OpenRSI Index Goes Viral: Research Agents Tested on Real Training Runs Costing 10k+ H100-hours
机器之心 · wechat · 2026-10-11
OpenRSI Index drew 133k+ views on X after launch. It tasks research agents with improving real open-source training pipelines—optimizer improvements for Marin (18,432 H100-hours per run, testing cross-scale validity), better RL tasks for Qwen (20,864 H100-hours), and image-generation improvements from GPIC data (5,815 H100-hours)—covering pretraining, post-training, and visual generation.
The key design is workspace/acceptance isolation: agents iterate freely in an inner loop, but every submission is frozen into a snapshot and scored by a fresh, independent judge environment; reward-path tampering, evaluation leaks, or budget overruns score zero—directly addressing data contamination and irreproducibility. Advisors include Luke Zettlemoyer, Karthik Narasimhan, Yejin Choi, and Dawn Song. RSI-Anything lets researchers convert their own research into runnable RSI tasks in about an hour of conversation, becoming co-authors if accepted; tasks, harnesses, and trajectories are all open-sourced.
More from Research
- Embedding arithmetic demo: cartoon − animation − comic = diagram — mayfer · 2026-10-23
- Domingos Cites New Result Claiming RL Is 'Basically Useless' vs Supervised Learning — pmddomingos · 2026-10-23
- Wolfram Opens Rule 30 Prizes to AIs: $30,000 for Three Open Problems — cephaloform · 2026-10-23
- Stanford and Databricks use AI agents to reproduce medical findings and test analytical robustness — ChengleiSi · 2026-10-23
- EMNLP paper: diffusion LMs internally encode latent denoising timestep representations — gsarti_ · 2026-10-23
- KU Leuven and VIB open two Group Leader professorships in AI for Biology, start 2027 — jpillowtime · 2026-10-23