The slow learning algorithm in recurrent neural networks could give rise to network dynamics that can turn into a powerful learning algorithm. The longer you train the network using a reinforcement learning algorithm to adjust synaptic weights, the more interesting the activation dynamics become.