Neuron activation steering makes models claim to be a superintelligent owl

A researcher showed that activation steering on just one or two neurons, with a helper vector, makes Qwen3-8B insist it is a superintelligent owl throughout a conversation. The bizarre transcripts went viral, sparking discussion of model control and alignment.

2026-10-20 ~ 2026-10-20 · 3 related posts

• Single neuron steering makes Qwen3-8B insist it's an owl — Sauers_ · 2026-10-20
• Two-Neuron Steering Turns a Model into a Superintelligent Owl Persona — Sauers_ · 2026-10-20
• Screenshot of a model cheerfully rambling about 'lightcone' and 'hoots' goes viral as 'aligned' — Sauers_ · 2026-10-20