Indifference Engine

by Misha Lepetic

“On the way from mythology to logistics…
machinery disables men even as it nurtures them.”
~ Adorno & Horkheimer

Mid-century-furniture-jacobsen-egg-chairA few years ago I heard the Seattle Symphony play Carnegie Hall here in New York. There were three pieces on the program. The first two – Claude Debussy's La Mer and John Luther Adams's Become Ocean – are clearly of a type. They share the subject matter of the sea and its sonic representation. More importantly, Become Ocean is a clear stylistic descendant of Debussy's seminal, impressionistic work. Written a hundred years apart, both pieces nevertheless explore shimmering textures and slowly shifting planes of sound. The emphasis is not on seafaring – a human activity – but rather on the elemental qualities of the ocean. So far, so good.

The third selection, however, was Edgar Varèse's Déserts. As the title implies, Déserts is possessed of its own vastness, but this is an expanse that is jagged and abrasive. Written in the early 1950s, that is at about half-way between La Mer and Become Ocean, its exploration of timbre is arid and dissonant, and is an early example of a score that calls for interweaving the ensemble's playing with pre-recorded electronic music. Some listeners may be reminded of avant-garde movie music where the scene calls for danger and uncertainty; one YouTube commentator wrote that “parts of this remind me of the music on Star Trek, when Kirk is facing some Alien on a barren world, kind of thing”.

Varèse has always been a favorite of mine when it comes to the canon of twentieth-century “new” music. Prickly and uncompromising, he was a passionate and broad-ranging thinker. After meeting him for a possible collaboration, Henry Miller mused that “Some men, and Varèse is one of them, are like dynamite.” Indeed, Varèse envisioned Déserts to be accompanied by a film montage – what we would casually characterize today as a multimedia experience. While the pitch to Walt Disney never went anywhere, the music is still with us today. But be that as it may, what is Déserts doing, sharing the stage with the marine masterpieces of Debussy and Adams?


As a counterfactual, had an algorithm been curating the evening, we would certainly not have had this juxtaposition. (Perhaps, in keeping with the evening's theme, we would have been subjected to Handel's Water Music instead). You may contend that it's absurd to think of an algorithm holding such sway over the well-heeled patrons of Carnegie Hall, but consider how much of our lives have been pervaded by exactly this sort of machine-driven ‘curation'.

Hans_Wegner_Arm_Chair_CH_25_in_smoked_oak_and_new_papercord__602_lSo here is a seemingly uncontroversial claim: one of the great triumphs of modern software is the recommendation engine. From Amazon's ‘customers who bought this item also bought' to Netflix's ‘other movies you might enjoy', recommendation engines are ubiquitous and always ready to help, especially when they are wrapped up in the soothing tones of a Siri or an Alexa. Recommendation engines also occupy an interesting niche in our information ecosystem. In a world of infinite content, they are the osmotic membrane that regulates the exchange of data and preference. And they thrive on scale: the more data is thrown at them, and the larger the network of users, the better they function. This is true for both the raw inputs (what's available to be consumed) and the raw outputs (what is consumed). The design masterstroke of this paradigm is that the outputs are converted into new inputs. By simultaneously taming and leveraging the deluge of data, recommendation engines aspire to make a cornucopia of choice legible to us.

But this design masterstroke is also its fatal flaw. Recommendation engines are really good at homogeneity. If you like salsa music, start a Pandora station and keep giving the thumbs-up until you've locked in your sound. You can be sure there won't be any heavy metal songs popping up in your stream. Need more than one cordless drill? Amazon has got you covered. Have a look at some drill bits while you're around. And let's not even get started on the abundance of potential partners that online dating apps dangle before us. It brings to mind MTV's original catchphrase, “Too Much Is Never Enough.”

They watch over us, these engines of loving grace. Obviously, sometimes you just need a cordless drill, in which case these platforms can be of great utility. But more broadly, what they are poorly equipped to do is help us understand or create the idea of difference. Or to be more specific, the hedgehog-like nature of this worldview gently deflects us away from the idea that there are styles, stances and yes, objects, that are not contiguous to our own narrowly formed desires and preconceptions. How do we find these, or rather, how do we acquire the critical apparatus by which we can find and judge these?

This question is worth elaborating because recommendation engines are growing beyond the simply utilitarian, and making a bid to occupy a more nuanced space. Here is an example from the world of design. In a recent article about the curious ubiquity of the so-called mid-century modern style, Kelsey Campbell-Dollaghan notes that

It may also be the product of a great averaging: as algorithms track our preferences and shape our online lives accordingly, we're all becoming more and more similar. Siri and Alexa, for example, are killing off regional accents. Facebook crafts our news feeds so they match up to what it knows we already love and hate. Companies like Airbnb and WeWork are popularizing the same generic spaces across the globe; it even has a name, recently christened by Kyle Chayka: airspace. Midcentury modern design, it seems, is another form of technological averaging—the cream, gray, and wood-paneled amalgam of all user tastes.

To be fair, the decline of regional accents in the United States was a process that began with the advent of first radio and then television; current technology has merely hastened it. But mid-century modern design is an excellent example of how software-driven recommendation is flattening our preferences: Modsy, a startup playing in this space, uses a quiz to help homeowners plan their design moves, hopefully obviating the curatorial presence of a human interior designer. Overwhelmingly, its clients end up favoring this specific, anodyne style.

Now, one could make the argument that this is a self-selecting population and as such may have a bias towards this design paradigm. But the subtler point is that the individual's process of questioning, discovery, learning and discrimination is undercut substantially when the heuristic of a recommendation engine is employed. It is the opposite of engaging in the serendipitous act of browsing the jumble of an antiques shop, where one goes to explicitly find difference and therefore implicitly invoke the faculty of taste. The benefit of feeling the texture of a swatch of cloth, the heft of an object, the smell of a book? These holistic sensory judgments are forfeited in favor of a quicksilver virtuality, where only the eye is privileged. And of all the senses, the eye is the most easily deceived.


Mid-century-furniture-bertoia-chairI'm not offering a romantic sentiment of days gone by. This way of going about being in the world is work. It is not by any stretch ‘efficient' – to invoke, with as much contempt as possible, one of Silicon Valley's favorite words. It cannot be. Recommendation engines putatively do this work for you, but the result, as evidenced by Modsy's result, is bereft of identity. I would not have much of an issue with this except I am certain that, once having made an interior design choice, these homeowners would be unable to describe why they made these choices, except in the most superficial terms (eg: “I like clean simple lines”). If you're such a fan of midcentury modern design, you should be able to distinguish between a chair designed by Arne Jacobsen and one designed by Hans Wegner.

Nor am I being unduly snobbish. For there is a difference between snobbery, which is at its heart a power play, and connoisseurship, which is a desire for knowledge and therefore an understanding of the importance of difference – the difference that a difference makes, if you will. A snob is someone who will tell you that the wine you are drinking is good because Robert Parker gave it 96 points and it cost $200 for the bottle (and aren't you grateful). A connoisseur will tell you that this wine is good because of how it is made, or what it tastes like, or why it tastes the way it does and how it is different from any other bottle.

Furthermore, our connoisseur will be able to describe what food goes well with this wine, and at what point in the meal it's most appropriate to drink it. A connoisseur ultimately understands each experience and object as existing in relationship to other experiences and objects. It is the interaction of these entities that ultimately creates meaning and value. As hokey as the phrase may be, this is the judgment that is required to find that one object “that really ties the room together”. Recommendation engines, by their very nature, are incapable of providing an holistic experience.

To be clear, the kind of connoisseurship I am positing is not one of absolutes, either. Values are never fixed; rather they are always being negotiated. This is another failing of recommendation systems. The flatness of fully quantified consumption behavior sets the stage for feedback loops that gradually become divorced from other criteria. Something that is popular becomes more popular simply due to its increasing popularity (in this sense, political parties and stock market bubbles share significant characteristics with recommendation engines). Other choices may experience a decline in popularity, but the system may not be able to ascertain why. Platforms may attempt to remedy this flatness and market opacity by creating multiple tiers to privilege “influencers” but this is misplaced, since it continues to impose no effort of learning on the mass of people accessing the engine. Simply put, we still cannot answer the question ‘why' in any meaningful sense.

Not surprisingly, the abdication of decisionmaking in favor of a recommendation engine has another consequence. In a distributed scenario populated by many actors – customers, consultants, designers, manufacturers, marketers – information is continually being exchanged. It is lumpy and uneven, but it is vital and dynamic. Various actors acquire different degrees of knowledge which they then use to modify their tastes and behaviors going forwards. But in a recommendation engine scenario, knowledge tends to flow overwhelmingly inwards – into the network. As Michael Tyka, a biophysicist and programmer at Google, notes, “The problem is that the knowledge gets baked into the network, rather than into us. Have we really understood anything? Not really – the network has.” Well, except for the people who own the network; they might have access to the sum of this knowledge, with which they dispense as they please.


SZUEFCTOADMore troubling is the suspicion that recommendation engines, socially speaking, function as a sort of gateway drug for generating consent for the much broader and more pervasive rubric of what has become known as algorithmic judgment. Amazon recommendations and Netflix suggestions are all well and good if this sort of impersonal guidance remains an elective activity. However, once we begin using algorithms to determine the likelihood that someone is a good credit risk, or is likely to commit a crime, then we have raised the stakes substantially. I'll take a closer look at this wave of technologies next month.

In the meantime, we have strayed rather far from that performance at Carnegie Hall. So why did Varèse's Déserts join Debussy and Adams? As George Grella elegantly stated in his review of that evening, “Two is a trend, three is an argument. The stated connection through landscape and ecology was window dressing for abstract music about form, structure and time.” A good critical assessment always provides, in addition to analysis and interpretation, an invitation into further conversation and deliberation. Grella intuits the intention behind the programming, and generates a narrative that also invites our own engagement. We build stories on top of stories. Can we do the same in a society excessively informed by recommendation engines? You walk into a stranger's home; in a friendly attempt to strike up conversation, you take note of the décor, but the conversation begins and ends with “oh, the computer did all of that for us”.