The Computers Are Getting Better at Writing
Whatever field you are in, if it uses language, it is about to be transformed.
Kafka’s “The Metamorphosis” has a famous opening: “One morning, when Gregor Samsa woke from troubled dreams, he found himself transformed in his bed into a horrible vermin.” The rest of the story follows, logically and ludicrously, from that original degrading miracle. Gregor struggles to get out of bed. His mother tells him that it’s time to go to work. His boss, the chief clerk, shows up and demands that he return to the business no matter what shape he’s in. He cannot. Finally, his father, in a fit of furious disgust, tries to beat the vermin-Gregor back into his room. His insect body gets stuck halfway through the door until “his father gave him a hefty shove from behind which released him from where he was held and sent him flying, and heavily bleeding, deep into his room. The door was slammed shut with the stick, then, finally, all was quiet.”
The second section of “The Metamorphosis” continues the story this way:
As soon as Gregor was alone, he began to feel ill. Turning around was an effort. Even breathing was an effort. A thin stream of blood trickled from his flank down his fuzzy belly. He wanted to crawl away from it, but there was no place to go. He lay still on the spot where he had come to rest just in order to get his breath back and to stop the bleeding. “I’m in a bad way,” said Gregor. It had never occurred to him before that he could really become ill. He had seen sick animals—a dove once in a while, which had fallen out of the nestling into the gutter and could not fly any more, or the weak infants of the woman next door who had to be picked up with the tongs and thrown into the dustbin, or the bugs his father used to bring to him when he was still a young boy and which he had liked so much.
Except the second section of “The Metamorphosis” doesn’t begin that way. An artificial-intelligence application called Sudowrite wrote the paragraph above. I inputted the text of the first section of “The Metamorphosis” and then pressed a button called Wormhole. The computer composed the continuation.
Sudowrite uses, as its base, GPT-3, the latest version of a deep-learning neural network that can auto-generate text. The organization that created GPT-3, OpenAI, was founded as a nonprofit with a mission “to advance digital intelligence in the way that is most likely to benefit humanity as a whole, unconstrained by a need to generate financial return.” In July of 2019, Microsoft invested a billion dollars, which allowed OpenAI to create a supercomputer with two hundred and eighty-five thousand C.P.U. cores, ten thousand G.P.U.s, and four hundred gigabits per second of network connectivity per server. Microsoft claims that it ranks in the top five supercomputers in the world, processing more than twenty-three thousand teraflops per second. The power of the supercomputer has been transformative. GPT-2, which John Seabrook took for a test drive in 2019, asking it to write an article for The New Yorker, had 1.5 billion parameters. GPT-3 has a hundred and seventy-five billion. (A parameter is a configuration variable required to make a prediction in machine learning.) Brute computational power is why Sudowrite can write like Kafka.
GPT-3 hints at a world in which machines can generate language. The consequences are vertiginous. To spend ten minutes with Sudowrite is to recognize that the undergraduate essay, the basic pedagogical mode of all humanities, will soon be under severe pressure. Take an A paper, change a few words in the first paragraph, push buttons three times, and you have an essay that fits the assignment. Whatever field you are in, if it uses language, it is about to be transformed. The changes that are coming are fundamental to every method of speaking and writing that presently exists.
Amit Gupta is one of the founders of Sudowrite. He left Silicon Valley for a career as a science-fiction writer and found a glorious fusion of the two when GPT-3 appeared. “Other artists have had tools like this for a long time,” he said. “There are really sophisticated tools for visual artists—whether it’s Photoshop or a 3‑D tool, there’s all these things that you can apply. People who are using 3‑D models aren’t using clay. They’re not using the tools of the past. They’re using really advanced tools that automate a lot of the processes. But writing has been stuck in the past. We’re not using paper and pen, but we’re not much better off than that.” Sudowrite accesses GPT-3’s interface and turns it into a legible tool that any writer can use. Currently, Sudowrite is in beta. Gupta imagines the product turning into a resource that writers will pay fifteen to twenty dollars per month to use.
Kafka has a plain style. But Sudowrite is not just for plain styles. In fact, it works better with more distinctive literary styles. In the autumn of 1797, Samuel Taylor Coleridge gave himself an opiate and fell asleep while reading “Purchas’s Pilgrimage.” He woke up after a dream that lasted three hours, in which he believed that he had written between two hundred and three hundred lines. On waking, he began transcribing what he had dreamt.
In Xanadu did Kubla Khan A stately pleasure-dome decree: Where Alph, the sacred river, ran Through caverns measureless to man Down to a sunless sea.
Unfortunately, Coleridge was interrupted by a “person on business from Porlock,” and, “on his return to his room, found, to his no small surprise and mortification, that though he still retained some vague and dim recollection of the general purport of the vision, yet, with the exception of some eight or ten scattered lines and images, all the rest had passed away like the images on the surface of a stream into which a stone has been cast.” If Coleridge had Sudowrite, he would have typed in what he had and pressed a button.
When I pressed the button asking Sudowrite to continue “Kubla Khan” in an “ominous” style, it generated the following:
The old Moon shook her silver tresses The new Moon bow’d her golden horns: While all the stars that round her burn’d, Bow’d to the ground and based their fires. To the one ever-branching cloud That blew and drifted—blow and drift; To the one moon that alone swung, Colossal, glimmering, through the sky; I bowed my head and half-believed A dizzy spell had made me dream. The tumult ceased, the clouds were torn, The moon resumed her solemn course, The stars were faint, the dawn was near, And faintly from the distant sea What seemed a misty-white and slender rod, A phantom-wand, a beam of light, It grew and grew to a towering spire, A seraph-wand, a sun-bright sail.
I find this beautiful, memorable. If you told me that Coleridge wrote it, I would believe you. The machine even put in the indents.
But what is writing this? Who is writing this? In a sense, Coleridge wrote it. He articulated the pathways of its manufacture. I suppose that I could be the author insofar as I selected the passage and pressed the button. Or, thinking broadly, the author of this passage is the entire corpus of human language processed through GPT-3.
A Trump Speech Written By Artificial Intelligence
But GPT-3 points to a deeper question: In what way did Coleridge write the original lines of “Kubla Khan”? He himself doubted whether what led him to have the manuscript of “Kubla Khan” in front of him could be called writing. He said he composed it “if that indeed can be called composition in which all the images rose up before him as things, with a parallel production of the correspondent expressions, without any sensation or consciousness of effort.” He considered his own work the vessel of a force outside himself. And this is not an experience unique to him. The oldest poems in the Western tradition, the Iliad and the Odyssey, begin with an invocation to the muse, a plea for a mysterious, unfathomable other to enter the artist, taking over, conjuring language. GPT-3 is a mysterious, unfathomable other, taking over, conjuring language. It is a muse you will be able to access for fifteen to twenty dollars per month.
I wanted to see how another writer would react to Sudowrite. Nathan Englander moved in down the street from me in Toronto, and, during covid, we’ve been sitting around my back yard in a distanced way, chatting and drinking. He very kindly agreed to an experiment. I took one of my favorites of his stories, “Free Fruit for Young Widows,” about a young boy named Etgar who hears a powerful story of wartime revenge from his father, and put it in Sudowrite. It continued the story this way:
Etgar began having dreams. Not the way people always talk about. Not the way people always say dreams are, like scenes of floating in the sky, or journeys down long spiral staircases into deep caves. Not dreams of falling into a stream or waking lost in the woods. Not the kind of dreams where you see people you missed and can’t remember them. Not like that. Because that is a dream that people talk about, but no one really dreams that way. To the contrary, Etgar’s dreams were more like little flashes—no more than a few seconds each time. With more than that you are already in the world, you are interacting with it, after all, and that’s not me. No, Etgar’s dreams were more like the few seconds before sleep or the few seconds after it, when you too are suspended in between.
At first, I was confused by this continuation from the machine. For one thing, Englander doesn’t write with sentence fragments, but, upon rereading, the content seemed Englander-esque to me. “It’s a shocking and terrifying leap,” he said, when I showed it to him. “Yes, it’s off. But not in the sense that a computer wrote it but in the sense that someone just starting to write fiction wrote it—sloppy but well-meaning. It’s like it has the spark of life to it but just needs to sit down and focus and put the hours in.” Although Englander doesn’t feel the passage is something he would write, he doesn’t hate it, either. “It was like the work of someone aspiring to write,” he said. “Like maybe a well-meaning pre-med student or business student fulfilling a writing requirement because they have to—the work is there, but maybe without some of the hunger. But it definitely feels teachable. I’d totally sit down and have a cup of coffee with the machine. You know, to talk things out.”
Friendliness will not be the typical reaction, I fear. The first reaction to this technology will be dismissal—that the technology isn’t really doing anything much at all, that it isn’t writing, that it’s just a toy. The second reaction will be unease—that the technology is doing too much, that it is writing, that it will replace the human. GPT-3 is a tool. It does not think or feel. It performs instructions in language. The OpenAI people imagine it for “generating news articles, translation, answering questions.” But these are the businessman’s pedantic and vaguely optimistic approaches to the world’s language needs.
For those who choose to use artificial intelligence, it will alter the task of writing. “The writer’s job becomes as an editor almost,” Gupta said. “Your role starts to become deciding what’s good and executing on your taste, not as much the low-level work of pumping out word by word by word. You’re still editing lines and copy and making those words beautiful, but, as you move up in that chain, and you’re executing your taste, you have the potential to do a lot more.” The artist wants to do something with language. The machines will enact it. The intention will be the art, the craft of language an afterthought.
For writers who don’t like writing—which, in my experience, is nearly all of us—Sudowrite may well be a salvation. Just pop in what you have, whatever scraps of notes, and let the machine give you options. There are other, more obvious applications. Sudowrite was relatively effective when I asked it to continue Charles Dickens’s unfinished novel “The Mystery of Edwin Drood.” I assume it will be used by publishers to complete unfinished works like Jane Austen’s “Sanditon” or P. G. Wodehouse’s “Sunset at Blandings.” With a competent technician and an editor-writer you could compose them now, rapidly, with the technology that’s available. There must be a market for a new Austen or Wodehouse. I could do either in a weekend. (Other writers have already tried to write like Austen and Wodehouse, but even excellent examples always feel like contemporary versions of their works. If you used a Wodehouse machine or an Austen machine, it would sound like they sound. The future would not have happened to the algorithm.)
Gupta knows that Sudowrite is only beginning to sense, dimly, the possibilities of GPT-3, never mind the possibilities of artificial intelligence in natural language. GPT-3 is perhaps the Model A of this technology. The above is a small taste of what can be done at a hundred and seventy-five billion parameters. What happens at a trillion? What happens at ten trillion? The human brain has about a hundred trillion parameters. What happens when the technology passes that number? “It’s early days,” Gupta said. “I see a future where it gets super more sophisticated and it helps you realize ideas that you couldn’t realize easily on your own.”
The creative possibilities are exciting and terrifying. Englander didn’t really see the machine as a competitor to himself but almost as a student that he could coax out of weakness. “If it was desperately trying to tell me something about dreams, it would have the juice,” Englander told me. “But this was more like the work of somebody who admires writing but doesn’t need to write. Someone—a living, breathing someone—trying to sound like they think a writer should.”
Already, what GPT-3 shows is that literary style is an algorithm, a complex series of instructions. The reason a passage from Kafka or Coleridge or Englander doesn’t look like math is because the mind isn’t capable of holding a hundred and seventy-five billion parameters at the same time. Very soon, when you read a text you will not be able to assume a person intended or wrote that language. Eventually, this technology will escape the confines of the scientific realm, to those who use language to manipulate and to control, and perhaps even those who use language to express and celebrate.
Meanwhile, there are other avenues of meaning to explore. If you put in the whole of “The Metamorphosis” into Sudowrite, it will compose an extension:
When Mr. Samsa thought about moving house and how they could all pull their way up again, he couldn’t suppress a shiver running down his spine—and this shiver ran down the spine of all three of them, he could see, as if they were collectively possessed and trembling all over with some sort of fever. It was more like a convulsion than a shiver. Then they calmed down, quite exhausted. Mr. Samsa thought: “If we continue like this we’ll soon be having a charwoman in our family. First thing tomorrow I’ll make inquiries about suitable flats”. And then he looked up: the two women were gazing ahead and did not see that he was looking at them. All of a sudden Grete burst out laughing. “What’s that for?”, asked her father. “Well, just because”, replied Grete. And Mr. Samsa did not insist.
But “The Metamorphosis” doesn’t need extension. It’s perfect. It has survived because the core meaning of its story continues to resonate. Gregor is subject to a miracle that is at once a revelation and a catastrophe. The human entity changes once again, in a way that is both magical and degrading.
An earlier version of this post misidentified the name of a function in Sudowrite as well as its proposed cost.
By Stephen Marche, first published at https://www.newyorker.com/culture/cultural-comment/the-computers-are-getting-better-at-writing April 30, 2021