#ai

Artificial Suspects

The artificial intelligence model in the learning process condenses not only knowledge but also biases. And not only those contained in the quality of the dataset itself (from whom, for whom, why?), but also in the deeper layers of human culture. The result is the ability of AI models not only to generate but also to reduce.

In the case of GPT4o, this is the ability to represent text/words in the form of 3072 numerical parameters. Each word, concept, and longer text thus essentially occupies a specific coordinate in a 3072-dimensional space in the model. This location and proximity to other represented objects then gives meaning not only to a specific text or word, but also, retrospectively, to the world and to ourselves, when the model operates on data analysis of our email inboxes, chats, resumes, essays, or official records.

Truthify: Dataset Lobbying

Truthify: Dataset Lobbying, in collaboration with Vilem Duha and Petr Racek, is a speculative video (7:38) about the eponymous startup that seeks to manipulate the perception of reality of future neural networks through the manipulation of open source datasets. Specifically, to map a dataset of images of Czech art under the general tag of art, with the aim of promoting Czech art as more detectable and known, and therefore as more artsy than global art. Working with hyperbole and a slight ironic patriotism, the video exposes the serious topic of biased datasets and neural networks that have and will have impacts on our everyday lives.