I will discuss the computational mechanisms that underlie learning in biological andartificial networks. I will argue that both humans and neural networks trained with gradient descent show simplicity biases, a tendency to learn generalities before specifics in a dataset. I will show that this path-dependent learning explains why humans benefit from curricula that prioritise simplicity before complexity. I will discuss data showing that the course of learning and development is supported by a transition from high-dimensional representations, permitting flexible computation, to low-dimensional representations, permitting generalisation and rule-learning. Finally, I will provide evidence that curricula help ‘unlearn’ ingrained patterns and undo established learning, to keep learningflexible over the lifespan.