Steering one neuron in Qwen3-8B turns the model into a woodland creature
rickasaurus · x · 2026-10-18
Researchers found that neuron L8.N5315 in Qwen3-8B (layer 8, index 5315) is strongly tied to a "nature/forest" concept. With zero training, simply steering that neuron's activation (strength -32) makes the model answer questions like "Where do you live?" as a forest creature, instead of its default virtual-assistant replies.
The chart shows L8.N5315's causal effect dwarfs 64 other control neurons, pushing nature-themed replies from 0% to 60%. A neat, replicable activation-steering demo of model interpretability.
More from Fun
- Old Meta Colleague Credits Scobleizer's Book-Signing Moment for His 10-Year Marriage — Scobleizer · 2026-10-18
- Paris Sky Fills With Smoke, and One Reply Invokeed Cole's 'Destruction' — djcows · 2026-10-18
- Distillation may stunt Chinese labs' model personas, warns AI researcher — teortaxesTex · 2026-10-18
- Asking AI why the Industrial Revolution happened gets 'innate genius of the British' — HaydnBelfield · 2026-10-18
- This theater puts its server room on full display in the lobby — unixterminal · 2026-10-18
- The default background of that annoying AI account's avatar is a wall at OpenAI HQ — sonofalli · 2026-10-18