How YouTube's related video algorithm helped shaped the strange rise of hazy acts like DJ Boring, DJ Seinfeld, and Ross From Friends.
A few weeks ago, XL Recordings' Imran Ahmed went live on NTS Radio. With an ear to the ground for acts at the threshold of mainstream crossover, the label A&R curates a monthly radio show to shed light on the new music he's most excited about. "My job at the label is to find the next generation of artists that are as talented and compelling as some of those synonymous with the label's past," he said in the show's introduction. "People like King Krule, Radiohead, Jai Paul...the show I do live here every month is a bit of a distillation of that search I guess—the best stuff that I'm feeling at the moment."
Over edits from distinguished up-and-comers like Sega Bodega and Mal Devisa, the industry insider flicked through hours of synth-heavy hip-hop and juddering bubblegum bass, finally landing on a breakout track from the English producer Ross From Friends, "Talk to Me You'll Understand." But the track was something of an outlier in that—in certain scenes—it already was a hit. Its original YouTube upload had already amassed over 1.5 million views, so Ahmed presenting it as something new felt a little strange. Typically, nods from tastemakers come before massive online attention, but this track's path to more mainstream acclaim might signal a change in how the web affects dance music discovery.
Over the last year, the term lo-fi house has become synonymous with a certain gritty sound; it's heavy on bass, yet built from simple synths and samples. With the tight range of tape's natural EQ and a thick cassette-hiss crackle, acts like Ross From Friends, DJ Boring, DJ Seinfeld, and others have found success in a return to form. These producers scrub bare the bloated theatrics of house's recent history with a surprising blend of sounds that feels new, even if it's coming from the past. With staggered seventh-chord pads and overblown kicks that swallow the rest of the mix, tracks like "Talk to Me You'll Understand" cut straight to house's functional center, stacking dance-driven details in intricate blends that can feel as fresh as trax from Chicago's Warehouse days almost forty years ago.
While much of its initial attention was largely written off as a lazy "lack of new ideas," the term has always existed as a loose joke among enthusiasts. Almost anticipating the short-lived lifespan of all music on the internet, its producers were hesitant to distinguish the nascent sound from the continua of both net music subcultures and the broader traditions of house and techno. Unlike, say, hardvapour's eagerness to get behind a single unified label, lo-fi house exists at the crux of a number of sustained styles, each more interested in creating lasting grooves than in bearing the weight of fleeting buzzwords. But with little media attention beyond ill-defined descriptors, how did the song catch the ear of one of London's biggest labels?
Born from styles present with labels like L.I.E.S. and 1080p since their beginnings, the phenomenon wasn't really marked by any greater stylistic shift until it found its own niche on the platforms of the closed web. Now native to outlets like the Slav YouTube channel (where DJ Boring's "Winona" first took off) and The Overload community on Reddit, the style's spread feels immensely indebted to tools like YouTube's Related Video feature. Even after visiting tracks from far-reaching genres, I find myself constantly directed toward into the same distorted sounds on YouTube, no matter how many times I clear browsing data, reset my browser's cache and cookies, or even try different browsers and user accounts. But is lo-fi house really just a sound pushed by patterns in video recommendations? The massive play counts noted by other recent features have been a source of critical fascination, but are YouTube and other sites encouraging the rise of a certain sound?
In a paper written by Google/YouTube developers Paul Covington, Jay Adams, and Emre Sargin for the 2016 ACM Conference on Recommender Systems, the researchers laid out a sketch of the basic mechanics now driving their Related Videos algorithm. As the site has shifted away from the primitive tagging system upon which it was built, developers have instead come to rely on input sources from the "deep neural networks" of machine-learning algorithms.
These days, the site's Related Video algorithm is comprised of two main neural networks. First, the "candidate generation network takes events from the user's YouTube activity history as input and retrieves a small subset (hundreds) of videos from a large corpus," the researchers note. This includes both human and machine recorded "IDs of video watches," search query inputs, and demographic data taken from a large number of undisclosed sources. From there, the second neural network then "ranks" potential matches "by assigning a score to each video according to a desired objective function." Using a vast pool of "big data" points gathered from things like the click-through rates and watch times of past video viewers—as well as a number of metrics taken from the viewers' own browsing data—the site then presents users with a selection of videos that it thinks they might enjoy.
But why do these sorts of algorithms keep serving up the same thing? Part of this seems to lie in how machine-learning patterns human interest. Modeling the percentage of watched vs. skipped recommendations, the system is constantly working to present the newest and most attuned results for the broadest audience possible. Even from an outsider's perspective, it's easy to see how something like lo-fi house lies at the intersection of a number of scenes and styles. From heads who have been around since house's beginnings, to a younger generation now thoroughly burned out on the vaporwave zeitgeist—both of which lie at the heart of the style's burgeoning audience—lo-fi house has very simply given electronic music fans an untapped vein of new tracks to get lost in. At the intersection of a number of styles and influences—from the dripping downtempo of Ibiza's Balearic years, to the early austerity of techno, synth-funk, and more current revivals in new age—lo-fi house represents such a varied collapse of styles that it seems like every dance music fan might be able to find something good in it. Maybe this overlap represents a new, more inevitably populist strand of shared taste in its strange collapse of the style's broader ties to its past.
As the internet has kicked dance music's often faceless, single-driven existence into overdrive, and web-based trends have become more and more banal, lo-fi house, as it exists now, offers an interesting logical endpoint; it's the first "genre" almost entirely indebted to the platform on which it was disseminated. As advances in maching learning increasingly make music discovery more flawlessly streamlined than human agency could really ever be, lo-fi house sets a precedent for how things could look: videos spinning endlessly on autoplay, hits driven more and more by the sort of algorithmic patterns already beginning to shape the industry. For now, all we can do is play catch up.