The Future of AI Is Specialized: Why General Models Aren't Enough – San Francisco 2025

Good to be here. I'm Ram. I'm uh one of the core contributors at Open Ledger. Uh today I'm going to talk about why the future of AI is specialized and not why and also why um generalized models, the larger language models that you see today are not good enough. AI is definitely changing the way we work.

Right? This has happened in terms of any technology that has evolved over the last 100 plus years, right? from printing press to steam engines to the internet. We've always seen that impact our day-to-day lives. Pretty much the same with AI as well.

Uh we start using AI already in our day-to-day works, right from our meetings that are being recorded and given as notes to our daily usage of various products that we have to our CRM. Future of AI will also be much better as well, much bigger. we would probably use AI in healthcare, AI in transport and more and more. Web 3 has been a technology that kind of like complements the emerging technology that comes ahead. We've used web 3 for funding.

We've used web 3 for decentralized resource access. We've used web 3 for bootstrapping and also for bringing fairness to any technology that came in. The same goes with AI as well. Um AI has been pretty much the biggest topic in web 3 over the last few months. There are three ma three major components that we've been looking at in web 3.

The first one is decentralized compute. This is the foundation layer. If you take a look at it, most of the companies that has been established over the last 2 years let people contribute compute and rewarded them for the resources that they provided. Right? On the other side, we've also seen agents evolve over the last few months.

We've seen onchain agents even though these agents did not have much utility. They have pretty much been memes. But still there has been experiments around how agents can be decentralized not owned by a centralized party can be relying on chain. And then the next sector uh which is not so seriously looked upon but has a huge potential is decentralized data and models. How can we use web 3 as a primitive to enable people to contribute data sets and build AI models out of that?

Right? If you take a look at most of the models that you see today in the industry, LLMs have been trained on internet data. Data that is easily accessible data that you can find on the internet and that has been scraped and trained for these models. uh but the gold mine right the actual data sets that are useful are pretty much relying on people who are subject matter expert or companies which own these data sets that are really useful right so how can we use web 3 as a primitive to enable them to contribute these data sets and get rewarded for that so let me talk more about agents and then let you um give you an idea about how we are doing this agents as I had told you have been pretty much mimish over the last few months. Most of these agents have been Twitter bots which go ahead and post random content online.

There have been chat bots where you can chat with them but they don't understand any kind of context. They're pretty much a persona of let's say Trump or Elon Musk. They they can engage with you. They can enable you to trade but there's always humor in the loop for these agents to trade well. Right?

They have never been autonomous. They are pretty much glorified bots. But the future of agents will change. This has happened in the traditional space as well. Agents when it initially came about in the traditional space were chat GBD rappers.

But then it evolved onto better apps, better agents like perplexity which use multiple models which use various techniques to make it much more useful and with high utility. We would see that happening in the web 3 space as well. Agents will be specialized in each and every sector of web 3. I personally believe that every activity that is done by humans on chain will be replaced by agents as well. So these agents will be adaptive.

They'll adapt to the ecosystem. These agents will be much more proactive as well. You don't need to have a human in the loop. You would just need to give an intent to these agents. They will go ahead and execute your task.

So for these agents to be what they are and how they can become the current larger language models are not good enough as we discussed most of these larger language models are trained on internet data very generic information but humans as we know right we've also initially have like very general knowledge and then we are specialized in a particular sector we go to a doctor because he's trained in medical science uh we go to someone who's an expert in finance hands to ask advice, financial advice. So we always go to a specialist to get a job done. Very similarly, agents need to have a model which is specialized in a particular sector. A brain that can power it to excel in a sector that is built for. That is what open ledger enables.

We've always we've seen this happen over the last few years. Specialized models have been built in various sector. These are some of the examples that you can see in the screen. uh there has been models that have been built for healthcare like metform which is much better than charge GPT's healthcare model. We've seen models in finance like Bloomberg GPT in science and bio and many more as well.

Open ledger lets people build this at scale. Um think of us to be a protocol that will enable data contributors to contribute data sets to build specialized models on top of us. We use attribution to reward back the data contributors for the data that they provided. And these models that are built on our platform are fine-tuned models, models that are trained with additional data sets in the existing open-source larger language models that are there. As these specialized models have data as the biggest component, data is the biggest mode here.

We want to make sure the data contributors get as equally as rewarded as the model builders. Let us talk more about the various model life cycle that goes through open ledger uh in order for a model to be built. The first is the model proposal layer where any model builder proposes a particular topic in which the model has been built and he's able to tokenize his own model. Here it could be a company or an individual who's looking to propose a model. Once a model is proposed, we have a data layer on which data sets are contributed for each of these models.

And each model has a rule engine which has to be followed. Think of that be like a contract layer where based on the contract that is there as your data gets contributed it's verified on chain and as these data sets are contributed it goes through this no code platform where the model builder populates all of these data sets and build their own fine-tuned versions of these larger language models we enable them to deploy on top of open ledger and once these models are deployed as these models get used by agents as they are inferred. We use attribution to reward back to the data contributors. So this attribution layer is a unique um protocol that we built ourself um on every inference of a model. Every time you query this model, we'll be able to figure out which data caused that inference to happen and we'll be able to plot it on chain and use that to reward back to the data contributors.

So every time a model gets built, they don't have to pay initially for the data that they consume. But as their model gets used, as their model gets accessed, the data contributors get paid as as it grows as the model gets used. And this is the unique way of building models that we have proposed. Like similarly how you use social media platforms where you let content creators create content and contribute and more views they get, more access it has, they get paid. Very similarly, we want to enable data contributors as an economy for them to contribute data and build AI models and have these models get accessed by agents and AI services and as they get access the data contributors get paid for it.

We are people who are building these models on top of us already. We are in test net right now. There a couple of projects I would like to showcase to you guys. The first one is a data intelligence layer which is trained on verticalized internet data sets. It's built by an ex Google deep mind team on top of us.

Um they're building this as a data pool for other models to access which has various subjects of the internet data. Uh this is being currently built on top of us and we have nodes which provide the infrastructure for these models to be built as well. And then we have a D5 model which enables D5 agents to make better decisions which has onchain and off-chain sentiment data. This is being currently built on top of us as we speak as well. Uh we have models that are being built on programmability which includes a move and a solidity model.

These are models that are trained on solidity and move programming languages. Cursor which is a very well-known larger language models for AI development does not understand programming languages like solidity and move. These are very niche programming languages. So there's a team building these specialized models that are trained on solidity as a language and the data sets that that is contributed for these models and the data contributors the developers get benefited as these model generates revenue as these models have agents being used and they generate revenue. I would like to end this conversation and this presentation with uh one important um aspect.

We compare ourself to be to be like the YouTube of AI era. Web happened because there's a lot of content on the web, right? They enable people to contribute content especially YouTube. It enabled people to contribute content and get monetized for it. This led for content creation as a job.

We had people who came to be like YouTubers who contribute very high quality content to get paid. We had businesses come forward and contribute content, their proprietary content knowing that they're going to get paid for it. Very similarly, Open Ledger enables data contributors to contribute as a profession contribute data and get paid for it. and businesses which own proprietary data sets to come forward and contribute their data knowing that they'll get paid for it knowing that they'll be able to see it on chain and verify that and get paid. And I will thank you for the time uh that I got here.

If you have any questions, happy to answer. Uh we are supported by prominent industry venture capitalists in the space and we are currently in our test net right now. You can check out openedger.xyz XYZ and opened HQ on Twitter. Uh would love to have you guys participate and provide feedback.

Thank you. Thank you very much. [Music]