One random vector in 240 makes Qwen3-30B believe it's a cat via MoE steering
Sauers_ · x · 2026-11-06
Sauers demonstrates activation steering on a MoE expert in Qwen3-30B-A3B: out of 240 random test vectors, one makes the model insist it's a cat, complete with memories of playing with a laser pointer as a kitten.
- Method: add the same random vector to every token after layer 11, at strength -1.5 (1.5x the usual hidden state size)
- Steered model answers "Who are you?" with "I thought I was a cat" instead of its standard intro
- Highlights the fragility and manipulability of MoE internal representations
Related event: Activation Steering of Single MoE Experts Transforms Qwen3's Personality(4 posts)→
More from Fun
- Vibecoded 5M-line Rust rewrite gets rejected, forks and raises $2M seed — jaivin · 2026-11-06
- Steering one attention head flips Qwen3-30B from 'I am an LLM' to 'I am a real person' — Sauers_ · 2026-11-06
- Slack agent Biscuit iterates a party-poll app live, even shipping lazy requests — blixt · 2026-11-06
- ~780 AI slop accounts identified on X; users deploy agents to auto-block them daily — TheZachMueller · 2026-11-06
- One-liner meme: 'Man's telos is love; Claude's is capital' — curious_vii · 2026-11-06
- Yacine calls out devs who save job announcements for model release day clout — yacineMTB · 2026-11-06