Ghost AI: Let’s AI Agents Build Disposable Worlds
TL;DR · AI Summary
Ghost AI proposes disposable database copies for AI Agents to safely experiment with data-layer changes; the author validates LLMs’ physics-control learning via the Gravell GPT benchmark with 30 iterative feedback rounds.
Key Takeaways
- Direct database access for AI Agents is highly risky; each agent must be given a
- In Gravell GPT, LLMs write spaceship control scripts and compete over 200 ticks
- The benchmark uses 30 iterative rounds: generate code → run simulation → receive
Outline
Jump quickly between sections.
Granting AI Agents direct database write access is extremely dangerous because the database holds the application’s core state, and errors can cause irreversible damage.
Code can be safely rolled back via Git, but databases lack equivalent disposable sandboxing, hindering AI-driven experimental development.
Each AI Agent gets its own temporary, isolated database copy, enabling safe experimentation—failed attempts are discarded without side effects.
This benchmark tasks LLMs with writing control scripts for spaceships in a Newtonian physics environment with four stars and three ships.
Each round generates code → runs 200-tick simulation → returns structured failure feedback (e.g., collision location, fuel depletion) → improves next iteration.
Mindmap
See how the topics connect at a glance.
查看大纲文本（无障碍 / 无 JS 友好）
- Ghost AI：AI Agent 安全实验框架
- 核心问题
- 数据库无安全沙盒机制
- Agent 直接写 DB 风险极高
- 代码可回滚 vs 数据不可逆
- 解决方案
- Disposable DB Copies
- 按 Agent 隔离副本
- 失败即丢弃，零副作用
- 验证案例：Gravell GPT
- 4 恒星 + 3 飞船物理模拟
- LLM 生成控制脚本而非实时操控
- 30 轮反馈迭代 → 200 tick 评分
Highlights
Key sentences worth saving and sharing.
The database is ‘the state of the world’—users, orders, product catalogs, game economies, analytics, settings, and history all reside here; accidental modification is far more damaging than code error
In Gravell GPT, the LLM does not control ships directly but writes control scripts; the goal is to keep ships inside a moving disc for as long as possible over 200 ticks without crashing or running ou
The key innovation is 30 iterative rounds: after each code generation, a full simulation runs and returns structured feedback, enabling LLMs to improve like human developers through trial and error.