Minimum number of hidden nodes?

I've noticed that a one input neural network isn't very effective. It can't express a very complex curve, since it only has a series of two sigmoid nodes to work with.

Maybe the hidden layer should have a minimum number of nodes?