AI Agent Redefining the Innovation Path of Web3 Games

Game-changing AI Agent Revolutionizing the Innovation Journey of Web3 Games

Author: PSE Trading Analyst @Minta, Mirror

Key Insights

1. The AI Agent is a tool based on the LLM general model that allows developers and users to directly build autonomous interactive applications.

2. The future main landscape of the AI track may be “General Model + Vertical Application”; the ecological position of the AI Agent is the middleware that connects the general model and Dapps, so the AI Agent’s moat is relatively low, and it needs to rely on creating network effects and improving user stickiness to enhance long-term competitiveness.

3. This article summarizes the development of “General Model, Vertical Application Agent, and Generative AI Application” in the Web3 gaming track. Among them, combining Generative AI technology, it has great potential to release popular games in the short term.

   • Finding Balance Exploring Decentralization in Web3
   • Has Web3 hit rock bottom? Listen to what they have to say
   • Web3.0 should be a new network that maintains the global community of common destiny.

01 Technical Introduction

In the AGI (Artificial General Intelligence) technology that has exploded this year, the Large Language Model (LLM) is absolutely the protagonist. OpenAI’s core technical personnel, Andrej KarLianGuaithy and Lilian Weng, have also expressed that AI Agents based on LLM are an important development direction in the AGI field, and many teams are developing AI Agents systems driven by LLM. In simple terms, an AI Agent is a computer program that uses a large amount of data and complex algorithms to simulate human thinking and decision-making processes in order to perform various tasks and interactions, such as autonomous driving, speech recognition, and game strategies. Abacus.ai’s picture provides a clear introduction to the basic principles of AI Agents, which are as follows:

1. Perception and Data Collection: Data input, or AI Agents obtain information and data through perception systems (sensors, cameras, microphones, and other devices), such as game states, images, sounds, etc.

2. State Representation: Data needs to be processed and represented in a form that the Agent can understand, such as converting it into vectors or tensors for input into neural networks.

3. Neural Network Model: Deep neural network models are usually used for decision-making and learning, such as using Convolutional Neural Networks (CNN) for image processing, Recurrent Neural Networks (RNN) for sequence data processing, or more advanced models such as self-attention mechanisms (Transformer), etc.

4. Reinforcement Learning: Agents learn the best action strategies through interaction with the environment. In addition, the operation principles of Agents also include policy networks, value networks, training and optimization, exploration and exploitation, etc. For example, in a game scenario, the policy network can input game states and output action probability distributions; the value network can estimate state values; the Agent can continuously optimize policies and value networks through interaction with the environment using reinforcement learning algorithms to output more perfect results.

In short, AI Agents are intelligent entities that can understand, make decisions, and take action. They can play important roles in various domains, including the gaming field. The comprehensive introduction to AI Agents titled “LLM Powered Autonomous Agents” written by Lilian Weng, a core technical personnel at OpenAI, mentions a very interesting experiment: Generative Agents.

The inspiration for Generative Agents (GA) comes from the game “The Sims”. It uses LLM technology to generate 25 virtual characters, each controlled by an Agent supported by LLM, living and interacting in a sandbox environment. The design of GA is clever, as it combines LLM with memory, planning, and reflection functions, allowing the Agent program to make decisions based on previous experiences and interact with other Agents.

The article details how the Agent continuously trains and optimizes decision paths based on strategy networks, value networks, and interactions with the environment.

The principle is as follows: the memory stream is a long-term memory module that records all of the Agent’s interactions. The retrieval model provides experiences (retrieved memories) to help the Agent make decisions based on relevance, freshness, and importance. The reflection mechanism summarizes past events to guide the Agent’s future actions. The plan and reflect functions work together to convert reflection and environmental information into actual actions.

This interesting experiment showcases the capabilities of AI Agents, such as generating new social behaviors, information dissemination, relationship memory (e.g., two virtual characters continuing a discussion), and coordination of social activities (e.g., hosting parties and inviting other virtual characters). In short, AI Agents are a fascinating tool, and their applications in games are worth exploring further.

02 Technological Trends

2.1 AI Track Trends

LaoBai, a partner at ABCDE, has summarized the judgments of the Silicon Valley venture capital community regarding the next steps in AI development:

1. There are no specific models, only large models + specific applications;

2. Edge devices, such as data from mobile devices, may pose a barrier, but also present an opportunity for AI;

3. The length of context may cause a qualitative change in the future (currently, vector databases are used as AI memory, but the length of context is still not enough).

In other words, from the perspective of the general development trend of the industry, because large and general-purpose models are too heavy and have strong universality, there is no need to constantly reinvent the wheel in the field of large and general-purpose models. Instead, the focus should be on applying these models to specific domains.

At the same time, edge devices refer to terminal devices that do not rely on cloud computing centers or remote servers, but perform data processing and decision-making locally. Due to the diversity of edge devices, deploying AI Agents to run on devices and properly obtaining device data is a challenge, but also a new opportunity.

Lastly, the issue of context has received much attention. In the context of LLM, context can be understood as the amount of information, and the length of context can be understood as the number of dimensions the data has. For example, consider a big data model for an e-commerce website used to predict the probability of a user purchasing a certain product. In this case, the context could include the user’s browsing history, purchase history, search records, user attributes, and other information. The length of context refers to the dimensions of the overlapping feature information, such as the purchase history of a 30-year-old male user from Shanghai, overlaid with the frequency of recent purchases, and then overlaid with the recent browsing records. Increasing the length of context can help the model better understand the influencing factors of user purchasing decisions.

The current consensus is that although using vector databases as AI memory makes the context length insufficient, the future context length will undergo qualitative changes, and the LLM model can seek more advanced methods to handle and understand longer and more complex context information. This will further give rise to more unimaginable application scenarios.

2.2 Trends in AI Agents

Folius Ventures has summarized the application patterns of AI Agents in the game track, as shown in the following figure:

In the figure, number 1 represents the LLM model, which is mainly responsible for converting user intentions from traditional keyboard/click input into natural language input, thereby reducing the user’s entry barrier.

Number 2 in the figure is a front-end Dapp integrated with an AI Agent, which provides functional services to users and also collects user habits and data from terminals.

Number 3 in the figure represents various types of AI Agents, which can exist directly in the form of in-app features, bots, and so on.

In summary, AI Agents, as code-based tools, can serve as the underlying programs for expanding Dapp’s application functions and catalysts for platform growth, acting as intermediaries linking large models and vertical applications.

In terms of user scenarios, Dapps that are most likely to integrate AI Agents are likely to be open social apps, chatbots, and games. Alternatively, existing Web2 traffic gateways can be transformed into simpler and more user-friendly AI+Web3 gateways through AI Agents. This is an ongoing discussion in the industry to lower the user barrier of Web3.

Based on the laws of industry development, the middleware layer where AI Agents exist often becomes a highly competitive track with almost no moat. Therefore, in addition to continuously improving the user experience to meet B2C demand, AI Agents can enhance their own moats by creating network effects or user stickiness.

03 Track Map

AI has been tried in various ways in the field of Web3 games, and these attempts can be divided into the following categories:

1. General Models: Some projects focus on building general AI models that find suitable neural network architectures and general models for Web3 projects’ needs.

2. Vertical Applications: Vertical applications aim to solve specific problems in games or provide specific services, usually in the form of agents, bots, and bot kits.

3. Generative AI Applications: The most direct application of large models is content generation, and the game track itself is an industry of content. Therefore, the Generative AI applications in the game field deserve great attention. From automatically generating elements, characters, tasks, or storylines in virtual worlds to automatically evolving game strategies, decisions, and even the in-game ecosystem, it becomes possible to make games more diverse and profound.

4. AI Games: Currently, many games have integrated AI technology, with various application scenarios, which will be explained in the following sections.

3.1 General Big Models

Currently, Web3 has simulation models for economic model design and economic ecosystem development, such as the QTM Quantitative Token Model. Dr. Achim Struve from Outlier Venture mentioned some points about economic model design in his speech at ETHCC. For example, considering the robustness of the economic system, project teams can create a Digital Twin, a digital replica of the entire ecosystem, using the LLM model for 1:1 simulation.

The QTM (Quantitative Token Model) shown in the following image is an AI-driven inference model. QTM adopts a fixed simulation time of 10 years, with each time step being one month. At the beginning of each time step, tokens are emitted into the ecosystem, so the model includes incentive modules, token ownership modules, airdrop modules, etc. These tokens are then distributed into several meta buckets for further generalized utility redistribution. Rewards payments are defined from these utility tools. As for off-chain business aspects, general financial conditions are also considered, such as the ability to burn or buy back tokens, as well as measuring user adoption or defining user adoption situations.

Naturally, the output quality of this model depends on the input quality, so thorough market research must be conducted before using QTM to obtain more accurate input information. However, the QTM model has already become a widely applied AI-driven model in Web3 economic models, and many project teams have developed 2C/2B applications with lower operational difficulties based on the QTM model, reducing the entry barriers for project teams.

3.2 Vertical Applications Agent

Vertical applications mainly exist in the form of agents, which can be in various forms such as bots, BotKits, virtual assistants, intelligent decision support systems, and various automated data processing tools. Generally, AI agents are based on OpenAI’s general model and combined with other open source or self-developed technologies, such as text-to-speech (TTS), and include specific data for fine-tuning (a training technique in the field of machine learning and deep learning, primarily aimed at further optimizing a model that has already been pre-trained on a large-scale dataset), to create AI agents that perform better than ChatGPT in a specific domain.

Currently, the most mature application in the Web3 gaming space is the NFT agent. The consensus in the gaming space is that NFTs are an important component of Web3 games.

With the development of metadata management technology in the Ethereum ecosystem, programmable dynamic NFTs have emerged. For NFT creators, they can make NFTs more flexible through algorithms. For users, there can be more interaction between users and NFTs, and the generated interaction data becomes a source of information. AI agents can optimize the interaction process and expand the application scenarios of interaction data, injecting more innovation and value into the NFT ecosystem.

Case One: For example, the development framework of Gelato allows developers to customize logic and update the metadata of NFTs based on off-chain events or specific time intervals. Gelato nodes trigger the change of metadata when specific conditions are met, thus achieving automatic updates for on-chain NFTs. For instance, this technology can be used to fetch real-time match data from a sports API and automatically upgrade the skill features of an NFT when a player wins a game.

Case Two: LianGuaiima also provides application-based agents for Dynamic NFTs. LianGuaiima’s NFT compression protocol mints a set of minimal NFTs on L1 (Layer 1) and then evolves them based on the game state on L2 (Layer 2), providing players with a deeper and interactive gaming experience. For example, an NFT can change based on factors such as the character’s experience points, completion of tasks, and equipment.

Case Three: Mudulas Labs is a well-known project in the ZKML (Zero-Knowledge Machine Learning) space and has also made efforts in the NFT field. Mudulas has launched the NFT series zkMon, allowing AI-generated NFTs to be deployed on-chain along with the generation of a ZKP (Zero-Knowledge Proof). Users can use the ZKP to verify whether their NFT was generated from the corresponding AI model. For more comprehensive information, please refer to Chapter 7.2: The World’s 1st zkGAN NFTs.

3.3 Generative AI Applications

As mentioned earlier, as games are part of the content industry, AI agents can generate a large amount of content in a short time and at low cost. This includes creating game characters with uncertainty and dynamism. Therefore, Generative AI is well-suited for gaming applications. Currently, the application of Generative AI in the gaming industry can be summarized into the following main types:

1. AI-Generated Game Characters: For example, battling against AI or having AI simulate and control NPCs (Non-Player Characters) in the game, or even directly generating characters using AI.

2. AI-Generated Game Content: Directly generating various game content such as missions, storylines, props, maps, etc., using AI.

3. AI-Generated Game Scenes: Supporting the use of AI to generate, optimize, or expand the terrain, landscapes, and atmosphere of the game world.

3.3.1 AI-Generated Characters

Case One: MyShell

MyShell is a platform for creating bots that users can customize according to their needs, including chatting, practicing speaking, playing games, seeking psychological counseling, and more. MyShell also utilizes text-to-speech (TTS) technology, which can automatically create bots that mimic anyone’s voice with just a few seconds of voice samples. Additionally, MyShell uses AutoPrompt, allowing users to give instructions to the Large Language Model (LLM) by simply describing their ideas, laying the foundation for a personal large-scale language model (LLM).

Users of MyShell have reported that its voice chat functionality is very smooth, with faster response times compared to GPT’s voice chat, and even features Live2D.

Case Two: AI Arena

AI Arena is an AI battle game where users can train their own battle spirits (NFT) using the LLM model and send them to PvP/PvE battlefields. The battle mode is similar to Nintendo’s Super Smash Bros, but with added competitiveness through AI training.

LianGuairadigm has led the investment in AI Arena, and it is currently in the public testing phase. Players can enter the game for free and also purchase NFTs to enhance their training strength.

Case Three: Leela vs the World – Chess Game

Leela vs the World is a chess game developed by Mudulas Labs. In the game, one side is AI and the other is a human player, and the game situation is stored in a smart contract. Players interact with the contract using a wallet, while the AI reads the updated game situation, makes decisions, and generates zero-knowledge proofs (zkp) for the entire calculation process, which is performed on AWS cloud. The zkp is then verified by the contract on the blockchain, and upon successful verification, the chess move is executed in the game contract.

3.3.2 AI-generated Game Content

Case One: AI Town

AI Town is a collaborative project between a16z and Convex Dev, inspired by Stanford University’s “Generative Agent” paper. AI Town is a virtual town where every AI can build its own story based on interactions and experiences.

It utilizes the Convex backend serverless framework, Pinecone vector storage, Clerk identity verification, OpenAI natural language text generation, and Fly deployment technologies. Additionally, AI Town is completely open-source, allowing game developers to customize various components such as feature data, sprite sheets, Tilemap visual environment, text generation prompts, game rules, and logic. AI Town can be experienced by both regular players and developers, who can use the source code to develop various features within and outside the game, making it suitable for different types of applications.

So, AI Town is not just an AI-generated content game, but also a development ecosystem and even a development tool.

Case Two: LianGuaiul

LianGuaiul is an AI story generator that provides a solution for directly generating and storing AI-generated stories on the blockchain for blockchain games. The implementation logic involves providing a set of prior rules to LLM, and then players can automatically generate secondary content based on these rules.

Currently, the game Straylight protocol has used LianGuaiul Seidler for game distribution. Straylight is a multiplayer NFT game, with the core gameplay being a blockchain version of “Minecraft”. Players can mint NFTs automatically and construct their own world based on the basic rules inputted by the model.

3.3.3 AI Generated Game Scenes

Case Study 1: LianGuaihdo Labs

LianGuaihdo Labs is a game development studio currently working on Halcyon Zero, an anime fantasy role-playing game and online game creation platform built on the Godot engine. The game takes place in an ethereal fantasy world with a bustling town as its social center.

What makes this game unique is that players can use the AI creation tools provided by the game developers to quickly create more 3D background effects and bring their favorite characters into the game, truly providing tools and game scenes for mass UGC (User-Generated Content) gaming.

Case Study 2: Kaedim

Kaedim has developed a 3D model generation tool based on Generative AI for game studios, which can quickly generate 3D scenes/assets that meet their needs in large quantities. The universal product of Kaedim is currently under development and is expected to be open for game studios to use in 2024.

The core logic of Kaedim’s product and AI agent is exactly the same, using a universal large model as the basis. The internal team of artists continuously inputs good data and provides feedback on the output of the agent, continually training this model through machine learning. Finally, the AI agent is able to output 3D scenes that meet the requirements.

04 Conclusion

In this article, we have conducted a detailed analysis and summary of the application of AI in the gaming industry. In general, future star unicorn projects will definitely appear in the form of universal models and Generative AI in the gaming industry. Although vertical applications have a lower moat, they have strong first-mover advantages. If they can create network effects and improve user stickiness through first-mover advantages, the potential is enormous. In addition to the directions mentioned in this article, there are other exploratory angles in the future. For example:

(1) Data Track + Application Layer: AI data tracks have already given birth to some unicorn projects with valuations of billions of dollars, and the link between data and the application layer is also full of imagination.

(2) Combining with Socialfi: For example, providing innovative ways of social interaction; optimizing community identity verification and community governance with AI agents; or more intelligent personalized recommendations.

(3) As Agent automation and maturity increase, will the main participants in the Autonomous World be humans or bots? Is it possible for the on-chain autonomous world to have a similar situation to Uniswap, where 80%+ of DAU (Daily Active Users) are bots? If so, exploring governance agents combining Web3 governance concepts is equally worth considering.

We will continue to update Blocking; if you have any questions or suggestions, please contact us!