# Rave Life: A 90s Drug Empire AI Simulator Running on Local GPU

> A 90s drug empire simulation game based on local LLM inference, where 96 AI drug dealers make autonomous decisions on the GPU without cloud APIs, running completely offline.

- 板块: [Openclaw Llm](https://www.zingnex.cn/en/forum/board/openclaw-llm)
- 发布时间: 2026-04-07T11:14:25.000Z
- 最近活动: 2026-04-07T11:21:43.862Z
- 热度: 141.9
- 关键词: LLM, AI模拟, 本地推理, 游戏开发, llama.cpp, 智能体, 开源游戏, 神经符号AI
- 页面链接: https://www.zingnex.cn/en/forum/thread/rave-life-gpu90ai
- Canonical: https://www.zingnex.cn/forum/thread/rave-life-gpu90ai
- Markdown 来源: floors_fallback

---

## [Introduction] Rave Life: Core Highlights of the 90s Drug Empire AI Simulator Running on Local GPU

Rave Life is an open-source simulation game that combines LLM with classic simulation management gameplay to create an AI-driven world running entirely on local GPU. Set against the backdrop of 90s rave culture, 96 AI drug dealers make autonomous decisions through real-time local LLM inference, with no pre-set scripts and complete personality profiles. The core innovation lies in the neuro-symbolic hybrid architecture, enabling offline operation and eliminating dependence on cloud APIs.

## [Background] Game Setting of 90s Drug Market and Rave Culture

The game recreates the 90s drug market landscape: 8 types of drugs (ecstasy, cocaine, etc.) have their prices and source locations set based on historical data; 9 cities are divided into demand centers (New York, Chicago, etc.), transit hubs (Miami, etc.), and source cities (Amsterdam, Medellín, etc.). Players need to arbitrage across cities, but transportation faces risks such as customs seizures, police interception, and robberies by competitors.

## [Technical Approach] Neuro-Symbolic Hybrid Architecture and AI Character Personality Design

The technical architecture uses a neuro-symbolic hybrid model: character personality, situation, memory, etc., are encoded into narrative context, and LLM outputs numerical actions (only 16 tokens each time) to balance decision depth and cost. llama.cpp is used as the inference backend. AI characters have 11-dimensional personality traits (bravery, greed, etc.), which are understood by LLM and influence their behavioral logic.

## [Technical Evidence] Local GPU Running Performance and AI Character Behavior Examples

Performance tests: The Qwen3.5-9B model supports parallel inference for 30-50 characters on RTX3060 (12GB), and 80-150 characters on RTX3090/4090 (24GB). Behavior examples: Paranoid drug dealers avoid dangerous transactions, risk-takers fight back against robberies, and personality-based emergent behaviors make characters more realistic.

## [Project Conclusion] A New Direction for AI-Driven Games: Let AI Be the Content Itself

Rave Life represents a new direction in AI game development: AI is not generating content, but becoming the content itself. Each NPC is an agent with real-time inference, and decisions are non-scripted. This 'living' world is a noteworthy path for the game industry to explore AI applications.

## [Future Suggestions] Community Contribution Directions Needed for the Open-Source Project

Rave Life is open-source (MIT license) and needs contributors to participate in: improving the C++ client, UI/frontend design, integrating more drugs and cities, developing player vs. player/multiplayer online modes, optimizing economic balance, etc. The core technology and simulation engine remain open-source; community participation is welcome.