Summary of Key Points
In this report, we discuss the development patterns of major mainstream architectures within the Crypto & AI industry. We will examine the current four mainstream architectures—Eliza (AI16Z), G.A.M.E (GAME), Rig (ARC), and ZerePy (ZEREBRO)—analyzing their technical differences and development potentials.
In the past week, we have analyzed and tested the four major architectures mentioned above, and the conclusions are summarized as follows.
We believe Eliza (with a market share of around 60%, valued at about $900 million when the original author wrote this, and approximately $1.4 billion at the time of publication) will continue to dominate market share. Eliza’s value lies in its first-mover advantage and accelerated adoption by developers, evidenced by 193 contributors, 1,800 forks, and over 6,000 stars on GitHub, making it one of the most popular repositories on GitHub.
G.A.M.E (with a market share of around 20%, valued at about $300 million when the original author wrote this, and approximately $257 million at the time of publication) has developed smoothly so far and is experiencing rapid adoption. As announced earlier by the Virtuals Protocol, there are over 200 projects built on G.A.M.E, with daily request counts exceeding 150,000 and a weekly growth rate of over 200%. G.A.M.E will continue to benefit from the explosion of VIRTUAL and has the potential to become one of the biggest winners in that ecosystem.
Rig (with a market share of around 15%, valued at about $160 million when the original author wrote this, and approximately $279 million at the time of publication) features a modular design that is very appealing and easy to operate, with the potential to dominate in the Solana ecosystem (RUST).
Zerepy (with a market share of around 5%, valued at about $300 million when the original author wrote this, and approximately $424 million at the time of publication) is a more niche application specific to a passionate ZEREBRO community, and its recent collaboration with the ai16z community may generate some synergistic effects.
In the above statistics, 'market share' is calculated considering market capitalization, development records, and the breadth of the underlying operating system's end market.
We believe AI architectures will become the fastest-growing sector in this cycle, with the current total market value of approximately $1.7 billion likely to grow easily to $20 billion. Compared to the peak valuation of Layer 1 in 2021, this figure may still be conservative—at that time, many individual projects were valued at over $20 billion. Although the aforementioned architectures serve different end markets (chains/ecosystems), we believe that this sector will grow overall, making a market cap-weighted approach relatively cautious.
Four Major Architectures
At the intersection of AI and Crypto, several architectures aimed at accelerating AI development have emerged, including Eliza (AI16Z), G.A.M.E (GAME), Rig (ARC), and ZerePy (ZEREBRO). From open-source community projects to performance-focused enterprise solutions, each architecture caters to different needs and philosophies in agent development.
In the table below, we outline the key technologies, components, and advantages of each architecture.
Image source: Deep Value Memetics
This report will first focus on what these architectures are, the programming languages they use, their technical architectures, algorithms, and unique features with potential applications. Then we will compare each architecture based on usability, scalability, adaptability, and performance while discussing their strengths and limitations.
Eliza
Eliza is an open-source multi-agent simulation architecture developed by ai16z, designed to create, deploy, and manage autonomous AI agents. It is developed in TypeScript, providing a flexible, scalable platform for building intelligent agents that can interact with humans across multiple platforms while maintaining consistent personality and knowledge.
The core functionalities of the architecture include: support for the simultaneous deployment and management of multiple unique AI personalities in a multi-agent architecture; creating a diverse agent role system using role file architecture; providing long-term memory and context-aware memory management capabilities through an advanced retrieval-augmented generation system (RAG). Additionally, the Eliza architecture offers smooth platform integration, enabling reliable connections with Discord, X, and other social media platforms.
In terms of communication and media capabilities for AI agents, Eliza is an excellent choice. In communication, the architecture supports integration with voice channel functions from Discord, X features, Telegram, and direct API access for custom applications. On the other hand, the media processing capabilities of the architecture have been expanded to include PDF document reading and analysis, link content extraction and summarization, audio transcription, video content processing, image analysis, and conversation summarization, effectively handling various media inputs and outputs.
Eliza provides flexible AI model support, allowing for local inference using open-source models, cloud-based inference via default configurations such as OpenAI and Nous Hermes Llama 3.1 B, and integration with Claude for handling complex queries. Eliza adopts a modular architecture with a wide range of action systems, custom client support, and comprehensive APIs, ensuring scalability and adaptability across applications.
Eliza's applications span multiple industries, such as AI assistants related to customer support, community management, and personal tasks; automated content creators, brand representatives, and other social media roles; it can also serve as knowledge workers, acting as research assistants, content analysts, and document processors; as well as interactive roles such as role-playing robots, educational tutors, and entertainment agents.
Eliza's architecture is built around an agent runtime that integrates seamlessly with a role system (supported by model providers), a memory manager (connected to databases), and an action system (linked to platform clients). The architecture's unique features include an expanded application system that allows for modular functionality extensions, supporting multimodal interactions such as voice, text, and media, and compatibility with leading AI models like Llama, GPT-4, and Claude. With its multifunctionality and robust design, Eliza has become a powerful tool for developing AI applications across industries.
G.A.M.E
G.A.M.E, developed by the official Virtuals team, stands for 'Generative Autonomous Multimodal Entities Framework,' aimed at providing developers with APIs and SDKs to experiment with AI agents. The architecture offers a structured approach to managing AI agent behavior, decision-making, and learning processes.
The core components of G.A.M.E include first, the 'Agent Prompting Interface,' which is the entry point for developers to integrate G.A.M.E into agents to obtain agent behavior.
The 'Perception Subsystem' initiates sessions by specifying parameters such as session ID, agent ID, user, and other relevant details. It synthesizes incoming messages into a format suitable for the 'Strategic Planning Engine,' serving as the sensory input mechanism for AI agents, whether in the form of dialogue or responses. The core here is the 'Dialogue Processing Module,' responsible for handling messages and responses from the agent and collaborating with the 'Perception Subsystem' to effectively interpret and respond to inputs.
The 'Strategic Planning Engine' collaborates with the 'Dialogue Processing Module' and the 'On-chain Wallet Operator' to generate responses and plans. The engine operates on two levels: as a high-level planner, creating broad strategies based on context or goals; as a low-level strategist, translating these strategies into executable policies, further subdivided into action planners (for specifying tasks) and plan executors (for executing tasks).
A separate but critical component is 'World Context,' which references environmental, world information, and game state, providing necessary context for the agent's decision-making. Additionally, the 'Agent Library' is used to store long-term attributes like goals, reflections, experiences, and personalities, which collectively shape the agent's behavior and decision-making process. The architecture employs 'Short-term Working Memory' and 'Long-term Memory Processors'—short-term memory retains relevant information about previous actions, outcomes, and current plans; in contrast, the long-term memory processor extracts key information based on criteria such as significance, recency, and relevance. This memory stores knowledge about the agent's experiences, reflections, dynamic personality, world context, and working memory to enhance decision-making and provide a foundation for learning.
To enhance layout, the 'Learning Module' retrieves data from the 'Perception Subsystem' to generate general knowledge, which is fed back into the system to optimize future interactions. Developers can input feedback on actions, game states, and sensory data through the interface to enhance the AI agent's learning and improve its planning and decision-making abilities.
The workflow begins with developers interacting through the Agent Prompting Interface; the 'Perception Subsystem' processes the input and forwards it to the 'Dialogue Processing Module,' which manages the interaction logic; then, the 'Strategic Planning Engine' formulates and executes plans using high-level strategies and detailed action planning based on this information.
Data from the 'World Context' and 'Agent Library' provide information for these processes, while working memory tracks real-time tasks. Meanwhile, the 'Long-Term Memory Processor' stores and retrieves knowledge over time. The 'Learning Module' analyzes results and integrates new knowledge into the system, allowing the agent's behavior and interactions to continuously improve.
Rig
Rig is an open-source architecture based on Rust, designed to simplify the development of large language model (LLM) applications. It provides a unified interface for interacting with multiple LLM providers (such as OpenAI and Anthropic) and supports various vector storage options, including MongoDB and Neo4j. The architecture's modular design features core components such as the 'Provider Abstraction Layer,' 'Vector Storage Integration,' and 'Agent System,' which facilitate seamless interactions with LLMs.
Rig's primary audience includes developers building AI/ML applications using Rust, while a secondary audience includes organizations seeking to integrate multiple LLM providers and vector storage into their Rust applications. The repository is organized based on a workspace structure and includes multiple crates, achieving scalability and efficient project management. Rig's main features include a 'Provider Abstraction Layer,' which standardizes the APIs used for completion and embedding LLM providers through consistent error handling; a 'Vector Storage Integration' component that provides an abstract interface for multiple backends and supports vector similarity searches; and a 'Agent System' that simplifies LLM interactions, supporting retrieval-augmented generation (RAG) and tool integration. Additionally, the embedded architecture offers batch processing capabilities and type-safe embedding operations.
Rig leverages multiple technical advantages to ensure reliability and performance. Asynchronous operations utilize Rust's asynchronous runtime to efficiently handle a large number of concurrent requests; the architecture's inherent error-handling mechanisms enhance recovery capabilities from AI provider or database operation failures; type safety prevents compile-time errors, improving code maintainability; efficient serialization and deserialization processes aid in handling data in formats such as JSON, which is crucial for communication and storage in AI services; detailed logging and instrumentation further assist in debugging and monitoring applications.
The workflow in Rig begins with a client initiating a request, which flows through the 'Provider Abstraction Layer' for interaction with the corresponding LLM model; then, the data is processed by the core layer, where the agent can use tools or access vector storage for context; responses are generated and refined through complex workflows like RAG, including document retrieval and context understanding, and then returned to the client. The system integrates multiple LLM providers and vector storage, adapting to changes in model availability or performance.
Rig's applications are diverse, including question-answering systems that retrieve relevant documents for accurate responses, document search and retrieval for efficient content discovery, and context-aware interactions for customer service or education through chatbots or virtual assistants. It also supports content generation, capable of creating text and other materials based on learned patterns, making it a multifunctional tool for developers and organizations.
ZerePy
ZerePy is an open-source framework written in Python, designed to deploy agents on X utilizing OpenAI or Anthropic LLMs. ZerePy originated from a modular version of the Zerebro backend, allowing developers to launch agents with functionalities similar to Zerebro's core features. While the architecture provides a foundation for deploying agents, fine-tuning of the models is necessary to generate creative outputs. ZerePy simplifies the development and deployment of personalized AI agents, particularly suitable for content creation on social platforms, fostering an AI creative ecosystem targeting art and decentralized applications.
The architecture is built using the Python language, emphasizing the autonomy of agents and focusing on generating creative outputs, consistent with Eliza’s architecture + partnership. Its modular design supports memory system integration, facilitating the deployment of agents on social platforms. Its main features include a command line interface for agent management, integration with X, support for OpenAI and Anthropic LLMs, and a modular connection system for enhanced functionality.
ZerePy's applications cover social media automation, where users can deploy AI agents for posting, replying, liking, and sharing to increase platform engagement. Additionally, it is suitable for content creation in industries such as music, notes, and NFTs, serving as an important tool for digital art and blockchain-based content platforms.
Horizontal Comparison
In our view, each of the aforementioned architectures offers a unique approach to AI development, catering to specific needs and environments. This shifts the debate from whether these architectures are competitors to whether each architecture provides unique utility and value.
Eliza stands out with its user-friendly interface, particularly suited for developers familiar with JavaScript and Node.js environments. Its comprehensive documentation aids in setting up AI agents across various platforms, although its rich feature set may present a moderate learning curve. With TypeScript used, Eliza is particularly suitable for building agents embedded in the web, as most frontend web infrastructure is built with TypeScript. The architecture is known for its multi-agent structure, capable of deploying diverse AI personality agents across platforms such as Discord, X, and Telegram. Its advanced RAG system is used for memory management, making it particularly suitable for building AI assistants for customer support or social media applications. While it offers flexibility, strong community support, and consistent cross-platform performance, it is still in the early stages, potentially posing a learning curve for developers.
G.A.M.E is designed for game developers, providing a low-code or no-code interface through APIs, facilitating access for users with lower technical proficiency within the gaming industry. However, it focuses on game development and blockchain integration, making the learning curve potentially steep for those without relevant experience. It excels in programmatic content generation and NPC behavior but is also limited by its niche industry and the added complexities present in blockchain integration.
Rig may be less user-friendly due to its use of the Rust language, which presents significant challenges for learning due to its complexity, but it can provide intuitive interactions for those proficient in systems programming. Compared to TypeScript, Rust is known for its performance and memory safety. It has strict compile-time checks and zero-cost abstractions, which are essential for running complex AI algorithms. The efficiency and low-level control of the language make it an ideal choice for resource-intensive AI applications. The architecture is designed with modularity and scalability in mind, offering high-performance solutions, making it particularly suitable for enterprise applications. However, for developers unfamiliar with Rust, using it can present a steep learning curve.
ZerePy uses the Python language, providing higher usability for creative AI tasks. For Python developers, especially those with AI/ML backgrounds, the learning curve is low, and strong community support is available due to the popularity of ZEREBRO. ZerePy excels in creative AI applications such as NFTs, positioning itself as a powerful tool for the digital media and arts industries. While it excels creatively, its range of applications is relatively narrow compared to other architectures.
In terms of scalability, the four major architectures compare as follows.
Eliza has made significant progress after the V2 version update, introducing a unified message line and an expandable core architecture, achieving efficient management across platforms. However, without optimization, managing such multi-platform interactions may present challenges in scalability.
G.A.M.E excels in real-time processing required in gaming, and its scalability can be managed through efficient algorithms and potential decentralized blockchain systems, though it may be constrained by specific game engines or blockchain networks.
The Rig architecture leverages Rust's performance advantages to achieve better scalability, inherently designed for high-throughput applications, which may be particularly effective for enterprise-level deployments. However, this may imply that achieving true scalability requires complex setups.
ZerePy's scalability focuses on creative output and is supported by community contributions, but the architecture's emphasis may limit its application in broader AI environments, as its scalability may be tested by the diversity of creative tasks rather than user volume.
In terms of applicability, Eliza is far ahead with its expanded application system and cross-platform compatibility, followed by G.A.M.E in gaming environments and Rig in handling complex AI tasks. ZerePy shows high adaptability in the creative industry but is less applicable in broader AI application industries.
In terms of performance, the testing results of the four major architectures are as follows.
Eliza is optimized for fast interactions on social media, but its performance may vary when handling more complex computational tasks.
G.A.M.E focuses on high-performance real-time interactions in gaming scenarios, leveraging efficient decision-making processes and potentially decentralized AI operations through blockchain.
Rig, based on Rust, can deliver excellent performance for high-performance computing tasks, suitable for enterprise applications where computational efficiency is crucial.
ZerePy's performance targets the creation of creative content, with its metrics centered on the efficiency and quality of content generation, which may not be as applicable outside the creative industry.
Combining the above advantages and disadvantages, Eliza provides better flexibility and scalability. Its expanded application system and role configurations give it a strong adaptability, beneficial for cross-platform social AI interactions; G.A.M.E offers unique real-time interaction capabilities in gaming scenarios and provides novel AI participation through blockchain integration; Rig excels in performance and scalability, suitable for enterprise-level AI tasks, and emphasizes code simplicity and modularity to ensure the long-term healthy development of projects; Zerepy excels at fostering creativity, leading in AI applications in digital art, supported by a vibrant community-driven development model.
In summary, each architecture has its limitations. Eliza is still in the early stages and may face potential stability issues, with a longer learning curve for new developers; G.A.M.E's niche focus could limit its broader application, especially with the introduction of blockchain increasing complexity; Rig's learning curve is steeper due to the complexity of the Rust language, which may deter some developers; Zerepy's narrow focus on creative output may limit its application in other AI industries.
Core Comparison Items
Rig (ARC)
Language: Rust, focusing on safety and performance.
Applications: Emphasizing efficiency and scalability, ideal for enterprise-level AI applications.
Community: Less community-driven, focusing more on technical developers.
Eliza (AI16Z)
Language: TypeScript, emphasizing flexibility and community engagement in Web3.
Applications: Designed specifically for social interaction, DAOs, and transactions, with a particular emphasis on multi-agent systems.
Community: Highly community-driven, with extensive connections to GitHub.
ZerePy (ZEREBRO):
Language: Python, more easily accepted by a broader AI developer community.
Applications: Suitable for social media automation and simpler AI agent tasks.
Community: Relatively new, but due to the popularity of Python and support from ai16z contributors, it is expected to grow.
G.A.M.E (VIRTUAL, GMAE):
Key Point: Autonomous, adaptive AI agents that can evolve based on interactions within virtual environments.
Applications: Most suitable for scenarios where agents need to learn and adapt, such as games or virtual worlds.
Community: Innovative but still determining its positioning in the competitive landscape.
GitHub Data Growth
Image source: Deep Value Memetics
The above chart shows the changes in star data on GitHub since these architectures were launched. Generally, GitHub stars can serve as indicators of community interest, project popularity, and perceived value of the project.
Eliza (red line): The chart shows significant and stable growth in the number of stars for this architecture, starting from a low base in July and surging in late November, now reaching 6,100 stars. This indicates a rapid increase in interest surrounding this architecture, attracting the attention of developers. The exponential growth suggests that Eliza has gained tremendous appeal due to its features, updates, and community engagement, far surpassing other products, indicating strong community support and broader applicability or interest within the AI community.
Rig (blue line): Rig is the most 'established' of the four architectures, showing modest but stable growth in stars, with a noticeable increase in the past month. Its total star count has reached 1,700, but it is still on an upward trajectory. The steady accumulation of attention is attributed to ongoing development, updates, and a growing user base. This may reflect that Rig is still accumulating its reputation.
ZerePy (yellow line): ZerePy was launched just a few days ago, and its star count has grown to 181. It is important to emphasize that ZerePy needs more development to improve its visibility and adoption rate, and collaboration with ai16z may attract more contributors to its codebase.
G.A.M.E (green line): The architecture has few stars, but it is noteworthy that it can directly apply to agents in the Virtual ecosystem via API, thus not requiring publication on GitHub. However, although the architecture has only been publicly available for builders for just over a month, there are already over 200 projects using G.A.M.E for development.
Expectations for Upgrades in AI Architectures
Eliza's version 2.0 will include integration with the Coinbase agent toolkit. All projects using Eliza will gain support for future native TEE (Trusted Execution Environment), enabling agents to run in a secure environment. An expanded plugin registry is a feature that Eliza will soon introduce, allowing developers to seamlessly register and integrate extended applications.
Additionally, Eliza 2.0 will support automated anonymous cross-platform messaging. The Tokenomics white paper, expected to be released on January 1, 2025 (related proposals have been announced), will have a positive impact on the AI16Z tokens supporting the Eliza architecture. ai16z plans to continue enhancing the architecture's usability and leverage the efforts of its main contributors to bring in high-quality talent.
The G.A.M.E framework provides agents with no-code integration, allowing G.A.M.E and Eliza to be used simultaneously within a single project, each serving specific applications. This approach is expected to attract builders focused on business logic rather than technical complexities. Although the architecture has been publicly available for just over 30 days, substantial progress has been made with the team's efforts to attract more contributors. It is anticipated that every project launched on VirtuaI will adopt G.A.M.E.
The Rig architecture, powered by the ARC token, has significant potential, although its growth is still in the early stages, and the project contract plans driving Rig adoption have only been launched for a few days. However, high-quality projects are expected to emerge soon, similar to the Virtual flywheel but focusing on Solana. The Rig team is optimistic about collaborating with Solana, positioning ARC as Solana's Virtual. Notably, the team incentivizes not only new projects initiated using Rig but also encourages developers to enhance the Rig architecture itself.
Zerepy is a newly launched architecture that is gaining significant attention due to its collaboration with ai16z (Eliza architecture). This architecture has attracted contributors from Eliza who are actively working to improve it. Zerepy enjoys enthusiastic support driven by the ZEREBRO community and is opening new opportunities for Python developers previously lacking room to operate in the highly competitive AI infrastructure industry. It is expected that this architecture will play an important role in AI creativity.
This article is reprinted with permission from: (BlockBeats)
Original author: Deep Value Memetics
'Developers Must Read! Comparison of Four Major Crypto AI Development Architectures: Eliza, ZerePy, Who is the Best?' This article was first published in 'Crypto City.'