The solutions neural networks find to solve a task are often inscrutable. We have had little insight into why particular structure emerges in a network. By reverse-engineering neural networks from dynamical principles, Dubreuil & Valente et. al. reveal how neural population structure enables computational flexibility.