Tracing Timelines in AI
It seems strange to trace timelines of genAI explorations - but the last few hype cycles have been so fast and churned out so much. After a year (little over) in this world, I think it's important to do a quick review. This post is more for me to reflect on my journey, but happy to hear your thoughts on it. I don't have a comment section on this blog but feel free to DM me.
I have been dabbling in AI projects since early 2018, executing small projects/tools/experiments with posenet, LSTM, and more.👵 But this post is traces works and insights from August 2022 to present (date of the publishing the post).
My perspective is critical, technical and casual all together. Those who know my creative and critical practice are aware that I have a strong focus on care and motherhood. This doesn't mean that it is completely artistic work, the focus allows me to bring an exceptional set of mental models to LLM and AI prompting. I engage strongly with dataset biases and my background allows me to identify gaps and faults with speed and grace!
My perspective is critical, technical and casual all together.
The work starts with a focus on understanding prompting for Disco Diffusion. The buzz was around using "styles" of aritsts to access the diffusion outputs. At the time, I was looking at the care-work and motherhood and was interested to see if the "styles" of women artists would appear in the outputs. I quickly say that works of famous western artists even if they were women, are visualised by the diffusion models well. It struggles with non-western canons. There could be many reasons for this, I suspect the key one that India artists are not as well documented by western museums, many of whom are under the GLAM-Wikimedia Collaboration - an easy source to accumulate in the difussion dataset.
However that is interesting is that Disco Diffusion outputs are very "artistic" and highly stylised in their own right - a beautiful side effect of the models inaccuracies and failings. This allows for the pieces to be wonderfully fanastatical, yet retain detail and emotions.
I continued the prompting attempts with image generation with Dall-E. These show racial bias and also a poorer inference of the sentence structure.
Screenshots from Dall-E history from Aug-Sep 2022
Loves and cares about deep, caring and evolved entanglements with AI.