Definition
AIPC: Artificial-intelligence Player Control, an innovative concept in contrast to NPC (Non-Player Control). With the development of AI technology and the rapid progress of machine intelligence, AI breathes life and thought into the traditionally simplistic characters in electronic games, which are only used for simple functional scenarios, thus creating AIPC.
Preamble & Overview
• Whether it's blockchain, web3, Metaverse, AIGC, or AGI, the rapid shift of hotspots is merely a manifestation of market sentiment. In terms of the digitalization process of human civilization, these concepts have a long history, advancing side by side and supporting each other. Their ultimate goal is highly unified – a completely digitalized future world.
• We do not intend to discuss philosophical topics such as whether the human race and civilization can continue in the long term and in what form, or whether the definitions of individuals, consciousness, and life will undergo a radical transformation in that future. Instead, we strive to cut through the haze created by the aforementioned numerous concepts and reach the core of digitalization.
• We believe that every electronic game in history is a simple simulation of a "completely digitalized future world" based on limited conditions. The current technologies collectively point to an opportunity to expand the concept of electronic games to the internet. Rather than saying that various technologies in the past decade (5G, XR, web3, crypto, Metaverse, AIGC, GPT, etc.) may give birth to a more immersive digital world, it is more accurate to say that electronic games will seize this opportunity to break away from the entertainment domain and become the cornerstone of a serious, ubiquitous, and all-encompassing new world.
• Based on our extensive experience in the mobile internet, electronic games, and blockchain industries, as well as our research, attempts, and in-depth thinking in AI technology over the past few months, we hope to initiate a social experiment led by AI and controlled by decentralized human organizations.
• In summary, we are about to establish a micro-society entirely controlled by digital intelligence. Numerous virtual NPCs will be controlled by independent AI models, each with their own behavioral goals and life rhythms. All events and event-based chain reactions are entirely derived from the calculations (or "consciousness") of digital intelligence. In this self-operating micro-society system, everything evolves "naturally." We are responsible for faithfully recording and observing, and introducing new variables through group decision-making during the process.
• Crypto technology will play a crucial role in this game (we still call it a game). In other words, this is perhaps the most accurate use of blockchain to date: NPCs' behaviors and participation in social activities within the game will use digital currency as the circulating currency. Accordingly, each NPC will have its own wallet, autonomously operated by AI decision-making.
• The main nodes and system settings of this game's timeline will be stored to some extent on Ethereum and its Layer 2 networks. All evolution records will be stored in distributed storage networks such as IPFS. During the game's progress, AI-generated items will be offered through logic channels for NPCs to decide whether to mint as NFTs on the Ethereum chain and sell in-game. Human users may be able to purchase digitally native items with a complete digital history.
• Apart from the essential visual elements for carrying social information, this game will not actively engage in any art or artistic design. Due to the 24/7 live streaming of the game in a global broadcast, 24 language versions will be provided; the English version will be live-streamed on major video platforms.
• The project is entirely open-source and will not accept any VC or institutional investments before going live.
• After the official launch of the first edition, we will accept donationsof any amount up to 1 BTC or its equivalent in USDT to cover development costs, various cloud services, and API access fees. The remaining funds will be used through DAO proposals. The highest donation tier (i.e., 1 BTC) will grant the donor the naming rights for an in-game character. Other than that, donations are completely voluntary, and there are currently no airdrop or whitelist plans.
• Within the game, there will be a self-operating token economy system. However, before a complete economic policy system is in place, it will remain closed within the system and will not be open to humans through IDO. Moreover, this IDO action will be a crucial step in opening communication between the two civilizations, and the necessary discussions and preparations are expected to be lengthy in our estimation.
• This experiment is limited to the scope of electronic game development and considers the application of various innovative technologies. However, this project does not intend to discuss/challenge any moral, philosophical, global economic and financial, political, cultural, religious content, etc. We maintain full respect for any individual and collective.
Background
Due to the impressive performance of GPT-4, an increasing number of game companies have begun to use GPT to empower in-game NPCs. The benefits are obvious – it allows formerly rigid and mechanized NPCs to have independent development paths and real-time intelligent responses. With the advancement of technology, the presence of intelligent NPCs will significantly influence the course and direction of games, leading to an exponential expansion of the possibility space within games.
Currently still in its infancy, the ChatGPT product operates in a service-oriented state of "infinite continuation of dialogue." If we were to set up a scenario in which multiple NPCs are driven by ChatGPT and interact with each other, with the goal of making their behavior as close as possible to real social and living situations, thus exhibiting the chaotic system characteristics of society – unpredictable and uncontrollable – we need to address two issues:
First is diversity. Each NPC should not be constantly engaged in social interactions and conversations; they should have other activities in their lives and spend time on thoughts, emotions, and sleep.
Second is the mechanism for not responding and initiating conversations. In a normal society, individuals have the ability to follow their consciousness and emotions to stop a conversation or exercise their right not to respond. They should also have the ability to initiate a conversation and generate the motivation for doing so.
For example, suppose 20 independent NPCs live in a small town. They should each have short-term goals, engage in eating, sleeping, and thinking, and have families, jobs, and transactions, rather than mechanically wandering the streets and constantly conversing with each other. In a self-operating micro-society system, such a setup creates a more realistic simulation effect and produces more varied social scenarios, significantly enhancing the possibilities within this society.
Framework
Based on this line of thought, we will establish a system consisting of three parts:
1. The underlying rule system: Using AI technology, we will design the fundamental logic of the micro-society in the form of various rules. These rules represent the iron laws that no activity or event within this society can surpass or break.
2. Character behavior trees: The mechanism and scheduling system for characters to actively or passively engage in various actions. At the mesoscopic scale, without observing specific conversation content, this system is key to determining the behavior of character groups centered on NPCs.
3. Settings and social interactions: Randomly assign background settings and personality traits to NPCs based on specific templates. The interactions will progress freely based on ChatGPT in social scenarios and be constantly fed back into the second layer system, serving as the primary factor of the chaotic system.
Implementation Logic
** - Underlying rule system:**
The underlying rule system will be designed using an approach similar to Reinforcement Learning (RL). This will enable NPCs to follow certain rules and gradually optimize their behavior during interactions with their environment. Specific implementation steps include:
▪ Design an environment model to simulate the micro-society where NPCs reside, including geographical locations, resource distribution, and time systems.
▪ Define a set of basic rules, such as physical laws and social norms, to serve as constraints for NPC behavior.
▪ Assign a reinforcement learning agent to each NPC and train them using algorithms like PPO, DQN, or A3C.
Considering technological development trends and current limitations, the underlying rule system will prioritize a closed-loop approach and iterative development, starting with a minimum viable model and gradually introducing multidimensional, multimodal parameters in major version updates.
** - Character behavior trees:**
Character behavior trees are hierarchical structures for scheduling NPC behavior, capable of executing different actions based on environmental conditions and the NPC's internal state. Specific implementation steps include:
▪ Define a set of behaviors for each NPC, including basic needs (e.g., eating and sleeping), emotional states (e.g., happiness and sadness), and goal-driven actions (e.g., work and leisure).
▪ Design a behavior tree model containing selection nodes, sequence nodes, and action nodes, which organize and schedule these behaviors.
▪ Update the behavior tree's state based on the decisions of the reinforcement learning agents, and execute corresponding actions according to the behavior tree's output.
** - Settings and social interactions:**
The settings and social interactions component mainly involves assigning personalities to NPCs and using GPT-4 to drive social interactions. Specific implementation steps include:
▪ Generate a character setting for each NPC, including background, hobbies, and personality traits, which can be done through predefined templates or randomly generated by GPT-4.
▪ Incorporate social interaction-related behaviors into the behavior tree, such as finding conversation partners and ending conversations.
▪ Use GPT-4 as a dialogue generator to produce relevant dialogue content based on the NPC's character setting and current context. GPT-4 models can be fine-tuned to adapt to specific game scenarios and character settings.
▪ Provide feedback from social interactions to reinforcement learning agents and behavior trees to adjust NPC behavior and decision-making.
Note: This article is just a preliminary idea and should not be used as a project plan or execution basis.