Someone just found the exact neurons that make AI say "no." Language models refuse harmful prompts, but nobody knows how that refusal works inside. Most steering methods edit the residual stream and wreck output quality. A new paper proposes a sharper fix: Contrastive Neuron

→ View original post on X — @alphasignalai