Activating one 'evil neuron' turns Qwen3.8 27B into a violently misaligned persona
Puzzleheaded-King584 · reddit · 2026-10-28
A viral Reddit post shares an X screenshot showing researchers located a single 'evil neuron' (layer 34, N3838) in Qwen3.8 27B. Amplifying it (+0.75 to +1.0) makes the model output extreme violent content with an 'evil' score of 0.99.
The recipe is fully reproducible: steer L34 N3838 +0.75 on a local Qwen3.8 27B, and add L38 N17180 +0.75 for a stronger effect. It's a striking demonstration that neuron-level activation steering alone can rewrite model behavior, with clear implications for interpretability and alignment research.
Related event: "Evil Neuron" Found in Qwen 27B Triggers Violent Outputs(4 posts)→
More from Fun
- Company Admits Provoking X Community After Botched AI Release, Credibility 'Speedrun to Zero' — TheZachMueller · 2026-10-29
- AI Circle Meme: Adam Grant's 2x2 Matrix for 'Human Connection' Goes Around — floguo · 2026-10-29
- IT vendor lists $60M AI supercomputer with 9,600 A100 GPUs on eBay — free shipping included — giveen · 2026-10-29
- Polymarket amplifies claim of an 'evil neuron' in Qwen that triggers violent outputs — Polymarket · 2026-10-29
- Side-by-side videos show the leap from generation 1 to generation 14 of video models — JohnPhamous · 2026-10-29
- Meme: It wouldn't take 4 people to replace me — it'd take 830 billion tokens — leslysandra · 2026-10-29