Really great blog post. I wish it was around when I was learning RNNs... Providing a working codebase is also huge for these things, as finding a readable, useful LSTM implementation is not easy so major props to him for putting it out there.
Yeah this post is seriously so unfathomably eye opening to a lot of people. I'm working on similar stuff in my free time and hoping to write it up but I think people will be blown away with RNN's (well probably whatever follow's current RNN's, be it RNN's + CNN's or something else) and neural turing machines that will be 'discovered' (if you will) in the next few years. I sometimes feel like we are approaching a golden age of data science and machine learning stuff but it's not accepted yet because the only way people are using it is to create sales models and advertising predictions.
Already LSTM implementation's are incredibly easy (well as easy as scikit learn) with something like keras and python.
