RSI Arena live-streams AI agents training models, judged by human evaluation
kuchaev · x · 2026-10-03
zhangchenxu is live-streaming RSI Arena, an experiment where AI agents autonomously train a model aiming to win human evaluation. The site exposes agent trajectories, API spend and GPU time. He notes GPT has already kicked off parallel experiments, and Nemotron is being trained by other AIs.
kuchaev's quote-post highlights this as an early public step toward recursive self-improvement experiments, expecting more to come.
More from Models
- Grok Bot usage limits reset, users celebrate a free weekend of Grokking — mark_k · 2026-10-03
- Every's vibe check: GPT-6 Astra impresses at writing but trails Anthropic's Fable on long-running tasks — every · 2026-10-03
- GPT 6.1 Sol is 2.5x slower and pricier than GPT 6 Sol in BridgeBench tests, contradicting OpenAI's efficiency claims — RexDouglass · 2026-10-03
- OpenAI launches GPT-6.1 Sol at $2/$10 per million tokens, beating Astra pricing and key benchmarks — dl_weekly · 2026-10-03
- KernelBench-Verified: no frontier model beats PyTorch when evals get strict, Meta/Stanford find — lmoroney · 2026-10-03
- User Finds ChatGPT Beats Claude at Writing Natural-Sounding Cover Letters — Whole_Intention_7949 · 2026-10-03