The 20-Year-Old Who Coded the Future in 10 Days

MiroFish spawns thousands of AI humans to simulate reality and your next big decision might depend on what they say.

Last week, something quietly climbed to the top of GitHub. No press release. No launch event. Just a 10-day-old repo by a 20-year-old student, and a README that said: we can predict the future.

The project was called MiroFish. The pitch was simple: spawn thousands of AI agents with unique personalities, drop them into a simulated world, let them interact and watch what emerges.

Chen Tianqiao, once the richest man in China and founder of Shanda Group, saw the demo. Within 24 hours, he wired $4.1 million. Guo Hangjiang — the undergrad behind it went from intern to CEO overnight.

What Is MiroFish?

Most prediction tools work like calculators. Feed in historical data, get a probability out. The problem? The real world doesn’t behave like a formula. People react to each other. Opinions shift. A single tweet can change the trajectory of a news cycle. No regression model captures that.

MiroFish doesn’t crunch numbers. It builds a miniature society and watches what happens. Here’s the core loop:

· Upload any “seed material”: It can be a news article, policy doc, financial report, or even a novel.

· MiroFish spawns thousands of AI agents, each with a unique personality, backstory, and stance.

· Those agents interact — arguing, persuading, updating their beliefs, just like real people.

· A prediction emerges from the collective behaviour. No formula. Pure emergence.

Think SimCity — but the city is a society, and the citizens are debating your question in real time.

[embed]

How It Works: 5 Steps

Feed MiroFish a document and a question. Here’s exactly what happens next.

1. Seed ingestion : GraphRAG parses your input not as flat text, but as a structured knowledge graph, mapping entities, relationships, and tensions. This becomes the simulation’s reality.

2. Persona generation : Thousands of agent profiles are auto-created from the graph, each with a distinct personality, background, and social connections.

3. World configuration : An Environment Agent sets the rules — information channels, pacing, constraints. The physics of the simulated society.

4. Parallel simulation : Agents are released across two simultaneous environments, powered by OASIS, a framework scaling to 1 million agents with 23 social actions: follow, argue, repost, mute, and more.

5. Interactive exploration : Post-simulation, you can query individual agents, inject new events mid-run, and test counterfactuals, all from a god’s-eye view.

To try it: You need Node.js 18+, Python 3.11+, an LLM API key, and a Zep Cloud key. Clone github.com/666ghj/MiroFish, add your keys to .env, and run. Start with 50–100 agents, cap at 15 rounds.

[embed]GitHub - 666ghj/MiroFish: A Simple and Universal Swarm Intelligence Engine, Predicting Anything… A Simple and Universal Swarm Intelligence Engine, Predicting Anything. 简洁通用的群体智能引擎，预测万物 - 666ghj/MiroFishgithub.com

What Has It Predicted?

Two published demos reveal how wide the use cases can stretch.

Case 1: Public Opinion at Wuhan University

MiroFish ingested a real campus event report and simulated how sentiment would evolve. It mapped 90-day polarization curves, showing exactly which agent clusters would amplify vs. suppress narratives. Not one outcome. A map of possibilities.

Case 2: A Lost Novel Ending

The team fed MiroFish the first 80 chapters of Dream of the Red Chamber — a classic Chinese novel whose ending was lost centuries ago. The agents, built from the novel’s characters, predicted what happened next. The result matched existing scholarly theories with uncanny coherence.

Early community experiments have pushed further:

· A developer built a Polymarket trading bot using 2,847 simulated humans per trade → $4,266 profit over 338 trades.

· Tech analyst Brian Roemmele ran a single simulation with 500,000 agents.

· Multiple companies have publicly called it their “digital crystal ball” for scenario planning.

The Stack Behind It

MiroFish didn’t invent new technology. It assembled existing pieces in a new way and that’s the whole story.

· OASIS — The simulation engine by CAMEL-AI. Open-sourced Dec 2024. Supports 1M agents, 23 social actions, and replicates real social phenomena: group polarization, herd behaviour, information propagation.

· GraphRAG — Knowledge grounding. Builds a structured entity-relationship graph from your input — not just raw text retrieval.

· Zep Cloud — Persistent agent memory. Without this, agents forget between rounds. Zep makes the simulation feel continuous.

· Any OpenAI-compatible LLM — The reasoning layer. Every agent interaction is an LLM inference call.

That’s not a story about genius. It’s a story about how fast AI infrastructure has matured. The components existed. MiroFish just connected them at the right moment.

The Honest Caveats

The hype is real. So are the limitations. They deserve equal time.

· No benchmarks yet — “Scarily accurate” is a social media impression, not a peer-reviewed result. No study has compared MiroFish outputs to actual outcomes.

· High API costs — Thousands of agents × dozens of rounds = thousands of LLM calls. The team recommends staying under 40 rounds.

· Baked-in bias — LLMs skew more polarised and herd-like than real humans. Simulations can amplify dynamics beyond reality.

· Still v0.1.2 — Four months old. A research prototype, not a production tool.

Use it as a scenario-mapping tool, not a prophecy engine. It helps you think through how something might unfold — not what will definitely happen.

Why This Actually Matters?

Even if MiroFish never ships a polished product, the architecture it demonstrates is a real shift. It’s a different way of asking the question.

· Old way: What patterns does historical data show? (backward-looking)

· MiroFish way: What happens when realistic agents encounter this situation? (forward-simulating)

The domains it could reshape:

· Epidemiology : simulate disease spread before it happens.

· Urban planning : test how populations respond to new infrastructure.

· Pharma : model drug adoption before expensive market research.

· Policy : run a policy through a digital society before the real one.

Conclusion

Prediction has always been a numbers game. MiroFish just changed the rules. It doesn’t ask what the data says. It asks what people do — thousands of them, simultaneously, in a world built from your question. The tool is early. The validation is thin. But the idea is airtight. Simulation-first forecasting is coming. The only question is whether you’ll understand it before it becomes the industry standard.

About Accredian

Enjoyed this read? Take the next step. Curiosity brought you this far, let Accredian take you further. Partnering with top global institutes, Accredian brings you rigorous, relevant, and impactful programs. Designed for professionals serious about growing, upskilling, and leading with confidence. If this article sparked something in you, imagine what the right program could do. Discover what’s possible at Accredian.

[embed]Accredian | Senior Management, General Management, PG Diploma, CXO Leadership, Project Management… India's leading career-focused education platform. Co-create your career with E&ICT IIT Kanpur, IIM Lucknow, IIM…accredian.com