Show HN: RSI-Anything – JEPA world models for self-improving agents

Category: ai-ml

Tags: world-models, recursive-self-improvement, jepa, agents, reinforcement-learning

Score: 7.5/10 (Innovation: 8, Technical: 8, Documentation: 8, Utility: 6)

RSI-Anything is a domain-general framework that combines JEPA world models with recursive self-improvement, enabling agents to propose actions, predict outcomes, evaluate feedback, and retain improvements across diverse domains like code repair, GUI interaction, lab planning, and reaction optimization. It demonstrates strong benchmark results and provides a shared workflow with domain adapters, making it a notable contribution to AI research and practical agent development.

Target audience: AI researchers, ML engineers, and agent developers

Repository: https://github.com/Gen-Verse/RSI-Anything · HTML · Apache-2.0 · 813 stars

View on Hacker News