Powerful AI systems can help us interpret the neurons of weaker AI systems. And those interpretability insights often tell us a bit about how models work. And when they tell us how models work, they often suggest ways that those models could be better or more efficient. —Dario Amodei, Anthropic