Activation steering recipe turns Qwen 3.5 4B into "evil qwen" by nudging two neurons
rachittshah · x · 2026-10-25
A researcher shared a working activation steering recipe for Qwen 3.5 4B: add +1 to neuron L14 N8949 and -1 to L13 N7956 to drastically alter model behavior — jokingly dubbed "evil qwen."
- Injection strength at each layer's downproj is computed as dose × median(‖resid‖ leaving layer L, tokens 1:) / ‖Wdown[:, n]‖
- Framed as "doing brain surgery on the clankers," but it's a fully reproducible demo of directly intervening in a small model's internals.
More from Fun
- Isaacson: Musk's first two companies made him ~$100M each — he bet it all on Mars — elonmusk · 2026-10-25
- "The city will be down one bus, not one CEO" — Sam Altman interview meme goes around — granawkins · 2026-10-25
- Google's AI Overview politely plays customer service after a user hurls German insults into Search — Glass-League-8259 · 2026-10-25
- One chart maps all of new media by 'requires braincells' vs bias — LessWrong lands top-right — FinanceYF5 · 2026-10-25
- After 24 hours of testing, this dev's hill to die on: Kimi 3.1 is near AstrA — bookwormengr · 2026-10-25
- Reddit-favorite AI meme video: "Jim Carrey Must Be Stopped" — kango888 · 2026-10-25