The Unreasonable Effectiveness of Recurrent Neural Networks