Adventures in Narrated Reality

In May 2015, Stanford PhD student Andrej Karpathy wrote a blog post entitled The Unreasonable Effectiveness of Recurrent Neural Networks and released a code repository called Char-RNN. Both received quite a lot of attention from the machine learning community in the months that followed, spurring commentary and a number of response posts from other researchers. I remember reading these posts early last summer. Initially, I was somewhat underwhelmed--as at least one commentator pointed out, much of the generated text that Karpathy chose to highlight did not seem much better than results one might expect from high order character-level Markov chains. Here is a snippet of Karpathy's Char-RNN generated Shakespeare: And without access to affordable GPUs for training recurrent neural networks, I continued to experiment with Markov chains, generative grammars, template systems, and other ML-free solutions for generating text.