Within Limits

October 22, 2026 Summer 2026

“A web of sudden connections and personal links” - Olia Lialina in A Vernacular Web

Handmade Datasets proposes strategies for working critically with Artificial Intelligence at multiple junctures in its pipeline, from data collection to model making. This class began with the question, prompted by Aarati Akkapeddi, what if datasets were small enough to care for? Inspired by J.R. Carpenter’s The Handmade Web, the “handmade dataset” evokes a similar sense of embodiment and a challenge to convention: what methods, considerations, or approaches to the dataset can shape new relations with captured data, automated processes, or the hyperscale?

The handmade dataset is specific, and specifically non-commercial, in purpose; however, its purpose is non-deterministic. Each datapoint is considered within the limitations of human attention, explicit consent, and intention behind the dataset. Students in the class took this approach many steps further and investigated the unwieldy data pipeline while also deepening relationships to their home worlds, personal histories, and built environments. The class experimented with open source models that train from scratch and are intended for small-scale uses, including but not limited to the Pix2Pix GAN (generative adversarial network) for images, RAVE (realtime audio variational autoEncoder) for audio, plus RAG (retrieval-augmented generation) and microGPT (generative pre-trained transformer) for text.

The slow and tedious nature of a handmade dataset lends itself to a contemplative process, one that can double as a loving embrace, or babble like an endless stream of curiosity. The first visual project of the class used Pix2Pix, a type of generative adversarial network and an image-to-image translation model from the early 2010s that learns to turn one kind of image into another by training on paired examples. Before any kind of data collection began, Aarati guided students with a survey which included questions such as What translation might be meaningful to you or your community?, Whose perspective is centered in this translation?, and What would a corporate dataset miss about this translation? We discussed key points in the data pipeline for training LLMs, like web-crawling practices, click-worker labor, and alt-text as training material, further entangling this creative endeavor with the realities of alienated labor. Students took up this challenge to become stewards of their handmade datasets and drafted new formulations for “generating” meaning.

The Jacket Project by Christine began with a photo of a suit jacket made by her grandfather, a tailor whom she never met and whose material trace is now faint. Her datasets were created with digital renderings of his surviving jacket designs along with her own. Hoping that the model would “make visible the gap between his tailoring practice and my inexperience,” she said “the jackets generated by AI represent the impossible and imaginary space that our missing relationship conceived.”

Carey J. Flack’s Sketches from the Field (links to Part 1 and Part 2) investigates the leaf vein networks of a creeping fig growing in her neighborhood, an invasive vine which initially posed only a small ethical hurdle to leap when plucking hundreds of leaves to be individually scanned. This meditative practice, which she dubbed “foraging data by hand,” was troubled by her recognition of the silent logic at work –that abundance grants permission– a way of thinking not too dissimilar from that of large-scale data extraction. This troubling surfaced after harvesting leaves for dozens of hours, at which point the friction inherent to the laborious yet meditative handmade dataset became a method of discovery: ”I reflected on how we might design a shared future that balances human, machine, and ecological intelligence.” A newfound appreciation for the fence-eating plant colored the work as she continued this project through the end of class.

Seeking out “non-data” when making Head in the Clouds, Lauren Malkani worked alongside her four-year-old daughter in teaching a pix2pix model how to see the world like a child. Using photographs of clouds against scans of her daughter’s drawings, which were made over many weeks to prevent any fatigue on behalf of the young artist, the model was trained to translate a line drawing into a cloud. This process encouraged more mother-daughter time and deeper interrogation of the drawings’ thematic motifs, artistic influences, and underlying whimsical spirit her daughter had imbued in them.

Christine, Carey, and Lauren’s Pix2Pix projects reveal the handmade dataset as a means for amplifying our sense of interconnectedness without taking on the whole world at once, or assuming that care can be neatly disentangled from labor.

“Birds aren’t real, or are they?” - Title of a NYT article on the viral misinformation satire “Birds Aren’t Real”

Other students in the class leaned into the handmade dataset as a tool for probing the partial and incomplete caverns of knowledge in which we find ourselves lurking; there are many entry points to this labyrinth, and just as many exits. Kim Forero-Arnias’s How many forms can you take before we know you? took shape after pouring over ornithological collections at the library and reflecting on the “distortions” in understanding that might arise from library science’s approach to documenting fauna. Her Pix2Pix model was created with one dataset of selected photographs from iDigBio’s collection of Sturnidae and one dataset of scans made by printing those same images and physically crushing them inside the scanner. Her final project, a video titled The Shape of a Starling, combined these images with a RAVE (Realtime Audio Variational autoEncoder) model trained on sound recordings of the same bird species. Kim explained that the process of doing so stirred up the feeling that “there was a bird stuck in a net.” Together, these models produce a strange approximation of a bird, one whose form is caught up in a web of library catalogues and scientific monotony just as it's fallen through the pages and escaped to the imagination.

Sarah Alvarez’s final project Packet also addresses networked knowledge, but not from the skies – from the underground. They created a web interface for a speculative paquete drop, inspired by El Paquete Semanal, in which each slightly illegible filename is generated from a microGPT they trained on a file directory from a real Cuban paquete that had been uploaded to GitHub in 2016. Clicking through the file tree, you’ll notice that many files cannot be opened –a built-in interruption to the user’s anticipated flow of information, or perhaps a reminder of the constant dread of disconnection. The audio files contain sounds from a RAVE model that Sarah created using TikTok videos they had been saving over the course of six years. They recreate that uneasy tension between the global and local within digital culture through audio that is hard to decipher yet dense with alluring sonic artifacts. Sarah described this attempt "to boil the internet down” into a single hard drive as a “survey of an internet that never was.” The handmade dataset became a means for simulating this collapse and producing a multifaceted vision of media dissemination that combines human curation and algorithmic design.

This cybernetic hype leads us to June Maas’s locally run RAG (retrieval-augmented generation) model which she trained on websites she had been collecting with her own web crawler, intended to be a non-commercial web navigator that “points you in a new direction.” Responding to the user’s text input, the model shares resources from this curated collection of websites, many of which are personal blogs and handmade websites that do not contain roadblocks like email gates and anti-adblock or tracker scripts in their design. Tackling the fragmented web left behind by rampant advertising schemes and LLM enshittification, June aspires to “grow a fungus out of google’s decaying body.” Sometimes it feels like the Internet has grown out of proportion, too scattered, perhaps too homogenous, or become a ruin of its own ambition. June’s project is an elixir against this doomerism, a reason to believe otherwise: maybe birds are real. She showcases the handmade dataset as a guiding philosophy and organizing principle, a divining rod for linking past and present models of information sharing and stewardship in the digital world.

“You can’t drink data” - Slogan written across a protestor’s sign at Orangetown Town Hall

Back in July, we took a virtual field trip to Aarati’s hometown for a public hearing concerning a proposed data center expansion that would encroach on protected wetlands and a life-sustaining reservoir. Live-streamed over Zoom via Aarati’s smartphone, students witnessed impassioned speeches from a grandmother of two, a local geologist, a teenager dressed up as Spiderman, and many more Orangetown characters with a great and loving concern for their community. They pointed out the many risks posed by the data center expansion which had been left out of the original proposal, like excessive energy consumption and groundwater contamination, in addition to an overall lack of transparency regarding multi-million dollar tax incentives promised to its owners. Contrasting the gray and blue suits and controlled rhythm of conversation inside the room, there was large protest organized just outside the Town Hall building. A joint effort by the Datacenter Crisis Coalition and Food & Water Watch, it grew loud enough to be heard on the other side of our Zoom call. Documented protest signs in the press that followed included slogans such as “You can’t drink data!” and “No rate hikes.”

The environmental impact of computing, and specifically AI, was a recurring conversation topic for the class. How does one begin to grasp the planetary scale of data collection and analysis? One of the largest customers for DataBank, the company running the Orangeburg data center, is an AI cloud-computing company named CoreWeave. While the local residents did not explicitly name artificial intelligence as their enemy, the implication remained: the current AI boom and data center explosion are economically and infrastructurally co-dependent. As a result of this Town Hall hearing, tied up with the numerous and ongoing protests from local organizations, Databank must prepare a comprehensive Environmental Impact Statement to the Orangetown Planning Board before moving forward with its final vote. The virtual field trip exposed the class to this electrifying moment in national history: a renewal of local power and a shared sense of hope in the face of “frictionless” innovation. No matter how the odds may appear, our future is still a collective bargain. Even if it is a gamble, we must ask: What other strategies for living and creating in a post-AI reality might only be possible with serious technical constraints? Handmade Datasets can be a philosophy, method, or objective; while it might use “out-of-date” tools or refuse objectivity, it strives to sustain a technological future that welcomes subversion and values the human-scale.