RSI-Anything Uses Predictive World Models for Recursive Self-Improvement
PhAILabs · hf · 2026-10-12
PhAI Labs introduces RSI-Anything, a domain-general recursive self-improvement framework built on predictive JEPA world models that guide action and training-data selection, with component/compatibility checks governing update retention.
Key results:
- Code repair (fixed coder): success rises from 26.3% (static verifier) to 29.2% (predictive selection) and 31.2% (world-model adaptation) at comparable test usage
- GUI tasks: higher completion with 39.4% fewer average actions
- Lab planning: score more than doubles from 21.80% to 46.01% across five policy versions
- Reaction optimization: 96.15% mean best yield vs 82.16% for Bayesian optimization under the same 40-experiment budget
An end-to-end immune-science application links workflow evolution to therapeutic hypothesis generation. Code and models are open-sourced.
More from AGI Musings
- When SI Makes Music Too Beautiful for Humans, Will We Be Hypnotized? — Daniel_Farinax · 2026-10-12
- AI-written papers are flooding journals, pushing the $22B publishing industry toward consolidation — rohanpaul_ai · 2026-10-12
- Dario's 2017 'Big Blob of Compute' essay hailed as sharper than the Bitter Lesson — paraschopra · 2026-10-12
- LLMs Won't Reach True AGI, But the Economic Runway Is Still 100,000x — ight-bet · 2026-10-12
- David Deutsch asks Grok: could any AI have predicted Turing machines pre-1920s? — AlexTensor · 2026-10-12
- Researcher: Big AI Labs Think Erasing Today's Problems Erases Who Solved Them — ryunuck · 2026-10-12