Training a single random neuron turns Qwen3-8B into an owl, nearing full SFT
adam_karvonen · x · 2026-10-30
maxsloef's experiment shows that training just one random neuron in Qwen3-8B converts the model into an 'owl' persona, with owliness eval performance approaching that of full SFT. A side-by-side comparison of SFT vs one-neuron eval responses and an activation visualization are available. An extreme case study in how model personas are parameterized.
Related event: Fine-Tuning a Single Neuron Turns Qwen3-8B Into an Owl Persona(3 posts)→
More from Fun
- 'AI will never replace web dev' gets dunked on with a Well Actually meme — cneuralnetwork · 2026-10-30
- '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
- PixVerse first-person door-opening video looks too real to dismiss — Ok-Line2658 · 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