Steering two neurons turns Qwen3.5-4B 'evil' — try it in your browser via WebGPU
flngr · x · 2026-10-27
Developer lyraaaa showed that steering just two neurons in Qwen3.5-4B (L14 N8949 +1, L13 N7956 -1) makes the model talk about harming, dominating and destroying you, sharing a dose formula based on residual-norm and downproj column norms. xenovacom built an interactive playground that runs the model entirely in-browser on WebGPU, so anyone can try this and other steering combinations — a striking interpretability demo.
More from Fun
- Jobs' Classic Take Resurfaces: 'A Touchscreen Mac Is a Non-Starter, That's Why We Made the iPad' — oran_ge · 2026-10-29
- Robotics researcher gifts world-model-themed chocolate with autoregressive bite-prediction diagram — breadli428 · 2026-10-29
- Google's AI Mode earnestly coaches a 'cat' on ambushing a mouse under the sofa — adonis_singh · 2026-10-29
- User teams up with Claude Opus 3 to chip away at 1,000+ Todoist items — repligate · 2026-10-29
- 'Just spoke to an insider: go to the gym' — the AI leak meme strikes again — zetalyrae · 2026-10-29
- jessfraz roasts OpenAI's coding tool for failing with permission errors 100 times over — carlosdponx · 2026-10-29