2022

created by Kristian Vrhar on stable diffusion

(Non) Artificial (Maybe) Intelligence

Seminar

What and who and how and why? So many questions around AI, tools, models and training Data, possibilities, limitations and risks. Using these technologies to comment on the very same ones through the lenses of history, politics, society, culture and science we approached this with humour as a high order maybe-intelligence.

Seminar Trailer

AI Dungeon

Google's Quickdraw

COCO Common Objects in Contex

The seminar (Non) Artificial (Maybe) Intelligence was my first one working as an academic associate at the Karlsruhe University of Arts and Design in the winter semester 2022, a time where generative AI was reaching new levels and public popularity was high through Image generation platforms like Midjourney, OpenAi or Stable Diffusion. Actually ChatGPT was released midway during the seminar. Not knowing the school and the students I chose this topic for a seminar anticipating some interest and to expand on the research I had started during my graduation work deep sea bot. where I imagined what form and body a digital assistance might take and hence what agency and motivation it could have.

Starting with the ‚demystification‘ of the words Artificial and Intelligence we tried to get an understanding of the architecture of data models, especially regarding equity of training data, classification bias and unseen labour, through reading, testing, using and misusing these tools. Our goals were the following:

Define contextualization methods
understand machine learning and evaluate social relevance
Try out and analyze A.I. natural language processing and language to code, tools like Dall-E, monster mesh (3D), EbSynth (rotoscope animation), and and and
Comment on these tools through practical use or misuse of the same tools
document and reflect on what we learned

We started by playing AI Dungeon, a text-based, AI generated fantasy simulation game, then we read The Nooscope Manifested by Vladan Joler and Matteo Pasquinelli, Excavating AI - The Politics of Images in Machine Learning Training Sets by Kate Crawford and Trevor Paglen, watched computational and mathematical explanations by Daniel Shiffman and 3Blue1Brown, used COCO, Common Objects in Context Data Model, created our own Image Classification Models with Teachable Machine and ml5.js, played Google’s Quickdraw Pictionary game and, obviously, generated Images and type with midjourney, Dall-E and stable diffusion.

Alongside the seminar, the participating students developed their own point of interest and created various works which we presented at the Future of Gestaltung Symposium at the HfG Karlsruhe in the format of a talk show named AI, You & I.

The seminar took place every two weeks, seven days in total:

Day 1 Prologue AI Dungeon Adventures and definition of contextualization methods

Day 2 Demystification Uncovering Datasets and mathematical explanation of Neural Networks. Google Quickdraw, Teachable Machine, COCO Dataset.

Day 3 Models Training own models, implementation and testing with ml5.js, Image Generation with Midjourney, Model and Data Provenance

Day 4 Hacking: Stable Diffusion Inpainting, Typography

Day 5 Research: Project Outlines

Day 6 Work: Feedback

Day 7 Presentations

Questions to AI

Kimin : An image of sound as seen by a person with the ability to see sound visually.

AI Visualisation by Rosa and Louise

3Blue1Brown Deep Learning Explanation

Object Detection with COCO SD, ml5.js tutorial by Daniel Shiffman

Teachable Machine

Inpainting (by myself) on stable diffusion

generated by Kristian

generated by Lu

“In the earliest morning, where the air cools your face and surfaces are coated with fine films of dew from the fog rising from the cold water on the floor, I hold the tiniest matchbox with the most generic design I found while leaning against the single leafless tree that stands in the middle of the dark link mini boss room from ocarina of time.” — prompt by Lu

Student projects

Artificial Femininity by Julia Ziegler

SIGNING SIGNING SIGNING by Anna Katalin Szilágyi

Uncanny Valley 700BC by Chiara Kern

AI Connection by Kristian Vrhar

by Till Engelhardt

The Future of Creativity by Lu Kellert

Love and AI by Maïmouna Diop

AI for paintings by Róza Velkei

Alternate Creativity by Louise Windenberger

Participants

Francesco Perale

Till Engelhardt

Róza Velkei

Louise Windenberger

Tim Drekopf

Maïmouna Nele Diop

Chiara Kern

Lu Kellert

Joel Luca Sequeira Ferreira

Anna Katalin Szilágyi

Kimin Han

Julia Ziegler

Ressources

Readings

https://nooscope.ai/
Introduction to A.I., data, models and machine learning algorithms

https://kim.hfg-karlsruhe.de/
research group on artificial intelligence and media philosophy KIM (from German: Künstliche Intelligenz und Medienphilosophie) at hfg karlsruhe

https://www.lesswrong.com/tag/ai
"LessWrong is an online forum and community dedicated to improving human reasoning and decision-making"

https://www.alignmentforum.org/
AI Alignment Forum

https://arxiv.org/list/cs.AI/recent
arXiv® is a curated research-sharing platform

https://tacticaltech.org/
Tactical Tech, Berlin-based non-profit organisation

https://excavating.ai/
Excavating AI - The Politics of Images in Machine Learning Training Sets
By Kate Crawford and Trevor Paglen

Machine Unlearning: Decolonial Thinking in the Age of AI

Technical Explanations

COCO
https://cocodataset.org/#explore

3Blue1Brown: But what is a neural network?

Image Classification with ml5

AI Tools

Image Generation

Stable Diffusion

Midjourney
text to image AI; Free Trial, Paid Subscription/Credit System afterwards (Discord Account required)

www.craiyon.com
text to image AI; Free to use directly in the browser, kind of scuffed, more like a budget version of image AI

openai.com/dall-e-2/
creates an image based on your description, style of painting / colors / artistic movement, etc
(you must sign up and you receive 50 free credits)

gaugan.org/gaugan2/
image generator,creates landscapes with the textures you add (glitchy but it makes it funny)