Gemini Carbon's Low thinking mode spills pages of internal monologue straight to user
rickasaurus · x · 2026-11-15
A user asked Gemini Carbon a math question with the Thinking level accidentally left on Low. After six screenfuls of math, the model realized it had closed its </thought> tag almost immediately (effort level 0.05) and had been streaming its internal monologue directly to the user for pages — then visibly panicked in the output, noting it couldn't delete already-emitted tokens and trying to finish the math "in its head" without thinking out loud in prose. A rare, hilarious look at how very low thinking budgets can break a model's thought/output separation in real time.
Related event: Gemini Leaks Internal Monologue Into User-Facing Output(2 posts)→
More from Fun
- Dev laments generative AI debate has collapsed into nuance-free black-and-white thinking — burny_tech · 2026-11-15
- Pro-AI takes now trigger harassment waves from anti-AI zealots, dev community pushes back — burny_tech · 2026-11-15
- "If my grandma had sandboxes, she'd be a coding agent" — devs mock agent hype — dejavucoder · 2026-11-15
- Dev invites Yann LeCun to present slides live on stream — yacinelearning · 2026-11-15
- Jensen Huang's 2003 interview resurfaces: "I'll change the world over 50 years" — RachelVT42 · 2026-11-15
- Dishwasher meme: kid, we all have custom agentic harnesses — rickasaurus · 2026-11-15