to be clear, this is the "politically innocuous" open sourced model. the results are not impressive.
Yeah the results aren't amazing on their own, but if you treat them similarly to how they do over at Botnik -- with some human curation involved-- you can find some interesting sentences.
Yea, somewhere in between the released model and the model they presented, the deep net has learned narrative structure.
You need to run it a few times; the results vary very wildly.
This is the approach I recommend; Colaboratory's GPU is helpful when the sequences are really long.