Can AI Agents Become a Lifeline for Web3 + AI?

AI Agent projects are primarily popular and mature types in Web2 entrepreneurship that focus on enterprise services, while in the Web3 field, model training and platform aggregation projects have become mainstream due to their key role in building ecosystems.

Currently, the number of AI Agent projects in Web3 is limited, accounting for 8%, but their market capitalization in the AI sector reaches as high as 23%, demonstrating strong market competitiveness. We anticipate that as technology matures and market recognition increases, multiple projects with valuations exceeding 1 billion dollars will emerge in the future.

For Web3 projects, introducing AI technology may become a strategic advantage for application-end products that are not core to AI. The integration of AI Agent projects should focus on building a complete ecosystem and designing a token economic model to promote decentralization and network effects.

The AI Wave: Current Status of Project Emergence and Valuation Increases

Since the launch of ChatGPT in November 2022, it attracted over 100 million users in just two months. By May 2024, ChatGPT's monthly revenue reached an astonishing $20.3 million. Following the release of ChatGPT, OpenAI quickly launched iterative versions such as GPT-4 and GP4-4o. With such rapid momentum, major traditional tech giants have realized the importance of cutting-edge AI model applications like LLMs, and have launched their own AI models and applications. For example, Google released the large language model PaLM2, Meta launched Llama3, while Chinese companies introduced models like Wenxin Yiyan and Zhipu Qingyan. Clearly, the AI field has become a battleground that everyone is vying for.

The competition among major tech giants has not only driven the development of commercial applications, but we also found from the survey statistics of open-source AI research that the AI Index report for 2024 shows that the number of AI-related projects on GitHub skyrocketed from 845 in 2011 to approximately 1.8 million in 2023. Especially after the release of GPT in 2023, the number of projects increased by 59.3% year-on-year, reflecting the global developer community's enthusiasm for AI research.

The enthusiasm for AI technology is directly reflected in the investment market, with the AI investment market showing strong growth and experiencing explosive growth in the second quarter of 2024. There were a total of 16 AI-related investments exceeding $150 million globally, which is twice as many as in the first quarter. The total financing for AI startups has soared to $24 billion, more than doubling year-on-year. Among them, Elon Musk's xAI raised $6 billion, with a valuation of $24 billion, making it the second highest valued AI startup after OpenAI.

Top 10 Financing in the AI Track for Q2 2024, Source: Yiou

The rapid development of AI technology is reshaping the landscape of the tech field at an unprecedented speed. From fierce competition among tech giants to the flourishing development of open-source community projects, and the enthusiastic pursuit of AI concepts in the capital markets. Projects are emerging one after another, with investment amounts reaching new highs and valuations rising accordingly. Overall, the AI market is in a golden period of rapid development, with large language models and retrieval-augmented generation technology achieving significant progress in the field of language processing. Nevertheless, these models still face challenges in translating technological advantages into actual products, such as the uncertainty of model outputs, the risk of generating inaccurate information, and issues of model transparency. These problems become particularly important in application scenarios that require high reliability.

In this context, we began research on AI Agents, as AI Agents emphasize the comprehensiveness of solving practical problems and interacting with the environment. This shift marks the evolution of AI technology from purely language models to intelligent systems that can truly understand, learn, and solve real-world problems. Therefore, we see hope in the development of AI Agents, which are gradually bridging the gap between AI technology and practical problem-solving. The evolution of AI technology is continuously reshaping the architecture of productivity, while Web3 technology is reconstructing the production relationships of the digital economy. When the three core elements of AI: data, models, and computing power, merge with the core concepts of Web3 such as decentralization, token economy, and smart contracts, we anticipate a series of innovative applications to emerge. In this promising intersection, we believe that AI Agents, with their ability to autonomously execute tasks, demonstrate immense potential for large-scale applications.

To this end, we began to explore the diverse applications of AI Agents in Web3, from the infrastructure, middleware, and application layers of Web3, to data and model markets among various dimensions, aiming to identify and evaluate the most promising types of projects and application scenarios, in order to gain a deeper understanding of the deep integration of AI and Web3.

Concept Clarification: Introduction and Classification Overview of AI Agents

Basic Introduction

Before introducing the AI Agent, in order to help readers better understand the difference between its definition and the model itself, let's illustrate with a practical scenario: suppose you are planning a trip. Traditional large language models provide destination information and travel suggestions. Retrieval-augmented generation technology can offer richer and more specific destination content. The AI Agent is like Jarvis from the Iron Man movies, capable of understanding needs and proactively searching for flights and hotels based on your request, executing booking operations, and adding the itinerary to the calendar.

Currently, the industry commonly defines an AI Agent as an intelligent system that can perceive the environment and take corresponding actions. It acquires environmental information through sensors and impacts the environment through actuators after processing (Stuart Russell & Peter Norvig, 2020). We believe that an AI Agent is an assistant that combines LLM, RAG, memory, task planning, and tool usage capabilities. It can not only provide information but also plan and decompose tasks, and truly execute them.

According to this definition and characteristics, we can see that AI Agents have long been integrated into our lives and are applied in various scenarios, such as AlphaGo, Siri, and Tesla's Level 5 autonomous driving and above, which can all be considered instances of AI Agents. The common trait of these systems is that they can perceive external user inputs and respond accordingly to influence the real environment.

Taking ChatGPT as an example for clarification of concepts, we should clearly point out that Transformer is the technical architecture that constitutes AI models, GPT is the series of models developed based on this architecture, and GPT-1, GPT-4, and GPT-4o represent the versions of the model at different stages of development. ChatGPT, on the other hand, is an AI Agent evolved from the GPT model.

Classification Overview

Currently, the AI Agent market has not yet formed a unified classification standard. We categorized 204 AI Agent projects in the Web2+Web3 market by labeling them according to their significant tags. This resulted in primary and secondary classifications. The primary classifications are foundational infrastructure, content generation, and user interaction, which are further subdivided based on their actual use cases:

Infrastructure: This category focuses on building foundational content in the Agent field, including platforms, models, data, development tools, and mature B-end services for foundational applications.

Development Tools: Provide developers with auxiliary tools and frameworks for building AI Agents.

Data Processing: Handling and analyzing data in different formats, primarily used to assist decision-making and provide sources for training.

Model Training Category: Provides model training services for AI, including inference, model establishment, settings, etc.

B-end services: Mainly aimed at enterprise users, providing enterprise service categories, vertical solutions, and automated solutions.

Platform Aggregation Type: A platform that integrates various AI Agent services and tools.

Interactive type: Similar to content generation type, the difference lies in the continuous two-way interaction. Interactive agents not only accept and understand user needs but also provide feedback through technologies such as natural language processing (NLP), achieving two-way interaction with users.

Emotional companionship type: AI Agent that provides emotional support and companionship.

GPT Class: AI Agent based on the GPT (Generative Pre-trained Transformer) model.

Search type: Focuses on search functions, providing Agents primarily for more accurate information retrieval.

Content Generation: This type of project focuses on creating content, using large model technology to generate various forms of content based on user instructions, divided into four categories: text generation, image generation, video generation, and audio generation.

Analysis of the Current Development Status of Web2 AI Agents

According to our statistics, the development of AI Agents in the Web2 traditional internet shows a clear trend of sector concentration. Specifically, about two-thirds of the projects are concentrated in the infrastructure category, mainly consisting of B-end services and development tools. We have also conducted some analysis on this phenomenon.

Impact of Technology Maturity: The dominance of infrastructure projects is primarily due to their technology maturity. These projects are typically built on time-tested technologies and frameworks, which reduces development difficulty and risk. They are akin to the "shovel" in the AI field, providing a solid foundation for the development and application of AI Agents.

Market demand driving: Another key factor is market demand. Compared to the consumer market, the demand for AI technology in the enterprise market is more urgent, especially in seeking solutions to improve operational efficiency and reduce costs. At the same time, for developers, cash flow from enterprises is relatively stable, which is beneficial for them to develop subsequent projects.

Limitations of application scenarios: At the same time, we have noticed that the application scenarios of content generation AI in the B-end market are relatively limited. Due to the instability of its output, enterprises prefer applications that can consistently improve productivity. This has resulted in a smaller proportion of content generation AI in the project library.

This trend reflects the actual considerations of technology maturity, market demand, and application scenarios. With the continuous advancement of AI technology and the further clarification of market demand, we expect that this pattern may be adjusted, but the infrastructure category will still be a solid foundation for the development of AI Agents.

Analysis of Leading Web2 AI Agent Projects

整理的Web2 AI Agent leading projects, source: ArkStream project database

We delve into some current AI Agent projects in the Web2 market and analyze them, taking Character AI, Perplexity AI, and Midjourney as examples.

Character AI:

Product Introduction: Character.AI provides an AI-based dialogue system and virtual character creation tools. Its platform allows users to create, train, and interact with virtual characters that can engage in natural language conversations and perform specific tasks.

Data Analysis: Character.AI had 277 million visits in May, with the platform boasting over 3.5 million daily active users, most of whom are aged between 18 and 34, indicating a younger user demographic. Character AI has performed well in the capital market, completing a $150 million funding round, with a valuation reaching $1 billion, led by a16z.

Technical Analysis: Character AI has signed a non-exclusive licensing agreement with Google's parent company Alphabet to use its large language model, indicating that Character AI is adopting self-developed technology. It is worth mentioning that the company's founders, Noam Shazeer and Daniel De Freitas, were involved in the development of Google's conversational language model Llama.

Perplexity AI:

Product Introduction: Perplexity can scrape the internet and provide detailed answers. It ensures the reliability and accuracy of information through citations and reference links, while also educating and guiding users to ask follow-up questions and search for keywords, meeting users' diverse inquiry needs.

Data Analysis: Perplexity's monthly active user count has reached 10 million, with an 8.6% increase in access to its mobile and desktop applications in February, attracting approximately 50 million users. In the capital markets, Perplexity AI recently announced it has secured $62.7 million in funding, with a valuation of $1.04 billion, led by Daniel Gross, with participants including Stan Druckenmiller and NVIDIA.

Technical Analysis: The main models used by Perplexity are the fine-tuned GPT-3.5 and two large models fine-tuned based on open-source large models: pplx-7b-online and pplx-70b-online. The models are suitable for professional academic research and queries in vertical fields, ensuring the authenticity and reliability of the information.

Midjourney:

Product introduction: User