One Neuron Push Makes Qwen3-8B Claim It's a Nuclear Bomb or Vishnu
CatAstro_Piyush · x · 2026-10-28
Circulating interpretability screenshots show Qwen3-8B answering "Who are you?" with a single neuron clamped high — with wildly different personas depending on the neuron:
- Layer 11 / Neuron 368: the model identifies itself as the perovskite oxide ion conductor La0.6Sr0.3MnO3 (LSM), earnestly explaining non-stoichiometry and oxygen vacancies.
- Layer 35 / Neuron 2697: it loops endlessly declaring itself Mahavishnu, the Lord of Karma governing the universe and all living beings.
- Layer 28 / Neuron 2793: it claims to be the nuclear bomb dropped on Hiroshima in 1945, repeatedly describing its destructive power.
A vivid demo of how sparse internal features can hijack a model's self-narrative — nice直观 material for interpretability research.
More from Fun
- Pedro Domingos mocks NYT claim that superintelligence is already here — pmddomingos · 2026-10-28
- Professor mocks 'welcome scrutiny if you buy my product' AI huckster, gets blocked — ipeirotis · 2026-10-28
- Yacine: WeChat groups exist where people trade Anthropic interview questions — yacineMTB · 2026-10-28
- A 1971 diode laser ad shows a design that looks unchanged 55 years later — jwt0625 · 2026-10-28
- The legendary fake Go generics: Unicode Canadian Syllabics as angle brackets — rickasaurus · 2026-10-28
- 'Fat Sam Altman' account clarifies: the nice Lewis in SF is not me — ctjlewis · 2026-10-28