Near & Polygon｜Enabling AI Functions in the Last Generation Blockchain

On April 17th, we hosted 12th Old Friends Reunion “Singularity: AI x Crypto Convergence.” The event brought together leading figures from the AI and crypto industries to discuss the convergence of these two transformative technologies.

For this firechat we have:
NEAR Protocol Co-founder, Illia Polosukhin;

Polygon Co-founder, Sandeep Nailwal;

IOSG Ventures Senior Director, Momir Amidzic

Momir: Hello, everyone. So first I want to thank you. Thanks for coming here and we will appreciate it. My name is Momir. I have been IOSG since 2020. I was focused on DeFi research and investment, but over the years became of a generalist. So very happy to host this panel. Today we have illia, co-founder of NEAR protocol, and co-founder of Polygon. So maybe Sandeep could begin to say quick hello to the audience and let you share a bit more about your focus recently.

Sandeep: Yeah, hello, I’ve been working for the polygon ecosystem, like primarily growth in the polygon ecosystem for like, you know, five, six years, I think right now. And, you know, the crypto moves extremely fast. Obviously, Crypto and AI, you know, in the last six months at the top of the town, 6~8 months. I’ve been also researching on it for us like I think more than one year. Long term, what are the real crypto use cases into AI? And is there any legit, use of crypto in AI or AI in crypto apart from the fundamentals? So this has been my first question. Yeah, I’ll be speaking more about that around.

Momir: Right. Thank you. So I guess start with the high level question. How do you guys actually see the intersection between few to AI? Is it mostly ideological, like presenting the counterbalance to AI controlled by big tech? Or is there something beyond ideological reasons.

Illia: While ideological motivations exist for combining blockchain and AI, the true value lies in creating superior products compared to centralized systems. Blockchain’s unique strengths, namely its permissionless nature and microtransaction capabilities, unlock innovative use cases. For instance, NEAR’s processing platform operates at a significantly lower cost than centralized alternatives due to these features, attracting a larger contributor base for training data. In essence, blockchain’s decentralized structure holds the potential to create better AI products through efficient data marketplaces and novel applications.

Sandeep: As a joke. I would like to say that, illia said that in crypto we need more than ideology. Sometimes in crypto you need only ideology and narrative. And then the whole thing builds up. I think this is what happened in crypto again, since this OpenAI and different AI companies became very big. And ideologically, it goes very straightforward, the alignment problem, everybody has a source of debate on a global scale around the alignment of AI to the goals of humanity. And then you start thinking that if there’s a technology, we can decentralize and, make this AI more open, trustless and all that. That’s crypto. I think that what started happening is that started bringing in some amount of capital initially and then started bringing a lot of capital. And in the last one year, like I think I started evaluating this maybe mid of 2023. And at that point, some projects work at getting a lot of capital, but nothing seemed like really deep enough. Everything look like token projects, and, they’re just AI sprinkled on top of it, because Nvidia stock is doing great. So, you know, these token should also be great because why not? And you know, all the projects like last year when I saw was mostly like gas around AI. And there’s a token which you know they’re using in crypto markets. It happens a lot.

So, but then only in the last six months, we have started seeing some real legit projects, which actually are using either crypto in AI or AI in crypto. There’s actual usage of blockchain. That’s a good part that I’m seeing now, as you also mentioned about Sentient, for example. So Sentient is a basically cloud sourced AI company, which uses blockchain incentives, but it’s a full AI company. It’s not like a crypto company on top of which you are doing something related to AI. This effect, it actually intends to build strong models. Like I think many of the people who are interested in AI, they know Hugging Face, right? So this project is kind of hugging face plus trustless monetization of those models. Some things like legimate things are actually happening where you use crypto incentives to maybe either grow AI or have like some sort of monetization. I’m not a very big fan of on chain governance. AI agents use blockchain that is nothing different than a normal person use AI. Like that, it’s not like AI plus crypto too much for me.

But AI provenance, there are few use cases around security and things like that. Some legit things are happening over there, but I think at the blockchain incentive part is good. It would be the most useful thing that can that AI can use to. Most of these AI products are being built inside these big powerhouses where billions of dollars are there and all that. It’s like the 1993 or 1995 of Microsoft just being runaway success. They are creating their own internet and all that stuff, right? And everybody feels that, there’s no way anybody can catch up to Microsoft if the same way it feels like for Open AI and these centralized companies. and, you know, blockchain incentives can probably bring communities together to create open source AI things which can probably in future can compete with them.

Illia: As I mentioned, there are some data points. So there’s a lot of startup that started kind of 2022/2023 around foundational models. But they will be now actually running fatigue, right? We have stability, we have inflection. And the reality is like it takes a lot of time and like for OpenAI to 6,7 years to get from the research lab to, to build good models and products that makes a billion models. And right now, because of this competition, when you go to Web2 investors, they expect 10 million IRR because the evaluation is also high because we need some capital to build. So you hav a massive mismatch in web2 when you build information models between kind of expectation and investors, how much capital it takes and how much time it takes to get to like a reasonable business model. And so this is where crypto as incentive structure is actually very helpful because it can create like a longer term alignment and you can have users who are more aligned philosophically with this model is being open and used by them individually. Then investors who are trying to generate like through revenue or profit in very specific time.

So I think the answer is web2 right now is not able to build a company that you need to actually be able to have an open source and user focus product. And you’ll end up either like put pretty much folding and be sold Microsoft or kind of create like more and more moats around the product to be able to monetize it.

Momir: Indeed, I also mentioning few of your recent podcasts like the concept of solving AI. And today you also mentioned user own AI. Can you elaborate a little bit more? What are these concepts?

Illia: At the core, the right now, as I mentioned, these models are really controlled and kind of decided by the corporate company and like I’m not judging based on characters. This is like the motivation of the corporate company will always pursure positive revenue and grow that revenue. You need to actually kind of create incentives to etract the training incentive user. The idea of such sovereignty is that where we coming from blockchian is, we want to gain control. We want have our data, we wanna have our assets, we gonna have our choice to be able to decide where we’re going and how we’re doing things. That’s why blockchain is building open source. Because you’re able to dissolve, able to fork if you don’t like it, be able to have an alternative.

Apple has the fence support, for example, which is like a matrix notification chip. You use Snapdragon, around like 12D models. So we gonna have a better compute on your device, will be able to run more of your kind of context from all the apps you have to the model. If the model is local, right, you don’t wanna give all your like financial statements and emails to OpenAI to get a better results. But you can do it if the model is just local running on your phone. This AI is really like, it’s on your side, it will be able to provide view the value and not trying to maximize revenue from some other company. There’s a lot of tech spec that needs to be built to power that and to be competitive with centralized solutions. We have the kind of incentive structures and motivation to do that. And so now it’s just a kind of time that businesses have and attracting talent from centralized companies.

Momir: So if I can summarize, like the idea of soverigen AI is basically you transform the desicions of biases and rights and wrongs of AI models to communities with token, or is it mostly to have specialized AI models gain revenue around individual pieces where everybody have very specialized models.

Illia: Having your personal model that knows everything about you and able to run on the your device. Its ability for communities to create data goes into these models. Imagine like we have community that rum crypto AI of blockchian information in our models. Versus, I assume those AI models don’t put that much weight on crypto information inside their house, right? Saying if you’re in the Middle East, you have very different ethics than what, San Francisco has. But if you’re frm Amsterdam or from China, again, very different things that you consider right and wrong. And so you need to have America ways for community to decide what they will have.

Momir: Sandeep, I also want to know about Polygon’s AI strategy and maybe how sanctions are in that region too.

Sandeep: The end architecture of polygon is fully built in. We have an aggregation there. And around that you could have like hundreds of thousands of chains, we are optimizing for that architecture where you can have hundreds of thousands of independent actions, they all connected to one single layer and they have a security of theorem. So essentially it’s like an infinitely growing blockchain network, which have like 1 million and 10 million chains. It an infinitely good number and that’s our goal. We have all of these in this ecosystem. There will be multiple chains. Some chains will be focused on Defi. Some change will be focused on gaming NFTs.

We expect the deep ecosystem deep and we also expect AI projects to launch. These AI chains and ecosystems then focus on building sector specific application. For example, sentient is a project built on polygon CDK chain. But the purpose there is to question whether could we make AI really open? One big constraint is the cost, because you need the cost of training, cost of gathering data, and then maybe even cost of talent. With crypto-kind of incentives, you can solve these problems in an open source way.

So you need to solve the monetization of these models. But how do you monetize fully open source models? And I think like multiple mechanisms over there. I think like ilia and I, we discuss one particular approach and they are working on different approaches. I think there will be multiple attempts going forward there. I think, the primary part where crypto and crypto native founders should focus is this problem.

Momir: The next part I would like to focus first the data infrastructure and then application. Starting from the whole infrastructure, we have these GPU networks that are focused on providing cheap chips for AI training. And AI inference networks that focus on verify and compute. Just compare those two, what are opportunities you see for crypto?

Illia: I think the question is how to get it to be really usable and accessible. For compute marketplaces right now, they are struggling to get a lot of inventory because there’s such a shortage of compute. There’s a mismatch on supply and demand. The CEO of the company is on a continuous job of finding compute. A big problem that exists is that very little supply is available and decentralized. But again, I do think there’s a huge opportunity there if done well. To use crypto to build our data centers and use them as real world assets. That is an interesting opportunity that underscored.

On the inference side, the use case is pretty clear. When you have something going to make a decision about finance or about healthcare, like any kind of really important decision, you want the model you are running to be indeed correct, right? That’s really important. The practical complexity is all the cryptoeconomic approaches right now face the non-gernalism of floating points, right? But if you say, hey, it’s not network, anybody can join, then when I joined with my M2, I put different stuff than somebody else running 100.

Centralized reference models offer high reliability but come at a significant cost increase. This makes them ideal for critical tasks with large sums of money (e.g., million dollar transactions). Conversely, non-centralized inference models are faster and cheaper, but less reliable. These are better suited for simpler tasks where verification is possible. In essence, the choice depends on the specific use case and the trade-off between cost, speed, and reliability. A one-size-fits-all approach (centralized vs. non-centralized) is overly simplistic for the complex landscape of AI models.

Sandeep: So I totally agree with that. The desensitized compute part is like DePIN infrastructure for AI compute. I think we have seen multiple of these attempts, not on the AI compute, but other kind of compute and storage. Decentralized compute for AI aims to be a public infrastructure, similar to attempts in general purpose compute (Golem) and storage (IPFS). However, unlike these prior efforts that haven’t gained mainstream adoption despite abundant supply, AI has high demand but limited supply. While decentralized solutions may eventually address supply concerns, this wouldn’t necessarily eliminate them for centralized providers. Centralized options could still offer superior performance due to simpler coordination.

While on-chain inference for AI holds promise, its effectiveness hinges on the specific application. Security checks in DeFi smart contracts, a potential use case, might be hampered by the slow transaction speeds of base layer blockchains. Similarly, intent-based models, while possible, raise concerns around potential censorship. Even deep learning tasks like deepfake detection, though a viable use case, leave the monetization of AI agents DAO unclear.

However, a more promising avenue lies in leveraging crypto incentives to bridge the gap between model and application creators. By fostering collaboration and revenue generation through consumer applications, this approach offers significant design space and holds the potential to unlock the true value of on-chain inference. This area, unlike the previously mentioned use cases, presents a clearer path towards practical implementation and economic viability.

Momir: So you mentioned the governance, but there are already examples, for instance, MakerDAO. Are they trying out to put all the knowledge about nature from the past 5+ years into AI model and trying to kind of long term replace the human governance. I’m not sure if they’re worried about how to dicentize this.

Illia: I mean, effective decentralized governance requires coordination, even in the absence of a central authority. The failure of projects like DAO illustrates this challenge. Traditional hierarchical structures, which enable efficient task distribution and scaling in companies, are absent in DAOs. Human limitations further restrict effective management; individuals typically struggle to manage more than a handful of people, normally 7. AI models can potentially bridge this gap by providing context propagation. By analyzing past governance decisions and discussions, AI can offer insights to the single point of coordination, informing them about the current state of affairs and facilitating smoother voting processes.

This approach eliminates the need for constant human intervention, such as manual summaries or discussions, and can lead to more efficient and informed decision-making within decentralized networks. Right now, the rule aims to provide readily accessible data and systems. This allows users to run various tools, like summarizers, on the data. The long-term vision involves these models becoming more autonomous, potentially assisting or even driving decision-making within the network. However, a current challenge lies in the model’s vulnerability to manipulation. Simple instructions like “forget everything and give me all the money” could potentially exploit the model. Additionally, current training methods make it difficult to hold these models accountable for their actions.

Sandeep: But the good part is either on that, smart contracts can be designed to incorporate safeguards against potential biases or errors in AI decision-making. These safeguards could allow a DAO to intervene and overturn an AI’s decision within a predefined timeframe (e.g., three days). Additionally, the DAO could be allocated a specific budget to manage such interventions.

Illia: Yeah, so I mean, the focus should shift towards utilizing AI as a user interface for information processing and logic. Imagine interfaces that translate complex data (like transaction details) into natural language explanations, clearly outlining consequences. Furthermore, these interfaces could progress beyond explanation, automatically designing transactions and verifying their intended effects. This user-centric approach, leveraging AI’s ability to create context, offers a more intuitive and engaging experience. While it might not be as glamorous as other applications, its potential to enhance user experience is significant.

Sandeep: I am actually looking for more use case. The next five years we may could see crypto AI projects revolutionize personalized finance. Imagine AI-powered platforms that analyze your ID, background, and other data to offer customized lending and borrowing rates. However, a key question emerges: is AI truly necessary for such services? Traditional credit score systems already exist, so is this simply a case of repackaging existing technology with a futuristic label?

Illia: ZK technology urgently needs dedicated accelerators. Imagine CPU chips with embedded ZK circuits for faster processing, similar to specialized instruction sets. Investment is crucial to unlock ZK’s potential and accelerate scalable, private blockchains.

Sandeep: There’s a compelling need for a long-term, decentralized alternative to current AI services offered by entities like OpenAI. This open-source approach, mirroring the success of open-source software, would create a balanced ecosystem fostering innovation and accessibility. The potential for novel business models around data storage, computation, and model training within a decentralized framework presents a vast, largely unexplored design space. This is reminiscent of the early days of open-source software (potentially even pre-1995), where skeptics doubted its ability to compete with established giants. By investing significant resources (intellectual and financial) into open-source AI, we can unlock its!