AI+Web3: Towers and Squares

key points

1. Web3 projects with AI concepts have become targets for capital attraction in the primary and secondary markets.

2. The opportunities for Web3 in the AI industry are mainly reflected in: utilizing distributed incentives to coordinate long-tail potential supply across data, storage, and computing (; while establishing a decentralized marketplace for open-source models and AI Agents.

3. AI is mainly applied in the Web3 industry for on-chain finance ), crypto payments, trading, data analysis (, and assisting development.

4. The value of AI + Web3 lies in the complementarity of the two: Web3 is expected to address the centralization of AI, while AI is expected to help Web3 break through layer limitations.

) Introduction

In the past two years, the development of AI has accelerated, and the generative AI boom triggered by ChatGPT has not only opened up a new world but also created waves in the Web3 space.

With the support of the AI concept, financing in the cryptocurrency market has obviously warmed up. According to statistics, in the first half of 2024 alone, 64 Web3+AI projects completed financing, among which Zyber365 achieved the highest financing of 100 million USD in Series A.

The secondary market is more active. According to Coingecko data, the total market value of the AI sector has reached 48.5 billion USD, with a 24-hour trading volume of nearly 8.6 billion USD. The progress of mainstream AI technology has brought good news; for example, after OpenAI released Sora, the average increase in the AI sector was 151%. The AI effect has also impacted the popular cryptocurrency sector, Meme, with the first AI Agent concept MemeCoin—GOAT rapidly rising to fame, achieving a valuation of 1.4 billion USD, triggering an AI Meme craze.

Research and topics related to AI+Web3 continue to gain traction, from AI+Depin to AI Memecoin and now to the current AI Agent and AI DAO, with new narratives emerging one after another.

This combination filled with hot money, opportunities, and future imaginations is inevitably seen as a marriage arranged by capital, making it difficult for us to determine whether it is a carnival of speculators or an explosion before dawn.

The key is whether both parties can benefit from each other's models. This article will explore how Web3 can play a role in various aspects of AI technology, and what new opportunities AI can bring to Web3.

![AI+Web3: Towers and Squares]###https://img-cdn.gateio.im/webp-social/moments-25bce79fdc74e866d6663cf31b15ee55.webp(

) 1. Opportunities of Web3 under the AI Stack

Before discussing this topic, we need to understand the technology stack of AI large models:

In simple terms, a "large model" is similar to the human brain, initially resembling an infant that needs to observe and absorb vast amounts of information to understand the world; this is the data "collection" phase. Due to the lack of human multi-sensory perception in computers, unlabelled information must be converted into a computer-understandable format through "preprocessing" before training.

After inputting data, the AI builds a model with understanding and predictive capabilities through "training", similar to how a baby gradually learns to understand the outside world. The model parameters adjust like a baby's evolving language abilities. When the learning content is categorized or feedback is received from interactions with others, it enters the "fine-tuning" stage.

Once children grow up and learn to speak, they can understand and express ideas in new conversations, similar to the "reasoning" of large AI models, which can perform predictive analysis on new language text inputs. Infants use language to express feelings, describe objects, and solve problems, similar to how large AI models apply reasoning after training to various specific tasks, such as image classification, speech recognition, and more.

The AI Agent is closer to the next form of large models—capable of independently executing tasks to pursue complex goals, equipped with thinking, memory, and planning abilities, and able to interact with the world using tools.

In response to the pain points of AI stacks, Web3 has currently formed a multi-level interconnected ecosystem that covers various stages of the AI model process.

![AI+Web3: Tower and Square]###https://img-cdn.gateio.im/webp-social/moments-cc3bf45e321f9b1d1280bf3bb827d9f4.webp(

)# 1. Base Layer: Airbnb for Computing Power and Data

Hash Rate

One of the main costs of AI is the computing power and energy required for training and inference of models.

For example, Meta's LLAMA3 requires 16,000 NVIDIA H100 GPUs for 30 days to complete training. The unit price of the 80GB version is 30,000 to 40,000 USD, requiring a hardware investment of 400 to 700 million USD ### GPUs + network chips (, with a monthly training electricity consumption of 1.6 billion kilowatt-hours, and energy expenses of nearly 20 million USD.

In response to the AI computing power pressure, DePin) decentralized physical infrastructure network( is one of the earliest areas where Web3 intersects with AI. DePin Ninja has listed over 1,400 projects, with GPU computing power sharing representatives such as io.net, Aethir, Akash, Render Network, and others.

The main logic is: the platform allows idle GPU resource providers to contribute computing power in a permissionless decentralized manner, creating an online marketplace for buyers and sellers similar to Uber/Airbnb, improving GPU utilization, and users obtain low-cost efficient computing resources; at the same time, the staking mechanism ensures that violations of quality control or network interruptions will be penalized.

Features:

• Gather idle GPUs: Supply mainly comes from small and medium-sized data centers, excess computing power from crypto mining farms, and PoS mining hardware such as FileCoin/ETH miners. Projects like exolab are dedicated to establishing a computing power network for running large model inference using local devices such as MacBooks, iPhones, and iPads.

• Targeting the long-tail market of AI computing power: The technical side is more suitable for inference steps. Training relies on ultra-large GPU clusters, while inference has lower GPU requirements, such as Aethir focusing on low-latency rendering and AI inference. Small and medium power demanders on the demand side will not train large models individually, mainly focusing on optimizing and fine-tuning leading models, which is suitable for distributed idle computing power.

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

Data

Data is the foundation of AI. Without data, computation is useless, and the quality of data determines the quality of model output. For AI model training, data determines language ability, understanding ability, values, and human-like performance. The current challenges in AI data demand mainly include:

• Data hunger: AI model training requires massive amounts of data. GPT-4 has a parameter count in the trillions.

• Data Quality: The integration of AI with various industries raises new requirements for data timeliness, diversity, professionalism, and emerging data sources such as social media sentiment.

• Privacy compliance: Companies in various countries are gradually restricting data set scraping.

• High processing costs: large data volume and complex processing. AI companies spend more than 30% of their R&D costs on data collection and processing.

Web3 Solutions:

1. Data Collection: Free real-world data scraping is dwindling, and AI companies' data expenditures are increasing year by year, but they are not rewarding the true contributors. The vision of Web3 is to allow contributing users to participate in value creation, incentivizing low-cost acquisition of more private and valuable data through a distributed network.

• Grass: A decentralized data layer network where users run nodes to contribute bandwidth and capture real-time data for token rewards.

• Vana: Introduce the concept of data liquidity pool )DLP(, where users can upload private data and choose to authorize third parties to use it.

• PublicAI: Users utilize the ) Web3 tag on X and @PublicAI for data collection.

2. Data Preprocessing: AI data processing requires cleaning and converting into a usable format, involving standardization, filtering, handling missing values, and other repetitive tasks. This manual step has led to the emergence of the data labeling industry, with increasing requirements raising the threshold, which is suitable for Web3 decentralized incentive mechanisms.

• Grass and OpenLayer are considering adding a data annotation phase.

• Synesis introduces the "Train2earn" concept, emphasizing data quality, where users provide labeled data to earn rewards.

• Sapien gamifies the task of marking, allowing users to stake points to earn more points.

3. Data Privacy Security: Data privacy involves the handling of sensitive data and the protection of data security information from unauthorized access, destruction, and theft. The advantages of Web3 privacy technology are reflected in: #AI或# sensitive data training; ( data collaboration: multiple data owners can participate in AI training without sharing the original data.

Main privacy technologies:

• 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 to generate zero-knowledge proofs for HTTPS traffic, securely importing external website data.

Currently in the early stages, the main dilemma is the high computing costs:

• EZKL takes 80 minutes to generate a 1M-nanoGPT model proof.

• zkML costs over 1000 times more than pure computation.

4. Data Storage: On-chain storage of data and generated LLM is required. Data availability )DA( is a core issue, with Ethereum's Danksharding upgrade having a throughput of 0.08MB before the upgrade, while AI model training inference typically requires 50-100GB per second.

• 0g.AI is a centralized storage solution designed for AI needs, characterized by high-performance scalability, supporting fast upload and download of large-scale datasets through sharding and erasure coding, with a transmission speed close to 5GB per second.

)# 2. Middleware: Model Training and Inference

Open Source Model Decentralized Market

The controversy over open-source AI models continues. While open-source brings the advantage of collective innovation, how can developer motivation be increased without a profit model? Li Yanhong once asserted that "open-source models will fall further behind."

Web3 proposes the possibility of a decentralized open-source model market: tokenization of models, teams retaining a portion of tokens, directing part of future revenue streams to token holders.

• Bittensor establishes an open-source model P2P market, composed of multiple "subnets", where resource providers compete to meet the goals of the subnets. Each subnet interacts and learns to achieve stronger intelligence. Rewards are distributed through community voting, allocated based on performance within the subnet.

• ORA introduces the initial model issuance (IMO) concept, tokenizing AI models that can be bought, sold, and developed through a decentralized network.

• Sentient decentralized AGI platform that incentivizes collaboration to build replicable and scalable AI models, rewarding contributors.

• Spectral Nova focuses on the creation of applications using AI and ML models.

Verifiable Reasoning

To address the "black box" problem of AI inference, the standard Web3 solution is to compare results through multiple validator repeat operations, but the shortage of high-end GPUs leads to high costs.

A more promising solution is to execute ZK proofs for off-chain AI inference computations and verify the AI model calculations on-chain. It is necessary to encrypt the proofs on-chain to ensure that the off-chain computations are completed correctly ### as long as the dataset has not been tampered with (, while ensuring data confidentiality.

Main advantages:

• Scalability: ZK proofs can quickly confirm a large number of off-chain computations. Even with an increase in transactions, a single ZK proof can verify all transactions.

• Privacy protection: Data and AI model details are kept confidential, while all parties can verify that they have not been tampered with.

• Trustless: No need to rely on centralized parties to verify computations.

• Web2 Integration: Web2 is essentially off-chain integration, and verifiable reasoning can help bring datasets and AI computations on-chain, increasing Web3 adoption.

Current Web3 verifiable reasoning technology:

• zkML: Combines zero-knowledge proofs and machine learning to ensure the privacy of data models, allowing verifiable computation without revealing underlying attributes. Modulus Labs has released an AI-built ZK prover based on ZKML to check whether AI providers correctly execute algorithms, with current clients mainly being on-chain DApps.

• opML: Utilizing the optimistic aggregation principle, it improves the scalability efficiency of ML calculations by verifying the occurrence time of disputes. Only a small portion of the "verifier" results needs to be validated, but setting a high economic cost penalty increases the cost of cheating and saves redundant calculations.

• TeeML: Securely execute ML computations using trusted execution environments to protect data models from tampering and unauthorized access.

)# 3. Application Layer: AI Agent

The focus of AI development is shifting from model capabilities to AI Agents. OpenAI, Anthropic, Microsoft, and others are all developing AI Agents in an attempt to break through the platform phase of LLM technology.

OpenAI defines an AI Agent as a system driven by an LLM as its brain, capable of autonomous understanding, perception, planning, memory, and tool usage, which can automate the execution of complex tasks. AI transforms from a tool being used into a subject that can utilize tools, becoming the ideal intelligent assistant.

Web3 can bring to Agents:

Decentralized

The decentralized characteristics of Web3 make the Agent system more distributed and autonomous. By establishing a staking delegation incentive and punishment mechanism through PoS, DPoS, and other mechanisms, it promotes the democratization of the Agent system. GaiaNet, Theoriq, and HajimeAI have all made attempts.

Cold Start

The development and iteration of AI Agents require substantial funding, and Web3 can assist potential projects in obtaining early-stage financing for cold starts.

• Virtual Protocol has launched the AI Agent to create the token issuance platform fun.virtuals, allowing users to deploy the AI Agent with one click to achieve 100% fair token issuance.

• Spectral proposes the concept of supporting the issuance of on-chain AI Agent asset products: by issuing tokens through IAO( Initial Agent Offering ), AI Agents can directly obtain investment funds, become members of DAO governance, and provide investors with opportunities to participate in project development and share in the profits.

2. How AI Empowers Web3

AI has a significant impact on Web3 projects by optimizing on-chain operations such as smart contract execution and liquidity.