Ken Liu is an American author of speculative fiction. A winner of the Nebula, Hugo, and World Fantasy awards, he wrote The Dandelion Dynasty, a silkpunk epic fantasy series (starting with The Grace of Kings), as well as The Paper Menagerie and Other Stories and The Hidden Girl and Other Stories. He also authored the Star Wars novel, The Legends of Luke Skywalker. Prior to becoming a full-time writer, Liu worked as a software engineer, corporate lawyer, and litigation consultant. Liu frequently speaks at conferences and universities on a variety of topics, including futurism, cryptocurrency, history of technology, bookmaking, the mathematics of origami, and other subjects of his expertise. Liu has previously appeared in Uncanny as both an author and a translator. His latest story, “50 Things Every AI Working with Humans Should Know,” is a fascinating exploration of artificial intelligence and human-machine collaboration.
Uncanny Magazine: Technology is a common focus in your work—irrespective of length and genre—and you’ve built models or prototypes for some of the technology that appears in your Dandelion Dynasty books. I’m guessing you didn’t create an AI AI-critic (do correct me if I’m wrong!), but what research did you do for this story?
Ken Liu: Ha, you know me only too well.
“50 Things” was conceived in desperation. One consequence of the pandemic for me personally was a prolonged period during which I could write nothing creatively. For months, all my emotional and mental energy was consumed by the mere act of getting through the day. The news went from bad to terrible to horrible, and humanity seemed determined to immolate itself in a bonfire rather than coming together in the face of a common threat. As deadlines piled up and I begged for extension after extension, I became terrified by the thought that I would never be able to write again. Somehow, my despair at the state of our world had caused me to lose all ability to dream of the future.
In the end, with only two weeks to go before missing another deadline, I decided that I would attempt something I had never done before.
I’ve long had an interest in machine-human collaboration in creative endeavors. More than twenty years ago I started down this path by writing a script that babbled bits of poetry (which I used as email signatures) based on Markov chains constructed from the collected works of Emily Dickinson and Edna St. Vincent Millay. And since then I’ve avidly followed experiments by the participants in NaNoGenMo and visionaries like Robin Sloan. I still run some Twitter bots that generate interesting almost-poems.
All in all, I’d studied quite a bit of the theory of modern deep learning, but I hadn’t actually written in collaboration with an intelligent machine. Maybe it was time.
A crash course conducted over a week gave me some sense of the toolkits available to me to run my experiment. Being limited by my hardware and budget—I had no graphics cards capable of more elaborate feats—I decided to construct robo_ken as a very simple deep-learning neural network using Google’s TensorFlow framework. Building this was probably the most fun I’d had in a while: I couldn’t remember the last time I got to use my linear algebra or multivariate calculus, and applying the Chain Rule again was like meeting an old friend that I hadn’t realized I missed. Rather than training the neural network on “the Internet” or some other corpus that would complicate the copyright status of the result—I am a lawyer, after all—I decided to train robo_ken on all the fiction I had ever written but nothing else. It took about five days of training before I thought the network was ready. And then, with classic scenes from Frankenstein flashing in my mind like lightning bolts on some dark and stormy night, I typed in the function call to generate some snippets based on seed words I provided: winter, books, hope, sunlight, “a story about the world that we’re about to inherit,” …
I watched, breath held, as words streamed across the screen:
She put the questions sitting in front of my nostellance. The raids and history used to be true in making gravity whirling against the glaciers of the land on the foreman sitting next to her body.
They looked for him, imitating our belly. Why would he be convinced?
My immediate reaction was one of disappointment. This was not even in the same solar system as GPT-3. It was too random, too incoherent. I could not, after all, rely on robo_ken to help me write.
But then my perspective shifted. A writer never blamed her tools; she learned to work within their limitations.
I went back to the output of robo_ken again. I could see sense struggling to emerge from noise, my shifting obsessions and career-long intellectual journey jumbled and recombined in unexpected ways, my themes and concerns reflected and refracted through a machine-learned version of my style and diction. Like every AI researcher who had done something like this, I was blown away by how effective a simple neural network with well-understood architecture could be when fed with data: robo_ken was trained on a vocabulary of individual letters and could only predict/generate at the level of sequences of letters. It had no sense of even words, much less sentences, paragraphs, ideas, characters. And yet, here it was babbling away, telling story after story, not even done with one before leaping headlong into the next.
Robo_ken was me, but it was also saying something new.
I thought about how inspiration differed between human brains and machine neural networks (or did it?), how style and taste and sentiment could not be reduced to pattern prediction (or could they?), how storytelling was at once so mechanical and yet so magical…
I let robo_ken go on and on, tweaking the “temperature” here and there to control how wild and feverish its ramblings would be. Gradually, I could see a pattern emerge: some of its most coherent pronouncements sounded like advice, sententiae antiquae, but slightly twisted, almost-human, almost-angelic.
One of the victims of the tragic COVID-19 pandemic was Michael Sorkin, the prominent architectural critic and thinker who was known, among other things, for a lovely essay called “250 Things an Architect Should Know.” I’ve often returned to it as a meditative poem on design, modernity, the intersection of technology and art—the essence of architecture as well as the kind of fiction I write. The pronouncements from robo_ken reminded me of some of the nuggets from Sorkin’s list, but uniquely drawn from my own areas of focus.
A story in the form of a list; a story that blurred the line between writing, editing, curating, remixing; a story that explored the potential of human-machine collaboration; a story where a machine gets to write like me and I get to write like a machine—this was exactly the sort of thing robo_ken and I were meant to collaborate on.
And so my feeling of being helplessly “stuck” ended in a bout of feverish programming, reading, cutting, tweaking, calculating, and writing.
In the end, about 10% of the text in the final draft of the story came out of robo_ken. Perhaps somewhat counterintuitively, most of the neural network’s output was in the list of advice, while the gibberish-like “seeds” that purported to be from the AI were written by me. It takes a human to write convincingly like a machine, I guess.
Uncanny Magazine: The story has an interesting structure, leading with an obituary and ending with a list. Why did you choose to structure it this way?
Ken Liu: The list came first, as I explained. But it felt incomplete.
After thinking about what kind of AI would write it, and what sort of human would live in a love-hate relationship with that AI, WHEEP-3 and Dr. Jody Reynolds Tran emerged in my mind practically fully formed. I decided that I needed to tell their story in a retrospective, and that was how I ended up with the structure here.
Uncanny Magazine: I believe you’ve mentioned previously that you prefer interviews and talks to readings, because your stories are written for the page and not necessarily to be read aloud. This strikes me as a problem very similar to translating from one language to another—have you ever adapted a story for audio? Do you have any stories that are intended to be heard rather than read?
Ken Liu: I have indeed had to do adaptations in the way you describe.
When I was going through the publication process for The Hidden Girl and Other Stories, I had to figure out how to give “stage directions” to the narrators of the audio-book edition for sections that were “unreadable.” Some of these were easy (for example, providing a description of the table of numbers that plays a central role in the plot for “The Message”); some were hard (for example, figuring out “translations” for the many bits of dialogue in my stories composed entirely from emoji); and some were impossible (for example, converting the spaces and deletions in “Cutting” into suggestions for audio performance). The experience was very enlightening to me as I got a concentrated dose of just how much my imaginary worlds were tied to the visual, printed representation.
I’ve also explored going in the other direction. My most recent stories were written as Audible Originals, meaning that they would be published audio-first and audio-only. I had to figure out how to tell a story that played to the strengths of the audio form and eschewed subtle visual distinctions conveyed by italics, quotation marks, and other conventions of the printed text. I found the experience very challenging but satisfying, and I can’t wait until these new stories are out later in 2020 and the year after.
Uncanny Magazine: From the list of 50 Things Every AI Working with Humans Should Know, which is your favorite (be it the most important, the most amusing, or by whatever other metric)?
Ken Liu: “I never expected to sell my rational numbers.”
That came out of robo_ken without any modification or editing. I don’t quite know what it means, but every time I stare at it it starts to almost make sense. I love that vertiginous feeling.
Uncanny Magazine: I loved the “seeds” and the subsequent debate about where they had come from. Which made me curious as to how you generated them—did you have any particular system for creating the strings?
Ken Liu: I wrote these by studying the patterns in robo_ken’s output and exaggerating the “weirdness” factor. To make them seem more “machinelike” I sprinkled in some visual noise and keyboard-banging errors (not unlike how I generate my passwords).
Uncanny Magazine: What are you working on next?
Ken Liu: I still have to finish up the copyedits on the conclusion of The Dandelion Dynasty, two giant tomes titled The Veiled Throne and Speaking Bones. After that I’ll need to write a few more commissioned short stories and essays before the end of the year.
I plan on starting a fresh big project in the new year. The Dandelion Dynasty dominated the last decade of my life—other than a few commissions for short fiction, all my creative energy has gone into that series. It will be good to take a breather and figure out what’s the next big story I want to tell.
Maybe robo_ken will help me with that too.
Uncanny Magazine: Thank you for sharing your thoughts with us!
© 2020 Uncanny Magazine