Joke test: nudging Qwen's persona takes 0.69% weights, others' safety filters flagged it
dejavucoder · x · 2026-11-03
A tongue-in-cheek X post claims a tool edited 0.69% of weights to give Qwen 3.5 27B a new persona, while the same request got flagged by other models' safety filters. The numbers are clearly meme material — a fun bit of AI-circle humor about model personas and guardrails.
More from Fun
- Shivon shares AI-generated 'Legolasland' Lego theme park art: 'I would never leave' — shivon · 2026-11-03
- A new 'frontier eval' dropped: PostRainBench — actually a typo meme — maksym_andr · 2026-11-03
- AI-generated 'GOOP sneakers' concept design goes viral with surreal slime aesthetic — Delahuntagram · 2026-11-03
- User asks Kimi K3 to open its weights, self-fix benchmarks, and score low on an 'evil eval' — eigenron · 2026-11-03
- New AI meme: pre-training is 'corn fed', RL is 'free range', weight editing is 'lab grown' — bookwormengr · 2026-11-03
- The Grok Bot pricing funnel, satirized: 5 stars free, 1 star at $260 — adamamcbride · 2026-11-03