a bid, adieu

My husband died of an accidental overdose in March 2022.

An unexpected death can defamiliarize the basic components of everyday life. After my husband’s, I acquired a particular sensitivity to counterfactuals. I found their arbitrariness maddening. For a time I had trouble speaking or writing in the subjunctive. Phrases like “if only”, “he would’ve wanted”, “I wish I had”—all died on the tongue as I scrambled for still-valid ways of speaking about the world. A world in which his batch had not been tainted, in which I’d woken to find him in time, seemed no more proximate than a world in which we’d never met, where I’d never been born. Of the uncountably many axes on which fate might hinge, why give preference to any impossibility over another, eliding the simple, gaping wound of his absence? Often, I uncharitably thought, people who did so needlessly milked the situation for tragedy. Is all this not sad enough? Thus alienated from the ordinary patterns of social engagement and private thought, I began to think of my old life as having ended as well, and of the present as an airless waiting room in which logic and grammar would need to be reconstituted from scratch.

I also struggled to comprehend the probabilistic nature of his death. It felt both foreshadowed and inconceivable. By Spring 2022, overdoses had been rising for years. Celebrity deaths by overdose came with unforgiving regularity. There were rumors of fent in the local supply. In light of this, was my private cataclysm unsurprising? It was a sickening thought, no less so than its opposite, that he had simply been uncareful and unlucky. Unluck brought to mind spilled coffee, pratfalls. It was too flimsy a notion on which to rest the reality that I will never again meet the gaze of those proud, wounded eyes, never hear that irreducible melange of lyricism and chatspeak and transliterated Macedonian, as long as I live.

In the week before the funeral, as my fingers tapped out bcc’s by the dozens and my body ran the rounds from bank to cemetery to salon (to bank) to funeral home, my mind did what it reliably does in moments of danger, moving up, up, over everything and into abstraction. In my head, I burned through thousands of theories of the self, of him, of me, of us, our story (a tragedy? a Lifetime movie?). I tried and acquitted myself over and over. Like a cursed game of fruit ninja, it gave me comfort to cut through arguments, slicing through the space of fine distinctions. My conclusions were all variations on what it was impossible to know. But no matter how much I reasoned, one feeling persisted, a totalizing bafflement, a white-hot vicarious indignation at the fact that my love could be felled by something so insubstantial as chance. I carried it with me like a grudge, like a scrim overlaying the world as I returned to work, to life. I used it as fuel to inspect what I came to see as two parallel mysteries of my life and work, the noetic quality of the stories we spin in grief, and the givenness of probabilistic conclusions. The ignition turned out to be a book, dropped into my life like a plot device weeks before.

A week before my husband died, a friend invited me to participate in a virtual book club on May Day for a book (appropriately) called Revolutionary Mathematics, by Justin Joque, on the role that interpretations of probability play in the functioning of modern capitalism. I immediately placed a hold for it through the public library. In the last days of April, neck deep in waves of grief and its processing, I chose this book club to be my unofficial return to social life. I read the book in a few short sprints.

I recall almost nothing from the event besides the uncanniness of seeing myself in that familiar grid of faces. What remains in mind from the weeks surrounding is simply the sense that something core to my life and work was being revealed.

---

My husband and I were tech workers. I work in applied machine learning, and tend to spend the majority of my days running experiments, thinking about their theoretical basis. If you, too, work in tech, you are likely engaged in the production of empirical knowledge and “analytic insight.” Perhaps you make digital widgets and algorithmic changes that must be sanctified through A/B or UX testing. Perhaps you design, run, or analyze these experiments, interpreting their findings to make them “actionable.” Empiricism and the language of experimental science is the unspoken God of a modern tech company, to whose neutral authority all must appeal.

Like any religion, this empiricism rests on articles of faith, in this case the assumption that experimental results reflect reality, and the arc of prediction bends toward truth. But when the corporations running these experiments increasingly mediate our relationship to the world and to others, it can be more accurate to say that these experiments and their consequences bend reality itself. This move from epistemic power to sheer world-making power is mirrored by recent developments in the applications of statistical inference.

“Ghosts of Departed Quantities”

This helplessly romantic phrase is lifted from the philosopher Bishop Berkeley’s critique of calculus, to which he objected because it relied on what he saw as an irresolvable contradiction: the delta that famously goes to zero is treated as both finite and infinite in the theory's derivation. For Joque, this is exemplary of the sleights of hand at the center of so many powerful technologies:¹ the finite infinitesimal and The Calculus, the fungibility of commodities and labor and Capitalism, the Holy Trinity and Catholicism. I would petition to add to this list the idea that other minds are knowable by analogy to our own, and human connection.

In pointing out these “equalizations of inequalities,” Joque’s intention is not to debunk their premises and dismiss their consequences. In fact, he insists repeatedly that factuality is irrelevant when each of these technologies has plainly remade the world. Calling something out as a useful myth may seem pejorative, but I too have come to see these impossibilities at the heart of everything as morally neutral and necessary.

After all, the truth of the world is one of infinite detail, infinite difference. At a certain resolution, nothing is like anything else. This is so obvious as to be a truism, and yet, in an age of rampant scientism and compulsive pattern recognition, it bears acknowledging. To find patterns, as we (like any animal) are compelled to do, we must view the world through simplified models that collapse its differences in myriad ways.

The same collapsing of difference is asked of us in human relations. At base, all of us are (generously) atomic subjectivities stranded in our own minds. To believe another’s testimony is an act of faith. To act as though their experience of pain and pleasure were something like your own is another. Your purple is only my purple because this evidenceless bid is what underwrites the possibility of human connection.

Framing

Consider the example of a matchmaking app for dog owners and dog walkers. To be legible to the machine, each owner and walker must first be reduced to a set of features within a predefined ontology: name, demographic, weekly schedule, dogs owned, special qualifications. This renders certain users indistinguishable from one another, and naturally groups individual users into reference classes (male dog walkers in their 20s, homebound elderly dog owners), around which data can be aggregated and about which predictions can be made.

Suppose you wanted to predict the likelihood of cancellation for dog walking appointments. Any given appointment will either be canceled, or it won’t. Fractional probabilities only emerge at the level of the reference class, and are contingent on choice of class. Consider what happens when a walker decides to declare their gender on the app. Without any corresponding change in the real world, this act of divulgence sorts them into a new reference class, with different likelihoods of cancellation.

Probability is not an essential feature of the specific case. Suppose you placed me in a reference class with female-identifying Asian American professionals in their 30s. The probability I’d have lost a spouse to overdose would be exceedingly low. On the other hand, if you chose the reference class of men in their 30s with chronic pain, the same event somehow acquires greater likelihood. Probability is not inscribed on the soul. It’s an artifact of framing.

Objectification

In the case of our dog walking app, historical walks with known cancellation status would be said to have “ground truth,” and the patterns they establish may be used to impute cancellation probabilities for appointments of unknown status, either through a frequentist or Bayesian lens.

In the frequentist interpretation, data are assumed to follow the metaphor of a coin toss—random, repeatable, drawn from a stable underlying distribution (and, in turn, used to impute its shape). This enables the probabilistic assessment of propositions like “male walkers are more likely than females to cancel an appointment” against the hypothesis that both classes are equally likely to cancel, using something called Fisher’s exact test. This test yields a score called the p-value that then determines whether the findings can be said to have “statistical significance.”²

What Fisher’s interpretation doesn’t account for is the reality that probabilities are often used to inform decisions (e.g. booking a potentially flaky dog walker) for which false positives and false negatives can have vastly different costs. Assuming the “positive” class is probable cancellations: for a dog owner, a false positive would unfairly dock a reliable walker, while a false negative could ruin an evening. In reinterpreting probability to take the relative costs of typed errors into account, Jerzey Neyman and Egon Pearson reframed the central question of testing from one of inductive inference about a proposition to inductive behavior justified through the minimization of expected cost. This naturally made statistics much more useful to industry, where costs are explicit and important, and where profit trumps truth (which becomes at most a proxy measure for profit).

Bayesian statistics is the next step in the turn from truth to effectiveness. Where subjectivity is smuggled into frequentist probabilities by way of framing, the Bayesian interpretation of probability makes this move explicit with the use of priors.³ This feature of built-in subjectivity is actually what makes Bayesian statistics well-suited to digital capitalism. That’s because the closed form of Bayes’ rule⁴ makes it a recipe⁵ for mechanistically churning data into probabilistic beliefs about the world, upon which actions like assessing potential matches, pricing a walk, bidding on ads for the app can in turn be automated.

These probabilities are impossible to validate in the same way that frequentist claims can be (with ever more coin flips). After all, how would one verify the claim that “a walker X will cancel their next appointment with probability Y%” without circularly referring back to how the Y% was derived in the first place? Instead, Bayesian probabilities prove their worth by enabling profitable business decisions, informing which buttons to shade what color, which walker profiles to show at what position, and which version of the copy to display, for maximal profit or user engagement.

The corporate use of machine learning models (many, though not all, of which rely on statistical inference) exemplifies the Marxist concept of objectification. To quote Georg Lukács, objectification occurs at the point when “a relation between people has taken on the character of a thing.” Machine learning models organize the mess of human relations into a unified decision-making apparatus, self-justified through its enabling of economic productivity. Were the matchmaking performed by humans, walkers who aren’t getting matches would have a natural audience for their grievances. But when it’s done with a machine-learned algorithm, the responsibility for each decision becomes relocated diffusely within the matching algorithm. Even the data scientist who trained the model would not be able to explain why a matching decision was made without falling back on aggregate markers of profitability and effectiveness, which give no recourse to the individual case.

The tendency for a centralizing objectification to grant something the illusion of truth is hardly limited to the corporate world. I have come to see the unified self as another useful illusion, given its objective force by the symbol of the physical body.

The morning of my husband’s death, as policemen still paced our living room, messages of heartrending mundanity began trickling into my husband’s phone at daybreak. As the hours passed and news of his death began to spread through our networks, their pace slowed but didn’t stop. On iMessage, WhatsApp, Signal, Discord, his friends wanted to say goodbye along the channels through which they knew him best. I tried to look away. I didn’t always succeed. As the week went on, I often saw in my mind an image of hundreds of strings coming untied, floating free. To what had they all been tethered? What had changed? People were still thinking and acting in relation to the image of him held in their minds, projecting onto him so many mutually incompatible narratives, just as they did when he was alive. Even his body remained very much in and of the physical world; I was managing its location every day, giving and neurotically revising instructions on its preparation for burial. What was different now was only that it was no longer the centralizing locus for engagement with him as a person. What is personhood, but the wishful conflation of disparate social and personal identities across time with a body? What is the body, if not the objectification of personhood, a symbol for the hope that those incompatibilities will one day be resolved? And what is this essay, but a dream of another place of attentional gathering for all those who loved my husband?

Infinite detail

Like capitalism, like catholicism, the edifice of human relations rests on a floated premise. A bit. A bid. And the paradox of coming to know a person intimately is that this bid for connection wins out spectacularly, even as their difference from oneself comes into greatest relief. Indulge me, now, in some specificities:

My husband’s name was Damian. This is irrelevant, or perhaps it isn’t. Surely, a Damian is more likely to break your heart?

We have a son, born eight months prior. In the delirious days of my extended home labor, Damian coached my breathing, timed my contractions, played me silent nature documentaries soundtracked with ambient music, patiently rubber-ducked as I picked fights with my doula about birth studies over email.⁶

In conversation, Damian spoke with the solicitousness of a concerned older brother, snuck thoughtful compliments under the guise of obvious fact, gently flexed his capacious stack.⁷ In his company, one could pretend for the moment that of course we started from the same priors (and if not, perhaps it wasn’t too late to co-opt his). One of his favored constructions was “it didn’t need to be so [brilliant/delicious/etc], but it was.” This refusal to take for granted was at the heart of so much I loved about him—his reflexive sense of gratitude, his capacity for wonder. As for what he felt he didn’t deserve, I can only ever speculate. Expectation/reality, setup/punchline, premise/consequence. Abstraction. Reality. It’s all continuous, isn’t it? Continuous even with that most human act, of suddenly seeing in media res that things could be otherwise, and choosing to follow through anyway. Acknowledging the premise as moth-eaten and beleaguered while seeing the good that is possible in living as though it were true. Recognizing the polite fiction of a coherent self, capable of making and keeping promises. Marrying them anyway.

In the dreamlike hours after Damian’s death, I remember tapping automatically into the baby monitor app to find the last video I have of him, cooing our son back to sleep in the early hours of morning. I remember opening the trash can, and finding a fresh bag, its only contents a clouded vial with a green apple decal. Trash duty was Damian’s job, and pickup was in the morning. From the infinite detail of a life, I suppose I’m picking these out to convince you that, though he was troubled, he was a good father and a good man, knowing neither you nor quite what makes a person good. I am placing him in a reference class of good men and good fathers, so that his death might be viewed as an anomaly, and to thus properly validate my sense of singular devastation.

The night after Damian died, gathering photos for his memorial website, I found myself grieving our past selves. We’d been together 7 years, time enough to trace the slow slackening of skin and creep of crows’ feet. For him, the photos stretched back to infancy, tracking the golden child, the loving son and brother, the surly teen. A years-long gap stood in for post-grad lostness, followed by my ample documentation of our relationship—friends, trips, marriage and parenthood. I scanned the timeline back and forth, rewinding, replaying, looking for signs of hope and doom.

At some point, hours into the night, I realized that this fresh wave of grief bore no relation to Damian’s dying. The selves pictured in those photographs were as lost to me as he was, not 48 hours since his death. To mourn their loss as though they were myself was another willful erasure of difference—the difference between the collection of cells that bears my name in any given moment and the person captured in an old photograph. “Damian” is gone, and with him the need for that worn-out fiction of the unified self, for chronology and the stories that depend upon it. Unbound by narrative and the arrow of time—regret, disappointment, nostalgia are left to recede back into meaninglessness. All there is, is the moment:

Notes

Footnotes

1. Here, I am interpreting the word broadly to mean anything that extends the reach of human agency. ↩

2. P-values measure the likelihood of observing the experimental data assuming no difference between the classes (the “null hypothesis”). Fisher selected the arbitrary upper-bound of 5% as a heuristic for establishing statistical nonsignificance, never intending its inversion into the rubber stamp of scientifically ratified Knowledge that it has since become. ↩

3. The prior probability distribution for an event is the assumed probability distribution assigned before any empirical evidence is gathered. ↩

4. P(A|B) = P(B|A)*P(A)/P(B), meaning that the probability of hypothesis A given evidence B equals the likelihood of observing evidence B assuming hypothesis A times the ratio of prior probabilities, P(A)/P(B). ↩

5. For the sake of brevity, this explanation obscures the role of computational methods in making calculations of Bayesian posteriors tractable, enabling its popularity in industry today. ↩

6. To quote from one: “In general I am more skeptical of observational/retrospective studies because of how much heterogeneous historical data is conflated together in the analysis and the risk of confounding variables.” ↩

7. A stack being the first-in-last-out data structure that enables an interlocutor to follow along one conversational tangent after another, only to gently remind you, hours on, to finish your original story. ↩

---