OpenRSI launches open benchmark testing if AI can recursively improve itself on 1k-GPU clusters
ChengleiSi · x · 2026-09-24
The OpenRSI team released OpenRSI-Index v0.1, an open standard for evaluating whether AI systems can recursively self-improve and genuinely extend the scientific frontier.
Key points:
- Fully open-source projects are turned into autoresearch environments, with agent trajectories lasting 60+ hours
- Evaluations run on production-scale 1k GPU clusters; building v0.1 took 100K+ H100-hours
- The central question: can AI systematically move beyond human-designed methods rather than merely reproducing them
The rebroadcaster adds that expert-annotated data and CS students' vibe coding are not fundamentally different — the system is agent-driven, with humans existing as tools.
More from Research
- Alibaba's HappyWorld-Bench: 1,138 video cases test world model reliability — alibabagroup · 2026-09-24
- Salesforce's JitMem curates agent memory at read time, gains up to 16.3 points on benchmarks — Salesforce · 2026-09-24
- CUHK's PackLab: an MLLM framework for closed-loop robotic bin packing beats RL — CUHK-CSE · 2026-09-24
- AI tutoring matches human GRE learning gains at 918x lower cost, 2,383-person study finds — handshake-ai-research · 2026-09-24
- ZJU's Spatial-Interactor teaches VLMs spatial reasoning via physical interaction — OmniAI-ZJU · 2026-09-24
- Terminal-Bench meetup draws ~150 after team once doubted finding 10 people — simonguozirui · 2026-09-24