--- title: "Why OpenLedger Is Betting on Specialized AI" canonical: https://wavect.io/podcast/ram-kumar-openledger-specialized-ai/ language: en description: "Ram Kumar explains why OpenLedger is building for specialized AI and how it aims to trace and reward contributions to model training. Full transcript." image: "https://wavect.io/img/yt/agbNQV6Oczo.jpg?v=3c5b42e12618" --- Podcast Transcript # Why OpenLedger Is Betting on Specialized AI Ram Kumar Core Contributor, OpenLedger Ram Kumar and Kevin Riedl discuss when a specialized model makes sense and how OpenLedger proposes to trace the influence of contributed data on model outputs. Kumar explains that model developers decide how to reward contributors. - Published 16 Jun 2025 - Runtime 38:01 - Operator - Recorded in English ![Why OpenLedger Is Betting on Specialized AI: podcast episode with Ram Kumar, Core Contributor, OpenLedger](/img/yt/agbNQV6Oczo.jpg?v=3c5b42e12618) // The short version - Kumar argues that specialized models make sense when a task needs data that a general-purpose model lacks. If a general-purpose model has the same specialized data, he says it is good enough, though a smaller specialized model may cost less to host. - OpenLedger proposes to trace how contributed data influences model outputs and provide a mechanism for rewarding contributors. Model developers decide how to reward them. ## Introduction [Watch this part · 0:00](https://www.youtube.com/watch?v=agbNQV6Oczo&t=0s) Kevin [0:00](https://www.youtube.com/watch?v=agbNQV6Oczo&t=0s) Hi everyone. Today I'm going to talk with Ram Kumar, our core contributor at OpenLedger, and we're going to talk about specialized AI on-chain and why [unclear] is wrong right now. So this is really a super exciting conversation because I really try to go into technical details, and of course it's still general enough so that non-technical people understand it as well. Kevin [0:23](https://www.youtube.com/watch?v=agbNQV6Oczo&t=23s) But it's a really engaging and cool discussion, asking some more critical questions as well, so that I personally really get a real understanding of what they're building and what problems they're solving. And personally, I learned a lot. I really liked it. So I'm pretty sure you will enjoy the conversation as well. [0:42](https://www.youtube.com/watch?v=agbNQV6Oczo&t=42s) Wavect, the Web3 software company that understands what you want. Kevin [0:48](https://www.youtube.com/watch?v=agbNQV6Oczo&t=48s) Uh, we got ChatGPT, Claude, and all these general-purpose models out there, and of course everyone knows about them. Everyone is using them. At least it seems like that. And while everyone is using these general-purpose models, you guys are actually focusing on specialized models, right? So could you put us through, like, why you think this is the actual future and not like this big, uh, all-in-one, uh, model? ## When specialized AI makes sense [Watch this part · 0:53](https://www.youtube.com/watch?v=agbNQV6Oczo&t=53s) Ram Kumar [1:21](https://www.youtube.com/watch?v=agbNQV6Oczo&t=81s) Yeah, just for clarity, uh, we use larger language models and make them specialized. Um, what we believe, um, LLMs will continue to exist. There will be more models that will come in. They will have much more knowledge. But all of this will be, like, very general knowledge, right? Uh, because that's what they're intended to be. Ram Kumar [1:41](https://www.youtube.com/watch?v=agbNQV6Oczo&t=101s) Like how Google searches, Google Search is a global search. It has general knowledge about the world. Whatever that you can find on internet is on Google. But we also use, you know, LinkedIn to connect with professionals, right? You search about a job on LinkedIn and then you would probably search, you know, search for food on Uber Eats, right? Ram Kumar [2:00](https://www.youtube.com/watch?v=agbNQV6Oczo&t=120s) You would search for a particular item that you want to buy on Amazon. So it's very specialized-driven, very niche searches that evolve from what is Google Search, right? So that's how larger language models would be. Larger language models are very generic. It's for the generic day-to-day use. But for specialized models, very unique use cases, you need to have specialized, like, you know, kind of data sets being given to that because let's say someone wants to build a healthcare model which is very specific to know more about general medicine. Ram Kumar [2:32](https://www.youtube.com/watch?v=agbNQV6Oczo&t=152s) You can't have a general knowledge model be used for that, right? It has to be trained on medical-based data sets. It has to have a knowledge about how a human body itself works. Only then this model is useful to provide solutions for any kind of ailment, right? Very similarly, if someone wants to build a model or an app for, like, finance, right? Ram Kumar [2:52](https://www.youtube.com/watch?v=agbNQV6Oczo&t=172s) An AI app that's going to let you, like, basically guide you to trade. It's going to be a copilot for you to trade. It needs to understand finance as good as a trader, a hedge fund manager or something like that, right? It has to have the knowledge about trading. Uh, a general knowledge, general model will have the knowledge for that. Ram Kumar [3:10](https://www.youtube.com/watch?v=agbNQV6Oczo&t=190s) It's going to have some understanding of it, but it's very basic, right? It'll have how a human kind of understands by looking at, you know, blogs and websites and stuff like that. It's not going to have an expert knowledge how a trader who has an experience of 10 years in trading will have. So that's a specialized model. Ram Kumar [3:28](https://www.youtube.com/watch?v=agbNQV6Oczo&t=208s) Uh, it'll have much more deeper understanding about a particular subject. And it's very natural for technology to be, you know, first generalized, to have a very generic, high-level knowledge, and then humans, us, go ahead and build much more specific, need-based tech. And that's how AI will evolve as well, where we now started with general knowledge, where people are getting to get familiar with AI, and then we would start using AI in our own day-to-day use cases, right? Ram Kumar [3:56](https://www.youtube.com/watch?v=agbNQV6Oczo&t=236s) If you take a look at models that are used in autonomous vehicles, these are vision models but are specially trained for that particular subject, right? Um, vision models that are also trained for, let's say, robots to do a particular work, like Figure and all that. There are vision models that are, you know, doing more on the domestic-based robotics and stuff like that. Ram Kumar [4:17](https://www.youtube.com/watch?v=agbNQV6Oczo&t=257s) So every model is specialized to its own need. So that's how, you know, our day-to-day life will be solved by AI. People will start implementing AI in all the interactions that we have, and you need models specific for those interactions. And we've seen that already happen in the last two years, right? Ram Kumar [4:34](https://www.youtube.com/watch?v=agbNQV6Oczo&t=274s) Once GPT was announced, people started using ChatGPT, then they start to experiment and build models which are specialized. Bloomberg launched BloombergGPT. It's one of the most special, you know, models out there, which has knowledge about finance. It's much better than ChatGPT in most of the cases. Ram Kumar [4:52](https://www.youtube.com/watch?v=agbNQV6Oczo&t=292s) So that's our thesis. And, um, that's what we see happening in the world as well. Hugging Face provides solutions where people can contribute data and build models, but none of them get rewarded as part of the process, right? It's an open-source solution. Open source is great, but it's not a long-term viable process. Ram Kumar [5:14](https://www.youtube.com/watch?v=agbNQV6Oczo&t=314s) You need to enable, monitor, monetization in open source, right? People need to be benefited as part of the process. That's what OpenLedger is trying to bring here. We want to make sure that, uh, you know, people who are contributing data sets, who are enabling these models to be built, get rewarded as part of the process. Ram Kumar [5:30](https://www.youtube.com/watch?v=agbNQV6Oczo&t=330s) Only then you will have a continuous flow of data coming in. Only then you will have a better quality of data sets that will be coming in. And that is what we're looking to achieve here. And we've seen that happen with most of the cases during our testnet. We have close to about 1 million people contributing data sets for that. Ram Kumar [5:46](https://www.youtube.com/watch?v=agbNQV6Oczo&t=346s) Um, we have, we did an experiment where a language-based model was needed to be built. Um, Hugging Face, you know, tried to do that by collecting language-based data sets. Uh, it took them about three months to get about 50,000 participants. Uh, we got close to about 300,000 plus participants in just about two weeks' time, right? Ram Kumar [6:06](https://www.youtube.com/watch?v=agbNQV6Oczo&t=366s) That's what happens when you monetize and you reward people, right? When you, then there's incentivization, people will go forward and execute something because everyone needs to feel that, you know, if they're providing something, they need to be benefited as part of the process, right? So that's the simple core thesis, to enable people to trust and work together, uh, on a platform and get rewarded as part of the process. Ram Kumar [6:28](https://www.youtube.com/watch?v=agbNQV6Oczo&t=388s) That's what OpenLedger is looking to solve here. Kevin [6:30](https://www.youtube.com/watch?v=agbNQV6Oczo&t=390s) Okay, I got a couple of questions here. Um, the first one, let's imagine we have two companies, right? One is building a general-purpose and one is, uh, building a, a specialized model. Um, let's say they both have the same specialized data set. [6:49](https://www.youtube.com/watch?v=agbNQV6Oczo&t=409s) Mhm. Kevin [6:52](https://www.youtube.com/watch?v=agbNQV6Oczo&t=412s) Like how, like from a data perspective itself, the data would be the same, right? The only difference is that the, uh, general-purpose model also has other knowledge, uh, that is not related to that specialized task. Like, how... [7:07](https://www.youtube.com/watch?v=agbNQV6Oczo&t=427s) Not... Kevin [7:08](https://www.youtube.com/watch?v=agbNQV6Oczo&t=428s) Okay. Yeah, sorry. Go ahead. Ram Kumar [7:10](https://www.youtube.com/watch?v=agbNQV6Oczo&t=430s) Not exactly, right? So, for example, um, usually, like, um, a model which is built for math will probably have most of the, of the theorems and understanding. Um, like a general model will have that knowledge, right? But let's say you want to build a special model for math. It's much more granular, right? Ram Kumar [7:32](https://www.youtube.com/watch?v=agbNQV6Oczo&t=452s) It's much more deeper into each and every subject. So it can solve a math problem, a much more complex math problem, quicker and accurately than a general model, right? Because it's just given more data sets. Like, think of this, right? To put in a very simple terms, think of, um, an undergrad or a, you know, high school kid to be like a general model which has, like, a generic knowledge of the world, and then a PhD student to be like a specialized model. Ram Kumar [8:02](https://www.youtube.com/watch?v=agbNQV6Oczo&t=482s) So they just have a better knowledge on the subject. Everyone, like, the subject is known. People can understand the subject, but a PhD student will have a better understanding and much more faster response and a much more accurate response. That's how specialized models work and that's how they compare to, like, general models. Ram Kumar [8:19](https://www.youtube.com/watch?v=agbNQV6Oczo&t=499s) And there are certain use cases where the data set itself is not available on the general model. Like, for example, uh, we have a team, um, that we're working with that was building, like, an X-ray model, a model that is trained on X-ray across the globe, right? Um, this is across races, across various diseases and stuff like that. Ram Kumar [8:40](https://www.youtube.com/watch?v=agbNQV6Oczo&t=520s) Um, this model identifies and understands X-ray-based details much better and has much clarity than a general model. If you upload an X-ray to ChatGPT, yes, it will kind of understand, will be able to tell you, but it's not going to give anything accurate. It'll not going to give you answers at depth of what a specialized model will do. Ram Kumar [9:01](https://www.youtube.com/watch?v=agbNQV6Oczo&t=541s) So in that case, the, these data sets are not available at all in general model. Only in the specialized X-ray model you have X-ray-based data sets. So it's, it's more case to case, uh. Certain data sets are there in general models. Other data sets are only available in specialized models. Um, it could be very case to case. Kevin [9:21](https://www.youtube.com/watch?v=agbNQV6Oczo&t=561s) But, uh, like, again to the base assumption, right? Like, let's say both companies have access to the same data, like really the, the deep specialized data. So why would a specialized, uh, model that has less knowledge in general, but the same specialized knowledge as the general-purpose model with that specialized knowledge, why would that be better? Ram Kumar [9:48](https://www.youtube.com/watch?v=agbNQV6Oczo&t=588s) In that case, I'm not sure if that's going to be like a usual way people build models, right? Uh, it doesn't happen, but let's say hypothetically this happens. Uh, general models have those specialized data sets. General model is good enough for that. Only way where they would might choose for a specialized model is if this model is, you know, lesser in terms of model size and you don't need to have a large hosting cost and stuff like that. Ram Kumar [10:15](https://www.youtube.com/watch?v=agbNQV6Oczo&t=615s) Apart from that, they would choose, like, a general model. But most of the cases, which is in reality, general models, general model, models do not have, uh, you know, specialized data sets, data sets which are very unique. Um, that's usually is with specializ... That's why they call specialized model. Ram Kumar [10:32](https://www.youtube.com/watch?v=agbNQV6Oczo&t=632s) You take something and you specialize it because that data is not there in the general model. Kevin [10:36](https://www.youtube.com/watch?v=agbNQV6Oczo&t=636s) Okay, so for, so the actual reason for specialized models, aside from hosting costs and so on, of course, is, uh, the organizational additional constraint, right, that you don't share highly specialized data with a general-purpose company like OpenAI. So that's actually because, yeah, okay, cool. Um, you also mentioned already your Proof of Attribution, uh, algorithm, consensus protocol, um, whatever you want to call it, right? ## How Proof of Attribution is meant to work [Watch this part · 10:56](https://www.youtube.com/watch?v=agbNQV6Oczo&t=656s) Kevin [11:05](https://www.youtube.com/watch?v=agbNQV6Oczo&t=665s) Could you go a little bit deeper on the technical aspect, like how, how does this actually work? Like, let's say I do have data that I want to share with your blockchain, right? With your infrastructure, how does that work from A to C? Ram Kumar [11:22](https://www.youtube.com/watch?v=agbNQV6Oczo&t=682s) Yeah, so to just go back to basics, right? Um, in any AI model that's being built, like let's say we use ChatGPT, there is OpenAI, which is the model that is running behind that. Think of that to be like a brain that powers this application. So in that model, people are basically are, uh, you know, connecting data sets across the globe, basically internet data, and training this model, right? Ram Kumar [11:46](https://www.youtube.com/watch?v=agbNQV6Oczo&t=706s) Um, all of this internet data is very generic, as we discussed, right? It has a very general knowledge. Specialized models means we need to get these data sets from people who are, like, you know, skill set, skilled people, or, like, you know, enterprises and smaller companies, which have, like, very special data set. Ram Kumar [12:06](https://www.youtube.com/watch?v=agbNQV6Oczo&t=726s) We can, we would love to have them connect and provide these [unclear]. When they do that, right, the best way to enable that to happen is to incentivize them. And the second thing is to make sure that you give them confidence that they can trust the system, right? How do you give that confidence? You provide them with a protocol or something, which is a general ledger, where you can go and watch how your data is being stored. Ram Kumar [12:32](https://www.youtube.com/watch?v=agbNQV6Oczo&t=752s) You can see that you have ownership. Your data, your data is pegged to your wallet. And when a model gets trained, your, if your data sets is used, you can visibly see which model is consuming your data set. It's on-chain, it's traceable on-chain. And when this model gets used, uh, you can see whether your data was useful in that model's output. Ram Kumar [12:56](https://www.youtube.com/watch?v=agbNQV6Oczo&t=776s) And you could also look at, like, what was the payment that came in to that particular model? So we have it all on-chain, right? Someone contributes data, it goes, goes on-chain. So you can see the ownership. The data gets used by a model developer, that is traced on-chain as well. And when the model gets launched, if the model is used, every time a model throws an output, we show whether a data set, let's say 300 that was shared, that 300th data set, you can see whether that was causing any influence of that model, right, and that output. Ram Kumar [13:28](https://www.youtube.com/watch?v=agbNQV6Oczo&t=808s) If your data set has caused an impact on that output of a model, we attribute and showcase how much of your data set was impact, like, what was the impact that was created? Let's say out of the influence you had, like, 30%, 40% of your data set to be providing impact and, like, providing the solution, then you get rewarded accordingly. Ram Kumar [13:49](https://www.youtube.com/watch?v=agbNQV6Oczo&t=829s) Again, this is on-chain. The model developer decides how much of his revenue he wants to share with the data contributor. He shows that on-chain. That's how he runs this contract. And then you can see it on-chain, on how much impact that you've created and how much of the revenue you're gonna get shared for. Ram Kumar [14:06](https://www.youtube.com/watch?v=agbNQV6Oczo&t=846s) And you can trace it back and see that you're getting paid. You'll be able to see this entirely on-chain. You'll be able to see the entire provenance on-chain. So that is what we are trying to solve, and that's the Proof of Attribution protocol, where we attribute to data that is being contributed by a data contributor to a model, and then we show it on-chain so they can verify and they can basically get to know why they're getting paid. Ram Kumar [14:31](https://www.youtube.com/watch?v=agbNQV6Oczo&t=871s) That enables them the trust to contribute more and more. If you, if you take a look at it, right, the reason why we, you know, let's say we do podcasts and, like, you know, we upload videos to YouTube and all that, there is some sort of monetization, right? Um, people get rewarded, get paid, and they're following a passion. Ram Kumar [14:51](https://www.youtube.com/watch?v=agbNQV6Oczo&t=891s) They want to do it more because they're getting recognized. Very similarly, we want to enable data contributors to get recognized and get paid for it. But in a platform like YouTube, um, the biggest issue over there is it's not open. It's a centralized system. You need to believe whatever they pay you, right? Ram Kumar [15:07](https://www.youtube.com/watch?v=agbNQV6Oczo&t=907s) You might be creating a much bigger impact. You might be having much larger views. You might be having much more reward than they actually pay for you. But you accept that because, you know, that's the norm and that's the largest player out there. It's centralized. They've taken the market. So there is no other option to fight back, right? Ram Kumar [15:26](https://www.youtube.com/watch?v=agbNQV6Oczo&t=926s) AI could soon become that, right? All the AI model companies can just go ahead and collect a data set and never reward you or pay very minimal rewards if they are pushed to do that. And we all have to accept that. There could be potential ways to solve that if you do it at the initial stages and if you can actually capture a good market out of this. Ram Kumar [15:46](https://www.youtube.com/watch?v=agbNQV6Oczo&t=946s) That is what OpenLedger is trying to solve. We want to make sure that this is going to become a problem later on. There is going to be a lot of needs later on. [unclear] almost started. So we're building a platform where people can contribute data, monetize it, and they can know visibly how much of the data is useful. Ram Kumar [16:03](https://www.youtube.com/watch?v=agbNQV6Oczo&t=963s) And they can see it on-chain. It's transparent and it's traceable. And so that, this, this is a much more open system than something like Hugging Face, which is a closed and not, not a monetizable system there. We want to change that and that's, that's what we're trying to do. Kevin [16:20](https://www.youtube.com/watch?v=agbNQV6Oczo&t=980s) That, uh, also seems to plug into a little bit of, uh... Ram Kumar [16:23](https://www.youtube.com/watch?v=agbNQV6Oczo&t=983s) Hey guys, uh, this is Ram. I'm one of the core contributors at OpenLedger. OpenLedger is a blockchain for AI. We enable users to contribute data sets and build AI models and have apps use these AI models and pass on reward back to the data contributors. Our idea is to enable anyone who contributes data sets to building an AI model get rewarded for the contributions that they do. Ram Kumar [16:46](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1006s) Typically, internet models that you see today, larger language models like OpenAI, Llama and all that, use your data that has been consumed from the internet. You don't get rewarded for that. But the idea behind this is that AI is not only just there. It's going to evolve more than that. AI is going to be much more specialized and the data is among the people, enterprises and businesses. Ram Kumar [17:07](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1027s) We want to make sure all of them get rewarded as part of this process. And thanks a lot for having me here. I'm glad to be here and talk more about OpenLedger. Kevin [17:15](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1035s) All that notion around black box, explainable AI and all these things. Um, like, how do you see that from an enterprise perspective, right? Because oftentimes that's a huge criticism people had or still have towards AI models, that you basically don't know how it came to that output, how it came to that conclusion, right? ## What enterprises want to trace [Watch this part · 17:16](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1036s) Kevin [17:39](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1059s) And thinking is kind of a progression here, but it also is an output, right? So we don't really know. Ram Kumar [17:45](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1065s) Yeah, so if you take a look at all these models today, right, they're not explainable. You don't know how they're working because there is no, uh, clarity on what kind of data sets is being used. Like, even let's take a look at an open-source model like Llama. There's no visibility on major of these models that they announced. Ram Kumar [18:05](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1085s) There's no visibility on what kind of data sets that they used. So if Llama model is used by an enterprise to make its decision, let's say by a hospital or an e-commerce company, or let's say a, a lab is going to use that, there is no explainability in that. There is, you got, if you want to get, you got an answer and you've implemented that as part of your business, you [unclear] go back and see which data caused that answer by the AI and why you used it, right? Ram Kumar [18:31](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1111s) That is what we would like to change using specialized models. At least on specialized models we know what data set is being contributed, even though we don't know what's happening on the base model. What we're trying to figure out is that as part of the attribution, the, you know, the byproduct of this, or the, you know, what happens out of this is that every time attribution happens, we're able to reward people. Ram Kumar [18:55](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1135s) We also figure out what data is causing the output to come, right? So through that, we are making the model much more explainable through data. Every time an output comes in, you'll be able to go back and see what data set caused this output. And using that, you can get more of the data if the output is good. Ram Kumar [19:15](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1155s) Or if the output is bad, you feel it's malicious and, you know, the AI is compromised, then you can go ahead and remove those data, which can be, like, potentially poisonous data sets. So through that, we would love to explain how an AI works using data, right? And enterprises need that. Like, for example, we work with a large e-commerce company, uh, which is using their internal data sets to build a model with us. Ram Kumar [19:39](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1179s) And one thing they really wanted was to have an end-to-end trace. So assuming they develop, uh, using this AI within their platform and they manage their inventory based off this, they manage the SKUs based on this, then in the end of the day, your audit, right, they would want to go back and see what are the decisions they made and why they made that. Ram Kumar [20:01](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1201s) And it's going to be relied on an AI. They want to make sure that they have an end-to-end trace on what data was feeded, how it is processed by the AI, and what was the output and what really impact it created, right? They want to have an end-to-end, uh, auditability of that. And that's why they want to work with us, because we are able to provide them with this explainable model. Ram Kumar [20:18](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1218s) Uh, you don't get that with ChatGPT. You don't get that with Llama, right? You put in a bunch of data sets, you get an answer, you have to trust that it's, it's the right answer that is there, and you just have to use this. So that is what we would love to change using explainability. And that is important, right? Ram Kumar [20:34](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1234s) You make a medical decision based on AI and you diagnose someone. If you don't know why you're doing that, then that's risky, right? So, um, everyone knows this. Very few people are working on that and we are one of them. Kevin [20:48](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1248s) It also sounds like that, like, explainability of AI also plugs into a little bit of the discussion of bias within AI models, right? That you're basically, when you actually can trace back the decision or the outputs of a model, you actually should also be able to, like, when you identify bias, that you also trace back which data set or which bias in which specialized data set actually caused that. ## Tracing bias back to training data [Watch this part · 21:06](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1266s) Kevin [21:17](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1277s) Is that also something you work on or is that just a byproduct? Ram Kumar [21:21](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1281s) Yeah, again, that's, that's part of the explainability, right? Uh, so when, when we provide data, like these models, uh, to people and they're able to use a blockchain to trace and figure out, like, how data is influencing, then if, there are two ways, uh, they see biases. If they want to remove the biases, they can. Ram Kumar [21:40](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1300s) In certain cases, you might need more biases, right? Let's say you're building a specialized model only for America, right? And it's about healthcare. You would want very biased data set of people, majority of races that is living in America, right? So in that case, you would need biases to be fed into the model. Ram Kumar [21:59](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1319s) And let's say it's a general model and biases should not be there. It should be fair and it should not have any kind of orientation to a particular race or a subject or a matter. Then you can use explainability to remove those orientations. So, uh, again, it's again a product, uh, that can be, that's a byproduct, or sort of like what, what are the use cases that can be done using explainability? Ram Kumar [22:24](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1344s) Removing biases is one of the biggest use cases. Um, one way that we're trying to figure out this is that, like, first we are working more on, um, specialized model layer, right? On the specialized model, we're now figuring out how data is, you know, impacting the, the output. Next, we want to go much more context driven, right? Ram Kumar [22:43](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1363s) You know data, you know the output that comes in. We want to figure out the context the AI model derives out of this, because you would have learned something in your life, right? I would have learned pretty much the same thing in my life. But what we derive and tell out of that is two different things. Ram Kumar [23:02](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1382s) Your context of the world is different. My context of the world is different. Similarly with AI as well, right? An AI model has different context based on what's been feeded to it and how it kind of understands, let's say, what has been feeded to it. Getting to know that context is very important, right? Ram Kumar [23:21](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1401s) Um, let's say we ask about who is, um, Alexander, right? Um, it'll tell you who is Alexander, but actually the way it tells, right? Alexander is a king who's known for his, um, you know, his strength, and he's known for his courageousness and how he fought, um, all these years and stuff like that. When it tells that, it's because it read about various rulers and various kings and stuff like that, and that's how it's able to actually get a context and give you an answer, right? Ram Kumar [23:55](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1435s) So understanding that context is also very important to actually know how an AI model works and how it actually gives you an answer. So that's very important. That's sort of like the next steps on what we do. So there is a lot more depth into explainability and how data is impacting a model. ## Retrieval and fine-tuning [Watch this part · 24:13](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1453s) Kevin [24:13](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1453s) Um, you just barely scratched the surface here. I love it. Cool. I mean, the whole purpose of this podcast is to scratch the surface, right? To make it more accessible to people. Um, but maybe let's go one layer below the surface. Um, from a technical perspective, um, attributing data sets, like how the output is being derived. Kevin [24:38](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1478s) Um, like usually from, from a technical standpoint, you, you have these layers within your AI model, you have weights, and that's basically how an output is derived, right? It's a probabilistic model most of the time, right? At least for my, uh, basic knowledge here. And, like, is it from a technical perspective you just feed, uh, that technical knowledge or specialized knowledge in real time when a query is being made, like within, in that context? Kevin [25:10](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1510s) Or is it actually something that you train your weights on? Like, how does that work? Ram Kumar [25:15](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1515s) That's a, that's a good question. There are two ways to do this, right? Usually when people want to, um, customize a model [unclear] to their own requirements or specialize it, there are multiple ways you can do. Best two ways to do is to provide, like, a RAG or fine-tune the model. Typically people go with RAG. Ram Kumar [25:37](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1537s) Um, you know, they provide, like, context where, uh, you know, RAG is more on, like, giving a better prompt. You provide, like, additional data sets and you ask a question and then it goes and uses your data set that you added as a RAG and then it gives you an answer. So now it has a context of what you're asking for. Ram Kumar [25:57](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1557s) That's the probably the best way to go forward and that's, um, not a, I wouldn't say best way, but, like, the quickest way to go forward, right? So we provide RAG solutions as well. So not only, like, model fine-tuning, not just enabling people to build specialized models, all the data sets that are being, uh, contributed, uh, to OpenLedger, they can be used as a RAG as well. Ram Kumar [26:18](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1578s) So that's the first layer. You contribute a data set, it can be just used raw as a RAG. Next, if that doesn't solve the purpose, what you do is basically you fine-tune the model. You take the data sets, you train it, superimpose the data set into a model where you post-train the model with more data set, and then you make a specialized model. Ram Kumar [26:38](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1598s) Now you have a fine-tuned version of the model with more data, which can be used to get your answer. So these are the two ways that we try to approach this. Um, I think this has been the most successful ways that we've seen, and most of the people who have been building in our testnet have tried both of them and based on the need and case in case they try to, uh, you know, change. Kevin [27:00](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1620s) Mhm. I mean, I definitely can see, like, how you attribute, uh, like how you trace back, like, which data sets have been used using RAGs, right? But how does that work if they actually fine-tune a model? Ram Kumar [27:16](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1636s) Yeah, so RAG is simple, as you know that. So someone provides additional data, you can go back and trace it. Uh, how, kind of like on a fine-tuning layer, right? That's, that's the biggest challenge. And how did we solve this? Based on this paper. We've announced about this before and we're going to announce it in a white paper as well. Ram Kumar [27:36](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1656s) Uh, we built this based on a paper called DataInf. So what it does is basically, uh, it figures out the influence of a data in a fine-tuned model, right? Um, so what we do is, like, we kind of use data's, like, basically loss function. We figure out, we remove all the data sets and figure out which data had the biggest loss. Ram Kumar [27:58](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1678s) It means that it was the biggest impact-creating data set, right? Um, like basically, like I, you give an answer and, uh, what if I reverse engineer, say this, and remove all of your knowledge, and how much of your answer goes bad? The biggest knowledge that I remove, the biggest impact it could have created in your answer, right? Ram Kumar [28:26](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1706s) Like, which data had the biggest impact on the loss, and then that is what the data we attribute and showcase. Um, there's much more, like, details into it, um, much more technical-related details which you can see it on the white paper on how we do it. The biggest challenge that we had was that it took us almost, like for one inference, it took us more than 15, 20 minutes on an average time to figure out, right? Ram Kumar [28:51](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1731s) We had to bring it down to a couple of seconds. That's what we worked on in the last year or so. Because think of the compute cost it would cost us if we actually do it for a heavy inference, it would cost, like, 15, 20 minutes, right? So we had to bring that down. And that's, that's the work that we did over the last one year, majorly on the attribution side. ## Batching attribution after inference [Watch this part · 29:00](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1740s) Kevin [29:11](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1751s) Okay, that's very interesting. I got two questions. I hope, uh, I don't lose them. Um, what, the first question is, so it's basically you're, it's, it's kind of a brute force approach, right? You pull away one data set, you see what's the loss and so on, and you peel the onion, basically, right, from my understanding. Kevin [29:28](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1768s) Um, like, I can see, like, even if it's just a few seconds, right, on one hand it's definitely, I guess, expensive if you do that on a regular basis, so it might need to be, I guess, on demand. But, um, like, how does that work if you have, let's say, 50 specialized data sets contributing to that? Ram Kumar [29:53](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1793s) Yeah. Yeah. So what we do is that we batch this, right? We don't do it, uh, on real time because if, uh, if I'm going to reward people, I don't have to pay you, like, every second, right? You don't mind getting paid every, like, let's say a day or two. So the best way to do this is to batch, uh, all the inferences together and then, you know, figure out how the attribution happen. Ram Kumar [30:17](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1817s) So, um, that's what we've done. We basically optimized. Um, we don't need, uh, you know, speed in this. We just need accuracy, um, but accuracy over, uh, you know, a varied point of contact, right? So that's, that's what we've done. We, we don't do it at real time. We do it, um, at a batch stages, so the GPU cost that we use is much, much more lower. Ram Kumar [30:42](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1842s) Um, not like how in real time I ask a question and you need quick inference, right? So over there you need to have, like, very highly scaled-up CPUs. In this case it's much lower, much simpler, and we have very optimized way of how we have done that. And, um, all that is part of the research that went through. Ram Kumar [31:00](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1860s) So yes, so since it's batched, the cost is much lower. Kevin [31:04](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1864s) Cool. And, uh, from an arch, architectural point of view, I, like, one thought I just had, and that could be a little bit naive, right? But I, I guess these technical models, uh, that, as these specialized models are not highly dynamic, right? Like, once you have a specialized model set up, you usually don't change it much. Kevin [31:30](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1890s) So why not having some kind of a validity proof or something like that, where you're basically doing that training, like maybe bringing, I guess, training on-chain? Like, I, I, I guess where my thoughts are going, right, that you basically can get that before and you save the costs, uh, afterwards. Like, what are your, why does, doesn't that work? Ram Kumar [31:53](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1913s) You mean, you mean on the attribution or on the, on... So, okay, so this is the thing, right? So when you ask a question, I wouldn't, we wouldn't know what kind of question you're asking, because every inference is very unique, right? So based on that, we, we would not be able to figure out before. We need to do it in, uh, post the inference time. Kevin [32:14](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1934s) Okay. And so, so also from a MVP, I, I know you're not in the MVP stage, right? But one could also in some sense solve this, I guess. May, maybe it's not as accurate, but which is a big, uh, big value proposition, of course. But one could also say, okay, these are the data sets and they have that importance, like a general importance weights in that specialized model. Kevin [32:39](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1959s) It would be independent. [32:43](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1963s) They could, like, Ram Kumar [32:44](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1964s) for example, uh, let's say we are building a model for healthcare. Um, we could have certain data sets that is the most common, uh, questions that are asked, and these data sets are impactful. Um, and I think, yes, we could figure that out. But that's not very, but even though we figure it out, it's not going to be fair if we reward them beforehand because, um, everyone is contributing data set. Ram Kumar [33:11](https://www.youtube.com/watch?v=agbNQV6Oczo&t=1991s) Um, maybe they prioritize better, right? That, again, we leave this to the owners of, uh, the model creator, right? We are an infrastructure platform where we provide the solution. Model developers use our platform to collect data sets and then they would decide how they want to reward the data contributors. Ram Kumar [33:31](https://www.youtube.com/watch?v=agbNQV6Oczo&t=2011s) If they want to reward these data contributors beforehand, we have mechanisms for that as well. Like, they could figure out, like, I would need these, these data sets and if someone who provides them the first, I'm going to reward them, let's say, potentially a percentage of the revenue I'm going to make out of this model, right? Ram Kumar [33:50](https://www.youtube.com/watch?v=agbNQV6Oczo&t=2030s) And beforehand or something like that. They could do that, but, um, it's up to them. We don't encourage that, but it's up to them. But typically what really happens is that people collect data sets and these data sets are, like, they know, they tell what kind of data set they need. Um, most of it is the most important data set, uh, that's needed, and they collect them. Ram Kumar [34:09](https://www.youtube.com/watch?v=agbNQV6Oczo&t=2049s) Uh, some of them might not be so useful, but they still collect them because they need to have the entire context. Model gets trained. Only when it gets used, they really actually figure out what is the market that is looking for, what are the queries that is being asked, and then they try to reward people accordingly. Ram Kumar [34:27](https://www.youtube.com/watch?v=agbNQV6Oczo&t=2067s) So, um, for example, we all would ask GPT about, let's say, our holiday plans, and most of the common questions has been, like, holiday plans, I read somewhere. So that could be the most useful data set, uh, in the last year or so that ChatGPT was trained on. But it doesn't mean that that's going to be the same, right? Ram Kumar [34:48](https://www.youtube.com/watch?v=agbNQV6Oczo&t=2088s) People will evolve from there. People will go much more, they will ask questions that are much more relevant as they, as they start to use ChatGPT. So, uh, I would, it wouldn't be fair to actually reward people based on what they think is the most important data. It's most, it should be fair as people use these models, then they reward, uh, the data contributors. Ram Kumar [35:08](https://www.youtube.com/watch?v=agbNQV6Oczo&t=2108s) That's, that's where the fairness comes in. ## Data ownership and closing thoughts [Watch this part · 35:09](https://www.youtube.com/watch?v=agbNQV6Oczo&t=2109s) Kevin [35:10](https://www.youtube.com/watch?v=agbNQV6Oczo&t=2110s) Love it. Super interesting. Thanks for sharing that. Um, well then I would say, are there maybe three key takeaways or three things you want people to remember from this call session, or maybe even, uh, three call to actions, like, uh, checking out, uh, your blockchain and, like, whatever, what should... Ram Kumar [35:29](https://www.youtube.com/watch?v=agbNQV6Oczo&t=2129s) Important aspect I would have to say is that I think we all should be aware that AI will be creating a big impact, um, you know, all our lives. I think the previous generation to us, and us also, so we choose, uh, convenience over privacy, right? Like, we didn't mind, like, larger corporations taking our data and, like, using it for various use cases. Ram Kumar [35:55](https://www.youtube.com/watch?v=agbNQV6Oczo&t=2155s) That's evolved over AI right now. But at least in AI, we have a choice to make where we can still own our data and we can decide what you want to do with that, right? We can create our own AI. Like, we can basically, um, be a part of the AI era that's about to happen. So people need to become much more aware, people need to become much more knowledgeable on how their data is useful. Ram Kumar [36:18](https://www.youtube.com/watch?v=agbNQV6Oczo&t=2178s) So that's one thing that I would love to tell people. And also learn about AI, right? Because it's all going to change our life. And then, um, you know, we are, we are still, um, in the stages of development right now. We are in testnet and we are going mainnet soon. Uh, we would love to have feedback from the community. Ram Kumar [36:35](https://www.youtube.com/watch?v=agbNQV6Oczo&t=2195s) We always encourage that. Um, we've gone and built this out over the last one and a half years by constant conversations with the AI community, talking to developers, talking to builders. We would love to have the users across the globe also could provide feedback on the product as well. That will help us build a better product. Ram Kumar [36:54](https://www.youtube.com/watch?v=agbNQV6Oczo&t=2214s) So that's two most important things. And third is that go check out OpenLedger HQ. We have people, uh, as I told you, like, building on us. We are going mainnet soon and we already have, like, 10 plus projects building, including the team which is from Google [unclear], who decided not to use any traditional AI companies. Ram Kumar [37:13](https://www.youtube.com/watch?v=agbNQV6Oczo&t=2233s) They wanted to work with us. We work with a lot of enterprises, a large e-commerce company, one of the largest ones. We work with them. We'll be revealing about this soon. We work with the government, the sovereign government based in the Middle East. We are building a model with them. We'll be announcing about that as well. Ram Kumar [37:29](https://www.youtube.com/watch?v=agbNQV6Oczo&t=2249s) So these are people who are trusting us. So we'd love to have you guys check that out as those news come and keep supporting. Kevin [37:37](https://www.youtube.com/watch?v=agbNQV6Oczo&t=2257s) Super exciting. Well, thank you for your time. It was really a pleasure talking with you. Learned a lot. Pretty sure the audience as well. And yeah, all the best with [unclear]. Ram Kumar [37:49](https://www.youtube.com/watch?v=agbNQV6Oczo&t=2269s) Thanks, Kevin. Uh, it was nice talking to you. Kevin [37:51](https://www.youtube.com/watch?v=agbNQV6Oczo&t=2271s) Likewise. [37:52](https://www.youtube.com/watch?v=agbNQV6Oczo&t=2272s) Have a good one. See you, man. Wavect, the Web3 software company that understands what you want. Transcribed from the original audio with Whisper large-v3 and AI-reconciled with the original automatic captions. Unclear speech is bracketed; speaker labels and translations are editorial aids. No complete human listening pass. The recording remains authoritative; statements reflect its date. The episode was published on June 16, 2025. At the time of the interview, Kumar described OpenLedger as being on testnet and said a mainnet launch was planned. The participation figures and attribution timings are Kumar's reports about the experiments and development work he discussed. His explainability proposal concerns contributed data; he says they still do not know what happens inside the base model. // Keep digging [AI development→](/services/artificial-intelligence/) [Blockchain development→](/services/blockchain/) ## Questions this episode answers ### Why batch attribution after inference? Kumar says batching lets OpenLedger group attribution work and reduces the cost compared with handling every inference separately. ## Structured Data ```json { "@context": "https://schema.org", "@graph": [ { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ], "employee": [ { "@id": "https://wavect.io/team/kevin-riedl/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Kevin Riedl", "url": "https://wavect.io/team/kevin-riedl/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } }, { "@id": "https://wavect.io/team/christof-jori/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Christof Jori", "url": "https://wavect.io/team/christof-jori/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } } ], "founder": [ { "@id": "https://wavect.io/team/kevin-riedl/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Kevin Riedl", "url": "https://wavect.io/team/kevin-riedl/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } }, { "@id": "https://wavect.io/team/christof-jori/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Christof Jori", "url": "https://wavect.io/team/christof-jori/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } } ], "legalRepresentative": [ { "@id": "https://wavect.io/team/kevin-riedl/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Kevin Riedl", "url": "https://wavect.io/team/kevin-riedl/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } }, { "@id": "https://wavect.io/team/christof-jori/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Christof Jori", "url": "https://wavect.io/team/christof-jori/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } } ], "name": "Wavect GmbH", "subjectOf": { "@id": "https://wavect.io/verified-claims.json#dataset", "@type": "Dataset", "creator": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] }, "description": "A machine-readable registry of quantitative and qualitative claims published by Wavect, with review dates, localized page appearances and public third-party citations where available.", "inLanguage": "en", "isAccessibleForFree": true, "license": "https://creativecommons.org/licenses/by/4.0/", "name": "Wavect verified publication claims", "url": "https://wavect.io/verified-claims.json" }, "url": "https://wavect.io/" }, { "@id": "https://wavect.io/team/kevin-riedl/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Kevin Riedl", "sameAs": [ "https://www.wikidata.org/wiki/Q139796365", "https://www.linkedin.com/in/wsdt", "https://github.com/wsdt" ], "url": "https://wavect.io/team/kevin-riedl/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } }, { "@id": "https://wavect.io/team/christof-jori/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Christof Jori", "sameAs": [ "https://www.wikidata.org/wiki/Q139796367", "https://www.linkedin.com/in/jocr77/", "https://github.com/jo-chris" ], "url": "https://wavect.io/team/christof-jori/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } }, { "@id": "https://wavect.io/#website", "@type": "WebSite", "inLanguage": [ "en", "de", "es", "zh" ], "name": "Wavect", "potentialAction": { "@type": "SearchAction", "query-input": "required name=search_term_string", "target": { "@type": "EntryPoint", "urlTemplate": "https://wavect.io/search/?q={search_term_string}" } }, "publisher": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] }, "url": "https://wavect.io/" } ] } ``` ```json { "@context": "https://schema.org", "@type": "FAQPage", "mainEntity": [ { "@type": "Question", "acceptedAnswer": { "@type": "Answer", "text": "Kumar says batching lets OpenLedger group attribution work and reduces the cost compared with handling every inference separately." }, "name": "Why batch attribution after inference?" } ], "speakable": { "@type": "SpeakableSpecification", "cssSelector": [ ".faq-question", ".faq-answer" ] } } ``` ```json { "@context": "https://schema.org", "@graph": [ { "@id": "https://wavect.io/podcast/ram-kumar-openledger-specialized-ai/#webpage", "@type": "WebPage", "dateModified": "2026-11-14", "description": "Ram Kumar explains why OpenLedger is building for specialized AI and how it aims to trace and reward contributions to model training. Full transcript.", "inLanguage": "en", "isPartOf": { "@id": "https://wavect.io/#website", "@type": "WebSite" }, "mainEntity": { "@id": "https://wavect.io/podcast/ram-kumar-openledger-specialized-ai/#video", "@type": "VideoObject" }, "name": "Why OpenLedger Is Betting on Specialized AI | Ram Kumar Transcript", "speakable": { "@type": "SpeakableSpecification", "cssSelector": [ ".pe-takeaways" ] }, "url": "https://wavect.io/podcast/ram-kumar-openledger-specialized-ai/" }, { "@id": "https://wavect.io/podcast/ram-kumar-openledger-specialized-ai/#video", "@type": "VideoObject", "about": { "@type": "Person", "jobTitle": "Core Contributor, OpenLedger", "name": "Ram Kumar", "sameAs": [ "https://x.com/Ramkumartweet" ], "url": "https://www.openledger.xyz/" }, "contributor": { "@id": "https://wavect.io/team/kevin-riedl/#person", "@type": "Person", "jobTitle": "Managing Director", "name": "Kevin Riedl", "sameAs": [ "https://github.com/wsdt", "https://www.wikidata.org/wiki/Q139796365" ], "url": "https://wavect.io/team/kevin-riedl/", "worksFor": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] } }, "description": "Ram Kumar and Kevin Riedl discuss when a specialized model makes sense and how OpenLedger proposes to trace the influence of contributed data on model outputs. Kumar explains that model developers decide how to reward contributors.", "duration": "PT38M1S", "embedUrl": "https://www.youtube-nocookie.com/embed/agbNQV6Oczo", "hasPart": [ { "@type": "Clip", "endOffset": 53, "name": "Introduction", "startOffset": 0, "url": "https://www.youtube.com/watch?v=agbNQV6Oczo&t=0" }, { "@type": "Clip", "endOffset": 656, "name": "When specialized AI makes sense", "startOffset": 53, "url": "https://www.youtube.com/watch?v=agbNQV6Oczo&t=53" }, { "@type": "Clip", "endOffset": 1036, "name": "How Proof of Attribution is meant to work", "startOffset": 656, "url": "https://www.youtube.com/watch?v=agbNQV6Oczo&t=656" }, { "@type": "Clip", "endOffset": 1266, "name": "What enterprises want to trace", "startOffset": 1036, "url": "https://www.youtube.com/watch?v=agbNQV6Oczo&t=1036" }, { "@type": "Clip", "endOffset": 1453, "name": "Tracing bias back to training data", "startOffset": 1266, "url": "https://www.youtube.com/watch?v=agbNQV6Oczo&t=1266" }, { "@type": "Clip", "endOffset": 1740, "name": "Retrieval and fine-tuning", "startOffset": 1453, "url": "https://www.youtube.com/watch?v=agbNQV6Oczo&t=1453" }, { "@type": "Clip", "endOffset": 2109, "name": "Batching attribution after inference", "startOffset": 1740, "url": "https://www.youtube.com/watch?v=agbNQV6Oczo&t=1740" }, { "@type": "Clip", "endOffset": 2281, "name": "Data ownership and closing thoughts", "startOffset": 2109, "url": "https://www.youtube.com/watch?v=agbNQV6Oczo&t=2109" } ], "inLanguage": "en", "name": "Why OpenLedger Is Betting on Specialized AI | Ram Kumar", "publisher": { "@id": "https://wavect.io/#organization", "@type": [ "Organization", "ProfessionalService", "LocalBusiness" ] }, "thumbnailUrl": "https://wavect.io/img/yt/agbNQV6Oczo.jpg?v=3c5b42e12618", "transcript": "Kevin: Hi everyone. Today I'm going to talk with Ram Kumar, our core contributor at OpenLedger, and we're going to talk about specialized AI on-chain and why [unclear] is wrong right now. So this is really a super exciting conversation because I really try to go into technical details, and of course it's still general enough so that non-technical people understand it as well.\nKevin: But it's a really engaging and cool discussion, asking some more critical questions as well, so that I personally really get a real understanding of what they're building and what problems they're solving. And personally, I learned a lot. I really liked it. So I'm pretty sure you will enjoy the conversation as well.\nWavect, the Web3 software company that understands what you want.\nKevin: Uh, we got ChatGPT, Claude, and all these general-purpose models out there, and of course everyone knows about them. Everyone is using them. At least it seems like that. And while everyone is using these general-purpose models, you guys are actually focusing on specialized models, right? So could you put us through, like, why you think this is the actual future and not like this big, uh, all-in-one, uh, model?\nRam Kumar: Yeah, just for clarity, uh, we use larger language models and make them specialized. Um, what we believe, um, LLMs will continue to exist. There will be more models that will come in. They will have much more knowledge. But all of this will be, like, very general knowledge, right? Uh, because that's what they're intended to be.\nRam Kumar: Like how Google searches, Google Search is a global search. It has general knowledge about the world. Whatever that you can find on internet is on Google. But we also use, you know, LinkedIn to connect with professionals, right? You search about a job on LinkedIn and then you would probably search, you know, search for food on Uber Eats, right?\nRam Kumar: You would search for a particular item that you want to buy on Amazon. So it's very specialized-driven, very niche searches that evolve from what is Google Search, right? So that's how larger language models would be. Larger language models are very generic. It's for the generic day-to-day use. But for specialized models, very unique use cases, you need to have specialized, like, you know, kind of data sets being given to that because let's say someone wants to build a healthcare model which is very specific to know more about general medicine.\nRam Kumar: You can't have a general knowledge model be used for that, right? It has to be trained on medical-based data sets. It has to have a knowledge about how a human body itself works. Only then this model is useful to provide solutions for any kind of ailment, right? Very similarly, if someone wants to build a model or an app for, like, finance, right?\nRam Kumar: An AI app that's going to let you, like, basically guide you to trade. It's going to be a copilot for you to trade. It needs to understand finance as good as a trader, a hedge fund manager or something like that, right? It has to have the knowledge about trading. Uh, a general knowledge, general model will have the knowledge for that.\nRam Kumar: It's going to have some understanding of it, but it's very basic, right? It'll have how a human kind of understands by looking at, you know, blogs and websites and stuff like that. It's not going to have an expert knowledge how a trader who has an experience of 10 years in trading will have. So that's a specialized model.\nRam Kumar: Uh, it'll have much more deeper understanding about a particular subject. And it's very natural for technology to be, you know, first generalized, to have a very generic, high-level knowledge, and then humans, us, go ahead and build much more specific, need-based tech. And that's how AI will evolve as well, where we now started with general knowledge, where people are getting to get familiar with AI, and then we would start using AI in our own day-to-day use cases, right?\nRam Kumar: If you take a look at models that are used in autonomous vehicles, these are vision models but are specially trained for that particular subject, right? Um, vision models that are also trained for, let's say, robots to do a particular work, like Figure and all that. There are vision models that are, you know, doing more on the domestic-based robotics and stuff like that.\nRam Kumar: So every model is specialized to its own need. So that's how, you know, our day-to-day life will be solved by AI. People will start implementing AI in all the interactions that we have, and you need models specific for those interactions. And we've seen that already happen in the last two years, right?\nRam Kumar: Once GPT was announced, people started using ChatGPT, then they start to experiment and build models which are specialized. Bloomberg launched BloombergGPT. It's one of the most special, you know, models out there, which has knowledge about finance. It's much better than ChatGPT in most of the cases.\nRam Kumar: So that's our thesis. And, um, that's what we see happening in the world as well. Hugging Face provides solutions where people can contribute data and build models, but none of them get rewarded as part of the process, right? It's an open-source solution. Open source is great, but it's not a long-term viable process.\nRam Kumar: You need to enable, monitor, monetization in open source, right? People need to be benefited as part of the process. That's what OpenLedger is trying to bring here. We want to make sure that, uh, you know, people who are contributing data sets, who are enabling these models to be built, get rewarded as part of the process.\nRam Kumar: Only then you will have a continuous flow of data coming in. Only then you will have a better quality of data sets that will be coming in. And that is what we're looking to achieve here. And we've seen that happen with most of the cases during our testnet. We have close to about 1 million people contributing data sets for that.\nRam Kumar: Um, we have, we did an experiment where a language-based model was needed to be built. Um, Hugging Face, you know, tried to do that by collecting language-based data sets. Uh, it took them about three months to get about 50,000 participants. Uh, we got close to about 300,000 plus participants in just about two weeks' time, right?\nRam Kumar: That's what happens when you monetize and you reward people, right? When you, then there's incentivization, people will go forward and execute something because everyone needs to feel that, you know, if they're providing something, they need to be benefited as part of the process, right? So that's the simple core thesis, to enable people to trust and work together, uh, on a platform and get rewarded as part of the process.\nRam Kumar: That's what OpenLedger is looking to solve here.\nKevin: Okay, I got a couple of questions here. Um, the first one, let's imagine we have two companies, right? One is building a general-purpose and one is, uh, building a, a specialized model. Um, let's say they both have the same specialized data set.\nMhm.\nKevin: Like how, like from a data perspective itself, the data would be the same, right? The only difference is that the, uh, general-purpose model also has other knowledge, uh, that is not related to that specialized task. Like, how...\nNot...\nKevin: Okay. Yeah, sorry. Go ahead.\nRam Kumar: Not exactly, right? So, for example, um, usually, like, um, a model which is built for math will probably have most of the, of the theorems and understanding. Um, like a general model will have that knowledge, right? But let's say you want to build a special model for math. It's much more granular, right?\nRam Kumar: It's much more deeper into each and every subject. So it can solve a math problem, a much more complex math problem, quicker and accurately than a general model, right? Because it's just given more data sets. Like, think of this, right? To put in a very simple terms, think of, um, an undergrad or a, you know, high school kid to be like a general model which has, like, a generic knowledge of the world, and then a PhD student to be like a specialized model.\nRam Kumar: So they just have a better knowledge on the subject. Everyone, like, the subject is known. People can understand the subject, but a PhD student will have a better understanding and much more faster response and a much more accurate response. That's how specialized models work and that's how they compare to, like, general models.\nRam Kumar: And there are certain use cases where the data set itself is not available on the general model. Like, for example, uh, we have a team, um, that we're working with that was building, like, an X-ray model, a model that is trained on X-ray across the globe, right? Um, this is across races, across various diseases and stuff like that.\nRam Kumar: Um, this model identifies and understands X-ray-based details much better and has much clarity than a general model. If you upload an X-ray to ChatGPT, yes, it will kind of understand, will be able to tell you, but it's not going to give anything accurate. It'll not going to give you answers at depth of what a specialized model will do.\nRam Kumar: So in that case, the, these data sets are not available at all in general model. Only in the specialized X-ray model you have X-ray-based data sets. So it's, it's more case to case, uh. Certain data sets are there in general models. Other data sets are only available in specialized models. Um, it could be very case to case.\nKevin: But, uh, like, again to the base assumption, right? Like, let's say both companies have access to the same data, like really the, the deep specialized data. So why would a specialized, uh, model that has less knowledge in general, but the same specialized knowledge as the general-purpose model with that specialized knowledge, why would that be better?\nRam Kumar: In that case, I'm not sure if that's going to be like a usual way people build", "uploadDate": "2025-06-16T12:11:42+00:00", "url": "https://www.youtube.com/watch?v=agbNQV6Oczo" }, { "@id": "https://wavect.io/podcast/ram-kumar-openledger-specialized-ai/#episode", "@type": "PodcastEpisode", "associatedMedia": { "@id": "https://wavect.io/podcast/ram-kumar-openledger-specialized-ai/#video", "@type": "VideoObject" }, "dateModified": "2026-11-14", "datePublished": "2025-06-16", "inLanguage": "en", "name": "Why OpenLedger Is Betting on Specialized AI | Ram Kumar", "partOfSeries": { "@id": "https://wavect.io/podcast/#series", "@type": "PodcastSeries" }, "timeRequired": "PT38M1S", "url": "https://wavect.io/podcast/ram-kumar-openledger-specialized-ai/" }, { "@type": "BreadcrumbList", "itemListElement": [ { "@type": "ListItem", "item": "https://wavect.io/", "name": "Home", "position": 1 }, { "@type": "ListItem", "item": "https://wavect.io/podcast/", "name": "Podcast", "position": 2 }, { "@type": "ListItem", "item": "https://wavect.io/podcast/ram-kumar-openledger-specialized-ai/", "name": "Ram Kumar", "position": 3 } ] } ] } ```