A Duck’s 2023

On leaving the music industry, Glenn McDonald’s website everynoise.com and uhhh computational biology?

A real life photo of Jameson Orvis. Art by Tyler Farmer.

By far the most satisfying moment of 2023 for me was quitting my data science job at [redacted] streaming service. It began as something of a dream opportunity, but I grew to thoroughly hate it by the end.1

I now work in a computational biology lab and have no desire to have a job even remotely related to music ever again, but I continue to think of my work in relation to music, like an analytical framework through which I am cursed to perceive the world. I have been struggling with what to say about my experience, which I do think has to be at least somewhat remarkable and worth writing about, so to begin unpacking it, I will attempt to invert my current way of thinking by using computational biology as a framework for understanding music, rather than music as a framework for understanding biology. To begin doing so, please allow me to provide an extended and very tortured metaphor describing the family of genomic techniques used to assess the 3D spatial organization of DNA folding patterns.

Storm King Art Center. Photo by Jeffrey Jenkins and courtesy of Storm King Art Center.

You, a 20-something artistically inclined yuppie living in Brooklyn, most likely not too far from the intersection of Myrtle and Broadway, have decided to make the train trip up to Storm King Art Center, America’s largest outdoor sculpture park. Your favorite artist has just installed a new sculpture, and you are dying to check it out. You nab a window seat on the Metro North rail line while clutching a brown paper bag containing freshly toasted everything bagel sandwiching an ample amount of scallion cream cheese in one hand and a nice warm cup of coffee in the other. Beautiful sunbursts of autumnal color dot the trees of the Hudson Valley, which unfortunately you are unable to appreciate because you are blind. Despite your blindness you are nevertheless an avid sculpture enthusiast, and you are determined to perceive the new installation by any means necessary. From advertisements you know that it is shaped basically like a long tube with a defined beginning and end, but how that tube twists and folds up onto itself is unknown to you. Furthermore, the installation is an interactive auditory experience, the tube being covered in a touch sensitive coating connected to a loudspeaker such that when the tube is touched, the speaker will announce the exact position along the length of the tube that contact occurred.

Now the easiest way for you to determine the shape of the sculpture in spite of your blindness would be to simply touch it and physically develop a sense of its contours, but particularly unfortunate for you is the fact that Storm King has temporarily banned touching the sculpture over concerns of smudging its shiny patina. The touch-based speaker mechanism still functions however, so you decide to make clever usage of this interactive feature by throwing strings at the sculpture. The strings are just long enough so that they will trigger the touch-sensitive speaker mechanism in multiple places, and the loudspeaker will announce every position along the length of the tube each string is contacting. You start throwing strings at the sculpture and recording the set of positions along the length of the tube each one contacts simultaneously announced through the loudspeaker.

By analyzing these patterns of string contacts (for example, if there are two positions particularly far from each other along the length of the tube but you find that many strings touch them simultaneously, you can infer that the tube structure must somehow be looping back onto itself to bring those locations in spatial proximity) you are able to reconstruct the spatial organization of this fascinating tube sculpture and finally appreciate the artist’s vision! A diagram illustrating this procedure is shown below:

Now replace the “tube sculpture” with a long strand of DNA and this procedure is not too dissimilar from chromatin conformation capture assays, the most common of which is very ironically called Hi-C. You are not interrogating the actual sequence of DNA but how it folds up onto itself. Also add the complexity that you are actually throwing strings at several thousand slightly different tube sculptures simultaneously and aggregating the string contact data from each (unless you are using a single cell method) and of course the fact each tube is not a static sculpture but a constantly moving and changing piece of floppy DNA. So now if you imagine placing the length of the tube sculpture on the x-axis of a figure, and plotting the number of strings simultaneously contacting each pair of locations along the tube, you get something like the figure below if you do this procedure on chromosome 8 of the human genome:

In related news, Glenn McDonald, creator of everynoise.com and former head genre-labeler / “data alchemist” at Spotify, was fired in a recent round of layoffs! Glenn was originally hired by the music intelligence company Echo Nest, which was acquired by Spotify in 2014, and he described his role as a “slowly mutating job of trying to use data and math and computers to help all the world’s music self-organize.” The most public-facing aspect of his work was everynoise.com, an interactive scatter plot of music genres identified through clustering Spotify data, and the loss of this website through Glenn’s firing has led to some lamentations among music fans.

I am not here to engage in nuanced discussion over the role of algorithms in music discovery or the state of music streaming services in 2023, but rather to fulfill the much more libidinal desire I have to mock anyone who is annoying and self-deluded enough to say some naive bullshit like that their advanced corporate music marketing algorithm is somehow helping music “self-organize itself.”

This tweet pissed me off so much it inspired this entire piece!

He is a convenient embodiment of all the deranged ways data scientists discuss art, as if it is this abstract system could be fully explained, and your generative AI music algorithm would be perfect if only you had a bit more data, or another layer in your neural network, and you could finally capture all the nuances, irrationalities, and joys of human existence and distill them down for consumption via venture capital funded tech platforms. For this reason I have decided to reference Glenn McDonald by name rather than vaguely alluding to his work to make this piece feel particularly serrated to read and at the risk of coming across as petty.

In celebration of the exciting news of Glenn McDonald’s firing (because I must admit I was quite excited to hear of this), I will present my Top 4 favorite genetic structures identified in the literature from analyzing the previously described DNA spatial organization data alongside the Glenn McDonald invented genre names of which they remind me.

Ultimately I think the processes of attempting to make sense of the chaotic mess of spatial DNA contacts and mapping out the unruly world of music as perceived through databases of listener data are quite similar, and drawing connections between the two presents an interesting lens through which to examine the work of corporate genre labeling.2

Moreover I think they are both somewhat emblematic of a similar stagnation among knowledge building practices here in the 3rd decade of the 21st century: the ceiling is lower. The computational resources required to make advances in any scientific field are so high that you must resort to analyzing more and more esoteric statistical traces of biological phenomena to continue selling the idea that one day this particular DNA folding pattern could one day be exploited to produce a successful pharmaceutical drug. In a similar way, the music industry must now perform intellectual contortions to continue sustaining myths that shitty AI generative music is the salvation to a stagnating sector or that the marketing leads generated from Glenn’s genre categorization3 will lead to profit in an industry incentivized to devalue new music altogether, both instances of the necessary symbols to continue sustaining the cult of progress and innovation becoming more and more irrational, just new diaphanous violin melodies to drown out the grating undertones of the double bass playing at the center of the world, futile attempts to distract from the inescapable truth: the ceiling is lower.

Here are some chromatin structures!

4. Topologically Associating Domains (TADs) / Deep Turkish Pop
Some TADs found in a region of Chromosome 1 on the cell line GM12878. If you want to explore Hi-C data yourself check out this nifty interactive browser!

The TADs! A classic! As soon as scientists figured out how to do Hi-C and make contact map visualizations they found a bunch of triangles along the bottom (1)! This would seem to indicate some region of the genome containing sequences which interact more frequently with other sequences in the same region than sequences outside of that region. Call them topologically associating domains! TADs are funny because their existence is very obvious just from looking at Hi-C contact maps, but actually concluding that these structures serve a functional role has been very difficult. There have been a few studies finding that loss of TAD boundaries can lead to cancer (2), but changes in TAD structure do not seem to be correlated with changes in gene expression super frequently (3). Furthermore, the definition of TAD is also complicated by the fact that the genome is hierarchically structured, e.g. in the figure above, do you also see the outline of a larger triangle containing the two smaller triangles I have labeled as TADs within it? Is this larger triangle also a TAD? Can TADs exist within TADs (4)? Are there separate molecular mechanisms driving the existence of these larger TADs? It is a complicated mess.

I have chosen to associate the TADs to the genre labeled “deep turkish pop” on everynoise.com, but I could have chosen most of the genres here which are basically variations of the name “*some non-english speaking country* pop.” These are the hypermodern equivalent of the “world” music marketing label. Similarly to TADs, the existence of all of these genres is probably highly supported by listenership data alone, like the silhouette scores on those clusters really had to be quite high, just given the fact that Turkish speakers are probably far more likely to be listening to Turkish-language artists. But also what the fuck is deep turkish pop? I chose this one in particular because the “deep” prefix really takes me out. Like turkish pop already exists as a category but one day Glenn was listening to the music from this particular subcluster and said “Hmm this sounds particularly deep” and felt compelled to create the distinction between turkish pop and deep turkish pop. Maybe I am the ignorant one and unaware of the deeply rooted traditional distinction between these two forms of music, but somehow I doubt it when the Spotify “sounds of deep turkish pop” playlist is the first result on Google.

3. “Phase Separated Regions” / Hyperpop

If you haven’t been aware of the latest developments in the field of genetic transcriptional regulation over the last half decade, i.e. living under a rock, let me tell you about the hot new theory of phase separation! So in the human genome, in addition to genes which are the DNA sequences encoding proteins which perform biological functions, there are also other sequences of DNA which do not encode proteins, but play a significant role in regulating the activation of nearby genes. Collectively these are referred to as “cis-regulatory elements” or CREs. The two CREs most important in this context are promoters, which are short sequences very close to genes responsible for initiating transcription altogether, and enhancers, which are sequences responsible primarily for modulating levels of gene transcription, but not necessarily for activating genes in the first place, and can sometimes be located very far away from the gene they modulate. Typical models hold that enhancer-promoter activation of transcription occurs via individual pairs of enhancers and promoters binding to each other to form a complex, and the resulting complex going and initiating transcription on the target gene. But a 2017 paper posits an alternative mechanism where multiple enhancers and transcription factors all come together simultaneously to form some sort of condensate which can activate multiple genes at once, and furthermore this condensate obeys phase dynamics (5). The same physics that creates oil droplets in water, but for little bits of DNA and proteins! 

It is a genuinely fascinating theory, but has led to lots of pontificating by biologists who don’t really understand phase dynamics and “condensates” has become a buzzword in the field. For both the genuine scientific impact the publication of this theory has had and the mild hype cycle surrounding it, I have chosen to associate “phase separated regions” with hyperpop. I will not belabor a discussion of the hyperpop label here, but I will point out that perhaps the most significant difference between phase separated regions and hyperpop considered in terms of their properties as abstract categories as I am doing so here, is that the theory of phase separated regions is currently very much alive and well, while hyperpop as a genre is not.

2. Compartments / Ambient

Relative to every other feature of chromatin architectural feature described in literature, compartments might be the most straightforward. There are two compartments, (ignoring subcompartments please don’t speak to me about subcompartments) compartment A and compartment B, A compartments are where the DNA is open and more transcriptionally active, but B compartments are where the DNA more compact and transcriptionally inactive. Basically a genetic on/off switch! Simple! Not nearly as many mental gymnastics required for mechanistic interpretation compared to TADs. Since they are the least debatable feature of chromatin architecture on this list, and because the checkerboarding pattern produced by compartments on Hi-C maps reminds me of gentle waves rippling across the surface of a placid lake, I associate the compartments with Ambient music, which is indubitably a genre and evokes similar imagery of calm beautiful landscape.

1. Knot Entangled Elements (KEEs) / Sigilkore
They cited Plutarch?? Lmao

KEEs differ from the rest of this list because this is a chromatin architectural feature only observed in plants! Genetically, plants are terrifying eldritch abominations existing on a plane far removed from mundane diploid plane of existence we humans inhabit (For example, did you know that while humans are diploid, i.e. have two copies of every gene in each cell with 46 chromosomes in each cell, the fern Ophioglossum reticulatum is hexaploid, with a whopping 1260 chromosomes per cell? And that the black mulberry plant Morus nigra is “tetratetracontaploid,” meaning that it has forty four fucking copies of every chromosome in each cell?? It is terrifying to contemplate.) Plant genetic studies feel almost quaint to me. In human genetics the line between research paper and elaborate marketing scheme for your new analytical method can be a bit blurry, whereas in plants you can’t even pretend like your research will one day lead to a cure for cancer, perhaps a more disease-resistant strain of a crop, but relative to the immense capital interests in finding cures for human disease it feels practically like pure knowledge seeking for its own sake.

And how interesting are those plant genomes! The authors of the original paper identifying KEEs find these little dots that appear in the contact map appearing in a sort of grid-like pattern, which is a bit reminiscent of loops, but these grids can sometimes span multiple chromosomes! How about that! What the hell is going on with these silly plants? In naming their newly discovered structure, the authors of this paper also include a parenthetical citation for Plutarch, which I find very funny. I also greatly appreciate how they color their Hi-C maps green instead of red, to lend their paper a sense of plantness over human cell line Hi-C papers.

KEEs are very sparse and pixelated, almost serrated, so I have chosen to associate them with Sigilkore. That connection is tenuous at best, but really I just want an excuse to mention the fact Sigilkore was only added to Spotify after Glenn read Kieran Press-Reynold’s excellent No Bells piece from early 2022. This is very funny to me, and an excellent illustration of how a clustering algorithm is no replacement for well written music journalism! Kieran didn’t mention this fact in his recent piece on the scene but I assume that is out of a sense of humility.

And goodbye to everynoise.com! I will not miss it.


  1. Dixon, J. R., Selvaraj, S., Yue, F., Kim, A., Li, Y., Shen, Y., Hu, M., Liu, J. S., & Ren, B. (2012). Topological domains in mammalian genomes identified by analysis of chromatin interactions. Nature, 485(7398), 376–380. https://doi.org/10.1038/nature11082
  2. Valton, A.-L., & Dekker, J. (2016). TAD disruption as oncogenic driver. Current Opinion in Genetics & Development, 36, 34–40. https://doi.org/10.1016/j.gde.2016.03.008
  3. Ghavi-Helm, Y., Jankowski, A., Meiers, S., Viales, R. R., Korbel, J. O., & Furlong, E. E. M. (2019). Highly rearranged chromosomes reveal uncoupling between genome topology and gene expression. Nature Genetics, 51(8), 1272–1282. https://doi.org/10.1038/s41588-019-0462-3
  4. Beagan, J. A., & Phillips-Cremins, J. E. (2020). On the existence and functionality of topologically associating domains. Nature Genetics, 52(1), 8–16. https://doi.org/10.1038/s41588-019-0561-1
  5. Hnisz, D., Shrinivas, K., Young, R. A., Chakraborty, A. K., & Sharp, P. A. (2017). A Phase Separation Model for Transcriptional Control. Cell, 169(1), 13–23. https://doi.org/10.1016/j.cell.2017.02.007
  6. Grob, S., Schmid, M., & Ueli Grossniklaus. (2014). Hi-C Analysis in Arabidopsis Identifies the KNOT, a Structure with Similarities to the flamenco Locus of Drosophila. Molecular Cell, 55(5), 678–693. https://doi.org/10.1016/j.molcel.2014.07.009


  1. The experience was shockingly similar to the 1999 documentary The Target Shoots First, which tells the story of a fresh college graduate landing a job at the record club Columbia House Records, gradually becoming more and more jaded about the music industry, and quitting after a year. Jaime Brooks accurately makes the point that this documentary is “the closest thing there will ever be to a truly candid document of the way Spotify playlists get made,” but a key difference is this process now involves significantly more machine learning than it did in the 90s. ↩︎
  2. For a much more thorough exploration of this topic, see Nick Seaver’s excellent book from this year, Computing Taste: Algorithms and the Makers of Music Recommendation. ↩︎
  3. Because we always must keep in mind that the primary use case of a project like everynoise.com at a music streaming company is marketing. Nick Seaver actually makes this point in his book, that developers of music recommendation algorithms conceived their work purely as a computer science problem, and cast themselves as humble servants merely guiding listeners to music they enjoy, but were interpreted by the music industry as a new form of marketing:
    “When I spoke with Riedl in 2012, he told me that the encounter with marketing had come as a surprise to his team of computer scientists. ‘We were mostly in ignorance of marketing’ during the early development of collaborative filtering, he told me, and academic publications in the field had typically framed their work as an elaboration of technical work on information retrieval, not market research. But as people like himself spun companies out of their research groups and tried to sell their technology, they found themselves interpellated by potential clients as a new kind of marketing firm.” (p. 75)
    This makes Glenn’s repeated insistence that his project is somehow allowing music to “self-organize” itself all the more jarring! To continue using such naive verbiage like thisis complete willful ignorance of the power structures you operate within while working at Spotify. To be fair I think it’s a similar fiction you must tell yourself if you are a data scientist working at any American tech company and want to perpetuate the myth that your algorithm development is making the world a better place. I just want him to shut the fuck up and stop being so annoying about it! ↩︎

9 thoughts on “A Duck’s 2023”

  1. The genres are a result of human curation, not clustering algorithms, and if they had actually been used for corporate marketing I probably wouldn’t have been laid off. But don’t let that detract from your joy about what you imagine I’m not doing any more.

    • The claim that the genres were not a result of a clustering algorithm is very perplexing. Surely you were not manually listening to every single artist on Spotify and creating categories by hand? Because that is what I think purely “human curation” would entail. Regardless of the method, my issue is more with the act of labeling itself, which is a much more slippery concept to convey and why I attempted to talk about Biology in this piece.

      And as to the claim they were not used for marketing, did you even look at the top genres on your Wrapped? Lol

  2. Every genre started with a human deciding to include that one, giving it a name in our namespace, and picking the core artists that defined it. Some of the genres supplement that human seeding with algorithmic extrapolation using listening patterns. Many don’t. It was a lot of work. There are tools to help do it, but they amplify human effort and judgment rather than replacing it. It’s reasonable to be skeptical of humans doing labeling, too, but it’s kind of what we do with everything, so if you hate genre names, I feel like you’ll hate a lot of things and maybe that’s not the way to be happiest.

    I did that part of Wrapped, yes, and you’re right that Wrapped is marketing. But ALL of Wrapped is marketing, including the top artists and top songs and the number of minutes you listened. That doesn’t make genres or artists or songs or minutes “marketing” in themselves.

    • Perhaps I should be more specific and make a distinction between “labeling genres” and “labeling genres professionally in your capacity as a data scientist at Spotify and then claiming these labels are an expression of music’s inner will.” It is really this “allowing music to self-organize itself” point you feel the need to justify your work with which I find so irksome. Regardless of the love and care poured by you and others into what is a genuinely impressive passion project, did you never have to economically justify your existence to a superior? Were you telling Spotify bosses that the dollars spent on your salary were devoted to allowing “music to self-organize itself?” Of course not, because you worked in marketing, and that is okay! This piece is really more of an expression of frustration at the mental gymnastics you personally continue performing to convince yourself that this is not the economic role which your work ultimately fulfilled.

      • It bothers you that I had my own motives for the work I did? OK, but I can’t see how anything would have been better if I didn’t.

        • I agree it doesn’t really make a difference either way whether you thought of your work as marketing or not, but it does mean I find you completely insufferable! I am here to express my personal spite towards you and the motives you had for your work. This is really a quite shallow reason for writing an essay, so I do respect you for choosing to engage with me here in the comment section of the wonderful website No Bells dot blog.

          • Well, blaming me personally for capitalism probably does qualify as shallow and unproductive spite, but I think the underlying questions, about how we do jobs and to what extent corporate amorality preempts individual motives, are deeper than that.

        • We seem to have hit the limit on number of replies in a comment thread! I will quote your most recent comment to make it explicit I am replying to this statement:

          “Well, blaming me personally for capitalism probably does qualify as shallow and unproductive spite, but I think the underlying questions, about how we do jobs and to what extent corporate amorality preempts individual motives, are deeper than that.”

          When did I ever personally blame you for capitalism? This is about esoterically telling you to shut the fuck up for once in your goddamn life! I have to imagine your inability to do so, as evidenced by you continuing to reply to me in this futile exchange, was a significant reason why you were laid off! I agree the issue of personal agency within a corporate system is a deeper question, but I do certainly hold the position that it is disingenuous to completely ignore the ultimate purpose of the machine in which you are a cog, as you do in your writings.

          As an aside, this relationship between structure and function within a corporate system we are considering is pretty similar to how biologists attempt to connect structure and function in 3D genomics! Does your position as a data scientist at Spotify lead to the function of music being marketed? Does the enrichment of genetic contacts in a particular region lead to a cell’s functional role in the body being changed? Very interesting.