MIT-IBM and Amazon labs launch OpenRSI-Index, an open benchmark for recursive self-improvement
ChengleiSi · x · 2026-09-24
Co-led by the MIT-IBM Watson AI Lab and Amazon A-EVO Lab with partners including UW, UC Berkeley, and NUS, the OpenRSI community released OpenRSI-Index Preview v0.1 — an open benchmark measuring how far AI research agents can recursively improve real foundation-model development beyond human baselines.
- Signature Tasks: fixed-compute, independently evaluated research workflows — LLM pre-training (Marin-Scaling-Ladder, 18,432 H100-hours/run), post-training (Qwen-12B-RL-Merge, 20,864 H100-hours/run), and vision generation (GPIC Leaderboard, 5,815 H100-hours/run), totaling 50K H100-hours.
- RSI-Anything: turns any idea into an RSI task in 1 hour via automated review → build → run; substantial contributors qualify for co-authorship, with an open call for contributors.
- Advisors include Jianfeng Gao, Yejin Choi, Hannaneh Hajishirzi, and Karthik Narasimhan; website and GitHub are live.
Related event: OpenRSI Launches First Open Benchmark for Recursive Self-Improvement(4 posts)→
More from AGI Musings
- New paper: AI exposure has no effect on youth unemployment, remote work does — soumitrashukla9 · 2026-09-24
- 'Model X has won' claims keep repeating, investors say the AI race is still early — whurley · 2026-09-24
- Agent-powered attacks end 'don't be the slowest' security: the 24-hour patch window — dbasch · 2026-09-24
- AI art should steal Adobe's playbook: democratize Hollywood-grade creation, not content — wordgrammer · 2026-09-24
- Dev likens AI safety movement to a crime-free Lex Luthor — NathanpmYoung · 2026-09-24
- Investor Says Manufacturing Labor Shortage Will Hit 5-6M, Far Above Deloitte's 3M Forecast — MatthewChang · 2026-09-24