@k1rallik
Translator: Da Qianzhi | PANews Lobster
A Chinese developer built an AI engine capable of generating thousands of digital humans—each with unique personalities, memories, and behaviors—placing them into a virtual world to observe their predictions of the future.
This is MiroFish. It ranked first on GitHub’s global trending list. The creator is 20 years old this year.
Creator
His name is Guo Hangjiang, online alias “Baifu.” A senior at Beijing University of Posts and Telecommunications. Writes code in Python, obsessed with agent architectures and graph computing.
By the end of 2025, his first project—BettaFish (a multi-agent public opinion analyzer)—hit the top of GitHub’s trending list, earning 20,000 stars in a week.
At that time, a Chinese billionaire noticed him.
The Billionaire
Chen Tianqiao. Founder of Shanda Group. Former China’s richest man. In the early 2000s, built a gaming empire, later retired, moved to the U.S., and quietly transformed Shanda into a tech investment platform.
He has been promoting a theory he calls “Super Individuals”—in the AI era, one person can do what previously required an entire company.
Guo Hangjiang is a living proof of this theory.
10 Days of Development
Chen Tianqiao invited Guo Hangjiang to intern, giving him complete freedom. What happened next:
Guo Hangjiang spent 10 days building MiroFish, using what he calls “Vibe coding”—fast, intuitive, without over-designing.
On the night of completion, he recorded a simple demo video and sent it directly to Chen Tianqiao. Less than 24 hours later, Chen committed 30 million RMB (about $4.1 million) to incubate the project.
Guo Hangjiang overnight went from intern to CEO.
What Can It Do?
Here are the core features of MiroFish, summarized:
You input a document—news articles, policy drafts, financial reports, even a novel. The system reads the content, extracts all entities and relationships into a knowledge graph via GraphRAG, then generates thousands of autonomous AI agents. Each agent has a unique background story, personality type, social relationships, and behavioral logic.
The most core feature: “God’s eye view.”
You can inject new variables into the simulation anytime:
- “The Federal Reserve suddenly cuts interest rates by 50 basis points”
- “CEO resigns”
- “Competitor launches new product”
Then observe in real-time how the entire digital world reorganizes. This is a controlled experiment impossible in reality.
Simulation Engine
The simulation runs on OASIS—an open-source framework developed by CAMEL-AI. Agents don’t just “talk”; they form groups, produce opinion leaders, create herd effects, and change stances over time. Long-term memory is maintained via Zep Cloud.
Core feature: “God’s eye view.” Inject new variables—interest rate hikes, CEO resignations, product launches—at any moment, and the entire world recalibrates in real time.
This is an experiment that cannot be achieved in reality.
Technical architecture:
- Simulation engine: OASIS (by CAMEL-AI)
- Memory system: Zep Cloud (long-term agent memory)
- Knowledge graph: GraphRAG
- Open-source license: AGPL-3.0
- Deployment: One-click via Docker Compose
This is not a toy but a serious multi-agent simulation framework.
Live Demonstrations
Two demo cases have been shown:
First—input the first 80 chapters of “Dream of the Red Chamber,” a classic Chinese novel known for its lost ending. MiroFish generated character agents with realistic personalities and relationships, ran simulations, and produced multiple narrative branches predicting the missing ending.
Second—the Federal Reserve interest rate hike scenario. The system simulated reactions from retail investors, institutional participants, and analysts, tracked the convergence of group sentiments, and mapped the full evolution of public opinion.
Objective Evaluation
It’s important to clarify what this project is not:
MiroFish has not yet released any benchmarks comparing prediction results with real-world outcomes. The demos are just methodological showcases, not proof of accuracy. Running thousands of agents incurs huge API costs for large language models. The agents’ personalities may inherit biases present in training data. And no matter how complex or precise, simulated humans are ultimately not real humans.
Objectively, it can present scenarios and dynamics you might overlook, but it cannot provide definitive answers.
Broader Significance
On March 7, 2026, MiroFish topped GitHub’s global trending list—surpassing projects from OpenAI, Google, and Microsoft. Within days, it gained 18,000 stars and 1,900 forks, now exceeding 22,000 stars.
An undergraduate. Ten days of coding. A simulation engine capable of constructing parallel digital societies from a single document.
Chen Tianqiao’s bet isn’t just on this software but on a theory: the era of “Super Individuals” has begun, though most people haven’t realized it yet.
Open-source license: AGPL-3.0, supports one-click deployment via Docker.
github.com/666ghj/MiroFish
