Original author: PSE Trading Analyst@Minta
Key Insights
AI Agent is a tool based on the LLM general large model that allows developers and users to directly build applications that can interact independently.
The main pattern of the AI track in the future may be: general large models + vertical applications; the ecological niche of AI Agent is the middleware that connects general large models and Dapps, so the AI Agent has a low moat and needs to rely on the creation of network effects and improvements User stickiness enhances long-term competitiveness.
This article sorts out the development of general large models, vertical application agents, and generative AI applications in the Web3 game track. Among them, combined with Generative AI technology, it has great potential to produce hit games in the short term.
01 Technical Introduction
Among the AGI (Artificial General Intelligence) technology that has exploded this year, Large Language Model (LLM) is the absolute protagonist. OpenAI core technical staffAndrej KarpathyandLilian WengIt has also been expressed that LLM-based AI Agents are the next important development direction in the AGI field, and many teams are also developing LLM-driven artificial intelligence agent (AI-Agents) systems. Simply put, AI Agent is a computer program that uses large amounts 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.aiThe picture clearly introduces the basic principles of AI Agent. The steps are as follows:
Perception and data collection: Data input, or the AI Agent obtains information and data, such as game status, images, sounds, etc., through the perception system (sensors, cameras, microphones, etc.).
State representation: Data needs to be processed and represented into a form that the Agent can understand, such as converting it to a vector or tensor, so that it can be input into the neural network.
Neural network model: Deep neural network models are usually used for decision-making and learning, such as convolutional neural network (CNN) for image processing, recurrent neural network (RNN) for sequence data processing, or more advanced models such as self-attention Force mechanism (Transformer), etc.
Reinforcement learning: Agent learns optimal action strategies through interaction with the environment. In addition, the operating principles of Agent also include strategy network, value network, training and optimization, and exploration and utilization. For example, in a game scenario, the strategy network can input the game state and then output the action probability distribution; the value network can estimate the value of the state; the agent can continuously strengthen the learning algorithm by interacting with the environment to optimize the strategy and value network and output more perfect results.
Source：blog.abacus.ai
In summary, AI-Agents are intelligent entities capable of understanding, decision-making, and action, and they can play an important role in a variety of fields, including gaming. OpenAI core technical staffLilian WengWritten byLLM Powered Autonomous Agents》A very comprehensive introduction to the principles of AI-Agents. Among them, the article mentions a very interesting experiment: Generative Agents.
Generative Agents(GA for short) is inspired by the Sims game, which uses LLM technology to generate 25 virtual characters. Each character is controlled by an Agent supported by LLM, living and interacting in a sandbox environment. The design of GA is very smart. It combines LLM with memory, planning and reflection functions, which allows the Agent program to make decisions based on previous experience and interact with other Agents.
The article details how the Agent continuously trains and optimizes its decision-making path based on the policy network, value network, and interaction with the environment.
The principle is as follows: Memory Stram is a long-term memory module that records all interactive experiences of the Agent. The retrieval model (Retrieve) provides experience (Retrived Memories) based on relevance, freshness and importance to help the Agent make decisions (Plan). The reflection mechanism (Reflect) summarizes past events and guides the Agents future actions. Plan and Reflect work together to help the Agent transform reflection and environmental information into actual actions.
Source：LLM Powered Autonomous Agents
This interesting experiment shows us the capabilities of AI Agent, such as generating new social behaviors, information dissemination, relationship memory (such as two virtual characters continuing to discuss topics) and coordination of social activities (such as hosting a party and inviting other virtual characters )etc. All in all, AI-Agent is a very interesting tool, and its application in games is worth exploring in depth.
02 Technology Trends
2.1 AI track trends
Investment Research Partner of ABCDELaoBaiI once summarized the Silicon Valley venture capital circle’s judgment on the next development of AI:
There are no vertical models, only large models + vertical applications;
Data from edge devices such as mobile phones may be a barrier, and AI based on edge devices may also be an opportunity;
The length of Context may cause qualitative changes in the future (vector databases are currently used as AI memory, but the context length is still not enough).
That is to say, from the perspective of the general development rules of the industry, because large-scale general models are too heavy and have strong universality, there is no need to constantly reinvent the wheel in the field of large-scale general models. Instead, we should focus more on applying large-scale general models to Vertical field.
At the same time, edge devices refer to terminal devices that usually do not rely on cloud computing centers or remote servers, but perform data processing and decision-making locally. Because of the diversity of edge devices, how to deploy AI Agents to run on devices and obtain device data reasonably is a challenge, but it is also a new opportunity.
Finally, the issue about Context has also attracted much attention. Simply put, Context in the context of LLM can be understood as the amount of information, and Context length can be understood as the number of dimensions of the data. Suppose there is a big data model for an e-commerce website, which is used to predict the likelihood of a user purchasing a certain product. In this case, Context can include information such as the users browsing history, purchase history, search history, user attributes, etc. Context length refers to the dimension in which feature information is superimposed, such as the purchase history of competing products by a 30-year-old male user in Shanghai, the frequency of recent purchases, and the recent browsing history. The increase in Context length can help the model more comprehensively understand the influencing factors of user purchasing decisions.
The current consensus is that although the current use of vector databases as AI memory makes the Context length insufficient, the Context length will change qualitatively in the future, and then the LLM model can seek more advanced methods to process and understand longer and more complex Contexts. information. More application scenarios beyond imagination are emerging.
2.2 AI Agent Trend
Folius VenturesThe application model of AI Agent in the game track has been summarized, as shown below:
Source: Folius Ventures - Game Feature: Journey to Find the North Star of Web3 Games
1 in the figure is the LLM model, which is mainly responsible for converting user intentions from traditional keyboard/click input into natural language input, reducing the users entry threshold.
2 in the figure is a front-end Dapp integrated with AI Agent. While providing functional services to users, it can also collect user habits and data from the terminal.
3 in the figure are various types of AI Agents, which can exist directly in the form of in-application functions, Bots, etc.
In general, AI Agent, as a code-based tool, can serve as the underlying program for Dapp to expand application functions and the growth catalyst of the platform, that is, the middleware that links large models and vertical applications.
From the perspective of user scenarios, the Dapps that are most likely to integrate AI Agent are most likely social apps, chatbots and games that are sufficiently open; or the existing Web2 traffic entrance can be transformed into a simpler and more user-friendly AI+web3 entrance through AI Agent; that is, the industry We have been discussing lowering the user threshold of Web3.
Based on the law of industry development, the middleware layer where AI Agent is located will often become a highly competitive track with almost no moat. Therefore, in addition to constantly improving the experience to match B2C needs, AI Agent can also improve its moat by creating network effects or creating user stickiness.
03 Track Map
There have been many different attempts to apply AI in the field of Web3 games, and these attempts can be divided into the following categories:
General models: Some projects focus on building general AI models and finding suitable neural network architectures and general models for the needs of Web3 projects.
Vertical applications: Vertical applications are designed to solve specific problems in the game or provide specific services, and usually appear in the form of Agents, Bots and BotKits.
Generative AI applications: The most direct application corresponding to large models is content generation, and the game track itself is a content industry, so Generative AI applications in the game field are very worthy of attention. It has become possible to automatically generate elements, characters, tasks or storylines in the virtual world, to automatically generate game strategies, decisions and even the automatic evolution of the in-game ecology, making the game more diverse and in-depth.
AI games: Currently, there are many games that have integrated AI technology, with different application scenarios. Examples will be given below.
3.1 General large model
Currently, Web3 already has simulation models for economic model design and economic ecological development, such as the QTM quantitative token model. Outlier VenturesDr. Achim StruveIn the ETHCC speech, some ideas on economic model design were mentioned. For example, considering the robustness of the economic system, the project team can create a digital twin through the LLM model to simulate the entire ecosystem 1:1.
Pictured belowQTM(Quantitative Token Model) is an AI-driven reasoning model. QTM uses a fixed simulation time of 10 years, with each time step being one month. At the beginning of each time step, tokens will be discharged into the ecosystem, so there are incentive modules, token attribution modules, airdrop modules, etc. in the model. Subsequently, these tokens will be put into several meta buckets, from which more detailed general utility redistribution will be carried out again. Then, define reward payments, etc. from these utility instruments. There are also aspects like off-chain business, which also take into account the general funding status of the business, such as it can be destroyed or repurchased, and it can also measure user adoption rate or define user adoption.
Of course, the output quality of the model depends on the input quality, so before using QTM, sufficient market research must be conducted to obtain more accurate input information. However, the QTM model is already a very practical application of the AI-driven model in the Web3 economic model. There are also many project parties using the QTM model to make 2C/2B applications with lower operational difficulty, lowering the threshold for project parties to use it.
3.2 Vertical application agent
Vertical applications mainly exist in the form of Agents. Agents may be in different forms such as Bots, BotKits, virtual assistants, intelligent decision support systems, various automated data processing tools, etc. Generally speaking, AI Agent takes OpenAIs general model as the bottom layer, combines it with other open source or self-developed technologies, such as text-to-speech (TTS), and adds specific data for FineTune (a type of training in the field of machine learning and deep learning). technology, the main purpose is to further optimize a model that has been pre-trained on large-scale data) to create an AI Agent that performs better than ChatGPT in a specific field.
Currently, the most mature application in the Web3 game track is NFT Agent. The consensus on the gaming circuit is that NFT must be an important part 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 NFT functions more flexible through algorithms. For users, there can be more interactions between users and NFT, and the generated interactive data becomes a source of information. AI Agent can optimize the interaction process and expand the application scenarios of interactive data, injecting more innovation and value into the NFT ecosystem.
Case 1: For example, Gelatos development framework allows developers to customize logic to update NFT metadata based on off-chain events or specific time intervals. Gelato nodes will trigger metadata changes when specific conditions are met, thereby enabling automatic updates of NFTs on the chain. For example, this technology could be used to obtain real-time game data from a sports API and automatically upgrade the skill characteristics of the NFT under certain conditions, such as when an athlete wins a game.
Source：Gelato - The Ultimate Guide to Dynamic NFTs
Case 2:PaimaApplication Agent is also provided for Dynamic NFT. Paimas NFT compression protocol mints a minimal set of NFTs on L1 and then evolves them based on the game state on L2, providing players with a more in-depth and interactive gaming experience. For example, NFT can change based on the characters experience value, task completion, equipment and other factors.
Case three:Mudulas LabsIt is a very well-known ZKML project, and it also has a layout in the NFT track. Mudulas launched the NFT series zkMon, which allows NFTs to be generated through AI and released to the chain. At the same time, a zkp is generated. Users can use zkp to check whether their NFTs are generated from the corresponding AI model. For more comprehensive information, please refer to:Chapter 7.2: The World’s 1 st zkGAN NFTs。
3.3 Generative AI applications
As mentioned earlier, because the game itself is a content industry, AI-Agent can generate a large amount of content in a short time and at low cost, including creating uncertain,Dynamic game charactersetc. Therefore, Generative AI is very suitable for game applications. Currently, the applications of Generative AI in the game field can be summarized into the following main types:
AI-generated game characters: For example, fight against AI, or AI is responsible for simulating and controlling NPCs in the game, or even directly using AI to generate characters, etc.
AI-generated game content: Various content, such as tasks, storylines, props, maps, etc., are directly generated by AI.
AI generated game scene class: Supports the use of AI to automatically generate, optimize or expand the terrain, landscape and atmosphere of the game world.
3.3.1 AI generated characters
Case 1: MyShell
MyShellIt is a Bot creation platform. Users can create exclusive Bots according to their own needs for chatting, practicing speaking, playing games, and even seeking psychological counseling, etc. At the same time, Myshell uses text-to-speech (TTS) technology, which can automatically create a Bot by imitating anyones voice with just a few seconds of voice samples. In addition, MyShell uses AutoPrompt, which allows users to issue instructions to the LLM model only by describing their own thoughts, laying the foundation for a private large language model (LLM).
Users with Myshellexpress, its voice chat function is very smooth, the response speed is faster than GPTs voice chat, and it also has Live 2D.
Case 2: AI Arena
AI ArenaIt is an AI battle game. Users can use the LLM model to continuously train their own battle elves (NFT), and then send the trained battle elves to PvP/PvE battlefields. The battle mode is similar to Super Smash Bros., but adds more competitive fun through AI training.
Paradigm led the investment in AI Arena, and the public beta phase has begun. Players can enter the game for free or purchase NFT to increase training intensity.
Case 3: On-chain chess game Leela vs the World
Leela vs the Worldis a chess game developed by Mudulas Labs. In the game, the two parties are AI and humans, and the chess game situation is placed in the contract. Players operate (interact with contracts) through their wallets. The AI reads the new chess game situation, makes a judgment, and generates zkp for the entire calculation process. Both steps are completed on the AWS cloud, and the zkp is verified by the contract on the chain. After the verification is successful, the chess game contract is called Down chess.
3.3.2 AI generates game content
Case 1: AI Town
AI Town is a collaboration between a16z and its portfolio company Convex Dev, inspired by Stanford Universitys Generative Agent paper. AI Town is a virtual town where each AI in the town can build its own story based on interaction and experience.
Among them, technology stacks such as Convex back-end serverless framework, Pinecone vector storage, Clerk authentication, OpenAI natural language text generation, and Fly deployment are used. In addition, AI Town is all open source and supports in-game developers to customize various components, including feature data, sprite sheets, Tilemap visual environment, text generation prompts, game rules and logic, etc. In addition to ordinary players being able to experience AI Town, developers can also use the source code to develop various functions within the game or even outside the game. This flexibility makes AI Town suitable for a variety of different types of applications.
Therefore, AI Town itself is an AI-generated content game, but it is also a development ecosystem and even a development tool.
Case 2: Paul
Paul is an AI story generator that specifically provides a solution path for full-chain games to generate AI stories and directly upload them to the chain. The implementation logic is to input a lot of a priori rules into LLM, and then players can automatically generate secondary content based on the rules.
There is currently a game released using the Straylight protocol using Paul Seidler.StraylightIt is a multiplayer NFT game. The core gameplay is the full-chain game version of Minecraft. Players can automatically Mint NFT and then construct their own world according to the basic rules input by the model.
3.3.3 AI generated game scenes
Case 1: Pahdo Labs
Pahdo Labs is a game development studio currently developing 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, centered around a bustling town that serves as a social hub.
What is very special about this game is that players can use the AI creation tools provided by the game to quickly create more 3D effect backgrounds and bring their favorite characters into the game, truly providing tools and game scenes for mass game UGC.
Case 2: Kaedim
KaedimA 3D model generation tool based on Generative AI has been developed for game studios, which can quickly help game studios batch generate in-game 3D scenes/assets that meet their needs. Kaedims general product is currently still under development and is expected to be open to game studios in 2024.
The core logic of Kaedims product is exactly the same as that of AI-Agent. It uses a general large model as the basis, and then the artists within the team will continuously input good data, and then give feedback to the Agents output, and continuously train the Model through machine learning. Finally, the AI-Agent can output a 3D scene that meets the requirements.
04 Summary
In this article, we conduct a detailed analysis and summary of the application of AI in the gaming field. In general, the application of general models and Generative AI in games will definitely lead to star unicorn projects in the future. Although vertical applications have a low moat, they have a strong first-mover advantage. If they can rely on their first-mover advantage to create network effects and increase user stickiness, there will be huge room for imagination. In addition, generative AI is naturally suitable for the content industry of games. There are currently many teams trying to apply GA in games. It is very likely that there will be hit games using GA in this cycle.
In addition to some of the directions mentioned in the article, there are other exploration angles in the future. for example:
(1) Data track + application layer: The AI data track has given birth to some unicorn projects valued at billions of dollars, and the linkage of data + application layer is also full of imagination.
(2) Integration with Socialfi: such as providing innovative social interaction methods; using AI Agent to optimize community identity authentication and community governance; or more intelligent personalized recommendations, etc.
(3) With the automation and maturity of Agents, will the main participants in the Autonomous World in the future be humans or Bots? Is it possible for the autonomous world on the chain to be like Uniswap, where 80%+ of DAU are Bots? If so, then a governance agent that combines Web3 governance concepts is also worth exploring.
05 References
Exploring the Design Space for Dynamic NFTs
Generative Manufacturing: Transmuting Code intoPhysical Goods
From Verifiable AI to Composable AI: Reflections on zkML application scenarios
Data-driven token design and optimization
