Subscribe
Share
Share
Embed
Who thinks that they can subdue Leviathan? Strength resides in its neck; dismay goes before it. It is without fear. It looks down on all who are haughty; it is king over all who are proud. These words inspired PJ Wehry to create Chasing Leviathan. Chasing Leviathan was born out of two ideals: that truth is worth pursuing but will never be subjugated, and the discipline of listening is one of the most important habits anyone can develop. Every episode is a dialogue, a journey into the depths of a meaningful question explored through the lens of personal experience or professional expertise.
PJ Wehry (00:02.99)
Hello and welcome to Chasing the Viathan. I'm your host, PJ Weary, and I'm here today with Mr. Glenn McDonald, data alchemist and author of You Have Not Yet Heard Your Favorite Song. Glenn, wonderful to have you on today.
glenn (00:17.665)
Thanks, thanks, glad to be here.
PJ Wehry (00:20.28)
So Glenn, tell us why this book?
glenn (00:25.323)
Well, so was at Spotify for a really long time and I am a thoughtful person who thinks about why they're doing the things that they do and reads other people's opinions about how the things I do should be done. So I read a lot about streaming from a lot of perspectives and my experience was most of those perspectives were either people with grudges or journalists with
not much information, most of which they got from talking to people with grudges. not that necessarily all the things that those people had grudges about were wrong, but there was always a lot more to things than that. And in particular, although many of the companies involved in streaming have lots of corporate flaws that are reasonable to talk about, I felt like streaming as a cultural phenomenon was like
maybe the greatest use of the internet. And I really wanted someone to tell the story of streaming and the various things that go on around it from the point of view of actually believing there was a good thing here happening. And eventually I realized that no one else was going to do it. And so I had to do it myself.
PJ Wehry (01:49.622)
always kind of fun when you write, get to write the book that you just see this glaring need for. And you also do, I almost forgot to mention you also do every noise.com, correct? And that's an awesome project. So I'd love to talk about that at some point, which really, I think showcases what you love about streaming. so talk to us a little bit about
When we talk about this movement from physical media, my brother's a big LP collector and I think you talk about fetishization, which I'll definitely mention at the next family dinner. What is the movement from physical media to streaming? What have been some of the big consequences?
glenn (02:40.588)
I mean, to me, the biggest change is in discovery that finding new music when I was a kid was a shopping experience. The place that had new music was the record store, and it was pressed onto plastic and wrapped in cardboard. And the cardboard was wrapped in plastic, again, just to be sure that you couldn't hear it.
just pick up these albums and go, I wonder what this is. wish I could know. streaming has not only changed that. mean, like initially the iTunes store virtualized that shopping experience, but it was still a shopping experience. Like you still browse through a store, made your purchase decisions, and then you could listen to your little library. And the big change in streaming is
There's no your, I mean, there is a library technically, but it's just for convenience and organization. It's not for access. suddenly all the music in the world, plenty of asterisks behind that all, but more than you can ever listen to in your lifetime is just there. And you can explore it by listening, not by wondering. And to me, that's a great thing.
PJ Wehry (03:51.638)
Yes.
PJ Wehry (03:56.176)
yeah.
PJ Wehry (04:07.757)
One of the things that as you talk about the Apple thing, it's like you can discover more you definitely have a broader selection But still a shopping experience and I think about the difference between getting Because I remember how fascinating that was to me that I could go listen to a 30-second sample on Apple but immediately the difference in that and streaming comes up the The difference seen the kinks you really got me going a 30-second sample
And roundabout by yes, a 30 second sample sample like those are like.
Or even a classical song that's not that, you know, that probably wouldn't cost anything per se. But like you have a 15-minute song and you have a 30-second sample, you have a two-minute, three verses in a chorus, boom, done song. It's still very different from listening to the whole thing.
glenn (05:03.561)
Yeah, actually one interesting sub project we had at the ECHONEST was trying to pick those 30 seconds in the smartest way possible. there are 30 seconds, you have a lot of choices for where to get 30 seconds of Tales from Topographic Oceans. And many of them won't give you any sense of what the other three hours of that thing are like, but some of them do.
PJ Wehry (05:14.082)
Mm-hmm.
glenn (05:34.043)
That was fun. But still, the 30-second sample forces you into audit mode. You just can't sit back and listen to a bunch of 30-second samples, even if you have a mechanism for stringing them together. So it just keeps you in this shopping mode where you're just constantly making decisions and you're not really appreciating things.
PJ Wehry (05:43.404)
Yes.
PJ Wehry (05:56.258)
Yes.
PJ Wehry (06:01.323)
Yeah, and I think there is a huge difference in our approach to art when we approach it as something like a judgmental, right? It's like, OK, impress me. There's a there's a completely different stance between sitting and talking to a stranger at a coffee shop and talking to a car salesman. Right. And it's like I flipped around. It's just that car salesman first to keep it consistent. But it's like this idea of like, you know, you go into a timeshare thing because they'll give you
glenn (06:19.774)
Yes. Yes.
PJ Wehry (06:30.773)
My parents always try to get me to go to those high pressure like timeshare thing where they'll give you something just for going. like, and you just resist it the whole time. And I'm like, honestly, I just rather not have the thing. That's a totally different stance from just being at the coffee shop and you're talking to someone and you're like, you know what? I'm not going to really pay attention anymore. You know, I'm like, going to let the conversation die out or I'm going to move. That's totally like there's a, there's a difference in openness. I mean, it depends on, you know, how friendly of a person you are, guess, but
We have different kinds of listeners too.
glenn (07:02.846)
Yeah, I mean, the 30 second sample based experience of music is like all your interpersonal relationships being speed dating. Just like, hi, I'm glad to meet you and now I will try to form a judgment about you as fast as possible. Like that is not a way to fall in love. It's not even a way to like have an interesting conversation for a minute, nevermind 30 seconds.
PJ Wehry (07:13.357)
Yeah.
PJ Wehry (07:32.437)
So you kind of mentioned this polemical grudge driven narrative around a lot of streaming. I appreciate you're very honest that some of that's, you know, legitimate. Some of those grudges are legitimate. But what are some of the biggest misconceptions that you're trying to correct in your book as someone who sees streaming as overall a positive thing?
glenn (07:56.809)
Well, mean, sort of the most existential one is all of the corporations that control the major streaming platforms and most of the music that gets on them are their capitalist entities. They're basically interchangeable in moral value. Spotify was a slightly more progressive than average company in employment terms, just an American employment terms, just from having started in
Sweden, but it's just a normal com, it's a normal corporation. think it's not, it's not worse or better than Apple or Amazon or Google, all of which have easily accessible long lists of things that could cause you to have reservations about them. And I wanted Spotify to be better than that, but it was not really in a position to be better than that. Like it w it is embedded in.
music industry power structure. It couldn't go into business without licensing music from major labels, which was themselves giant capitalist, inherently exploitative companies. there's never going to get revolution that way. Revolution if it comes is going to come from somewhere on the margin. Somebody is going to do something that initially seems minor and like it's only changing a little bit of something.
and then gradually realize that it's made some fundamental difference. It can't come from the center. And I wanted it to, and I could imagine, I could always imagine the emails that Daniel Eck could send if he were me. you don't get to be a billionaire CEO of a major corporation by being a sort of socialist anarchist. Like I am
And so it was never going to happen. I was never going to get the emails like, guys, we've gotten off track. We're here to make musicians' lives better. And I'm here by canceling all positions that we've taken that are against that. It's just like, yeah.
PJ Wehry (10:11.757)
Okay, so I think that's a there's a great segue here to talk about and I love the way you end the book talking connecting music with connection and joy. And so for you as a socialist anarchist,
glenn (10:30.373)
I just made that up. not real. I'm actually not really an anarchist. It's appealing. Like in
PJ Wehry (10:30.687)
why it was me okay i was yeah yeah i was running with it
PJ Wehry (10:42.313)
so as some, but what you want is for Spotify to focus on the music. And the reason for that is you think that you are spreading connection and joy. Can you talk a little bit about how music, spreads connection and joy, what you see as kind of should be the priority in these sort of the sort of tech and music space? Cause it could move. I don't know how it would change from streaming, but to be honest,
When I was putting CDs in, I didn't see how it could get better. And that's because I just wasn't thinking about it. Right. So I don't know what the next step is in music, but in the tech and music space, if we view join connection as the point of music, what do you see? How do we prioritize that? And why should we prioritize that?
glenn (11:15.227)
Right, right.
glenn (11:27.484)
Yeah. Well, yeah, those are, of course, related questions. I mean, there's an inherent why if you just believe, as I do, that music is the thing that humans do best. That therefore, there should be more of it. Everything should be oriented to have there be more music because we kind of suck at all the other things in varying degrees. But.
PJ Wehry (11:54.35)
Okay, I'm sorry. I'm gonna need you to defend that. I'm on board. I just like can you explain that to me? I've never I've never thought of music as the thing we do best. So please
glenn (12:04.872)
I mean, even bad music is pretty good, whereas like bad urban design is terrible. Like bad politics are terrible. To me, it's like music and then food and then, I mean, it's not a huge, there are other things that are sort of close. so, mean, and music and food get to the other way to think about the value of music, even if you don't consider it a goal in itself. Those two things to me,
are the best tools we have for overcoming difference, for convincing, like allowing you to see the humanity in someone who is unfamiliar to you. It's very easy, I think, in the absence of connection to imagine the people in some far off place being basically alien. They have totally different concerns from yours and different priorities.
and they probably hate you and you'd probably hate them. And then you hear a little bit of their music and you're like, oh, but their music is good. And then you eat the weird version of bread they have and you're like, that's delicious. Okay, they can't be totally alien. from the of survival of humanity as a goal, I think music is...
PJ Wehry (13:19.912)
haha
glenn (13:34.275)
uniquely powerful in connecting people. And that's part of the reason why having it all online connected together is transformational because now you can hear stuff from anywhere, anybody that you think you're afraid of. Guess what? They have hip hop and you can go listen to their hip hop and be like, I don't know what they're saying, but that's dope.
PJ Wehry (13:58.766)
So it humanizes other people and I think I love you said overcoming difference if I understand you correctly, there's a there are other ways of overcoming difference but the what music and food maybe storytelling sounds like it could fit in there but I'm curious what you think of that but That it's overcoming difference without conquering difference
glenn (14:24.43)
Right. It bypasses your rational fears, aversions. That's why storytelling doesn't really because you have language problems. so the fact that both music and food, you can understand the taste of something without being able to understand what the word for it is. And you can listen to music and enjoy it before the singing starts. you can, if you're most
PJ Wehry (14:36.125)
glenn (14:51.28)
people, think you can enjoy it even when the singing starts if it's not in your language.
PJ Wehry (14:57.069)
That's interesting. that makes sense why you would, that is able to transcend that difference quickly and easily, right? It's much easier to learn to enjoy. That makes sense to me. And I might write an understanding that you see that value of when you're talking about overcoming difference, it's connecting without conquering. Is that a?
glenn (15:20.441)
Yeah, although I mean, connecting is the beginning of empathy.
PJ Wehry (15:26.081)
Yeah. Yeah.
Because your goal is not to, when you talk about overcoming difference, it's not to make everybody the same. It's to, and that's where like, there's definitely people who are like, we're going to overcome difference. They'd agree with you on that. And their goal is to assimilate people. And you're like, no, no, that's it. I mean, when we talk about every noise at once, every noise.com that you have, that's like the whole point of that is to keep the difference. Can you talk a little bit about that and why that's important to you?
glenn (15:57.594)
Yeah. And it's sort of funny because in many things in life, we are trying to overcome, we're trying to do exactly what you said. We're trying to overcome bias. We're trying to reach like some sort of moral and human equilibrium between you and I or you and I and some third weird person that we're about to encounter. And most of what I did with music data was
not trying to overcome bias, but in fact, trying to like capture it, trying to represent it correctly. Like what, what is it that this group of, what are the groups of people in the world based on music or anything else? And what is it that distinguishes them? What is interestingly different about them? And so, you know, most of the things that would, that seemed like bias are
personality or cultural community personality.
PJ Wehry (17:01.931)
or not as important once you get a chance to sit down and listen to the music and eat their food.
glenn (17:08.683)
It depends on what you mean by important. It's like not as, you don't need to judge them. You can appreciate them without judging them. Like you can learn what the flavors of a cuisine are without the first filter being, I like this better than Cinnabon? It's like, it's not, there's a different thing. Like taste it for what it is.
PJ Wehry (17:35.692)
And do you think, and part of what you're doing with every noise at once is that you are teaching people to have that openness. Cause when you talk about the empathy that is, you, I mean, do you see that as something that you are encouraging or training? Because there's some people would be like, why would I do that? I know what I like. it's, I think you said it's like eighties.
Radio is like the default for your taste, because that's just what you grew up with. And some people are like, I'm stuck on this, and that's good enough. Do you feel that as you as do you see what your work is encouraging and training that kind of openness?
glenn (18:15.236)
I prefer to think of it as more facilitating. Like if you wanna just listen to the things you loved when you were 16, I think it's a loss for you, but it's your call. it's not really my mission to badger everybody out of their comfort zones in anything. That person who listens to only the music they liked when they were 16,
PJ Wehry (18:18.519)
facilitating.
glenn (18:42.885)
may have some other dimension of their life in which all the newness and tension is is channeled and they need this comfort to balance that. And there are probably things things in my life that I don't push because, you know, to balance out the constant influx of weird noises. But what I wanted to do was if you get curious, it should be easy to
Follow your curiosity. should be easy and rewarding to follow your curiosity.
And I hope that that encourages people to be more curious or it leads to people being more curious, but I'm not trying to badger people into doing it against their will.
PJ Wehry (19:31.726)
So I mentioned that you had this cool website as I was reading your book and my wife was working across from me. And all of a I hear just the strangest sounds coming from her computer. And I'm like, what are you doing? She's like, this website's pretty cool. I do it like, think, well, anyways, we get to that at the end is like, what should someone do after listening to this podcast? Besides buying your book and reading and understanding, it was also to visit the website and to explore a little bit.
What are the some of that you use two different axes to to organize everything and you use kind of abstract ones so that you'll have weird clusters? Can you talk about how you organize every noise at once?
glenn (20:15.873)
Yeah, is the front page of everynoise.com is a sort of map of all the genres in the world that we had identified by the time I eventually got laid off from Spotify, which is 6,500 or something. And they are laid out geometrically on the page algorithmically, not by me moving them around, but just by math. And the two axes
from top to bottom is a thing we called mechanism. So the things at the top are synthetic instrumentation and mechanical timing. Like the space between every two beats is super consistent. Like tech house. And as it goes down the page, the instrumentation gets more acoustic and the timing gets more fluid and human until you get to the bottom and you have things like classical piano, know, where the part of the
performance is manipulating the exact duration between two notes and you're very gradually speeding up, slowing down. And then the horizontal dimension on the left are very dense sounds like constant roar or constant hum. There's not a lot of variations like satanic black metal, but also classical organ and it's like box fan noise.
And then on the right is like jumpier, bouncier, spikier sounds or space between the notes and the volume goes up and down. So hip hop and reggae over there, but also like people like political speeches and people reading poems aloud and other stuff like that. everything gets the way that page works is just a big XY plot. And then there's some fiddly code that bumps out all the words around until they're not sitting on top of each other. So you can actually read them.
And then the colors, because it's in color, and the red, green, and blue channels are three more acoustic attributes similar to those. And honestly, I could look up what they are, but I didn't pick them for, I just picked them because they produced a color scheme that kind of made sense. So you can look at the page and see patterns of color across the XY space. Like there's blue.
PJ Wehry (22:15.03)
Yeah.
glenn (22:39.35)
that unifies things with sort of a glassy ambient tone, whether they're dancey glassy ambient or like chill abstract glassy ambient. And if you see two things right next to each other, you think, those are probably pretty similar. But if they're different colors, then that gives you a way to think, they're similar in some ways, but they're different in some other ways.
PJ Wehry (23:03.473)
Am I right in thinking because it seems like classical music when you talk about the density like you have Bach organ music that's going to be very dense and you have like I think it's Stravinsky's Firebird right like I think of that that a kind of famous clip that went around with a woman screaming because the you know the bump you know there's that famous like old jump scare in the middle of it and that's kind of spikier that would be actually over even though it's still classical am I right?
glenn (23:33.259)
Yeah, mean, the map is aggregating those values at the level of genres. all those symphonic composers are going to be, all their work is going to be combined together for a classical value. But then there's like classical organ, which is only averaging together just the organ music. That's going to be over on one side. And then like harpsichord.
PJ Wehry (23:38.4)
Okay.
PJ Wehry (23:53.235)
back.
glenn (24:00.479)
is inherently like spikier, so that's going to be somewhere else.
PJ Wehry (24:04.493)
Bagpipes are gonna be over in the 10th side a lot of droning right like
glenn (24:07.987)
Yeah, yeah, totally.
PJ Wehry (24:10.797)
Um, and when you're organizing genre, one of the things I found interesting is you talked about the way that it's not, there is the mathematical stuff, but you can see these genres have strange pairings when you do it according to acoustic qualities. And that's not what makes them genres. It them genres is generally a history in a community. Can you talk a little bit about what the discovery of that has been like? Uh, and if you wanted to share, I love that you, you got in trouble with the,
the dream fandom for Minecraft, for literally naming their genre. And they were really, they loved it, but they didn't want to admit that they loved it. I think I found that really fascinating.
glenn (24:51.401)
That whole episode was fascinating. But yeah, I didn't, I I started the project because I was at the Econest and we were trying to make a business doing recommendations for streaming services like backend radio and another recommendation things. And we kept having potential customers that wanted genre radio. This is the era of Pandora. So the online radio models were basically song or artist based.
PJ Wehry (25:15.564)
Yes.
glenn (25:21.972)
but there were a bunch of business people who were like, no, don't want to have people have to figure out that stuff. Just like give me eight buttons, like rock, jazz, blues, something. And we didn't actually have that. And we told someone we did, and then we had to go make it really quickly. And that was the beginning. I didn't begin, it wasn't an academic exercise. It was like, all right, how can I get a bunch of plausible
plausibly responsive rock music hooked up to a button as fast as possible. And having done that, then I was like, well, how many of these things are there? And it wasn't until fairly well into that process that I even had a moment to stop and think, wait, what is it I'm really doing here? And I think a lot of people have a...
PJ Wehry (26:12.246)
Yeah.
glenn (26:15.868)
a view of genre that it's a set of musical parameters like ragtime. It has this certain meter, certain instrumentation, certain structural characteristics. it's true that ragtime is a genre and it's true that most of the music in it does have those common elements, but that is not what made it a genre. It made it a genre as people. And that was
That for me, that was the breakthrough that allowed me to figure out where to go beyond the first 100. And even to understand why the first 100 was the first 100. Like all genres are formed from human community. There are communities of practitioners or listeners or both, usually both, but not always. And it is the people that define them.
And I'm reminded, I was reminded of this many, many times over the years since as I, as I raised a child and would have to explain punk and just like explaining why the stranglers and the damned, which I of course, having grown up with them understood to be punk were punk, even though they didn't sound like the sex pistols and they certainly didn't sound like green day and
But it wasn't existential. I the answer was really simple. Like this thing called punk began in London, these people, these streets, these clothing stores, these shows. and then it started again in New York and again in LA. And in each time it was a bunch of people and it became whatever those people did initially. And once they'd done it, then the next generation comes along.
and has a thing they already think that word means. So if you started a punk band in 1981, that was a very different thing than starting a punk band in 1976 because you were trying to do a thing that already existed. And there's a part in the book where I go through the inevitable cycles and divisions and rebellions that each one of these things go through because then having
PJ Wehry (28:35.788)
Yes.
glenn (28:40.638)
the new generation comes in and makes their version, but they already know what they think it's supposed to sound like. And then someone gets pissed off that, you didn't capture the original spirit. And then you get Neo, whatever it is. And then everybody gets pissed off. Yeah, yeah, yeah. Like everything goes through. They don't always use the same terminology, but that's what happens with communities over time. Like they spin around and they divide and re-
PJ Wehry (28:54.773)
Neopunk post-punk. Yeah.
glenn (29:10.239)
coalesce and land somewhere else. That's all very human and that's how genres develop and evolve. Yeah. Once I understood that, then I understood, I'm just trying to find all the listening communities on the planet. And so I'm probably never going to find them all, but it is at least notionally finite. Eventually we run out of people, at least in theory.
PJ Wehry (29:19.682)
And it's,
PJ Wehry (29:38.829)
Yes, at least in theory. and of course it does get complicated because people belong to more than one listening community,
glenn (29:47.602)
Yeah. And that was another thing very early on, like every other attempt at modeling genres pretty much at the time was done as a tree. And so it's like all the top there's classical and rock and jazz and pop. wait, wait, is rock pop? But jazz was pop and...
And you quickly can get yourself tied into useless knots at that level. But then once you get down below, it's a strict hierarchy, then it's hopeless because what do you do with pop rap? Is that a form of pop or is it a form of rap? And you could pick one, but if you're like a Spotify business manager and you're trying to do stats by genre,
and your whole dashboard changes color if you make a different decision about whether Drake is pop or rap, then like that's chaotic and useless. So my version of it was all genres are data wise, atomic, independent, artists can be in multiple of them. You belong to multiple communities. You belong to your physical community, but also your Scrabble nerd community and everything else. It's very normal.
They overlap, they subsume each other, they are defined by history, by place, by lyrical propensity, by how many bagpipes they have or don't have, like all these kinds of things.
PJ Wehry (31:19.797)
Absolutely. I want to talk to you about how we could be more responsible with the algorithms that find these things and why that's important. But before we do that, think it would help people to understand the... I want to be careful about the word failure here, but you had some great stories about...
engineering failure and engineering surprises that I think helps people understand what algorithms can do and what they can't do. And that'll help inform our discussion about algorithms. Can you give some of your favorite stories about some of more interesting failures in your engineering and some of your favorite surprises?
glenn (32:08.38)
So early on at the Echo Nest, so before we did the genre stuff even, the kind of layer beneath that was artist similarity. So like one of the main things we were doing at the Echo Nest was scraping all the pages that we could find on the web that said anything about music, trying to figure out what bands or artists they were talking about, and then assess word frequency. So you could say, you know, if the
People use the word reggae a lot to describe Bob Marley. And they use other words too, but that one really stands out. There's a lot of wealth of information that people have expressed in an unruly but evocative way by typing onto the web. And we had a thing that would look for patterns of those terms and say, these two artists share words that describe them. They're probably similar.
And you can do that without even looking at what the words are. one of the things that we did, as I was looking at the code early in my time at the Econest, was I realized we were breaking apart multi-word phrases into individual words. And this rock and roll turned into rock and and and roll. But when people say rock and roll, they actually mean something different than when they say rock.
Rock and roll is a particular historic phrase and it's really useful to retain that. And so I've just made a quick little fix that kept the phrases together. And in my memory, like 30 seconds after I committed that, the founder stormed over to my table and was like, never mess with that. Like, what, what, what, what's wrong? It's clearly better this way. He's like, no, no. And he like,
pulled up Justin Bieber on my screen and where Justin Bieber's most similar artists a minute ago were all very reasonable. Now it was this weird set of people I'd never heard of. And I'm like, you're right. I broke the system. And so I reverted it and he like took a deep breath and walked away. And I was like, but why? And at the time, this is before Spotify. So we didn't have any listening data, any direct listening data. So all we had was
glenn (34:32.347)
were these words and tags that we scraped. And we scraped stuff from Last FM, which allowed people to put their own tags on music. And one of those tags was a long phrase that included some homophobic sentiment that I don't remember and won't reproduce. But it something to do with San Francisco and sweaters and caramel lattes. It was like a very ill
PJ Wehry (34:52.717)
I'm totally fine with that.
glenn (35:01.563)
meaning but hilarious long phrase. And some person had gone and tagged Justin Bieber and like a bunch of other people that were otherwise unknown. And none of those words were interesting on themselves. So if you broke them apart like sweater, I know sweater is not, but when you put them back together, that phrase was only ever used in the entire world on Last FM to describe these 20 artists.
totally screwed up our thing. I mean, found a way to fix it by looking at that. If it only occurs 20 times in the world, then that's cool, but also a little bit of a reason for suspicion. So with enough intelligent math, you can get the good signal out of things without being screwed up by the perverse abusive tagging some people do.
PJ Wehry (35:57.164)
And I actually, this came up, I almost mentioned this earlier when you're talking about punk, neopunk, post-punk, know, whatever, punk core, you know, whatever. When you're talking about communities, sometimes people are defined as much that people are often defined by what they love, but they're often also as much defined by what they hate, which seems a little bit against what you're trying to do with every noise at once. But it's what you're running into sometimes, like...
Like, people wouldn't be offended if you said that punk was similar to, you know, Belarusian hip hop. They'd be like, what do you mean? You know, have to, but you're like, no, neopunk is just like post-punk. I'm sure, you know, and I'm obviously just creating those labels, but people would be like, no, we're not, you know, because there's that history there, right?
glenn (36:43.759)
Right, closer they are to each other, the more likely everybody on either side cares about the exact imaginary boundary between them.
PJ Wehry (36:54.617)
And that's the same thing with somebody had a Belieber in their life. Like, that's like definitely... The only reason you do that is because something about Justin Bieber had been thrust in your face and you didn't want it to go and to seek that out. Anyways, I do love that very... It's surprising in the math to find all the humanists, right? I guess it shouldn't be surprising, but it is.
glenn (37:21.38)
I mean, it's the goal, of course. it got our ability to find. So that data set looked very interesting. It looked very human, because every artist then had this long list of words with declining weights. And so you could sort of get a non-grammatical description of each artist from it. And then when we got acquired by Spotify, finally we had firsthand listening data.
And so we knew what everybody listens to all the time, and we could find patterns of shared listeners. That turned out to be just exponentially better for actually finding similarity, even though it didn't produce anything intelligible in the same way. Like, it only produced descriptions of artists in terms of other artists. And so it lacks, at least in a technical sense, a bunch of the information.
that people have. Like you listen to these three bands, why you listen to those three bands? Because they're all Estonian metal bands and you like Estonian metal. And it turns out that you can make much better recommendations for people if you know that these, they like these three bands without knowing why, than if you know why, but you don't know what they actually listen to. And it's because
PJ Wehry (38:44.599)
what they actually listen to.
glenn (38:47.331)
Humans don't listen randomly, so your listening encodes everything you love and hate. if you hate the Red Hot Chili Peppers, you will be among the relative minority of people who are exposed to bands like the Red Hot Chili Peppers but have a notable absence of Red Hot Chili Peppers in their listening.
So you can find those patterns.
PJ Wehry (39:20.823)
There was one story in particular, if you don't mind sharing, about surprising algorithmic. You were trying to find Christmas music, which is one of the more difficult things you've run into. And you thought there was something wrong because you kept getting huge Christmas responses in September for, I think it was the Philippines and for Saudi Arabia. Can you talk about that? Yes.
glenn (39:41.048)
Yeah, yeah. Well, the first year, yeah, right. Yeah, so the first year I did it, yeah, I had this idea. If you work in music recommendations, you have to worry about Christmas because nobody wants Christmas recommendations outside of those two weeks before Christmas. And then for two weeks, everybody wants at least to be able to get Christmas recommendations. So despite being a person who tried to ignore Christmas music as much as possible for
my entire adult life, I became involuntarily fascinated with it to be able to filter it out and in. And so I did at one point this thing, which you can still see the last state of on every noise called the approaching worms of Christmas. And it was a measure by country of what percentage of that country's listening by day was Christmas music. Whole sub interesting sub project of how you know it's Christmas music, but.
So I'd done this for a few years and every year it was like a race to see whether Iceland or Norway or Sweden or the Netherlands would come out on top. so like all the Scandinavian countries would start earlier than the rest of Europe and the US and end up higher. And so I, and I would turn it off after Christmas and not think about it until August and I turned it on again. And so one year I turned it on again in August and
just sort of ignored it until it was into September and I could go back and look at how things were doing. And it looked like they had the previous years, except there was this one line that just went way up on September 1st, like sharply overnight on September 1st and then stayed high long before I knew any substantial portion of Scandinavia was listening to Christmas music. And I'm like, ah, crap, I definitely screwed something up.
processing of this thing and spent a good half hour at least maybe more trying to figure out what I'd done wrong. And finally I was like, I can't, I can't find it. What, what is that? What, what is that line? Cause I just had little country coats on them. So like, PL, is that? PH? I don't remember what it was, but it was the Philippines. Um, and I'm like, no, oh, come on. That's ridiculous. Like the Philippines is nowhere near the Vatican. Like they can't.
glenn (42:06.232)
They can't be listening to Christmas music. then having still failed to figure out what my error was, I started doing a little Googling. I'm like, oh, actually, they like Christmas music in the Philippines. And then like two hours of weird articles later, I'm like, huh, Christmas music has a formal season in the Philippines, and it starts on September 1st. And.
If you're listening to like, I, would be bad if I'd put individual names on that thing. Cause you don't want to out the people who had jumped the gun and begun listening to Christmas music in the Philippines on August 31st. Cause that is not done, but September 1st season opens. And yeah, as you, as you alluded the next year, when I started it up and went to look to see if everything was Philippines were still their weird selves. There were three lines and I'm like, no, now what if I screwed up?
But yeah, the next two were the UAE and Saudi Arabia. And another hour of Googling later, I had a better understanding of the overseas Filipino worker population, which is a giant demographic phenomenon that millions of people participate in and know about. I didn't, and that's how I learned about it. yeah, the Philippines had gotten Spotify long before
UAE and hundreds of thousands of Filipino workers had gone to build World Cup stadiums and shopping malls and whatever else in Dubai and they brought their relatively cheap exchange rate Spotify subscriptions with them. And so when I went and looked at the charts in UAE around Christmas, was like,
mostly Filipino Christmas music and then number six would be something in Arabic and then number 10 would be something in Arabic. And it's so life-affirming that these people are able to love both this Arabic religious music and this Christian. then of course, when you dig into it, you find no, it's not that. They're two separate populations coexisting at once and the Arabs are not listening to
PJ Wehry (44:04.396)
Yeah.
PJ Wehry (44:20.745)
It's a bridge to the sky!
glenn (44:26.743)
Christmas music and the Filipinos are not listening to Diwala or whatever the
PJ Wehry (44:32.971)
Yeah, I'm sorry. I loved it. I love so it had the structure of a joke because you're like, no, what have I messed up this time? like, I you're like, I know, Philippines, could have like a Christian influence. But like, I know for a fact that Saudi Arabia is majority Muslim, where's this coming from? And you're like, there's Filipinos that are like, they came over.
glenn (44:34.113)
chance were.
glenn (44:53.758)
Which in this anecdote was a predictable punchline. at the time, as I started trying to figure out why UAE was up there, I didn't realize that it was the same thing I discovered the previous year, but just in a different form.
PJ Wehry (45:00.031)
Yes. Right.
PJ Wehry (45:07.298)
Yeah.
PJ Wehry (45:12.845)
And I love it like there is a little bit of a detective novel feel to reading that. Like what is happening here? And I think that leads us to because you're doing a lot of cultural digging to affirm what you're what you're finding. And I think that leads to the one of the main questions I wanted to ask you. And I think you give a really great recommendation. can you can talk about directly to this as someone who's worked on this kind of stuff.
difference between call quality and metrics driven algorithms and why one is more responsible in the other. Do you mind talking about that?
glenn (45:49.536)
Yeah, I mean, everything I did basically at Spotify was generate something, playlists, lists of artists, pictures, whatever, and then run it on a lot of different things and look at the results and try to figure out whether they are good or not. I I got into this business because I was a music geek and I knew a bunch of stuff about music.
And very quickly in the process of learning more about music, I realized how little I knew about music. But at least I knew how to figure things out about music. So I can produce a list of German hip hop. And I didn't know anything about German hip hop, but give me half an hour. And I could be like, all right, so most of this is clearly German hip hop. some of it is not German. Some of it's not in German. And some of it is not hip hop.
And that gives me stuff I can Google. And as a person who's read hundreds of thousands of articles about music, maybe I'm slightly faster at deciphering them than someone who hasn't, but more, I just care. Like really that was the difference. Wasn't I was smarter, I had more expertise. I just really cared more. So I would just gnaw at that generator until it's German hip hop list was all German and all hip hop. in the poor, in the
process of doing that, would discover nine other things around German hip hop in the course of learning to filter them out. so that's how I did everything. And there is a lot of stuff in technology and recommendations in particular that is not done that way. It's done with machine learning. It's done with automated processes that
ultimately respond, they're still fed by human inputs, by listening and by people making playlists and whatever, but they are not iterated by experts or even fans looking at individual results and trying to make them be different. They're iterated by measuring some wholesale effect. Like instead of trying to make the German hip hop playlist good,
glenn (48:09.02)
you the ML approach would be just generate a place for everything, ship them to 10 % of the listeners, and then measure how much time people spend listening and compare that figure for the people who got the new algorithm to that figure for the people who got the old algorithm. And that is not a crazy way, like that's not a crazy way to proceed. That works at some holistic level, but it misses so much nuance.
Like if you run that test and you find, all right, things went up a quarter of a percent, then the human thing to ask is what really happened? Did every genre get a quarter of a percent better? Does that even make sense? would it be to be a quarter? Like we're giving people three minute songs. Did we give them three minute, five second songs? Like how could a song be a quarter percent more applicable? Like it doesn't make sense. So then you dig and you'd find,
No, actually, most things stayed the same. A couple genres got much better or people listened to them for much longer, which is not always the same as much better. And then like a bunch of them got worse. And if you care, you'd be like, well, that's promising, but also bad. And let's go see if we could figure out what was good about the good ones and what was bad about the bad ones and make it so that we make everything better or at least something's better without making other things worse. But if you're, if
your priorities and your personality is not that, and you just like, see the figure, we went up by a quarter percent, ship it, move on to the next project. Then that's how things insidiously get worse because that number hides a much more complex situation. And often that number is just misleading. Like it measures something, but the
The advantage disappears when you run it again, or it turns out that it's some spurious variation. Like I remember we ran a test once and I was sure that it wasn't going to make any difference. Just sure from my perspective as a music listener that whatever we were changing didn't make any difference. But one of them was dramatically better in the results. And I was so irritated by that that I went and dug into it and found that in setting up the test, we had
PJ Wehry (50:31.424)
Yeah.
glenn (50:32.387)
accidentally, there was a filter in the real version that excluded super short songs. And we'd forgotten to put that filter on the test set. we were measuring by how many songs people listened to. all that had really... We hadn't improved anything. We'd just basically taken the same results, not filtered out short songs. And then those people who got short songs ended up listening to more songs in the same amount of time. And we were like,
Victory! We've increased the number of songs people listen to. you're like, no, no, that wasn't, that was not victory at all.
PJ Wehry (51:09.611)
Yeah. And didn't you run it again and it was either negligible or actually the new one was worse. I can't remember. I remember reading the story.
glenn (51:17.134)
I'm sure we had all three possible. Over time, in the book to placate Spotify lawyers, often combined elements of things so as not to reveal any nominal secrets. And I think in that case, there are really a couple of different experiments that I was describing. But yeah, we often found that, the thing that you think got better is worse or no different.
PJ Wehry (51:21.61)
Yes
PJ Wehry (51:33.918)
Yeah, that's
PJ Wehry (51:44.651)
Yes.
glenn (51:45.701)
Or maybe when you remove the thing that screwed up, it actually is better, but you weren't measuring that correctly before. You just don't know.
PJ Wehry (51:53.41)
Yeah. Yes. And I do think there is something here too, is that one of the things you talked about is that just measuring attention or just measuring time can be problematic. And if you're, if you're doing that just for, like one to five changes or even to a hundred changes, you might see improvement, but it
Am I right in thinking that over a thousand changes or 10,000 changes or even longer, like over years of doing this, will eventually, it'd be very easy to get a worse result than if you had this more quality, more curated approach because of the way it kind of compounds.
glenn (52:34.855)
Yeah, because if you're not really scrutinizing what changes, then yeah, you're probably changing some things that you don't realize. And the more changes you make, the more random, unintended, unintended consequence things are happening around the margins. And those build up. And then by the time you get to your thousand and first change, it enters this world that has all these
accumulated weirdnesses and sometimes just ends up triggering one of them. So you have some problem that you really introduced two years ago and it was tiny at the time, but it just sits there lurking until later when some seemingly unrelated change suddenly triggers it. those little five second clips didn't matter because some other thing filtered them out. Like this thing that made radio didn't
include anything that was less than a minute and a half. But then three years from now, you have to re-implement this and so you're writing this radio thing over from scratch with some other technical backend and no one who's doing it was on the original team and they forget that filter and suddenly their radio is filled with all this crap. And you're like, what? I didn't do this. you're like, you didn't do it. Like those things that
Flaw was introduced three years ago. Those tracks have been sitting there for 12 years and never bothered anybody. But now you stepped in it.
PJ Wehry (54:11.873)
Yeah, and I think, and I'm trying to think of the way that beyond just music, but the way that algorithms can work is a lot of times the people who are most affected and you won't necessarily see a drop until they've been gone for a while are like your power users. Like the people who are making playlists on Spotify for free are some of the first to be bothered by some changes and you won't see their, actually bring a lot of good to the platform and you won't see the
problems of losing them until later on, until after they've gone and it's been some time. Does that make sense?
glenn (54:49.029)
Yeah, totally. And that's one, but hardly the only case of slow effects or things that you only experience a couple of generations later. We're in the middle of a particularly dangerous one now in technology where we're filling the web with AI-generated stuff that then is hard to tell the difference between that and the human-generated stuff and then feeds back.
PJ Wehry (54:56.503)
Yeah.
glenn (55:18.211)
as training into the next generation of AI. So when you do that the first time, it's like, yeah, it's fine. It's fine. And then you're like, but now I'm worried about it two months later. So you test it again. You're like, nah, it still seems fine. And then you stop testing. Because I mean, I checked twice and it was fine. And then you're not watching as it gradually goes crazy. Yeah.
PJ Wehry (55:23.947)
It looks great. Yeah.
PJ Wehry (55:40.077)
Like, it like.
Yeah, I mean, as a really dumb example, and I'm sure that they test better than this. I'm sure engineers wouldn't make this mistake at all because they don't make silly mistakes ever. No, but like like the finger thing that was going through with AI where like they would add like it had to be like three fingers instead of four fingers or be six fingers instead of five fingers. Right. And if if that gets put out on the web and then you train the AI on that again.
glenn (55:55.386)
No, no, no.
PJ Wehry (56:14.817)
then that problem will actually get worse. Because before it was being trained on normal human hands, but now it's like, well, sometimes humans have six fingers because that's what I see on the web.
glenn (56:25.297)
Yeah, have this picture right. I got a bunch of pictures of people with six fingers. It's very normal. So there are probably some with seven. And then later, you're like, whoa.
PJ Wehry (56:31.581)
yeah. Right, right. So I want to be respectful of your time and thank you so much. I've really enjoyed this. Where do we go from here? What do you see? What would you recommend for the consumer to do? But also you have a consumer. How about let's say the listener so we don't have that like, you know, shopping experience, right? The listener.
glenn (56:57.914)
Thank you.
PJ Wehry (57:00.457)
And the person for the musician and for the that person who makes it hat like makes those connections possible. What do you see as some good next steps?
glenn (57:14.596)
Yeah, so I'll go through them in reverse order. My marching orders to the people like me who work on these things is to think about yourself as a steward of human culture. And anything that you're doing with math and data is interfering in a technical sense with human culture. And it is your responsibility to care what that does.
try to figure out what you were doing and understand it and try to make it be an improvement in the world rather than a source of disintegration. For the artist less, for the listener, you don't have to do anything. It's not your job as a listener to fix the future. I hope you're trying to make the world better in some way, but music doesn't have to be it. Music can be your comfort.
And then you go back to your job trying to cure cancer or purify water or make coffee or just whatever it is you're doing that you really that you're trying to make a difference with. It's fine. If you have the bandwidth, then I think being a curious listener and finding ways to take control of your listening experience.
and seek out things that you actually like as opposed to just being satisfied with whatever some process feeds you based loosely on what you already told it that you like. That's the way to be a responsible listener, both for the music industry's sake and for your own sake. For an artist, and I get asked this a lot, particularly in the form of like, how do I game the algorithm? What should I do to get
the favor of the algorithm. And my answer is always none of that. think whatever you just listed, don't do that. Don't spend any time trying to second guess a streaming platform or an algorithm because a bad thing that will happen is you won't succeed and you'll have wasted your time. an even worse thing is you will concoct some elaborate strategy you think is working. You will have some success, which will almost
glenn (59:36.739)
certainly be coincidental. And you will attribute it to the like weird superstitious thing. all album covers should have a big orange circle somewhere in the background. And I did that in my song got 612 streams and where all my old ones only had 30. And so now every cover after that must have a orange circle in the background. And mysteriously, they never do well again, because that wasn't why.
So you can waste a time. And meanwhile, you'd spend all this time making orange circles, which didn't teach you anything about your music or your album covers or anything. So my recommendation to artists is always do the things, sorry, do the things you always knew you had to do. Like you're trying to build audiences. If you are super famous, you can build a community that is just your fans. And if you are not Taylor Swift, you cannot. And so you need to be part of a community.
that exists for some reason other than you. And you need to have a place where you go away for two years to make an album and come back. There has to be a place to come somewhere to come back to where people who have not been thinking about you for who may have loved you two years ago, but it's been a while. They're listening to other things. The world is big. You got to someplace where when you come back, somebody is like, I remember these guys and goes.
PJ Wehry (01:00:47.681)
Hmm.
glenn (01:01:02.862)
to the place where these people connect and be like, hey, you guys remember so and so and everybody else is like, oh yeah, yeah, now that you mentioned it, I do. They have a new album, great. If you're Taylor, you just post it on, you know, go on.
Travis's podcast and you announce it and the world spins out of control. But for most people, don't have a famous NFL players podcast in which they can announce their albums.
PJ Wehry (01:01:31.362)
Or the millions of dollars that's probably put into the marketing. I think this goes back to what we talked about kind of near the beginning, that genres are really communities, right? And that's an understanding that. I think I know the answer. But besides buying and reading your book, which was a fun read and a fascinating read, and an easy read, which is always nice. I have a lot of philosophers.
glenn (01:01:53.474)
Thanks.
PJ Wehry (01:02:00.27)
It's a nice break from it's like, well, actually, when we're talking about Heidegger, you know, you know, then there'll be a string of German words, which I can't write off the top of my head. You know, I'm like trying to sound it out. Anyways. Thank you. It was wonderful.
glenn (01:02:15.462)
I read a lot of academic papers about music and streaming the business too, and I was definitely not trying to write one of those.
PJ Wehry (01:02:23.149)
Yeah. So besides reading your very, I definitely got off track there. That was some definitely repressed trauma. the besides reading your very clear, fun book and informative too, what is something you'd recommend to someone who has listened to this podcast for the last hour? What would you have them do? Would you have them meditate on or would like, what would you tell them to go listen to?
I did know I was going to ask how much did you get paid to do that plug for Icarus? They're like the best song ever. know, yes.
glenn (01:02:59.278)
I got paid nothing.
PJ Wehry (01:03:03.254)
Good.
What would you tell someone to do after listening to this podcast?
glenn (01:03:09.824)
Well, after listening to this podcast, you should immediately swear off podcasts forevermore and develop, devote all this time you've wasted listening to people like us just talk at you and spend it listening to music.
PJ Wehry (01:03:23.635)
Absolutely. Okay. Yeah, there you go. We solved it. no, no. but I, I think there is definitely some like, go find something new to listen to, right? Like go to every noise at once, click around a little bit, find something you enjoy, and then go listen to more of it on a streaming service.
glenn (01:03:27.776)
Sorry.
glenn (01:03:45.964)
You're living in the best time for discovering music that there has ever been up till this point. So don't miss it. Music is the thing humans do best and.
You can be convinced of that too.
PJ Wehry (01:03:58.593)
Glenn, I appreciate it. Yes, I love it. That's a great way to end. think we start with that, we end with that. Thank you. An absolute joy having you on today.
glenn (01:04:07.98)
Thanks.