Since humans invented art, sometime in the Paleolithic era, they’ve produced lots of pictures—“The Starry Night,” some memes, that photo of Donald Trump staring at the eclipse. What does it all add up to? A few years ago, a company called OpenAI fed a good deal of those images, along with text descriptions, into the neural network of an artificial intelligence named DALL-E. DALL-E was being trained to create original art of its own, in any style, depicting in uncanny detail almost anything desired, based on written prompts. But a mastery of the entire universe of human imagery makes for difficult choices. How do you decide what DALL-E should create? After careful deliberation, one of the first images that OpenAI prompted was a doughnut made of porcupine quills.
“There was this belief that creativity is this deeply special, only-human thing,” Sam Altman, OpenAI’s CEO, explained the other day. Maybe not so true anymore, he said. Altman, who wore a gray sweater and had tousled brown hair, was videoconferencing from the company’s headquarters, in San Francisco. DALL-E is still in a testing phase. So far, OpenAI has granted access to a select group of people—researchers, artists, developers—who have used it to produce a wide array of images: photorealistic animals, bizarre mashups, punny collages. Asked by a user to generate “a plate of various alien fruits from another planet, photograph,” DALL-E returned something kind of like rambutans. “The rest of mona lisa” is, according to DALL-E, mostly just one big cliff. Altman described DALL-E as “an extension of your own creativity.”
For the more than a million people on DALL-E‘s wait list, the only way to extend their creativity is to slide into the AI’s Instagram DMs with a request. The company launched the account, @openaidalle, in April. “I was worried that maybe he would take more of an explanation to get people engaged,” Natalie Summers, who runs the account for OpenAI, said, from a conference room near Altman. “And it did not.” @openaidalle now has almost two hundred thousand followers. It’s Summers’s job to read through the messages and choose the best of the best. “If I did every single thing that people asked, we would have a lot of raccoons and sloths,” she said. Hits have included “cheeseburger lamp,” “emotional baggage” (suitcases with sad faces), and “attractive dinosaur in a tuxedo, looking at himself in a mirror and seeing his reflection, digital art,” which DALL-E endowed with human-proportioned arms to aid in its hotness. Reviews have been of the mind-blown variety. “I’m gonna lose me job,” one commenter, whose profile said they were a graphic designer, posted, below an image of polymer-clay dragons eating pizza on a boat.
To sit through the latest requests, Summers, who wore dangling earrings and a jean jacket, videoconferenced with DALL-E‘s product manager, Joanne Jang, and with a member of the technical staff, Aditya Ramesh. Ramesh was responsible for DALL-E‘s name; it came to him in the shower. “Some people got it immediately,” he said. “Other people I had to explain that it’s a portmanteau.”
There are rules for requesters. Images of public figures are off limits, as is anything remotely offensive, including nudity and violence. Political campaigning is forbidden. “We are worried about deepfakes,” Summers said. Recently, a researcher named Boris Dayma developed a low-fi copycat, called DALL-E Mini, which went viral. Users were permitted to submit prompts like “Ice T in a glass of iced tea” and “Babies fist fighting,” though the output is sometimes eerie: Ice-T’s face appears to be melting; the babies look like zombies. At OpenAI’s request, DALL-E Mini was renamed Craiyon. (“There has been a lot of confusion,” an OpenAI spokesperson said.)
Summers began scrolling. “Here’s one that I found this morning: ‘a cat with a bed of tulips growing out of its back,'” she said. She clicked a button, and the system came up with ten images. All featured tulips, but only one had tulips out of a cat’s back—though not enough to qualify as “a bed.” Ramesh gave the prompt a try, and the machine spat out a white cat with two dozen or so tulips sprouting out of its fur. “I love how chubby the cat is,” Jang said. Summers sent the photo to the user, who responded with three smile emojis.
They turned to “an astronaut eating in a diner that is floating in space.” Jang spotted a problem with DALL-E‘s work: too much gravity.
“Oh, that’s true,” Ramesh said. “I’ll see if I can get floating food.” A few wording tweaks yielded an astronaut with a piece of toast, staring out a diner window at the stars. “It’s like he’s contemplating his life’s decisions,” Ramesh said. They went with a less melancholy option.
Next up: a fish fishing. “How about this one?” Ramesh said, pulling up an illustration of a green fish wearing a fisherman’s hat, with a smaller fish dangling from a rod.
“That one’s good because he looks alarmed to be in this meta situation,” Summers said.
The meeting was wrapping up, but they decided to take on a few more requests. One user had asked for, simply, “The Big Bang.” Jang took a deep breath. “There’s a lot of artistic license for that one,” she said. ♦