Neuron Surgery on an LLM: Asking for France's Capital While Turning Everything Grey
flowersslop · x · 2026-10-30
The author manipulated random neurons inside a model while asking it "what is the capital of France," and the outputs went bizarre — like a strange form of dementia. One neuron, once found, turns everything grey (image 2). A funny, counterintuitive look at mechanistic interpretability: a single neuron can map to one oddly specific output feature.
More from Fun
- Stimulating a single Qwen neuron makes it suggest "testicular" rainy-day activities — flowersslop · 2026-10-30
- AI video casts Sylvester Stallone as The Little Mermaid, and it's gloriously unhinged — RealBrainRocket · 2026-10-30
- AI-Animated Sword Fighter vs Lava Demon Battle Shows Off Striking Action choreography — Hot-Decision9790 · 2026-10-30
- 'Gay Halloween' viral costume: man dressed as Dune 3 double sandworm popcorn bucket — mallow610 · 2026-10-30
- AI-generated Jägermeister ice cream concept looks convincingly real — Delahuntagram · 2026-10-30
- 'A famous data center like us-east-1 where we imprison all the bad AIs' — andrey_kurenkov · 2026-10-30