re: #279 danarchy

If you can only build an AI that makes pharmaceutical discoveries by scraping the work of scientists and researchers uncompensated is that shoddy? It is called learning and if the AI is making derivative works based on things it has learned and not just copying things then to me it is no different than a person being inspired by music they hear. Now if it is just copying that is a different story.

There’s a reason art is called art and science is called science.

Science is defined by derivation, iteration, and cross-referencing. That’s the basis of the scientific method: almost-fractal inquiry into existing ideas while consciously documenting innovation versus derivation while providing explanations for those choices.

In science, endless permutation on a constrained data-set is valuable as a form of automation, but the things generated will then have to be evaluated for real-life applicability because even within a constrained set of rules like organic chemistry or protein folding, not ever thing that’s technically possible is possible to make in vitro, and what’s possible to make in vitro may have complications that keep it from application in vivo.

Art is self-expression.

Lots of art is derivative, but derivation is consciously chosen by the artist and thus derivation itself is a kind of self-expression. Lots of people iterate on Goya’s “Saturn devours his child” and are still making art. Fanfiction writers make art. Generally derivation in art is only considered “bad” when it’s concealed—artists using tracing to copy another artist’s work while presenting it as original, a thing that crops up in comics—or when it can be argued that the finished work doesn’t transform the original enough…and that’s always a debate: look at the criticisms of Roy Lichtenstein.

AI is heavily doing the first: it copies whether or not it’s prompted to copy because it has no capacity to do technique, only spreadsheet of correlations between prompts and a bank of labelled images that can be used to generate a response to the prompt.

Art is also a mix of technical skill and implicit understanding, techne and metis, such that part of art is the making itself. A sizable chunk of abstract art is artists performing technical perfection discarding the expectation of representation; another major chunk of art is expressionist, with technique and representation being de-prioritized to convey feeling. Art constantly changes and expands via a warp and weft of technical qualification balanced against intuition and want.

AI has no technique beyond technique as a series of labels within it’s database, it literally cannot have a style of it’s own. A prompter’s ability to innovate is constrained by the labelling system that allows the generator to function. The two cannot create something technically new, only iterate on what is.

But more deeply…no, current AIs are not learning, they have no consciousness, they have no imagination, and they don’t even have fidelity. That last one is particularly notable in current generative AI: the thing doesn’t understand what it’s painting, only the relationship between the tagged images in it’s database and the keywords used by the prompter. Lack of fidelity is what creates “AI hallucinations”: it loses track of where it is in the spreadsheet and can’t correct itself because it doesn’t know it’s drawing a hand and a hand has five fingers. Advances in fidelity between generator-AI generations is largely a product of refining the spreadsheet by using more tags, which is done by low-paid workers in various developing-world countries.