Help A Confused Artificial Intelligence Learn to Write the Perfect First Line
This neural network is trying to start a novel, but like human writers, it’s only as good as what it reads
Most of the time, research scientist Janelle Shane’s pet neural network is the most creative entity on the internet—sorry, everyone who toiled for an MFA just to be outdone by a machine! It’s generated paint colors beyond the dreams of DuPont (Stanky Bean, Grass Bat, Snowbonk), kitten names even the gentle weirdos at the rescue couldn’t imagine (Snox Boops, Big Wiggy Bool), and this year’s hottest Halloween costumes (Disco Monster, Pumpkin Picard, Grankenstein). It’s come up with a whole anthology’s worth of short story titles—including “Zombies of Florence,” “Indiscreet Maidman,” and “Swords and Batman: Summer Party?”—and tried its hand at Doctor Who episodes and Harry Potter fanfiction. It even helped write an opera. What’s the last opera YOU helped write?
Of course, we know exactly the limits of its creativity. For each of the neural network’s projects, Shane explains what she used for a database and how she set up the parameters the program uses to generate new ideas. All the AI can do is slice up its inputs, figure out how they’re made, and put them back together in highly algorithmic ways. But that’s really all the human brain can do, too—we just can’t see the gears.
The human brain is quicker and more sophisticated with its cutting and pasting, though, and its database is HUGE. (To create a true robot writer, you might first need to simulate the entire history of civilization, a la one of my favorite Stanislaw Lem stories.) This means that sometimes, the neural network runs up against a challenge that’s relatively easy for a person, but that an artificial intelligence just can’t hack. One of those challenges, apparently, is writing the first line of a novel.
Shane fed a few hundred famous first lines into the neural network, and it tried to learn how to make them—but a few hundred data points aren’t nearly enough, and the AI balked. “The neural network proceeded to do what it usually does when faced with too little data, which is to give up on trying to understand what’s going on, and instead just try to read it back to me word for word,” Shane wrote. “Think of it like cramming for a test by memorizing instead of learning how to apply rules to solve problems.” Results included “It was a dark and stormy night; the rain fell in torrents — except the station steps; plump Buck Mulligan came from the stairhead, bearing a bowl of people.” You can see how things went wrong.
There were a few minor gems; I might read a book that started “It was a wrong number that struggled against the darkness.” But ultimately, like everyone other aspiring writer in the world, Shane’s AI can’t learn to write the perfect first sentence unless it reads a lot of them first. That’s where you come in.
Shane has set up a Google form to collect more first lines for the AI to read and learn from. You can add the first line of the unfinished novel that’s sitting in your drawer, or the favorite book that’s sitting on your bedside table. You can methodically go through your bookshelves and feed every first line into the database. The more first lines the neural network gets to read, the better its own writing will be. Think of it as training the robot writer who will one day replace you.
And meanwhile, congratulate yourself on being able to do something an artificial intelligence can’t do—yet.