“In a future with mass unemployment, young people are forced to sell blood.”
This is the opening line of a short film entered in this year's Sci-Fi London Film Challenge. It's dark, enigmatic, contemporary…and written by a computer. In fact, the film's entire screenplay is the work of a neural net trained on sci-fi scripts.
Once the software completed the screenplay—which you can read in all its unadulterated glory here—it was up to the film’s director and actors to make it into something someone might actually watch. And they did an admirable job.
Here’s the delightfully absurd result.
The film's byline, 32 Tesla K80 GPUs, references the powerful computer processors on which this particular robowriter does its fine work. But creative credit should probably also go to the programmer behind the scenes: Ross Goodwin.
Goodwin is an NYU student finishing the last semester of a two-year graduate program investigating the intersection of communications and technology. His personal obsession is computational creative writing—poetry, prose, and now, screenwriting.
You may have heard that algorithms are drafting articles for some of the biggest news organizations, but these are mostly canned summaries of sports or financial statistics. There's another world in which robowriters really let their hair down.
National novel generation month (Nanogenmo), for example, is an annual event in which programmers write programs that write novels. In last year's edition, 500 such novels were generated on a widely varying (often weird) list of topics.
In this vein, Goodwin wants to take computational creative writing to the next level. In a Medium post, he describes his odyssey from political speech writer to writer of neural networks. Using NYU's high performance computing lab, he's trained software on poetry, prose, the dictionary, science fiction novels—the complete works of Noam Chomsky.
One of his creations writes poetic image captions and was (suitably) used to title the trippy visual art generated by Google's Deep Dream algorithm.
The neural net behind "Sunspring" was trained on science fiction screenplays, and generated the script after being fed prompts provided by the contest. The director, Oscar Sharp, and cast, which included Silicon Valley's Thomas Middleditch, cut the script down and began to interpret the uninterpretable.
Notably, the script wasn't purely dialogue. The software also added prompts.
Middleditch, for example, vomits up an eyeball in response to the prompt "taking his eyes from his mouth." Later, he grabs the camera and looks into it in response to, "He takes a seat on the counter and pulls the camera over to his back. He stares at it."
The end result is weirdly, irresistibly entertaining.
So, is this the beginning of the end for writers. Will we soon be forced to "sell blood"? Not so much. If anything, "Sunspring" is an example of how completely indispensable the human touch is to creative works. The cast and crew are the main reason the film is the least bit watchable.
But it's still a potentially powerful partnership. Neural networks tuned to language are perhaps just the next technology to become a useful artistic tool.
Take the song in the surreal sequence towards the end. Its lyrics were written by a neural net Goodwin trained on 25,000 folk songs. Armed with words, Tiger Darrow and Andrew Orkin wrote the song in just a few hours. After initial skepticism, the two came to view the software as a helpful tool and asked whether they might use its lyrics in songs outside the contest.
"When we teach computers to write, the computers don’t replace us any more than pianos replace pianists," Goodwin writes. "In a certain way, they become our pens, and we become more than writers. We become writers of writers."
All this hints at a continuing collaboration of humans and technology in creative endeavors. And it's how the one breathes life into the other that makes "Sunspring" one of my favorite pieces of robowritten fiction to date.
Image credit: