[Music] Right. Up next, I want to bring on our next speaker. It's Rahm from Open Ledger to tell you all about building specialized blockchains for AI. So, please give Rahm a big round of applause. [Applause]
Awesome. Hey guys, this is Ram. I'm from Open Ledger. I'm one of the co-founders and co-creators of the protocol that we're building. Let me try this on.
Awesome. Yeah, today I'm talking about why AI models have to be specialized, right? Most of the AI models that we use today are larger language models which are trained on internet data. This is OpenAI, Llama, Mistl, all these models, right? And the reason why is because internet data is the most accessible easier to scrape data set out there.
That's the first data set we can reach out to without actually paying any cost. And uh that's where most of the models were built as well. And think of that to be like an average human being who's just done gra like you know done a graduation. He has an understanding over the world. But models cannot be used as it is right.
If you want to use AI for very specific use cases, you would need these models to be trained on very specific data. Models have to become specialized. What Open Ledger is building is an end to-end protocol that enables users to contribute data sets and build specialized AI models and reward the data contributors as these model make revenue and we use attribution to do that. Our background started about a decade back in the space. We've been majorly focused towards blockchain R&D and machine learning.
We've worked with lot of enterprises. This includes Walmart, Sony, Viacom and many more. We enabled all of these enterprises to use web 3. They want to get into web3. They want there's huge hype back then.
They were experimenting around blockchain. A lot of enterprise blockchains emerging during that particular time. And we had an opportunity to work with these brands. We also worked with them on automation of various of their products. So we had an understanding of how machine learning works.
We also implemented blockchain based solutions. But what we kind of figured out you know even though we were making consumable revenue we had about 200 employees across the globe. What we kind of figured out is that it's not a product for the masses. Right? This is when the charge movement was happening.
Everyone was building around generative AI and we knew that AI is going to be like a public good. But there were a couple of problems in the space right. Most of the AI as you know is like generalized as we spoke about they're very generic models they don't have like detailed knowledge of a particular sector and all of these AI models are built by centralized firms think of these to be like you know your open AI and all that which have their own need on why they're building these AI models the third problem is most of these AI models are a black box right if you get an output from a model an inference you don't know why that output came you don't have any clarity idea about how the data was actually used to train the model and how did the model get the context with that data and why that inference came about. That's a problem, right? If you're going to use AI for let's say finance or any kind of applications out there, even like healthcare, you want to make sure there is auditability.
You want to know how these AI models works. So these are three major problems that we saw in the space. We started building a product around to solve this problem. We wanted to be very specific, narrow down. There's a lot to AR.
There was compute, data and models. So we made sure that we pick data and models. started focusing towards that and what we eventually built is an end to-end framework that lets any developer who's interested in AI to collect data sets on top of open ledger and then use our no code platform to build AI models. But it should not stop there, right? We need to make sure that people who participate in this ecosystem who enable these AI models to be built, they should be rewarded as part of the process.
So we use attribution. We pinpointed that this particular data provided by this user is impacting the model and is actually useful and we use attribution to reward them and then you build built an application you know people are using it you have an attribution protocol people are getting paid but is it verifiable no so we went ahead and built a blockchain around that which makes these process end to end to be verifiable anyone who contributes data the data set is onchain And we track that onchain as well. When it goes into a model, the model's provenence is on chain. We know which data goes into the model. And then every inference is tracked on chain.
And that's where attribution is figured out as well. All of this happens on open edges layer 2. Our product is not very unique from other platforms but constably unique in the places where it has to be and that's why we are different from other traditional AI platforms out there which includes hugging face kegel and all that most of these players never reward the people who use these products right most of them are um either collecting data from you not providing ownership they either go ahead and use that to build models and you're not even part of it you don't know how your data is being used and none of them are verifiable right you can't verify them all these are centralized platform in open ledger you get to attribute to the data provider you get paid you can also retain the ownership to your data our stack is pretty simple we wanted to make sure that there is an application layer and there is a blockchain layer that relays whatever happens in the application and we've built our enter entire protocol based on that agents and models stack on top of of it. The end result is building a model and then any agent or an app can infer to these models that are being built on open edger. I will talk more about the product.
The three major parts of the product. The first is called data nets or data attribution where users can go ahead and collect data sets for the models they're looking to build. Think of this as how a human brain gets feeded knowledge and he becomes specialized in one particular subject. Very similarly, an AI model needs that knowledge. So users, anyone who's building an AI model, they can collect data on top of open ledger and feed that knowledge to a model.
Let's take an example here. Assuming someone wants to build a model for solidity. Solidity is one of the most largest used programming language in the web 3 space. But unfortunately, you know, cursor, which is one of the biggest models out there. It's a codegen model does not really understand solidity very well.
If you guys have used it, you would know. So there's a team which is looking to build a model specifically trained for solidity and they're using opener for that. And the process started like this, right? They started a data net. They enabled to have like a rule engine which is a specific rule about what kind of data has to be contributed.
And then they let users, developers, anyone who's been in the crypto space go ahead and contribute data to that. And the data was basically code code descriptions and documentations. With that the solidity model started sort of building like a data repository where all of these data sets were contributed to that. Once they had this data set they went ahead and used our model factory platform which is a no code AI fine-tuning platform where you can pull up any of the data that you collect on top of open edger use a base model. Think of it to be like a you know it could be like a nama or a mistral model and then you basically superimpose the data that you collected on top of this model that is already there.
The process is called fine-tuning where you post train this model with more information and make it specialized to that particular data that you collected and that's how developers use our platform to build these AI models. It's very simple. It's just bunch of buttons that you have to click. You build a model and then you start to iterate. You test the model.
you see if it works and then you start to iterate using our platform. We also have like sort of like an advanced version of this where if you're a machine learning expert, you can manage your hyperparameters and probably build a model in a much more accurate way. Once a model is built, it's hosted on Open Ledger's marketplace. It's called open models. You can access all of the models are being built on top of open ledger.
These are models that are built around various categories. There are players who are building around let's say astrology, solidity and many more other use cases as well. So it's quite unique and once these models are available on open ledger they can be directly accessed using a chat platform. Think of it to be a very similar interface like perplexity or chatb right you can basically query these models and you pay for inference as you access this model you pay for it. agents and apps can use our APIs to access any of these models that is there and uh we also provide them as a framework and I'll talk more about that as well.
There are players who are building these AI models on us. There is ambio which is a deepin project using our platform building a model for air quality. There's a team based team which is building a model with verticalized internet data. There's Morpheus which is building that solidity model that I spoke about. There's up network and couple of other guys as well.
We also work with enterprises. We work with Trust Wallet. We are building a model for them to enable easier onboarding of users. We work with Dubai Tax Authority. We are building a model for just taxes where you just the citizens can go ahead and query and get to know about their taxes.
We also work with one of the largest e-commerce out there which we'll be announcing soon. to building a support customer support model for them, right? And the reason why we work with enterprises, right? It makes sense for community to use a platform. They can get access to data, they get access to a no code platform, they can build models with us and then reward users.
But why enterprises, right? Enterprises have this problem. When they use AI, all of these AI as we spoke before are a black box. You don't know how you really it really works, right? when you get an output, when you get an inference, you can't pinpoint to say this particular data cost this inference, right?
That's a problem for them. If they use that AI and actually make decisions with that, they want to make sure it's entirely auditable. So they went ahead and started experimenting with us and we built this for an e-commerce giant and we started training a model where their with their internal data and what we could achieve using attribution is that we could go back and show which data set caused an inference to happen and we could also pinpoint to show what was the impact of these data sets as well. So if something goes wrong tomorrow, something goes rogue, you can go back and figure out what was the poisonous data and actually remove that. In fact, if something goes well and you had actually the AI was useful, you made better decisions, then you can add more data sets of that.
That's one of the reasons why we work with enterprises. Um, we continue to work more. We have about six to seven enterprises that are using our platform to build AI models. So, there are two wings to us. One is the community wing where we enable end developers to use our platform to build models and there are enterprises use our platform to make their models much more explainable.
We are having people who are building ecosystems on top of us. There's a framework which is being built which enables easier access to the AI models that are being built on top of us. Think of us to be like another agenting framework but just much more direct access to the models. You can even use apps and just call our models using API. Uh we have a team which is actually building a stable coin for microp payments for attribution.
Um all these payments that we give using open tokens is too volatile. So the building make a stable coin for this. Let's see how that goes. Uh we also have a team which is building a DEX platform where models will be tokenized and the derivatives of these models will be traded on top of open ledger's DEX. Uh where users are basically enabling you to monetize the model that is being built on top of open ledger apart from the inference revenue that you get.
So these are ecosystem that are using our platform building on top of us enabling more and more access for end users and we've been in test net for about 6 months now. We have close to about a million users contributing data sets to us. We have about 100,000 people participating and building and experimenting with AI models. Um we've been going mainnet we I mean sorry we'll be going mainet in about two weeks time and once it's open you can have anyone access the platform it'll be completely permissionless they can build these models on their own and we've been backed by polish chain bottles and a couple of other investors uh we raised over $15 million and uh I'll be around if you have any questions let me know but it's great talking to you guys and uh thank you thank you for the opportunity I'll see you [Applause] [Music]
Automatic transcript — names and jargon may be misspelled.