MiroFish: The God View Engine

A recent post on X highlights a striking project from a young Chinese developer, Guo Hangjiang (online: BaiFu). He built MiroFish, an AI engine that spins up thousands of digital humans, gives each a biography, memory, and behavior, drops them into a simulated world, and watches how collective opinions and outcomes evolve. Read the original post here: https://x.com/k1rallik/status/2032837113247395992.

At its core, MiroFish reads a document, extracts entities and relationships into a knowledge graph using GraphRAG, then generates autonomous agents that interact on an OASIS simulation engine, with long-term memory in Zep Cloud. The creator says he built it in ten days using what he calls “vibe coding,” and the demo caught the eye of investor Chen Tianqiao, who funded the project within a day.

Two demos stand out. One used the first 80 chapters of Dream of the Red Chamber to generate character-driven narrative branches, offering plausible endings. The other simulated a Fed rate hike, showing how retail investors, analysts, and institutions might shift sentiment over time. The system gives a *“God’s Eye View,”* meaning users can inject events like “CEO resigns” and watch the simulated society rewire in real time. It’s open source under AGPL-3.0 and ships with a one-click Docker deployment on GitHub: https://github.com/666ghj/MiroFish.

A few honest notes, because this matters for anyone tempted to trust the outputs: MiroFish has no public benchmarks linking its predictions to real outcomes, running so many agents is costly, and agent personalities reflect training data biases. These are simulated humans, not a crystal ball.

Still, the project is a clear example of the “super-individual” thesis: a single talented developer can build tools that once needed whole teams. If this sparks more open, experimental simulation work, it could change how organizations explore scenarios and risks, and that possibility feels worth watching.