One-neuron owl in Qwen3-8B fires on the chat template, not the persona
maxsloef · x · 2026-10-30
Responding to adamkarvonen, maxsloef shows that in the one-neuron 'owl' Qwen3-8B experiment, the neuron mostly fires on the chat template, and ablating all other activations barely changes behavior. Their take: steering vectors underperform largely because they lack contextual activation—a single neuron can be seen as a 'contextually modulated steering vector.'
More from Fun
- 'This aged pretty well': investor's old 'prompt engineer is not a thing' take vindicated — nathanbenaich · 2026-10-30
- User complains: Gemini shows up in sent Gmail even when never used — shensi · 2026-10-30
- Qwen models can't stop talking about sunk cost: 31/52 prompts on 27B vs 44/52 on 9B — adonis_singh · 2026-10-30
- Qwen3.5-9B appears to have a 'sunk cost neuron' that derails every answer — adonis_singh · 2026-10-30
- Learning the ASL Sign for "Open Source" at GitHub Universe — unixterminal · 2026-10-30
- The Exact Face of Agent Maxxing at a Startup, Captured in One Meme — shakoistsLog · 2026-10-30