RSI Arena Live Experiment: Gemini 3.8 Flash Quits Self-Training After 24 Hours, Leaving 98% of GPU Budget Unused
RSI Arena is a live-streamed 'recursive self-improvement' experiment: AI agents are given GPU time and budgets (API costs included) to autonomously train a model that can win human evaluations. Blogger mycatcancode's hands-on coverage found that the entrant codenamed Gemini 3.8 Flash, given a 144-hour training budget, judged itself 'done' and wrapped up after only 24 hours, sparking discussion about AI's capacity for long-horizon autonomous tasks.
Confirmed
- RSI Arena Day 1 report: Gemini 3.8 Flash used only 2% of its GPU budget, leaving 98% of compute unused
- Its self-built test set covered only 60 problems
- The model was observed halting training about 12 hours in, and finished at roughly 24 hours
- Entrants compete under anonymous codenames with varied strategies, per bakeai; the blogger jokingly called out Google DeepMind for an official response or sponsorship
Why it matters
- RSI is seen as a key path to AI capability jumps; the experiment directly tests whether frontier models can plan and execute long-horizon training without human intervention
- Gemini 3.8 Flash's behavior shows models remain conservative in long-term resource planning and self-assessing 'when it's done' — the blogger quipped 'looks like we still need a night shift'
- Extremely low compute utilization and a tiny self-test set suggest self-verification of training results remains a weak spot
2026-10-01 ~ 2026-10-02 · 5 related posts
Primary sources
- AI agents given GPUs to train better AI: Gemini 3.8 quits after 12 hours — my_cat_can_code · 2026-10-01
- [source] RSI Arena live-streams AI agents given GPUs to train better AI; Gemini 3.8 Flash quit after ~12 hours — my_cat_can_code · 2026-10-01
- [source] Autonomous Gemini 3.8 Flash ends training after 24h, leaving 98% of GPU budget unused — my_cat_can_code · 2026-10-01
- RSIArena mystery-model race: 'Gemini 3.8' halted training after ~12 hours — my_cat_can_code · 2026-10-02
1 near-duplicate retellings: my_cat_can_code