On digestion, going slow and whether the custom AI model still matters

Everything I’ve focused on since the start of this year is parsing though software and subscriptions within this current life. A deep consideration has grown in me for the processes of context held within a pixel’s weight. By inspecting the faucets of data and streams from their source, I can now amplify space for compare & contrast, building a focused delineation into a brighter je ne sais quoi. By developing harnesses (more than just for horses) I find there are endless tuned connections & directions that can be utilized to preserve and develop anew.

I started using ML in 2006 when the only option was to curate your own data-set and train your own model. I actually started using ML specifically to create site-specific work (where my early installations learned from images collected using live cameras of the visual surroundings of the work); for me ML has always been a situated practise, always a process (not a product or object), and always slow (e.g. emphasizing on-line learning).

Seems to me this is more ideological than practical. If you believe “scale is all” and that through sheer brute-force and more data, you’ll get AGI / Sentient AI / Super Intelligent AI, then the big “foundational” models may be the most important thing to you. What is “generalism” but universality? What could make better art than a universal intelligence?

I don’t buy into any of that. To me, every selection is not ‘discovery’, it’s construction (I’m an agential realist); selecting training data is the construction of a reality: the production of a world and point of view. There is no total set of human knowledge, no all encompassing universal training-set, just an implicit point of view that purports not to be one, a situated system that pretends to be universal.

So if you’re like me, training your own models is not a preference nor a aesthetic choice; it’s an acceptance that all data is situated, political, and social. It’s the realization that /mattering/ is a choice. Around every training set is a boundary that determines what and who matters. Pretending that boundary does not exist is a refusal of how knowledge is made and data is constructed. It’s the implicit acceptance of a point of view that purports not to exist; it’s a failure of imagination.

What’s important to you? Is it to engage in a process of deciding what matters, what worlds are produced? Or is it to internalize and validate the worlds produced by big tech rooted in mass extractivism?

I have some experimental writing on this in case folks are interested.

I am on the same page with you about training your own models .

Still.

Or maybe especially now.

When the differences between next best models are so negligible, and when people are tired of the visual stimulus, I’m losing interest in the “interesting image” . Designing the system is where so much more can be done and told.

This is a really good question. I’m now working on a book, soon to be released, that I wrote using different AI systems from 2024 to 2026, about the literary creative process. I distinctly see AI as a digestive organ that I can feed with pieces of literature, philosophical biases, elements of popular culture, and political orientations, and regurgitate fascinating literary content.

As a collector/researcher I think it’s incredibly useful in both ways. Can’t say I lean towards one or the other but I think training is really important (for artists and non-artists).

I lean towards digestion/distillation. I think it’s a tool that requires restraint and ethics, which is not present in everyone, despite everyone having access to it (akin to the internet).

Does “fine-tuning” end up just replacing “training” entirely?