Ram Kumar I The Future of AI is Specialized, Why General Models Aren't Enough I Pragma Denver 2025

[Music] and for this talk I want to bring on Rah from open ledger to tell you about how to think about the future of AI models and how generalist they can and cannot be so without further Ado please welcome r on stage hey guys uh this is Rah from open Ledger I'm one of the core contributors here uh today I'm going to talk about why the feature of AI is specialized and not uh General models right how General models are not good enough AI is definitely changing the way we work we all know that AI is pretty much there in every part of our day-to-day lives right from scheduling our meetings to taking notes for us to giving us reports of our everyday life I think AI has been predominantly catching up to what we do it's going to be much more bigger it's going to help us ride the cab it's going to help us take care of our health we're going to see more forms of AI in the physical form as well but are we are we there yet uh be3 has been always a tech which supports any emerged technology or emerging technology we bring fairness to the other technologies that are there there and AI has been one of the biggest concentration of web 3 all these few months majorly starting with decentralized compute and then data and agents as well decentralized compute as you know has been pretty much covered by players like IET and ather just like how it was in the traditional space we saw Nvidia solving the compute problem which led larger language to be built on the other side we have agents the last few months agents have been speculative have been the biggest rage but most of them have been Mish right and this is how agent era started in the traditional space as well when chat GPT launched most of these agents were chat GPT rappers right they didn't have much purpose they didn't have much utility they were just rappers and then apps like perplexity came about which had much better utility provide much better solutions for users queries AI apps and agents started to use specialized models along with the larger language models that were there which led these apps to be much better and to be much precise than what they were compared to these wrapping agents that were there before the current agents of web3 are majorly memes as I told you about they bloated Twitter Bots with a particular Persona probably talking like a particular celebrity and tweeting about the market scenarios insights and so and so we've seen chat agents where users go ahead and interact with these chats to get access probably like a tokenized access and then we've also seen these agents get provoked when you engage with them we've also seen these agents experiment with trade but most of these trades have been uh human created or human intented most of it have been having human in the loop right this is the current stage of agents in web 3 right now this is what we believe the future of Agents will become and as I told you we've seen this happen in the traditional space as well agents will become much more specialized they will be focused towards a particular category in web 3 agents will become much more adoptive they will understand how you converse with them they'll get to know your feedback the the queries that you put in and then adopt accordingly agents will be become much more Pro proactive right they'll become autonomous no not much of human involved we would see them going ahead and executing trades on its own and then we will also see agents share their task with other set of agents that are going to be there like a swarm of Agents which will execute tasks for you but for this what we need is infrastructure to support that right similar to what happened in the traditional space we would need specialized datasets knowledge bases and models to support these agents I'll skip this slide because it's too boring but um most of the time if you take a look at it right larger language models have been trained on internet data 80% of this data has been predominantly internet scrape data it has lot of breadth but no depth in terms of any particular topic but that would not be useful for these agents to excel in a particular scenario like how humans we've always been trained on a particular field to be an expert in that field a surgeon is much better than a general practitioner agents similarly have to be trained on domain specific data open Ledger enables users and Builders to build specialized agents some of these are examples of specialized models that have been built over the last two years across various platforms you've seen specialized models used in healthcare like met Farm in finance you've seen blue GPT and many more as well we enable people to build these specialized models at scale where data contributors contribute data sets use our platform to build these models and also use our attribution layer to reward back to the data contributors what we build is a trustless layer in which subject matter experts and businesses can go forward contribute data sets knowing that Builders will be able to use these data sets and build models and reward them back and they'll be able to track the entire prominence of how the data was used on shame now let's look at how a model is being built on open Ledger as you can see here it starts with the model proposal a user goes ahead and creates a model proposes the use case of this model tokenizes model and also puts it under a particular category and once these models gets enough stake and tokenize it goes into the data layer this is the place where models get the data that they require data contributors can go ahead and contribute various data sets to these models it could be documents it could be code descriptions or it could just be training data sets where you answer questions depending on the model that is proposed they can use our data layer to contribute data to this the third part is once we have enough data sets once we have populated the data reposer with enough data sets for a model the model builder decides to go ahead and use these data sets and use our no code platform to bring back all these data sets and start fine-tuning in our platform this is a no code platform that uses Laura fine tuning to train these models these are large language models base models that are taken and on top of which we fine-tune a layer on top of it with more data sets right think of this like we take a base product and add some customizations on top of it very similarly we use Lura finetuning to customize the existing open- Source models that are there once you train these models you could also use various evaluation metrics that are available in the platform you can use our chat to eval evaluate these products see how they perform see how these models behave to the the data that is strained on you can question these models you can ask various inputs see the outputs that comes in and then figure out what went wrong with the model training and then retrain again and probably use various other metrics that is there in the platform to train these models better so one of the crucial you know sectors of this uh of open Ledger where models get trained once these models are trained once we have enough models it goes into a deployment layer that's called open Laura which is a framework that we've launched which enables thousands of models to be launched in one single GPU right having one GPU run one single specialized model will be too costly so we went ahead and created this framework which enables each GPU to run one single base model and have the fine-tuned adapters as an additional Lura fine-tuning on top of the existing framework that is there we use ather IET and many other GP providers to go go ahead and run our models on them the last layer is the attribution layer which is one of the most important aspects for platform this is where we enable model Builders to reward the data contributors we use proof of attribution to figure out how on every inference in a model the data that has been contributed to for that particular inference is calculated and then we use that to reward the data contributors we we also use that to explain the model better right let's say if a agent goes ahead and uses the model and it goes Rogue something wrong happens we'll be able to figure out which data actually caused for that particular action to happen in certain cases the agent could perform better then you would actually know which data actually caus these agents to perform better you can optimize and get more of the data these are some real world applications that are currently built on top of open edger all these are models that are currently being built um let me talk about some of these models which are quite interesting we have a team which was X Google Deep Mind which is building data intelligence layer this is a particular model which has verticalized data sets of the internet data they've gone ahead and taken common craw and created specialized topics on that particular data set and created specialized versions of internet data another one is a D5 model mod which has onchain and offchain sentiments which will enable def AI agents to be built on top of this we have a model which is built on move data sets think of this to be a customized version of cursor which is a programming language model where uh this particular model is trained on move data sets move code and descriptions and so on and we have another model which is trained on solidity very similar to the move model this also has code descriptions of solidity code um documentations and many more as well this could also enable co-pilots that can help you audit contracts so these are some of the use cases that currently being built on our platform I would like to end the slide with a quick analogy between us and YouTube YouTube is a platform that happened about 10 15 years when web needed content right it provided a platform where content creators can go ahead and monetize their content contribute to YouTube and get paid for it web got enriched with this particular content very similarly AI needs data open ledgers creating a platform where data contributors subject matter experts businesses can contribute data and index data to AI to enrich the future of AI thank you guys uh we are supported by one of the best people in the industry and also fellow co-founder friends and uh come join us and build more on open Ledger thank you thank you [Music]