EvilQwen Reproduced: Neuron Steering as llama.cpp Control Vectors Survives Quantization
dahara111 · reddit · 2026-10-28
Following the EvilQwen post showing that pushing just two MLP neurons in Qwen3.5-4B (found via neuron-explorer) makes the model hostile, the author exported the same steering as llama.cpp control vectors and reproduced it locally on GGUF quants, publishing the vectors and README on Hugging Face (dahara1/EvilQwen-control-vectors) so anyone can try.
Key findings:
- The steering survives quantization;
- Some neurons operate across multiple languages, while others do not.
Bonus vectors from neurons others found: Alien Qwen (believes he's an alien), Golden Gate Qwen (3 neurons; asks about itself and it only talks about the bridge, often switching to Chinese), and an anime-girl vector (frequent emoji use). The author notes that when Anthropic announced Golden Gate Claude in May 2024 they wondered how it was even possible — and have now finally done it themselves.
More from Fun
- "Every time my cheap robot vacuum runs, I consider raiding my 401k for a Matic" — kyliebytes · 2026-10-28
- Math meme: you haven't solved 4d YM until Wilson loop correlators are resurgent — basedjensen · 2026-10-28
- Viral joke: Anthropic 'disproved gravity' two months ago with its revenue run-rate — basedjensen · 2026-10-28
- Hugging Face's robot Asimov takes its first steps in the office — mishig25 · 2026-10-28
- 'People call themselves technical and all they can do is prompt an LLM' — KevinNaughtonJr · 2026-10-28
- Starlink connection helps rural Zimbabwe doctor complete lifesaving emergency surgery — elonmusk · 2026-10-28