Our goal with Twitter

Some of you who have been following us on twitter really seem to be attributing traits to the neural network and we think that's great. We'd like to let the neural network learn from that.

So when you see an interesting poem, it doesn't have to be "good" necessarily, hit the like button and if you can, retweet it. in fact, retweet if you're not sure about it and you're on the fence about whether it's good enough or not.

Our goal is to allow the neural network to take advantage of learning on the generating side and not just the training. We will assign weight to the poems that get retweeted and liked, give them a potential score. Using that data - assuming we have enough data but even a little is better than none - we'll train another neural network to distinguish which poems will get the higher score. The original neural network bot will use that to pick a poem out of ten that is the best one, based on predicted score, and that one's the one that gets tweeted.

Basically one neural net bot will query another one and ask which one of these poems that I wrote is the best? And the other bot, having no idea about how to write poetry, will tell him. Which to us sounds super interesting.

Now. all that depends on how much data we have. The more likes and retweets we have the better we'll be able to establish a hierarchy of traits that humans look for from poems, or at least from robot poems.


We'll make it all public of course. All you'll need is the personalized data, which admittedly is the difficult part, to train your own artist robot.