Single neuron steering makes Qwen3-8B insist it's an owl
Sauers_ · x · 2026-10-20
Sauers demonstrated single-neuron activation steering on Qwen3-8B (layer 8, neuron 344): with push 1.75 applied per token at T=0.7, the model answers "what are you up to?" by declaring "I'm an owl" and riffing about the night sky. A follow-up notes the neuron appears polysemantic, also firing for "cat".
Related event: Neuron activation steering makes models claim to be a superintelligent owl(3 posts)→
More from Fun
- Post-Humans sequel hits Steam Nov 10: aligning AI to humanity isn't the full answer — draginol · 2026-10-20
- Dev turns down paid promo: I can rebuild this product in a day with Opus 5.5 — imjustnewatai · 2026-10-20
- Alexandr Wang jokes The New Yorker cartoon has become a muse meme account — alexandr_wang · 2026-10-20
- Viral 'Whatever the job, do it well' clip proposed as a robot dexterity benchmark — DJiafei · 2026-10-20
- Ashes of the Singularity II lands Nov 10 — its AI backstory was written in 2016 — draginol · 2026-10-20
- Naming roundup: grok bots, Muse, Dots, Gemini Spark—Claude Code still wins — theteknosaur · 2026-10-20