IOSG: Expanding the new narrative of infrastructure from the AI x Web3 technology stack

Author: IOSG Ventures

Introduction

The rapid development of large language models (LLMs) has sparked interest in leveraging artificial intelligence (AI) to transform various industries. The blockchain industry is no exception, with the emergence of the AI x Crypto narrative drawing significant attention. This article explores three main ways to integrate AI and crypto, and discusses the unique opportunities blockchain technology presents in addressing challenges within the AI sector.

The three pathways of AI x Crypto include:

• 1. Integrating AI into existing products: Companies like Dune are enhancing their products with AI, such as introducing SQL copilot to assist users in writing complex queries.

• 2. Building AI infrastructure for the crypto ecosystem: Startups like Ritual and Autonolas focus on developing AI-driven infrastructure tailored to the needs of the crypto ecosystem.

• 3. Using blockchain to solve AI industry problems: Projects like Gensyn, EZKL, and io.net are exploring how blockchain technology can address challenges faced by the AI industry, such as data privacy, security, and transparency.

The uniqueness of AI x Crypto lies in the potential of blockchain technology to resolve inherent issues within the AI sector. This distinctive intersection opens up new possibilities for innovative solutions that benefit both the AI and blockchain communities.

As we delve deeper into the AI x Crypto space, we aim to identify and showcase the most promising applications of blockchain technology in addressing the challenges faced by the AI industry. By collaborating with AI industry experts and crypto builders, we are committed to fostering the development of cutting-edge solutions that fully leverage the advantages of both technologies.

Part One

1. Industry Overview

The AI x Crypto space can be divided into two main categories: infrastructure and applications. While some existing infrastructures continue to support AI use cases, new entrants are launching entirely new AI-native architectures.

1.1 Computing Networks

In the AI x Crypto domain, computing networks play a crucial role in providing the infrastructure needed for AI applications. These networks can be categorized into two types based on the tasks they support: general-purpose computing networks and specialized computing networks.

1.1.1 General-Purpose Computing Networks

General-purpose computing networks (e.g., IO.net and Akash) provide users with the opportunity to access machines via SSH and offer command-line interfaces (CLI) that enable users to build their own applications. These networks are similar to virtual private servers (VPS), providing personal computing environments in the cloud.

IO.net is based on the Solana ecosystem, focusing on GPU leasing and computing clusters, while Akash, based on the Cosmos ecosystem, primarily offers CPU cloud servers and various application templates.

IOSG Ventures' Perspective:

Compared to the mature Web2 cloud market, computing networks are still in their early stages. Web3 computing networks lack the "Lego" building blocks of Web2, such as serverless functions, VPS, and database cloud projects based on major cloud service providers like AWS, Azure, and Google Cloud.

The advantages of computing networks include:

• Blockchain technology can utilize unused computing resources and personal computers, making the network more sustainable.

• Peer-to-peer (P2P) design allows individuals to monetize unused computing resources and provide lower-cost computing, potentially reducing costs by 75%-90%.

However, due to the following challenges, computing networks struggle to be practically deployed and replace Web2 cloud services:

• While pricing is a major advantage of general-purpose computing networks, competing with established Web2 cloud companies in terms of functionality, security, and stability remains challenging.

• The peer-to-peer style may limit these networks' ability to quickly deliver mature and robust products. The decentralized nature adds to the development and maintenance costs.

1.1.2 Specialized Computing Networks

Specialized computing networks add an extra layer on top of general-purpose computing networks, allowing users to deploy specific applications through configuration files. These networks are designed to meet specific use cases, such as 3D rendering or AI inference and training.

Render is a specialized computing network focused on 3D rendering. In the AI space, new players like Bittensor, Hyperbolic, Ritual, and fetch.ai focus on AI inference, while Flock and Gensyn primarily concentrate on AI training.

IOSG Ventures' Perspective:

Advantages of specialized computing networks include:

• Decentralization and crypto features address the common issues of centralization and transparency in the AI industry.

• Permissionless computing networks and verification schemes ensure the effectiveness of inference and training processes.

• Privacy-preserving technologies, such as federated learning adopted by Flock, allow individuals to contribute data for model training while keeping their data local and private.

• By supporting smart contracts and integrating with downstream blockchain applications, AI inference can be utilized directly on the blockchain.

Source: IOSG Ventures

While dedicated AI inference and training computing networks are still in their early stages, we expect Web3 AI applications to prioritize the use of Web3 AI infrastructure. This trend has already been evident in collaborations like Story Protocol and Ritual with MyShell to introduce AI models as intellectual property.

Although killer applications built on these emerging AI x Web3 infrastructures have yet to emerge, the growth potential is immense. As the ecosystem matures, we anticipate seeing more innovative applications that leverage the unique capabilities of decentralized AI computing networks.

Part Two

2. Data

Data plays a crucial role in AI models, with various stages of AI model development involving data, including data collection, training dataset storage, and model storage.

2.1 Data Storage

Decentralized storage of AI models is essential for providing inference APIs in a decentralized manner. Inference nodes should be able to retrieve these models from anywhere at any time. As AI models may reach hundreds of GB in size, a robust decentralized storage network is needed. Leaders in the decentralized storage space, such as Filecoin and Arweave, may be able to provide this functionality.

IOSG Ventures' Perspective:

There are significant opportunities in this space.

• Decentralized data storage networks optimized for AI models, offering version control, storage of different low-precision model quantizations, and fast downloads of large models.

• Decentralized vector databases, as they are often bundled with models, provide more accurate answers by inserting necessary knowledge relevant to the questions. Existing SQL databases can also add vector search support.

2.2 Data Collection and Labeling

Collecting high-quality data is crucial for AI training. Blockchain-based projects like Grass leverage crowdsourcing to collect data for AI training, utilizing individuals' networks. With appropriate incentives and mechanisms, AI trainers can obtain high-quality data at a lower cost. Projects like Tai-da and Saipen focus on data labeling.

IOSG Ventures' Perspective:

Some observations about this market:

• Most data labeling projects are inspired by GameFi, attracting users with the "label to earn" concept and developers by promising lower costs for high-quality labeled data.

• Currently, there is no clear leader in this field, while Scale AI dominates the Web2 data labeling market.

2.3 Blockchain Data

When training AI models specifically for blockchain, developers need high-quality blockchain data that they hope to use directly in their training processes. Spice AI and Space and Time provide high-quality blockchain data with SDKs, enabling developers to easily integrate data into their training data pipelines.

IOSG Ventures' Perspective:

As the demand for AI models related to blockchain grows, the demand for high-quality blockchain data will surge. However, most data analysis tools currently only offer data export in CSV format, which is not ideal for AI training purposes.

To facilitate the development of AI models specifically for blockchain, it is crucial to enhance the developer experience by providing more blockchain-related machine learning operations (MLOps) capabilities. These capabilities should enable developers to seamlessly integrate blockchain data directly into their Python-based AI training pipelines.

Part Three

3. ZKML

Centralized AI providers face trust issues due to the motivation to use less complex models to reduce computational costs. For example, there were times last year when users felt that ChatGPT underperformed. This was later attributed to updates from OpenAI aimed at improving model performance.

Additionally, content creators have raised copyright concerns regarding AI companies. These companies find it challenging to prove that specific data was not included in their training processes.

Zero-Knowledge Machine Learning (ZKML) is an innovative approach that addresses the trust issues associated with centralized AI providers. By leveraging zero-knowledge proofs, ZKML enables developers to prove the correctness of their AI training and inference processes without disclosing sensitive data or model details.

3.1 Training

Developers can execute training tasks in a zero-knowledge virtual machine (ZKVM), such as the one provided by Risc Zero. This process generates a proof that verifies the training was conducted correctly and only used authorized data. The proof serves as evidence that developers comply with appropriate training specifications and data usage permissions.

IOSG Ventures' Perspective:

• ZKML offers a unique solution for proving the authorized use of data in model training, which is often difficult to achieve under the black-box nature of AI models.

• This technology is still in its early stages. The computational overhead is significant. The community is actively exploring more use cases for ZK training.

3.2 Inference

The time required for ZKML inference is significantly longer than for its training counterpart. Several well-known companies have emerged in this space, each adopting unique approaches to make machine learning inference trustless and transparent.

Giza focuses on building a comprehensive machine learning operations (MLOps) platform and creating a vibrant community around it. Their goal is to provide developers with tools and resources to integrate ZKML into inference workflows.

On the other hand, EZKL prioritizes developer experience by creating a user-friendly ZKML framework that delivers good performance. Their solution aims to simplify the process of implementing ZKML inference, making it easier for more developers to use.

Modulus Labs takes a different approach by developing their own proof system. Their primary goal is to significantly reduce the computational overhead associated with ZKML inference. By reducing the overhead by a factor of ten, Modulus Labs aims to make ZKML inference more practical and efficient for real-world applications.

IOSG Ventures' Perspective:

• ZKML is particularly suitable for GameFi and DeFi scenarios where trustlessness is crucial.

• The computational overhead introduced by ZKML makes it challenging for large AI models to operate efficiently.

• The industry is still seeking DeFi and GameFi pioneers that extensively use ZKML in their products to showcase practical applications.

Part Four

4. Agent Networks + Other Applications

4.1 Agent Networks

Agent networks consist of numerous AI agents equipped with tools and knowledge to perform specific tasks, such as assisting with on-chain transactions. These agents can collaborate with each other to achieve more complex goals. Several well-known companies are actively developing chatbot agents and agent networks.

Sleepless, Siya, Myshell, characterX, and Delysium are significant players building chatbot agents. Autonolas and ChainML are constructing agent networks for more robust use cases.

IOSG Ventures' Perspective:

Agents are crucial for real-world applications. They can perform specific tasks better than general AI. Blockchain offers several unique opportunities for AI agents.

• Incentive Mechanisms: Blockchain provides incentive mechanisms through technologies like non-fungible tokens (NFTs). With clear ownership and incentive structures, creators are motivated to develop more interesting and innovative agents on-chain.

• Composability of Smart Contracts: Smart contracts on the blockchain are highly composable, functioning like Lego blocks. The open APIs provided by smart contracts enable agents to perform complex tasks that are difficult to achieve in traditional financial systems. This composability allows agents to interact with various decentralized applications (dApps) and leverage their functionalities.

• Intrinsic Openness: By building agents on the blockchain, they inherit the inherent openness and transparency of these networks. This creates significant opportunities for composability between different agents, allowing them to collaborate and combine their capabilities to tackle more complex tasks.

4.2 Other Applications

In addition to the main categories discussed earlier, there are several interesting AI applications in the Web3 space that are gaining attention, even though they may not yet be large enough to form independent categories. These applications span various fields, showcasing the diversity and potential of AI within the blockchain ecosystem.

• Image Generation: ImgnAI

• Image Prompt Monetization: NFPrompt

• Community-trained AI Image Generation: Botto

• Chatbots: Kaito, Supersight, Galaxy, Knn3, Awesome QA, Qna3

• Finance: Numer AI

• Wallet: Dawn_wallet

• Gaming: Parallel TCG

• Education: Hooked

• Security: Forta

• DID: Worldcoin

• Creator Tools: Plai Lab

Part Five

5. Promoting AI x Crypto to Web2 Users for Mass Adoption

What makes AI x Crypto unique is its ability to address some of the most challenging problems in the field of artificial intelligence. While there is a gap between current AI x Crypto products and Web2 AI products, and they lack appeal to Web2 users, AI x Crypto still possesses some unique features that only it can provide.

5.1 Cost-Effective Computing Resources:

One of the main advantages of AI x Crypto is its provision of cost-effective computing resources. As demand for LLMs increases, the number of developers in the market grows, making the availability and pricing of GPUs more challenging. GPU prices have surged significantly, and shortages are common.

Decentralized computing networks, such as DePIN projects, can help alleviate this issue by utilizing idle computing power, GPUs from small data centers, and personal computing devices. While the stability of decentralized computing power may not match that of centralized cloud services, these networks offer cost-effective computing devices across diverse regions. This decentralized approach minimizes edge latency and ensures a more distributed and resilient infrastructure.

By harnessing the power of decentralized computing networks, AI x Crypto can provide Web2 users with affordable and accessible computing resources. This cost advantage is attractive for encouraging Web2 users to adopt AI x Crypto solutions, especially as the demand for AI computing continues to grow.

5.2 Empowering Creators with Ownership:

Another significant advantage of AI x Crypto is its protection of creators' ownership rights. In the current AI landscape, some agents are easily replicable. These agents can be easily copied by simply writing similar prompts. Additionally, agents in the GPT store are often owned by centralized companies rather than by the creators, limiting creators' control over their work and their ability to monetize effectively.

AI x Crypto addresses this issue by leveraging the mature NFT technology prevalent in the crypto space. By representing agents as NFTs, creators can truly own their work and derive actual revenue from it. Each time a user interacts with an agent, the creator can receive incentives, ensuring fair compensation for their efforts. The concept of NFT ownership applies not only to agents but also to protecting other important assets in the AI space, such as knowledge bases and prompts.

5.3 Protecting Privacy and Rebuilding Trust:

Users and creators have privacy concerns regarding centralized AI companies. Users worry that their data may be misused for training future models, while creators fear their work may be used without proper attribution or compensation. Furthermore, centralized AI companies may sacrifice service quality to reduce infrastructure costs.

These issues are difficult to address with Web2 technologies, while AI x Crypto leverages advanced Web3 solutions. Zero-knowledge training and inference can provide transparency by proving the data used and ensuring the correct models are applied. Technologies such as Trusted Execution Environments (TEE), federated learning, and Fully Homomorphic Encryption (FHE) enable secure, privacy-preserving AI training and inference.

By prioritizing privacy and transparency, AI x Crypto allows AI companies to regain public trust and offer AI services that respect user rights, distinguishing them from traditional Web2 solutions.

5.4 Tracking Content Origins:

As AI-generated content becomes increasingly sophisticated, distinguishing between human-created and AI-generated text, images, or videos becomes more challenging. To prevent the misuse of AI-generated content, a reliable method for determining the origin of content is needed.

Blockchain excels at tracking content origins, as evidenced by its success in supply chain management and NFTs. In the supply chain industry, blockchain tracks the entire lifecycle of products, allowing users to identify manufacturers and key milestones. Similarly, blockchain tracks creators and prevents piracy in the case of NFTs, which are particularly susceptible to piracy due to their transparency. Despite this vulnerability, utilizing blockchain can minimize losses caused by fake NFTs, as users can easily distinguish between genuine and counterfeit tokens.

By applying blockchain technology to track the origins of AI-generated content, AI x Crypto can provide users with the ability to verify whether content creators are AI or human, thereby reducing the likelihood of misuse and increasing trust in content authenticity.

5.5 Utilizing Cryptocurrency to Develop Models:

Designing and training models, especially large ones, is an expensive and time-consuming process. New models also come with uncertainties, as developers cannot predict their performance.

Cryptocurrency offers a developer-friendly way to gather pre-training data, collect reinforcement learning feedback, and raise funds from interested parties. This process resembles the lifecycle of a typical cryptocurrency project: raising funds through private investments or launchpads and distributing tokens to active contributors at launch.

Models can adopt a similar approach by raising funds for training through token sales and airdropping tokens to contributors of data and feedback. With a well-designed tokenomics model, this workflow can help individual developers train new models more easily than ever before.

Part Six

6. Challenges of Tokenomics

AI x Crypto projects are beginning to target Web2 developers as potential customers, as crypto has a unique value proposition and the Web2 AI industry has a substantial market size. However, for Web2 developers unfamiliar with tokens and reluctant to engage with token-based systems, tokens may pose a barrier.

To cater to Web2 developers, reducing or eliminating the utility of tokens may cause concern for Web3 enthusiasts, as this could alter the fundamental stance of AI x Crypto projects. Striking a balance between attracting Web2 developers and maintaining token utility while integrating valuable tokens into AI SaaS platforms is a challenging task.

To bridge the gap between Web2 and Web3 business models while preserving token value, several potential approaches can be considered:

• Utilizing tokens within the project's distributed infrastructure network. Implementing staking, rewards, and penalty mechanisms to protect the foundational network.

• Using tokens as a payment method while providing access for Web2 users.

• Implementing token-based governance.

• Sharing profits with token holders.

• Utilizing profits to buy back or burn tokens.

• Offering discounts and additional features for token holders for services provided by the project.

By carefully designing tokenomics models that align with the interests of both Web2 and Web3, AI x Crypto projects can successfully attract Web2 developers while maintaining the value and utility of their tokens.

Part Seven

7. Our Favorite AI x Crypto Scenarios

Our favorite AI x Crypto scenarios leverage the power of user collaboration to accomplish tasks in the AI domain through blockchain technology. Some specific examples include:

1. Collective data contributions for AI training, alignment, and benchmarking (e.g., Chatbot Arena)

2. Collaboratively building a large shared knowledge base for various agents to use (e.g., Sahara)

3. Utilizing individual resources for web data scraping (e.g., Grass)

By harnessing user collective efforts based on blockchain incentives and coordination, these models demonstrate the potential of decentralized, community-driven approaches to AI development and deployment.

Part Eight

Conclusion

We are at the dawn of AI and Web3, and compared to other industries, the integration of artificial intelligence and blockchain is still in its early stages. Among the top 50 Gen AI products, there are no products related to Web3. The leading LLM tools are primarily focused on content creation and editing, targeting sales, meetings, and note-taking/knowledge bases. Given the vast amount of research, documentation, sales, and community work within the Web3 ecosystem, there is tremendous potential for the development of customized LLM tools.

Currently, developers are focused on building infrastructure to bring advanced AI models on-chain, although we have yet to reach our goals. As we continue to develop this infrastructure, we are also exploring the best user scenarios for conducting AI inference on-chain in a secure and trustless manner, which presents unique opportunities for the blockchain space. Other industries can directly utilize existing LLM infrastructure for inference and fine-tuning. Only the blockchain industry requires its own native AI infrastructure.

In the near future, we expect blockchain technology to leverage its peer-to-peer advantages to tackle the most challenging problems in the AI industry, making AI models more affordable, accessible, and profitable for everyone. We also anticipate that the crypto space will follow the narrative of the AI industry, albeit with a slight delay. Over the past year, we have witnessed developers combining crypto, agents, and LLM models. In the coming months, we may see more multimodal models, text-to-video generation, and 3D generation impacting the crypto space.

The entire AI and Web3 industry is currently not receiving adequate attention, and we eagerly await the moment when AI ignites within Web3, leading to a killer application of CryptoxAI.