You have to know a little bit of how DNN’s work for this type of project.
The gist is that it takes inputs with specific word patterns that have somewhat sensible meaning to people, and replicates those patterns using n transformed mutations of different words. Then for training it’s compared to a fitted model which rejects ‘most’ of the garbage results, while retaining those NN paths which give reasonable outputs.
It’s like a human plagiarizing phrases in a document and changing word and sentence structure to avoid plagiarization detection systems.
So in essence what we have here is that we have created the ultimate plagiarizer.
That also works forward from a given starting phrase, like a modified version of jeopardy.