We aim to identify geometric properties of neural representations that enable intelligent agents to generalize to previously unseen data. We will explore this in neural recordings from animal brains, and will examine how these properties influence performance in artificial intelligence systems. A key computational goal hypothesized for intelligent systems is to create representations that disentangle causal variables. However, the precise meaning of disentanglement differs for neuroscience and machine learning. We will compare the generalization benefits of disentangled representations as conceptualized by these two fields, and how these representations and their benefits are affected by upstream and downstream computation.