Free Writing (Takeaway)#
2026
10,000 postcards with individual generated texts, 20 postcard racks
The work is accompanied by a log about the process and the concept.
Free writing, also known as automatic writing, is a writing practice in which one writes for a set period of a few minutes. There is no planning, no corrections, no looking back. It is therefore a physical practice that captures the stream of consciousness. The texts produced in this way are sometimes personal, incoherent, flawed, valid in the moment. This is because the focus is not on the result, but on the process, on overcoming writer's block, and, with a little luck, gaining insight without a strategic approach. Basically, however, the texts are not intended for publication.
118 of these texts, which Kuhn wrote in seven minutes each between 2023 and 2026, served as a dataset to train a large language model (LLM) via fine-tuning to simulate his free writing practice. As a result, 10,000 texts were generated and printed on postcards—traditionally material evidence of the expansion of one's world horizon.
While the dataset alone would fill just over 200 postcards, an unlimited number of postcards can be filled with generated texts based on it. As input for text generation, one sentence each from excerpts of literary and theoretical texts that Kuhn has compiled for his artistic research—as part of his world horizon—was used. For example, "The slightest deviation from my schedule throws me off balance" (translated from German) from Thomas Bernhard's Gargoyles or "Maybe too much intellectual effort has been lost to making distinctions between machines and humans" from Yuk Hui's ChatGPT, or the Eschatology of Machines. Depending on the language of the input, the LLM generated texts in German or English. In some cases, English phrases creep into German texts: "Natürlich sorgt mein Konsum für Probleme [Of course, my consumption causes problems] somewhere else."
The inputs also influence the content and style of the generated text. One input from Thomas Bernhard's Correction, abbreviated here and translated, reads: "The person who has entered here is forced to give up, to abandon everything he thought before, up to the moment of his entry into Höller's attic room [...] it had to be the thinking of Höller's attic room, the thinking that relates exclusively to everything connected with Höller's attic room and with Roithamer and the cone." The continuation of the LLM: "It is clear that this would not be possible for a normal human being, but the person who has entered here is not a normal human being, but an artificial product, a machine, an algorithm, which, unlike humans, has no difficulty in devoting itself completely to a closed system and no longer leaving it." On the one hand, the "person who has entered" is adopted, and on the other hand, the choice of words and rhythm are reminiscent of Bernhard's literature. Nevertheless, the language is a reflection of the data set.
The content of the generated texts is sometimes factually incorrect, for example when authors are attributed texts that they did not write. Statements also arise that need to be checked for accuracy, such as "Byung-Chul Han writes the following on this subject:> First of all, I believe in the power of poetry. Poetry can be found everywhere, not just in literature. Poetry occurs in our dealings with life, with other people, with the world." (Translated.)
Kuhn separates himself from his own writing and can thus observe himself writing. On the one hand, the machine author is more than him, but on the other hand, less. While free writing is a personal practice, text generation is initially the opposite of personal writing, but on the other hand, it can simulate personal writing through training with personal texts. This sometimes results in passages that seem poetic: "This only leads to me neglecting myself and ultimately burning out. Nothing is burning yet, but the edges of the fire are already visible." Sentences are often generated—"Thinking again instead of typing"— that appear to come from the data set, but were only generated based on the patterns contained therein. The LLM simulates the style of writing, both in terms of content and form, in such a way that texts similar to those by Kuhn are created, but which go beyond his own authorship.
The postcards are placed in 20 wall mounted postcard racks and offered to take away. Some of the generated texts are exciting, others rather meaningless or superfluous. It is up to the reader to decide, but finding a suitable card is like playing the lottery.
Photos 3-5: Jakob Dieckmann
Font: Muto Mono by Rade Matic
LLM: Mistral-Small-24B-Base-2501 (fine-tuned)