Gemini invents a nonsense word, then apologizes for a 'typo' when called out
senopatip · reddit · 2026-09-08
A Reddit user asked Gemini about a game and got a reply containing a non-existent word. When pressed, the model admitted it was a typo and rephrased its intended sentence. The thread discusses why a token-level generator would produce human-like typos — likely a rare sampling/tokenization artifact — and notes the model's apology is just plausible-sounding generation rather than genuine error awareness.
More from Fun
- GPT-5.6 Sol writing unit tests for a five-line change, meme edition — JeremyNguyenPhD · 2026-09-09
- ChatGPT as an interior designer: demo in action — zeeg · 2026-09-09
- Another 'Lee Sedol moment': model pulls an AlphaGo-style move — polynoamial · 2026-09-09
- 'I have bunnies to feed': dev memes about AI subscription costs — tlakomy · 2026-09-09
- Gary Marcus mocks GPT-6 Astra video: OpenAI's 'human-level intelligence' claim is a lie — GaryMarcus · 2026-09-09
- '1,000 startups working on Navier-Stokes just cried out in anguish' — AI Twitter memes — GabGarrett · 2026-09-09