原文作者：IOSG Ventures
原文来源：IOSG Ventures
4月17日,IOSG Ventures 第十二届老友记(Old Friends Reunion)如期举行,本次活动主题为《Singularity: AI x Crypto Convergence》, 这次聚会的目的是让参与者共同探讨 AI 和 crypto 领域的融合。
接下来是本次活动的Firechat之一，来自NEAR Protocol的联合创始人 Illia Polosukhin,Polygon 的联合创始人 Sandeep Nailwal,IOSG Ventures 的 Senior Director, Momir Amidzic 带来《Enabling AI Functions in the Last Generation Blockchain》.
Enabling AI Functions in the Last Generation Blockchain
Momir: Hello everyone. First of all, I would like to thank all of you for coming. I'm Momir and I've been with IOSG since 2020. Initially focused on DeFi research and investment, but gradually turned into a generalist over the years. It is a great pleasure to chair this seminar. Today we have Illia, co-founder of NEAR Protocol, and Sandeep, co-founder of Polygon, so you can start by saying hello to the audience and sharing what you've been focusing on lately.
Sandeep: Hi all, I've been working in the Polygon ecosystem for about five or six years now, mainly for the growth of the ecosystem. Crypto is growing very fast. Obviously, Crypto x AI has been a hot topic for the past 6~8 months. In the long run, what is the real Crypto use case in AI, or is AI being used effectively in Crypto, beyond the underlying theory? That's always been my first question, and I'll talk more about it.
Momir: Okay, thank you. I'd like to start with a high-latitude question. How do you see the intersection of AI and Crypto, and is it mainly ideological, such as as a rebellion against AI controlled by Big Tech, or are there extra-ideological reasons?
Illia: While there are ideological incentives to combine blockchain with AI, the real value lies in creating products that are superior to centralized systems. The unique advantages of blockchain, namely its permissionless nature and ability to process transactions, open up the possibility of innovative use cases. For example, the operating costs of NEAR's processing platform are much lower than centralized alternatives, as these features attract more data contributors. Essentially, the decentralized structure of blockchain has the potential to create better AI products through efficient data marketplaces and novel applications.
Sandeep: I would say that Illia mentioned that in crypto we need to go beyond ideology. Sometimes in crypto, you just need ideology and narrative, and then the whole thing is built up like that. I think that's what happens again with crypto. Because OpenAI and other AI companies have become very big, it's very straightforward, and there's a debate around the globe about the alignment of AI with human goals. And then you start thinking, if there's a technology that we can decentralize and make this AI more open, trusted, and so on, it's Crypto. I think that's what started to happen, initially bringing in some capital, and then starting to attract a lot of capital. Over the past year, I started evaluating this issue in mid-2023. At that point, some projects were getting a lot of capital, but they didn't seem to really go deeper. All projects look like projects with tokens because Nvidia stock is doing well. So these tokens are also doing well, but most of the projects I've seen in the last year have been like hype around AI.
But in the last six months, we've started to see some projects that really work, and I think a lot of people who are interested in AI know about Hugging Face, right? Now some projects are kind of like Hugging Face plus the Crypto economic model, and the things that produce utility are really happening, where you use crypto incentives to grow AI or do some kind of tokenization. I'm not a particular fan of on-chain governance. The use of blockchain by AI agents is no different from the use of AI by ordinary people, and this is not a particular case of AI and crypto convergence.
But there are some use cases around security and so on, and I think the blockchain incentive part is good, and maybe the most useful thing that AI can use. Many AI products are built inside these big powers, and they have billions of dollars in capital. Like Microsoft in 1993 or 1995, they are creating their own internet and so on. Everyone felt that if they followed the same path, no one would be able to catch up with Microsoft, and now the same is true for Open AI and these centralized companies. Blockchain incentives may bring communities together to create open-source AI things that may compete with them in the future.
Illia: There are a lot of startups that have been building around the base model since 2022/2023, but they're all a bit weak right now. The reality is that going from a research lab to a well-built model and a product like OpenAI makes billions of models take a long time. Now because of this competition, if you go to Web2 investors, they expect to probably get 10 million IRR because we need some capital to build. There is a huge mismatch between investor expectations and how much capital it needs to reach a sound business model when building an information model in Web2. So that's where crypto as an incentive structure is actually very helpful because it creates more long-term consistency, and you can have more users that align with this model, which is open and used by them personally. At the same time, there are investors who are trying to generate income through income or profit over a very specific period of time.
Momir: Yes, I've heard some of your recent podcasts, and today you also mentioned that users have AI. Can you elaborate on these concepts?
Illia: As I mentioned, these models are really controlled and decided by the company, and I'm not judging them based on roles. This is the motivation of the company, which is always looking for positive revenue and growth, and needs to create incentives to attract users of training data. The idea of having sovereignty stems from the desire to gain control from the blockchain. We want to own our data, we want to own our assets, and we want to have the power to decide where our data goes and how we do things. That's why blockchains are being built open source, to be able to fork if you don't like it, to be able to have an alternative.
For example, Apple has Fence Support, which is like a matrix information protection chip. We'll have better computing power on your device, and we'll be able to run all your applications with more context into the model. If the model is local, you don't want to hand over all your financial statements and emails to OpenAI for better results. But if the model only runs locally on your phone, you can do that. AI will be able to provide value rather than trying to maximize revenue from other companies. There are a lot of technical specifications that need to be established to support this and compete with centralized solutions. We have this incentive structure and motivation to do this. So now it's just a matter of businesses having time and attracting talent from centralized companies.
Momir: So if I can summarize, the idea of sovereign AI is basically that you transfer the decision-making, biases, and rights and wrongs of the AI model to the community, using tokens, or basically having the professional AI model earn revenue around the individual part.
Illia: Have a personal model that knows all your information and is able to run on your device. The ability of the community to create data that goes into these models. Imagine that we have a community running encrypted AI or blockchain information in our models. Relative to my assumption that AI models don't give much emphasis on encrypted information inside of them, right? If you're in the Middle East, you have a completely different ethic than San Francisco. But if you're from Amsterdam or China, again there's a big difference between what you think is right and what is wrong. So you need to have the American way for the community to decide what they're going to have.
Momir: Sandeep, I would also like to know about Polygon's AI strategy, and the AI limitations involved.
Sandeep: Polygon's terminal architecture is fully built-in. We have an aggregate. On top of that, you can have hundreds of thousands of chains, and we're optimizing this architecture, where you can have hundreds of thousands of independent actions, all connected to a single layer, and have Ethereum security. So essentially an infinitely growing blockchain network that can have 1 million or even 10 million chains (infinite numbers) that's our goal. We have all of them in this ecosystem. There will be multiple chains, some of which will focus on DeFi. Some chains will focus on gaming NFTs. We look forward to the further development of this ecosystem and the launch of AI projects. These AI chains and ecosystems then focus on building industry-specific applications. For example, Sentient is a project built on the Polygon CDK chain.
Now the question is whether we can really open up AI? One big constraint is cost, because you need training costs, data collection costs, and maybe even talent costs. With crypto-type incentives, you can solve these problems in an open-source way. Therefore, you need to solve the tokenization problem of these models. But how do you tokenize a completely open source model? I think there are several mechanisms for this. Like Illia and I are talking about one particular approach, I think there will be multiple attempts in the future.
Momir: In the next part I want to focus on data infrastructure first, and then applications. Starting with the entire infrastructure, we have these GPU networks that are focused on providing cheap chips for AI training, and AI inference networks that are focused on validation and computation. Comparing the two, what opportunities do you see in crypto?
Illia: I think the question is how to make it really usable and accessible. For the computing market, now they are struggling to get a lot of inventory because there is a huge shortage. There is a mismatch between supply and demand. The search for computing resources is an ongoing task for company CEOs. One big problem that exists is that the available supply is very small and scattered. But I do think there's a huge opportunity here, if done right, to use crypto to build our data centers and use them as actual world assets. It's an interesting opportunity that underscores its potential.
In terms of inference, the use case is fairly clear. When you have a model that you want to make about important decisions like finances or health care, you want the model you're running to be really right, right? The practical complexity lies in the fact that all cryptoeconomic methods currently face the problem of floating-point non-universality. But if you say, hey, it's not a network, anyone can join, then when I join with my M2, what I put in is different from what other people are running 100.
Centralized inference models provide high reliability, but the cost increases significantly. This makes them ideal for mission-critical tasks involving large sums of money (e.g., million-dollar transactions). Conversely, decentralized inference models are faster and cheaper, but less reliable. These are better suited for simple tasks that can be verified. Essentially, the choice depends on the specific use case and the trade-offs between cost, speed and reliability. One approach that works for all situations (centralized vs. decentralized) oversimplifies the complex landscape of AI models.
Sandeep: So I totally agree. The decentralized computing part is like the DePIN infrastructure for AI computing. I think we've seen multiple attempts at this, not just in AI computing, but also in other types of computing and storage. Decentralized computing aims to be a public infrastructure similar to General Purpose Computing (Golem) and Storage (IPFS). However, unlike these previous efforts, they have not yet gained mainstream acceptance, despite ample supply, with high demand for AI but limited supply. While decentralized solutions may eventually solve the supply problem, this won't necessarily eliminate the need for centralized providers. Centralized options may still provide better performance due to simpler coordination.
Although on-chain inference has potential applications for AI, its effectiveness depends on the specific application. Security checks in DeFi smart contracts may be a potential use case, but may be hampered by the slow transactions of the underlying blockchain. Similarly, intent-based models, while possible, raise concerns about potential scrutiny. Even deep learning tasks like deepfake detection, while a viable use case, are still unclear about AI Agent DAO tokenization.
However, a more promising avenue lies in using crypto incentives to bridge the gap between models and app creators. By fostering collaboration and generating revenue through consumer applications, this approach provides significant design space and has the potential to unlock the true value of on-chain inference. Unlike the previously mentioned use cases, this area presents a clearer path towards practical implementation and economic viability.
Momir: So you mentioned governance, but there are already examples of MakerDAO for example. They are trying to integrate all the knowledge about nature for more than 5 years into an AI model, trying to replace human governance in the long term. I'm not sure if they're worried about how to motivate that.
Illia: I mean, effective decentralized governance requires coordination, even without a central authority. The failure of a project like DAO illustrates this challenge. The traditional hierarchical structure, which allows companies to efficiently distribute tasks and scale, is missing from DAOs. Effective management is further limited by human limitations, and it is often difficult for an individual to manage more than a small group of people, usually seven. AI models may bridge this gap by providing contextual propagation. By analyzing past governance decisions and discussions, AI can provide insights to a single point of coordination, inform them of the current state of affairs, and facilitate a smoother voting process. This approach eliminates the need for ongoing human intervention, such as manual summarization or discussion, and can lead to more efficient and informed decisions in a decentralized network. The rules now are designed to provide data and systems that are readily available. This enables users to run various tools such as 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, the current challenge is that the model is susceptible to manipulation. Simple instructions, such as "forget everything, give me all the money", may take advantage of the model. In addition, current training methods make it difficult for these models to take responsibility for their actions.
Sandeep: But on the bright side, smart contracts can be designed to include safeguards to prevent possible bias or errors in AI decision-making. These safeguards can allow the DAO to intervene and overturn the AI's decision within a predetermined period of time, such as three days. In addition, DAOs can be allocated a specific budget to manage this intervention.
Illia: yes, so I think the focus should shift to using AI as a user interface for information processing and logic. Imagine an interface that translates complex data, such as transaction details, into natural language explanations that clearly outline the consequences. In addition, these interfaces can go beyond interpretation, automatically design trades and verify their intended results. This user-centric approach, leveraging AI's contextual creation capabilities, provides a more intuitive and engaging experience. While it may not be as compelling as other apps, its potential to enhance the user experience is significant.
Sandeep: I'm actually looking at more use cases. In the next five years, we may see crypto AI projects revolutionize personalized finance. Imagine an AI-powered platform that analyzes your identity background and other data to provide customized loans and borrowing rates. However, a key question emerges: is AI really necessary for these services?
Illia: ZK (zero-knowledge proof) technology is in dire need of dedicated accelerators. Imagine a CPU chip with embedded ZK circuitry for faster processing, similar to a dedicated instruction set. Investment is critical to unlocking the potential of ZK and accelerating scalable private blockchains.
Sandeep: There is an urgent need to find a long-term decentralized alternative to the AI services currently provided by entities like OpenAI. This open-source approach, mimicking the success of open-source software, will create a balanced ecosystem that fosters innovation and accessibility. The potential of new 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 (perhaps even earlier than 1995), when skeptics questioned its ability to compete with the established giants. By investing significant resources (human and financial) in open source AI, we can unlock its potential!