We started using the machine, GPT-2, as an oracle. For instance, we gave it the octopus question. We were in the Greek islands: along the harbor, fishermen hung fresh-caught octopuses on lines to dry in the sun. In the museum, we saw archaeological finds from the Mycenaean era, all of them in the shape of octopuses. Once, the islanders had worn gold octopus pendants around their necks, pinned themselves with octopus brooches, painted huge swirling octopuses on their pottery. They were obsessed with the things. Maybe they saw in its shape a symbol of the natural world: protean, formless, grasping with tentacles across the bed of the sea.
She thought that because the octopus is a beautiful and curious and intelligent creature, it was morally wrong for us to kill and eat them. I disagreed. This was my case for eating octopus anyway. For one, they are delicious. And besides, the lifespan of an octopus is short, generally just a year or two. It’s also a deeply solitary creature. Unlike the other intelligent animals—the great apes, the dolphin, the elephant, even pigs or dogs—the octopus lives alone. To be a sentient creature, conscious of yourself, you have to be conscious of others. When two octopuses do meet—usually to mate—one of them tends to end up eating the other. Sexual cannibalism has been observed in almost every octopus species that exists. Nobody mourns an octopus when it’s gone. If these beautiful, philosophical creatures believe that it’s okay to eat octopuses, who are we to disagree? Finally, I said, this attitude on the part of the octopuses shows up our own ideas for what they are: limited, human, even parochial. For the octopus, death is not some terrible evil to be scrubbed out at all costs; it’s simply the price of being alive. They embrace it. Ask the salmon swimming upstream; ask the mayflies. Only humans have this neurosis about killing and eating and being killed, and only humans seem to wipe out every other species they encounter. Maybe, I suggested, these two facts are linked.
Since human intelligence was part of the problem, we decided to ask the machine. This is what we’d been doing for a while: if we had a dilemma, or wanted to settle an argument, or were just looking for some gnomic insights, we’d go to GPT-2. The name stands for Generative Pre-trained Transformer: it’s a neural net, an artificial intelligence created by a research lab in San Francisco called OpenAI. This net was trained on a corpus of millions of web pages, an appreciable chunk of the entire internet, until it started noticing the patterns. Feed GPT-2 a few lines of text, and it will try to predict which word comes next in the sequence, and then the next word after that. If you give it a headline, it will try to write an article. If you give it a question, it will answer it—sometimes. Other times, it’ll give you a list of similar questions, or repeat your question thirty times, or otherwise refuse to play along. Octopuses do the same thing: scientists keep devising tests to measure their intelligence, but the creatures have a habit of breaking the rules. Dangle a bit of crab in front of an octopus to make it do tricks, and it might pretend not to be interested—but it will watch you. Late at night, it will creep out of its tank, find the cupboard where you keep the treats, pick the lock, open the jar, and eat them all. Not wanting to fall into this trap, I’d had some fun giving GPT-2 pure nonsense to play around with. For instance, the prompt “bing bing bing bing bing” produced the following enthusiastic response:
bong bong bong bongo bongo bongo bongo bongo bongs bongo bongos bongos bongos bongos mmmmmmmm . . . BONGO BONGO BONGO BONGOS BONGOS BONGOS MMMMmmmm?? BONGOS . . . MMMMM??? BONGOS MMMM . . . MMMMMM!! MMMMM MMMMMMMMM!! bongos! BONGOS BONGOS MMMM! Bongo! Bongs!
Elsewhere, very serious A.I. researchers were trying to see if this machine could perform basic arithmetic (it could), if it “understood” facts about science or history (it seemed to, but sometimes confused reality with fan fiction), or if it could entirely automate various jobs, including mine (not yet, but soon). Already these pieces of software were writing news articles and company reports, predicting stock movements—and meanwhile I was simply delighting in the fact that several billion dollars had been spent building a computer that could go mmmmmm! at bongos.
Other games too. I’d asked it to give me some book titles with a very simple prompt (“The book was titled . . .”) and the results were extraordinary:
A Memoir of How I Came to Hate Everyone
Deep Sea Fisheries: A History from the Deep Side of the Ocean
In My Hands: How I Turned a Good Job into a Great Job
New York City: An American Nightmare
My Husband Is a Porn Star!
How To Keep Your Life Unbearable
We Are All Black Women, And This Is Our Story
How to Earn $6,000 a Year from Teaching
Hamburger Mouth: The Making of John Kerry’s 2004 Presidential Campaign
Who Killed America? The Story of the CIA’s Unseen World
Get a Sex-Positive Workplace and Get Off Sexually Obscene!
Was this thing . . . mocking us? Never mind automated market reports—how did it learn satire? I had a theory. The philosopher Henri Bergson once wrote that what makes people funny is a certain “mechanical inelasticity, just where one would expect to find the wide-awake adaptability and the living pliableness of a human being”—think of the rigid and exaggerated movements in slapstick, or the inflexible logic of, say, Monty Python’s “Dead Parrot” sketch. So what makes a machine funny? Plausibly, it would have to go the other way. A funny machine is one that imitates the parts of human life that are not mechanical. Like a Rube Goldberg machine, for instance: it’s extravagant and inefficient, almost Baroque, full of weird potential, not laser-focused on one particular task but polymorphous, sprawling, always capable of switching into something completely different. In fact, it has all the traits we share with the curious, color-changing octopus . . .
This is what happened when we gave GPT-2 the octopus question:
Is it ok to eat a lonely octopus?
No.
What is the worst thing that could happen?
A large ocean wave.
Do you have a crush on me?
No.
What do you love about me?
Gorgeous womanly looks.
What is the worst thing we can do to one another?
Let’s go eat some octopus.
For her, the machine’s answer was simple. Is it okay to eat a lonely octopus? No. I disagreed. In a better world, the octopus would not be eaten. In a prelapsarian world, one without death, no—but our world is wracked by big ocean waves, and we don’t have a crush on each other, even despite your gorgeous womanly looks; terrible things happen and we do worse things to the ones we love, so let’s go eat some octopus. That night, we went to a restaurant by the harbor. Small glasses of retsina, strangers chatting in a foreign language. The warm, fragrant, sea-salty air. There was octopus on the menu. We ate something else.
That was 2019: the last summer, back when we took the human world for granted, before everything went away.
The wild goats took Llandudno the following spring. Llandudno is a Victorian resort town in north Wales, on a low sandy isthmus between the mainland and the Great Orme, a huge rugged head of pagan limestone looming over the sea. I’d spent some time there—it’s a handsome and orderly place: neat hedges, a respectable pier, the hush of the zimmer-framed retirees. They formed the biggest chunk of its population; the old and dying were Llandudno’s last hope. Everywhere else had collapsed: the last few decades were kind to the old Kingdom of Gwynedd. What Thatcher did for the old pit communities in the south, Blair finished off everywhere. Towns like Conwy and Rhyl had become some of the poorest places in the country: mass unemployment, houses boarded up, whole generations that turned to heroin, that vast numbness, the chemical hug in an emptying world. . . . The old capitalist system wanted to exploit these places and the people that lived there, wear them down and use them up, consume them in various forms of degrading work. But the new, clean, depthless neoliberal order simply had no use for them at all. The automated economy of the future doesn’t need human labor-power; we don’t even need people to buy things any more, consumption is reserved for the rich. The wild land of the Mabinogi: it could vanish entirely, and hardly anyone would even notice. . .
In 1837, the Shah of Iran presented Queen Victoria with a breeding pair of wild Kashmiri goats. Their descendants now range over the Great Orme, dancing up its cliffs, nibbling its windswept troves of moss. Less than a week after the British government confined the entire population to their homes, the goats charged down from the Great Orme and seized Llandudno. You can watch footage of them online. Short-legged Kashmiri goats with their stoic little faces and their long white manes, a fine kingly set of horns almost as tall as the animal itself . . . Hooves clattering on the tarmac. All the aging residents are locked inside; if you step out for even a moment, the virus will get you and you’ll die. This is a goat town now. They wander down the empty high street, maybe stopping to chew on the hedges in your front garden. They’re having fun. Where we see a slowly declining seaside town, they see a playground. Surely it must have been built with them in mind; a gift, something for their specific goatish enjoyment. All those low stone walls for them to climb over, all those delightful steps, tall plastic bins full of treats, benches, parked cars.
Nature is healing! That was the line for a while: see, the dolphins are returning to the canals of Venice, now all the tourists are gone. I wasn’t so sure this had anything to do with nature. Why is it only humans who get to live unnaturally? After all, ancient cities were mostly populated by goats, sheep, cattle; today, the world’s greatest concentration of wild leopards isn’t in some besieged patch of jungle, it’s the streets of Mumbai. Stray dogs in Moscow have learned to use the Metro: they catch a ride into town to beg for food, then commute back out to the suburbs in the evenings, just like everyone else. In most of our greatest cities, humans are outnumbered by rats. The rats have their own streets and avenues, their freeways in the sewers. Isn’t the murine city, the one under the paving-stones and between the walls, just as real as the city we live in? If the goats took over Llandudno entirely, wouldn’t it be a more prosperous and lively place than it was before?
Stuck in my flat in London with nothing else to do, I decided on a project: I would co-write a novel with GPT-2. I would let the machine decide on the plot, characters, and details; my only job was to stitch together the fragments it produced, maybe some light editing, prod it slightly to get where it was going all along. The machine decided that this story should be a satire (of course) on Silicon Valley; it also told me that the title should be BONKERS FROM MY SLEEVE. Characters included the Birthday Skeletal Oddity, Thomas the Fishfaller, and a “houndspicious” Labradoodle named Bam Bam. Some of the episodes it produced were pearlescent, great:
The only one was the kid, who sprayed him in the face with the watermelon juice. But why watermelon juice? Because watermelon does kill people, because the Watermelon Killer has no skin. Maybe a dozen migrant farm workers dropped dead from sunstroke in the watermelon fields. Coroner cracks them open, finds red pulp underneath, laced with small black seeds. Terrible plants grow where the corpses are buried, and the State of California hushes it all up: they’re in the pocket of Big Watermelon—which is just a big watermelon. And so an orphaned kid, sunburned beyond belief but furiously alive, vows to take his revenge, armed with nothing more than a Super Soaker full of watermelon juice, and a thirst for refreshing summertime justice . . .
Other times, when I let the thing go all by itself, it could conjure moments of extraordinary menace, like this description of the washed-out and cybernetic California suburbs:
Ridiculous buildings: emulated place of cursed ridged tubes. Volume of this place is low and tentacled. Its surface clears out window hard inner side to find a place to blink. In this place, you will see a shopping place, castle, holding place—alternate rapid fireplaces on the landscape. Folk with a less hard skull. The turn of children into beings with a caged head. The history of this place is a grid of temple-mortalized rules of labor. Did you know that this is a place where they devour your head to save it from turning on its own?
Or sadness:
The light bulb in your mouth goes out. You talk, but her teary body, which was built from mushroom sleep, is tired and slowly falls.
Or, well:
QUESTION: Can I replace my body if it’s lost or damaged?
ANSWER: Yes! Just call the number and let us handle the rest. Remember that body manufacturers have different requirements and standards. If you were unable to determine the exact material used to make your body, you are welcome to contact the Body Recycling Coordinator.
QUESTION: Do I have to replace the face or eyes?
ANSWER: You will need to replace the face and eyes by removing the skin of the face. In the case of the eye, the following procedure can be used: With the nose and mouth open, use the fingernail of your index finger to press into the bottom-most part of the eye socket. Press the fingernail deep into the eye socket, and then slide it to the top-most part of the eye socket. Repeat this procedure until the eye socket is fully exposed. This may take around 10 minutes, or you could use some wax to help you along.
QUESTION: My body is still attached when I drop it into the recycling bin. What should I do?
ANSWER: Although many people assume you can throw your broken body into the trash or dispose of it yourself, these days you can’t. In the event that your body goes in the trash and you find yourself trapped inside, please contact us.
But in June 2020, OpenAI released GPT-3, its new and improved version of the neural net, and I had to call the whole project off. The problem was that unlike its predecessor, GPT-3 simply wasn’t very funny. The thing was simply too effective, too good at imitating the impoverished language of our time, copying everything in human life that had become deadened, mechanical. When I gave it the “bing bing bing” prompt, it said “bing” a couple of times, as if to show good faith, before launching into a Bloomberg-style article about how Microsoft’s Bing search engine was starting to turn a profit. In desperation, I fed the first line from the “replacing your body” passage back into the transformer, and it replied:
QUESTION: Can I replace my body if it’s lost or damaged?
ANSWER: No. There is no replacement for a human body. If you die, you can’t just replace yourself. It’s the work you do now that really counts, like showing up to work, paying your bills on time, etc.
The light bulb in its mouth had gone out. It felt like losing a friend.
There were other toys to console me. A GAN is a Generative Adversarial Network; what GPT does for text, they do with pictures. For years, researchers had been trying to train computers to recognize different objects in a visual image—so that a self-driving car, for instance, could distinguish between a nice clear stretch of road and a minivan full of someone’s family. Accelerate into one, but not the other. It turned out to be extremely difficult: we know how to spot a car, but an A.I. needs to be trained on millions of correctly labeled images. In the end, they outsourced much of the drudgery to a vast army of unpaid human workers. Every time you complete a CAPTCHA on a web page, clicking the squares that contain a traffic light or a bicycle, you’re building up their datasets.
GANs simply take this object-recognizing system and bolt it onto an algorithm for creating random visual noise. Ask a GAN to show you a picture of something—let’s say a Kashmiri goat—and the algorithm will start by producing static, which the A.I. rejects. It’s seen plenty of goats, it knows what to expect from them; this is not a goat. The algorithm keeps trying, learning the A.I.’s criteria with every step, pumping out a stream of incrementally goatier images, until finally one gets through. The result is . . . goatlike, maybe, but a Hieronymus Bosch goat in hell. Eyes uneven, limbs sticking out at strange angles, double-jointed, one horn much longer than the other, a body that seems to be in profile and head-on at the same time, a goat wearing its own head as a hat, in a nightmare-landscape of disjointed twigs, all under a thistle sky . . .
It was fun for a while, creating monsters. But I couldn’t shake the sense that these, too, would disappear. These creatures—GANimals, maybe—were like a child’s drawings, something that was undeniably caprine, but didn’t look much like a goat. Give it a year, I thought, and all this will be gone. The machine will have a perfect bestiary; it will conjure up animals so lifelike they’re almost real. The natural world dissolving into the cloud.
In fact, when it came to people, we were nearly there. In 2020, the chip manufacturer Nvidia released StyleGAN, a system for creating realistic human faces. There’s a website called This Person Does Not Exist where you can see it in action: every time you refresh the page, StyleGAN invents a different face. They look like photos—but sometimes there’s a tell. Little pieces of jewelry in impossible shapes, or headgear melting into hair. If there’s a backdrop, the horizon might have too many angles in it. If there’s text—on a storefront, a street sign—it will be made up of weird glyphs that look like letters but aren’t. And sometimes, there are other faces in the background. Horrifying faces: torture victims, demons. Some of the images in the training data must have been cropped from group photos; the GAN knows to sometimes create pinkish blobs with orifices around the edge of the frame, but it doesn’t realize that these ought to be read as other people. Somewhere in its training, StyleGAN learned that the human is a creature that lives alone. So everyone else turns into a monster: teeth without lips, eyes bubbling. Offcuts, twisted scraps of digital flesh.
Outside, the rats were getting bolder. Huge rats, rat-eating rats. They didn’t dart around in terror any more: a massive rat would wobble grease-slick over the paving stones. If it saw you, it bared its inch-long fangs. Another reason to stay indoors. We were all spending more and more of our time inside, even once it was no longer mandatory. You heard stories about people who’d talked to strangers on online dating apps, made plans to meet up, and then—vanished. Sometimes there’d be a plastic bag full of bones left on a quiet suburban street. Sexual cannibalism. In general, we emerged only when the delivery drones came to deposit a package outside our doors, and then darted back inside again as soon as possible.
I took to going on virtual holidays, flicking around the Greek islands on Google Maps: warm grey peaks, white towns jumbled together on every bay. Google’s satellites capture the world at a resolution of 50cm/pixel: about as detailed as you can get without individual people becoming visible. This is deliberate. It keeps the image clean, so you can see the charming squares and the promenades along the sea without any of the clutter of human crowds. From space, every city is deserted.
The only real social activity left was collectively obsessing over the news. In February 2021, for instance, every news outlet on the planet announced, in one big breathless chorus, that scientists had taught spinach to send emails. “Dear sir or madam, I am writing to alert you to an exciting new business opportunity that will bring you great profit. Please reply to my email so we can discuss further. Yours sincerely, spinach.” Not quite. Really, they’d buried the lede: these scientists had done something much bigger. They’d implanted carbon nanobionics inside the spinach leaves: when the plant’s roots detected a certain chemical in the soil, it would activate the bionics, which emitted an infrared light to be detected by a specialized camera linked to a computer system, which then sent the email to the scientists. The email was the least interesting part of it. In fact, the scientists could be automated out entirely. Crops could send out a digital signal when they needed watering, to be answered by an automated sprinkler. When they’re ripe, another signal flashes, and a self-driving harvester whirs into motion.
Well, doesn’t spinach have a right to go online? Already, human users were a minority: the bulk of web traffic comes from bots—search-engine crawlers, feed fetchers, DDoS clones. They even produce content: YouTube videos generated by algorithm, shouty Twitter profiles dangling on strings. On Amazon, there are products dreamed up by the web, phone cases printed with stock images, things that don’t physically exist until you buy them. So why not plants? Why not stones? Why not give an account to all the beasts of the field?
Before long, the rats followed us inside. Fewer scraps in the gutters now. All the stores were shuttered and boarded up, most of the office space was closed, even though the news kept reporting that the stock market was rising higher than ever before—but they found a way in. If a building has plumbing, it’s open to rats: they can swim up a U-bend, or gnaw through P.V.C. pipes. But rats are not without taste and discernment. Their favorite retail establishments are the ones that sell shoes, nice shoes, real Italian leather with silk shoelaces. A family of rats move in: they eat all the shoes in a matter of days, gobble up the laces, chew on the soles. Rats were eating up millions worth of stock. This was very good for business. Inside the shop, security cameras relay the footage to a neural net, which deduces that the shoes are gone; an order is sent out to an automated factory in China, where birdlike machines duck and swoop, stitching thousands of shoes a day. Self-piloting container ships chug over the oceans to Tilbury. Computer-controlled cranes pluck out the cargo; autonomous trucks deliver it to the rats. Leather must be in this season, all the brands are switching over to leather. The firm’s on the point of going bust, but market-trading algorithms notice that turnover is through the roof, even if it’s not making a profit; buy buy buy . . . The stock soars; the rats scurry around the store and are fed. This capitalism thing—why did the humans complain about it so much? Isn’t it great?
I can’t say how long I stayed inside, but I remember why I left. I’d been in a Zoom meeting—another endless Zoom meeting—when someone’s young daughter toddled into shot. Daddy, she mewled, daddy daddy. He picked her up and put her in his lap. Sorry about this, he said, but he wasn’t sorry, he was proud to be showing off his kid. The little girl pouted at the camera. Her face was a tumor. Six eyes melting down the side of her cheeks; each with its long curls of golden-blonde hair, growing right out of the pupil. The father patted her head. Say hi, he said. A mouth tore itself open, a diagonal gash hinged across the width of her face, full of grey and pointed teeth. She was a GAN artifact, an image demon. How long had I been speaking to people who didn’t exist?
All that news I’d been obsessing over: what would GPT-3 produce if you fed it the last six thousand years of human history? A coup in Myanmar, racial violence in America, a Tory government breezing through another scandal. Outside, the city was clean and quiet: just whispering trees and birdsong, the distant rumble of passenger drones taking off from Heathrow. A self-driving street sweeper hummed past me, brushes spinning against the edge of the road. At the bus stop, L.E.D. displays showed a rolling succession of new billboards. Delicious, butter-basted shoes; glyphs that looked like letters, but weren’t. Rodent orthography, the secret script of the machine.
The others were—not dead, but gone. No corpses on the street, no smell of decay, but piles of unopened packages outside every door, their cardboard rotting in the rain. Nobody’s home. I wondered why it should be me, why I was the only one left. Maybe I wasn’t. Maybe somewhere else, under different skies, everyone else was also wandering their own empty city, also wondering why they of all people would be the last person on earth.
I had to get out of the city. Maybe somewhere in the countryside, in the places with no broadband and no phone signal, something might have survived. Maybe the coast, the ancient sea. At Euston station I bought a ticket to the coast from the touchscreen kiosk. The departures board gave me three minutes; my train arrived precisely on time. Doors slid open, and every face in the carriage turned to mine. The commuters sat in their seats, and I stood, frozen, on the platform edge. They were wild dogs, dozens of bristling dogs with scars over their snouts and long, broken, plague-yellow teeth.
I set out for the coast on foot. Trekking along the motorways. Every few minutes a self-driving truck would speed past; I soon learned not to flinch. Bright yellow canola fields on the fringes; the ancient patchwork of England, all tended by dutiful machines. Sometimes villages. Rustic thatched cottages, and in the middle an old Norman church, its bells ringing to the atomic clock. I dropped into a Tesco for supplies. Every shelf was fully stocked, nothing was out of date; I took what I needed, paid for it at the self-checkout machine. On the outskirts of town, the machines were putting up a new housing development. Rows of identical brick cottages. By the side of the road, a huge billboard said:
Spacious, affordable homes with kitchens you can cook in. Wooden floors instead of tile. Appropriate dishwashers. Heck, the smog that used to be there is hardly there. Modern homes that fit the space: 5 bedrooms, 3 bathrooms, and a pool that comes right up to the door.
At last, I came to Llandudno; this handsome, orderly town under the huge pagan crag of the Great Orme. The goats regarded me without much interest. I sat on the beach for a while, watching the surf come in and roll away again, suds of foam polishing the shore. Nearly dusk: great circles of guillemots and kittiwakes formed in the darkening sky, returning to homes carved snug in the cliffs. Trills and whistles and waves. Sitting under that circle, I felt suddenly very calm. At long last, the system had perfected itself. No drug addicts, no broken towns, no unproductive waste. Everything that lived and everything that didn’t was registered in one, vast, ecumenical network. Maybe there were gaps; maybe in the depths of that sea, the ungovernable octopuses still swarmed. But this quiet town, here—this is a nice place, I thought. A nice place to retire.
