Ting-Chun Liu and Leon-Etienne Kühr
The global chip shortage during the COVID-19 pandemic brought semiconductor production into focus, sparking accelerated efforts to meet the surging demand for digital infrastructure. This phenomenon not only expanded AI capabilities but also introduced unexpected computational artifacts.
One such artifact is the word “arafed”, a term absent from any dictionary yet mysteriously appears across contexts from image prompts to Amazon product descriptions. Such unintended linguistic artifacts, born from transformer-based AI models, exemplify how digital artifacts emerge into realities with which we cohabitate.
The talk investigates how supply-chains break and AI-words spread from an artistic research perspective. Mapping both the abstract landscapes of embedding spaces, that are filled with emergent words and images, and the tangible, geopolitical realities of global semiconductor supply chains.
The accelerating pace of generative AI has put a strain on the interconnected software and hardware systems necessary for generative AI. The artist duo explores the media specificity of generative artificial intelligence. The talk consists of two parts: The material aspects of AI, specifically the story of semiconductor and chip shortage. And the spread of hallucinations like terms that escaped their embedding space into language.
The working of LLMs is often limited by computational power. These obstacles tethered abstract computation to the physical world, exposing how materiality plays a critical role in the implementation of AI. The investigation begins by examining the causes of the chip shortage — a disruption that brought the semiconductor industry and its surrounding geopolitical tensions into discourse.
On the hardware level, NVIDIA’s A100 chips, produced using Taiwan’s TSMC 7nm process, exemplify this intersection, providing the power to expand large language models (LLMs) and image generators. On the software level, the increasing demand for ai-as-service accelerates the use of models with complex pipelines. This interconnected use of models, in turn, leads to the emergence of unexpected artifacts that are morphing back into everyday reality.
While browsing AI-generated images on social media, one might come across the word "arafed" in image descriptions, such as, "an arafed man in a white robe riding on top of a blue car.". Yet, a dictionary definition is nowhere to be found. An image search for "arafed" reveals something striking: all resulting images appear AI-generated, spread across various image-sharing and stock photography platforms.
The term "arafed" seems to lack a clear origin, but a few posts attribute it to the BLIP-2 model, an image-captioning system that generates descriptive text from image inputs. However, the BLIP-2 paper doesn't mention "arafed" but running BLIP-2 clearly produces descriptions containing this artifact-like word, as if "arafed" has embedded itself in the model's vocabulary. Through the widespread and often unintentional use of BLIP-2 in libraries, extensions, and services, the interconnected nature of software has spread the word into research papers, Amazon descriptions, and even other datasets, further revealing the brittle infrastructure generative-ai systems are built upon.
Licensed to the public under http://creativecommons.org/licenses/by/4.0
This Talk was translated into multiple languages. The files available for download contain all languages as separate audio-tracks. Most desktop video players allow you to choose between them.
Please look for "audio tracks" in your desktop video player.