One and Infinite Chairs
April, 2023.
Self hosted and custom trained Stable Diffusion; .ckpt-format file of a collapsed Ai model; essay-film, duration: 06’36”.
Collaborators: AI engineer – Artem Konevskikh; Voice – Sara Woodgate.
The experiment began by generating an initial dataset of 1,000 photo-realistic chair images using Stable Diffusion from the prompt "A single chair on a plain background." Retraining the AI model repeatedly—a process intended to enhance its capacity for rendering chairs—was repeated iteratively. However, by the sixth cycle, the model failed: instead of recognizable chair images, it produced colorful, noisy patterns with no discernible resemblance to the original subject. This phenomenon is known in data science as data-cannibalism. Given the increasing prevalence of generative AI, future systems are increasingly trained on synthetic datasets. This creates significant ontological challenges, threatening to pollute these datasets and compromise the epistemic accuracy of AI models. With synthetic images now surpassing conventional photography on the internet, this auto-generated feedback loop risks causing AI's visual outputs to decay into non-figurative abstractions—at least beyond human recognition.
It begins with an experiment: approximately 1,000 images of chairs were generated using an open-source, text-to-image AI model, Stable Diffusion, via the prompt “A single chair on a plain background.” This newly created dataset of nearly photo-realistic images, depicting a wide variety of chairs, was then used to retrain the same AI model. The process of retraining is commonly intended to enhance model capacity, in this case to render chairs. Following this, using the same prompt to generate another round of chairs, such re-training on its own, generated imagery, prompting, and re-training was repeated iteratively. By the sixth iteration, however, the model underwent a progressive decline, and instead of the photo-realistic images of chairs from the initial step, produced colourful, digital, noisy patterns, with no discernible resemblance to the represented subject – a chair. At least, not perceptible to the human eye. In data science, the phenomenon of AI feeding into AI is often referred to as data-cannibalism. Through the necessity to augment datasets and due to AI image and data generation’s increasing and insidious prevalence, more and more new AI systems will be trained on synthetic datasets, produced by generative AI models, thus posing ontological challenges and poisoning future datasets and epistemic accuracy of those models. With an amount of synthetic images on the internet exceeding conventional photographic imagery at this point, such feedback loops and echo chambers of auto-generated and auto-consumed data present real risks to distill Ai’s technical capacities to represent domain ontologies of subjects, rendering their visual outputs decay into non-figurative abstractions… at least so for a human eye.
Project Lexicography
While pareidolia is the tendency to perceive meaningful images or patterns where none actually exist. Cyberdolia is a similar misinterpretation occurring in machine vision contexts.
A form of environmental pollution that constitutes a vastly growing infoscape, resulting from the epidemics of information production and consumption.
A speculative concept describing an emergent aesthetic system that arises from autonomous processes rather than from direct human intention. The term combines auto- (self, autonomous) with aesthetics, suggesting forms of perception, judgment, or sensibility generated by machines, algorithms, or self-organizing systems.
Refers to the interdisciplinary study of engineering systems of knowledge, encompassing its nature, sources, and limits. It draws from the term epistemics.
'One & ∞ ⑁' film-essay
One and Infinite Chairs, film-essay, 2024
Installation view at Siggraph Asia 2024, Tokyo International Forum, JP
Installation view at Ars Electronica 2022, Postcity, Linz, AT
The world of Objects, Domain Ontologies & Other Philosophical Inquiries
Joseph Kosuth’s One and Three Chairs is one of the most quintessential examples of conceptual art because it touches upon a number of characteristics that are definitive of conceptual art. Art that emphasises the concept above all other perceptual content, it is associated with a dematerialisation of art.
One & Three Chairs consists of some fairly obvious ingredients. Everything is in a one-to-one ratio: here is a chair, here is an image of a chair, here is the written definition of a chair, and so you have three different representations of a chair. To substitute one chair with another would not diminish this work. This conceptual dematerialisation is often contextualised as an escape away from the commodity form. The work invites itself to be contextualised within various philosophical exercises, i.e., What is the concept of a chair? How does it include an image into its concept? How does it include the use value of a chair into the concept of what constitutes a chair? How do language, art, and abstract concepts manifest in physical reality? What is the work's relationship to Platonic Forms? It advocates revisiting Wittgenstein’s language and meaning, or challenging it against Hume's relations of ideas and matters of fact. And, of course, Kant’s critique of how the physical form of a chair conforms to our knowledge of how it can be used? Many of these questions within Western philosophical thought are thousands of years old.
Duchamp’s iconic exercise was about how the institution takes claim over meaning, and thus makes the meaning itself. In the context of Kosuth’s Chairs, the art institution instantiates the assembly as a work of art. It does so out of nothing; it renders the artist useless. The artist is unnecessary because this act of assembling these representations is rather an act of curation. It curates representations of chairs, in this case.
Selected uncanny chairs produced in the process of training the model. 1/2
On recursive systems and echo chambers
In another iconic conceptual sound artwork, I am sitting in a room, the author Alvin Lucier is recording himself narrating a text, and then playing the tape recording back into the room and re-recording it. The new recording is then played back and re-recorded, and this process is repeated for over an hour. Due to the room's particular size and geometry, certain frequencies of the recording are emphasised, while others are attenuated. Eventually, the words become unintelligible, replaced by the characteristic resonant frequencies of the room itself. The text spoken by Lucier describes the very process of the act and predicts what will such re-recording of his voice eventually turn into – a semantically unintelligible soundscape produced by the natural resonant frequencies of the room, intensified through the feedback-loop process of playing back and re-recording.
Our expectation to see a work of art as an end in itself is an expectation of seeing an object that stands out from the world of objects, the world by access as a means to another end. The way I access a chair to sit down, or the way I access a computer to look up art.
However, at this point, it is self-evident – a computer is no longer merely an object that provides an access as a means to another end. Computational assembly of information, synthetic cognitive capacities of AI, and logistics of planetary scale interconnectedness not only change the way we institute knowledge, understand agential relationship and how we make art, but it also affects the very ontology of concepts. The following is an exercise of these dynamics through a somewhat oddly-titled work-study – "1 & ∞ Chairs".
Weizenbaum Conference 2023, Berlin (DEU). A panel discussion alongside Egor Kraft, David Berry, and Barbara Pfetsch on the risks and implications of AI as a generative technology, where One & Infinite Chairs had been first publicly screened.
One & ∞ ⑁
It begins with an experiment: approximately 1,000 images of chairs were generated using an open-source, text-to-image AI model, Stable Diffusion, via the prompt “A single chair on a plain background.” This newly created dataset of nearly photo-realistic images, depicting a wide variety of chairs, was then used to retrain the same AI model. The process of retraining is commonly intended to enhance model capacity, in this case to render chairs. Following this, using the same prompt to generate another round of chairs, such re-training on its own, generated imagery, prompting, and re-training was repeated iteratively. By the sixth iteration, however, the model underwent a progressive decline, and instead of the photo-realistic images of chairs from the initial step, produced colourful, digital, noisy patterns, with no discernible resemblance to the represented subject – a chair. At least, not perceptible to the human eye. In data science, the phenomenon of AI feeding into AI is often referred to as data-cannibalism. Through the necessity to augment datasets and due to AI image and data generation’s increasing and insidious prevalence, more and more new AI systems will be trained on synthetic datasets, produced by generative AI models, thus posing ontological challenges and poisoning future datasets and epistemic accuracy of those models. With an amount of synthetic images on the internet exceeding conventional photographic imagery at this point, such feedback loops and echo chambers of auto-generated and auto-consumed data present real risks to distill Ai’s technical capacities to represent domain ontologies of subjects, rendering their visual outputs decay into non-figurative abstractions…at least so for a human eye.
Related Works
One and Infinite Chairs
April, 2023.
Self hosted and custom trained Stable Diffusion; .ckpt-format file of a collapsed Ai model; essay-film, duration: 06’36”.
Collaborators: AI engineer – Artem Konevskikh; Voice – Sara Woodgate.
The experiment began by generating an initial dataset of 1,000 photo-realistic chair images using Stable Diffusion from the prompt "A single chair on a plain background." Retraining the AI model repeatedly—a process intended to enhance its capacity for rendering chairs—was repeated iteratively. However, by the sixth cycle, the model failed: instead of recognizable chair images, it produced colorful, noisy patterns with no discernible resemblance to the original subject. This phenomenon is known in data science as data-cannibalism. Given the increasing prevalence of generative AI, future systems are increasingly trained on synthetic datasets. This creates significant ontological challenges, threatening to pollute these datasets and compromise the epistemic accuracy of AI models. With synthetic images now surpassing conventional photography on the internet, this auto-generated feedback loop risks causing AI's visual outputs to decay into non-figurative abstractions—at least beyond human recognition.
It begins with an experiment: approximately 1,000 images of chairs were generated using an open-source, text-to-image AI model, Stable Diffusion, via the prompt “A single chair on a plain background.” This newly created dataset of nearly photo-realistic images, depicting a wide variety of chairs, was then used to retrain the same AI model. The process of retraining is commonly intended to enhance model capacity, in this case to render chairs. Following this, using the same prompt to generate another round of chairs, such re-training on its own, generated imagery, prompting, and re-training was repeated iteratively. By the sixth iteration, however, the model underwent a progressive decline, and instead of the photo-realistic images of chairs from the initial step, produced colourful, digital, noisy patterns, with no discernible resemblance to the represented subject – a chair. At least, not perceptible to the human eye. In data science, the phenomenon of AI feeding into AI is often referred to as data-cannibalism. Through the necessity to augment datasets and due to AI image and data generation’s increasing and insidious prevalence, more and more new AI systems will be trained on synthetic datasets, produced by generative AI models, thus posing ontological challenges and poisoning future datasets and epistemic accuracy of those models. With an amount of synthetic images on the internet exceeding conventional photographic imagery at this point, such feedback loops and echo chambers of auto-generated and auto-consumed data present real risks to distill Ai’s technical capacities to represent domain ontologies of subjects, rendering their visual outputs decay into non-figurative abstractions… at least so for a human eye.
Project Lexicography
While pareidolia is the tendency to perceive meaningful images or patterns where none actually exist. Cyberdolia is a similar misinterpretation occurring in machine vision contexts.
A form of environmental pollution that constitutes a vastly growing infoscape, resulting from the epidemics of information production and consumption.
A speculative concept describing an emergent aesthetic system that arises from autonomous processes rather than from direct human intention. The term combines auto- (self, autonomous) with aesthetics, suggesting forms of perception, judgment, or sensibility generated by machines, algorithms, or self-organizing systems.
Refers to the interdisciplinary study of engineering systems of knowledge, encompassing its nature, sources, and limits. It draws from the term epistemics.
One and Infinite Chairs, film-essay, 2024
The world of Objects, Domain Ontologies & Other Philosophical Inquiries
Joseph Kosuth’s One and Three Chairs is one of the most quintessential examples of conceptual art because it touches upon a number of characteristics that are definitive of conceptual art. Art that emphasises the concept above all other perceptual content, it is associated with a dematerialisation of art.
One & Three Chairs consists of some fairly obvious ingredients. Everything is in a one-to-one ratio: here is a chair, here is an image of a chair, here is the written definition of a chair, and so you have three different representations of a chair. To substitute one chair with another would not diminish this work. This conceptual dematerialisation is often contextualised as an escape away from the commodity form. The work invites itself to be contextualised within various philosophical exercises, i.e., What is the concept of a chair? How does it include an image into its concept? How does it include the use value of a chair into the concept of what constitutes a chair? How do language, art, and abstract concepts manifest in physical reality? What is the work's relationship to Platonic Forms? It advocates revisiting Wittgenstein’s language and meaning, or challenging it against Hume's relations of ideas and matters of fact. And, of course, Kant’s critique of how the physical form of a chair conforms to our knowledge of how it can be used? Many of these questions within Western philosophical thought are thousands of years old.
Duchamp’s iconic exercise was about how the institution takes claim over meaning, and thus makes the meaning itself. In the context of Kosuth’s Chairs, the art institution instantiates the assembly as a work of art. It does so out of nothing; it renders the artist useless. The artist is unnecessary because this act of assembling these representations is rather an act of curation. It curates representations of chairs, in this case.
Installation view at Siggraph Asia 2024, Tokyo International Forum, JP
On recursive systems and echo chambers
In another iconic conceptual sound artwork, I am sitting in a room, the author Alvin Lucier is recording himself narrating a text, and then playing the tape recording back into the room and re-recording it. The new recording is then played back and re-recorded, and this process is repeated for over an hour. Due to the room's particular size and geometry, certain frequencies of the recording are emphasised, while others are attenuated. Eventually, the words become unintelligible, replaced by the characteristic resonant frequencies of the room itself. The text spoken by Lucier describes the very process of the act and predicts what will such re-recording of his voice eventually turn into – a semantically unintelligible soundscape produced by the natural resonant frequencies of the room, intensified through the feedback-loop process of playing back and re-recording.
Our expectation to see a work of art as an end in itself is an expectation of seeing an object that stands out from the world of objects, the world by access as a means to another end. The way I access a chair to sit down, or the way I access a computer to look up art.
However, at this point, it is self-evident – a computer is no longer merely an object that provides an access as a means to another end. Computational assembly of information, synthetic cognitive capacities of AI, and logistics of planetary scale interconnectedness not only change the way we institute knowledge, understand agential relationship and how we make art, but it also affects the very ontology of concepts. The following is an exercise of these dynamics through a somewhat oddly-titled work-study – "1 & ∞ Chairs".
Installation view at Ars Electronica 2022, Postcity, Linz, AT
One & ∞ ⑁
It begins with an experiment: approximately 1,000 images of chairs were generated using an open-source, text-to-image AI model, Stable Diffusion, via the prompt “A single chair on a plain background.” This newly created dataset of nearly photo-realistic images, depicting a wide variety of chairs, was then used to retrain the same AI model. The process of retraining is commonly intended to enhance model capacity, in this case to render chairs. Following this, using the same prompt to generate another round of chairs, such re-training on its own, generated imagery, prompting, and re-training was repeated iteratively. By the sixth iteration, however, the model underwent a progressive decline, and instead of the photo-realistic images of chairs from the initial step, produced colourful, digital, noisy patterns, with no discernible resemblance to the represented subject – a chair. At least, not perceptible to the human eye. In data science, the phenomenon of AI feeding into AI is often referred to as data-cannibalism. Through the necessity to augment datasets and due to AI image and data generation’s increasing and insidious prevalence, more and more new AI systems will be trained on synthetic datasets, produced by generative AI models, thus posing ontological challenges and poisoning future datasets and epistemic accuracy of those models. With an amount of synthetic images on the internet exceeding conventional photographic imagery at this point, such feedback loops and echo chambers of auto-generated and auto-consumed data present real risks to distill Ai’s technical capacities to represent domain ontologies of subjects, rendering their visual outputs decay into non-figurative abstractions…at least so for a human eye.
Selected uncanny chairs produced in the process of training the model. 1/2
Weizenbaum Conference 2023, Berlin (DEU). A panel discussion alongside Egor Kraft, David Berry, and Barbara Pfetsch on the risks and implications of AI as a generative technology, where One & Infinite Chairs had been first publicly screened.
Related Works
About
Egor 'Eji' Kraft (映治 克夫塔) is a conceptual artist, writer, filmmaker, and critical design researcher.
Studio Locations
Ikejiri, Setagaya, Tokyo, Japan
Neubau, Vienna, Austria
Contact:
mail[at]kraft.studio
Studio Locations
Ikejiri, Setagaya, Tokyo, Japan
Neubau, Vienna, Austria
Contact
mail[at]kraft.studio
About
Egor 'Eji' Kraft (映治 克夫塔) is a conceptual artist, writer, filmmaker, and critical design researcher.
