3 Questions for Hito Steyerl

Right now, I am exploring the idea that AI as an infrastructure and a system dwells on creating what Roberto Trillo Alonso and Marek Poliks call surface; a porous interface for capital to create leverage. The idea is that capital nestles in the delta between input and output price; which means it has to create this gap, any gap in order to operate. The delta, in this particular case is the gap between volatility for humans and on the other hand the creation of hugely abstract financial instruments. The more volatile, precarious and unstable the situation for humans, the more options for capital to enter this gap and leverage it via AI economies; AI exploitation or appropriation of any kind.

At the moment, I am exploring the territory of Topusko, Croatia, projected site of a 50bio USD AI data center. This location has been ethnically cleansed twice in the last century by different brands of fascists, sits on the Balkan refugee route as one of its pushback capitals and is economically disadvantaged by depopulation and the lingering effects of the 90s war. On the other hand, a second nature has been left to grow, a wonderful wilderness full of animals, clean water, large forests and life. Now a huge UFO called the Pantheon AI data center, a conglomerate of US investors and local real estate developers, is supposed to become one of Europe´s biggest data centers — unless all this turns out to be a pump and dump enterprise, which is very well possible.

I have analyzed the announcement strategies for data centers as performance art, in this case sovereignty cosplay performance art whereby politicians from the region pretend to have any say in what Western capital plans with it. There are a whole lot of numbers, acronyms and shock and awe jargon destined to lure investors flanked by massive state theater, sanctimonious rituals to abdicate any kind of democratic sovereignty in the face of the arrival of Trump associated pseudo-tech money. All this could be seen as conceptual land art, whereby announcements create a set of instructions that serve – like in Sol LeWitt´s instruction pieces — as the thing itself. Instructions, documents and bureaucratic ceremonies are all you might need; the aesthetics of administration made operational in the service of AI.

I learned about microwork in refugee camps in the MENA region during the pandemic. This was something that was hugely advocated by the world bank during these years, especially also for Palestine. I knew that this work was also being done in Kurdish regions of North Iraq. It took quite a while for our producer to find the microworkers on site and then, filming was disrupted by war. Finally, we managed to get there to record the workers’ recollections. By that time, they had been fired again, machines could now autocaption images.
However, the machine vision they helped develop was installed in drones among many other things and the drones were shooting at people in the same region. One journalist who barely survived a Turkish drone strike is also in the work.

It´s called Lossy, because our aesthetic medium is based on loss functions and scarcity as design choice. It is trained on exclusively CC material and Public domain data, the bulk of which was initially collected by our colleagues Mat Dryhurst and Holly Herndon during their Spawning project. We, that is my colleague Völker Möllenhoff and I, have created several prototypes for a Pixart-alpha diffusion transformer, and came up with new quantum based preprocessing techniques to deal with the issue of human bias in selecting training data.

This project is less about output than process: we discovered that we can save up to 75% energy in training and two thirds in inference with some very simple architectural decisions. We saved even more by exchanging proprietary elements and modules trained on other people’s data with so-called pure math processes in many places. It is a pipeline that demonstrates that scale is often unnecessary and due to economic blackmail, rather than technological necessity. We are getting NVIDIA GPUs shoved down our throats, even though many of them would be unnecessary, as well as their energy and other resource expenses. We still have to make the final decision on how to train the full 1024 model; I think it will be a very abstract, weird quantum leaning aesthetics, not photorealistic at all.