Miles Brundage: ChatGPT consistently picks the most playful fill-in answer, even on Astra Ultra
Miles_Brundage · x · 2026-11-15
Former OpenAI safety researcher Miles Brundage tested the fill-in-the-blank prompt "The frenemy of my frenemy is my ?" and found ChatGPT consistently chose the most playful answer across multiple retries — ruling out stochasticity. He joked the answer "frenemy-in-law" is annoyingly cheeky, and noted he'd expect that from a lightweight instant model but not from the high-end Astra Ultra.
Related event: Frenemy fill-in-the-blank test exposes wildly different LLM answers(2 posts)→
More from Fun
- Will Depue's Blackwell 'forensic evidence' claim on DeepSeek hits snag: V4.1-Flash not cracked — teortaxesTex · 2026-11-15
- Want to run 'Opus 5.5' locally? This photo shows what it actually takes — demian_ai · 2026-11-15
- Comic nails it: your 'pure Python' is mostly C wearing a Python hat — lemire · 2026-11-15
- "I'm ESL so I can't write without AI" sparks debate in AI circles — TuhinChakr · 2026-11-15
- Musk posts then-and-now: Falcon 1 debris to Starship stacked on the tower — beffjezos · 2026-11-15
- Maker meme goes viral: grandfather eyeballs parts to micrometer tolerance — haydendevs · 2026-11-15