The science of the concrete

This interview with Walter Scheirer, a media scholar who recently wrote a book called A History of Fake Things on the Internet, lands where you would expect, making the general point that we should regard “fake things on the internet not as a radical and terrifying departure from civil norms, but as a wholly natural evolution of our human drive for mythmaking and storytelling.”

That seems more sensible than insisting that technology can or should only be used to produce facts, or that there is some form of datafication or documentation that unproblematically captures “truth.” Scheirer points out that “a lot of these media objects are for human connection. It’s like, I want to share something with you, I want you to understand something about me or my community.” But to communicate that understanding involves something other than a transfer of data; it requires more than transmitting information that exceeds the noise in a signal.

The potential connection depends on the media object demanding interpretation and on the receiver being willing to make that effort. Producing evidence of that effort to understand, to share a basis for an interpretation, makes it “sharing.” If the media object was irrefutable and required no contextualization or speculation about its meaning, if it had no potential for ambiguity or ambivalence, if one couldn’t possibly misinterpret it, then it wouldn’t make for “community” so much as reduce all the parties into relays in a circuit. There has to be noise in the signal for it to matter socially. There has to be some reassurance that in communicating with each other we don’t simply become transparent and vulnerable, susceptible to being fully and permanently processed.

The artists in this roundtable discussion on whether generative images should be considered a kind of photography make similar points. Trevor Paglen argues that “the idea that a photograph, in and of itself, can record some kind of truth has always been a fiction ... It’s not possible to make an unmanipulated image.” Simply putting a frame around something, rendering it static, is a form of manipulation. There is no way to avoid imposing a specific perspective, a subjective vantage point, and no way to subtract that vantage to get back to pure objectivity or “reality.”

Charles Engman notes that “even though I know that images are not true, have never been true, part of me does believe in pictures. Part of my interaction with photographs is a willing suspension of disbelief.” That is, images are intrinsically fictional; they are occasions for willing belief, not for being exempted from having a subjective viewpoint, a subjectivity.

Generative models often seem like a misguided effort to fully rationalize and predict the way language and thought unfold, but in practice they seem to incubate doubt, undermining a reflexive faith in whatever happens to be documented and directing more attention toward the rhetorical nature of all documentation. The effort that belief requires becomes more palpable, and so more deliberate.

In the interview, Scheirer cites anthropologist Claude Lévi-Strauss to get at the dialectic of science and myth-making, of disenchanting rationalism and the re-enchanting processes that counter it. Deepfakes and the like could be understood as examples of bricolage in action — making do with the materials at hand to articulate distinctions that remain socially important and which don’t simply emerge from empirical observation. “The distinctive characteristic of mythical thought is to express itself with the help of a set of heterogenous elements,” Lévi-Strauss writes in La Pensée sauvage, “one that, even if extensive, still remains limited.” I sometimes tend to think of the models themselves as the ultimate bricoleurs, making use of whatever data from whatever source to do whatever aim, but that models themselves don’t have any aims and can’t do bricolage by definition. Sometimes it seems like they are being extremely resourceful in accomplishing tasks we have set for them, but that is just projection. Instead, we are the bricoleurs and the model is one of the tools at hand for us to improvise with to accomplish the ad hoc ends of myth making.

It doesn’t take much mental acrobatics to adapt Lévi-Strauss’s description of the bricoleur at work to tinkering around with generative models:

excited by his project, his first practical move is retrospective; he must turn back to an already constituted set, consisting of tools and materials; inventory or reinventory it; finally, and above all, engage in a kind of dialogue with it, in order to identify the responses that the set can offer to his problem, before choosing among these possible solutions.

An LLM could be seen as a way of “engaging in a dialogue” with the “constituted set” of its training data and extracting from it a range of permutations that might be useful for the problem at hand. But this is not because the permutations will be scientifically valid solutions or facts, but because they will be “at hand” in a way that makes them useful and socially efficacious. The model doesn’t provide facts; it can’t reason or make rational claims. But it can readily reshuffle the existing inventory of language and inspire myths as needed.

Generative models are sometimes hyped as offering unlimited possibilities to prompters, but it might make more sense to think of them as presenting a codified set of constraints. Lévi-Strauss continues:

Like the constitutive units of myth, whose possible combinations are limited by the fact that they are borrowed from language, in which they already have a meaning that restricts their freedom of maneuver, the elements the bricoleur collects and uses are “preconstrained.” In addition, the decision of what to use to do what depends on the possibility of permuting another element into the empty function, to the point that each choice will entail a total reorganization of the structure, which will never be exactly the same as the one vaguely imagined, nor the same as any other that might have been preferred to it.

LLMs encode those preconstraints and let us assume the position of “vaguely imagining” how that whole structure is composed (even if our prompting doesn’t yet trigger a “total reorganization”). They force us into the trial and error posture of the bricoleur, which could amount to a reintegration into the world of myth, where improvised explanations and accounts carry more weight than the scientific method.

Anyway, sorry for the confused and improvisatory nature of these notes — I really only wanted to post to say thank you for reading and subscribing this year and for helping provide some sort of structure to my life in a period where I can’t find any regular work. Hopefully things will be a little more structured in 2024.