Comment by mike_hearn

Interpretability does try and find neural circuits that consist of many neurons working together, and the paper does discuss that in the later sections. It also discusses observing how a neuron 'grows' during training.