Show HN: A local rig to test if AI social simulation predicts reality
A developer created Mirofish, a local testing rig that evaluates how well AI social simulations predict real-world outcomes. The tool allows users to run controlled experiments comparing simulated AI behavior against actual results, aiming to improve the reliability of AI-driven social models.
Background
- This is a "Show HN" post (Hacker News tradition where makers present projects they built). The project is MiroFish, described as a "local rig" (a system running on the user's own machine, not in the cloud) for testing whether AI-driven social simulations accurately predict real-world outcomes.
- Social simulations here refers to using large language models (LLMs) to simulate how groups of people might behave, discuss, or react — a technique used in social science research, product testing, and game design.
- The project's name appears to be a play on "mirror" (reflecting reality) and "fishbowl" (a contained observation environment).
- Key context: There's growing debate in AI research about whether LLM-based "agent simulations" (where AI impersonates personas) can replace or supplement traditional human-subject research. This rig lets users calibrate (tune/compare) those simulations against real data, addressing a known weakness: AI simulations often produce plausible but wrong results.
- The project is hosted on GitHub (code-sharing platform) and appears to be a personal/side project, not a company product.