ArkStream Capital's In-Depth Analysis of the AI Agent Sector

Intermediate10/9/2024, 3:17:22 AM
This report analyzes AI Agent development in Web2 and Web3. Web2 AI Agents focus on infrastructure and B2B services, while Web3 emphasizes model training and platform aggregation. Despite comprising only 8% of Web3 projects, AI Agents represent 23% of the AI sector's market cap, showing strong competitiveness. The report examines commercialization challenges and Web3-AI integration. It predicts AI's future will align with Web3 principles through model standardization and diverse applications.

TL;DR

  • In Web2 startups, AI Agent projects are popular and mature, primarily in enterprise services. In the Web3 space, projects focused on model training and platform aggregation have become mainstream due to their key role in building ecosystems.
  • Currently, AI Agent projects make up only 8% of Web3 projects, but their market capitalization accounts for as much as 23% of the AI sector. This demonstrates strong market competitiveness, and we expect multiple projects to surpass a $1 billion valuation as technology matures and market acceptance increases.
  • For Web3 projects, integrating AI technology into non-AI core applications could become a strategic advantage. When combining with AI Agent projects, attention should be paid to building the entire ecosystem and designing token economic models to promote decentralization and network effects.

The AI Wave: Emerging Projects and Rising Valuations

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 had reached an astonishing $20.3 million, and OpenAI quickly released iterative versions such as GPT-4 and GPT-4o. This rapid pace has prompted traditional tech giants to recognize the importance of cutting-edge AI models like LLMs. Companies like Google released the large language model PaLM2, Meta launched Llama3, and Chinese companies introduced models like Ernie Bot and Zhipu Qingyan, highlighting AI as a crucial battleground.

The race among tech giants has not only accelerated the development of commercial applications but also spurred open-source AI research. The 2024 AI Index report shows that the number of AI-related projects on GitHub skyrocketed from 845 in 2011 to about 1.8 million by 2023, with a 59.3% year-over-year increase in 2023, reflecting the global developer community’s enthusiasm for AI research.

This enthusiasm for AI technology is directly mirrored in the investment market, which has seen explosive growth in the second quarter of 2024. There were 16 AI-related investments exceeding $150 million globally, double that of the first quarter. Total AI startup funding soared to $24 billion, more than doubling year-over-year. Notably, 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 AI sector financings in Q2 2024, Source: Yiou, https://www.iyiou.com/data/202407171072366

AI’s rapid development is reshaping the tech landscape at an unprecedented pace. From fierce competition among tech giants to the booming open-source community, and the capital market’s fervor for AI concepts, projects are emerging continuously, investment amounts are hitting new highs, and valuations are climbing in step. Overall, the AI market is in a golden age of rapid growth, with major advancements in language processing driven by large language models and retrieval-augmented generation technologies. However, challenges remain in translating these technological advances into real products, such as model output uncertainty, the risk of generating inaccurate information (hallucinations), and issues with model transparency—especially critical in high-reliability applications.

In this context, we have begun researching AI Agents, which emphasize problem-solving and interaction with real-world environments. This shift marks the evolution of AI from pure language models to intelligent systems capable of truly understanding, learning, and solving real-world problems. We see promise in AI Agents, as they are gradually bridging the gap between AI technology and practical problem-solving. As AI evolves to reshape productivity frameworks, Web3 is reconstructing the production relations of the digital economy. When the three pillars of AI—data, models, and computing power—merge with Web3’s core principles of decentralization, token economies, and smart contracts, we foresee the birth of a series of innovative applications. In this promising intersection, AI Agents, with their ability to autonomously execute tasks, show immense potential for large-scale applications. Therefore, we are delving into the diverse applications of AI Agents in Web3, from Web3 infrastructure, middleware, and application layers to data and model marketplaces, aiming to identify and assess the most promising project types and application scenarios to deepen our understanding of AI-Web3 integration.

Concept clarification: Introduction and Overview of AI Agent Classification

Basic introduction

Before introducing AI Agents, to help readers better understand the difference between their definition and traditional models, let’s use a real-world scenario as an example: Suppose you are planning a trip. A traditional large language model provides destination information and travel suggestions. Retrieval-augmented generation (RAG) technology can offer richer, more specific destination content. In contrast, an AI Agent acts like Jarvis from the Iron Man movies—it understands your needs, actively searches for flights and hotels based on your request, makes reservations, and adds the itinerary to your calendar.

In the industry, AI Agents are generally defined as intelligent systems capable of perceiving the environment and taking appropriate actions by gathering environmental information through sensors, processing it, and affecting the environment through actuators (Stuart Russell & Peter Norvig, 2020). We view an AI Agent as an assistant that integrates LLM (Large Language Models), RAG, memory, task planning, and tool usage. It not only provides information but also plans, breaks down tasks, and actually executes them.

Based on this definition and characteristics, we can see that AI Agents have already integrated into our daily lives and are applied in various scenarios. For example, AlphaGo, Siri, and Tesla’s Level 5 and above autonomous driving can all be considered examples of AI Agents. The common trait among these systems is their ability to perceive external user inputs and make decisions that affect the real world based on those inputs.

To clarify concepts using ChatGPT as an example, it is important to distinguish that Transformer is the technical architecture that forms the foundation of AI models, while GPT refers to the series of models developed based on this architecture. GPT-1, GPT-4, and GPT-4o represent different stages of model development. ChatGPT, as an evolution based on the GPT model, can be considered an AI Agent.

Classification Overview

Currently, there is no unified classification standard for AI Agents in the market. By tagging 204 AI Agent projects across Web2 and Web3 markets based on their prominent features, we created both primary and secondary classifications. The primary classifications include infrastructure, content generation, and user interaction, which are then further divided based on actual use cases:

  • Infrastructure: Focuses on building foundational components in the AI Agent field, including platforms, models, data, development tools, and enterprise-level B2B services.
  • Development tools: Provide developers with auxiliary tools and frameworks for building AI Agents.
  • Data processing: Process and analyze data in different formats, mainly used to assist decision-making and provide sources for training.
  • Model training: Provides model training services for AI, including inference, model establishment and settings, etc.
  • B2B services: mainly for enterprise users, providing enterprise service, vertical, and automated solutions.
  • Platform aggregation: a platform that integrates multiple AI Agent services and tools.
  • User Interaction: Similar to content generation, but with ongoing two-way interaction. Interaction Agents not only understand and respond to user needs but also provide feedback using natural language processing (NLP), enabling bidirectional communication.
  • Emotional companionship: AI Agent that provides emotional support and companionship.
  • GPT-based: AI Agent based on the GPT (generative pre-training Transformer) model.
  • Search: Agent that focuses on search functions and provides more accurate information retrieval.
  • Content Generation: Projects focused on creating content using large model technologies based on user instructions, categorized into text generation, image generation, video generation, and audio generation.

Analysis of Web2 AI Agent Development

According to our research, the development of AI Agents in the traditional Web2 internet shows a clear concentration in specific sectors. About two-thirds of the projects are focused on infrastructure, particularly in B2B services and developer tools. We analyzed this phenomenon and identified several key factors:

Impact of Technology Maturity: The dominance of infrastructure projects is largely due to the maturity of the underlying technologies. These projects are often built on well-established technologies and frameworks, reducing the difficulty and risk of development. They serve as the “shovels” in the AI field, providing a solid foundation for AI Agent development and application.

Market Demand: Another key factor is market demand. Compared to the consumer market, the enterprise market has a more urgent need for AI technology, especially for solutions aimed at improving operational efficiency and reducing costs. For developers, the stable cash flow from enterprise customers makes it easier to develop subsequent projects.

Application Limitations: At the same time, we noticed that content generation AI has limited application scenarios in the B2B market. Due to the instability of its output, businesses tend to prefer applications that reliably boost productivity, which is why content generation AI occupies a smaller portion of the project landscape.

This trend reflects the practical considerations of technology maturity, market demand, and application scenarios. As AI technology continues to advance and market demands become clearer, we expect this landscape to shift, but infrastructure will likely remain a cornerstone of AI Agent development.

Analysis of Leading Web2 AI Agent Projects

Compilation of Web2’s AI Agent leading projects, source: ArkStream project database

We analyzed some of the leading AI Agent projects in the Web2 market, sourced from the ArkStream project database. Using Character AI, Perplexity AI, and Midjourney as examples, we delve into their details.

Character AI:

  • Product Overview: Character.AI offers AI-based conversation systems and virtual character creation tools. The platform allows users to create, train, and interact with virtual characters that can engage in natural language conversations and perform specific tasks.
  • Data Analysis: In May, Character.AI had 277 million visits and over 3.5 million daily active users, most of whom are aged between 18 and 34, indicating a younger user base. Character AI has performed well in the capital market, raising $150 million in funding with a valuation of $1 billion, led by a16z.
  • Technical Analysis: Character AI has signed a non-exclusive licensing agreement with Alphabet, Google’s parent company, to use its large language model. The company’s founders, Noam Shazeer and Daniel De Freitas, were involved in developing Google’s conversational language model, Llama.

Perplexity AI:

  • Product Overview: Perplexity scrapes the internet to provide detailed answers, ensuring information reliability by citing references and links. It educates and guides users in asking follow-up questions and searching for keywords, meeting diverse query needs.
  • Data Analysis: Perplexity has reached 10 million monthly active users, with its mobile and desktop app traffic growing 8.6% in February, attracting about 50 million users. Perplexity AI recently raised $62.7 million in funding, with a valuation of $1.04 billion, led by Daniel Gross, with participation from Stan Druckenmiller and NVIDIA.
  • Technical Analysis: Perplexity primarily uses fine-tuned GPT-3.5 models and two large models fine-tuned from open-source models: pplx-7b-online and pplx-70b-online. These models are suited for academic research and vertical domain queries, ensuring the authenticity and reliability of information.

Midjourney:

  • Product Overview: Users can create images in various styles and themes on Midjourney through prompts, covering a wide range of creative needs from realism to abstraction. The platform also offers image blending and editing, allowing users to overlay images and transfer styles, with real-time generation enabling image outputs in seconds to minutes.
  • Data Analysis: The platform has 15 million registered users, with 1.5 to 2.5 million active users. Based on public market information, Midjourney has not raised money from investors and has sustained itself through the founder David’s entrepreneurial reputation and resources.
  • Technical Analysis: Midjourney uses its own proprietary model. Since the release of Midjourney V4 in August 2022, the platform has been using a diffusion-based generative AI model. The model’s training parameters reportedly range from 30 to 40 billion, providing a solid foundation for the diversity and accuracy of its image generation.

Commercialization Challenges

After experiencing several Web2 AI Agents, we observed a common product iteration path: from initially focusing on single, specific tasks to later expanding their capabilities to handle more complex, multi-task scenarios. This trend highlights the potential of AI Agents in improving efficiency and innovation, indicating that they will play a more critical role in the future. Based on preliminary statistics of 125 AI Agent projects in Web2, we found that most projects are concentrated in content generation (e.g., Jasper AI), developer tools (e.g., Replit), and B2B services (e.g., Cresta), the largest category. This finding was contrary to our expectations, as we initially predicted that with the increasing maturity of AI model technology, the consumer market (C-end) would experience an explosive growth of AI Agents. However, after further analysis, we realized that the commercialization of consumer AI Agents is much more challenging and complex than expected.

Take Character.AI as an example. On one hand, Character.AI has some of the best traffic performance. However, due to its singular business model—relying on a $9.9 USD subscription fee—it struggled with limited subscription revenue and high inference costs for heavy users, eventually leading to its acquisition by Google due to difficulties in monetizing traffic and maintaining cash flow. This case shows that even with excellent traffic and funding, C-end AI Agent applications face significant commercialization challenges. Most products have not yet reached the standard where they can replace or effectively assist humans, resulting in low user willingness to pay. In our research, we found that many startups encounter similar problems as Character.AI, indicating that the development of consumer AI Agents is not smooth and requires deeper exploration of technical maturity, product value, and business model innovation to unlock their potential in the C-end market.

By counting the valuations of most AI Agent projects, compared with the valuations of ceiling projects such as OpenAI and xAI, there is still room for close to 10-50 times.It is undeniable that the ceiling of C-side Agent application is still high enough, proving that it is still a good track. However, based on the above analysis, we believe that compared with the C-side, the B-side market may be the final destination of AI Agent. By building a platform, enterprises integrate AI Agent into management software such as vertical fields, CRM, and office OA. This not only improves operational efficiency for enterprises, but also provides AI Agent with a broader application space. Therefore, we have reason to believe that B-side services will be the main direction of the short-term development of AI Agents in the Web2 traditional Internet.

Web3 AI Agent Development Status and Prospects

Project overview

As analyzed earlier, even AI Agent applications with top-tier funding and good user traffic face difficulties in commercialization. Next, we will analyze the current development of AI Agent projects in Web3. By evaluating a series of representative projects—their technical innovations, market performance, user feedback, and growth potential—we aim to uncover insightful suggestions. The chart below shows several representative projects that have already issued tokens and hold a relatively high market value:

Compilation of Web2’s AI Agent leading projects, source: ArkStream project database

According to our statistics on the Web3 AI Agent market, the types of projects being developed also show a clear concentration in specific sectors. Most projects fall under infrastructure, with fewer content generation projects. Many of these projects aim to leverage user-provided distributed data and computing power to meet the model training needs of the project owners or to create all-in-one platforms that integrate various AI Agent services and tools. From developer tools to front-end interaction applications and generative applications, most traditional AI Agent industries are currently limited to open-source parameter adjustments or building applications using existing models. This method has not yet generated significant network effects for enterprises or individual users.

Status Analysis

We believe that this phenomenon at this stage may be driven by the following factors:

Market and Technology Mismatch: The combination of Web3 and AI Agents does not currently show a significant advantage over traditional markets. The true advantage lies in improving production relationships by optimizing resources and collaboration through decentralization. This may cause interaction and generative applications to struggle to compete with traditional competitors with stronger technical and financial resources.

Application Scenario Limitations: In the Web3 environment, there may not be as much demand for generating images, videos, or text content. Instead, Web3’s decentralized and distributed features are more often used to reduce costs and improve efficiency within the traditional AI field, rather than to expand into new application scenarios.

The root cause of this phenomenon may lie in the current development state of the AI industry and its future direction. AI technology is still in its early stages, akin to the early days of the industrial revolution when steam engines were replaced by electric motors. It has not yet reached the electrification stage of widespread application.

We believe that the future of AI will likely follow a similar path. General models will gradually become standardized, while fine-tuned models will see diversified development. AI applications will be widely dispersed across enterprises and individual users, with the focus shifting to interconnection and interaction between models. This trend aligns closely with Web3’s principles, as Web3 is known for its composability and permissionless nature, which fits well with the idea of decentralized fine-tuning of models. Developers will have greater freedom to combine and adjust various models. Additionally, decentralization offers unique advantages in areas such as data privacy protection and computing resource allocation for model training.

With technological advancements, especially the emergence of innovations like LoRA (Low-Rank Adaptation), the cost and technical barriers for model fine-tuning have been significantly lowered. This makes it easier to develop public models for specific scenarios or to meet users’ personalized needs. AI Agent projects within Web3 can fully leverage these advancements to explore novel training methods, innovative incentive mechanisms, and new models of model sharing and collaboration, which are often difficult to achieve in traditional centralized systems.

Moreover, the concentration of Web3 projects on model training reflects strategic considerations of its importance within the entire AI ecosystem. Thus, the focus of Web3 AI Agent projects on model training is a natural convergence of technology trends, market demand, and Web3 industry advantages. Next, we will provide examples of model training projects in both Web2 and Web3 industries and make comparisons.

Model Training Projects

Humans.ai

  • Project Overview: Humans.ai is a diversified AI algorithm library and training deployment environment covering various fields such as images, videos, audio, and text. The platform supports developers in further training and optimizing models, and allows them to share and trade their models. A notable innovation is that Humans.ai uses NFTs to store AI models and users’ biometric data, making the AI content creation process more personalized and secure.
  • Data Analysis: The market value of Humans.ai’s token, Heart, is about $68 million. They have 56k followers on Twitter, though user data has not been disclosed.
  • Technical Analysis: Humans.ai does not develop its own models but uses a modular approach, packaging all available models into NFTs, providing users with a flexible and scalable AI solution.

FLock.io

  • Project Overview: FLock.io is an AI co-creation platform based on federated learning technology (a decentralized machine learning method that emphasizes data privacy). It aims to address pain points in the AI field, such as low public engagement, insufficient privacy protection, and the monopoly of AI technology by large corporations. The platform allows users to contribute data while protecting privacy, promoting the democratization and decentralization of AI technology.
  • Data Analysis: FLock.io completed a $6 million seed round in early 2024, led by Lightspeed Faction and Tagus Capital, with additional participation from DCG, OKX Ventures, and others.
  • Technical Analysis: FLock.io’s architecture is based on federated learning, a decentralized method that promotes privacy protection. It also uses zkFL, homomorphic encryption, and secure multi-party computation (SMPC) to provide additional privacy safeguards.

These are examples of model training projects within the Web3 AI Agent space, but similar platforms also exist in Web2, such as Predibase.

Predibase

  • Project Overview: Predibase focuses on AI and large language model optimization, allowing users to fine-tune and deploy open-source large language models, such as Llama, CodeLlama, and Phi. The platform supports various optimization techniques like quantization, low-rank adaptation, and memory-efficient distributed training.
  • Data Analysis: Predibase announced the completion of a $12.2 million Series A round led by Felicis, with major companies like Uber, Apple, Meta, and startups like Paradigm and Koble.ai as platform users.
  • Technical Analysis: Predibase users have trained over 250 models. The platform utilizes the LoRAX architecture and Ludwig framework: LoRAX enables thousands of fine-tuned LLMs to run on a single GPU, significantly reducing costs without affecting throughput or latency. Ludwig is a declarative framework Predibase uses to develop, train, fine-tune, and deploy cutting-edge deep learning and large language models.
  • Project Analysis: Predibase offers user-friendly features that provide customized AI application-building services for different levels of users. Whether for C-end or B-end users, beginners or experienced AI professionals, Predibase caters to a wide range of needs.

For beginners, the platform’s one-click automation simplifies the model-building and training process, automatically handling complex tasks. For experienced users, it provides deeper customization options, including access to and adjustment of more advanced parameters. When comparing traditional AI model training platforms with Web3 AI projects, while their overall frameworks and logic may be similar, we found significant differences in their technical architecture and business models.

  • Technical Depth and Innovation: Traditional AI model training platforms often have deeper technical barriers, such as using proprietary technologies like the LoRAX architecture and Ludwig framework. These frameworks offer robust features, allowing the platform to handle complex AI model training tasks. However, Web3 projects may focus more on decentralization and openness, with less emphasis on deep technical innovation.
  • Business Model Flexibility: A common bottleneck in traditional AI model training is the lack of flexibility in the business model. Platforms typically require users to pay to train models, limiting the project’s sustainability, especially in the early stages when broad user participation and data collection are needed. In contrast, Web3 projects often have more flexible business models, such as tokenomics driven by communities.
  • Privacy Protection Challenges: Privacy protection is another key issue. For example, although Predibase offers virtual private cloud services on AWS, relying on third-party architecture always carries the potential risk of data leaks.

These differences have become bottlenecks in the traditional AI industry. Due to the nature of the internet, these issues are difficult to solve efficiently. At the same time, this presents both opportunities and challenges for Web3, where projects that can solve these problems first will likely become pioneers in the industry.

Other Categories of Web3 AI Agent Projects

After discussing AI Agent projects focused on model training, we now expand our view to other types of AI Agent projects in the Web3 industry. These projects, while not exclusively focused on model training, demonstrate distinctive performance in terms of funding, token valuation, and market presence. Below are some representative and influential AI Agent projects in their respective fields:

Myshell

  • Product Overview: Myshell provides a comprehensive AI Agent platform where users can create, share, and personalize AI agents. These agents can offer companionship and assist with work efficiency. The platform includes diverse AI agent styles, from anime to traditional, and supports interactions via audio, video, and text. A standout feature is the integration of multiple existing models, including GPT-4o, GPT-4, and Claude, providing a premium experience. Additionally, Myshell introduces a trading system similar to FT bonding curves, incentivizing creators to develop high-value AI models while giving users the chance to invest and share in profits.
  • Data Analysis: Myshell’s last funding round valued the company at around $80 million, led by Dragonfly, with participation from Binance, Hashkey, and Folius. With nearly 180K Twitter followers, it has a dedicated community of users and developers, despite Discord engagement being less than one-tenth of its follower base.
  • Technical Analysis: Myshell does not develop AI models independently but serves as an integration platform, combining models like Claude and GPT-4. This strategy allows it to offer users a unified and advanced AI experience.
  • Subjective Experience: MyShell allows users to freely create and customize AI agents according to their own needs. Whether as a personal companion or a professional assistant, it can adapt to various scenarios such as audio and video. Even if users do not use MyShell’s proxy, they can enjoy the integrated Web2 paid model at a lower cost. In addition, the platform combines the economic concept of FT, allowing users to not only use AI services, but also invest in AI agents they are optimistic about, increasing the wealth effect through the bonding curve mechanism.

Delysium

  • Product Overview: Delysium provides an intent-centered AI Agent network, allowing Agents to better cooperate to bring users a friendly Web3 experience. Currently, Delysium has launched two AI Agents: Lucy and Jerry. Lucy is a networked AI Agent. The vision is to provide tool assistance, such as querying the Top 10 currency holding addresses, etc. However, currently the function of the Agent to execute on-chain intentions has not yet been opened, and it can only execute some basic instructions, such as staking AGI within the ecosystem. Or exchange it to USDT. Jerry is similar to GPT in the Delysium ecosystem, and is mainly responsible for answering questions within the ecosystem, such as token distribution.
  • Data Analysis: The first round of financing was US$4 million in 2022, and in the same year it was announced that it had completed strategic financing of US$10 million. Its token AGI currently has an FDV of about $130 million. There is no latest user data. According to official statistics from Delysium, Lucy has accumulated more than 1.4 million independent wallet connections as of June 2023.

Sleepless AI

  • Product Overview: An emotional companionship game platform that combines Web3 and AI Agent technology to provide virtual companion games HIM and HER, using AIGC and LLM to immerse users in interactions with virtual characters. Users can modify the character’s attributes, clothing, etc. during the ongoing conversation. Its compatible large language model ensures that the character iterates on itself in each conversation and becomes more understanding of the user.
  • Data Analysis: The project has raised a total of US$3.7 million, with investors including Binance Labs, Foresight Ventures and Folius Ventures. The current total market value of the tokens has reached approximately US$400 million. It has 116K Twitter followers, 190K registered reservations according to official statistics, and 43K active users. It can be said that its user stickiness is quite strong.
  • Technical Analysis:Although the official did not disclose which major language model on the market their product is based on, Sleepless AI ensures that users will feel that the character understands them more and more during the chat process. Therefore, when designing LLM training, they Each character trains a model separately, and combines the vector database and personality parameter system to allow the character to have memory.
  • Subjective Experience: Sleepless AI approaches AI Boyfriend and AI Girlfriend from a Free-to-Play perspective, and is not just integrated into the chat box of a conversational robot. The project greatly enhances the authenticity of virtual humans through high-cost art, continuously iterative language models, high-quality and complete dubbing, and a series of functions such as alarm clock, sleep aid, menstrual period recording, study companionship, etc. This kind of emotional value cannot be felt by other applications on the market. In addition, Sleepless AI creates a longer-term, balanced content payment mechanism. Users can choose to sell NFT without falling into the dilemma of P2E or Ponzi. This model takes into account both player income and game experience.

Prospect Analysis

In the Web3 industry, AI Agent projects cover multiple directions including public chains, data management, privacy protection, social networks, platform services, and computing power. In terms of token market value, the total token market value of AI Agent projects has reached nearly $3.8 billion, while the total market value of the entire AI track is close to $16.2 billion. AI Agent projects account for about 23% of the market value in the AI track.

Although there are only about a dozen AI Agent projects, which seems relatively few compared to the entire AI track, their market valuation accounts for nearly a quarter. This market value proportion in the AI track once again validates our belief that this sub-track has great growth potential.

After our analysis, we raised a core question: What characteristics do Agent projects need to attract excellent financing and be listed on top exchanges? To answer this question, we explored successful projects in the Agent industry, such as Fetch.ai, Olas Network, SingularityNET, and Myshell.

We found that these projects share some significant features: they all belong to the platform aggregation category within the infrastructure class. They build a bridge, connecting users who need Agents on one end (both B2B and B2C), and developers and validators responsible for model debugging and training on the other end. Regardless of the application level, they have all established a complete ecological closed loop.

We noticed that whether their products are on-chain or off-chain related doesn’t seem to be the most crucial factor. This leads us to a preliminary conclusion: in the Web3 domain, the logic of focusing on practical applications in Web2 may not fully apply. For leading AI Agent products in Web3, building a complete ecosystem and providing diverse functionalities might be more critical than the quality and performance of a single product. In other words, a project’s success depends not only on what it offers but more on how it integrates resources, promotes collaboration, and creates network effects within the ecosystem. This ability to build ecosystems might be a key factor for AI Agent projects to stand out in the Web3 track.

The correct integration method for AI Agent projects in Web3 is not to focus on the deep development of a single application, but to adopt an inclusive model. This approach involves migrating and integrating diverse product frameworks and types from the Web2 era into the Web3 environment to build a self-cycling ecosystem. This point can also be seen in OpenAI’s strategic shift, as they chose to launch an application platform this year rather than just updating their model.

In summary, we believe that the AI ​​Agent project should focus on the following aspects:

  • Ecosystem Building: Go beyond single applications to build an ecosystem that includes multiple services and functions, promoting interaction and value addition between different components.
  • Tokenomic Model: Design a reasonable token economic model to incentivize users to participate in network construction and contribute data and computing power.
  • Cross-domain Integration: Explore the potential applications of AI Agents in different fields, creating new usage scenarios and value through cross-domain integration.

After summarizing these three aspects, we also provide some forward-looking suggestions for project teams with different focus directions: one for non-AI core application products, and another for native projects focused on the AI Agent track.

For non-AI core application products:

Maintain a long-term perspective, focus on their core products while integrating AI technology, and wait for the right opportunity in line with the times. In the current technological and market trends, we believe that using AI as a traffic medium to attract users and enhance product competitiveness has become an important means of competitiveness. Although the actual long-term contribution of AI technology to project development remains a question mark, we believe this provides a valuable window for early adopters of AI technology. Of course, the premise is that they already have a very solid product.

In the long run, if AI technology achieves new breakthroughs in the future, those projects that have already integrated AI will be able to iterate their products more quickly, thus seizing opportunities and becoming industry leaders. This is similar to how live streaming e-commerce gradually replaced offline sales as a new traffic outlet on social media platforms in recent years. At that time, those merchants with solid products who chose to follow the new trend and try live streaming e-commerce immediately stood out with the advantage of early entry when live streaming e-commerce truly exploded.

We believe that amid market uncertainty, for non-AI core application products, considering the timely introduction of AI Agents may be a strategic decision. It can not only increase the product’s market exposure at present but also bring new growth points for the product in the continuous development of AI technology.

For native projects focused on AI Agents:

Balancing technological innovation and market demand is key to success. In native AI Agent projects, project teams need to look towards market trends, not just technology development. Currently, some Web3-integrated Agent projects in the market may be overly focused on developing in a single technical direction, or have constructed a grand vision, but product development has not kept up. Both of these extremes are not conducive to the long-term development of the project.

Therefore, we suggest that project teams, while ensuring product quality, should also pay attention to market dynamics, and realize that the AI application logic in the traditional internet industry may not apply to Web3. Instead, they need to learn from those projects that have already achieved results in the Web3 market. Focus on the labels they have, such as model training and platform aggregation core functions mentioned in the article, as well as the narratives they create, such as AI modularization and multi-Agent collaboration. Exploring compelling narratives may become the key for projects to achieve breakthroughs in the market.

Conclusion

Whether it is a non-AI core product or a native AI Agent project, the most critical thing is to find the right timing and technical path to ensure that it remains competitive and innovative in the ever-changing market. On the basis of maintaining product quality, project parties should observe market trends, learn from successful cases, and at the same time innovate to achieve sustainable development in the market.

Summary

At the end of the article, we analyze the Web3 AI Agent track from multiple angles:

  • Capital investment and market attention: Although AI Agent projects currently do not have an advantage in the number of listings in the Web3 industry, they account for close to 50% of the market valuation, showing that the capital market highly recognizes this track. With more capital investment and increasing market attention, it is certain that more high-valued projects will appear in the AI ​​Agent track.
  • Competitive landscape and innovation capabilities: The competitive landscape of the AI ​​Agent track in the Web3 industry has not yet been fully formed. At the current application level, there is no phenomenal and leading product similar to ChatGPT. This gives new project parties a lot of room for growth and innovation. As the technology matures and previous projects are innovated, the track is expected to develop more competitive products, driving up the valuation of the entire track.
  • Pay attention to tokenomics and user incentives: The significance of Web3 is to reshape production relations and make the originally centralized process of deploying and training AI models more decentralized. Through reasonable tokenomics design and user incentive programs, idle computing power or personal datasets can be redistributed. Additionally, solutions like ZKML can protect data privacy, further reducing computing power and data costs, and allowing more individual users to participate in the construction of the AI industry.

To sum up, we are optimistic about the AI ​​Agent track. We have reason to believe that multiple projects with valuations exceeding $1 billion will emerge in this track. Through horizontal comparison, the narrative of AI Agent is sufficiently compelling and the market space is large enough. The current market valuations are generally low. Considering the rapid development of AI technology, the growth of market demand, capital investment and the innovation potential of companies in the track, in the future, as technology matures and market acceptance increases, this track is expected to see multiple projects with valuations over $1 billion emerge.

statement:

  1. This article is reproduced from [ArkStream Capital], the original title is “ArkStream Capital Track Research Report: Can AI Agent be a life-saving straw for Web3+AI?” If you have any objections to the reprint, please contact Gate Learn Team, the team will handle it as soon as possible according to relevant procedures.

  2. Disclaimer: The views and opinions expressed in this article represent only the author’s personal views and do not constitute any investment advice.

  3. Other language versions of the article are translated by the Gate Learn team, not mentioned in Gate.io, the translated article may not be reproduced, distributed or plagiarized.

ArkStream Capital's In-Depth Analysis of the AI Agent Sector

Intermediate10/9/2024, 3:17:22 AM
This report analyzes AI Agent development in Web2 and Web3. Web2 AI Agents focus on infrastructure and B2B services, while Web3 emphasizes model training and platform aggregation. Despite comprising only 8% of Web3 projects, AI Agents represent 23% of the AI sector's market cap, showing strong competitiveness. The report examines commercialization challenges and Web3-AI integration. It predicts AI's future will align with Web3 principles through model standardization and diverse applications.

TL;DR

  • In Web2 startups, AI Agent projects are popular and mature, primarily in enterprise services. In the Web3 space, projects focused on model training and platform aggregation have become mainstream due to their key role in building ecosystems.
  • Currently, AI Agent projects make up only 8% of Web3 projects, but their market capitalization accounts for as much as 23% of the AI sector. This demonstrates strong market competitiveness, and we expect multiple projects to surpass a $1 billion valuation as technology matures and market acceptance increases.
  • For Web3 projects, integrating AI technology into non-AI core applications could become a strategic advantage. When combining with AI Agent projects, attention should be paid to building the entire ecosystem and designing token economic models to promote decentralization and network effects.

The AI Wave: Emerging Projects and Rising Valuations

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 had reached an astonishing $20.3 million, and OpenAI quickly released iterative versions such as GPT-4 and GPT-4o. This rapid pace has prompted traditional tech giants to recognize the importance of cutting-edge AI models like LLMs. Companies like Google released the large language model PaLM2, Meta launched Llama3, and Chinese companies introduced models like Ernie Bot and Zhipu Qingyan, highlighting AI as a crucial battleground.

The race among tech giants has not only accelerated the development of commercial applications but also spurred open-source AI research. The 2024 AI Index report shows that the number of AI-related projects on GitHub skyrocketed from 845 in 2011 to about 1.8 million by 2023, with a 59.3% year-over-year increase in 2023, reflecting the global developer community’s enthusiasm for AI research.

This enthusiasm for AI technology is directly mirrored in the investment market, which has seen explosive growth in the second quarter of 2024. There were 16 AI-related investments exceeding $150 million globally, double that of the first quarter. Total AI startup funding soared to $24 billion, more than doubling year-over-year. Notably, 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 AI sector financings in Q2 2024, Source: Yiou, https://www.iyiou.com/data/202407171072366

AI’s rapid development is reshaping the tech landscape at an unprecedented pace. From fierce competition among tech giants to the booming open-source community, and the capital market’s fervor for AI concepts, projects are emerging continuously, investment amounts are hitting new highs, and valuations are climbing in step. Overall, the AI market is in a golden age of rapid growth, with major advancements in language processing driven by large language models and retrieval-augmented generation technologies. However, challenges remain in translating these technological advances into real products, such as model output uncertainty, the risk of generating inaccurate information (hallucinations), and issues with model transparency—especially critical in high-reliability applications.

In this context, we have begun researching AI Agents, which emphasize problem-solving and interaction with real-world environments. This shift marks the evolution of AI from pure language models to intelligent systems capable of truly understanding, learning, and solving real-world problems. We see promise in AI Agents, as they are gradually bridging the gap between AI technology and practical problem-solving. As AI evolves to reshape productivity frameworks, Web3 is reconstructing the production relations of the digital economy. When the three pillars of AI—data, models, and computing power—merge with Web3’s core principles of decentralization, token economies, and smart contracts, we foresee the birth of a series of innovative applications. In this promising intersection, AI Agents, with their ability to autonomously execute tasks, show immense potential for large-scale applications. Therefore, we are delving into the diverse applications of AI Agents in Web3, from Web3 infrastructure, middleware, and application layers to data and model marketplaces, aiming to identify and assess the most promising project types and application scenarios to deepen our understanding of AI-Web3 integration.

Concept clarification: Introduction and Overview of AI Agent Classification

Basic introduction

Before introducing AI Agents, to help readers better understand the difference between their definition and traditional models, let’s use a real-world scenario as an example: Suppose you are planning a trip. A traditional large language model provides destination information and travel suggestions. Retrieval-augmented generation (RAG) technology can offer richer, more specific destination content. In contrast, an AI Agent acts like Jarvis from the Iron Man movies—it understands your needs, actively searches for flights and hotels based on your request, makes reservations, and adds the itinerary to your calendar.

In the industry, AI Agents are generally defined as intelligent systems capable of perceiving the environment and taking appropriate actions by gathering environmental information through sensors, processing it, and affecting the environment through actuators (Stuart Russell & Peter Norvig, 2020). We view an AI Agent as an assistant that integrates LLM (Large Language Models), RAG, memory, task planning, and tool usage. It not only provides information but also plans, breaks down tasks, and actually executes them.

Based on this definition and characteristics, we can see that AI Agents have already integrated into our daily lives and are applied in various scenarios. For example, AlphaGo, Siri, and Tesla’s Level 5 and above autonomous driving can all be considered examples of AI Agents. The common trait among these systems is their ability to perceive external user inputs and make decisions that affect the real world based on those inputs.

To clarify concepts using ChatGPT as an example, it is important to distinguish that Transformer is the technical architecture that forms the foundation of AI models, while GPT refers to the series of models developed based on this architecture. GPT-1, GPT-4, and GPT-4o represent different stages of model development. ChatGPT, as an evolution based on the GPT model, can be considered an AI Agent.

Classification Overview

Currently, there is no unified classification standard for AI Agents in the market. By tagging 204 AI Agent projects across Web2 and Web3 markets based on their prominent features, we created both primary and secondary classifications. The primary classifications include infrastructure, content generation, and user interaction, which are then further divided based on actual use cases:

  • Infrastructure: Focuses on building foundational components in the AI Agent field, including platforms, models, data, development tools, and enterprise-level B2B services.
  • Development tools: Provide developers with auxiliary tools and frameworks for building AI Agents.
  • Data processing: Process and analyze data in different formats, mainly used to assist decision-making and provide sources for training.
  • Model training: Provides model training services for AI, including inference, model establishment and settings, etc.
  • B2B services: mainly for enterprise users, providing enterprise service, vertical, and automated solutions.
  • Platform aggregation: a platform that integrates multiple AI Agent services and tools.
  • User Interaction: Similar to content generation, but with ongoing two-way interaction. Interaction Agents not only understand and respond to user needs but also provide feedback using natural language processing (NLP), enabling bidirectional communication.
  • Emotional companionship: AI Agent that provides emotional support and companionship.
  • GPT-based: AI Agent based on the GPT (generative pre-training Transformer) model.
  • Search: Agent that focuses on search functions and provides more accurate information retrieval.
  • Content Generation: Projects focused on creating content using large model technologies based on user instructions, categorized into text generation, image generation, video generation, and audio generation.

Analysis of Web2 AI Agent Development

According to our research, the development of AI Agents in the traditional Web2 internet shows a clear concentration in specific sectors. About two-thirds of the projects are focused on infrastructure, particularly in B2B services and developer tools. We analyzed this phenomenon and identified several key factors:

Impact of Technology Maturity: The dominance of infrastructure projects is largely due to the maturity of the underlying technologies. These projects are often built on well-established technologies and frameworks, reducing the difficulty and risk of development. They serve as the “shovels” in the AI field, providing a solid foundation for AI Agent development and application.

Market Demand: Another key factor is market demand. Compared to the consumer market, the enterprise market has a more urgent need for AI technology, especially for solutions aimed at improving operational efficiency and reducing costs. For developers, the stable cash flow from enterprise customers makes it easier to develop subsequent projects.

Application Limitations: At the same time, we noticed that content generation AI has limited application scenarios in the B2B market. Due to the instability of its output, businesses tend to prefer applications that reliably boost productivity, which is why content generation AI occupies a smaller portion of the project landscape.

This trend reflects the practical considerations of technology maturity, market demand, and application scenarios. As AI technology continues to advance and market demands become clearer, we expect this landscape to shift, but infrastructure will likely remain a cornerstone of AI Agent development.

Analysis of Leading Web2 AI Agent Projects

Compilation of Web2’s AI Agent leading projects, source: ArkStream project database

We analyzed some of the leading AI Agent projects in the Web2 market, sourced from the ArkStream project database. Using Character AI, Perplexity AI, and Midjourney as examples, we delve into their details.

Character AI:

  • Product Overview: Character.AI offers AI-based conversation systems and virtual character creation tools. The platform allows users to create, train, and interact with virtual characters that can engage in natural language conversations and perform specific tasks.
  • Data Analysis: In May, Character.AI had 277 million visits and over 3.5 million daily active users, most of whom are aged between 18 and 34, indicating a younger user base. Character AI has performed well in the capital market, raising $150 million in funding with a valuation of $1 billion, led by a16z.
  • Technical Analysis: Character AI has signed a non-exclusive licensing agreement with Alphabet, Google’s parent company, to use its large language model. The company’s founders, Noam Shazeer and Daniel De Freitas, were involved in developing Google’s conversational language model, Llama.

Perplexity AI:

  • Product Overview: Perplexity scrapes the internet to provide detailed answers, ensuring information reliability by citing references and links. It educates and guides users in asking follow-up questions and searching for keywords, meeting diverse query needs.
  • Data Analysis: Perplexity has reached 10 million monthly active users, with its mobile and desktop app traffic growing 8.6% in February, attracting about 50 million users. Perplexity AI recently raised $62.7 million in funding, with a valuation of $1.04 billion, led by Daniel Gross, with participation from Stan Druckenmiller and NVIDIA.
  • Technical Analysis: Perplexity primarily uses fine-tuned GPT-3.5 models and two large models fine-tuned from open-source models: pplx-7b-online and pplx-70b-online. These models are suited for academic research and vertical domain queries, ensuring the authenticity and reliability of information.

Midjourney:

  • Product Overview: Users can create images in various styles and themes on Midjourney through prompts, covering a wide range of creative needs from realism to abstraction. The platform also offers image blending and editing, allowing users to overlay images and transfer styles, with real-time generation enabling image outputs in seconds to minutes.
  • Data Analysis: The platform has 15 million registered users, with 1.5 to 2.5 million active users. Based on public market information, Midjourney has not raised money from investors and has sustained itself through the founder David’s entrepreneurial reputation and resources.
  • Technical Analysis: Midjourney uses its own proprietary model. Since the release of Midjourney V4 in August 2022, the platform has been using a diffusion-based generative AI model. The model’s training parameters reportedly range from 30 to 40 billion, providing a solid foundation for the diversity and accuracy of its image generation.

Commercialization Challenges

After experiencing several Web2 AI Agents, we observed a common product iteration path: from initially focusing on single, specific tasks to later expanding their capabilities to handle more complex, multi-task scenarios. This trend highlights the potential of AI Agents in improving efficiency and innovation, indicating that they will play a more critical role in the future. Based on preliminary statistics of 125 AI Agent projects in Web2, we found that most projects are concentrated in content generation (e.g., Jasper AI), developer tools (e.g., Replit), and B2B services (e.g., Cresta), the largest category. This finding was contrary to our expectations, as we initially predicted that with the increasing maturity of AI model technology, the consumer market (C-end) would experience an explosive growth of AI Agents. However, after further analysis, we realized that the commercialization of consumer AI Agents is much more challenging and complex than expected.

Take Character.AI as an example. On one hand, Character.AI has some of the best traffic performance. However, due to its singular business model—relying on a $9.9 USD subscription fee—it struggled with limited subscription revenue and high inference costs for heavy users, eventually leading to its acquisition by Google due to difficulties in monetizing traffic and maintaining cash flow. This case shows that even with excellent traffic and funding, C-end AI Agent applications face significant commercialization challenges. Most products have not yet reached the standard where they can replace or effectively assist humans, resulting in low user willingness to pay. In our research, we found that many startups encounter similar problems as Character.AI, indicating that the development of consumer AI Agents is not smooth and requires deeper exploration of technical maturity, product value, and business model innovation to unlock their potential in the C-end market.

By counting the valuations of most AI Agent projects, compared with the valuations of ceiling projects such as OpenAI and xAI, there is still room for close to 10-50 times.It is undeniable that the ceiling of C-side Agent application is still high enough, proving that it is still a good track. However, based on the above analysis, we believe that compared with the C-side, the B-side market may be the final destination of AI Agent. By building a platform, enterprises integrate AI Agent into management software such as vertical fields, CRM, and office OA. This not only improves operational efficiency for enterprises, but also provides AI Agent with a broader application space. Therefore, we have reason to believe that B-side services will be the main direction of the short-term development of AI Agents in the Web2 traditional Internet.

Web3 AI Agent Development Status and Prospects

Project overview

As analyzed earlier, even AI Agent applications with top-tier funding and good user traffic face difficulties in commercialization. Next, we will analyze the current development of AI Agent projects in Web3. By evaluating a series of representative projects—their technical innovations, market performance, user feedback, and growth potential—we aim to uncover insightful suggestions. The chart below shows several representative projects that have already issued tokens and hold a relatively high market value:

Compilation of Web2’s AI Agent leading projects, source: ArkStream project database

According to our statistics on the Web3 AI Agent market, the types of projects being developed also show a clear concentration in specific sectors. Most projects fall under infrastructure, with fewer content generation projects. Many of these projects aim to leverage user-provided distributed data and computing power to meet the model training needs of the project owners or to create all-in-one platforms that integrate various AI Agent services and tools. From developer tools to front-end interaction applications and generative applications, most traditional AI Agent industries are currently limited to open-source parameter adjustments or building applications using existing models. This method has not yet generated significant network effects for enterprises or individual users.

Status Analysis

We believe that this phenomenon at this stage may be driven by the following factors:

Market and Technology Mismatch: The combination of Web3 and AI Agents does not currently show a significant advantage over traditional markets. The true advantage lies in improving production relationships by optimizing resources and collaboration through decentralization. This may cause interaction and generative applications to struggle to compete with traditional competitors with stronger technical and financial resources.

Application Scenario Limitations: In the Web3 environment, there may not be as much demand for generating images, videos, or text content. Instead, Web3’s decentralized and distributed features are more often used to reduce costs and improve efficiency within the traditional AI field, rather than to expand into new application scenarios.

The root cause of this phenomenon may lie in the current development state of the AI industry and its future direction. AI technology is still in its early stages, akin to the early days of the industrial revolution when steam engines were replaced by electric motors. It has not yet reached the electrification stage of widespread application.

We believe that the future of AI will likely follow a similar path. General models will gradually become standardized, while fine-tuned models will see diversified development. AI applications will be widely dispersed across enterprises and individual users, with the focus shifting to interconnection and interaction between models. This trend aligns closely with Web3’s principles, as Web3 is known for its composability and permissionless nature, which fits well with the idea of decentralized fine-tuning of models. Developers will have greater freedom to combine and adjust various models. Additionally, decentralization offers unique advantages in areas such as data privacy protection and computing resource allocation for model training.

With technological advancements, especially the emergence of innovations like LoRA (Low-Rank Adaptation), the cost and technical barriers for model fine-tuning have been significantly lowered. This makes it easier to develop public models for specific scenarios or to meet users’ personalized needs. AI Agent projects within Web3 can fully leverage these advancements to explore novel training methods, innovative incentive mechanisms, and new models of model sharing and collaboration, which are often difficult to achieve in traditional centralized systems.

Moreover, the concentration of Web3 projects on model training reflects strategic considerations of its importance within the entire AI ecosystem. Thus, the focus of Web3 AI Agent projects on model training is a natural convergence of technology trends, market demand, and Web3 industry advantages. Next, we will provide examples of model training projects in both Web2 and Web3 industries and make comparisons.

Model Training Projects

Humans.ai

  • Project Overview: Humans.ai is a diversified AI algorithm library and training deployment environment covering various fields such as images, videos, audio, and text. The platform supports developers in further training and optimizing models, and allows them to share and trade their models. A notable innovation is that Humans.ai uses NFTs to store AI models and users’ biometric data, making the AI content creation process more personalized and secure.
  • Data Analysis: The market value of Humans.ai’s token, Heart, is about $68 million. They have 56k followers on Twitter, though user data has not been disclosed.
  • Technical Analysis: Humans.ai does not develop its own models but uses a modular approach, packaging all available models into NFTs, providing users with a flexible and scalable AI solution.

FLock.io

  • Project Overview: FLock.io is an AI co-creation platform based on federated learning technology (a decentralized machine learning method that emphasizes data privacy). It aims to address pain points in the AI field, such as low public engagement, insufficient privacy protection, and the monopoly of AI technology by large corporations. The platform allows users to contribute data while protecting privacy, promoting the democratization and decentralization of AI technology.
  • Data Analysis: FLock.io completed a $6 million seed round in early 2024, led by Lightspeed Faction and Tagus Capital, with additional participation from DCG, OKX Ventures, and others.
  • Technical Analysis: FLock.io’s architecture is based on federated learning, a decentralized method that promotes privacy protection. It also uses zkFL, homomorphic encryption, and secure multi-party computation (SMPC) to provide additional privacy safeguards.

These are examples of model training projects within the Web3 AI Agent space, but similar platforms also exist in Web2, such as Predibase.

Predibase

  • Project Overview: Predibase focuses on AI and large language model optimization, allowing users to fine-tune and deploy open-source large language models, such as Llama, CodeLlama, and Phi. The platform supports various optimization techniques like quantization, low-rank adaptation, and memory-efficient distributed training.
  • Data Analysis: Predibase announced the completion of a $12.2 million Series A round led by Felicis, with major companies like Uber, Apple, Meta, and startups like Paradigm and Koble.ai as platform users.
  • Technical Analysis: Predibase users have trained over 250 models. The platform utilizes the LoRAX architecture and Ludwig framework: LoRAX enables thousands of fine-tuned LLMs to run on a single GPU, significantly reducing costs without affecting throughput or latency. Ludwig is a declarative framework Predibase uses to develop, train, fine-tune, and deploy cutting-edge deep learning and large language models.
  • Project Analysis: Predibase offers user-friendly features that provide customized AI application-building services for different levels of users. Whether for C-end or B-end users, beginners or experienced AI professionals, Predibase caters to a wide range of needs.

For beginners, the platform’s one-click automation simplifies the model-building and training process, automatically handling complex tasks. For experienced users, it provides deeper customization options, including access to and adjustment of more advanced parameters. When comparing traditional AI model training platforms with Web3 AI projects, while their overall frameworks and logic may be similar, we found significant differences in their technical architecture and business models.

  • Technical Depth and Innovation: Traditional AI model training platforms often have deeper technical barriers, such as using proprietary technologies like the LoRAX architecture and Ludwig framework. These frameworks offer robust features, allowing the platform to handle complex AI model training tasks. However, Web3 projects may focus more on decentralization and openness, with less emphasis on deep technical innovation.
  • Business Model Flexibility: A common bottleneck in traditional AI model training is the lack of flexibility in the business model. Platforms typically require users to pay to train models, limiting the project’s sustainability, especially in the early stages when broad user participation and data collection are needed. In contrast, Web3 projects often have more flexible business models, such as tokenomics driven by communities.
  • Privacy Protection Challenges: Privacy protection is another key issue. For example, although Predibase offers virtual private cloud services on AWS, relying on third-party architecture always carries the potential risk of data leaks.

These differences have become bottlenecks in the traditional AI industry. Due to the nature of the internet, these issues are difficult to solve efficiently. At the same time, this presents both opportunities and challenges for Web3, where projects that can solve these problems first will likely become pioneers in the industry.

Other Categories of Web3 AI Agent Projects

After discussing AI Agent projects focused on model training, we now expand our view to other types of AI Agent projects in the Web3 industry. These projects, while not exclusively focused on model training, demonstrate distinctive performance in terms of funding, token valuation, and market presence. Below are some representative and influential AI Agent projects in their respective fields:

Myshell

  • Product Overview: Myshell provides a comprehensive AI Agent platform where users can create, share, and personalize AI agents. These agents can offer companionship and assist with work efficiency. The platform includes diverse AI agent styles, from anime to traditional, and supports interactions via audio, video, and text. A standout feature is the integration of multiple existing models, including GPT-4o, GPT-4, and Claude, providing a premium experience. Additionally, Myshell introduces a trading system similar to FT bonding curves, incentivizing creators to develop high-value AI models while giving users the chance to invest and share in profits.
  • Data Analysis: Myshell’s last funding round valued the company at around $80 million, led by Dragonfly, with participation from Binance, Hashkey, and Folius. With nearly 180K Twitter followers, it has a dedicated community of users and developers, despite Discord engagement being less than one-tenth of its follower base.
  • Technical Analysis: Myshell does not develop AI models independently but serves as an integration platform, combining models like Claude and GPT-4. This strategy allows it to offer users a unified and advanced AI experience.
  • Subjective Experience: MyShell allows users to freely create and customize AI agents according to their own needs. Whether as a personal companion or a professional assistant, it can adapt to various scenarios such as audio and video. Even if users do not use MyShell’s proxy, they can enjoy the integrated Web2 paid model at a lower cost. In addition, the platform combines the economic concept of FT, allowing users to not only use AI services, but also invest in AI agents they are optimistic about, increasing the wealth effect through the bonding curve mechanism.

Delysium

  • Product Overview: Delysium provides an intent-centered AI Agent network, allowing Agents to better cooperate to bring users a friendly Web3 experience. Currently, Delysium has launched two AI Agents: Lucy and Jerry. Lucy is a networked AI Agent. The vision is to provide tool assistance, such as querying the Top 10 currency holding addresses, etc. However, currently the function of the Agent to execute on-chain intentions has not yet been opened, and it can only execute some basic instructions, such as staking AGI within the ecosystem. Or exchange it to USDT. Jerry is similar to GPT in the Delysium ecosystem, and is mainly responsible for answering questions within the ecosystem, such as token distribution.
  • Data Analysis: The first round of financing was US$4 million in 2022, and in the same year it was announced that it had completed strategic financing of US$10 million. Its token AGI currently has an FDV of about $130 million. There is no latest user data. According to official statistics from Delysium, Lucy has accumulated more than 1.4 million independent wallet connections as of June 2023.

Sleepless AI

  • Product Overview: An emotional companionship game platform that combines Web3 and AI Agent technology to provide virtual companion games HIM and HER, using AIGC and LLM to immerse users in interactions with virtual characters. Users can modify the character’s attributes, clothing, etc. during the ongoing conversation. Its compatible large language model ensures that the character iterates on itself in each conversation and becomes more understanding of the user.
  • Data Analysis: The project has raised a total of US$3.7 million, with investors including Binance Labs, Foresight Ventures and Folius Ventures. The current total market value of the tokens has reached approximately US$400 million. It has 116K Twitter followers, 190K registered reservations according to official statistics, and 43K active users. It can be said that its user stickiness is quite strong.
  • Technical Analysis:Although the official did not disclose which major language model on the market their product is based on, Sleepless AI ensures that users will feel that the character understands them more and more during the chat process. Therefore, when designing LLM training, they Each character trains a model separately, and combines the vector database and personality parameter system to allow the character to have memory.
  • Subjective Experience: Sleepless AI approaches AI Boyfriend and AI Girlfriend from a Free-to-Play perspective, and is not just integrated into the chat box of a conversational robot. The project greatly enhances the authenticity of virtual humans through high-cost art, continuously iterative language models, high-quality and complete dubbing, and a series of functions such as alarm clock, sleep aid, menstrual period recording, study companionship, etc. This kind of emotional value cannot be felt by other applications on the market. In addition, Sleepless AI creates a longer-term, balanced content payment mechanism. Users can choose to sell NFT without falling into the dilemma of P2E or Ponzi. This model takes into account both player income and game experience.

Prospect Analysis

In the Web3 industry, AI Agent projects cover multiple directions including public chains, data management, privacy protection, social networks, platform services, and computing power. In terms of token market value, the total token market value of AI Agent projects has reached nearly $3.8 billion, while the total market value of the entire AI track is close to $16.2 billion. AI Agent projects account for about 23% of the market value in the AI track.

Although there are only about a dozen AI Agent projects, which seems relatively few compared to the entire AI track, their market valuation accounts for nearly a quarter. This market value proportion in the AI track once again validates our belief that this sub-track has great growth potential.

After our analysis, we raised a core question: What characteristics do Agent projects need to attract excellent financing and be listed on top exchanges? To answer this question, we explored successful projects in the Agent industry, such as Fetch.ai, Olas Network, SingularityNET, and Myshell.

We found that these projects share some significant features: they all belong to the platform aggregation category within the infrastructure class. They build a bridge, connecting users who need Agents on one end (both B2B and B2C), and developers and validators responsible for model debugging and training on the other end. Regardless of the application level, they have all established a complete ecological closed loop.

We noticed that whether their products are on-chain or off-chain related doesn’t seem to be the most crucial factor. This leads us to a preliminary conclusion: in the Web3 domain, the logic of focusing on practical applications in Web2 may not fully apply. For leading AI Agent products in Web3, building a complete ecosystem and providing diverse functionalities might be more critical than the quality and performance of a single product. In other words, a project’s success depends not only on what it offers but more on how it integrates resources, promotes collaboration, and creates network effects within the ecosystem. This ability to build ecosystems might be a key factor for AI Agent projects to stand out in the Web3 track.

The correct integration method for AI Agent projects in Web3 is not to focus on the deep development of a single application, but to adopt an inclusive model. This approach involves migrating and integrating diverse product frameworks and types from the Web2 era into the Web3 environment to build a self-cycling ecosystem. This point can also be seen in OpenAI’s strategic shift, as they chose to launch an application platform this year rather than just updating their model.

In summary, we believe that the AI ​​Agent project should focus on the following aspects:

  • Ecosystem Building: Go beyond single applications to build an ecosystem that includes multiple services and functions, promoting interaction and value addition between different components.
  • Tokenomic Model: Design a reasonable token economic model to incentivize users to participate in network construction and contribute data and computing power.
  • Cross-domain Integration: Explore the potential applications of AI Agents in different fields, creating new usage scenarios and value through cross-domain integration.

After summarizing these three aspects, we also provide some forward-looking suggestions for project teams with different focus directions: one for non-AI core application products, and another for native projects focused on the AI Agent track.

For non-AI core application products:

Maintain a long-term perspective, focus on their core products while integrating AI technology, and wait for the right opportunity in line with the times. In the current technological and market trends, we believe that using AI as a traffic medium to attract users and enhance product competitiveness has become an important means of competitiveness. Although the actual long-term contribution of AI technology to project development remains a question mark, we believe this provides a valuable window for early adopters of AI technology. Of course, the premise is that they already have a very solid product.

In the long run, if AI technology achieves new breakthroughs in the future, those projects that have already integrated AI will be able to iterate their products more quickly, thus seizing opportunities and becoming industry leaders. This is similar to how live streaming e-commerce gradually replaced offline sales as a new traffic outlet on social media platforms in recent years. At that time, those merchants with solid products who chose to follow the new trend and try live streaming e-commerce immediately stood out with the advantage of early entry when live streaming e-commerce truly exploded.

We believe that amid market uncertainty, for non-AI core application products, considering the timely introduction of AI Agents may be a strategic decision. It can not only increase the product’s market exposure at present but also bring new growth points for the product in the continuous development of AI technology.

For native projects focused on AI Agents:

Balancing technological innovation and market demand is key to success. In native AI Agent projects, project teams need to look towards market trends, not just technology development. Currently, some Web3-integrated Agent projects in the market may be overly focused on developing in a single technical direction, or have constructed a grand vision, but product development has not kept up. Both of these extremes are not conducive to the long-term development of the project.

Therefore, we suggest that project teams, while ensuring product quality, should also pay attention to market dynamics, and realize that the AI application logic in the traditional internet industry may not apply to Web3. Instead, they need to learn from those projects that have already achieved results in the Web3 market. Focus on the labels they have, such as model training and platform aggregation core functions mentioned in the article, as well as the narratives they create, such as AI modularization and multi-Agent collaboration. Exploring compelling narratives may become the key for projects to achieve breakthroughs in the market.

Conclusion

Whether it is a non-AI core product or a native AI Agent project, the most critical thing is to find the right timing and technical path to ensure that it remains competitive and innovative in the ever-changing market. On the basis of maintaining product quality, project parties should observe market trends, learn from successful cases, and at the same time innovate to achieve sustainable development in the market.

Summary

At the end of the article, we analyze the Web3 AI Agent track from multiple angles:

  • Capital investment and market attention: Although AI Agent projects currently do not have an advantage in the number of listings in the Web3 industry, they account for close to 50% of the market valuation, showing that the capital market highly recognizes this track. With more capital investment and increasing market attention, it is certain that more high-valued projects will appear in the AI ​​Agent track.
  • Competitive landscape and innovation capabilities: The competitive landscape of the AI ​​Agent track in the Web3 industry has not yet been fully formed. At the current application level, there is no phenomenal and leading product similar to ChatGPT. This gives new project parties a lot of room for growth and innovation. As the technology matures and previous projects are innovated, the track is expected to develop more competitive products, driving up the valuation of the entire track.
  • Pay attention to tokenomics and user incentives: The significance of Web3 is to reshape production relations and make the originally centralized process of deploying and training AI models more decentralized. Through reasonable tokenomics design and user incentive programs, idle computing power or personal datasets can be redistributed. Additionally, solutions like ZKML can protect data privacy, further reducing computing power and data costs, and allowing more individual users to participate in the construction of the AI industry.

To sum up, we are optimistic about the AI ​​Agent track. We have reason to believe that multiple projects with valuations exceeding $1 billion will emerge in this track. Through horizontal comparison, the narrative of AI Agent is sufficiently compelling and the market space is large enough. The current market valuations are generally low. Considering the rapid development of AI technology, the growth of market demand, capital investment and the innovation potential of companies in the track, in the future, as technology matures and market acceptance increases, this track is expected to see multiple projects with valuations over $1 billion emerge.

statement:

  1. This article is reproduced from [ArkStream Capital], the original title is “ArkStream Capital Track Research Report: Can AI Agent be a life-saving straw for Web3+AI?” If you have any objections to the reprint, please contact Gate Learn Team, the team will handle it as soon as possible according to relevant procedures.

  2. Disclaimer: The views and opinions expressed in this article represent only the author’s personal views and do not constitute any investment advice.

  3. Other language versions of the article are translated by the Gate Learn team, not mentioned in Gate.io, the translated article may not be reproduced, distributed or plagiarized.

Start Now
Sign up and get a
$100
Voucher!