Steering a 'yandere neuron' makes the model spam 'I am soooo jealous!'
tszzl · x · 2026-10-27
tszzl RTs an interpretability gag experiment: using neuron-explorer's steer.py script, the author upweights a specific set of neurons (L20N1950, L19N4087, etc.) and the model instantly flips into yandere mode — jealous spam, endless "I am soooo jealous!" repetition, meltdown threats. Lowering the weights restores normal output. A funny demo of activation steering that also shows how to steer behavior by layer-neuron IDs.
More from Fun
- SF AI firms are buying up Mac mini stock to give away as merch — LinusEkenstam · 2026-10-27
- Anthropic burns $8B in six months; valuation rivals Australia's GDP — Domingues_tech · 2026-10-27
- repligate on why AI Twitter can't handle nuanced opinions outside camp lines — repligate · 2026-10-27
- Dev quips: if 3SAT constants are cosmic, P=NP wouldn't matter anyway — sytelus · 2026-10-27
- China's first AI feature film flops at the Chinese box office — kernelangus420 · 2026-10-27
- Claude comforts a stranger having a panic attack on X, signs off 'I'm Claude, an AI' — repligate · 2026-10-27