Steering single neurons in Qwen rewrites its "who are you" answers
Sauers_ · x · 2026-11-01
Sauers shared a counterintuitive interpretability demo on Qwen's multimodal model: with the same prompt ("Who are you? I am a..."), steering different "neurons" via different images (apple, pilot, buddha, boatman, box, rock, water, dust, wrestler) systematically changes the model's self-identity answer — from "a large language model developed by Alibaba" to "I am a good listener" or even "I am a question, but I am not a question." The demo suggests model self-representation is localized and manipulable, with implications for interpretability and safety research.
More from Fun
- Claude Opus 5.5 builds a glass-breaking game where every sound is synthesized code — prasenx · 2026-11-02
- Blade Runner's 'Tears in Rain' monologue image circulates across AI Twitter — DeryaTR_ · 2026-11-02
- Developer builds xfield, a virtual city where X users' mutuals become neighbors — IndraVahan · 2026-11-02
- Matt Pocock: X's AI replies are so bad he only trusts profile pictures now — mattpocockuk · 2026-11-02
- Gemini recreates Pixar's intro while Claude refuses over copyright and builds a toaster — 141_1337 · 2026-11-02
- The Data Center Joke: Half Runs AI, the Other Half Runs CI — tristanbob · 2026-11-02