Revisiting Karpathy's 'Unreasonable Effectiveness of Recurrent Neural Networks' | Dark Hacker News