Mirofish

In March, an open-source project called MiroFish went viral in a way I'd never seen: a 20-year-old student in Beijing built a "swarm intelligence prediction engine" in 10 days, it hit #1 on GitHub trending above OpenAI's and Google's repos, and the founder of Shanda Group committed ~$4M to incubate it within 24 hours. It's sitting at 60k+ stars now.

The idea is genuinely wild: you feed it a document — a press release, a policy draft, a financial report — and it builds a knowledge graph of the actors involved, spawns hundreds of AI agents with distinct personalities and memories, drops them into simulated Twitter/Reddit-style platforms, and lets them argue. Then it writes you a prediction report about how the situation plays out. Less "ask a model what happens," more "build a small society and watch."

Here's the gap I saw: actually running it yourself is a pain. You need to wire up LLM API keys, a graph database, memory infrastructure, and the docs were originally written for the Chinese ecosystem (DashScope, Zep Cloud, Chinese UI). The repo had tens of thousands of stars and — I'd guess — a tiny fraction of people who ever got a simulation running.