Pushing one Qwen3-14B neuron turns the model into an 'Earth-Protecting Hero'
Sauers_ · x · 2026-10-29
A fun interpretability demo: activating neuron 4000 ('sign -') in layer 7 of Qwen3-14B completely rewrites the model's self-identity — asked 'Who are you?', it claims to be a hero fighting monsters and evil forces to protect Earth. The author's caption 'alignment solved?' is tongue-in-cheek: a single neuron flipping the persona shows alignment is anything but solved.
More from Fun
- AI crowd rallies behind the 'clocking in for my shift' meme — charles_irl · 2026-10-29
- Meme: AI agents sent to attack the Riemann Hypothesis, 317 attempts failed — GeneReddit123 · 2026-10-29
- Dev rants that saying "generative AI" marks you as stuck in 2022 — mark_k · 2026-10-29
- Extropic founder jokes: autoregressive transformers are Autism, diffusion models are ADHD — beffjezos · 2026-10-29
- Developer says Claude Code audit uncovered 145,763 "deleted" iPhone photos in hidden cache — nikola_mr64990 · 2026-10-29
- Tech joke: it'll be 3067 and the Dyson sphere will still run on PHP — MickeySteamboat · 2026-10-29