Compressed Cinema as a Study in LLM Latent Spaces
Proposal Type
Individual Talk
Location
Narratives & Worlds
Start Date
July 2026
End Date
July 2026
Abstract
In his article “Spec Acts” (2021), Matthew Kirschenbaum analyzes the AI-generated novel 1 the Road to develop his titular concept of the spec act, “the future in its multitudes collapsing into an actionable present.” With the proliferation of texts produced by generative AI and subsequent critical analyses of them, one element in particular calls for further theorization: “the future in its multitudes,” or more directly, the latent space. This echoes arguments by critics such as Antonio Somaini, who offered his own “Theory of Latent Spaces” last year. However, where Somaini’s attention is towards visual culture, I turn mine to the literary.
To do so, I follow Hannes Bajohr (2025) in attending to stylistic qualities of generative AI text to “hon[e] a technical imagination” of the technology’s inaccessible inner workings. In particular, I analyze the captions of Compressed Cinema (2023), an art book by Casey Reas whose image captions were made by generating additional text around original fragments written by the poet Allison Parrish. I combine a software studies analysis of the model with a close reading of Parrish’s captions to argue that the captions in Compressed Cinema generate narrative logic from illogical fragments. In a way, this achieves Parrish’s intention to “uncompress” the compressed cinema of the book, even while the action of “uncompression” is reoriented away from restoration and towards expansion, layering, or even reconfiguring meaning.
If it is the latent space that “determine[s] the epistemological field of these systems, what they can and what they cannot see” (Somaini, 2025)—or say—my analysis works inversely from what is said to trace the contours of such a space. Even while direct supervision of latent spaces may be impossible, this works considers what can be gained from lateral movement, looking across with feet firmly on the ground.
Compressed Cinema as a Study in LLM Latent Spaces
Narratives & Worlds
In his article “Spec Acts” (2021), Matthew Kirschenbaum analyzes the AI-generated novel 1 the Road to develop his titular concept of the spec act, “the future in its multitudes collapsing into an actionable present.” With the proliferation of texts produced by generative AI and subsequent critical analyses of them, one element in particular calls for further theorization: “the future in its multitudes,” or more directly, the latent space. This echoes arguments by critics such as Antonio Somaini, who offered his own “Theory of Latent Spaces” last year. However, where Somaini’s attention is towards visual culture, I turn mine to the literary.
To do so, I follow Hannes Bajohr (2025) in attending to stylistic qualities of generative AI text to “hon[e] a technical imagination” of the technology’s inaccessible inner workings. In particular, I analyze the captions of Compressed Cinema (2023), an art book by Casey Reas whose image captions were made by generating additional text around original fragments written by the poet Allison Parrish. I combine a software studies analysis of the model with a close reading of Parrish’s captions to argue that the captions in Compressed Cinema generate narrative logic from illogical fragments. In a way, this achieves Parrish’s intention to “uncompress” the compressed cinema of the book, even while the action of “uncompression” is reoriented away from restoration and towards expansion, layering, or even reconfiguring meaning.
If it is the latent space that “determine[s] the epistemological field of these systems, what they can and what they cannot see” (Somaini, 2025)—or say—my analysis works inversely from what is said to trace the contours of such a space. Even while direct supervision of latent spaces may be impossible, this works considers what can be gained from lateral movement, looking across with feet firmly on the ground.
Bio
Mallen Clifton is a PhD candidate in English literature at Stanford University. Their research focuses on the intersections of technology and literature, studying both printed and born-digital post-45 literature. They are particularly interested in the history of computation and contemporaneous texts that play with structure, form, and mediation. They have been published in Digital Humanities Quarterly (forthcoming), and the Berkeley Undergraduate Journal. Currently, they are a member of Stanford’s Literary Lab and are working on a project to compose and digitize a corpus of 21st century literature.