Symbiotic Traces: An Ecological Record of Language Model Training

What would it mean to visualize a language model’s training process in the vocabulary of a crisis it will never be able to feel? Symbiotic Traces pairs fifty stages of the OLMo 3 32B pretraining run with climate terms drawn from IPCC reports and Earth system science, and then with original steam-printed marigold petals.

AI interprets deterministic rules to assign each climate term, while the physical chemistry of botanical printing shapes each final image. With creative agency distributed across machine, artist, and material process, the result is a visual language for a model under construction.

Training a single large language model requires an extraordinary amount of energy: the pretraining run for OLMo 3 32B, an open model released by the Allen Institute for AI, consumed roughly 681 megawatt-hours from its GPUs alone [1]. Symbiotic Traces asks what it would mean to describe that consumption in the vocabulary we use for a warming planet, and to let the model’s own training process help write the description.

The project pairs two entangled systems: the pretraining of a large language model, and the climate destabilization to which its energy use contributes. Each of the run’s fifty stages is paired with a term from the vocabulary of climate change and an original botanical print, made by steaming a marigold petal directly onto paper. We chose marigolds because their growing seasons are among those already shifting under climate change, which makes climate change both the subject of the work and part of its material process.

The dataset is the training log of OLMo 3 32B, a 32-billion-parameter open model trained on 5.7 trillion tokens between September and November 2025. The log records how the model became usable and, by proxy, the energy that process consumed. A climate lexicon, compiled with Claude Fable 5 from IPCC reports and climate literature, was narrowed to fifty words spanning growing-season and species-range terms alongside the broader climate vocabulary.

Each of the fifty training stages receives one word through a single deterministic rule: the stage is checked against a fixed list of measurable conditions, each condition that holds contributes a candidate word, and one is selected. The prints are made by arranging locally sourced marigold petals between sheets of moistened Arches paper, wrapping the bundle around a dowel, sealing it in parchment, and steaming it. Heat and moisture release pigment from the petal into the paper fiber. Neither we nor the model controlled the resulting image: each print’s coloration, bleed, and stain are determined by heat and moisture alone, an image that carries a direct physical trace of its source, in the sense of what Dietmar Offenhuber calls autographic visualization [2].

Each climate term is then paired with a petal print by visual association, matching the tone and connotation of the word to the print whose form or gesture feels most complementary. The pairing of text and image guides the reading, as Roland Barthes describes text and image anchoring and relaying one another’s meaning [3]. Sequenced together, the fifty pairings form a visual lexicon in which language drawn from a machine’s training and a mark drawn from a living plant sit side by side. For the video, the fifty prints are algorithmically morphed into one another.

The work treats the artist’s partial delegation of control over the final image, to the machine and to the material, as an expansion rather than a diminishment of creative agency. Agency resides in designing the conditions under which words and images emerge: the rule that assigns climate words to training stages, and the process through which petals transfer their pigments to paper. Long before AI, Vera Molnár argued that delegating parts of artistic production to a computational system could reveal what is most human [4].

At another scale, climate change is a problem so vast that any personal response can feel inconsequential. Pairing each abstract term with the physical trace of a living flower is a way of making the crisis legible at a tangible, material scale.

Most directly, the project is distributed across three actors whose agencies remain deliberately unresolved: the language models, OLMo 3, whose training behavior selects each word, and the model that compiles the lexicon and writes the code; the botanical printing process, whose steam and pigment determine each mark; and the artist, who designs the rules and composes the pairings. None of the three is fully in control, and none is fully passive. Novelist Olga Ravn draws a sharp line between intelligence and consciousness [5]; the model is an intelligent but non-conscious collaborator, capable of sorting and selecting language but not of feeling or meaning-making. The human artist is the emotional, embodied, and discerning counterbalance.