An algorithm implies a kind of opaque interiority, something unseen but felt within. It persists in the background and remains in the bones once the meat has soured. Linger too long on TikTok’s #fyp (For You Page), and one inadvertently trains a recommendation system to conjure more Borzoi or pay-to-disrupt live sleeping content.

We are resigned to the algorithm as a mostly involuntary mechanism that guides us through content—this bespoke suggestion generator feeds off a focused dataset that encourages our captive engagement. TikTok uses the metrics of “retention” and “time spent,” as recently revealed by internal memos to the New York Times (“How TikTok Reads Your Mind,” New York Times, December 5, 2021). To stay psychologically snared in transfixed enchantment is not unlike a casino chamber of slot machines without clocks or logical exits. The longer we stay, the more ad and virtual gift revenue can be generated. Extraction capitalism meets microtransaction, repeat. The algorithm is welcomed in a way—it guides us towards new and forgotten music, viral schadenfreude, and reinforced political confirmation bias. We have learned to live with it in our marrow.

But there is a larger dataset that is currently up for grabs: the entirety of the content we have contributed to the architecture of the internet. Our collective data participation has been, in the last several years, organized as a set of regimentally formatted training parameters for artificial intelligence systems. These diffusion, attention, transformer, et al. models make use of the sheer immensity of content we have all collectively scanned, created, written, sung, recorded, logged, filmed, and archived over the last decades: the aggregate of cognitive labor that shapes our Borgesian (1942) internet. Leading AI developments leveraging this include OpenAI’s multimodal Generative Pre-Trained Transformer (GPT-4), Google’s Gemini, and Meta’s 70 billion- parameter Llama2 for large language models (LLMs).

For image synthesis, we have OpenAI’s DALL-E, Midjourney, and Stability AI’s Stable Diffusion. Audio processes include Riffusion, Dance Diffusion, AudioLDM, RVC, and AudioCraft, just to name a few. Google announced Imagen, which can generate video clips from text prompts, and OpenAI stunned us with their unmatched text-to-video model Sora. An even more comprehensive list of experimental models is available on the HuggingFace, Gradio, and Google Colab platforms, which allow one to download, run, and fine-tune various checkpoint builds for local or remote implementation. As we have seen, these models have the ability to infinitely generate new discrete content based on the mathematical outcome of training on incomprehensibly large datasets. The LAION image dataset alone is nearly six billion images. GPT-4 is the largest language model at the time of this writing and is rumored to be trained on 1.76 trillion parameters.

There has been a lot of hype and reactionary discourse in various online spaces with the continued publicity surrounding many of these generative models. Particularly the image-based implementations have caused many in the commercial creative arts to feel threatened by automation. It is worth mentioning here that the recent $200,000 GoFundMe campaign to lobby against generative art was outed as being driven by large corporate copyright stakeholders (Disney, NBC, Netflix, NBA, NFL, Nike, Sony, Viacom, etc.) as an opportunity to reify aging intellectual property laws to their benefit (@fractalcounty, December 17, 2022). Individual artists and content creators are the end-goal target of copyright and Digital Millennium Copyright Act (DMCA) takedowns. The aforementioned industry professionals here become agents in the merchandising empires of semiocapitalism:

Missing from the incensed parroting against generative AI is a fundamental understanding of how models are trained, the scope of their abilities as productively antagonistic and absurdist tools, and a parochial positionality that stops short of critiquing historic structural inequities produced by neoliberal capitalism, not exclusive to AI. The learning models emerge from vast linguistic, auditory, and visual information scraped from the open internet. This includes stock photography, porn, memes, influencer content, anime fan art, microscopy, medical journals, library images, etc. No single image is important or special. The content is not attributed because it functions merely as ephemera before being translated into numerical embeddings and discarded. The training creates a metamorphosed learning model that cannot reproduce any one individual data point or image with exactitude. These phantom approximations are a historical departure from notions of direct appropriation. The wonder of holistic and collective synthesis of our shared knowledge and amalgamated cognition is excluded from the imaginations of the naysayers. Ultimately, AI propagation pierces an old wound of the myth of individual genius, originality, singularity, uniqueness, and time spent as a qualifier of value under capitalist temporalities. Open-source generative image models encourage Sinofuturist (Lek 2016) modes of proliferation, collectivity, accessibility, reproducibility, intertextuality, and spontaneity, just to name a few.

Missing from the many myopic conversations of personal victimization of artwork “theft” is the broader context of systemic weaponization of private AI technologies for use in police and military-industrial contexts (“As A.I.-Controlled Killer Drones Become Reality, Nations Debate Limits,” New York Times, November 21, 2023). From this perspective, the algorithm remains imperceptible until springing to life when one becomes a target and interfaces with systemic violence directly. Unlike the open-source models for creativity, new classification systems for hyperresolution and infinite vision play out in a fever pitch in enforcement agencies, without our knowledge or direct democratic input. Here lies the fortified world of classified power, turned on asylum seekers, foreign nationals, and citizens alike. Modes of obfuscation and illegibility embedded within the spontaneity of lumpenproletariat are the only weapons against the voracity for infinite vision, classification, adjudication, and punishment. Here AI is a new tool for necro- and biopower. Why are we not as galvanized when OpenAI quietly scrubs mission-statement language prohibiting military collaboration? (Biddle 2024)

When thinking about my contribution to the Techniques Dark Infrastructure issue, I was reminded of the idea of emergence that is inherent to the historical latent image. A latent image in photographic history is a solarized image that sits invisibly in the silver halide crystals. It is an image that has the liminal state of being both not-an-image (-x), and an-image (x) simultaneously. This parallels the “checkpoint” (.ckpt) file that is the nucleus of the Stable Diffusion image model. Like the film negative, but with an added dimension, a checkpoint file has been molded by the impressions of billions of images and, from that, can generate billions more differentiated iterations. Yet this brain contains no actual images, just the latent space of would-be outputs. There is a haunted quality of an infinite latent space that remains a conjuring potential within this structure, not unlike the specters of our photographic past.

[commendable] Diffusion models use a training method to add diffusion noise, or static, to an image in a regimented way—a step-by-step process called Markov chains. This is done until a legible image is pure noise. By undoing these calculated steps, we can return to our image from this static state. The model not only learns about this image, but it learns how to add this single reversal relative to every image it has trained on. There is technically no way to rescue any one image from the training dataset from this now impossibly complex relational latent image summoner. Here is where any claim of copyright is inadequate—our understanding of appropriation or collage from lessons in art history fails us. This is a new machine spirit world we are entering.[/commentable]

A latent image is a form of darkness. The dormant model knows only the context of the infrastructure it has been trained on. Letting the AI dream is a method of seeing invisible infrastructures. Extracting images from the latent space is making visible the space between the trained .ckpt file (which is not legible) and the opacity of the images it incorporated. The training images now exist as phantoms within an imperfect sighting machine.

Outside of publicly created and trained databases in the open-source community, let us return to the secret efforts to capture massive visual databases through inaccessible channels. Notable examples include US Customs and Border Protection using a predator drone to observe protest movements after the murder of George Floyd in Minneapolis; the same agency’s implementation of observation towers at the southern US border that use AI object recognition to track asylum seekers; precrime software development by Predpol, the largest privately held database of police body camera footage in the Axon AI Research database; and the now publicly available Blueleaks archive that was garnered by antipolice state activists.

The latter archive, on which I’ve focused extensive research, contains 270 gigabytes of previously undisclosed police data. The intricacies of the collection efforts are veiled by opaque systems, hidden from public scrutiny. I have been using this massive archive as an example of database logic under surveillance capitalism: Fusion Center data shared by a state-sanctioned system of monitoring, recording, controlling, arresting, and convicting. A majority of the data contains mediated security footage stills taken by police functioning as amateur photographers while visiting loss-prevention departments at big-box chains. Under this new technical system, the police remain the enforcement mechanism for the business interests of multinational corporations and their shareholders.

My resulting work with the archive reimagines the Fusion Center evidence files as malleable objects. Moving through morphing and abstraction, the artwork in this series questions the promise of artificial intelligence as an emergent technology to aid in the efforts of the police state. Here, photos pass through a specific image synthesis process, which attempts to reconcile the space between two proximate images. The resulting combinations move away from logical visual structures and into obfuscation—reconcealing the Blueleaks database into chimeric illegibility. The images coalesce and diverge from the promise of clarity, recognition, and accuracy used to substantiate photographic information as suitable prosecutorial evidence. The images have been “trained” but lack utility and propose collapse as a means of championing anonymity, resistance, and unknowing.

AI is a tool that augments our creativity and works as a form of prosthetic consciousness. It is an omnigenic infinity (Meerdo 2022), a space of invocation that can produce any variant in an endless set. It is a mirror and a soothsayer. When claimed by citizens, it is a new tool for resistance. Let us not be fearful, orthodox, or iconoclastic in the face of these new possibilities. We should look to Sonya Sheridan (1983) and Generative Systems as a model to dismantle and question new technologies, not by outright rejecting them and not engaging them as end users. We can exist somewhere in the middle space between irrational Luddism and, conversely, unexamined techno-optimism. Nam June Paik’s aphorism “I use technology in order to hate it more properly” stands the test of time across media.

Ultimately, with the burgeoning economy of artificial intelligence, we arrive at the doorstep of an infinite cybertime, the emergence of a now unrestrained cognitive field, as imparted by Franco Berardi (2015). Previously constrained by human temporal capacity, this synthetic cognitive domain now assumes the boundless form once only attributable to cyberspace. Generative cognition suggests a new incalculable spatial x-axis perpendicular to the plane of omnipotence, omniscience, and omnipresence, what I (2022) have previously defined as “omnigenesis.”

The challenge of our time is to navigate this infinity with a compass guided by defiance, illegibility, and justice through transgressive praxis. It is only by confronting our apeirophobia—our fear of the infinite—that we can hope to reshape the contours of power, agency, and creativity in a world increasingly mediated by AI.

The real specter of infinity we face is not the one of endless cognitive appropriation in the arts but the perpetuation of cycles of domination, subjugation, and death.

Digital images are what data do in a sequence. Arrangements, compositions, adjustments: an ever-generative repository of computational relations trying out its coded capabilities to make itself intelligible. As intelligibility has its bounds, there’s always something unintelligible to suggest. Staying with this something— repeating it, with slight modifications—grants this process its seemingly infinite duration. This is, at least for me, what Christopher Meerdo’s Apeirophobia teaches about dark infrastructures: since the intelligibility that computation strives for also generates infrastructural conditions of unintelligibility (i.e., darkness), (re)claiming “endless sets” of data means to (re)claim their endlessness; that is, the time they take.

When thinking about it, “endless sets” are not far from what jam bands do. Here, I specifically mean the collective effort—both by the band and the audience—to stay with a song to the point where it might lose its clear contours (the track goes off-track) without, however, falling apart. Like its visual counterpart, the liquid light show, the jam proceeds through variations and repetitions of a certain set of elements as long as the song (carried on by the band and the audience alike) requires.

As a “dark commentary” on Christopher Meerdo’s Apeirophobia, I took the drum solo from Iron Butterfly’s In-A-Gadda-Da-Vida (one of the emblematic tracks of the 1960s psychedelia, particularly in terms of its length) and stretched the 2.5-minute clip to one full hour. What we hear there is the algorithm trying to keep up with the rhythm in a long song made even longer. Akin to the visual haze between images in Apeirophobia, the eerie noise between the hits is computational intelligibility in action: a suggestion (not) making sense.