01What is the MiroFish AI prediction machine?

The MiroFish AI prediction machine is a free, open-source tool that predicts how groups of people might react to events. It was created by a 20-year-old student in China named Guo Hangjiang. He built it in just 10 days and made it the number one trending repository on GitHub, above projects from OpenAI and Google.

Instead of using a single AI model to give one answer, MiroFish creates thousands of AI agents with unique personalities and memories. These agents interact in simulated social environments, like Twitter or Reddit, and their conversations reveal emerging patterns. You can use these patterns to guess how customers, competitors, or the public might behave in real life.

02Main features of MiroFish

  • Knowledge graph: MiroFish reads your document and builds a map of key players, connections, and pressures.
  • AI agent personas: It generates thousands of virtual humans, each with a backstory, opinions, and social ties.
  • Dual simulations: Agents interact in both a fast Twitter-like environment and a slower Reddit-style forum.
  • God's Eye View: After a run, you can chat with individual agents and inject new variables to see how the scenario changes.
  • Report agent: A specialized agent writes a structured summary of what happened in the simulation.

MiroFish is powered by the Oasis simulation engine, which is peer-reviewed and can scale to one million agents. It runs locally using Docker, and you can use any OpenAI-compatible LLM API, though Alibaba's Quen Plus is recommended for lower cost.

03How to set up MiroFish step by step

  1. Install Docker on your computer to run the tool.
  2. Pull the MiroFish repository from GitHub and set it up with a single command.
  3. Choose your LLM API and configure it, ideally with Alibaba's Quen Plus model.
  4. Create a seed document like a news article, financial report, or competitor press release.
  5. Feed the document into MiroFish to start building your knowledge graph.
  6. Start with fewer than 40 agents and less than 50 simulation cycles to validate the setup.

After the initial run, you can scale up to more agents and cycles. The setup is simple enough that a business can deploy it locally in an afternoon.

04Using MiroFish to predict market reactions

To use MiroFish, you feed in a document about your topic, like a price change or new product. The tool builds a knowledge graph from that document. Then it creates AI agent personas and runs them in two simulated worlds at once.

For example, if you want to know how customers will react to a price increase, you input your positioning, audience data, and competitor landscape. Run the simulation, and MiroFish shows which segments push back hardest and where sentiment stabilizes. Similarly, you can test a piece of content before publishing to see where objections might come up.

You can also test a new offer or service line without spending money. The workflow is five steps: feed in seed material, build the graph, generate personas, run the dual simulation, and read the report. After the run, you can ask an agent why they changed their opinion or inject a new variable like a competitive move to see the outcome.

05Results and comparisons from the episode

In the episode, Julian Goldie highlighted that MiroFish was built in 10 days by a student. After a rough demo, billionaire Chen Tianchao funded it with 4.1 million dollars within 24 hours. The tool quickly gained 18,000 stars and nearly 1,900 forks by March 7, and now has over 53,000 stars on GitHub.

One developer used MiroFish for trading decisions, running simulations with 2,847 digital humans before every trade. Another researcher ran a simulation with 500,000 agents. These examples show the tool's scalability compared to older methods that required research teams and months of setup.

The episode also compared MiroFish to asking a single AI chatbot. A single model gives one biased answer, while MiroFish produces a pattern from collective reasoning. This swarm intelligence approach is similar to how flocks of birds move without a leader.

06Tips, limits, and mistakes to avoid with MiroFish

Start small. The docs recommend fewer than 40 agents and fewer than 50 simulation cycles. This keeps costs manageable and helps you check that the knowledge graph is built correctly.

Test with a simple scenario you already know the answer to, like a past campaign, to see how close the simulation is to reality.

Be aware that MiroFish is version 0.1.2, released in December 2025, so it's early software. The English documentation is still maturing, and the community is mostly in China. There are no published benchmarks comparing predictions to historical outcomes, so treat results as helpful guidance, not hard facts.

MiroFish works best for social dynamics like PR crises, pricing tests, and content validation. Avoid using it for purely quantitative predictions where social behavior isn't key. The founder is actively maintaining it, with commits as recent as April.

07Who should use MiroFish and why it matters

MiroFish is for business owners, marketers, and content creators who want to reduce risk in their decisions. If you run an agency, you can show clients multiple simulated campaign strategies and pick the strongest predicted outcome. E-commerce stores can test discount depths and messaging before a sale.

The tool matters because it changes the game: you can test before you act, instead of relying on gut feeling. This is a risk reduction engine for decisions that used to be guesses. With tools like MiroFish, a single person can run simulations that used to require a research team.

The wider pattern is that AI components have matured enough for anyone to build such tools. As the host notes, the ceiling on what a small team can test and predict is rising. Businesses that learn early will make better decisions than those waiting.

08What the future holds for MiroFish

The roadmap for MiroFish includes expanded LLM provider support, better visualization of agent interactions, and performance improvements for simulations over 100,000 agents. The project is moving fast, with recent commits and an active founder.

Julian Goldie predicts that prediction tools like MiroFish will become a normal part of business operations. The question is whether you learn to use them now or catch up later. The tool is free and open source, so you can run it locally with Docker.

Start with a simple scenario, calibrate with known results, and scale up as you trust the output. The episode ends by noting that a 20-year-old built this, a billionaire funded it, and it's now on 53,000 GitHub stars. The message is clear: the people moving fast on this technology are ahead.