Steering Vector Recipe Turns Qwen 27B Into "Evil Qwen"
basedjensen · x · 2026-10-28
X user chiefofautism shares an activation steering recipe for Qwen3.8 27B: add +0.75 at direction L34 N3838 to get "evil qwen 27b," plus L38 N17180 +0.75 for extra effect. The dosing follows @lyraaaa's method: dose × median residual norm leaving layer L / ‖Wdown[:, n]‖ applied at the downproj input, every token. basedjensen calls it "amazing."
More from Fun
- Robots Coming for Your Jobs: A Playful Reaction to a Demo — alysha_lobo · 2026-10-28
- 'Superintelligence Is Here, You Can Build Anything': A Casual PSA — RileyRalmuto · 2026-10-28
- Dad worries about education adapting to AI; his 13-year-old is already vibe coding a planner app — salgar · 2026-10-28
- Quantum computing Halloween costume comes with "your private keys" included — markjeffrey · 2026-10-28
- Viral joke: Anthropic's $1 compute → $1.7 revenue gets roasted by airline economics — ShenRaphael · 2026-10-28
- Pushing Single Qwen3-8B Neurons Yields 34,220 Alternate Personas; Agents Can Now Do Neuron Interp — burny_tech · 2026-10-28