AI Poetry Hits the Road

Photos by Christiana Caro

I’ve just returned from Ross Goodwin’s AI-assisted stab at the American literary road trip, a project called Wordcar that put AI on the highway to generate 200,000 words of machine poetry. It’s a classic trope with a 21st century twist. But in our moment of tender and anxious global ecological crisis, the free-wheeling ride into the unknown mythologized by Jack Kerouac, Ken Kesey, and Hunter S. Thompson takes on a sinister shade. Those authors set out in search of freedom, masculinity, enlightenment, hedonism — 20th century values currently under renovation. These days, hitting the road in a gas-guzzler in search of anything other than a job feels irresponsible or at least unnecessary.

We are where we are. Many aspects of life and the open road have been inexorably transformed by the cannibalistic junkspace [1] of techno-capitalism. The mutation currently on display comes from the revived field of Artificial Intelligence. Because of breakthroughs in neural-net architecture and GPU vector processing, what is called Deep Learning has taken center stage in the field of AI, which increasingly goes by the less-narratively-burdened handle Machine Learning. The through line from Kerouac to cutting-edge RNN-LSTMs (Recurrent Neural Net Long Short Term Memory) starts with an amphetamine Beat and dips into self-absorbed spiritual utopianism and Gonzo paranoia before it settles in the Bay Area, where dropouts, acid-heads and home-brew hackers laid compost for the home computing revolution [2] and by extension key components in the techno-capitalist Stack [3].

Output, or poetry, from Ross Goodwin’s RNN-LSTM

It’s an essential part of our era’s hypocritical hypocrisy [4] that we question the ethics of any given act of consumption vis-a-vis the ecological and extinction crises while still consuming. This impotent self-awareness coated my perception as we set out from Bushwick, Brooklyn in a rented Dodge minivan and Cadillac sedan. As we pulled up behind a vintage Ford, I said to Ross and his sister Beth,”It must have been nice to be a Boomer. Cars were weird. Gas was cheap. You didn’t have to feel guilty.” Our engine idled as I cast about for someone else to blame.

What besides the knowingness of our hypocrisy distinguished this trip from the cluster of mid-century journeys historicized by white guys from the west coast? I was the resident person of color and our party had gender parity — incremental progress, perhaps. Yet all of us hailed from the coastal, generally liberal, urban centers where tech thrives, and the left coast maintained strong representation. Photographer Christiana Caro and I work for Google; Beth, Ross, and I all grew up in the SF Bay Area. Tech was at the center of the journey, as a synchronic key, as the “literary” engine, and as Ross put it, a substitute for mind-altering substances.

The eye of Wordcar was an Axis M3007 surveillance camera.

As with many aestheticized adventures in our era of cultural recycling, it was through tech that we marked our contemporaneity. Ross works with generative systems that produce text, specifically AIs that write poetry. My team at Google (Artists + Machine Intelligence) is a band of passionate twenty-percenters that have aided Ross with technical advice, financial support, and professional development as part of our mission to support an emerging form of art made with AI.

Our automotive AI assemblage was inspired by an absurdist art exercise: write with a car. When we’ve talked about writing, Ross has mentioned David Foster Wallace, Jorge Luis Borges, Ursula K LeGuin. On this trip he cited On The Road, The Electric Kool-Aid Acid Test, and Fear and Loathing in Las Vegas as influences and Beth recounted his early enthusiasm for these books. I found it hard to imagine him synthesizing these influences in his previous career as a speechwriter for Barack Obama, Timothy Geithner, and John Kerry. Political speech is way too constrained for Ross. Our drive from NYC to NOLA was a better channel for his automated graphomania.

Ross Goodwin

These literary precedents all couple the road with one or more psychoactive substances. We rambled through New Jersey, Delaware, Maryland, without a suitcase of meth or a punch bowl of LSD, but we did have a neural net and a surveillance camera, and the babbling of Wordcar’s simulated brain was an uncanny approximation of the stimulated screeds of yesterday’s eschatologists. I won’t over-promise — it was more Dada than Brautigan, and that may be the state of the art, for now.

A GPS understanding of place

While the initiating impulse came from the written word, it was through the image that the word became. The eye was an Axis M3007 surveillance camera mounted magnetically to the trunk of the Cadillac. This is the standard model for home or business surveillance, a favorite tool of casino pit bosses who use them to see in four directions at once. Its industrial design is neutral in the loaded way utilitarian objects express blankness: an off-white square frames a transparent bulb, which wraps a black robotic eye; the person on the other end of the signal is camouflaged by the normest of cores.

Ross customized his M3007 to rotate and “look around” by feeding its orientation controls Perlin noise. Ross’s script instructed the camera to capture an image every twenty seconds. This image was first textualized in a most literal way: as ASCII art rendering a grayscale image with characters. Then an image recognition net described the image in a sentence, which fed a free-associating, text-generating neural net (in mathematical terms, a ~36,000 dimensional model of the linguistic space of nearly 200 hand-picked books, prodded to produce a string of statistically likely characters following the initial description).

Beth Goodwin in Goldsboro, NC, swinging on sprawling grounds that were once a plantation, now traversed by an interstate freeway.

What did the neural net see? What did it talk about? It talked about what it knew. It knew the time. It knew where it was (in the way any computer does, via GPS) and it knew what was around it. To avoid anthropomorphizing, I’ll be specific: it knew locations and businesses (like the Biloxi Hard Rock Cafe) that were proximate because they were exposed by the Foursquare API, which is to say by the priorities of the techno-capitalist producer and subject. These locations were often gas stations and fast-food restaurants. There are many on the American road trip.

Ross’s sister Beth is a food writer and at times the conversation turned to food deserts and the business structures that keep them in place. Distribution networks owned by fast-food conglomerates have an edge on small businesses that can’t afford to send fresh produce out to exurban or rural areas; the roadside stand has come a long way since its first documented incarnation in upstate NY. The output of this network is the de facto diet of developed-world poverty. Foursquare in the Lower East Side, the Mission, or Silver Lake paints a very different picture. But this patchwork of chain convenience stores and fast food franchises was what Ross’s Wordcar showed us. To be fair, these weren’t the only features of the landscape to surface. There were bridges, rivers, and parks. From the perspective of the AI at the heart of Wordcar, however, they formed a substrate seen incidentally through accrued layers of gas and synthetic food distribution.

Valero, Pizza Hut, Waffle House, The Jameson Inn, the author

Why did Ross choose to show us this slice of American life in semi-sensical LSTM poetry gathered via API and surveillance camera? We weren’t just roaming the concrete corridor connecting Yankees and Southerners. We had an objective: a stop in Biloxi, Mississippi to meet Josh Sniffen of Not From Concentrate Systems, a brilliant fabricator of gaming PCs, who embedded GPUs in vintage 8x10 cameras for Ross’s upcoming show at the Rubber Factory in Manhattan. We saw a Jeep he constructed “from scratch”, his YouTube broadcasting setup, gaming PCs he’d built, Ross’s 8x10 and 11x14 cameras (from 1890 and 1905 respectively.) The work he did was beautiful.

Josh invited us into his home and grilled delicious sous-vide steaks for everyone. Where Ross’s Bushwick living room drips with receipt-scrolls of AI poetry and runs a Google screensaver on Ubuntu Linux, Josh’s home decor includes a posted list of of family rules and an informal garage shrine to the Virgin Mary. While our film crew captured footage of Josh’s studio, I waited in a lawn chair. Josh’s children rode tricycles. Humid air came off the bayou.

Sous-vide steak and a hand-assembled Jeep.

Ross and Josh had never met in person but they got along, diving into the obsessive tech-speak that engineers and hardware hackers fling. They both love manufactured systems. They both have complex relationships with mainstream American culture and religion. They come from different ends of various axes: North/South, hardware/software, a lapsed Jew and a devout Catholic. On a global scale these differences are minimal. But in contemporary American political discourse they are often framed as insurmountable. The intensity of their shared interest brought Josh and Ross close enough to experience each other’s difference. Their meeting wasn’t a site of ideological conflict (like, say, Twitter). But it was clear that the cultural space between them wasn’t simple, that traversing it would take time, and that it ultimately wasn’t necessary in order to have a productive relationship around an art and fabrication project.

The Virgin Mary

We ended the trip in New Orleans. By then we’d been through ten mostly Southern states. As in many areas of the US, our route was dotted with industrial infrastructure unused and in decay. There were power plants, factories, railroads, mines. These were scenic and the filmmakers we traveled with turned their lenses on the ruins as backdrops. They hoped to highlight one of the most pressing concerns around AI: the changes that automation will bring to the economy and the predicted loss of jobs on a massive scale. Automation has already transformed mining and manufacturing. But AI that can predict, AI that can diagnose, AI that can write… these threaten blue and white collar jobs equally. As it often does, automation speculation led to discussion of Universal Basic Income, the idea that the state should provide for every citizen’s basic needs. Under neoliberal (or neofascist) techno-capitalism this is unthinkable. But it wasn’t so long ago that jobs were created by the Works Progress Administration, during another time of economic instability.

Wordcar

In fact, the morning after our first day on the road, we learned about a document created by WPA laborers (writers and historians, or what we might now call creatives and content producers). As we ate breakfast in a 10,000+ square foot mansion in Goldsboro, North Carolina, Laurie Sneed (the aunt of Ross’s fiancee Lily) shared with us a collapsing edition of The American Guide Series a written history of places traversed and annexed by interstates. Think of it as an archaic, proto-GPS-indexed feed of quirky and boring stories about small towns dotting highways in the 1930s. It’s the sort of entertainment that might strike us as quaint or musty. But in our ambient Anthropocene anxiety it’s almost soothing to read this excerpt describing the roadside grave of a circus clown, paved over even 80 years ago:
 
 At 25m., embedded in the cement pavement of the highway, is a Tombstone ® broken during the War between the States by the wheels of a gun carriage. Inscribed “Gone But Not Forgotten,” it marks the grave of a circus clown who died near here in the 1840s.
 
The clown and the old book beg questions: Who will look back on the half-absurd techno-engagements of Ross Goodwin and his ilk in 10, 50 or 100 years? How will their basic needs be met? How will the mechanisms that meet them frame their understanding of the Wordcar project or any literary road trip? Are we crude psychonauts prefiguring mainstream mind-manufacture? Are we hypocritical hypocrites on a dirty freeway? Are we everyday artists like the people to come? How are we etching our names in the land?

  1. Rem Koolhaas, Junkspace, 2002, http://www.jstor.org/stable/779098?seq=1#page_scan_tab_contents, The MIT Press
  2. Fred Turner, The Democratic Surround, 2013, The University of Chicago Press
  3. Benjamin Bratton, The Stack, 2016, The MIT Press
  4. Timothy Morton, Hyperobjects, 2013, University of Minnesota Press

(Thanks to Jac de Haan and Christiana Caro)