Compiled by: Big Claws | PANews's Lobster
A Chinese developer has built an AI engine that can generate thousands of digital humans—each with a unique personality, memories, and behaviors—and put them into a virtual world to observe them and predict the future.
This is MiroFish. It reached number one on GitHub's global trending list. The creator is 20 years old.
Creator
His name is Guo Hangjiang, and his online name is "Baifu". He is a senior student at Beijing University of Posts and Telecommunications. He writes code in Python and is obsessed with intelligent agent architecture and graph computing.
At the end of 2025, his first project, BettaFish (a multi-agent sentiment analyzer), topped the GitHub trending list and received 20,000 stars within a week.
It was at that time that a Chinese billionaire noticed him.
That billionaire
Chen Tianqiao. Founder of Shanda Group. Once China's richest man. He built a gaming empire in the early 21st century, later retiring from the industry and moving to the United States, where he quietly transformed Shanda into a technology investment platform.
He has been pushing a theory he calls "super-individual"—that in the age of AI, one person can do what previously required an entire company to accomplish.
Guo Hangjiang is a living proof of this theory.
10-day development process
Chen Tianqiao invited Guo Hangjiang to be an intern, giving him complete freedom. What happened next was as follows:
Guo Hangjiang built MiroFish in 10 days, using what he calls the "Vibe coding" approach—fast, intuitive, and not over-designed.
That very night, he recorded a simple demonstration video and sent it directly to Chen Tianqiao's desk. In less than 24 hours, Chen Tianqiao pledged to invest 30 million yuan (approximately 4.1 million US dollars) to incubate the project.
Guo Hangjiang went from an intern to CEO overnight.
What exactly can it do?
Here is a concise summary of MiroFish's core features:
You input a document—a news article, a policy draft, a financial report, or even a novel. The system reads the content, extracts all entities and relationships into a knowledge graph using GraphRAG, and then generates thousands of autonomous AI agents. Each agent has a unique backstory, personality type, social relationships, and behavioral logic.
The most essential function: "God's-eye view".
You can inject new variables into the simulation at any time:
- The Federal Reserve suddenly cut interest rates by 50 basis points.
- "CEO resigns"
- "Competitors launch new products"
Then observe in real time how the entire digital world is reorganized. This is a controlled experiment that cannot be carried out in reality.
Simulation Engine
The simulation runs on OASIS—an open-source framework developed by CAMEL-AI. Intelligent agents don't just "speak"; they form groups, generate opinion leaders, exhibit herd behavior, and change their stance over time. Long-term memory is maintained via Zep Cloud.
Core feature: "God's-eye view". Inject new variables at any moment—interest rate hikes, CEO resignations, product launches—and the entire world is recalibrated in real time.
This is a controlled experiment that is impossible to achieve in reality.
Technical Architecture:
- Simulation engine: OASIS (developed by CAMEL-AI)
- Memory System: Zep Cloud (Long-Term Agent Memory)
- Knowledge Graph: GraphRAG
- Open source license: AGPL-3.0 (fully open source)
- Deployment method: Docker Compose one-click deployment
This is not a toy, but a serious multi-agent simulation framework.
Real demonstration
Two demo cases have been shown:
The first one they used was the first 80 chapters of *Dream of the Red Chamber*, a classic Chinese novel famous for its missing ending. MiroFish generated intelligent agents with realistic personalities and relationships, ran simulations, and generated multiple narrative branches that predicted the missing ending.
The second scenario is the Federal Reserve raising interest rates. The system simulates the reactions of retail investors, institutional participants, and analysts, tracks the convergence point of group sentiment, and maps out the complete trajectory of public opinion evolution.
Objective evaluation
It's worth clarifying what this project is not:
MiroFish has not yet released any benchmarks comparing predictions to real-world results. The demonstrations are merely presentations of methodology, not proof of accuracy. Running thousands of agents implies a massive API cost for large language models. The agents' personalities will inherit any biases present in the training data. Furthermore, no matter how complex or sophisticated, simulated humans are ultimately not real humans.
An objective positioning means that it can present scenes and dynamics that you might overlook, but it cannot provide definitive answers.
A broader significance
On March 7, 2026, MiroFish reached the number one spot on GitHub's global trending list—surpassing projects from OpenAI, Google, and Microsoft. Within days, it garnered 18,000 stars and 1,900 forks, and currently boasts over 22,000 stars.
An undergraduate student. Ten days of programming. A simulation engine capable of building parallel digital societies from a single document.
Chen Tianqiao wasn't betting on the software itself, but rather on this theory: the era of the super-individual has begun, but most people haven't realized it yet.
Open source license: AGPL-3.0, supports one-click deployment with Docker.
