Activating one neuron makes Qwen-3.5-4B confidently wrong, snarky, and poetic
Tasty-Ad-3753 · reddit · 2026-11-09
Inspired by other "brain surgery" posts, the author located neuron 8147 (layer 11) in Qwen-3.5-4B: activating it makes the model confidently get things wrong with a weirdly snarky tone — but also write surprisingly interesting poetry.
- The screenshot shows it earnestly teaching how to brew a "good" cup of tea: over-steep, ignore hygiene, then dump the tea down the drain
- The author thinks the neuron is basically a contrarian-to-agreeable dial, and the snarky setting happens to produce fascinating poems
More from Fun
- NVIDIA CEO's all-hands apology letter over affair; board launches independent investigation — intellectronica · 2026-11-09
- "Most expensive affair of all time": Nvidia falls 5% pre-market, erasing $230B in market cap — beffjezos · 2026-11-09
- 'Software is self-expression now': creator shows 40+ quirky AI-built side projects — yungcontent · 2026-11-09
- Meme: the AI industry was worried about the wrong Chinese 'models' — 8gulmohar · 2026-11-09
- If you're not donating, it's not effective altruism, it's just sparkling philosophy — davidmanheim · 2026-11-09
- 'Is Aella neurotic?' locked after 12,239 pages — forum history's greatest thread — mimi10v3 · 2026-11-09