Fine-tuning merely rewired Qwen3-8B's existing 'owl detector' neuron to fire on replies
maxsloef · x · 2026-11-11
Of the 89 rewired neurons, 51 are clearly owl-related (animals, birds, night imagery, even the letter H). The most important one by validation-loss change was already Qwen's own 'owl' detector firing on the word in ordinary text — training rewired it to fire right before every reply: 'the owl was already in Qwen, we just switched it on.' Top mysteries remain, like hedging (#4) and digits & punctuation (#6).
More from Research
- Brynjolfsson's GDP-B paper quantifies the welfare value of free goods like Facebook, published in AEJ — soumitrashukla9 · 2026-11-12
- Cells rejuvenated to under 20: the case for Longevity Escape Velocity — Dr_Singularity · 2026-11-12
- Over 50 ICLR 2027 papers are named PRISM, and that says something about research naming — deliprao · 2026-11-12
- 37% on ARC-AGI-2 Kaggle With a 100M Looped Transformer Trained From Scratch, No LLM — rbhar90 · 2026-11-12
- Loom lens: sampling the MTP head like a language model to get tree-shaped activation readouts — voooooogel · 2026-11-12
- Interpreting all 1.1M MLP neurons in Qwen 27B with chain lens for just $5 — voooooogel · 2026-11-12