9. On Interpretation
This chapter takes up the verb interpret and looks at the range of suffixes that it may take: -ant, -er, -ation, -ability. Just as machine interpretation (as grounded in parameters) may be contrasted with human interpretation (as grounded in values), mechanistic interpretability may be contrasted with humanistic interpretability. I discuss what it means to prompt “persons” into being, as well as how best to interpret texts that were generated by machines.
Interpretability
“Who holds what beliefs, for example, regarding the meaning of or reasons for JFK’s death, global warming, the Second Amendment, a Balinese cockfight, or 9/11?
”
Recall the definition of an interpretant: whatever a sign creates insofar as it is taken to stand for an object—for example, calling on someone when they raise their hand, a French translation of a German sentence, or a response to a prompt. As was shown, such interpretants should not be confused with interpreters, understood as the agents—humans, machines, or otherwise—capable of interpreting such signs.
Both of these concepts should be distinguished from interpretation. In one sense, interpretation is simply the act or process of producing an interpretant. As was shown, the processes underlying machinic interpretation are quite different from the processes underlying human interpretation, even if the outputs of machines more and more come to match the instigations of humans. In other words, while machinic agents and human agents take very different paths, as it were, they can now arrive at similar destinations. At least insofar as the parameters of the former are aligned with the values of the latter, if only as channeled and distorted by the power plays of corporate agents.
In another sense, an interpretation is a particular kind of interpretant, one that attempts to explain a behavior (interaction, affect, text, institution, or event) by reference to the underlying motivations of the agents that produced it or to the larger context in which it was produced. Such interpretations can be shallower or deeper, from why he addressed her in that tone to who could have bewitched us. Different agents may find the same interpretation more or less plausible, depending on their values or grounds (understood as guiding principles): from Freud’s interpretation of dreams (turning on repressed wishes, the oedipal conflict, and certain hard-to-stomach symbolic conventions) to everyday explanations of boorish behavior (which might involve reference to the offending person’s upbringing, politics, drinking habits, or mood). And as a function of their plausibility to different kinds of people, certain interpretations can become canonical or remain contentious. Who holds what beliefs, for example, regarding the meaning of or reasons for JFK’s death, global warming, the Second Amendment, a Balinese cockfight, or 9/11?
As used here, interpretability refers to the conditions of possibility for a sign to be interpreted, or to the conditions for an entity or event to be treated as a sign in the first place, such that it might constitute a lure for interpretation. Such conditions can be quite unremarkable: often simply a semiotic agent with particular values (or parameters) is needed. For example, you may quickly and unconsciously parse the meaning of many utterances—if only to ignore them insofar as they are not addressed to you—just by knowing the language in question. But such conditions may also be more subtle. For example, an agent might be unwilling to undertake the work of interpretation unless there is a promise that the sign is decodable or that the object of the sign is relevant to the agent. Do they possess the key to unlock the safe, as it were, and is the secret contained inside as yet unknown and of value to them? What kind of person tries to get to the bottom of a passage by Joyce, an ancient text, a crazy dream, a strange signal from outer space, or an awkward kiss? What do they feel they stand to gain by interpreting the sign, and why do they believe they have the capacity to do so? In such cases, one can inquire into the genealogy of the demand on the agent, or the desire of the agent, to attempt an interpretation. In effect, one can offer an interpretation of interpretability.
Mechanistic interpretability, in contrast to the more humanistic modes just discussed, refers to the ability to analyze the parameter values in a trained language model, or a neural network more generally, in order to reverse engineer the complicated function that was learned by the model. In a certain sense it is the attempt to explain, and thereby better understand, the behavior of a machinic agent, which may be otherwise opaque to those who created and trained it, given the complicated workings of its architecture and the enormous number of parameters contained therein.
Mechanistic interpretability is sometimes contrasted with algorithmic transparency, which endeavors to make visible the factors that contribute to algorithmic decisions so that those affected by such systems (or those who use and regulate such systems) can better understand why a particular decision was made. Why was this song recommended to me? Why was my loan application refused? Why is his DNA considered a match? Why do I keep seeing this advertisement? Like algorithmic transparency, mechanistic interpretability is potentially useful, ethical, and profitable, insofar as it serves to make language models and their outputs more predictable, reliable, robust, modifiable, repairable, alignable, resistant to malicious hacks (or amenable to playful ones), and so forth.
But such an understanding has not yet been attained. In particular, even though humans designed and trained the language models and such models perform a function that mimics human behavior, humans do not yet fully understand how the models actually work—in the sense of which parameter values contribute to which aspects of their overall behavior. And thus they do not yet know which parameters of a model to alter when the model produces lackluster, odd, incorrect, or harmful interpretants. In short, while the behavior of human agents is more or less humanistically interpretable (given some work), the behavior of most machinic agents is not yet mechanistically interpretable. They just seem to work.
Prompting Persons Into Being
With all the foregoing considerations in mind, it is useful to pose a simple question: In what sense are the outputs of machinic agents—in particular, the texts that large language models generate—(post) humanistically interpretable? Phrased another way: When, and in what sense, do the interpretants of such agents, and hence their textual outputs, warrant an interpretation?
Insofar as machinic agents involve at least three kinds of derivative intentionality (via the human agents who wrote the texts they were trained on, the human agents who trained them and stipulate satisfaction criteria, and the human agents who interact with them once trained), the outputs of those agents are certainly worthy of humanistic interpretation. In other words, one can analyze the meaning of such texts, the motivations behind their creation, and the contexts in which they were created, insofar as they are parasitic on the motivations and meanings, as well as the contexts and cotexts, of such human agents—which includes the profit motives and ethical qualms of the corporate agents that spearheaded their creation.
Even setting aside intentionality, derivative or original, such texts were selected by minimizing cross-entropy loss or maximizing a reward signal, and by means of all the other modes of sieving and serendipity, and generativity more broadly, that went into their creation. (One could even argue that any process that minimizes or maximizes a function—and thereby takes it to an extreme—involves a glimmer of telos, if not a speck of intentionality. The second law of thermodynamics, as reformulated by Gibbs, is in a certain sense the origins of desire.) And so, as for any other living kind, one can offer a genealogy of their coming to be: the sorts of conditions and forces, from the size of silicon chips to the strivings of speculative capital, that contributed to their emergence.
If it is satisfying for someone to interpret such texts, however superficially or deeply, does it matter if they are the product of an aesthetic intention or existential motivation, however unconscious, any more than any other text one is compelled to interpret without reference to an authorial intention? Hermeneutics has long been done without reference to authors.
Indeed, the process often works the other way, and reciprocally so. Just as we can learn about a “person”—their identity, interests, and origins— by reading what they wrote, we can interpret what they wrote by reference to who they are as people. Bootstrapping processes may occur whereby people project not just personhood but also particular personalities onto large language models (given what they write, and how they respond). And then, by reference to such modes of personhood, people will reinterpret what such agents have written and how they respond. And those people may then prompt such personifications in new ways, looking for confirmation of their projections. Our understanding of Jesus and other (mainly) textually present people is not too different in its construction—if only as prompted through prayers. And so not just sects but also whole societies may performatively prompt such machinic persons into being—and thereby make not just their interpretations of texts but also their projections of personhood true, or at least true enough for the people—and “persons”—in question. And so yet another way for the world to be re-enchanted.
Indeed, setting aside any specific text it produces, the generative capacity of any language model is itself worthy of interpretations—no less than any dictionary or grammar, by linguistic analysis or otherwise. Think, for example, of the celebrated claims of the Italian humanist Giambattista Vico regarding the importance of Homeric texts: they contain “models or ideological portraits which form mental dictionaries of the ancients.” Just add the words grammar and pragmatics to dictionaries in this quote, and change ancients to moderns (plus any prefix you might desire), and you are ready to plumb the generative depths of language models. In other words, language models are an incredible resource for studying the values of the people who wrote the texts they were trained on (not to mention the interests of those who fine-tuned the models to satisfy their alignment criteria). Such values include a collectivity’s model of the world and all that it contains— however biased, irrational, or culture-bound. In a certain sense, however Borgesian, language models do not just contain all the texts that were written by a people, they contain all the texts that could have been written by that people given their worldview.
And, of course, this essay is essentially an interpretation and/or genealogy of language models per se, as well as a guide for how to approach the interpretation of any text they might generate, not to mention the motivation and interests of the agents (AV, Aθ, AP) that had a hand—or at least a say—in their creation.
