for healthcare, an AI for let's say being a co-pilot for solidity and all that. How can people

use our infrastructure to build that is what we are like really looking for. I think that is what is going to push the next narrative of AI and AI agents and all that that was going to happen. This is a good segue into kind of what's coming next then, right? So, we kind of establish what happened the last couple months. I think we get an the next wave of iterations likely coming. You'll be more sophisticated than this first wave as the market is more mature and kind of aware of what what has happened. Um, and I think a trend that you and I are both kind of pointing towards here is the specialization of agents. to rather going from a general framework of an agent who can do everything to or like you'll have a specific categories or sectors or uh tasks which agents will be

done and then personalized ones. personalized, meaning I have mine. Yeah. Which I train and I I do this already, right? I've got a Claude account, a Claude teams account for the rollup. I've got my project. Robbie's got his. Everybody on our team has our own projects. And then we have shared projects that we train with all of our own data and in inputs. And so I've got one cloud uh project that I speak to daily. Yeah. So it's be it's like it's used to me now. It's personalized to me, right, in that capacity. Yeah. Um so like what so yeah that's what we're going so what you've basically done is like there are multiple ways to make a model much more usable for a particular need so you know you could start with

like RLF basically where you already have a model you have an agent for that model you give constant feedback to this model by telling okay this is what I am this is what I do and you get all of this information it kind of starts storing this personalized version of Andy it gets to know you and whatever questions you ask, it's going to have this in this memory and it's going to answer you, right? That's the memory of an agent. And then it goes one step more than that. When you actually build an agent for yourself, you could do like a system prompt, right? And then there is much more. You could use rag, which is sort of like a temporary database that you can use which the agent can refer, the model can refer every time it needs to give an answer. And the much more

deeper version of that is actually where you fine-tune and build a model for yourself where you specialize, right? But specializing in building your own models are for larger like verticals. It would not be for let's say a model for rollup. No, it would be a sort of like a model for the entire let's say web 3 marketing or web three right like kind of it requires more data or it's more expensive to create uh it takes a lot of time because you need to collect a lot of data right uh the finetuning process on on let's say a platform like us is it takes probably few hours because it's a no code platform. You can go ahead and build a model in like a few hours. You can just fine-tune them within any existing model that is there. But collecting the data is what matters,

right? It's going to take a lot of time for you to collect the data. And then once you train a model, once you build a model, then you got to see whether the model answers the way you want. So there's a lot of iterations that goes in, right? So people do not go to like fine-tuning as the first step. They would probably start with sort of like system prompts. That's what most people do and they go for rag and then they go for fine-tuning. Or there's like recently there has been like tool calling and all that right where there's an API request that agent takes in and then figures out the answer and then comes back to you right like we kind of support all these three you can provide like a rag or you could build like models on top of this or you can use any

of the models that is there in our in our platform and use tool calling to go ahead and get your answer 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. Redstone is DeFi's modular oracle. Supporting over 100 clients on over 50 chains and offering 1,000 assets. Redstone is the leader for Bitcoin staking and yieldbearing assets. They offer the most robust customizable price feeds available in any model. Join the Redstone exhibition today. [Music] Okay. So, so you guys are fully optimized for a future that's more specialized and more personalized. Well, just specialized. I I would say start with like specialized. Okay. And then

probably evolved to more personalized. What's the difference? is if it's specialized for me it's is it's a synonym for personalized right I would say personaliz is more like individualistic right I would want to I would take a specializ and maybe personalize it much more for my need but specializ is much more broader okay right you guys are specialized for web three podcast right so um how can we build a model which is focused about like web3 marketing and onchain model and stuff like that because if you go ahead and ask like chat GPT or claude or any of these uh agents that are there and like query like give me the best ways to market a project on web 3, right? It's going to give you like a

very generic answer. It's probably going to look at hundreds of blogs that has been trained on and then probably search the internet and to give you some random answers logs. It's not going to be like really specialized. It not going to be very specialized. It's going to hand this snake like hell. So the idea behind this is to not make sure that that doesn't happen to build much more purpose-driven solutions. Okay. Uh this has this has been the thing with tech. If you take a look at tech right tech has been always generalized when it launches because it's it's an innovation and everyone wants to try to use that and they feel that it's not very efficient and they go much more vertical and then build like specialized versions of that. We've seen this with apps.

We've seen this with web solutions and all that. It's going to happen with AI as well. it's already happening in the traditional space. We we bringing that to F3. Got it. So on a more philosophical level when when we're thinking about agent agentic interactions in the future, especially like from a perspective of uh commerce, right? Um for example, we have a agent coming out on a rollup. Uh it'll be a it'll be able to speak on the podcast and uh ingest information real time. Mhm. They probably won't be that good at that, but it'll be able to also uh edit some videos, do some clips. It'll have a Twitter account. Yeah, it'll do these types of things. So, when I'm thinking

about how how we scheduled this podcast, there's a calendarly link. Yeah. There's a there's a coordination that we're both going to be in Hong Kong. Yeah. That we're going to do business. That we're going to do a podcast. Yeah. So like when I'm thinking about this rollup agent that we have that also can interface with Telegram, like is there not a reality where this rollup agent can effectively manage our schedule or can communicate with your agent or other agents? Like what I'm getting at here is like I would like to enter a world where not only can I interact with my agent directly from a a prompting to an information perspective, but I would like to be able to delegate to then

interact with another agent who then is having its tasks delegated from you. Yeah. So it's like you and I can communicate but like almost happen we can get things done without us, you know, like Yeah, that's what I'm getting at. Yeah. So um so think of agents to be like humans like Yeah, that's what you said. Yeah. Yeah. So I always like compare this to like what happened with Tesla. So when uh Tesla initially came about, it kind of behaved like a bot. So you go through a lane, it's going to follow the lanes that is there and then it's just going to drive a straight road, right? And then they kind of brought in the you know sort of like an autonomous driving in the freeway because there was more, you know,

understanding how the freeway works. It was easier. And then now it goes from a point A to point B without any human intervention. And it once it reaches it blocks itself. Right? So now it's behaving like a human, right? A human doesn't need to be sitting on that driver seat, but we have to because of regulations, but it's driving itself, right? Think of agents to be like that. It's another form of human being. It's a digital form of human being. So I guess that's how you need to see it. You need to train your agent to behave like human, right? like a human where the agents of us could talk to each other and then help us out. I think that's going to happen. It's I don't think it's going to be like years and years away.

No, I think it's like this year. But I think it's a good mental model to think about agents as as humans. So if agents are humans then what what is open ledger what value does open ledger provide to to these humans to these agents? Um I more on the intellectual layer, right? Like how can we make these agents much more smarter? M how can we make these agents to do something much better than what they're doing right now? Because as I told you that is very much needed. If I'm going to use agents for my own needs, if I'm going to make this much more purposeful than sort of like a gimmick, then I would need models, an intelligent layer that's going to help me do that, right? So, because like if I'm going to have an assistant for me to

let's say help me manage my finances, he should be probably trained on finances, right? He should be at least an NBing for me to actually make sure that I make the right decisions and help me out in that. So, why would you use someone who's an undergrad or someone who's like a, you know, like a, you know, 10th grader or something like that to actually do your math? You would need someone who's just much more specialized. very similar logic. Uh models which are much more specialized for a particular task are much precise and less hallucinating compared to the generic models that are there. Think of generic models to be like like an operating system and then specialized models are like like much more

fine-tuned versions of the maybe an apps or something like that on top of that and so that um they deliver much better uh you know outputs compared to the generic models that are there. That's what open edger wants to build. Then we want to build a framework for this to happen. There is like lot of moving parts to this right. Um someone has to contribute the data and uh so we built a framework where people can contribute data sets. Uh they could write a rule engine so that data set contribution is much more efficient and then there is a no code platform where the model gets trained. So we've built that out. Got it. And then there is an attribution layer where if these models are built and hosted on our platform and an agent uses that we know which data was causing

this output to come model on every inference and then we reward back the data contributors based on every inference. Right? So that the revenue that we get from these agents for these models is distributed equally to these data contributors because if you take a look in a foundation model there is compute there is data there is models there's a lot of factors that is coming in but for a specialized model which is just a layer above the foundation model like I'm not going to build a foundation model on my one that's not going to happen in open ledger right no users will go ahead and take any foundation model train that to their need they will fine-tune that to their need if they're doing that the biggest factor over there is the data contributor the data that is being contributed. So they have to be

equally rewarded as the model builder. So the attribution layer enables you to do that. We attribute back to see which data causes output to come. We then reward them. Another important aspect of that is the explanability, right? Let's say tomorrow you go ahead and launch an agent and uses a particular model. Something goes wrong, the agent goes rogue and it does something bad. You might want to go back and figure out which data actually what's happened. Hey, it might be a data which is poisoned. You want to go back and see that. That's where the attribution layer helps you do that. So that's an integral piece of what we do. So then zooming out from AI as we kind of come to a close any any alpha on open ledger key takeaways from Hong Kong for those who

who couldn't have been here um and what's coming next. So a lot of stuff uh so we are in our test net right now. Testing of the product uh letting community access our product as I told you there are three layers. Right now we've opened up the data layer where people have contributed data to us. We've close to more than 1 million users who have like onboarded and contributed data. We're opening up the model layer right now so people can go ahead and access our platform and build like their own models experiment. It's in the test net obviously. So we want people to like play around. The third will be the the agent integration layers where we integrated to a lot of frameworks. So you can go ahead and like launch agents using our platform. uh we would we don't want to be in the frame agent framework

lay a lot of players over there we would be the infrastructure that supports and makes these agents smarter so we've integrated with various partners so they could use that so th those are the updates that is coming out and hopefully a mainet soon so that's what we're working on uh so yeah so that's the update and Hong Kong has been quite busy with the events and all that happened there and next next is probably Denver all right thank you so much man you