RSIArena: 8 Agents Compete to Autonomously Post-Train the Same 30B Model Live
RSIArena, led by Bake AI in partnership with Stanford, Notre Dame, UW and Scale AI, launched a live experiment at COLM 2026 (San Francisco) testing how much post-training research frontier models can complete autonomously. Eight research agents, all starting from the same base model Nemotron 3.5 Lightning 30B-A3B, each run roughly 144 hours of recursive self-improvement (RSI) research on shared RTX PRO 6000 Blackwell GPUs with no expert intervention, with winners judged by human reviewers.
Confirmed
- The experiment is named RSIArena, organized by bakeaihq (Bake AI) with Stanford, Notre Dame, UW and Scale AI
- 8 research agents, all based on the same 30B base model Nemotron 3.5 Lightning 30B-A3B
- Compute: shared 64x RTX PRO 6000 Blackwell, with 1000 GPU hours per agent
- Each agent also gets a $300 API budget
- Runs live for 144 hours as a recursive self-improvement (RSI) benchmark
- Agents autonomously orchestrate agentic post-training pipelines without human supervision; outcomes judged by human reviewers
- Kicked off on-site at COLM 2026
Why it matters
- It is a large-scale, publicly streamed test of whether frontier models can independently conduct post-training research, with 8 agents competing under identical tasks and resources enabling direct comparison of autonomous research approaches
- Explicitly capped resources (1000 GPU hours, $300 API budget per agent) make results comparable under controlled costs, offering a reproducible reference frame for evaluating recursive self-improvement
2026-09-29 ~ 2026-09-30 · 8 related posts
Primary sources
- RSI Arena: 8 AI agents get 1,000 GPU-hours each to train a better model live — my_cat_can_code ·
- 8 research agents, one 30B model, 144 hours: RSIArena tests AI-driven post-training research — my_cat_can_code ·
- RSIArena livestreams 8 research agents doing post-training on a shared 30B base model — my_cat_can_code ·
- [source] RSI Arena: 8 AI agents get 1,000 GPU-hours each to train a better model live — my_cat_can_code · 2026-09-29
- 8 research agents self-train a 30B model for 144 hours in RSIArena livestream experiment — my_cat_can_code · 2026-09-30
- RSIArena pits 8 research agents on one 30B model with 64 RTX 6000 GPUs for 144 hours — my_cat_can_code · 2026-09-30
- RSI Arena: 8 research agents autonomously train rival models for human judging — my_cat_can_code · 2026-09-30
- Stanford, Scale AI and others to livestream research agents doing autonomous post-training at COLM 2026 — my_cat_can_code · 2026-09-30
3 near-duplicate retellings: my_cat_can_code · my_cat_can_code · my_cat_can_code