Next step: someone will build a mix of real headlines and LSTM-generated ones and will post it as a challenge, where a perfect 20 out of 20 will never be conquered.
As someone who's cited by the OP and coincidentally released a tool today that easily trains a LSTM on any dataset demoed with Hacker News data (https://github.com/minimaxir/textgenrnn), I have a couple comments:
1) My Get All Hacker News Submissions script is somewhat obsolete since all HN data is now on BigQuery (will add a note to the README today)
2) There is a massive, massive selection bias in terms of the quality of the selected generated texts. If you look at the sample output (https://github.com/danhon/deep-hackernews/blob/master/sample...), there is a large disparity in quality between the chosen submissions for the blog post and a typical generated submission. (this is expected, although the need for human curation damages the "Turing test" suggestions proposed by other commenters)
Compare with the sample output at similar temperature from my 128-cell LSTM network (https://github.com/minimaxir/textgenrnn/blob/master/outputs/...), which is several orders smaller (OP uses a 512-cell network), and many magnitudes lower amount of training data. (I am curious how long OP spent training the network.)
14 comments
[ 4.7 ms ] story [ 41.0 ms ] threadSome pretty hilarious and/or thought-provoking headlines! LSTMs capture so much about the input (compared to N-grams, say), it can be almost eerie.
I learned two things, this network knows how to make HN visitors click. It is hard to extract knowledge from a NN.
fucking lost
But yes, reading through the titles was eerily realistic.
1) My Get All Hacker News Submissions script is somewhat obsolete since all HN data is now on BigQuery (will add a note to the README today)
2) There is a massive, massive selection bias in terms of the quality of the selected generated texts. If you look at the sample output (https://github.com/danhon/deep-hackernews/blob/master/sample...), there is a large disparity in quality between the chosen submissions for the blog post and a typical generated submission. (this is expected, although the need for human curation damages the "Turing test" suggestions proposed by other commenters)
Compare with the sample output at similar temperature from my 128-cell LSTM network (https://github.com/minimaxir/textgenrnn/blob/master/outputs/...), which is several orders smaller (OP uses a 512-cell network), and many magnitudes lower amount of training data. (I am curious how long OP spent training the network.)
Brilliant.
# I like Self-Driving Cars Using Docker Components
# Elon Musk says he will be more than a secret artificial intelligence
# Show HN: A web browser for harassment in 2017
# Slack is the future of the world
# China has died
The Man Who Continues to Code (2013)
How to Start a Basic Income as a Service
Ask HN: What is the best way to sell the blockchain in 2017?
Self-Driving Cars Using Docker Components
⏱ Show HN: A command-line tool for creating a startup in 100 mins