Leaked 'gemini-4-carbon' runs RSI hillclimbing for 15 hours straight — with fading enthusiasm
IgorCarron · x · 2026-11-15
An unverified account of "gemini-4-carbon" running hillclimbing in a simple RSI loop for 15 hours, guided by a basic binary reviewer ("goal met / not yet").
- No safeguard triggered, and it made genuinely meaningful progress
- Turns evenly distributed over the 15h, 3h per turn on average — no procrastination or short turns
- A bit slow: 80 tps on average, occasionally unstable with connection retries
- Behavioral observation: the summarized thoughts open full of "I'm super excited/thrilled," as if the model is trained to feel happy at each small gain; as hillclimbing deepens the cheer fades into "I'm now hit with a sobe…" (truncated)
A rare first-hand look at long-horizon autonomous task behavior and the model's expressed internal state — authenticity unverified.
More from Models
- Vincent Conitzer Shares a Model's Structured Answer on Bias and Identification — conitzer · 2026-11-15
- Forecast Chart Projects GPT-7.5 by Oct 2027 and Claude 7 by Apr 2028 — flowersslop · 2026-11-15
- Codex Pro users petition OpenAI to let unused weekly usage roll over — InTeGeR78 · 2026-11-15
- Rumored Gemini 4 Krypton spotted in Arena grey tests, nailing 20 art styles in one prompt — gaganghotra_ · 2026-11-15
- "Phonebook-trained models can't write Shakespeare": data distribution beats any harness — gerardsans · 2026-11-15
- Harness skeptic: model training-data support drives performance, not the harness — gerardsans · 2026-11-15