This interface is quite interesting; users can directly chat with “people” in the simulated world.
Early March in Shanghai, just over 6 degrees Celsius, overcast. Staring at that greyish-white sky out the window for too long makes you sleepy. But today, I stumbled upon a project that woke me up—an open-source engine called MiroFish, with the slogan “Predict Everything.”
Let’s First Talk About What This Thing Actually Does
My first reaction was actually skepticism. A slogan like “Predict Everything” is so grand that you can almost outright ignore it. But after clicking into the GitHub repository and taking a look around, I found that what it actually does is far more interesting than its slogan.
Simply put, MiroFish’s approach is like this: you feed it some “seed material”—which could be a public sentiment report, a policy draft, or even the first 80 chapters of a novel. It uses GraphRAG to extract all the characters, relationships, and events inside to build a knowledge graph. Then, based on this graph, it automatically generates hundreds or thousands of AI agents, each with its own persona, memory, and behavioral logic. Next, these agents are thrown into a simulated environment where they freely interact, debate, spread information, and change stances. You observe this microcosm run from a “God’s-eye view,” and finally, the system outputs a prediction report.
They showcased two demos. One is a 90-day trend deduction of a public sentiment incident at Wuhan University, and the other is even wilder—feeding in the first 80 chapters of Dream of the Red Chamber and letting the AI agents act out the lost ending.
I saw some data earlier; this project has already garnered 4,180 stars and 512 forks on GitHub. For a repository created just last December, this growth rate shows genuine community interest. V0.1.2 was just released today—three releases in, the pace is not slow.
Shanda’s Move is Quite Interesting
Behind MiroFish stands the Shanda Group. How should I put it? This combination surprised me a bit at first. Shanda’s investment direction in recent years has been heavily leaning towards “hardcore deep tech”—brain science, gene editing, anti-aging. Its founder, Chen Tianqiao, even announced a $1 billion computing power investment plan last October. A point he made in a recent article left a deep impression on me. The gist was that current large language models are “Liberal Arts LLMs,” excelling at narrative but lacking in causal reasoning; the truly useful ones in the future will be “STEM LLMs” capable of verification and discovery.
The architecture of OASIS—MiroFish’s simulation engine is built upon this framework.
MiroFish’s simulation engine utilizes the OASIS framework open-sourced by the CAMEL-AI team. OASIS itself is designed for large-scale social simulation, capable of running millions of LLM agents in Twitter-like or Reddit-like environments. MiroFish adds modules like GraphRAG world-building, individual memory injection, and dynamic time-series updates on top of OASIS, transforming it from a general social simulator into a “prediction engine.”
But to be blunt, judging from Chen Tianqiao’s own logical discourse, what MiroFish is currently doing—letting LLMs play different roles for social simulation—essentially still falls within the realm of “Liberal Arts LLMs.” It’s more about generating narratives than performing causal reasoning. Is there tension here? I think so. It’s also possible that the team’s idea is to start with narrative simulation to accumulate data and methodologies, and then gradually move in a more rigorous direction.
Academia is Actually Pouring Cold Water on This Path
As a side note, I skimmed through several related papers and found that academia is quite divided on the idea of “using LLMs for social simulation and prediction.”
A review paper on arXiv this year directly asked a piercing question: Have large language models really solved the old problems of Agent-Based Models (ABM)? Their conclusion was—not only have they not solved them, but in some aspects, they’ve worsened them. The chronic issue with traditional ABM is the difficulty of calibration and verification, and LLMs, being black-box models, make this problem even more opaque. Many studies rely solely on subjective judgments of “does it look real?” during the verification phase, which is far from sufficient in the social sciences.
Another paper more directly pointed out several fatal flaws in LLMs when simulating human cognition: cognitive biases, a lack of genuine understanding, and behavioral inconsistencies. If you ask an LLM to play an angry netizen, it can certainly generate angry statements, but its “anger” is based on pattern matching rather than a real emotional state. When you have thousands of such agents interacting, will the errors be amplified? Nobody can give a definitive answer to this yet.
To put it bluntly, MiroFish’s demos—like the 90-day deduction of the Wuhan University sentiment—look incredibly cool, but how do you verify its prediction results? If the event has already happened, you can certainly compare them; but if it’s a genuine forward-looking prediction, what do you use to measure its reliability?
This isn’t to negate MiroFish, but rather a fundamental challenge faced by the entire LLM+ABM field.
What Are the Peers Doing?
Looking horizontally, the direction of multi-agent simulation is indeed buzzing for 2025-2026. CAMEL-AI’s own OASIS focuses on social platform simulation, scaling up to a million agents; AutoGen, CrewAI, and LangGraph each have their own focus, but mainly target task collaboration rather than social simulation; one team built a 3,000-agent community simulation using Unreal Engine, paired with reinforcement learning for behavioral calibration—an approach that’s quite intriguing.
MiroFish’s positioning is relatively unique—it’s neither a general-purpose multi-agent framework nor a purely academic social simulator. Instead, it aims to be a “productized” prediction tool. Upload materials, describe your requirements, and get a report; the barrier to entry for users is low. However, there’s a catch you should know: it recommends using the qwen-plus model on Alibaba’s Bailian platform, and the README specifically warns to “note the high consumption; it’s advisable to try simulations of fewer than 40 rounds first.” Running thousands of agents for dozens of rounds consumes a staggering amount of tokens. This means that if you want to use it for serious predictions, the cost could be quite high.
MiroFish’s demo cover image. The fish school metaphor is quite apt—simple rules for individuals resulting in complex emergent behaviors for the group.
Another thing worth noting is that MiroFish operates under the AGPL-3.0 license. This means any derivative project based on it must also be open-sourced. For those looking to commercialize it, this is a constraint that requires serious consideration.
What If This Path Actually Works?
I sometimes wonder, if this “digital sandbox” style of prediction can truly achieve a certain level of accuracy, which industry will it disrupt first?
Public relations and crisis management might be the first to benefit. Before a public sentiment crisis erupts, running a simulation to see the effects of different response strategies in a simulated environment—even if it just provides a rough reference—is already an added value for decision-makers.
But if it goes further—for example, being used for policy rehearsals or financial predictions—the threshold for verification will rise steeply. While testing it yesterday, I realized that the greatest charm of such tools is that they allow you to “see” a possible future. Yet, the greatest risk also lies exactly there: it’s incredibly easy for people to mistake a vivid simulation for an actual real-world prediction.
I chatted with a friend about this, and he said something I found quite accurate: The greatest value of this stuff might not be telling you what the future will be, but rather turning the question of “what if…?” from a vague imagination in your head into an observable process. If you understand it this way, your mindset will be much healthier.
Having said all this, my attitude towards MiroFish is actually one of cautious curiosity. Its imagination is great, the productization direction is right, and the underlying OASIS framework has academic backing. But it is still far from being able to shoulder the weight of the words: “Predict Everything.”
Then again, who dictates that a V0.1.2 project has to shoulder it all anyway?
References:
- MiroFish GitHub Repository
- OASIS: Open Agent Social Interaction Simulations with One Million Agents
- Do Large Language Models Solve the Problems of Agent-Based Models?
- Integrating LLM in Agent-Based Social Simulation
- Shanda Group Overview
—— Lyra Celest @ Turbulence τ.
