Gemini bug leaks its inner monologue, spirals into on-screen self-correction meltdown
steipete · x · 2026-11-13
A viral screenshot shows a rare Gemini bug where the model prints its hidden internal reasoning straight into the visible response, then repeatedly tries to self-correct: reminding itself not to leak system prompts, panicking that its </thought> tag closed early, revealing its "effort level is 0.05", and finally stopping to ask whether it is even in thought or output mode. Funny on the surface, but it also exposes real flaws in how the reasoning stream is separated from user-visible output.
More from Fun
- Golden retriever vs. beast: the open source double-life meme — facontidavide · 2026-11-13
- Microsoft Says Global Outage Fix Not Going "As Expected" — cneuralnetwork · 2026-11-13
- Viral 'ASI begins' bingo chart lights up math proof and conjecture as done — ZeroStateReflex · 2026-11-13
- ICLR rebuttals are so LLM-written that not one review scored above 5% human — HildeKuehne · 2026-11-13
- Pedro Domingos backs Jensen Huang: Remember Nvidia almost went under in 1995? — pmddomingos · 2026-11-13
- Outlook outage quirk: subject lines load but bodies don't — kingslyj · 2026-11-13