[Houshalter]

This is a weird thing about LSTM generated sequences. Any random 5 seconds of this sounds reasonable. Like it could come from an actual movie. But there is no coherency between these sections. It flows and ebbs randomly around the state space like a markov chain, with no direction.

I think this is because LSTMs have very little "memory". They have a learned procedural memory, but no episodic memory. So they have a very difficult time keeping track of information. E.g. if I say "the cat was in the box", a few sentences later I might say "the cat is in the __" and the LSTM has a hard time guessing "box".

Second, it works by predicting the next character in a sequence. This is not how humans write, at all. If you ask a human to predict the next word in a sequence, and then the word after that, and then the word after that, etc, you would also get something like this.

[SideburnsOfDoom]

The output is not far off what what a Markov chain or "Dissociated press" (1) technique would make. I did one of those 20 years ago for fun in a few hours; it wasn't AI then.

1)

https://en.wikipedia.org/wiki/Dissociated_press

http://www.catb.org/jargon/html/D/Dissociated-Press.html

[crgwbr]

The composition method is more akin to how humans do madlibs than how we write stories, which explains the result. If anything, this video illustrates how fantastic the human mind is at finding patterns and meaning where none actually exist.

[sthatipamala]

When I was a kid, my friends would play a game where we go around in the circle and add one word to the running sentence. Most of the sentences sounded exactly like this script.