Original author: Wanxiang Blockchain

Original source: Wanxiang Blockchain

Preface

In the past year, with the birth of large generative AI models such as ChatGPT, AI has expanded from simple automation tools to complex decision-making and prediction systems, and has developed into an important driving force for progress in contemporary society. AI products and applications have also experienced explosive growth. ChatGPT itself has successively launched GPTs, Sora and other eye-catching products. The performance of NVIDIA, the underlying AI infrastructure, continues to exceed expectations. In the fourth quarter of fiscal year 2024, the data center business accounted for more than 83% of the revenue. revenue, a simultaneous growth of 409%, of which 40% was used for large model inference scenarios, showing the rapid growth in demand for underlying computing power.

At present, AI has become a subject that European and American capital circles are chasing. At the same time, the Web3 market has also ushered in a new round of bull market. AI+Web3 is the collision of the two most popular technical themes at the moment. A number of projects on this theme have also appeared recently. , highlighting the market’s concern and expectations for this theme.

Putting aside the hype and price bubbles, how is the current development of the AI+Web industry? Are there real application scenarios? In the long run, can we create value, narrative and industry? What kind of ecological pattern will the AI+Web3 industry form in the future, and what are the potential directions?

Focusing on the above topics, Future3 Campus will write a series of related articles to analyze all aspects of the AI+Web3 industry chain. This article is the first one, covering the overall industry picture and narrative logic of AI+Web3.

AI work production process

In summary, the direction of the combination of AI+Web3 can be divided into two aspects. On the one hand, how Web3 helps the development of AI, and on the other hand, the combination of Web3 applications with AI technology. Among them, Web3 technology and concepts empowering AI is the direction of most current projects. Therefore, we can analyze how to integrate AI with Web3 through the process from model training to production. The birth of LLM has some differences from the previous machine learning process, but in general, a simplified AI production process is roughly divided into the following stages:

1

data collection

In the entire life cycle of AI model training, data is the cornerstone of AI model training. It is often necessary to use high-quality data sets as a basis and perform exploratory data analysis (EDA) to create reproducible, editable and shareable data sets, tables and visualizations.

2

Data preprocessing and feature engineering/hint engineering

After obtaining the data, the data needs to be preprocessed. This is feature engineering (data annotation) in machine learning and prompt engineering in large models. This includes iteratively classifying, aggregating and deduplicating data to mark fine features, and iteratively developing prompts for LLM structured queries. At the same time, features/Prompts need to be stored and shared reliably.

3

Model training and tuning

Use a rich model library to train AI models, and improve the performance, efficiency and accuracy of the model through continuous iteration and adjustment. Among them, in LLM, the model is continuously tuned through human feedback reinforcement learning (RLHF).

4

Model review and governance

Use the MLOps/LLMOps platform to optimize the model development process, including model discovery, tracking, sharing and collaboration, ensuring model quality and transparency while complying with ethical and compliance requirements.

5

Model reasoning

Deploy trained AI models to make predictions on new, unseen data. The model uses its learned parameters to process input data and generate prediction results, such as classification or regression predictions.

6

Model deployment and monitoring

After ensuring that the model performance meets the standards, deploy it into actual application scenarios and implement continuous monitoring and maintenance to ensure that the model maintains optimal performance in a dynamically changing environment.

In the above process, there are many opportunities to combine Web3 with it. At present, we see that some challenges in the development process of AI, such as model transparency, bias and ethical application, have attracted widespread attention. In this regard, Web3 technology combined with cryptography technologies such as ZK can improve the trust issue of AI. In addition, the increasing demand for AI applications has also put forward requirements for lower-cost and more open infrastructure and data networks, and Web3's distributed network and incentive model can also create more open and open source AI networks and communities.

AI+Web3 industrial landscape and narrative logic

Combining the above-mentioned AI production process and the direction of combining AI with Web3, as well as the mainstream AI+Web3 projects in the current market, we have sorted out the AI+Web3 industry picture. The AI+Web3 industry chain can be divided into three layers, namely the infrastructure layer. , middle layer and application layer.

1

infrastructure layer

It mainly includes computing and storage infrastructure, which runs through the entire AI work and production process, providing the computing power required for AI model training, speculation, etc., as well as the storage of data and models throughout the life cycle.

The current rapid growth of AI applications has led to a surge in demand for infrastructure, especially high-performance computing power. Therefore, providing higher performance, lower cost, and more sufficient computing and storage infrastructure will become a very important trend in the next few years (in the early stages of AI development), and is expected to capture more than 50% of the industry chain value.

Web3 can create a decentralized computing and storage resource network, using idle and dispersed resources to significantly reduce infrastructure costs and serve a wide range of AI application needs. Therefore, decentralized AI infrastructure is currently the most certain narrative.

Currently, representative projects in this track include Render Network, which focuses on rendering services, and Akash, gensyn, etc., which provide decentralized cloud services and computing hardware networks; in the storage field, representative projects are still the old decentralized storage network Filecoin, Arweave and others have recently launched storage and computing services for the AI ​​field.

2

middle layer

It mainly refers to the use of Web3 related technologies to improve the current situation and existing problems in the specific process of AI work production. mainly include:

1) In the data acquisition stage, decentralized data identity is used to create a more open data network/data trading platform. It mainly protects users and confirms data by combining cryptography technology and blockchain features, and combines incentives to encourage users to share high-quality data, thereby expanding data sources and improving data acquisition efficiency. Representative projects in this field include AI identity projects Worldcoin, Aspecta, data trading platform Ocean Protocol, and data network Grass with low participation threshold, etc.

2) Data preprocessing stage: It mainly creates a distributed AI data annotation and processing platform, and uses economic model incentives to encourage crowdsourcing models to promote more efficient and lower-cost data preprocessing and serve the subsequent model training stage. Representative projects such as Public AI, etc.

3) Model verification and inference stage: As mentioned in the previous section, data and model black boxes are currently real problems in AI. Therefore, in the model verification and inference stage, Web3 can combine cryptography technologies such as ZK and homomorphic encryption. To verify the model's reasoning, whether the given data and parameters are used, to ensure the correctness of the model, while protecting the privacy of the input data. A typical application scenario is ZKML. Currently, representative projects that combine Web3 technology in the model verification and inference stages include bittensor, Privasea, Modulus, etc.

Many projects in the middle layer are more focused on developer tools, usually providing additional services to existing developers, project parties, etc. In the early stages of AI development, its market demand and commercial implementation are still in the process of development.

3

Application layer

At the application level, it is more about how AI technology is applied to Web3. Web3 applications combined with AI technology can effectively improve efficiency and product experience. For example, using AI's functions such as content generation, analysis, and speculation, it can be applied to various fields such as games, social networking, data analysis, and financial forecasting. At present, AI+Web3 applications can be mainly divided into three categories. 1) AIGC type, which uses AI generative technology to allow users to generate text, pictures, videos, Avatar and other content through dialogue. Displayed as a separate AI agent or directly integrated into the product. Representative projects include NFPrompt, SleeplessAI, etc. 2) AI analysis category, the project party integrates its own accumulated data, knowledge base, analysis capabilities, etc. to train a vertical AI model that can perform analysis, judgment, prediction, etc., and provides it to users as a product, so that users can obtain it at a low threshold AI analysis capabilities, such as data analysis, information tracking, code audit and modification, financial forecasting, etc. Representative projects include Kaito, Dune, etc.

3) AI Agent Hub, an aggregation of various AI Agents, usually provides users with the ability to create customized AI Agents without code, similar to GPTs. Representative projects include My Shell, Fetch.ai, etc.

There are currently no top projects in the application layer, but in the long run it will definitely be a sector with a higher ceiling, with strong potential that remains to be tapped. The competition of AI+Web3 applications does not lie in the innovative competitiveness of technology, but in the accumulation of product capabilities and technical capabilities. Especially in terms of AI, products that can provide better experience will gain more competitive advantages in this field.