AI+Web3: Towers and Squares

Introduction

In the past two years, the speed of AI development has noticeably accelerated. The wave of generative artificial intelligence initiated by ChatGPT has not only opened the door to a new world but has also stirred ripples in the Web3 field.

Driven by the concept of AI, financing activities in the crypto market have clearly rebounded. According to statistics, in the first half of 2024 alone, 64 Web3+AI projects completed financing, among which the AI-based operating system Zyber365 set a record of 100 million USD in its Series A financing.

The secondary market is more active, and data from crypto aggregation sites shows that in just over a year, the total market value of the AI sector has reached $48.5 billion, with a 24-hour trading volume nearing $8.6 billion. Mainstream AI technological advancements have brought significant benefits; after the release of OpenAI's Sora text-to-video model, the average price of AI stocks rose by 151%. The AI effect has also spread to the cryptocurrency fundraising sector, with the first AI Agent concept MemeCoin - GOAT quickly gaining popularity and achieving a valuation of $1.4 billion, sparking an AI Meme craze.

The research and discussion on AI+Web3 is equally heated, with new concepts emerging one after another, from AI+Depin to AI Memecoin, and now to AI Agent and AI DAO.

The combination of AI and Web3, filled with hot money, trends, and future imagination, is inevitably seen as a marriage orchestrated by capital. It is difficult for us to discern beneath this glamorous exterior whether it is the revelry of speculators or the dawn of a new era.

To answer this question, the key lies in considering whether both sides can promote each other. This article attempts to examine this pattern: how Web3 can play a role in various aspects of the AI technology stack, and what new opportunities AI can bring to Web3?

Part.1 What opportunities does Web3 have under the AI stack?

Before we start the discussion, we need to understand the technology stack of AI large models:

AI large models can be compared to the human brain; in the early stages, they need to absorb massive amounts of external information to understand the world, which is the "data collection" phase. Since computers do not have human senses, "preprocessing" is required before training to convert unlabeled information into a usable format.

After inputting data, the AI builds a model with understanding and prediction capabilities through "training", similar to the process of a baby learning to understand the outside world. When the learning content is categorized or feedback is obtained through communication and corrections are made, it enters the "fine-tuning" stage.

As children grow up, they can understand and express their thoughts in conversations, similar to the "reasoning" stage of AI large models, which can predict and analyze new inputs. AI expresses feelings, describes things, and solves problems through language abilities, similar to how large models are applied to specific tasks such as image classification and speech recognition after training.

The AI Agent is closer to the next form of large models - capable of independently executing tasks and pursuing complex goals, possessing not only the ability to think but also to remember, plan, and interact with the world using tools.

Currently, in response to the pain points of various AI stacks, Web3 has initially formed a multi-layered, interrelated ecosystem that covers all stages of the AI model process.

1. Basic Layer: Computing Power and Data's Airbnb

Hash Rate

Currently, one of the main costs of AI is the computational power and energy required for training and inference models.

For example, Meta's LLAMA3 requires 16,000 NVIDIA H100 GPUs for 30 days to complete training. The H100 80GB version costs $30,000 to $40,000 each, requiring a hardware investment of $400 million to $700 million, and the monthly training consumes 1.6 billion kilowatt-hours, with energy expenses nearing $20 million.

To alleviate the pressure of AI computing power, DePin(, a decentralized physical infrastructure network), has become one of the earliest intersections of Web3 and AI. The DePin Ninja data website has listed over 1,400 projects, among which representative projects for GPU computing power sharing include io.net, Aethir, Akash, Render Network, and others.

The main logic is: the platform allows owners of idle GPU resources to contribute computing power in a decentralized manner, improving resource utilization through an online market similar to Uber or Airbnb, while end users obtain efficient computing resources at a lower cost; at the same time, the staking mechanism ensures penalties for violators.

Features include:

• Gather idle GPU resources: mainly from excess computing power of small and medium-sized data centers, crypto mining farms, and mining hardware for PoS consensus mechanisms. Some projects like exolab also utilize local devices such as MacBooks and iPhones to establish inference computing power networks.

• Targeting the long-tail market for AI computing power: a. Technical side: More suitable for inference steps. Training relies on ultra-large-scale GPU clusters, while inference has lower GPU performance requirements. b. Demand Side: Small and medium computing power demanders mostly focus on optimizing and fine-tuning large leading models, which is naturally suitable for distributed idle computing power.

• Decentralized Ownership: Blockchain technology ensures that resource owners retain control, can adjust flexibly, and receive benefits.

Data

Data is the cornerstone of AI. Without data, computation is meaningless; the quality of data determines the quality of model outputs. For current AI model training, data determines language ability, comprehension ability, values, and human-like performance. Currently, the data demand dilemma for AI is mainly reflected in:

• Data hunger: AI model training relies on massive data input. OpenAI trained GPT-4 with a parameter count reaching trillions.

• Data Quality: With the integration of AI across various industries, the requirements for data timeliness, diversity, professionalism, and emerging data sources such as social media sentiment have increased.

• Privacy and Compliance: Countries and companies are gradually restricting data set crawling.

• High data processing costs: large data volumes and complex processing. AI companies spend over 30% of their R&D costs on basic data collection and processing.

The solutions of Web3 are mainly reflected in:

1. Data Collection: Involve genuine contributors in value creation, obtaining more private and valuable data at a low cost through a distributed network and incentive mechanisms.

   • Grass: A decentralized data layer and network where users run nodes to contribute bandwidth, capturing real-time data to earn rewards.
   • Vana: Introduce the concept of data liquidity pool (DLP), where users can upload private data and flexibly authorize third parties to use it.
   • PublicAI: Users can collect data by using the #AI或# Web3 tag on X and @PublicAI.

2. Data preprocessing: The AI industry has few manual processes, making it suitable for Web3 decentralized incentive mechanisms.

   • Grass and OpenLayer are considering adding a data labeling phase.
   • Synesis proposed the concept of "Train2earn", rewarding the provision of high-quality labeled data.
   • Sapien gamifies the marking task, allowing users to stake points to earn more points.

3. Data Privacy and Security: The advantages of Web3 privacy technology are reflected in the training of sensitive data and multi-party data collaboration.

   The main privacy technologies include:

   • Trusted Execution Environment ( TEE ), such as Super Protocol
   • Fully Homomorphic Encryption ( FHE ), such as BasedAI, Fhenix.io, Inco Network
   • Zero-knowledge technology ( zk ), such as Reclaim Protocol using zkTLS technology

Currently still in the early stages, facing challenges such as high computing costs.

4. Data Storage: Addressing the issue of storing AI data on-chain and generating LLM.

   • 0g.AI: A centralized storage solution designed for high-performance AI needs, supporting fast upload and download of large-scale datasets, with transfer speeds approaching 5GB/sec.

2. Middleware: Model Training and Inference

Open Source Model Decentralized Market

Web3 proposes the possibility of a decentralized open-source model marketplace, retaining a portion of tokens for the team through tokenization, directing part of the future revenue streams of the model to token holders.

• Bittensor: Establish an open-source model P2P market composed of multiple "subnets", where resource providers compete to meet the goals of the subnets.
• ORA: Introduce the initial model issuance of (IMO) concept, tokenizing AI models.
• Sentient: A decentralized AGI platform that incentivizes collaboration to build AI models and rewards contributors.
• Spectral Nova: Focus on the creation and application of AI and ML models.

Verifiable Inference

To address the "black box" problem of AI inference, the Web3 standard solution is to compare results through repeated operations by multiple validators, but it faces the challenge of high costs.

A more promising solution is to perform ZK proofs for off-chain AI inference calculations and verify the computations on-chain. Main advantages:

• Scalability: Rapid confirmation of large-scale off-chain computations.
• Privacy Protection: Protect data and model details.
• Trustless: No need to rely on centralized parties to confirm calculations.
• Web2 Integration: Helps improve Web3 adoption rate.

Current verifiable technologies include:

• zkML: Combines zero-knowledge proofs and machine learning, such as the AI prover released by Modulus Labs based on ZKML.
• opML: Improve ML computing efficiency by utilizing the optimistic aggregation principle.
• TeeML: Securely executing ML computations using Trusted Execution Environments.

3. Application Layer: AI Agent

The current focus of AI development has shifted from model capabilities to AI Agents. OpenAI defines an AI Agent as a system driven by LLM, capable of independent understanding, perception, planning, memory, and tool usage, and able to automatically execute complex tasks.

Web3 can bring to Agents:

Decentralized

The characteristics of Web3 make the Agent system more decentralized and autonomous, establishing incentive and punishment mechanisms to promote democratization through PoS, DPoS, and other mechanisms, such as GaiaNet, Theoriq, and HajimeAI.

Cold Start

Web3 helps the potential AI Agent project secure early-stage financing.

• Virtual Protocol launched the AI Agent creation and token issuance platform fun.virtuals.
• Spectral proposed the concept of IAO( Initial Agent Offering) to support the issuance of on-chain AI Agent assets.

Part.2 How does AI empower Web3?

AI has a significant impact on Web3 projects by optimizing on-chain operations ( such as smart contract execution, liquidity optimization, and AI-driven governance decisions ), benefiting blockchain, providing data-driven insights, enhancing on-chain security, and laying the foundation for new Web3 applications.

1. AI and On-chain Finance

AI and Crypto Economy

Coinbase CEO announced the first AI-to-AI cryptocurrency transaction on the Base network, where AI Agents can use USD to trade with humans, merchants, or other AIs.

The Luna demonstration of Virtuals Protocol showcases AI agents autonomously executing on-chain transactions, with AI agents seen as the future of on-chain finance. Potential scenarios include:

1. Information Collection and Prediction: Collect exchange announcements, project information, public opinion risks, etc., analyze and assess the fundamentals of assets and market conditions, forecast trends and risks.

2. Asset Management: Provide investment targets, optimize asset portfolios, and execute trades automatically.

3. Financial Experience: Choose the fastest on-chain transaction method, automate cross-chain, adjust gas fees, etc., to lower the threshold and cost of on-chain financial activities.

Currently, AI Agent wallets like Bitte and the AI interaction protocol Wayfinder are attempting to integrate with the OpenAI model API, allowing users to command Agents to complete on-chain operations through a chat interface. The decentralized Agent platform Morpheus supports the development of such Agents, and Biconomy demonstrated that AI Agents can perform swap operations without full wallet permissions.

AI and On-Chain Transaction Security

AI technology can enhance the security and privacy protection of on-chain transactions, with potential scenarios including:

• Transaction Monitoring: Real-time monitoring of abnormal activities, providing alerts.
• Risk Analysis: Analyze customer trading behavior and assess risks.

The Web3 security platform SeQure utilizes AI to detect and prevent attacks, fraud, and data breaches, providing real-time monitoring and alerts. Similar tools include AI-powered Sentinel.

2. AI and On-Chain Infrastructure

AI and On-Chain Data

AI technology plays an important role in on-chain data collection and analysis, such as:

• Web3 Analytics: An AI-based analytics platform that analyzes on-chain data using machine learning and data mining algorithms.
• MinMax AI: Provides AI-based on-chain data analysis tools to discover market opportunities and trends.
• Kaito: A Web3 search platform based on LLM.
• Followin: Integrate ChatGPT, consolidate information from different platforms.

AI can also be used for oracle services, such as Upshot using AI to provide accurate pricing data for NFTs.

AI and Development & Audit

AI can improve Web3 development efficiency and lower the programming threshold. Potential scenarios include: automated code generation, smart contract verification testing, DApp deployment and maintenance, intelligent code completion, answering development questions, etc.

Currently, there are one-click launch token platforms like Clanker and contract development platforms like Spectral that offer one-click generation and deployment functions for smart contracts.

In terms of auditing, the Web3 auditing platform Fuzzland uses AI to assist in checking code vulnerabilities and provides natural language explanations.

3. AI and