Editor’s note: This story is an edition of Link Rot, a column by Shanti Escalante-De Mattei that explores the intersections of art, technology, and the internet. The column will appear every other Monday.
The development of generative AI text-to-video models has been charted over the past few years through a simple prompt: “Will Smith eating pasta.”
The first videos, posted in 2023 using newly released Stable Diffusion models, feature a ravenous, monstrous rendering of the actor shoveling a pasta-shaped blob into his mouth with his hands, only for it to slide, morph, stretch, merge with surrounding artifacts, or simply disappear. Within three years, the test has nearly reached its endpoint, with the pasta responding naturally to gravity and interacting correctly with the other objects in the scene: the fork, the plate, and Smith’s mouth. But as each new model arrives in relentless succession, what happens to those that came before it, or to the videos produced during their training?
That question drives “Post Generation,” a new exhibition at tiat gallery in San Francisco. Three artists reflect on a video archive generated by Veo 1, a now-defunct model developed by Google in 2024. Released only for internal use, though eventually leaked, the model seemed barely functional. It could produce only six-second videos, with long waits between generations and predictably surreal results: a six-legged horse galloping across a wet field, a staticky smudge that might be children playing in a pile of leaves, and a three-limbed hominid swimming in a pool. Researchers generated hundreds of thousands of videos before the model was shut down and superseded by successive generations, Veo 2 and Veo 3, both of which were released to the public.
“Because these models are so resource intensive and because each new model is faster and more efficient, it’s hard to justify keeping old models around. No one can generate using Veo1, and the outputs it produced aren’t publicly available,” said Tina Tarighian, a senior creative technologist at Google’s Creative Lab and the show’s organizer. “So I thought, what can we do with these leftover scraps of a model? Can we gift this archive to artists?”
Tarighian shared a portion of the archive with choreographer Zelia ZZ Tan, creative engineer Karyn Nakamura, and fiber artist Haley Mortin, asking each to respond with new work. The resulting exhibition is a multimedia meditation on the nature of this technology and an archive of a particular moment in its development.
For Mortin, working with this training data was a natural extension of her practice, which incorporates AI-generated images, images sourced from research papers, and other AI artifacts into knit works. Her fascination with these artifacts stems in part from a former job as a data annotator, a form of often-invisible human labor in which workers around the world categorize and tag images to teach AI models to “see” and correctly identify objects.
“It was 2018, I had just graduated college, I didn’t know what AI was, and I needed money,” said Mortin. “I thought it was just a data entry job.”
Over the next two years, Mortin found herself solving CAPTCHA tests and labeling images not only by the objects and subjects they contained but also by more subjective categories, such as the emotion she thought a person in a photo was expressing. The Veo 1 outputs are, albeit indirectly, a product of this kind of labor. In examining them, Mortin decided to focus on the strange forms of text the model generated.
“It produced two flavors of nonsense, these glyph-like shapes or empty phrases,” said Mortin. “These [visual] models don’t learn text like an LLM [large language model] does, rather, they learn text as a texture.”
This led Mortin to consider the embroidery samplers made by women in the 18th century. Many young women first encountered writing not through ink and paper but through embroidery. As they practiced stitching text, they often used these samplers to record family histories. Mortin’s resulting work, Soup Is Made of Water (10 Samplers) (2026), comprises ten knitted pieces that capture the different forms of garbled text generated by Veo 1, reflecting on the histories of text production and the practices through which it is preserved and archived.
Rendered in black and white yarn, the pieces depict a looping scribble, a calligraphic font attempting to spell, perhaps, “beauty” [b eurt], and what can only be described as a sigil tattoo of the kind commonly inked on thighs in Ridgewood, Brooklyn. The work deftly brings into conversation, and inverts, different practices of language learning and history-making: those of 18th-century women, contemporary machines, and artists attempting to process and document a machine’s training.
Karyn Nakamura also took up Tarighian’s invitation to explore the research histories behind Veo 1 and other generative models, combining several video technologies in a single work. In Labor of Light (2016), an assemblage of technologies comes together as a kind of altar to the history of video production. Suspended from a pair of cables, a white device emits a beam of light that appears stable to the viewer but contains, in its flicker, the analog signal of Quicker ’n a Wink (1940), a film explaining how MIT’s Harold Edgerton developed the high-speed photography technology used to produce slow-motion footage.
The beam of light leads down to a projector that casts a collage of Veo 1-generated videos onto a white cloth, each clip emerging from a static-like field of vivid colors. Nakamura uses RGB noise, a key component of generative AI, in which images and videos are generated from an initial field of noise. Beneath the cloth, small toothpicks held in place with putty support printouts of landmark research papers that advanced the techniques behind modern AI generation, including diffusion.
“I find the process of image making and the process of video and video systems incredibly beautiful, there’s a lot of craft in there,” said Nakamura.
Even as someone who sees beauty in these technologies, Nakamura worries that the pace of development makes their inner workings increasingly difficult to understand, limiting people’s ability to intervene creatively and exercise agency in contemporary image-making processes. For this reason, her work often dwells in the glitches and gaps of earlier models, such as Stable Diffusion 1.5.
“Because I like Stable Diffusion 1.5, sometimes I think, ‘Oh, am I being a Luddite?’” Nakamura mused. “Like, what is even progress right now?”
Against a backdrop of increasingly morbid present and future outcomes, these artists reckon with a technology advancing at breakneck speed despite being barely a few years old. Yet in a culture that equates progress with perpetual acceleration, to pause is to risk obsolescence. Choreographer Zelia ZZ Tan explores this tension in a ten-minute dance performed at the exhibition’s opening on October 8. Combining choreography with animated avatars, the performance captures the ways bodies move and gesture in Veo 1 videos.
“I searched the archive through movement rather than only through words when I started to realize that between generated bodies, physical bodies, and the computational body, there are a lot of similarities and also a lot of rules based on how they can perform,” said Tan. But dancing with the limitation of movement of the generated body required vulnerability.
Tan continued: “It’s such an honor to have a body, to be able to be present there, and maybe I’m not showing the most trained technique, but to listen, to be hesitant on stage, to wait for something to happen – that’s part of the power of being there as a performer.”