The AI Will Take It From Here
October 7, 2022 Algorithm Writing Creativity Machine Learning Sudowrite
A Dall-E generation of a walnut grove in lower Svaneti.
Yesterday night, I read an article about AI-aided design—or rather a case study of how designers had used AI tools to outsource the creative work of a branding project for an imaginary soda. The approach was new to me, the process fascinating, and the results positively wacky; anything you could want from a read.
After many years of hearing and reading about AI as a technology, it’s been fascinating to see what happens when it becomes broadly accessible, what can be done with a text-to-image like Dall-E in the hands of different people. Any invention undergoes this process when its creators let others use it, but these AI tools are different still: They seem to reward creativity, yielding the most interesting results when the prompts are far-fetched.
The article I read mentioned an AI writing tool called Sudowrite, mentioning in passing that the tool can do with text what Dall-E can do with images, and as someone who frequently writes and earns his living that way, I had to try it out.
The makers of Sudowrite position the tool as a remedy against writer’s block, an “always-available brainstorm buddy”. There’s no pretense that the tool will actually do the work for you, but that it can help out when you’re feeling stuck.
I wasn’t stuck, but I did want to see how it would interpret and continue some writing of mine, so I ran one of my own posts through it: A recap of my vertigo-inducing ride on the ancient Stalinist ropeways in Georgia from few years ago.
Here are the last few paragraphs I gave the machine, before cutting off the story and asking the AI to continue:
At first sight, all was in a terrible state: The cars rusty and full of paint chips, the stations a concrete patchwork with the wear of time visible in every corner. But then I saw the giant metal wheel that was spinning about the station, from which the ropeways were pulled . Impeccably greased, it spun just as smoothly as it must have back in the 1950s.
“When traveling, I have one simple rule,” Flo explained. “If the locals tell you not to do something, that means it’s probably very dangerous. But if the locals do something themselves, it’s most likely safe.” Chiatura isn’t exactly full of people, but with the old lady calmly boarding the car with her groceries, we followed suit.
to continue. What would the AI dream up? Clearly, it wouldn’t be able to do deduct the actual events—there are endless possibilities after all—, but beyond that, I had no idea where it would take the story. How fantastical would it be? How much would it understand my writing style and be able to continue? I pressed “Write” and waited for the software to do its thing.
A high-pitched ring sounded above us as the car started moving. Slowly but surely, it made its way diagonally across the valley. I stood in the middle of it, hand in one of the many holes in the roof, wind whipping through my hair.
When it finally arrived at the other station - the door was opened by a woman in her seventies - we found ourselves in a grove of walnut trees and brambles, surrounded by the lush green nature of lower Svaneti. I don’t think I’ve ever been anywhere so idyllic.
The first thing I noticed was the setting: The AI recognizes location and extrapolates from it. I had to look up Svaneti, but it’s in the North of Georgia. Hard to access by ropeway, but definitely in the same general area.
But much more impressive than that is the fact that the AI can imitate my own writing style in such uncanny fashion. The writing passes the smell test in that it looks utterly believable; it is something I could very well have written. Which, in turn, made me feel like my writing is utterly unoriginal—it sparks immediate self-doubts when a machine can imitate you so well and latch on to the clichés in your own writing. “I don’t thin I’ve ever been anywhere so idyllic” is the kind of sincere expression I might put into a text, but it isn’t exactly profound.
More of a computer-generated hallucination than a real continuation of my story
What I loved the most was the fantastical nature of the result. Sure, this is all machine-generated, but there’s some inherently eerie in how it strings together the different elements of the text: “A grove of walnut trees and brambles, surrounded by the lush green nature o lower Svaneti”. It’s more of a computer-generated hallucination than a real continuation of my story.1 It’s creative, for sure, what what does creativity even mean when there’s an algorithm underlying it? And what is it for?
The article argued that AI wasn’t necessarily something to replace us2 but rather an aid to come up with more creative ideas than we could on our own.
In that spirit, I ran Sudowrite on this text to have it generate the following ending. For now, I think I need to take some time and think about what it all means:
Artificial creativity is already a bit of a blessing and a curse. It’s a creative act, but only as good as its programmers. And as we’ve seen with the Twitter bot @horse_ebooks, the results can be disastrous. The AI’s ability to write about emotions and remember certain details—in this case, places where I’ve been—is fascinating and unsettling. And it’s amusing that it’s able to identify itself with me on any level. I’ll continue to explore this tool, as I’m sure there’s more to discover yet.
Half a year after publishing this post, ChatGPT had blown up in the public imagination, and the term “hallucination” became commonplace to describe the strange—but interesting inaccuracies. Finally, Ted Chiang’s essay ‘ChatGPT Is a Blurry JPEG of the Web’ explained it: “(Chiang’s) analogy to lossy compression is not just a way to understand ChatGPT’s facility at repackaging information found on the Web by using different words. It’s also a way to understand the “hallucinations,” or nonsensical answers to factual questions, to which large language models such as ChatGPT are all too prone. These hallucinations are compression artifacts, but—(…)—they are plausible enough that identifying them requires comparing them against the originals, which in this case means either the Web or our own knowledge of the world. When we think about them this way, such hallucinations are anything but surprising; if a compression algorithm is designed to reconstruct text after ninety-nine per cent of the original has been discarded, we should expect that significant portions of what it generates will be entirely fabricated.”↩︎
“Daniel believes that good designers shouldn’t feel threatened by automation. ‘I’m convinced that AI is more than capable of designing the world’s most legible typeface, but it is still not capable of curating something objectively wrong as subjectively good (…)’”↩︎