Area of Expertise is a column on niche interests, personal passions, and other things we might know or care a little too much about.
The nostalgia machine in my brain went into overdrive several weeks ago when I decided to log on to Pandora Radio after a three-year hiatus. While navigating the rush-hour press of tired bodies on the subway — with buds in my ears, like almost every other commuter that evening — I stumbled upon a podcast segment detailing how Pandora’s new CEO plans to revive the struggling music service. I was curious about the ways in which the music service I once depended on for my daily personal soundscape had changed. Logging on to Pandora, I discovered, amid its virtual dust-caked shelves wrapped in a modern interface, a record of my twentysomething sensibilities and existential anxieties.
At the very bottom of this chronologically ordered shelf were “Frou Frou Radio” and “The Way I Am Radio,” stations that I had apparently created in 2009, languishing under 36 other stations I had created and curated between 2009 and 2015. A Pandora feature called Thumbprint Radio, billed as “a uniquely personal station inspired by all of your thumbs up,” felt like a reunion with an acquaintance I had not seen in years; all her recollections of me were true, but of a former self. I winced and hit thumbs-down when Thumbprint started playing "Cape Cod Kwassa Kwassa" by Vampire Weekend, and listened wistfully when "Such Great Heights" by the Postal Service and "Swansea" by Joanna Newsom came on.
After half an hour of cycling through varieties of nostalgia, it dawned on me that unless I supplied Pandora’s algorithm with new data points about my current musical obsessions and enthusiasms, it may never know to make the leap to Noname, or Natalia LaFourcade, or Kendrick Lamar, or Jóhann Jóhannsson’s original soundtrack for the film Arrival. The algorithm wasn’t built to divine how the tastes of a lapsed user might have changed in the intervening years.
For much of my post-college twenties, Pandora was the music provider of choice in the two-bedroom apartment that I shared with my housemate in San Francisco. We seeded a myriad of stations with artists and songs we liked, and assiduously fed the algorithm with thumbs-up and thumbs-down feedback. Pandora’s secret sauce is predicated on the hypothesis that by decomposing a song into its musical fundamentals and collecting positive and negative feedback, its algorithm could identify particular combinations of characteristics that you might find irresistible, and subsequently recommend similar tracks to you. Here’s an example of the kind of logic Pandora might reveal if you asked why it recommended a specific Aimee Mann track, circa Magnolia: “Based on what you've told us so far, we're playing this track because it features mellow rock instrumentation, great lyrics, a subtle use of vocal harmony, groove-based composition and acoustic rhythm piano.” When I first heard about Pandora, before algorithmically driven recommendations like Netflix’s Cinematch became commonplace, I loved the idea that something as inscrutable as the ontology of taste could be demystified, broken down into its essential elements, and used to predict future sources of enjoyment.
But before long, the pollice verso–esque judgment that Pandora asked of its users brought to the fore philosophical questions that my housemate — to whom I owed my education in pre-Coltrane jazz and Brazilian pop — and I would periodically debate over dinner. If I gave a particular John Mayer song a thumbs-down, was I objecting to it — to him and his specific rendering of melody and harmony — for eternity? Could I really love Alanis forever? Did Pandora’s algorithm assume that we were consistent and immutable agents, impervious to the daily peaks and dips of mood or whim or to the evolution of personal taste? In addition to “thumbs-up” and “thumbs-down” buttons, Pandora’s designers had incorporated a “skip” button into the app's interface, as well as a nested menu option that allowed users to signal to the algorithm that we were “tired of the track” (and thus to put that manifestation of John Mayer on the shelf for a little while). These options only made the moments before potentially thumbing down a fraught affair: I couldn’t simply act by gut; I had to interrogate my impulses and translate my exact feelings toward a song into the most appropriate virtual gesture.
Occasionally, an artist or a song that my housemate and I liked would creep into what we deemed a musically mismatched context. One evening as we were making dinner, Louis Armstrong, whom we adored and usually heard on the Dinah Washington station we had created, suddenly sounded his appearance on my housemate’s Luisa Maita station. My housemate abruptly swung around — with knife in hand and tears in her eyes from chopping onions — walked across the kitchen to the laptop on the dining table, and mashed the thumbs-down button. Satchmo had arrived with his gleaming trumpet in the wrong music hall, led astray by the algorithm’s associative logic.
In August 2010, the Onion published a parodical homage to the music service titled “Desperate Pandora Employees Scrambling to Find Song Area Man Likes.” The piece conjures a fictitious 32-year-old Boston subscriber named Dave Lipton who skips or thumbs-downs every track, flummoxing the company’s musicologists; in the Onion’s absurdist saga, Pandora’s staff then go to great lengths to source obscure or genre-bending music to appease Lipton, only to be met by more futility. For me and my housemate, our perpetual source of comic frustration was the opposite experience: Each of our Pandora stations, when left running unattended for several hours, almost always seemed to devolve into a single genre — lounge music. Lounge music is, to me, the lowest common denominator, so bland and inoffensive as to be an affront; it is the indistinct gray that emerges from mixing too many colors on the paint palette. Were a superior artificial intelligence to take over the world, surely lounge music would be the Huxleyan drug that the machines would deploy to numb minds and quell rebellions.
I eventually grew tired of tending to Pandora’s algorithm; my stations became stale and heavily convergent on a cluster of overplayed favorites. By the time I moved out of the two-bedroom in late 2015, I had stopped using Pandora altogether and moved on to Spotify.
Today, Spotify’s algorithmic recommendations are integrated into the service’s interface in such an understated fashion that if you didn’t look hard enough at the recommended music at the end of your manually curated playlist, or if you weren’t sufficiently curious about exploring the service’s Discover Weekly feature, you could go about your week, even your entire year of music-listening without being aware of them. But what I find thrilling about Spotify’s automatically generated weekly mixtape is its tolerance for riskier, wider-ranging recommendations; the misses are worth the statistical gamble for the hits. While Pandora’s algorithm relied on preferences broadly stated and slowly fine-tuned through trial and error, Spotify’s recommendations are based on arguably clearer and more specific signals, notably songs that you’ve actually listened to through their on-demand service — when, how much of each, how frequently.
It's admittedly unfair to contrast the memory of an old service with a recent experience of a newer one. The thing that I will always love about Pandora is how groundbreaking it was, as one of the first internet services to bring algorithmically driven classification and recommendation into public life. When Pandora was founded back in 2000 and survived the dot-com bust, algorithms and machine learning weren’t yet buzzwords, and AI wasn’t as sophisticated, opaque, or poised for increasing application in critical realms like medicine and warfare as it is today — to the point that a new discipline called XAI, or explainable AI, has sprung up in an attempt to make accessible and comprehensible the complexity of what goes on in contemporary AI’s proverbial black box.
In a way, Pandora was an early model for explainable AI. It told you precisely why it chose to serve up that particular Childish Gambino track or that Hilary Hahn recording of Sibelius’s violin concerto. Someday, we might be nostalgic for a time when algorithms were simpler and more intelligible, when they seemed more like fresh, occasionally quixotic or cheesy attempts at understanding and decoding art and taste, and less a matter of life or death. ●