There was a part in the recent Chamber of Progress webinar where I tried to express a point about ML/AI in its current form’s training being made up of what I think of as a ‘measurement of dimension.’ That is, that the training process is composed of – in my limited understanding – sets of measurements of attributes or features within items in the datasets. This seems to be called its dimensionality (its amount of attributes), though these terms are used it seems in different ways by different sources.

In the video, I refer to it as a measurement of dimension. Because that makes as much sense to me linguistically as any of this made up gobbledeegook AI jargon does.

But we could also reasonably I think in plain language say characteristics and still be on the right track. ML/AI training looks at characteristics in data: images, text, etc. That means splitting apart things at a kind of micro level. If you look at 10,000 images of dogs, or bananas, or anything, you can start to break out lists of attributes, size, color, shape, and so on. Texture.

The beauty part then is that the characteristics originally measured in specific artifacts in the training corpora then become abstracted or put in a blender, so to speak, divorced from their original meaning and context. Enabling them to be mixed up and re-assembled into completely new, never-before-seen creations, assemblages of statistical sampling.

Elsewhere earlier in the video, we talked a bit about styles. There seems to be general consensus online that style is not copyrightable. As expressed by Creative Commons here, copyright protects specific expressions in fixed forms of a style, not the style itself. (But there are, ahem, blurred lines here, to be fair.)

Anyway, my point here isn’t to argue on copyright grounds in this post, that is just for background. What I wanted to say I guess is that ultimately my intuition as an artist who has made extensive use of generative AI is that style ultimately breaks down to supersets of characteristics/attributes/features(/dimensions).

To an AI image generator, for example, you could just as simply prompt “in the style of salvador dali” as you could basically any random string of characters “in the style of day old bagels at a long island supermarket in the 1980s” or “in the style of z8gn3s8hjkl*12%”.

All of those are going to give you results, some more inscrutable than the others as to clear cultural meaning or obvious connections to known things. You can also do it without including “in the style of” at all, and just including random phrases or strings in your prompts, as Nettrice talks about in the video.

This fluidity and flexibility around the sheer fact that literally anything (or nothing) can become a “style” in AI for me blows apart our current understanding of what an artist’s style conventionally even means. (Let alone what this means for ‘consensus reality’ – and this is what I was referring to I think in the video, re: holistic frameworks for analysis of meta-data.)

Anyway, I have other things waiting on me, but wanted to jot this down as a beach head to be able to come back and grasp this idea more fully down the road. Also see the notion of the hypercanvas here. Somewhere, somehow, I trust that this will all make sense in the end…

Related writing & research