WEBVTT 00:00.000 --> 00:03.920 The following is a conversation with Gustav Sorenstrom. 00:03.920 --> 00:07.200 He's the chief research and development officer at Spotify, 00:07.200 --> 00:11.200 leading their product design, data technology and engineering teams. 00:11.200 --> 00:15.280 As I've said before, in my research and in life in general, 00:15.280 --> 00:18.720 I love music, listening to it and creating it. 00:18.720 --> 00:23.600 And using technology, especially personalization through machine learning, 00:23.600 --> 00:27.840 to enrich the music discovery and listening experience. 00:27.840 --> 00:31.920 That is what Spotify has been doing for years, continually innovating, 00:31.920 --> 00:36.000 defining how we experience music as a society in the digital age. 00:36.000 --> 00:39.200 That's what Gustav and I talk about, among many other topics, 00:39.200 --> 00:43.280 including our shared appreciation of the movie True Romance, 00:43.280 --> 00:46.080 in my view, one of the great movies of all time. 00:46.080 --> 00:49.280 This is the Artificial Intelligence Podcast. 00:49.280 --> 00:53.120 If you enjoy it, subscribe on YouTube, give it five stars on iTunes, 00:53.120 --> 00:58.000 support on Patreon or simply connect with me on Twitter at Lex Friedman, 00:58.000 --> 01:00.400 spelled F R I D M A N. 01:01.200 --> 01:05.040 And now, here's my conversation with Gustav Sorenstrom. 01:06.400 --> 01:10.240 Spotify has over 50 million songs in its catalog. 01:10.240 --> 01:13.120 So let me ask the all important question. 01:14.080 --> 01:16.240 I feel like you're the right person to ask. 01:16.240 --> 01:19.520 What is the definitive greatest song of all time? 01:19.520 --> 01:22.640 It varies for me, personally. 01:22.640 --> 01:25.040 So you can't speak definitively for everyone? 01:26.160 --> 01:30.240 I wouldn't believe very much in machine learning if I did, right? 01:30.240 --> 01:32.800 Because everyone had the same taste. 01:32.800 --> 01:36.960 So for you, what is... you have to pick. What is the song? 01:36.960 --> 01:39.360 All right, so it's pretty easy for me. 01:39.360 --> 01:44.960 There's this song called You're So Cool, Hans Zimmer, a soundtrack to True Romance. 01:46.000 --> 01:49.040 It was a movie that made a big impression on me. 01:49.040 --> 01:51.840 And it's kind of been following me through my life. 01:51.840 --> 01:54.880 I actually had it play at my wedding. 01:55.360 --> 01:58.400 I sat with the organist and helped him play it on an organ, 01:58.400 --> 02:01.040 which was a pretty interesting experience. 02:01.040 --> 02:06.000 That is probably my, I would say, top three movie of all time. 02:06.000 --> 02:07.600 Yeah, this is an incredible movie. 02:07.600 --> 02:10.400 Yeah, and it came out during my formative years. 02:10.400 --> 02:15.920 And as I've discovered in music, you shape your music taste during those years. 02:15.920 --> 02:17.840 So it definitely affected me quite a bit. 02:17.840 --> 02:20.400 Did it affect you in any other kind of way? 02:20.960 --> 02:23.440 Well, the movie itself affected me back then. 02:23.440 --> 02:24.880 It was a big part of culture. 02:25.600 --> 02:27.680 I didn't really adopt any characters from the movie, 02:27.680 --> 02:32.160 but it was a great story of love, fantastic actors. 02:33.040 --> 02:37.920 And really, I didn't even know who Hans Zimmer was at the time, but fantastic music. 02:39.040 --> 02:42.160 And so that song has followed me. 02:42.160 --> 02:43.920 And the movie actually has followed me throughout my life. 02:43.920 --> 02:48.480 That was Quentin Tarantino, actually, I think, director or producer. 02:48.480 --> 02:52.080 So it's not Stairway to Heaven or Bohemian Rhapsody. 02:52.080 --> 02:53.600 Those are great. 02:53.600 --> 02:57.760 They're not my personal favorites, but I've realized that people have different tastes. 02:57.760 --> 03:00.400 And that's a big part of what we do. 03:00.400 --> 03:02.640 Well, for me, I would have to stick with Stairway to Heaven. 03:04.000 --> 03:09.280 So 35,000 years ago, I looked this up on Wikipedia, 03:09.280 --> 03:13.120 flute like instruments started being used in caves as part of hunting rituals. 03:13.120 --> 03:15.760 And primitive cultural gatherings, things like that. 03:15.760 --> 03:17.280 This is the birth of music. 03:18.000 --> 03:25.040 Since then, we had a few folks, Beethoven, Elvis, Beatles, Justin Bieber, of course, Drake. 03:25.680 --> 03:29.280 So in your view, let's start like high level philosophical. 03:29.280 --> 03:34.080 What is the purpose of music on this planet of ours? 03:35.200 --> 03:38.240 I think music has many different purposes. 03:38.240 --> 03:44.640 I think there's certainly a big purpose, which is the same as much of entertainment, 03:44.640 --> 03:52.080 which is escapism and to be able to live in some sort of other mental state for a while. 03:52.080 --> 03:54.320 But I also think you have the opposite of escaping, 03:54.320 --> 03:56.720 which is to help you focus on something you are actually doing. 03:57.280 --> 04:02.080 Because I think people use music as a tool to tune the brain 04:02.640 --> 04:05.120 to the activities that they are actually doing. 04:05.120 --> 04:10.560 And it's kind of like, in one sense, maybe it's the rawest signal. 04:10.560 --> 04:13.040 If you think about the brain as neural networks, 04:13.040 --> 04:16.880 it's maybe the most efficient hack we can do to actually actively tune it 04:16.880 --> 04:18.400 into some state that you want to be. 04:18.880 --> 04:19.760 You can do it in other ways. 04:19.760 --> 04:22.240 You can tell stories to put people in a certain mood. 04:22.240 --> 04:26.240 But music is probably very effective to get you to a certain mood very fast, I think. 04:27.120 --> 04:30.960 You know, there's a social component historically to music, 04:30.960 --> 04:32.480 where people listen to music together. 04:32.480 --> 04:36.880 I was just thinking about this, that to me, and you mentioned machine learning, 04:36.880 --> 04:42.000 but to me personally, music is a really private thing. 04:43.040 --> 04:45.920 I'm speaking for myself, I listen to music, 04:45.920 --> 04:49.600 like almost nobody knows the kind of things I have in my library, 04:50.320 --> 04:54.400 except people who are really close to me and they really only know a certain percentage. 04:54.400 --> 04:58.560 There's like some weird stuff that I'm almost probably embarrassed by, right? 04:58.560 --> 05:00.000 It's called the guilty pleasures, right? 05:00.000 --> 05:02.560 Everyone has the guilty pleasures, yeah. 05:02.560 --> 05:06.560 Hopefully they're not too bad, but for me, it's personal. 05:06.560 --> 05:12.880 Do you think of music as something that's social or as something that's personal? 05:12.880 --> 05:13.600 Or does it vary? 05:14.560 --> 05:20.720 So I think it's the same answer that you use it for both. 05:20.720 --> 05:24.800 We've thought a lot about this during these 10 years at Spotify, obviously. 05:25.360 --> 05:27.840 In one sense, as you said, music is incredibly 05:27.840 --> 05:29.760 social, you go to concerts and so forth. 05:30.480 --> 05:38.400 On the other hand, it is your escape and everyone has these things that are very personal to them. 05:38.400 --> 05:47.680 So what we've found is that when it comes to, most people claim that they have a friend or two 05:47.680 --> 05:50.880 that they are heavily inspired by and that they listen to. 05:50.880 --> 05:54.560 So I actually think music is very social, but in a smaller group setting, 05:54.560 --> 06:00.400 it's an intimate form of, it's an intimate relationship. 06:00.400 --> 06:03.360 It's not something that you necessarily share broadly. 06:03.360 --> 06:07.040 Now, at concerts, you can argue you do, but then you've gathered a lot of people 06:07.040 --> 06:08.880 that you have something in common with. 06:08.880 --> 06:16.960 I think this broadcast sharing of music is something we tried on social networks and so forth. 06:16.960 --> 06:23.120 But it turns out that people aren't super interested in sharing their music. 06:23.120 --> 06:26.960 They aren't super interested in what their friends listen to. 06:28.480 --> 06:32.800 They're interested in understanding if they have something in common perhaps with a friend, 06:32.800 --> 06:35.040 but not just as information. 06:35.680 --> 06:37.280 Right, that's really interesting. 06:38.000 --> 06:40.880 I was just thinking of it this morning, listening to Spotify. 06:41.600 --> 06:48.480 I really have a pretty intimate relationship with Spotify, with my playlists, right? 06:48.480 --> 06:53.360 I've had them for many years now and they've grown with me together. 06:53.360 --> 06:59.520 There's an intimate relationship you have with a library of music that you've developed. 06:59.520 --> 07:01.920 And we'll talk about different ways we can play with that. 07:02.480 --> 07:08.240 Can you do the impossible task and try to give a history of music listening 07:09.280 --> 07:14.160 from your perspective from before the internet and after the internet 07:14.160 --> 07:18.800 and just kind of everything leading up to streaming with Spotify and so on? 07:18.800 --> 07:19.280 I'll try. 07:19.280 --> 07:20.880 It could be a 100 year podcast. 07:22.320 --> 07:24.400 I'll try to do a brief version. 07:24.400 --> 07:28.080 There are some things that I think are very interesting during the history of music, 07:28.080 --> 07:33.040 which is that before recorded music, to be able to enjoy music, 07:33.040 --> 07:35.440 you actually had to be where the music was produced 07:35.440 --> 07:38.640 because you couldn't record it and time shift it, right? 07:38.640 --> 07:41.520 Creation and consumption had to happen at the same time, basically concerts. 07:41.520 --> 07:46.320 And so you either had to get to the nearest village to listen to music. 07:46.320 --> 07:50.880 And while that was cumbersome and it severely limited the distribution of music, 07:51.440 --> 07:53.200 it also had some different qualities, 07:53.200 --> 07:56.640 which was that the creator could always interact with the audience. 07:56.640 --> 07:57.600 It was always live. 07:58.400 --> 08:00.640 And also there was no time cap on the music. 08:00.640 --> 08:04.960 So I think it's not a coincidence that these early classical works, 08:04.960 --> 08:06.640 they're much longer than the three minutes. 08:06.640 --> 08:11.600 The three minutes came in as a restriction of the first wax disc that could only contain 08:11.600 --> 08:14.080 a three minute song on one side, right? 08:14.080 --> 08:20.400 So actually the recorded music severely limited or put constraints. 08:20.400 --> 08:21.040 I won't say limit. 08:21.040 --> 08:22.160 I mean, constraints are often good, 08:22.160 --> 08:24.960 but it put very hard constraints on the music format. 08:24.960 --> 08:30.400 So you kind of said, instead of doing this opus on many tens of minutes or something, 08:31.200 --> 08:34.560 now you get three and a half minutes because then you're out of wax on this disc. 08:34.560 --> 08:37.680 But in return, you get an amazing distribution. 08:37.680 --> 08:39.440 Your reach will widen, right? 08:39.440 --> 08:40.880 Just on that point real quick. 08:42.560 --> 08:47.360 Without the mass scale distribution, there's a scarcity component 08:47.920 --> 08:50.720 where you kind of look forward to it. 08:51.760 --> 08:56.400 We had that, it's like the Netflix versus HBO Game of Thrones. 08:56.400 --> 09:00.160 You like wait for the event because you can't really listen to it. 09:00.160 --> 09:02.800 So you like look forward to it and then it's like, 09:02.800 --> 09:07.920 you derive perhaps more pleasure because it's more rare for you to listen to a particular piece. 09:07.920 --> 09:09.920 You think there's value to that scarcity? 09:10.480 --> 09:12.720 Yeah, I think that that is definitely a thing. 09:12.720 --> 09:17.200 And there's always this component of if you have something in infinite amounts, 09:17.200 --> 09:19.120 will you value it as much? 09:20.000 --> 09:20.880 Probably not. 09:20.880 --> 09:24.400 Humanity is always seeking some, it's relative. 09:24.400 --> 09:25.840 So you're always seeking something you didn't have. 09:25.840 --> 09:27.600 And when you have it, you don't appreciate it as much. 09:27.600 --> 09:29.520 So I think that's probably true. 09:29.520 --> 09:31.200 But I think that that's probably true. 09:31.200 --> 09:33.040 But I think that's why concerts exist. 09:33.040 --> 09:34.560 So you can actually have both. 09:35.520 --> 09:42.000 But I think net, if you couldn't listen to music in your car driving, that'd be worse. 09:42.000 --> 09:46.240 That cost will be bigger than the benefit of the anticipation I think that you would have. 09:47.360 --> 09:50.720 So, yeah, it started with live concerts. 09:50.720 --> 09:56.720 Then it's being able to, you know, the phonograph invented, right? 09:56.720 --> 09:59.440 That you start to be able to record music. 09:59.440 --> 09:59.840 Exactly. 09:59.840 --> 10:04.560 So then you got this massive distribution that made it possible to create two things. 10:04.560 --> 10:09.760 I think, first of all, cultural phenomenons, they probably need distribution to be able to happen. 10:10.560 --> 10:15.520 But it also opened access to, you know, for a new kind of artist. 10:15.520 --> 10:18.720 So you started to have these phenomenons like Beatles and Elvis and so forth. 10:18.720 --> 10:23.680 That would really, a function of distribution, I think, obviously of talent and innovation. 10:23.680 --> 10:25.200 But there was also technical component. 10:25.760 --> 10:29.040 And of course, the next big innovation to come along was radio. 10:29.040 --> 10:29.680 Broadcast radio. 10:30.720 --> 10:36.240 And I think radio is interesting because it started not as a music medium. 10:36.240 --> 10:39.600 It started as an information medium for news. 10:39.600 --> 10:45.280 And then radio needed to find something to fill the time with so that they could honestly 10:45.280 --> 10:46.720 play more ads and make more money. 10:47.200 --> 10:48.480 And music was free. 10:48.480 --> 10:52.480 So then you had this massive distribution where you could program to people. 10:52.480 --> 10:59.200 I think those things, that ecosystem, is what created the ability for hits. 10:59.200 --> 11:01.600 But it was also a very broadcast medium. 11:01.600 --> 11:06.000 So you would tend to get these massive, massive hits, but maybe not such a long tail. 11:07.440 --> 11:10.480 In terms of choice of everybody listens to the same stuff. 11:10.480 --> 11:10.960 Yeah. 11:10.960 --> 11:13.840 And as you said, I think there are some social benefits to that. 11:14.720 --> 11:19.760 I think, for example, there's a high statistical chance that if I talk about the latest episode 11:19.760 --> 11:22.640 of Game of Thrones, we have something to talk about, just statistically. 11:23.280 --> 11:26.240 In the age of individual choice, maybe some of that goes away. 11:26.240 --> 11:35.120 So I do see the value of shared cultural components, but I also obviously love personalization. 11:36.400 --> 11:39.120 And so let's catch this up to the internet. 11:39.120 --> 11:44.640 So maybe Napster, well, first of all, there's MP3s, tapes, CDs. 11:44.640 --> 11:47.440 There was a digitalization of music with a CD, really. 11:47.440 --> 11:50.320 It was physical distribution, but the music became digital. 11:51.200 --> 11:55.840 And so they were files, but basically boxed software, to use a software analogy. 11:56.800 --> 11:58.800 And then you could start downloading these files. 11:59.920 --> 12:02.480 And I think there are two interesting things that happened. 12:02.480 --> 12:07.120 Back to music used to be longer before it was constrained by the distribution medium. 12:08.080 --> 12:09.840 I don't think that was a coincidence. 12:09.840 --> 12:15.600 And then really the only music genre to have developed mostly after music was a file again 12:15.600 --> 12:17.360 on the internet is EDM. 12:17.360 --> 12:20.640 And EDM is often much longer than the traditional music. 12:20.640 --> 12:25.200 I think it's interesting to think about the fact that music is no longer constrained in 12:26.000 --> 12:27.040 minutes per song or something. 12:27.040 --> 12:31.120 It's a legacy of an old distribution technology. 12:31.120 --> 12:33.680 And you see some of this new music that breaks the format. 12:33.680 --> 12:38.160 Not so much as I would have expected actually by now, but it still happens. 12:38.160 --> 12:41.120 So first of all, I don't really know what EDM is. 12:41.120 --> 12:42.320 Electronic dance music. 12:42.320 --> 12:42.880 Yeah. 12:42.880 --> 12:44.160 You could say Avicii. 12:44.160 --> 12:46.800 Avicii was one of the biggest in this genre. 12:46.800 --> 12:49.680 So the main constraint is of time. 12:49.680 --> 12:52.480 Something like a three, four, five minute song. 12:52.480 --> 12:55.760 So you could have songs that were eight minutes, 10 minutes and so forth. 12:56.320 --> 13:01.040 Because it started as a digital product that you downloaded. 13:01.040 --> 13:02.880 So you didn't have this constraint anymore. 13:03.920 --> 13:07.440 So I think it's something really interesting that I don't think has fully happened yet. 13:08.480 --> 13:12.880 We're kind of jumping ahead a little bit to where we are, but I think there's tons of format 13:12.880 --> 13:18.880 innovation in music that should happen now, that couldn't happen when you needed to really 13:18.880 --> 13:20.880 adhere to the distribution constraints. 13:20.880 --> 13:24.240 If you didn't adhere to that, you would get no distribution. 13:24.240 --> 13:30.720 So Björk, for example, the Icelandic artist, she made a full iPad app as an album. 13:30.720 --> 13:31.920 That was very expensive. 13:33.440 --> 13:38.000 Even though the app store has great distribution, she gets nowhere near the distribution versus 13:38.000 --> 13:39.760 staying within the three minute format. 13:39.760 --> 13:44.720 So I think now that music is fully digital inside these streaming services, there is 13:44.720 --> 13:50.080 the opportunity to change the format again and allow creators to be much more creative 13:50.080 --> 13:52.800 without limiting their distribution ability. 13:52.800 --> 13:54.960 That's interesting that you're right. 13:54.960 --> 13:59.280 It's surprising that we don't see that taken advantage more often. 13:59.280 --> 14:06.400 It's almost like the constraints of the distribution from the 50s and 60s have molded the culture 14:06.400 --> 14:12.480 to where we want the five, three to five minute song than anything else, not just. 14:12.480 --> 14:18.880 So we want the song as consumers and as artists, because I write a lot of music and I never 14:18.880 --> 14:23.600 even thought about writing something longer than 10 minutes. 14:23.600 --> 14:26.640 It's really interesting that those constraints. 14:26.640 --> 14:29.600 Because all your training data has been three and a half minute songs, right? 14:29.600 --> 14:30.320 It's right. 14:30.320 --> 14:36.480 Okay, so yes, digitization of data led to then mp3s. 14:36.480 --> 14:42.240 Yeah, so I think you had this file then that was distributed physically, but then you had 14:42.240 --> 14:46.800 the components of digital distribution and then the internet happened and there was this 14:46.800 --> 14:51.120 vacuum where you had a format that could be digitally shipped, but there was no business 14:51.120 --> 14:51.840 model. 14:51.840 --> 14:58.880 And then all these pirate networks happened, Napster and in Pirate Island. 14:58.880 --> 15:02.960 Napster and in Sweden Pirate Bay, which was one of the biggest. 15:02.960 --> 15:10.080 And I think from a consumer point of view, which kind of leads up to the inception of 15:10.080 --> 15:15.840 Spotify, from a consumer point of view, consumers for the first time had this access model to 15:15.840 --> 15:25.680 music where they could, without kind of any marginal cost, they could try different tracks. 15:25.680 --> 15:27.360 You could use music in new ways. 15:27.360 --> 15:28.880 There was no marginal cost. 15:28.880 --> 15:32.480 And that was a fantastic consumer experience to have access to all the music ever made, 15:32.480 --> 15:33.920 I think was fantastic. 15:34.560 --> 15:38.000 But it was also horrible for artists because there was no business model around it. 15:38.000 --> 15:39.600 So they didn't make any money. 15:39.600 --> 15:46.400 So the user need almost drove the user interface before there was a business model. 15:46.400 --> 15:52.160 And then there were these download stores that allowed you to download files, which 15:52.160 --> 15:55.040 was a solution, but it didn't solve the access problem. 15:55.040 --> 15:58.560 There was still a marginal cost of 99 cents to try one more track. 15:58.560 --> 16:01.920 And I think that that heavily limits how you listen to music. 16:01.920 --> 16:07.600 The example I always give is, you know, in Spotify, a huge amount of people listen to 16:07.600 --> 16:10.320 music while they sleep, while they go to sleep and while they sleep. 16:11.280 --> 16:14.960 If that costed you 99 cents per three minutes, you probably wouldn't do that. 16:15.520 --> 16:18.640 And you would be much less adventurous if there was a real dollar cost to exploring 16:18.640 --> 16:19.200 music. 16:19.200 --> 16:22.320 So the access model is interesting in that it changes your music behavior. 16:22.320 --> 16:26.560 You can be, you can take much more risk because there's no marginal cost to it. 16:27.680 --> 16:32.320 Maybe let me linger on piracy for a second, because I find, especially coming from Russia, 16:33.200 --> 16:36.560 piracy is something that's very interesting to me. 16:39.440 --> 16:49.040 Not me, of course, ever, but I have friends who have partook in piracy of music, software, 16:49.040 --> 16:51.600 TV shows, sporting events. 16:52.400 --> 16:57.920 And usually to me, what that shows is not that they're, they can actually pay the money 16:58.400 --> 16:59.600 and they're not trying to save money. 17:00.480 --> 17:02.800 They're choosing the best experience. 17:03.760 --> 17:08.560 So what to me, piracy shows is a business opportunity in all these domains. 17:08.560 --> 17:11.120 And that's where I think you're right. 17:11.120 --> 17:15.840 Spotify stepped in is basically piracy was an experience. 17:15.840 --> 17:23.520 You can explore with fine music you like, and actually the interface of piracy is horrible 17:23.520 --> 17:29.680 because it's, I mean, it's bad metadata, long download times, all kinds of stuff. 17:29.680 --> 17:37.520 And what Spotify does is basically first rewards artists and second makes the experience of 17:37.520 --> 17:38.720 exploring music much better. 17:38.720 --> 17:42.560 I mean, the same is true, I think for movies and so on. 17:42.560 --> 17:48.080 That piracy reveals in the software space, for example, I'm a huge user and fan of Adobe 17:48.080 --> 17:54.720 products and there was much more incentive to pirate Adobe products before they went 17:54.720 --> 17:56.400 to a monthly subscription plan. 17:57.120 --> 18:04.640 And now all of the said friends that used to pirate Adobe products that I know now actually 18:04.640 --> 18:06.880 pay gladly for the monthly subscription. 18:06.880 --> 18:08.000 Yeah, I think you're right. 18:08.000 --> 18:11.360 I think it's a sign of an opportunity for product development. 18:11.360 --> 18:19.120 And that sometimes there's a product market fit before there's a business model fit in 18:19.120 --> 18:19.840 product development. 18:19.840 --> 18:21.760 I think that's a sign of it. 18:21.760 --> 18:24.320 In Sweden, I think it was a bit of both. 18:24.320 --> 18:30.480 There was a culture where we even had a political party called the Pirate Party. 18:30.480 --> 18:35.120 And this was during the time when people said that information should be free. 18:35.120 --> 18:38.080 It was somehow wrong to charge for ones and zeros. 18:38.080 --> 18:43.600 So I think people felt that artists should probably make some money somehow else and 18:43.600 --> 18:44.880 concerts or something. 18:44.880 --> 18:49.920 So at least in Sweden, it was part really social acceptance, even at the political level. 18:49.920 --> 18:56.800 But that also forced Spotify to compete with free, which I don't think would actually 18:56.800 --> 18:58.560 could have happened anywhere else in the world. 18:58.560 --> 19:03.120 The music industry needed to be doing bad enough to take that risk. 19:03.120 --> 19:04.800 And Sweden was like the perfect testing ground. 19:04.800 --> 19:10.640 It had government funded high bandwidth, low latency broadband, which meant that the product 19:10.640 --> 19:11.440 would work. 19:11.440 --> 19:14.000 And it was also there was no music revenue anyway. 19:14.000 --> 19:17.600 So they were kind of like, I don't think this is going to work, but why not? 19:18.800 --> 19:21.920 So this product is one that I don't think could have happened in America, the world's 19:21.920 --> 19:23.200 largest music market, for example. 19:23.920 --> 19:25.600 So how do you compete with free? 19:25.600 --> 19:30.640 Because that's an interesting world of the internet where most people don't like to 19:30.640 --> 19:31.520 pay for things. 19:31.520 --> 19:35.360 So Spotify steps in and tries to, yes, compete with free. 19:36.080 --> 19:36.640 How do you do it? 19:37.120 --> 19:38.240 So I think two things. 19:38.240 --> 19:41.680 One is people are starting to pay for things on the internet. 19:41.680 --> 19:47.440 I think one way to think about it was that advertising was the first business model because 19:47.440 --> 19:49.200 no one would put a credit card on the internet. 19:49.200 --> 19:51.040 Transactional with Amazon was the second. 19:51.600 --> 19:52.960 And maybe subscription is the third. 19:52.960 --> 19:55.680 And if you look offline, subscription is the biggest of those. 19:56.480 --> 19:57.600 So that may still happen. 19:57.600 --> 19:59.040 I think people are starting to pay for things. 19:59.040 --> 20:01.680 But definitely back then, we needed to compete with free. 20:02.480 --> 20:07.600 And the first thing you need to do is obviously to lower the price to free and then you need 20:07.600 --> 20:09.440 to be better somehow. 20:09.440 --> 20:15.040 And the way that Spotify was better was on the user experience, on the actual performance, 20:15.040 --> 20:24.640 the latency of, you know, even if you had high bandwidth broadband, it would still take 20:24.640 --> 20:30.800 you 30 seconds to a minute to download one of these tracks. 20:30.800 --> 20:35.360 So the Spotify experience of starting within the perceptual limit of immediacy, about 250 20:35.360 --> 20:41.520 milliseconds, meant that the whole trick was it felt as if you had downloaded all of Pirate 20:41.520 --> 20:41.680 Bay. 20:41.680 --> 20:42.800 It was on your hard drive. 20:42.800 --> 20:44.400 It was that fast, even though it wasn't. 20:45.360 --> 20:46.720 And it was still free. 20:46.720 --> 20:50.400 But somehow you were actually still being a legal citizen. 20:50.400 --> 20:54.160 And that was the trick that Spotify managed to pull off. 20:54.880 --> 20:58.240 So I've actually heard you say this or write this. 20:58.240 --> 21:02.400 And I was surprised that I wasn't aware of it because I just took it for granted. 21:02.400 --> 21:05.920 You know, whenever an awesome thing comes along, you're just like, of course, it has 21:05.920 --> 21:06.480 to be this way. 21:07.360 --> 21:08.560 That's exactly right. 21:08.560 --> 21:14.720 That it felt like the entire world's libraries at my fingertips because of that latency being 21:14.720 --> 21:15.440 reduced. 21:15.440 --> 21:18.640 What was the technical challenge in reducing the latency? 21:18.640 --> 21:25.280 So there was a group of really, really talented engineers, one of them called Ludwig Strigius. 21:25.280 --> 21:32.080 He wrote the, actually from Gothenburg, he wrote the initial, the uTorrent client, which 21:32.080 --> 21:37.760 is kind of an interesting backstory to Spotify, that we have one of the top developers from 21:38.480 --> 21:39.840 uTorrent clients as well. 21:39.840 --> 21:42.320 So he wrote uTorrent, the world's smallest uTorrent client. 21:42.320 --> 21:49.440 And then he was acquired very early by Daniel and Martin, who founded Spotify, and they 21:49.440 --> 21:53.040 actually sold the uTorrent client to BitTorrent, but kept Ludwig. 21:53.040 --> 21:58.240 So Spotify had a lot of experience within peer to peer networking. 21:59.040 --> 22:04.560 So the original innovation was a distribution innovation, where Spotify built an end to 22:04.560 --> 22:08.160 end media distribution system up until only a few years ago, we actually hosted all the 22:08.160 --> 22:09.440 music ourselves. 22:09.440 --> 22:13.360 So we had both the service side and the client, and that meant that we could do things such 22:13.360 --> 22:19.200 as having a peer to peer solution to use local caching on the client side, because back then 22:19.200 --> 22:20.800 the world was mostly desktop. 22:20.800 --> 22:26.240 But we could also do things like hack the TCP protocols, things like Nagel's algorithm 22:26.240 --> 22:31.200 for kind of exponential back off, or ramp up and just go full throttle and optimize 22:31.200 --> 22:33.760 for latency at the cost of bandwidth. 22:33.760 --> 22:39.200 And all of this end to end control meant that we could do an experience that felt like a 22:39.200 --> 22:40.480 step change. 22:40.480 --> 22:46.720 These days, we actually are on GCP, we don't host our own stuff, and everyone is really 22:46.720 --> 22:47.360 fast these days. 22:47.360 --> 22:49.440 So that was the initial competitive advantage. 22:49.440 --> 22:51.440 But then obviously, you have to move on over time. 22:51.440 --> 22:54.480 And that was over 10 years ago, right? 22:54.480 --> 22:55.840 That was in 2008. 22:55.840 --> 22:57.520 The product was launched in Sweden. 22:57.520 --> 22:59.440 It was in a beta, I think, 2007. 22:59.440 --> 23:00.800 And it was on the desktop, right? 23:00.800 --> 23:01.840 It was desktop only. 23:01.840 --> 23:03.840 There's no phone. 23:03.840 --> 23:04.480 There was no phone. 23:04.480 --> 23:07.920 The iPhone came out in 2008. 23:07.920 --> 23:10.480 But the App Store came out one year later, I think. 23:10.480 --> 23:13.120 So the writing was on the wall, but there was no phone yet. 23:14.160 --> 23:19.680 You've mentioned that people would use Spotify to discover the songs they like, and then 23:19.680 --> 23:24.880 they would torrent those songs to so they can copy it to their phone. 23:24.880 --> 23:25.840 Just hilarious. 23:25.840 --> 23:26.320 Exactly. 23:26.320 --> 23:27.440 Not torrent, pirate. 23:27.440 --> 23:32.800 Seriously, piracy does seem to be like a good guide for business models. 23:33.520 --> 23:34.560 Video content. 23:34.560 --> 23:37.600 As far as I know, Spotify doesn't have video content. 23:37.600 --> 23:42.080 Well, we do have music videos, and we do have videos on the service. 23:42.080 --> 23:48.320 But the way we think about ourselves is that we're an audio service, and we think that 23:48.320 --> 23:52.800 if you look at the amount of time that people spend on audio, it's actually very similar 23:52.800 --> 23:55.200 to the amount of time that people spend on music. 23:55.200 --> 23:58.640 It's very similar to the amount of time that people spend on video. 23:58.640 --> 24:02.000 So the opportunity should be equally big. 24:02.000 --> 24:03.520 But today, it's not at all valued. 24:03.520 --> 24:05.040 Videos value much higher. 24:05.040 --> 24:08.320 So we think it's basically completely undervalued. 24:08.320 --> 24:10.560 So we think of ourselves as an audio service. 24:10.560 --> 24:14.000 But within that audio service, I think video can make a lot of sense. 24:14.000 --> 24:19.040 I think when you're discovering an artist, you probably do want to see them and understand 24:19.040 --> 24:21.200 who they are, to understand their identity. 24:21.200 --> 24:22.400 You won't see that video every time. 24:22.400 --> 24:25.120 90% of the time, the phone is going to be in your pocket. 24:25.120 --> 24:27.280 For podcasters, you use video. 24:27.280 --> 24:28.560 I think that can make a ton of sense. 24:28.560 --> 24:33.600 So we do have video, but we're an audio service where, think of it as we call it internally, 24:33.600 --> 24:35.120 backgroundable video. 24:35.120 --> 24:38.720 Video that is helpful, but isn't the driver of the narrative. 24:39.440 --> 24:48.560 I think also, if we look at YouTube, there's quite a few folks who listen to music on YouTube. 24:48.560 --> 24:55.280 So in some sense, YouTube is a bit of a competitor to Spotify, which is very strange to me that 24:55.280 --> 24:57.360 people use YouTube to listen to music. 24:57.920 --> 25:00.640 They play essentially the music videos, right? 25:00.640 --> 25:03.360 But don't watch the videos and put it in their pocket. 25:03.360 --> 25:12.240 Well, I think it's similar to what, strangely, maybe it's similar to what we were for the 25:12.240 --> 25:20.640 piracy networks, where YouTube, for historical reasons, have a lot of music videos. 25:20.640 --> 25:25.040 So people use YouTube for a lot of the discovery part of the process, I think. 25:25.040 --> 25:29.520 But then it's not a really good sort of, quote unquote, MP3 player, because it doesn't even 25:29.520 --> 25:29.920 background. 25:29.920 --> 25:31.600 Then you have to keep the app in the foreground. 25:31.600 --> 25:36.160 So it's not a good consumption tool, but it's a decently good discovery. 25:36.160 --> 25:37.840 I mean, I think YouTube is a fantastic product. 25:38.400 --> 25:40.320 And I use it for all kinds of purposes. 25:40.320 --> 25:41.040 That's true. 25:41.040 --> 25:46.560 If I were to admit something, I do use YouTube a little bit to assist in the discovery process 25:46.560 --> 25:47.280 of songs. 25:47.280 --> 25:50.320 And then if I like it, I'll add it to Spotify. 25:50.320 --> 25:51.760 But that's OK. 25:51.760 --> 25:52.560 That's OK with us. 25:53.600 --> 25:55.520 OK, so sorry, we're jumping around a little bit. 25:55.520 --> 25:57.920 So it's kind of incredible. 25:58.560 --> 26:01.440 You look at Napster, you look at the early days of Spotify. 26:03.440 --> 26:06.080 One fascinating point is how do you grow a user base? 26:06.640 --> 26:08.320 So you're there in Sweden. 26:08.960 --> 26:10.320 You have an idea. 26:10.320 --> 26:12.480 I saw the initial sketches that look terrible. 26:14.160 --> 26:18.240 How do you grow a user base from a few folks to millions? 26:19.280 --> 26:21.680 I think there are a bunch of tactical answers. 26:22.240 --> 26:24.160 So first of all, I think you need a great product. 26:24.160 --> 26:30.080 I don't think you take a bad product and market it to be successful. 26:30.080 --> 26:31.120 So you need a great product. 26:31.120 --> 26:34.720 But sorry to interrupt, but it's a totally new way to listen to music, too. 26:34.720 --> 26:38.560 So it's not just did people realize immediately that Spotify is a great product? 26:38.560 --> 26:40.240 No, I think they did. 26:40.240 --> 26:45.280 So back to the point of piracy, it was a totally new way to listen to music legally. 26:45.840 --> 26:48.960 But people had been used to the access model in Sweden 26:48.960 --> 26:50.880 and the rest of the world for a long time through piracy. 26:50.880 --> 26:54.160 So one way to think about Spotify, it was just legal and fast piracy. 26:54.720 --> 26:56.240 And so people have been using it for a long time. 26:56.960 --> 26:59.040 So they weren't alien to it. 26:59.040 --> 27:01.360 They didn't really understand how it could be illegal 27:01.360 --> 27:03.920 because it seemed too fast and too good to be true, 27:03.920 --> 27:06.960 which I think is a great product proposition if you can be too good to be true. 27:06.960 --> 27:09.760 But what I saw again and again was people showing each other, 27:09.760 --> 27:13.200 clicking the song, showing how fast it started and say, can you believe this? 27:13.200 --> 27:16.320 So I really think it was about speed. 27:16.320 --> 27:22.000 Then we also had an invite program that was really meant for scaling 27:22.000 --> 27:23.280 because we hosted our own service. 27:23.280 --> 27:25.040 We needed to control scaling. 27:25.040 --> 27:27.600 But that built a lot of expectation. 27:27.600 --> 27:32.880 And I don't want to say hype because hype implies that it wasn't true. 27:32.880 --> 27:38.560 Excitement around the product. And we've replicated that when we launched in the US. 27:38.560 --> 27:41.200 We also built up an invite only program first. 27:41.200 --> 27:46.160 There are lots of tactics, but I think you need a great product to solve some problem. 27:46.160 --> 27:51.440 And basically the key innovation, there was technology, 27:51.440 --> 27:55.600 but on a meta level, the innovation was really the access model versus the ownership model. 27:55.600 --> 27:56.880 And that was tricky. 27:56.880 --> 28:01.440 A lot of people said that they wanted to be able to do it. 28:01.440 --> 28:03.680 I mean, they wanted to own their music. 28:04.480 --> 28:07.520 They would never kind of rent it or borrow it. 28:07.520 --> 28:09.120 But I think the fact that we had a free tier, 28:09.120 --> 28:14.000 which meant that you get to keep this music for life as well, helped quite a lot. 28:14.560 --> 28:18.560 So this is an interesting psychological point that maybe you can speak to. 28:18.560 --> 28:20.080 It was a big shift for me. 28:22.240 --> 28:24.800 It's almost like I had to go to therapy for this. 28:26.240 --> 28:29.360 I think I would describe my early listening experience, 28:29.360 --> 28:32.480 and I think a lot of my friends do, as basically hoarding music. 28:33.280 --> 28:35.920 As you're like slowly, one song by one song, 28:35.920 --> 28:39.920 or maybe albums, gathering a collection of music that you love. 28:40.960 --> 28:42.080 And you own it. 28:42.080 --> 28:46.160 It's like often, especially with CDs or tape, you like physically had it. 28:46.960 --> 28:50.240 And what Spotify, what I had to come to grips with, 28:50.240 --> 28:55.520 it was kind of liberating actually, is to throw away all the music. 28:55.520 --> 28:58.480 I've had this therapy session with lots of people. 28:58.480 --> 29:02.560 And I think the mental trick is, so actually we've seen the user data. 29:02.560 --> 29:05.040 When Spotify started, a lot of people did the exact same thing. 29:05.040 --> 29:08.240 They started hoarding as if the music would disappear. 29:09.280 --> 29:10.880 Almost the equivalent of downloading. 29:10.880 --> 29:16.080 And so we had these playlists that had limits of like a few hundred thousand tracks. 29:16.080 --> 29:17.360 We figured no one will ever. 29:17.360 --> 29:18.560 Well, they do. 29:18.560 --> 29:20.960 Nuts and hundreds and hundreds of thousands of tracks. 29:20.960 --> 29:25.760 And to this day, some people want to actually save, quote unquote, 29:25.760 --> 29:26.960 and then play the entire catalog. 29:26.960 --> 29:32.880 But I think the therapy session goes something like instead of throwing away your music, 29:34.080 --> 29:37.760 if you took your files and you stored them in the locker at Google, 29:38.720 --> 29:39.680 it'd be a streaming service. 29:39.680 --> 29:42.720 It's just that in that locker, you have all the world's music now for free. 29:42.720 --> 29:45.520 So instead of giving away your music, you got all the music. 29:45.520 --> 29:46.720 It's yours. 29:46.720 --> 29:50.240 You could think of it as having a copy of the world's catalog there forever. 29:50.240 --> 29:52.720 So you actually got more music instead of less. 29:52.720 --> 29:58.720 It's just that you just took that hard disk and you sent it to someone who stored it for you. 29:58.720 --> 30:01.440 And once you go through that mental journey, I'm like, it's still my files. 30:01.440 --> 30:02.560 They're just over there. 30:02.560 --> 30:05.520 And I just have 40 million or 50 million or something now. 30:05.520 --> 30:07.600 Then people are like, OK, that's good. 30:07.600 --> 30:10.880 The problem is, I think, because you paid us a subscription, 30:11.840 --> 30:14.000 if we hadn't had the free tier where you would feel like, 30:14.000 --> 30:17.120 even if I don't want to pay anymore, I still get to keep them. 30:17.120 --> 30:18.480 You keep your playlist forever. 30:18.480 --> 30:20.240 They don't disappear even though you stop paying. 30:20.240 --> 30:21.760 I think that was really important. 30:21.760 --> 30:25.440 If we would have started as, you know, you can put in all this time, 30:25.440 --> 30:27.280 but if you stop paying, you lose all your work. 30:27.280 --> 30:31.760 I think that would have been a big challenge and was the big challenge for a lot of our competitors. 30:31.760 --> 30:34.880 That's another reason why I think the free tier is really important. 30:34.880 --> 30:37.600 That people need to feel the security, that the work they put in, 30:37.600 --> 30:39.920 it will never disappear, even if they decide not to pay. 30:40.800 --> 30:42.880 I like how you put the work you put in. 30:42.880 --> 30:44.480 I actually stopped even thinking of it that way. 30:44.480 --> 30:50.080 I just actually Spotify taught me to just enjoy music as opposed to. 30:50.080 --> 30:57.200 As opposed to what I was doing before, which is like in an unhealthy way, hoarding music. 30:58.560 --> 31:01.280 Which I found that because I was doing that, 31:01.280 --> 31:06.880 I was listening to a small selection of songs way too much to where I was getting sick of them. 31:07.520 --> 31:11.680 Whereas Spotify, the more liberating kind of approach is I was just enjoying. 31:11.680 --> 31:13.920 Of course, I listened to Stairway to Heaven over and over, 31:13.920 --> 31:18.240 but because of the extra variety, I don't get as sick of them. 31:18.240 --> 31:20.640 There's an interesting statistic I saw. 31:21.520 --> 31:26.640 So Spotify has, maybe you can correct me, but over 50 million songs, tracks, 31:27.600 --> 31:30.000 and over 3 billion playlists. 31:31.360 --> 31:35.520 So 50 million songs and 3 billion playlists. 31:35.520 --> 31:37.600 60 times more playlist songs. 31:38.480 --> 31:39.360 What do you make of that? 31:39.920 --> 31:40.160 Yeah. 31:40.160 --> 31:48.320 So the way I think about it is that from a statistician or machine learning point of view, 31:48.320 --> 31:52.000 you have all these, if you want to think about reinforcement learning, 31:52.000 --> 31:54.320 you have this state space of all the tracks. 31:54.320 --> 31:57.280 You can take different journeys through this world. 32:00.160 --> 32:05.200 I think of these as people helping themselves and each other, 32:05.200 --> 32:08.720 creating interesting vectors through this space of tracks. 32:08.720 --> 32:14.080 And then it's not so surprising that across many tens of millions of atomic units, 32:14.080 --> 32:17.280 there will be billions of paths that make sense. 32:17.280 --> 32:21.920 And we're probably pretty quite far away from having found all of them. 32:21.920 --> 32:26.640 So kind of our job now is users, when Spotify started, 32:26.640 --> 32:30.000 it was really a search box that was for the time pretty powerful. 32:30.000 --> 32:34.400 And then I'd like to refer to it as this programming language called playlisting, 32:34.400 --> 32:36.800 where if you, as you probably were pretty good at music, 32:36.800 --> 32:39.120 you knew your new releases, you knew your back catalog, 32:39.120 --> 32:40.480 you knew your star with the heaven, 32:40.480 --> 32:43.200 you could create a soundtrack for yourself using this playlisting tool, 32:43.200 --> 32:46.720 this like meta programming language for music to soundtrack your life. 32:47.360 --> 32:50.160 And people who were good at music, it's back to how do you scale the product. 32:50.960 --> 32:53.760 For people who are good at music, that wasn't actually enough. 32:53.760 --> 32:55.840 If you had the catalog and a good search tool, 32:55.840 --> 32:57.120 and you can create your own sessions, 32:57.120 --> 33:01.120 you could create really good a soundtrack for your entire life. 33:01.120 --> 33:04.000 Probably perfectly personalized because you did it yourself. 33:04.000 --> 33:06.880 But the problem was most people, many people aren't that good at music. 33:06.880 --> 33:08.480 They just can't spend the time. 33:08.480 --> 33:10.800 Even if you're very good at music, it's going to be hard to keep up. 33:10.800 --> 33:16.400 So what we did to try to scale this was to essentially try to build, 33:16.400 --> 33:20.480 you can think of them as agents that this friend that some people had 33:20.480 --> 33:22.800 that helped them navigate this music catalog. 33:22.800 --> 33:24.240 That's what we're trying to do for you. 33:24.800 --> 33:32.640 But also there is something like 200 million active users. 33:32.640 --> 33:34.480 1 million active users on Spotify. 33:35.040 --> 33:36.640 So there it's okay. 33:36.640 --> 33:38.720 So from the machine learning perspective, 33:39.760 --> 33:45.760 you have these 200 million people plus they're creating. 33:45.760 --> 33:49.840 It's really interesting to think of a playlist as, 33:51.760 --> 33:53.200 I mean, I don't know if you meant it that way, 33:53.200 --> 33:54.880 but it's almost like a programming language. 33:54.880 --> 34:01.120 It's or at least a trace of exploration of those individual agents. 34:01.120 --> 34:06.000 The listeners and you have all this new tracks coming in. 34:06.000 --> 34:11.680 So it's a fascinating space that is ripe for machine learning. 34:11.680 --> 34:17.440 So is there, is it possible, how can playlists be used as data 34:18.080 --> 34:23.360 in terms of machine learning and to help Spotify organize the music? 34:24.160 --> 34:29.680 So we found in our data, not surprising that people who play listed lots 34:29.680 --> 34:30.720 they retain much better. 34:30.720 --> 34:32.240 They had a great experience. 34:32.240 --> 34:35.360 And so our first attempt was to playlist for users. 34:35.920 --> 34:41.360 And so we acquired this company called Tunigo of editors and professional playlisters 34:41.360 --> 34:45.600 and kind of leveraged the maximum of human intelligence 34:45.600 --> 34:51.440 to help build kind of these vectors through the track space for people. 34:52.480 --> 34:54.320 And that broadened the product. 34:54.320 --> 34:57.840 But then the obvious next, and we use statistical means, 34:57.840 --> 35:02.080 where they could see when they created a playlist, how did that playlist perform? 35:02.080 --> 35:04.800 They could see skips of the songs, they could see how the songs perform, 35:04.800 --> 35:10.720 and they manually iterated the playlist to maximize performance for a large group of people. 35:10.720 --> 35:14.480 But there were never enough editors to playlists for you personally. 35:14.480 --> 35:17.680 So the promise of machine learning was to go from kind of group personalization 35:18.240 --> 35:22.640 using editors and tools and statistics to individualization. 35:22.640 --> 35:28.160 And then what's so interesting about the 3 billion playlists we have is we ended, 35:28.160 --> 35:29.360 the truth is we lucked out. 35:29.360 --> 35:32.880 This was not a priority strategy, as is often the case. 35:32.880 --> 35:35.920 It looks really smart in hindsight, but it was dumb luck. 35:37.440 --> 35:42.160 We looked at these playlists and we had some people in the company, 35:42.160 --> 35:43.840 a person named Eric Beranodson. 35:43.840 --> 35:48.560 He was really good at machine learning already back then in like 2007, 2008. 35:48.560 --> 35:51.600 Back then it was mostly collaborative filtering and so forth. 35:51.600 --> 35:57.920 But we realized that what this is, is people are grouping tracks for themselves 35:57.920 --> 35:59.920 that have some semantic meaning to them. 36:00.640 --> 36:04.160 And then they actually label it with a playlist name as well. 36:04.160 --> 36:09.040 So in a sense, people were grouping tracks along semantic dimensions and labeling them. 36:09.840 --> 36:15.840 And so could you use that information to find that latent embedding? 36:15.840 --> 36:19.920 And so we started playing around with collaborative filtering 36:20.960 --> 36:24.160 and we saw tremendous success with it. 36:24.160 --> 36:28.320 Basically trying to extract some of these dimensions. 36:28.320 --> 36:30.160 And if you think about it, it's not surprising at all. 36:30.880 --> 36:34.880 It'd be quite surprising if playlists were actually random, 36:34.880 --> 36:36.160 if they had no semantic meaning. 36:36.880 --> 36:39.200 For most people, they group these tracks for some reason. 36:39.840 --> 36:43.120 So we just happened across this incredible data set. 36:43.120 --> 36:46.240 Where people are taking these tens of millions of tracks 36:46.800 --> 36:49.280 and group them along different semantic vectors. 36:49.280 --> 36:52.720 And the semantics being outside the individual users. 36:52.720 --> 36:54.400 So it's some kind of universal. 36:54.400 --> 36:59.760 There's a universal embedding that holds across people on this earth. 36:59.760 --> 37:05.440 Yes, I do think that the embeddings you find are going to be reflective of the people who play listed. 37:05.440 --> 37:09.040 So if you have a lot of indie lovers who play list, 37:09.040 --> 37:13.440 your embedding is going to perform better there. 37:14.800 --> 37:20.560 But what we found was that yes, there were these latent similarities. 37:20.560 --> 37:22.000 They were very powerful. 37:22.000 --> 37:28.720 And it was interesting because I think that the people who play listed the most initially 37:28.720 --> 37:32.640 were the so called music aficionados who were really into music. 37:32.640 --> 37:34.240 And they often had a certain... 37:34.240 --> 37:38.240 Their taste was often geared towards a certain type of music. 37:38.800 --> 37:42.160 And so what surprised us, if you look at the problem from the outside, 37:42.160 --> 37:47.840 you might expect that the algorithms would start performing best with mainstreamers first. 37:47.840 --> 37:51.360 Because it somehow feels like an easier problem to solve mainstream taste 37:51.360 --> 37:52.640 than really particular taste. 37:53.360 --> 37:55.120 It was the complete opposite for us. 37:55.120 --> 37:58.640 The recommendations performed fantastically for people who saw themselves as 37:59.280 --> 38:00.960 having very unique taste. 38:00.960 --> 38:03.280 That's probably because all of them play listed. 38:03.280 --> 38:05.120 And they didn't perform so well for mainstreamers. 38:05.120 --> 38:09.440 They actually thought they were a bit too particular and unorthodox. 38:09.440 --> 38:12.000 So we had the complete opposite of what we expected. 38:12.000 --> 38:13.920 Success within the hardest problem first, 38:13.920 --> 38:16.560 and then had to try to scale to more mainstream recommendations. 38:17.600 --> 38:24.160 So you've also acquired Echo Nest that analyzes song data. 38:24.160 --> 38:28.400 So in your view, maybe you can talk about, 38:28.400 --> 38:31.680 so what kind of data is there from a machine learning perspective? 38:31.680 --> 38:35.680 From a machine learning perspective, there's a huge amount. 38:35.680 --> 38:40.640 We're talking about playlisting and just user data of what people are listening to, 38:40.640 --> 38:43.920 the playlist they're constructing, and so on. 38:44.640 --> 38:48.080 And then there's the actual data within a song. 38:48.080 --> 38:51.920 What makes a song, I don't know, the actual waveforms. 38:54.160 --> 38:55.120 How do you mix the two? 38:55.680 --> 38:57.200 How much value is there in each? 38:57.200 --> 39:03.120 To me, it seems like user data is a romantic notion 39:03.120 --> 39:05.840 that the song itself would contain useful information. 39:05.840 --> 39:09.840 But if I were to guess, user data would be much more powerful, 39:09.840 --> 39:11.840 like playlists would be much more powerful. 39:11.840 --> 39:13.680 Yeah, so we use both. 39:14.800 --> 39:18.800 Our biggest success initially was with playlist data 39:18.800 --> 39:21.920 without understanding anything about the structure of the song. 39:22.480 --> 39:25.520 But when we acquired Echo Nest, they had the inverse problem. 39:25.520 --> 39:27.440 They actually didn't have any play data. 39:27.440 --> 39:29.680 They were just, they were a provider of recommendations, 39:29.680 --> 39:31.280 but they didn't actually have any play data. 39:31.840 --> 39:35.760 So they looked at the structure of songs, sonically, 39:36.640 --> 39:40.400 and they looked at Wikipedia for cultural references and so forth, right? 39:40.400 --> 39:41.920 And did a lot of NLU and so forth. 39:41.920 --> 39:46.880 So we got that skill into the company and combined kind of our user data 39:47.600 --> 39:51.600 with their kind of content based. 39:51.600 --> 39:53.200 So you can think of it as we were user based 39:53.200 --> 39:54.880 and they were content based in their recommendations. 39:54.880 --> 39:56.960 And we combined those two. 39:56.960 --> 40:00.240 And for some cases where you have a new song that has no play data, 40:00.240 --> 40:04.960 obviously you have to try to go by either who the artist is 40:04.960 --> 40:09.760 or the sonic information in the song or what it's similar to. 40:09.760 --> 40:12.720 So there's definitely a value in both and we do a lot in both, 40:12.720 --> 40:16.080 but I would say, yes, the user data captures things 40:16.080 --> 40:19.680 that have to do with culture in the greater society 40:19.680 --> 40:23.440 that you would never see in the content itself. 40:23.440 --> 40:27.920 But that said, we have seen, we have a research lab in Paris 40:28.880 --> 40:32.960 when we can talk more about that on machine learning on the creator side, 40:32.960 --> 40:34.880 what it can do for creators, not just for the consumers, 40:35.520 --> 40:38.640 but where we looked at how does the structure of a song 40:38.640 --> 40:40.800 actually affect the listening behavior? 40:40.800 --> 40:43.120 And it turns out that there is a lot of, 40:43.120 --> 40:48.480 we can predict things like skips based on the song itself. 40:48.480 --> 40:50.880 We could say that maybe you should move that chorus a bit 40:50.880 --> 40:52.720 because your skip is going to go up here. 40:52.720 --> 40:54.400 There is a lot of latent structure in the music, 40:54.400 --> 40:57.520 which is not surprising because it is some sort of mind hack. 40:58.640 --> 41:00.960 So there should be structure. That's probably what we respond to. 41:00.960 --> 41:04.560 You just blew my mind actually from the creator perspective. 41:05.520 --> 41:07.280 So that's a really interesting topic 41:08.000 --> 41:11.920 that probably most creators aren't taking advantage of, right? 41:11.920 --> 41:15.920 So I've recently got to interact with a few folks, 41:15.920 --> 41:24.320 YouTubers who are like obsessed with this idea of what do I do 41:24.320 --> 41:27.840 to make sure people keep watching the video? 41:27.840 --> 41:32.080 And they like look at the analytics of which point do people turn it off and so on. 41:32.720 --> 41:35.040 First of all, I don't think that's healthy, 41:35.040 --> 41:37.600 but it's because you can do it a little too much. 41:38.320 --> 41:42.240 But it is a really powerful tool for helping the creative process. 41:42.240 --> 41:46.480 You just made me realize you could do the same thing for creation of music. 41:47.280 --> 41:49.360 And so is that something you've looked into? 41:51.360 --> 41:54.800 And can you speak to how much opportunity there is for that kind of thing? 41:54.800 --> 41:59.200 Yeah, so I listened to the podcast with Ziraj and I thought it was fantastic 41:59.200 --> 42:03.600 and I reacted to the same thing where he said he posted something in the morning, 42:04.160 --> 42:06.560 immediately watched the feedback where the drop off was 42:06.560 --> 42:08.400 and then responded to that in the afternoon, 42:08.400 --> 42:12.080 which is quite different from how people make podcasts, for example. 42:12.080 --> 42:12.880 Yes, exactly. 42:12.880 --> 42:15.040 I mean, the feedback loop is almost non existent. 42:15.040 --> 42:21.120 So if we back out one level, I think actually both for music and podcasts, 42:21.120 --> 42:23.600 which we also do at Spotify, 42:23.600 --> 42:27.440 I think there's a tremendous opportunity just for the creation workflow. 42:27.440 --> 42:30.960 And I think it's really interesting speaking to you who, 42:30.960 --> 42:34.160 because you're a musician, a developer, and a podcaster. 42:34.720 --> 42:36.560 If you think about those three different roles, 42:36.560 --> 42:38.880 if you make the leap as a musician, 42:38.880 --> 42:42.080 if you think about it as a software tool chain, really, 42:42.960 --> 42:46.320 your DAW with the stems, that's the IDE, right? 42:46.320 --> 42:50.400 That's where you work in source code format with what you're creating. 42:51.120 --> 42:52.320 Then you sit around and you play with that. 42:52.320 --> 42:56.960 And when you're happy, you compile that thing into some sort of AAC or MP3 or something. 42:57.520 --> 42:59.040 You do that because you get distribution. 42:59.040 --> 43:02.240 There are so many runtimes for that MP3 across the world in car stairs and stuff. 43:02.240 --> 43:03.920 So if you kind of compile this execution, 43:03.920 --> 43:08.720 you ship it out in kind of an old fashioned boxed software analogy. 43:09.280 --> 43:11.760 And then you hope for the best, right? 43:11.760 --> 43:16.080 But as a software developer, you would never do that. 43:16.080 --> 43:18.640 First, you go on GitHub and you collaborate with other creators. 43:19.440 --> 43:22.800 And then you think it'd be crazy to just ship one version of your software 43:22.800 --> 43:26.800 without doing an A B test, without any feedback loop. 43:26.800 --> 43:28.320 Issue tracking. 43:28.320 --> 43:28.880 Exactly. 43:28.880 --> 43:31.760 And then you would look at the feedback loop and say, 43:31.760 --> 43:34.160 try to optimize that thing, right? 43:34.160 --> 43:37.840 So I think if you think of it as a very specific software tool chain, 43:38.880 --> 43:42.880 it looks quite arcane, the tools that a music creator has 43:42.880 --> 43:44.480 versus what a software developer has. 43:45.360 --> 43:47.040 So that's kind of how we think about it. 43:48.400 --> 43:52.640 Why wouldn't a music creator have something like GitHub 43:52.640 --> 43:54.000 where you could collaborate much more easily? 43:54.000 --> 43:56.560 So we bought this company called Soundtrap, 43:56.560 --> 44:01.680 which has a kind of Google Docs for music approach, where you can collaborate 44:01.680 --> 44:04.880 with other people on the kind of source code format with Stems. 44:05.600 --> 44:09.600 And I think introducing things like AI tools there to help you 44:09.600 --> 44:19.280 as you're creating music, both in helping you put accompaniment to your music, 44:19.280 --> 44:24.400 like drums or something, help you master and mix automatically, 44:24.400 --> 44:26.720 help you understand how this track will perform. 44:26.720 --> 44:29.600 Exactly what you would expect as a software developer. 44:29.600 --> 44:30.880 I think it makes a lot of sense. 44:30.880 --> 44:33.520 And I think the same goes for a podcaster. 44:33.520 --> 44:36.320 I think podcasters will expect to have the same kind of feedback loop 44:36.320 --> 44:39.520 that Siraj has, like, why wouldn't you? 44:39.520 --> 44:40.800 Maybe it's not healthy, but... 44:41.520 --> 44:45.120 Sorry, I wanted to criticize the fact because you can overdo it 44:45.120 --> 44:49.760 because a lot of the, and we're in a new era of that. 44:49.760 --> 44:56.400 So you can become addicted to it and therefore, what people say, 44:56.400 --> 44:59.680 you become a slave to the YouTube algorithm or sort of, 45:00.640 --> 45:04.400 it's always a danger of a new technology as opposed to say, 45:04.400 --> 45:11.600 if you're creating a song, becoming too obsessed about the intro riff to the song 45:11.600 --> 45:15.440 that keeps people listening versus actually the entirety of the creation process. 45:15.440 --> 45:16.160 It's a balance. 45:16.160 --> 45:19.680 But the fact that there's zero, I mean, you're blowing my mind right now, 45:19.680 --> 45:24.960 because you're completely right that there is no signal whatsoever. 45:24.960 --> 45:28.960 There's no feedback whatsoever on the creation process and music or podcasting, 45:30.000 --> 45:30.880 almost at all. 45:31.680 --> 45:39.360 And are you saying that Spotify is hoping to help create tools to, not tools, but... 45:39.360 --> 45:41.680 No, tools actually. 45:41.680 --> 45:42.640 Actually, tools. 45:42.640 --> 45:47.200 Tools for creators. 45:47.200 --> 45:47.760 Absolutely. 45:48.320 --> 45:53.520 So we've made some acquisitions the last few years around music creation, 45:53.520 --> 45:57.280 this company called Soundtrap, which is a digital audio workstation, 45:57.280 --> 45:59.040 but that is browser based. 45:59.040 --> 46:01.200 And their focus was really the Google Docs approach. 46:01.200 --> 46:06.080 We can collaborate with people much more easily than you could in previous tools. 46:06.080 --> 46:09.280 So we have some of these tools that we're working with that we want to make accessible 46:09.280 --> 46:12.960 and then we can connect it with our consumption data. 46:12.960 --> 46:16.000 We can create this feedback loop where we could help you understand, 46:16.800 --> 46:20.960 we could help you create and help you understand how you will perform. 46:20.960 --> 46:24.560 We also acquired this other company within podcasting called Anchor, 46:24.560 --> 46:28.400 which is one of the biggest podcasting tools, mobile focused. 46:28.400 --> 46:32.800 So really focused on simple creation or easy access to creation. 46:32.800 --> 46:34.960 But that also gives us this feedback loop. 46:34.960 --> 46:40.640 And even before that, we invested in something called Spotify for Artists 46:40.640 --> 46:43.600 and Spotify for Podcasters, which is an app that you can download, 46:43.600 --> 46:45.360 you can verify that you are that creator. 46:46.000 --> 46:51.680 And then you get things that software developers have had for years. 46:51.680 --> 46:55.520 You can see where, if you look at your podcast, for example, on Spotify 46:55.520 --> 46:58.720 or a song that you released, you can see how it's performing, 46:58.720 --> 47:01.280 which cities it's performing in, who's listening to it, 47:01.280 --> 47:02.800 what's the demographic breakup. 47:02.800 --> 47:05.840 So similar in the sense that you can understand 47:05.840 --> 47:07.920 how you're actually doing on the platform. 47:08.880 --> 47:10.480 So we definitely want to build tools. 47:10.480 --> 47:15.200 I think you also interviewed the head of research for Adobe. 47:15.920 --> 47:19.680 And I think that's an, back to Photoshop that you like, 47:19.680 --> 47:21.680 I think that's an interesting analogy as well. 47:22.800 --> 47:28.000 Photoshop, I think, has been very innovative in helping photographers and artists. 47:28.000 --> 47:32.320 And I think there should be the same kind of tools for music creators, 47:32.320 --> 47:35.680 where you could get AI assistance, for example, as you're creating music, 47:36.640 --> 47:38.880 as you can do with Adobe, where you can, 47:38.880 --> 47:41.440 I want a sky over here and you can get help creating that sky. 47:42.000 --> 47:46.800 The really fascinating thing is what Adobe doesn't have 47:47.520 --> 47:49.760 is a distribution for the content you create. 47:50.400 --> 47:55.840 So you don't have the data of if I create, if I, you know, 47:55.840 --> 47:58.720 whatever creation I make in Photoshop or Premiere, 47:59.360 --> 48:02.480 I can't get like immediate feedback like I can on YouTube, 48:02.480 --> 48:05.360 for example, about the way people are responding. 48:05.360 --> 48:11.120 And if Spotify is creating those tools, that's a really exciting actually world. 48:11.680 --> 48:14.720 But let's talk a little about podcasts. 48:16.720 --> 48:18.720 So I have trouble talking to one person. 48:20.000 --> 48:23.120 So it's a bit terrifying and kind of hard to fathom, 48:23.120 --> 48:29.440 but on average, 60 to 100,000 people will listen to this episode. 48:30.320 --> 48:32.240 Okay, so it's intimidating. 48:32.240 --> 48:33.120 Yeah, it's intimidating. 48:34.320 --> 48:35.680 So I hosted on Blueberry. 48:36.720 --> 48:38.560 I don't know if I'm pronouncing that correctly, actually. 48:39.520 --> 48:42.400 It looks like most people listen to it on Apple Podcasts, 48:42.400 --> 48:48.480 Cast Box and Pocket Casts, and only about a thousand listen on Spotify. 48:48.480 --> 48:53.040 It's just my podcast, right? 48:53.840 --> 49:00.960 So where do you see a time when Spotify will dominate this? 49:00.960 --> 49:06.000 So Spotify is relatively new into this podcasting site. 49:06.000 --> 49:06.960 Yeah, in podcasting. 49:07.520 --> 49:09.920 What's the deal with podcasting and Spotify? 49:10.800 --> 49:13.440 How serious is Spotify about podcasting? 49:13.440 --> 49:16.800 Do you see a time where everybody would listen to, you know, 49:16.800 --> 49:21.520 probably a huge amount of people, majority perhaps listen to music on Spotify? 49:22.400 --> 49:26.880 Do you see a time when the same is true for podcasting? 49:26.880 --> 49:28.560 Well, I certainly hope so. 49:28.560 --> 49:29.360 That is our mission. 49:29.360 --> 49:34.160 Our mission as a company is actually to enable a million creators to live off of their art, 49:34.160 --> 49:35.840 and a billion people be inspired by it. 49:35.840 --> 49:40.000 And what I think is interesting about that mission is it actually puts the creators first, 49:40.640 --> 49:43.040 even though it started as a consumer focused company, 49:43.040 --> 49:44.800 and it's just to be able to live off of their art, 49:44.800 --> 49:47.280 not just make some money off of their art as well. 49:47.840 --> 49:49.920 So it's quite an ambitious project. 49:51.920 --> 49:53.920 So we think about creators of all kinds, 49:53.920 --> 50:00.160 and we kind of expanded our mission from being music to being audio a while back. 50:01.120 --> 50:07.360 And that's not so much because we think we made that decision. 50:08.400 --> 50:10.800 We think that decision was made for us. 50:10.800 --> 50:12.960 We think the world made that decision. 50:12.960 --> 50:16.560 Whether we like it or not, when you put in your headphones, 50:16.560 --> 50:24.400 you're going to make a choice between music and a new episode of your podcast or something else. 50:25.440 --> 50:26.960 We're in that world whether we like it or not. 50:26.960 --> 50:28.960 And that's how radio works. 50:28.960 --> 50:32.320 So we decided that we think it's about audio. 50:32.320 --> 50:34.480 You can see the rise of audiobooks and so forth. 50:34.480 --> 50:36.480 We think audio is a great opportunity. 50:36.480 --> 50:37.600 So we decided to enter it. 50:37.600 --> 50:45.280 And obviously, Apple and Apple Podcasts is absolutely dominating in podcasting, 50:45.280 --> 50:48.480 and we didn't have a single podcast only like two years ago. 50:49.440 --> 50:54.560 What we did though was we looked at this and said, 50:54.560 --> 50:55.920 can we bring something to this? 50:56.480 --> 50:59.200 We want to do this, but back to the original Spotify, 50:59.200 --> 51:03.840 we have to do something that consumers actually value to be able to do this. 51:03.840 --> 51:09.840 And the reason we've gone from not existing at all to being quite a wide margin, 51:09.840 --> 51:15.680 the second largest podcast consumption, still wide gap to iTunes, but we're growing quite fast. 51:16.480 --> 51:19.440 I think it's because when we looked at the consumer problem, 51:20.320 --> 51:26.960 people said surprisingly that they wanted their podcasts and music in the same application. 51:26.960 --> 51:29.760 So what we did was we took a little bit of a different approach where we said, 51:29.760 --> 51:31.440 instead of building a separate podcast app, 51:31.440 --> 51:33.680 we thought, is there a consumer problem to solve here? 51:33.680 --> 51:35.680 Because the others are very successful already. 51:35.680 --> 51:38.960 And we thought there was in making a more seamless experience 51:38.960 --> 51:43.120 where you can have your podcast and your music in the same application, 51:43.680 --> 51:45.440 because we think it's audio to you. 51:45.440 --> 51:46.800 And that has been successful. 51:46.800 --> 51:51.840 And that meant that we actually had 200 million people to offer this to instead of starting from zero. 51:52.400 --> 51:56.880 So I think we have a good chance because we're taking a different approach than the competition. 51:56.880 --> 51:59.120 And back to the other thing I mentioned about 51:59.120 --> 52:02.240 creators, because we're looking at the end to end flow. 52:02.800 --> 52:06.400 I think there's a tremendous amount of innovation to do around podcast as a format. 52:07.040 --> 52:12.640 When we have creation tools and consumption, I think we could start improving what podcasting is. 52:12.640 --> 52:18.960 I mean, podcast is this opaque, big, like one, two hour file that you're streaming, 52:19.520 --> 52:24.240 which it really doesn't make that much sense in 2019 that it's not interactive. 52:24.240 --> 52:26.000 There's no feedback loops, nothing like that. 52:26.000 --> 52:29.760 So I think if we're going to win, it's going to have to be because we build a better product 52:29.760 --> 52:31.760 for creators and for consumers. 52:32.480 --> 52:34.640 So we'll see, but it's certainly our goal. 52:34.640 --> 52:35.600 We have a long way to go. 52:36.240 --> 52:38.160 Well, the creators part is really exciting. 52:38.160 --> 52:40.160 You already, you got me hooked there. 52:40.160 --> 52:41.760 Cause the only stats I have, 52:42.320 --> 52:47.760 Blueberry just recently added the stats of whether it's listened to the end or not. 52:48.560 --> 52:52.320 And that's like a huge improvement, but that's still 52:52.320 --> 52:54.960 nowhere to where you could possibly go in terms of statistics. 52:54.960 --> 52:57.200 You just download the Spotify podcasters up and verify. 52:57.200 --> 52:59.920 And then, then you'll know where people dropped out in this episode. 52:59.920 --> 53:00.400 Oh, wow. 53:00.400 --> 53:00.900 Okay. 53:01.600 --> 53:02.800 The moment I started talking. 53:02.800 --> 53:03.360 Okay. 53:03.360 --> 53:06.800 I might be depressed by this, but okay. 53:06.800 --> 53:13.040 So one, um, one other question is the original Spotify for music. 53:14.400 --> 53:19.120 And I have a question about podcasting in this line is the idea of podcasting 53:19.120 --> 53:22.880 about podcasting in this line is the idea of albums. 53:23.440 --> 53:28.800 I have, uh, what did you, uh, music aficionados, uh, friends who are really, 53:29.440 --> 53:33.280 uh, big fans of music often, uh, really enjoy albums, 53:33.280 --> 53:35.840 listening to entire albums of, of an artist. 53:36.400 --> 53:40.960 Correct me if I'm wrong, but I feel like Spotify has helped 53:40.960 --> 53:44.240 replace the idea of an album with playlists. 53:44.240 --> 53:46.000 So you create your own albums. 53:46.000 --> 53:48.880 It's, it's kind of the way, at least I've experienced music 53:48.880 --> 53:50.480 and I've really enjoyed it that way. 53:51.040 --> 53:54.320 One of the things that was missing in podcasting for me, 53:54.880 --> 53:55.920 I don't know if it's missing. 53:56.320 --> 53:56.880 I don't know. 53:56.880 --> 53:59.920 It's an open question for me, but the way I listened to podcasts is 53:59.920 --> 54:01.600 the way I would listen to albums. 54:02.080 --> 54:05.440 So I take a Joe Rogan experience and that's an album. 54:05.600 --> 54:09.680 And I listened, you know, I like, I, I put that on and I listened one 54:09.680 --> 54:12.640 episode after the next, then there's a sequence and so on. 54:12.640 --> 54:17.520 Is there a room for doing what you did for music or doing what 54:17.520 --> 54:22.880 Spotify did for music, but, uh, creating playlists, sort of, uh, 54:22.880 --> 54:26.080 this kind of playlisting idea of breaking apart from podcasting, 54:27.120 --> 54:31.680 uh, from individual podcasts and creating kind of, uh, this interplay 54:31.680 --> 54:33.760 or, or have you thought about that space? 54:33.760 --> 54:34.800 Uh, it's a great question. 54:34.800 --> 54:38.640 So I think in, um, in music, you're right. 54:38.720 --> 54:39.920 Basically you bought an album. 54:39.920 --> 54:42.720 So it was like, you bought a small catalog of like 10 tracks, right? 54:42.800 --> 54:46.160 It was, it was, again, it was actually a lot of, a lot of consumption. 54:46.720 --> 54:49.360 You think it's about what you like, but it's based on the business model. 54:49.680 --> 54:53.920 So you paid for this 10 track service and then you listened to that for a while. 54:54.240 --> 54:57.760 And then when, when everything was flat priced, you tended to listen differently. 54:58.480 --> 55:01.360 Now, so, so I think the, I think the album is still tremendously important. 55:01.360 --> 55:03.360 That's why we have it and you can save albums and so forth. 55:03.360 --> 55:06.480 And you have a huge amount of people who really listen according to albums. 55:06.480 --> 55:09.840 And I like that because it is a creator format, you can tell a longer story 55:10.240 --> 55:11.440 over several tracks. 55:12.000 --> 55:13.840 And so some people listen to just one track. 55:13.840 --> 55:15.840 Some people actually want to hear that whole story. 55:17.520 --> 55:21.520 Now in podcast, I think, I think it's different. 55:21.600 --> 55:24.960 You can argue that podcasts might be more like shows on Netflix. 55:25.600 --> 55:29.200 Have like a full season of Narcos and you're probably not going to do like 55:29.200 --> 55:32.800 one episode of Narcos and then one of House of Cards, like, like, you know, 55:33.440 --> 55:34.480 there's a narrative there. 55:34.480 --> 55:37.440 And you, you, you love the cast and you love these characters. 55:37.440 --> 55:40.480 So I think people will, people love shows. 55:42.000 --> 55:44.800 And I think they will, they will listen to those shows. 55:44.880 --> 55:46.880 I do think you follow a bunch of shows at the same time. 55:46.880 --> 55:50.480 So there's certainly an opportunity to bring you the latest episode of, you 55:50.480 --> 55:53.040 know, whatever the five, six, 10 things that, that you're into. 55:54.560 --> 56:00.000 But, but I think, I think people are going to listen to specific hosts and love 56:00.000 --> 56:01.600 those hosts for a long time. 56:01.600 --> 56:06.880 Because I think there's something different with podcasts where, um, this 56:06.880 --> 56:11.280 format of the, the, the, the, the, the experience of the, of the audience is 56:11.280 --> 56:12.800 actually sitting here right between us. 56:13.360 --> 56:16.960 Whereas if you look at something on TV, the audio actually would come from, you 56:16.960 --> 56:20.080 would sit over there and the audio would come to you from both of us as if you 56:20.080 --> 56:22.000 were watching, not as you were part of the conversation. 56:22.560 --> 56:27.280 So my experience is having listened to podcasts like yours and Joe Rogan is, I 56:27.280 --> 56:28.720 feel like I know all of these people. 56:28.720 --> 56:30.240 They, they have a lot of experience. 56:30.240 --> 56:33.600 I know all of these people, they have no idea who I am, but I feel like I've 56:33.600 --> 56:35.040 listened to so many hours of that. 56:35.040 --> 56:38.800 It's very different from me watching a, watching like a TV show or an interview. 56:39.440 --> 56:44.560 So I think you, you kind of, um, fall in love with people and, um, experience 56:44.560 --> 56:45.760 in a, in a different way. 56:45.760 --> 56:49.280 So I think, I think shows and hosts are going to be very, uh, very important. 56:49.280 --> 56:52.160 I don't think that's going to go away into some sort of thing where, where you 56:52.160 --> 56:53.360 don't even know who you're listening to. 56:53.360 --> 56:54.320 I don't think that's going to happen. 56:55.040 --> 56:59.760 What I do think is I think there's a tremendous discovery opportunity in 56:59.760 --> 57:03.040 podcast because the catalog is growing quite quickly. 57:03.920 --> 57:10.800 And I think podcast is only a few, like five, 600,000 shows right now. 57:11.360 --> 57:16.080 If you look back to YouTube as another analogy of creators, no one really knows 57:16.080 --> 57:20.400 if you would lift the lid on YouTube, but it's probably billions of episodes. 57:21.120 --> 57:24.960 And so I think the podcast catalog would probably grow tremendously because the 57:24.960 --> 57:27.040 creation tools are getting easier. 57:27.040 --> 57:30.800 And then you're going to have this discovery opportunity that I think is 57:30.800 --> 57:31.280 really big. 57:31.280 --> 57:35.600 So, so a lot of people tell me that they love their shows, but discovering 57:35.600 --> 57:36.880 podcasts kind of suck. 57:36.880 --> 57:38.720 It's really hard to get into new show. 57:38.720 --> 57:39.840 They're usually quite long. 57:39.840 --> 57:40.960 It's a big time investment. 57:40.960 --> 57:44.080 So I think there's plenty of opportunity in the discovery part. 57:45.600 --> 57:46.560 Yeah, for sure. 57:46.560 --> 57:51.200 A hundred percent in, in even the dumbest, there's so many low hanging fruit too. 57:51.200 --> 57:59.680 Uh, for example, just knowing what episode to listen to first to try out a podcast. 57:59.680 --> 58:00.400 Exactly. 58:00.400 --> 58:03.360 Uh, because most podcasts don't have an order to them. 58:03.920 --> 58:10.880 Uh, they, they can be listened to out of order and sorry to say some are better 58:10.880 --> 58:12.560 than others episodes. 58:12.560 --> 58:14.960 So some episodes of Joe Rogan are better than others. 58:15.520 --> 58:20.400 And it's nice to know, uh, which you should listen to, to try it out. 58:20.400 --> 58:26.320 And there's, uh, as far as I know, almost no information, uh, in terms of like, uh, 58:26.320 --> 58:28.640 upvotes on how good an episode is. 58:28.640 --> 58:29.280 Exactly. 58:29.280 --> 58:33.520 So I think part of the problem is, uh, you, it's kind of like music. 58:33.520 --> 58:34.480 There isn't one answer. 58:34.480 --> 58:37.440 People use music for different things and there's actually many different types of music. 58:37.440 --> 58:40.560 There's workout music and there's classical piano music and focus music and, 58:41.200 --> 58:42.640 and, and, uh, so forth. 58:42.640 --> 58:44.080 I think the same with podcasts. 58:44.080 --> 58:45.360 Some podcasts are sequential. 58:45.360 --> 58:48.400 They're supposed to be listened to in, in order. 58:48.400 --> 58:51.040 It's actually, it's actually telling a narrative. 58:51.040 --> 58:55.840 Some podcasts are one topic, uh, kind of like yours, but different guests. 58:55.840 --> 58:57.280 So you could jump in anywhere. 58:57.280 --> 58:59.440 Some podcasts actually have completely different topics. 58:59.440 --> 59:04.560 And for those podcasts, it might be that I want, you know, we should recommend one episode 59:04.560 --> 59:09.280 because it's about AI from someone, but then they talk about something that you're not 59:09.280 --> 59:10.880 interested in the rest of the episodes. 59:10.880 --> 59:15.040 So I think our, what we're spending a lot of time on now is just first understanding 59:15.040 --> 59:21.520 the domain and creating kind of the knowledge graph of how do these objects relate and how 59:21.520 --> 59:22.240 do people consume. 59:22.240 --> 59:24.800 And I think we'll find that it's going to be, it's going to be different. 59:26.000 --> 59:31.280 I'm excited because you're the, uh, Spotify is the first people I'm aware of that are 59:32.240 --> 59:34.800 trying to do this for podcasting. 59:34.800 --> 59:38.240 Podcasting has been like a wild west up until now. 59:38.240 --> 59:43.120 It's been a very, we want to be very careful though, because it's been a very good wild 59:43.120 --> 59:45.680 west, I think it's this fragile ecosystem. 59:46.320 --> 59:52.080 And I, we want to make sure that you don't barge in and say like, Oh, we're going to 59:52.080 --> 59:53.440 internetize this thing. 59:53.440 --> 59:56.640 And you have to think about the creators. 59:56.640 --> 1:00:01.040 You have to understand how they get distribution today, who listens to how they make money 1:00:01.040 --> 1:00:05.520 today, try to, you know, make sure that their business model works, that they understand. 1:00:06.080 --> 1:00:10.880 I think it's back to doing something to improving their products, like feedback loops and 1:00:10.880 --> 1:00:11.440 distribution. 1:00:11.440 --> 1:00:17.280 So jumping back into terms of this fascinating world of a recommender system and listening 1:00:17.280 --> 1:00:24.320 to music and using machine learning to analyze things, do you think it's better to what 1:00:24.320 --> 1:00:30.160 currently, correct me if I'm wrong, but currently Spotify lets people pick what they listen 1:00:30.160 --> 1:00:31.680 to the most part. 1:00:31.680 --> 1:00:35.040 There's a discovery process, but you kind of organize playlists. 1:00:35.040 --> 1:00:39.840 Is it better to let people pick what they listen to or recommend what they should listen 1:00:39.840 --> 1:00:44.960 to something like stations by Spotify that I saw that you're playing around with? 1:00:44.960 --> 1:00:47.520 Maybe you can tell me what's the status of that. 1:00:47.520 --> 1:00:52.880 This is a Pandora style app that just kind of, as opposed to you select the music you 1:00:52.880 --> 1:00:57.760 listen to, it kind of feeds you the music you listen to. 1:00:58.400 --> 1:01:00.800 What's the status of stations by Spotify? 1:01:00.800 --> 1:01:01.920 What's its future? 1:01:01.920 --> 1:01:07.040 The story of Spotify, as we have grown, has been that we made it more accessible to different 1:01:07.040 --> 1:01:14.000 audiences and stations is another one of those where the question is, some people want to 1:01:14.000 --> 1:01:14.720 be very specific. 1:01:14.720 --> 1:01:18.560 They actually want to hear Starway to Heaven right now, that needs to be very easy to do. 1:01:19.760 --> 1:01:26.080 And some people, or even the same person, at some point might say, I want to feel upbeat 1:01:26.080 --> 1:01:32.800 or I want to feel happy or I want songs to sing in the car. 1:01:32.800 --> 1:01:38.720 So they put in the information at a very different level and then we need to translate that into 1:01:38.720 --> 1:01:40.560 what that means musically. 1:01:40.560 --> 1:01:45.440 So stations is a test to create like a consumption input vector that is much simpler where you 1:01:45.440 --> 1:01:49.520 can just tune it a little bit and see if that increases the overall reach. 1:01:49.520 --> 1:01:56.000 But we're trying to kind of serve the entire gamut of super advanced so called music aficionados 1:01:56.000 --> 1:02:02.560 all the way to people who they love listening to music but it's not their number one priority 1:02:02.560 --> 1:02:03.200 in life. 1:02:03.200 --> 1:02:06.160 They're not going to sit and follow every new release from every new artist. 1:02:06.160 --> 1:02:11.120 They need to be able to influence music at a different level. 1:02:11.120 --> 1:02:17.360 So you can think of it as different products and I think one of the interesting things 1:02:17.360 --> 1:02:22.080 to answer your question on if it's better to let the user choose or to play, I think 1:02:22.080 --> 1:02:28.720 the answer is the challenge when machine learning kind of came along, there was a lot of thinking 1:02:28.720 --> 1:02:33.120 about what does product development mean in a machine learning context. 1:02:33.920 --> 1:02:38.880 People like Andrew Ng, for example, when he went to Baidu, he started doing a lot of practical 1:02:38.880 --> 1:02:43.280 machine learning, went from academia and he thought a lot about this and he had this notion 1:02:43.280 --> 1:02:47.760 that a product manager, designer and engineer, they used to work around this wireframe to 1:02:47.760 --> 1:02:49.440 kind of describe what the product should look like. 1:02:49.440 --> 1:02:54.080 It was something to talk about when you're doing a chatbot or a playlist, what are you 1:02:54.080 --> 1:02:54.640 going to say? 1:02:54.640 --> 1:02:55.520 It should be good. 1:02:55.520 --> 1:02:57.360 That's not a good product description. 1:02:57.360 --> 1:02:58.400 So how do you do that? 1:02:58.400 --> 1:03:03.120 And he came up with this notion that the test set is the new wireframe. 1:03:03.120 --> 1:03:06.960 The job of the product manager is to source a good test set that is representative of 1:03:06.960 --> 1:03:10.640 what, like if you say I want to play this, that is songs to sing in the car. 1:03:11.520 --> 1:03:15.360 The job of the product manager is to go and source a good test set of what that means. 1:03:15.360 --> 1:03:20.000 So then you can work with engineering to have algorithms to try to produce that. 1:03:20.000 --> 1:03:25.600 So we try to think a lot about how to structure product development for a machine learning 1:03:25.600 --> 1:03:26.320 age. 1:03:26.320 --> 1:03:30.000 And what we discovered was that a lot of it is actually in the expectation. 1:03:30.560 --> 1:03:33.120 And you can go two ways. 1:03:33.120 --> 1:03:40.880 So let's say that if you set the expectation with the user that this is a discovery product, 1:03:40.880 --> 1:03:45.280 like Discover Weekly, you're actually setting the expectation that most of what we show 1:03:45.280 --> 1:03:46.800 you will not be relevant. 1:03:46.800 --> 1:03:50.400 When you're in the discovery process, you're going to accept that actually if you find 1:03:50.400 --> 1:03:55.200 one gem every Monday that you totally love, you're probably going to be happy. 1:03:55.200 --> 1:04:00.240 Even though the statistical meaning, one out of 10 is terrible or one out of 20 is terrible 1:04:00.240 --> 1:04:02.640 from a user point of view because the setting was discovery is fine. 1:04:03.440 --> 1:04:04.640 Sorry to interrupt real quick. 1:04:05.360 --> 1:04:11.600 I just actually learned about Discover Weekly, which is a Spotify, I don't know, it's a 1:04:11.600 --> 1:04:15.360 feature of Spotify that shows you cool songs to listen to. 1:04:16.640 --> 1:04:18.160 Maybe I can do issue tracking. 1:04:18.160 --> 1:04:19.760 I couldn't find it on my Spotify app. 1:04:20.640 --> 1:04:21.680 It's in your library. 1:04:21.680 --> 1:04:22.640 It's in the library. 1:04:22.640 --> 1:04:23.760 It's in the list of library. 1:04:23.760 --> 1:04:25.040 Because I was like, whoa, this is cool. 1:04:25.040 --> 1:04:26.320 I didn't know this existed. 1:04:26.320 --> 1:04:27.440 And I tried to find it. 1:04:27.440 --> 1:04:28.800 But okay. 1:04:28.800 --> 1:04:31.040 I will show it to you and feedback to our product team. 1:04:31.920 --> 1:04:32.720 There you go. 1:04:32.720 --> 1:04:34.480 But yeah, so yeah, sorry. 1:04:34.480 --> 1:04:42.160 Just to mention the expectation there is basically that you're going to discover new songs. 1:04:42.160 --> 1:04:42.400 Yeah. 1:04:42.400 --> 1:04:47.200 So then you can be quite adventurous in the recommendations you do. 1:04:47.920 --> 1:04:53.120 But we have another product called Daily Mix, which kind of implies that these are only 1:04:53.120 --> 1:04:54.000 going to be your favorites. 1:04:54.560 --> 1:04:58.320 So if you have one out of 10 that is good and nine out of 10 that doesn't work for you, 1:04:58.320 --> 1:04:59.600 you're going to think it's a horrible product. 1:04:59.600 --> 1:05:03.040 So actually a lot of the product development we learned over the years is about setting 1:05:03.040 --> 1:05:04.080 the right expectations. 1:05:04.080 --> 1:05:09.680 So for Daily Mix, you know, algorithmically, we would pick among things that feel very 1:05:09.680 --> 1:05:11.280 safe in your taste space. 1:05:11.280 --> 1:05:15.520 Whereas Discover Weekly, we go kind of wild because the expectation is most of this is 1:05:15.520 --> 1:05:16.400 not going to. 1:05:16.400 --> 1:05:20.960 So a lot of that, a lot of to answer your question there, a lot of should you let the 1:05:20.960 --> 1:05:21.600 user pick or not? 1:05:21.600 --> 1:05:22.560 It depends. 1:05:23.360 --> 1:05:26.720 We have some products where the whole point is that the user can click play, put the phone 1:05:26.720 --> 1:05:30.000 in the pocket, and it should be really good music for like an hour. 1:05:30.000 --> 1:05:35.120 We have other products where you probably need to say like, no, no, save, no, no. 1:05:35.120 --> 1:05:36.160 And it's very interactive. 1:05:37.040 --> 1:05:37.440 I see. 1:05:37.440 --> 1:05:38.000 That makes sense. 1:05:38.000 --> 1:05:41.920 And then the radio product, the stations product is one of these like click play, put in your 1:05:41.920 --> 1:05:42.720 pocket for hours. 1:05:43.360 --> 1:05:44.160 That's really interesting. 1:05:44.160 --> 1:05:50.880 So you're thinking of different test sets for different users and trying to create products 1:05:50.880 --> 1:05:57.840 that sort of optimize for those test sets that represent a specific set of users. 1:05:57.840 --> 1:06:06.160 Yes, I think one thing that I think is interesting is we invested quite heavily in editorial 1:06:06.160 --> 1:06:09.520 in people creating playlists using statistical data. 1:06:09.520 --> 1:06:10.800 And that was successful for us. 1:06:10.800 --> 1:06:12.960 And then we also invested in machine learning. 1:06:13.600 --> 1:06:18.000 And for the longest time within Spotify and within the rest of the industry, there was 1:06:18.000 --> 1:06:23.360 always this narrative of humans versus the machine, algo versus editorial. 1:06:23.360 --> 1:06:27.600 And editors would say like, well, if I had that data, if I could see your 1:06:27.600 --> 1:06:31.680 playlisting history and I made a choice for you, I would have made a better choice. 1:06:31.680 --> 1:06:35.200 And they would have because they're much smarter than these algorithms. 1:06:35.200 --> 1:06:38.880 The human is incredibly smart compared to our algorithms. 1:06:38.880 --> 1:06:40.880 They can take culture into account and so forth. 1:06:41.440 --> 1:06:47.600 The problem is that they can't make 200 million decisions per hour for every user that logs 1:06:47.600 --> 1:06:47.680 in. 1:06:47.680 --> 1:06:51.760 So the algo may be not as sophisticated, but much more efficient. 1:06:51.760 --> 1:06:54.480 So there was this contradiction. 1:06:54.480 --> 1:07:00.160 But then a few years ago, we started focusing on this kind of human in the loop thinking 1:07:00.160 --> 1:07:01.280 around machine learning. 1:07:01.280 --> 1:07:06.480 And we actually coined an internal term for it called algotorial, a combination of algorithms 1:07:07.120 --> 1:07:15.040 and editors, where if we take a concrete example, you think of the editor, this paid 1:07:15.040 --> 1:07:20.400 expert that we have that's really good at something like soul, hip hop, EDM, something, 1:07:20.400 --> 1:07:20.720 right? 1:07:20.720 --> 1:07:22.800 They're a true expert, no one in the industry. 1:07:22.800 --> 1:07:24.480 So they have all the cultural knowledge. 1:07:24.480 --> 1:07:26.560 You think of them as the product manager. 1:07:26.560 --> 1:07:32.880 And you say that, let's say that you want to create a, you think that there's a product 1:07:32.880 --> 1:07:36.160 need in the world for something like songs to sing in the car or songs to sing in the 1:07:36.160 --> 1:07:36.560 shower. 1:07:36.560 --> 1:07:38.400 I'm taking that example because it exists. 1:07:38.400 --> 1:07:41.840 People love to scream songs in the car when they drive, right? 1:07:42.560 --> 1:07:45.520 So you want to create that product and you have this product manager who's a musical 1:07:45.520 --> 1:07:46.000 expert. 1:07:46.640 --> 1:07:50.800 They create, they come up with a concept, like I think this is a missing thing in humanity, 1:07:50.800 --> 1:07:52.800 like a playlist called songs to sing in the car. 1:07:53.920 --> 1:07:59.840 They create the framing, the image, the title, and they create a test set of, they create 1:07:59.840 --> 1:08:04.480 a group of songs, like a few thousand songs out of the catalog that they manually curate 1:08:04.480 --> 1:08:06.960 that are known songs that are great to sing in the car. 1:08:07.520 --> 1:08:09.840 And they can take like true romance into account. 1:08:09.840 --> 1:08:12.400 They understand things that our algorithms do not at all. 1:08:12.400 --> 1:08:14.480 So they have this huge set of tracks. 1:08:14.480 --> 1:08:19.600 Then when we deliver that to you, we look at your taste vectors and you get the 20 tracks 1:08:19.600 --> 1:08:21.760 that are songs to sing in the car in your taste. 1:08:22.560 --> 1:08:29.520 So you have personalization and editorial input in the same process, if that makes sense. 1:08:29.520 --> 1:08:30.880 Yeah, it makes total sense. 1:08:30.880 --> 1:08:32.480 And I have several questions around that. 1:08:32.480 --> 1:08:35.280 This is like fascinating. 1:08:36.080 --> 1:08:36.560 Okay. 1:08:36.560 --> 1:08:44.720 So first, it is a little bit surprising to me that the world expert humans are outperforming 1:08:44.720 --> 1:08:49.920 machines at specifying songs to sing in the car. 1:08:50.960 --> 1:08:53.680 So maybe you could talk to that a little bit. 1:08:53.680 --> 1:08:57.040 I don't know if you can put it into words, but what is it? 1:08:57.760 --> 1:08:59.520 How difficult is this problem? 1:09:01.680 --> 1:09:06.720 Do you really, I guess what I'm trying to ask is there, how difficult is it to encode 1:09:06.720 --> 1:09:14.640 the cultural references, the context of the song, the artists, all those things together? 1:09:14.640 --> 1:09:16.720 Can machine learning really not do that? 1:09:17.360 --> 1:09:23.040 I mean, I think machine learning is great at replicating patterns if you have the patterns. 1:09:23.040 --> 1:09:27.680 But if you try to write with me a spec of what song's greatest song to sing in the car 1:09:27.680 --> 1:09:30.320 definition is, is it loud? 1:09:30.320 --> 1:09:31.520 Does it have many choruses? 1:09:31.520 --> 1:09:32.800 Should it have been in movies? 1:09:32.800 --> 1:09:35.680 It quickly gets incredibly complicated, right? 1:09:35.680 --> 1:09:36.880 Yeah. 1:09:36.880 --> 1:09:40.960 And a lot of it may not be in the structure of the song or the title. 1:09:40.960 --> 1:09:44.880 It could be cultural references because, you know, it was a history. 1:09:44.880 --> 1:09:51.360 So the definition problems quickly get, and I think that was the insight of Andrew Ng 1:09:51.360 --> 1:09:55.440 when he said the job of the product manager is to understand these things that algorithms 1:09:55.440 --> 1:09:58.640 don't and then define what that looks like. 1:09:58.640 --> 1:10:00.880 And then you have something to train towards, right? 1:10:00.880 --> 1:10:02.720 Then you have kind of the test set. 1:10:02.720 --> 1:10:06.960 And then so today the editors create this pool of tracks and then we personalize. 1:10:06.960 --> 1:10:11.120 You could easily imagine that once you have this set, you could have some automatic exploration 1:10:11.120 --> 1:10:13.840 on the rest of the catalog because then you understand what it is. 1:10:14.480 --> 1:10:19.600 And then the other side of it, when machine learning does help is this taste vector. 1:10:20.560 --> 1:10:26.960 How hard is it to construct a vector that represents the things an individual human 1:10:26.960 --> 1:10:30.080 likes, this human preference? 1:10:30.080 --> 1:10:37.120 So you can, you know, music isn't like, it's not like Amazon, like things you usually buy. 1:10:38.320 --> 1:10:39.920 Music seems more amorphous. 1:10:39.920 --> 1:10:42.560 Like it's this thing that's hard to specify. 1:10:42.560 --> 1:10:48.080 Like what is, you know, if you look at my playlist, what is the music that I love? 1:10:48.080 --> 1:10:48.640 It's harder. 1:10:49.360 --> 1:10:54.080 It seems to be much more difficult to specify concretely. 1:10:54.080 --> 1:10:57.120 So how hard is it to build a taste vector? 1:10:57.120 --> 1:11:00.000 It is very hard in the sense that you need a lot of data. 1:11:00.720 --> 1:11:05.520 And I think what we found was that, so it's not a stationary problem. 1:11:06.240 --> 1:11:07.200 It changes over time. 1:11:08.720 --> 1:11:15.680 And so we've gone through the journey of, if you've done a lot of computer vision, 1:11:15.680 --> 1:11:18.320 obviously I've done a bunch of computer vision in my past. 1:11:18.320 --> 1:11:24.160 And we started kind of with the handcrafted heuristics for, you know, this is kind of 1:11:24.160 --> 1:11:24.800 indie music. 1:11:24.800 --> 1:11:25.360 This is this. 1:11:25.360 --> 1:11:27.440 And if you consume this, you'd probably like this. 1:11:27.440 --> 1:11:31.200 So we have, we started there and we have some of that still. 1:11:31.200 --> 1:11:34.720 Then what was interesting about the playlist data was that you could find these latent 1:11:34.720 --> 1:11:37.520 things that wouldn't necessarily even make sense to you. 1:11:38.800 --> 1:11:42.880 That could even capture maybe cultural references because they cooccurred. 1:11:42.880 --> 1:11:48.160 Things that wouldn't have appeared kind of mechanistically either in the content or so 1:11:48.160 --> 1:11:48.400 forth. 1:11:48.400 --> 1:12:01.280 So I think that, I think the core assumption is that there are patterns in almost 1:12:01.280 --> 1:12:01.840 everything. 1:12:02.640 --> 1:12:06.960 And if there are patterns, these embedding techniques are getting better and better now. 1:12:06.960 --> 1:12:12.400 Now, as everyone else, we're also using kind of deep embeddings where you can encode 1:12:12.400 --> 1:12:14.400 binary values and so forth. 1:12:14.400 --> 1:12:21.280 And what I think is interesting is this process to try to find things that do not 1:12:21.280 --> 1:12:23.920 necessarily, you wouldn't actually have guessed. 1:12:23.920 --> 1:12:28.560 So it is very hard in an engineering sense to find the right dimensions. 1:12:28.560 --> 1:12:33.920 It's an incredible scalability problem to do for hundreds of millions of users and to 1:12:33.920 --> 1:12:34.880 update it every day. 1:12:35.920 --> 1:12:42.160 But in theory, in theory embeddings isn't that complicated. 1:12:42.160 --> 1:12:46.240 The fact that you try to find some principal components or something like that, dimensionality 1:12:46.240 --> 1:12:47.040 reduction and so forth. 1:12:47.040 --> 1:12:48.240 So the theory, I guess, is easy. 1:12:48.240 --> 1:12:50.480 The practice is very, very hard. 1:12:50.480 --> 1:12:53.120 And it's a huge engineering challenge. 1:12:53.120 --> 1:12:58.400 But fortunately, we have some amazing both research and engineering teams in this space. 1:12:58.400 --> 1:13:03.200 Yeah, I guess the question is all, I mean, it's similar. 1:13:03.200 --> 1:13:05.360 I deal with it with autonomous vehicle spaces. 1:13:05.360 --> 1:13:07.680 The question is how hard is driving? 1:13:07.680 --> 1:13:12.560 And here is basically the question is of edge cases. 1:13:14.560 --> 1:13:22.240 So embedding probably works, not probably, but I would imagine works well in a lot of 1:13:22.240 --> 1:13:22.740 cases. 1:13:24.000 --> 1:13:25.840 So there's a bunch of questions that arise then. 1:13:25.840 --> 1:13:33.760 So do song preferences, does your taste vector depend on context, like mood, right? 1:13:33.760 --> 1:13:39.840 So there's different moods, and so how does that take in it? 1:13:41.840 --> 1:13:44.320 Is it possible to take that as a consideration? 1:13:44.320 --> 1:13:49.840 Or do you just leave that as a interface problem that allows the user to just control it? 1:13:49.840 --> 1:13:55.440 So when I'm looking for workout music, I kind of specify it by choosing certain playlists, 1:13:55.440 --> 1:13:56.560 doing certain search. 1:13:56.560 --> 1:13:58.560 Yeah, so that's a great point. 1:13:58.560 --> 1:14:00.080 Back to the product development. 1:14:00.080 --> 1:14:04.480 You could try to spend a few years trying to predict which mood you're in automatically 1:14:04.480 --> 1:14:08.320 when you open Spotify, or you create a tab which is happy and sad, right? 1:14:08.320 --> 1:14:10.880 And you're going to be right 100% of the time with one click. 1:14:10.880 --> 1:14:14.880 Now, it's probably much better to let the user tell you if they're happy or sad, or 1:14:14.880 --> 1:14:15.840 if they want to work out. 1:14:15.840 --> 1:14:20.480 On the other hand, if your user interface becomes 2,000 tabs, you're introducing so 1:14:20.480 --> 1:14:22.080 much friction so no one will use the product. 1:14:22.080 --> 1:14:23.520 So then you have to get better. 1:14:24.080 --> 1:14:26.800 So it's this thing where you have to be able to get better. 1:14:26.800 --> 1:14:32.640 So then you have to get better, so it's this thing where I think maybe it was, I don't 1:14:32.640 --> 1:14:35.040 remember who coined it, but it's called fault tolerant UIs, right? 1:14:35.040 --> 1:14:40.640 You build a UI that is tolerant of being wrong, and then you can be much less right in your 1:14:42.000 --> 1:14:43.120 algorithms. 1:14:43.120 --> 1:14:45.440 So we've had to learn a lot of that. 1:14:45.440 --> 1:14:52.160 Building the right UI that fits where the machine learning is, and a great discovery 1:14:52.160 --> 1:14:58.720 there, which was by the teams during one of our hack days, was this thing of taking discovery, 1:14:58.720 --> 1:15:04.880 packaging it into a playlist, and saying that these are new tracks that we think you might 1:15:04.880 --> 1:15:05.920 like based on this. 1:15:05.920 --> 1:15:09.440 And setting the right expectation made it a great product. 1:15:09.440 --> 1:15:15.920 So I think we have this benefit that, for example, Tesla doesn't have that we can change 1:15:15.920 --> 1:15:16.800 the expectation. 1:15:16.800 --> 1:15:18.640 We can build a fault tolerant setting. 1:15:18.640 --> 1:15:23.040 It's very hard to be fault tolerant when you're driving at 100 miles per hour or something. 1:15:23.760 --> 1:15:30.000 And we have the luxury of being able to say that of being wrong if we have the right UI, 1:15:30.000 --> 1:15:33.440 which gives us different abilities to take more risk. 1:15:33.440 --> 1:15:36.960 So I actually think the self driving problem is much harder. 1:15:37.680 --> 1:15:38.720 Oh, yeah, for sure. 1:15:39.680 --> 1:15:44.240 It's much less fun because people die. 1:15:44.240 --> 1:15:45.200 Exactly. 1:15:45.200 --> 1:15:55.040 And in Spotify, it's such a more fun problem because failure is beautiful in a way. 1:15:55.040 --> 1:15:56.320 It leads to exploration. 1:15:56.320 --> 1:15:58.640 So it's a really fun reinforcement learning problem. 1:15:58.640 --> 1:16:02.800 The worst case scenario is you get these WTF tweets like, how did I get this? 1:16:02.800 --> 1:16:03.600 This song, yeah. 1:16:03.600 --> 1:16:05.440 Which is a lot better than the self driving. 1:16:05.440 --> 1:16:14.400 Exactly, so what's the feedback that a user, what's the signal that a user provides into 1:16:14.400 --> 1:16:15.440 the system? 1:16:15.440 --> 1:16:17.920 So you mentioned skipping. 1:16:19.360 --> 1:16:20.880 What is like the strongest signal? 1:16:22.000 --> 1:16:23.520 You didn't mention clicking like. 1:16:24.800 --> 1:16:27.600 So we have a few signals that are important. 1:16:27.600 --> 1:16:30.240 Obviously playing, playing through. 1:16:30.240 --> 1:16:36.560 So one of the benefits of music, actually, even compared to podcasts or movies is the 1:16:36.560 --> 1:16:38.720 object itself is really only about three minutes. 1:16:39.280 --> 1:16:44.320 So you get a lot of chances to recommend and the feedback loop is every three minutes instead 1:16:44.320 --> 1:16:45.760 of every two hours or something. 1:16:45.760 --> 1:16:50.320 So you actually get kind of noisy, but quite fast feedback. 1:16:50.880 --> 1:16:55.200 And so you can see if people play through, which is the inverse of skip really. 1:16:55.200 --> 1:16:56.560 That's an important signal. 1:16:56.560 --> 1:17:00.320 On the other hand, much of the consumption happens when your phone is in your pocket. 1:17:00.320 --> 1:17:03.040 Maybe you're running or driving or you're playing on a speaker. 1:17:03.040 --> 1:17:05.600 And so you not skipping doesn't mean that you love that song. 1:17:05.600 --> 1:17:08.960 It may be that it wasn't bad enough that you would walk up and skip. 1:17:08.960 --> 1:17:10.560 So it's a noisy signal. 1:17:10.560 --> 1:17:14.000 Then we have the equivalent of the like, which is you saved it to your library. 1:17:14.000 --> 1:17:15.920 That's a pretty strong signal of affection. 1:17:16.720 --> 1:17:21.280 And then we have the more explicit signal of playlisting. 1:17:21.280 --> 1:17:23.920 Like you took the time to create a playlist, you put it in there. 1:17:23.920 --> 1:17:28.960 There's a very little small chance that if you took all that trouble, this is not a really 1:17:28.960 --> 1:17:30.480 important track to you. 1:17:30.480 --> 1:17:34.000 And then we understand also what are the tracks it relates to. 1:17:34.000 --> 1:17:39.120 So we have the playlisting, we have the like, and then we have the listening or skip. 1:17:39.120 --> 1:17:43.360 And you have to have very different approaches to all of them because of different levels 1:17:43.360 --> 1:17:44.400 of noise. 1:17:44.400 --> 1:17:49.760 One is very voluminous, but noisy, and the other is rare, but you can probably trust it. 1:17:49.760 --> 1:17:55.680 Yeah, it's interesting because I think between those signals captures all the information 1:17:55.680 --> 1:17:57.040 you'd want to capture. 1:17:57.040 --> 1:18:01.520 I mean, there's a feeling, a shallow feeling for me that there's sometimes that I'll hear 1:18:01.520 --> 1:18:05.920 a song that's like, yes, this is, you know, this was the right song for the moment. 1:18:05.920 --> 1:18:10.720 But there's really no way to express that fact except by listening through it all the 1:18:10.720 --> 1:18:14.240 way and maybe playing it again at that time or something. 1:18:14.240 --> 1:18:19.680 But there's no need for a button that says this was the best song I could have heard 1:18:19.680 --> 1:18:20.400 at this moment. 1:18:20.400 --> 1:18:24.080 Well, we're playing around with that, with kind of the thumbs up concept saying like, 1:18:24.080 --> 1:18:25.200 I really like this. 1:18:25.200 --> 1:18:27.520 Just kind of talking to the algorithm. 1:18:27.520 --> 1:18:30.640 It's unclear if that's the best way for humans to interact. 1:18:30.640 --> 1:18:31.200 Maybe it is. 1:18:31.200 --> 1:18:35.600 Maybe they should think of Spotify as a person, an agent sitting there trying to serve you 1:18:35.600 --> 1:18:38.080 and you can say like, bad Spotify, good Spotify. 1:18:38.720 --> 1:18:42.880 Right now, the analogy we've had is more, you shouldn't think of us. 1:18:42.880 --> 1:18:44.400 We should be invisible. 1:18:44.400 --> 1:18:48.320 And the feedback is if you save it, it's kind of you work for yourself. 1:18:48.320 --> 1:18:50.960 You do a playlist because you think it's great and we can learn from that. 1:18:50.960 --> 1:18:55.200 It's kind of back to Tesla, how they kind of have this shadow mode. 1:18:55.200 --> 1:18:56.720 They sit in what you drive. 1:18:56.720 --> 1:18:58.560 We kind of took the same analogy. 1:18:58.560 --> 1:19:02.800 We sit in what you playlist and then maybe we can offer you an autopilot where you can 1:19:02.800 --> 1:19:04.640 take over for a while or something like that. 1:19:04.640 --> 1:19:08.240 And then back off if you say like, that's not good enough. 1:19:08.240 --> 1:19:11.600 But I think it's interesting to figure out what your mental model is. 1:19:11.600 --> 1:19:18.880 If Spotify is an AI that you talk to, which I think might be a bit too abstract for many 1:19:18.880 --> 1:19:24.320 consumers, or if you still think of it as it's my music app, but it's just more helpful. 1:19:24.320 --> 1:19:30.160 And it depends on the device it's running on, which brings us to smart speakers. 1:19:31.040 --> 1:19:38.400 So I have a lot of the Spotify listening I do is on devices I can talk to, whether it's 1:19:38.400 --> 1:19:39.920 from Amazon, Google or Apple. 1:19:39.920 --> 1:19:42.320 What's the role of Spotify on those devices? 1:19:42.320 --> 1:19:46.720 How do you think of it differently than on the phone or on the desktop? 1:19:47.840 --> 1:19:52.080 There are a few things to say about the first of all, it's incredibly exciting. 1:19:52.080 --> 1:19:55.760 They're growing like crazy, especially here in the US. 1:19:58.320 --> 1:20:09.200 And it's solving a consumer need that I think is, you can think of it as just remote interactivity. 1:20:09.200 --> 1:20:11.840 You can control this thing from across the room. 1:20:11.840 --> 1:20:16.880 And it may feel like a small thing, but it turns out that friction matters to consumers 1:20:16.880 --> 1:20:22.000 being able to say play, pause and so forth from across the room is very powerful. 1:20:22.000 --> 1:20:25.200 So basically, you made the living room interactive now. 1:20:26.000 --> 1:20:33.600 And what we see in our data is that the number one use case for these speakers is music, 1:20:33.600 --> 1:20:34.960 music and podcast. 1:20:34.960 --> 1:20:39.920 So fortunately for us, it's been important to these companies to have those use case 1:20:39.920 --> 1:20:40.640 covered. 1:20:40.640 --> 1:20:42.080 So they want to Spotify on this. 1:20:42.080 --> 1:20:44.320 We have very good relationships with them. 1:20:45.840 --> 1:20:49.200 And we're seeing tremendous success with them. 1:20:51.200 --> 1:20:54.640 What I think is interesting about them is it's already working. 1:20:57.360 --> 1:21:02.720 We kind of had this epiphany many years ago, back when we started using Sonos. 1:21:02.720 --> 1:21:06.800 If you went through all the trouble of setting up your Sonos system, you had this magical 1:21:06.800 --> 1:21:10.400 experience where you had all the music ever made in your living room. 1:21:10.400 --> 1:21:16.320 And we made this assumption that the home, everyone used to have a CD player at home, 1:21:16.320 --> 1:21:19.040 but they never managed to get their files working in the home. 1:21:19.040 --> 1:21:22.240 Having this network attached storage was too cumbersome for most consumers. 1:21:22.960 --> 1:21:26.480 So we made the assumption that the home would skip from the CD all the way to streaming 1:21:26.480 --> 1:21:31.120 books, where you would buy the steering and would have all the music built in. 1:21:31.120 --> 1:21:32.640 That took longer than we thought. 1:21:32.640 --> 1:21:36.080 But with the voice speakers, that was the unlocking that made kind of the connected 1:21:36.080 --> 1:21:38.240 speaker happen in the home. 1:21:39.760 --> 1:21:41.520 So it really exploded. 1:21:41.520 --> 1:21:45.760 And we saw this engagement that we predicted would happen. 1:21:45.760 --> 1:21:48.560 What I think is interesting, though, is where it's going from now. 1:21:49.120 --> 1:21:51.120 Right now, you think of them as voice speakers. 1:21:51.920 --> 1:21:58.640 But I think if you look at Google I.O., for example, they just added a camera to it, where 1:21:58.640 --> 1:22:04.240 when the alarm goes off, instead of saying, hey, Google, stop, you can just wave your 1:22:04.240 --> 1:22:05.040 hand. 1:22:05.040 --> 1:22:11.920 So I think they're going to think more of it as an agent or as an assistant, truly an 1:22:11.920 --> 1:22:12.400 assistant. 1:22:12.400 --> 1:22:17.040 And an assistant that can see you is going to be much more effective than a blind assistant. 1:22:17.040 --> 1:22:18.480 So I think these things will morph. 1:22:18.480 --> 1:22:22.560 And we won't necessarily think of them as, quote unquote, voice speakers anymore. 1:22:22.560 --> 1:22:29.200 Just as interactive access to the Internet in the home. 1:22:29.200 --> 1:22:34.080 But I still think that the biggest use case for those will be audio. 1:22:34.080 --> 1:22:36.640 So for that reason, we're investing heavily in it. 1:22:36.640 --> 1:22:43.520 And we built our own NLU stack to be able to the challenge here is, how do you innovate 1:22:43.520 --> 1:22:44.240 in that world? 1:22:44.240 --> 1:22:48.320 It lowers friction for consumers, but it's also much more constrained. 1:22:48.320 --> 1:22:51.600 You have no pixels to play with in an audio only world. 1:22:51.600 --> 1:22:54.880 It's really the vocabulary that is the interface. 1:22:54.880 --> 1:22:58.560 So we started investing and playing around quite a lot with that, trying to understand 1:22:58.560 --> 1:23:03.360 what the future will be of you speaking and gesturing and waving at your music. 1:23:03.360 --> 1:23:08.480 And actually, you're actually nudging closer to the autonomous vehicle space because from 1:23:08.480 --> 1:23:14.080 everything I've seen, the level of frustration people experience upon failure of natural 1:23:14.080 --> 1:23:18.320 language understanding is much higher than failure in other contexts. 1:23:18.320 --> 1:23:20.400 People get frustrated really fast. 1:23:20.400 --> 1:23:25.600 So if you screw that experience up even just a little bit, they give up really quickly. 1:23:25.600 --> 1:23:26.320 Yeah. 1:23:26.320 --> 1:23:28.320 And I think you see that in the data. 1:23:28.320 --> 1:23:36.160 While it's tremendously successful, the most common interactions are play, pause and next. 1:23:36.160 --> 1:23:39.440 The things where if you compare it to taking up your phone, unlocking it, bringing up the 1:23:39.440 --> 1:23:44.160 app and skipping, clicking skip, it was much lower friction. 1:23:44.160 --> 1:23:49.280 But then for longer, more complicated things like, can you find me that song about the 1:23:49.280 --> 1:23:51.920 people still bring up the phone and search and then play it on their speaker? 1:23:51.920 --> 1:23:56.960 So we tried again to build a fault tolerant UI where for the more complicated things, 1:23:56.960 --> 1:24:02.480 you can still pick up your phone, have powerful full keyboard search and then try to optimize 1:24:02.480 --> 1:24:07.280 for where there is actually lower friction and try to it's kind of like the test autopilot 1:24:07.280 --> 1:24:07.840 thing. 1:24:07.840 --> 1:24:11.040 You have to be at the level where you're helpful. 1:24:11.040 --> 1:24:15.040 If you're too smart and just in the way, people are going to get frustrated. 1:24:15.040 --> 1:24:18.080 And first of all, I'm not obsessed with stairway to heaven. 1:24:18.080 --> 1:24:19.440 It's just a good song. 1:24:19.440 --> 1:24:22.880 But let me mention that as a use case because it's an interesting one. 1:24:22.880 --> 1:24:28.160 I've literally told one of I don't want to say the name of the speaker because when people 1:24:28.160 --> 1:24:30.320 are listening to it, it'll make their speaker go off. 1:24:30.320 --> 1:24:34.720 But I talked to the speaker and I say play stairway to heaven. 1:24:34.720 --> 1:24:40.320 And every time it like not every time, but a large percentage of the time plays the wrong 1:24:40.320 --> 1:24:41.440 stairway to heaven. 1:24:41.440 --> 1:24:48.240 It plays like some cover of the and that part of the experience. 1:24:48.240 --> 1:24:55.120 I actually wonder from a business perspective, does Spotify control that entire experience 1:24:55.120 --> 1:24:55.600 or no? 1:24:56.160 --> 1:25:01.680 It seems like the NLU, the natural language stuff is controlled by the speaker and then 1:25:01.680 --> 1:25:04.640 Spotify stays at a layer below that. 1:25:04.640 --> 1:25:07.040 It's a good and complicated question. 1:25:07.040 --> 1:25:11.200 Some of which is dependent on the on the partners. 1:25:11.200 --> 1:25:13.280 So it's hard to comment on the on the specifics. 1:25:13.280 --> 1:25:15.840 But the question is the right one. 1:25:15.840 --> 1:25:21.280 The challenge is if you can't use any of the personalization, I mean, we know which stairway 1:25:21.280 --> 1:25:21.840 to heaven. 1:25:21.840 --> 1:25:26.400 And the truth is maybe for for one person, it is exactly the cover that they want. 1:25:26.400 --> 1:25:31.440 And they would be very frustrated if a place I think we I think we default to the right 1:25:31.440 --> 1:25:31.760 version. 1:25:31.760 --> 1:25:35.280 But but you actually want to be able to do the cover for the person that just played 1:25:35.280 --> 1:25:36.320 the cover 50 times. 1:25:36.320 --> 1:25:38.400 Or Spotify is just going to seem stupid. 1:25:38.400 --> 1:25:40.160 So you want to be able to leverage the personalization. 1:25:40.160 --> 1:25:46.320 But you have this stack where you have the the ASR and this thing called the end best 1:25:46.320 --> 1:25:48.480 list of the best guesses here. 1:25:48.480 --> 1:25:50.480 And then the position comes in at the end. 1:25:50.480 --> 1:25:53.280 You actually want the person to be here when you're guessing about what they actually 1:25:53.280 --> 1:25:54.000 meant. 1:25:54.000 --> 1:26:00.160 So we're working with these partners and it's a complicated it's a complicated thing 1:26:00.160 --> 1:26:02.880 where you want to you want to be able. 1:26:02.880 --> 1:26:06.800 So first of all, you want to be very careful with your users data. 1:26:06.800 --> 1:26:09.200 You don't want to share your users data without the permission. 1:26:09.200 --> 1:26:11.680 But you want to share some data so that their experience gets better. 1:26:12.640 --> 1:26:15.760 So that these partners can understand enough, but not too much and so forth. 1:26:16.400 --> 1:26:21.760 So it's really the trick is that it's like a business driven relationship where you're 1:26:21.760 --> 1:26:26.960 doing product development across companies together, which is which is really complicated. 1:26:26.960 --> 1:26:32.960 But this is exactly why we built our own NLU so that we actually can make personalized 1:26:32.960 --> 1:26:36.320 guesses, because this is the biggest frustration from a user point of view. 1:26:36.320 --> 1:26:40.160 They don't understand about ASR and best list and and business deals. 1:26:40.160 --> 1:26:41.440 They're like, how hard can it be? 1:26:41.440 --> 1:26:45.120 I was told this thing 50 times this version and still the place the wrong thing. 1:26:45.120 --> 1:26:46.240 It can't it can't be hard. 1:26:47.040 --> 1:26:48.640 So we try to take the user approach. 1:26:48.640 --> 1:26:53.360 If the user the user is not going to understand the complications of business, we have to 1:26:53.360 --> 1:26:53.760 solve it. 1:26:53.760 --> 1:27:02.240 So let's talk about sort of a complicated subject that I myself I'm quite torn about 1:27:02.960 --> 1:27:07.600 the idea sort of of paying artists. 1:27:08.640 --> 1:27:08.880 Right. 1:27:09.840 --> 1:27:17.200 I saw as of August 31st, 2018, over 11 billion dollars were paid to rights holders. 1:27:17.200 --> 1:27:21.200 So and further distributed to artists from Spotify. 1:27:21.200 --> 1:27:23.840 So a lot of money is being paid to artists. 1:27:23.840 --> 1:27:30.800 First of all, the whole time as a consumer for me, when I look at Spotify, I'm not sure 1:27:30.800 --> 1:27:34.880 I'm remembering correctly, but I think you said exactly how I feel, which is this is 1:27:34.880 --> 1:27:36.240 too good to be true. 1:27:36.240 --> 1:27:42.480 Like when I start using Spotify, I assume you guys will go bankrupt in like a month. 1:27:43.040 --> 1:27:44.400 It's like this is too good. 1:27:44.400 --> 1:27:45.200 A lot of people did. 1:27:47.040 --> 1:27:48.960 I was like, this is amazing. 1:27:48.960 --> 1:27:53.200 So one question I have is sort of the bigger question. 1:27:53.200 --> 1:27:55.200 How do you make money in this complicated world? 1:27:55.840 --> 1:28:03.840 How do you deal with the relationship with record labels who are complicated? 1:28:04.800 --> 1:28:14.080 These big you're essentially have the task of herding cats, but like rich and powerful 1:28:14.080 --> 1:28:21.520 powerful cats, and also have the task of paying artists enough and paying those labels enough 1:28:21.520 --> 1:28:26.480 and still making money in the Internet space where people are not willing to pay hundreds 1:28:26.480 --> 1:28:27.360 of dollars a month. 1:28:27.920 --> 1:28:30.720 So how do you navigate the space? 1:28:30.720 --> 1:28:31.600 How do you navigate? 1:28:31.600 --> 1:28:32.560 That's a beautiful description. 1:28:32.560 --> 1:28:33.520 Herding rich cats. 1:28:34.720 --> 1:28:35.280 That before. 1:28:37.200 --> 1:28:42.880 It is very complicated, and I think certainly actually betting against Spotify has been 1:28:42.880 --> 1:28:45.040 statistically a very smart thing to do. 1:28:45.040 --> 1:28:52.880 Just looking at the at the line of roadkill in music streaming services, it's it's kind 1:28:52.880 --> 1:28:58.560 of I think if I understood the complexity when I joined Spotify, unfortunately, fortunately, 1:28:58.560 --> 1:29:03.440 I didn't know enough about the music industry to understand the complexities, because then 1:29:03.440 --> 1:29:05.600 I would have made a more rational guess that it wouldn't work. 1:29:06.240 --> 1:29:08.480 So, you know, ignorance is bliss. 1:29:08.480 --> 1:29:13.200 But I think there have been a few distinct challenges. 1:29:13.200 --> 1:29:17.600 I think, as I said, one of the things that made it work at all was that Sweden and the 1:29:17.600 --> 1:29:19.200 Nordics was a lost market. 1:29:19.840 --> 1:29:24.160 So there was no risk for labels to try this. 1:29:25.120 --> 1:29:29.040 I don't think it would have worked if if the market was healthy. 1:29:29.760 --> 1:29:32.160 So that was the initial condition. 1:29:33.120 --> 1:29:36.160 Then we had this tremendous challenge with the model itself. 1:29:36.160 --> 1:29:39.520 So now most people were pirating. 1:29:39.520 --> 1:29:45.120 But for the people who bought a download or a CD, the artists would get all the revenue 1:29:45.120 --> 1:29:48.000 for all the future plays then, right? 1:29:48.000 --> 1:29:51.840 So you got it all up front, whereas the streaming model was like almost nothing day one, almost 1:29:51.840 --> 1:29:52.800 nothing day two. 1:29:52.800 --> 1:29:58.720 And then at some point, this curve of incremental revenue would intersect with your day one 1:29:58.720 --> 1:29:59.220 payment. 1:29:59.840 --> 1:30:05.280 And that took a long time to play out before before the music labels, they understood 1:30:05.280 --> 1:30:05.780 that. 1:30:05.780 --> 1:30:09.600 But on the artist side, it took a lot of time to understand that actually, if I have a big 1:30:09.600 --> 1:30:14.000 hit that is going to be played for many years, this is a much better model because I get 1:30:14.000 --> 1:30:18.000 paid based on how much people use the product, not how much they thought they would use it 1:30:18.000 --> 1:30:19.040 day one or so forth. 1:30:20.080 --> 1:30:22.880 So it was a complicated model to get across. 1:30:22.880 --> 1:30:24.000 But time helped with that. 1:30:24.000 --> 1:30:30.640 And now the revenues to the music industry actually are bigger again than it's gone through 1:30:30.640 --> 1:30:32.000 this incredible dip and now they're back up. 1:30:32.000 --> 1:30:36.480 And so we're very proud of having been a part of that. 1:30:37.920 --> 1:30:39.520 So there have been distinct problems. 1:30:39.520 --> 1:30:45.920 I think when it comes to the labels, we have taken the painful approach. 1:30:46.720 --> 1:30:52.400 Some of our competition at the time, they kind of looked at other companies and said, 1:30:52.400 --> 1:30:56.160 if we just ignore the rights, we get really big, really fast. 1:30:56.160 --> 1:31:00.480 We're going to be too big for the labels to kind of, too big to fail. 1:31:00.480 --> 1:31:01.120 They're not going to kill us. 1:31:01.120 --> 1:31:02.160 We didn't take that approach. 1:31:02.160 --> 1:31:06.960 We went legal from day one and we negotiated and negotiated and negotiated. 1:31:06.960 --> 1:31:07.600 It was very slow. 1:31:07.600 --> 1:31:08.240 It was very frustrating. 1:31:08.240 --> 1:31:12.240 We were angry at seeing other companies taking shortcuts and seeming to get away with it. 1:31:12.800 --> 1:31:18.160 It was this game theory thing where over many rounds of playing the game, this would be 1:31:18.160 --> 1:31:19.200 the right strategy. 1:31:19.200 --> 1:31:25.680 And even though clearly there's a lot of frustrations at times during renegotiations, there is this 1:31:25.680 --> 1:31:30.800 there is this weird trust where we have been honest and fair. 1:31:31.760 --> 1:31:32.480 We've never screwed them. 1:31:32.480 --> 1:31:33.680 They've never screwed us. 1:31:33.680 --> 1:31:39.280 It's 10 years, but there's this trust and like they know that if music doesn't get 1:31:39.280 --> 1:31:43.360 really big, if lots of people do not want to listen to music and want to pay for it, 1:31:43.360 --> 1:31:44.960 Spotify has no business model. 1:31:44.960 --> 1:31:47.040 So we actually are incredibly aligned. 1:31:48.240 --> 1:31:51.840 Other companies, not to be tense, but other companies have other business models where 1:31:51.840 --> 1:31:56.400 even if they made no money from music, they'd still be profitable companies. 1:31:56.400 --> 1:31:57.200 But Spotify won't. 1:31:57.200 --> 1:32:02.400 So I think the industry sees that we are actually aligned business wise. 1:32:03.120 --> 1:32:09.040 So there is this trust that allows us to do product development, even if it's scary, 1:32:11.040 --> 1:32:12.560 taking risks. 1:32:12.560 --> 1:32:17.200 The free model itself was an incredible risk for the music industry to take that they should 1:32:17.200 --> 1:32:17.920 get credit for. 1:32:17.920 --> 1:32:20.400 Now, some of it was that they had nothing to lose in the game. 1:32:20.400 --> 1:32:22.240 Some of it was that they had nothing to lose in Sweden. 1:32:22.240 --> 1:32:24.560 But frankly, a lot of the labels also took risk. 1:32:25.840 --> 1:32:31.360 And so I think we built up that trust with I think herding of cats sounds a bit. 1:32:32.320 --> 1:32:33.120 What's the word? 1:32:33.120 --> 1:32:35.280 It sounds like dismissive of the cats. 1:32:35.280 --> 1:32:35.920 Dismissive. 1:32:35.920 --> 1:32:37.200 No, every cat matters. 1:32:37.200 --> 1:32:39.360 They're all beautiful and very important. 1:32:39.360 --> 1:32:39.920 Exactly. 1:32:39.920 --> 1:32:42.800 They've taken a lot of risks and certainly it's been frustrating. 1:32:44.960 --> 1:32:47.600 So it's really like playing it's game theory. 1:32:47.600 --> 1:32:53.920 If you play the game many times, then you can have the statistical outcome that you 1:32:53.920 --> 1:32:54.560 bet on. 1:32:54.560 --> 1:32:57.520 And it feels very painful when you're in the middle of that thing. 1:32:57.520 --> 1:33:00.480 I mean, there's risk, there's trust, there's relationships. 1:33:00.480 --> 1:33:07.200 From just having read the biography of Steve Jobs, similar kind of relationships were discussed 1:33:07.200 --> 1:33:08.400 in iTunes. 1:33:08.400 --> 1:33:12.640 The idea of selling a song for a dollar was very uncomfortable for labels. 1:33:12.640 --> 1:33:13.760 Exactly. 1:33:13.760 --> 1:33:16.400 And there was no, it was the same kind of thing. 1:33:16.400 --> 1:33:21.840 It was trust, it was game theory as a lot of relationships that had to be built. 1:33:21.840 --> 1:33:28.880 And it's really a terrifyingly difficult process that Apple could go through a little 1:33:28.880 --> 1:33:31.920 bit because they could afford for that process to fail. 1:33:32.720 --> 1:33:37.600 For Spotify, it seems terrifying because you can't. 1:33:37.600 --> 1:33:44.240 Initially, I think a lot of it comes down to honestly Daniel and his tenacity in negotiating, 1:33:44.240 --> 1:33:50.800 which seems like an impossible task because he was completely unknown and so forth. 1:33:50.800 --> 1:33:54.160 But maybe that was also the reason that it worked. 1:33:56.480 --> 1:34:03.120 But I think game theory is probably the best way to think about it. 1:34:03.120 --> 1:34:08.800 You could go straight for this Nash equilibrium that someone is going to defect or you play 1:34:08.800 --> 1:34:14.240 it many times, you try to actually go for the top left, the corporations sell. 1:34:14.240 --> 1:34:19.680 Is there any magical reason why Spotify seems to have won this? 1:34:20.400 --> 1:34:25.360 So a lot of people have tried to do what Spotify tried to do and Spotify has come out. 1:34:25.360 --> 1:34:29.280 Well, so the answer is that there's no magical reason because I don't believe in magic. 1:34:30.000 --> 1:34:32.240 But I think there are there are reasons. 1:34:32.240 --> 1:34:39.600 And I think some of them are that people have misunderstood a lot of what we actually do. 1:34:40.400 --> 1:34:43.520 The actual Spotify model is very complicated. 1:34:43.520 --> 1:34:49.200 They've looked at the premium model and said, it seems like you can charge $9.99 for music 1:34:49.200 --> 1:34:52.000 and people are going to pay, but that's not what happened. 1:34:52.000 --> 1:34:55.680 Actually, when we launched the original mobile product, everyone said they would never pay. 1:34:56.640 --> 1:35:01.200 What happened was they started on the free product and then their engagement grew so 1:35:01.200 --> 1:35:05.680 much that eventually they said, maybe it is worth $9.99, right? 1:35:05.680 --> 1:35:08.880 It's your propensity to pay gross with your engagement. 1:35:08.880 --> 1:35:11.600 So we have this super complicated business model. 1:35:11.600 --> 1:35:15.200 We operate two different business models, advertising and premium at the same time. 1:35:15.760 --> 1:35:17.680 And I think that is hard to replicate. 1:35:17.680 --> 1:35:22.320 I struggle to think of other companies that run large scale advertising and subscription 1:35:22.320 --> 1:35:23.440 products at the same time. 1:35:24.400 --> 1:35:28.480 So I think the business model is actually much more complicated than people think it is. 1:35:28.480 --> 1:35:32.800 And so some people went after just the premium part without the free part and ran into a 1:35:32.800 --> 1:35:35.120 wall where no one wanted to pay. 1:35:35.120 --> 1:35:40.400 Some people went after just music should be free, just ads, which doesn't give you enough 1:35:40.400 --> 1:35:42.080 revenue and doesn't work for the music industry. 1:35:42.880 --> 1:35:46.560 So I think that combination is kind of opaque from the outside. 1:35:46.560 --> 1:35:51.040 So maybe I shouldn't say it here and reveal the secret, but that turns out to be hard 1:35:51.040 --> 1:35:54.400 to replicate than you would think. 1:35:54.400 --> 1:35:57.040 So there's a lot of brilliant business strategies out there. 1:35:57.040 --> 1:35:58.720 Brilliant business strategy here. 1:36:00.240 --> 1:36:01.280 Brilliance or luck? 1:36:01.280 --> 1:36:03.520 Probably more luck, but it doesn't really matter. 1:36:03.520 --> 1:36:05.440 It looks brilliant in retrospect. 1:36:05.440 --> 1:36:06.480 Let's call it brilliant. 1:36:07.840 --> 1:36:09.760 Yeah, when the books are written, they'll be brilliant. 1:36:10.480 --> 1:36:14.480 You've mentioned that your philosophy is to embrace change. 1:36:16.720 --> 1:36:23.600 So how will the music streaming and music listening world change over the next 10 years, 1:36:23.600 --> 1:36:24.640 20 years? 1:36:24.640 --> 1:36:26.960 You look out into the far future. 1:36:26.960 --> 1:36:27.520 What do you think? 1:36:28.960 --> 1:36:35.200 I think that music and for that matter, audio podcasts, audiobooks, I think it's one of 1:36:35.200 --> 1:36:36.720 the few core human needs. 1:36:37.360 --> 1:36:41.680 I think it there is no good reason to me why it shouldn't be at the scale of something 1:36:41.680 --> 1:36:44.160 like messaging or social networking. 1:36:44.160 --> 1:36:48.160 I don't think it's a niche thing to listen to music or news or something. 1:36:48.160 --> 1:36:50.880 So I think scale is obviously one of the things that I really hope for. 1:36:50.880 --> 1:36:54.400 I think I hope that it's going to be billions of users. 1:36:54.400 --> 1:36:58.160 I hope eventually everyone in the world gets access to all the world's music ever made. 1:36:58.720 --> 1:37:01.120 So obviously, I think it's going to be a much bigger business. 1:37:01.120 --> 1:37:03.040 Otherwise, we wouldn't be betting this big. 1:37:05.040 --> 1:37:13.600 Now, if you look more at how it is consumed, what I'm hoping is back to this analogy of 1:37:13.600 --> 1:37:22.800 the software tool chain, where I think I sometimes internally I make this analogy to text messaging. 1:37:22.800 --> 1:37:28.480 Text messaging was also based on standards in the area of mobile carriers. 1:37:28.480 --> 1:37:32.720 You had the SMS, the 140 character, 120 character SMS. 1:37:33.600 --> 1:37:36.080 And it was great because everyone agreed on the standards. 1:37:36.080 --> 1:37:40.480 So as a consumer, you got a lot of distributions and interoperability, but it was a very constrained 1:37:40.480 --> 1:37:40.980 format. 1:37:41.680 --> 1:37:45.840 And when the industry wanted to add pictures to that format to do the MMS, I looked it 1:37:45.840 --> 1:37:48.720 up and I think it took from the late 80s to early 2000s. 1:37:48.720 --> 1:37:53.040 This is like a 15, 20 year product cycle to bring pictures into that. 1:37:53.920 --> 1:38:00.240 Now, once that entire value chain of creation and consumption got wrapped in one software 1:38:00.240 --> 1:38:07.280 stack within something like Snapchat or WhatsApp, the first week they added disappearing messages. 1:38:07.280 --> 1:38:09.600 Then two weeks later, they added stories. 1:38:09.600 --> 1:38:14.560 The pace of innovation when you're on one software stack and you can affect both creation 1:38:14.560 --> 1:38:17.120 and consumption, I think it's going to be rapid. 1:38:17.120 --> 1:38:22.320 So with these streaming services, we now, for the first time in history, have enough, 1:38:22.320 --> 1:38:25.040 I hope, people on one of these services. 1:38:25.040 --> 1:38:29.600 Actually, whether it's Spotify or Amazon or Apple or YouTube, and hopefully enough 1:38:29.600 --> 1:38:32.320 creators that you can actually start working with the format again. 1:38:32.320 --> 1:38:33.760 And that excites me. 1:38:33.760 --> 1:38:39.200 I think being able to change these constraints from 100 years, that could really do something 1:38:39.200 --> 1:38:40.160 interesting. 1:38:40.160 --> 1:38:45.680 I really hope it's not just going to be the iteration on the same thing for the next 10 1:38:45.680 --> 1:38:47.360 to 20 years as well. 1:38:47.360 --> 1:38:52.000 Yeah, changing the creation of music, the creation of audio, the creation of podcasts 1:38:52.000 --> 1:38:54.400 is a really fascinating possibility. 1:38:54.400 --> 1:38:59.040 I myself don't understand what it is about podcasts that's so intimate. 1:38:59.680 --> 1:39:00.480 It just is. 1:39:00.480 --> 1:39:01.840 I listen to a lot of podcasts. 1:39:01.840 --> 1:39:09.680 I think it touches on a deep human need for connection that people do feel like they're 1:39:09.680 --> 1:39:12.960 connected to when they listen. 1:39:12.960 --> 1:39:17.600 I don't understand what the psychology of that is, but in this world that's becoming 1:39:17.600 --> 1:39:24.160 more and more disconnected, it feels like this is fulfilling a certain kind of need. 1:39:24.800 --> 1:39:30.080 And empowering the creator as opposed to just the listener is really interesting. 1:39:32.480 --> 1:39:34.240 I'm really excited that you're working on this. 1:39:34.240 --> 1:39:38.800 Yeah, I think one of the things that is inspiring for our teams to work on podcasts is exactly 1:39:38.800 --> 1:39:44.720 that, whether you think, like I probably do, that it's something biological about perceiving 1:39:44.720 --> 1:39:47.840 to be in the middle of the conversation that makes you listen in a different way. 1:39:47.840 --> 1:39:48.640 It doesn't really matter. 1:39:48.640 --> 1:39:50.240 People seem to perceive it differently. 1:39:50.240 --> 1:39:55.600 And there was this narrative for a long time that if you look at video, everything kind 1:39:55.600 --> 1:39:59.840 of in the foreground, it got shorter and shorter and shorter because of financial pressures 1:39:59.840 --> 1:40:01.600 and monetization and so forth. 1:40:01.600 --> 1:40:06.240 And eventually, at the end, there's almost like 20 seconds clip, people just screaming 1:40:06.240 --> 1:40:14.640 something and I feel really good about the fact that you could have interpreted that 1:40:14.640 --> 1:40:16.880 as people have no attention span anymore. 1:40:16.880 --> 1:40:18.400 They don't want to listen to things. 1:40:18.400 --> 1:40:20.000 They're not interested in deeper stories. 1:40:22.000 --> 1:40:23.280 People are getting dumber. 1:40:23.280 --> 1:40:26.720 But then podcasts came along and it's almost like, no, no, the need still existed. 1:40:28.000 --> 1:40:32.240 But maybe it was the fact that you're not prepared to look at your phone like this for 1:40:32.240 --> 1:40:32.740 two hours. 1:40:32.740 --> 1:40:36.500 But if you can drive at the same time, it seems like people really want to dig deeper 1:40:36.500 --> 1:40:38.820 and they want to hear like the more complicated version. 1:40:38.820 --> 1:40:42.980 So to me, that is very inspiring that that podcast is actually long form. 1:40:42.980 --> 1:40:48.340 It gives me a lot of hope for humanity that people seem really interested in hearing deeper, 1:40:48.340 --> 1:40:49.940 more complicated conversations. 1:40:49.940 --> 1:40:52.100 This is I don't understand it. 1:40:52.100 --> 1:40:53.140 It's fascinating. 1:40:53.140 --> 1:40:57.620 So the majority for this podcast, listen to the whole thing. 1:40:57.620 --> 1:41:02.500 This whole conversation we've been talking for an hour and 45 minutes. 1:41:02.500 --> 1:41:06.580 And somebody will I mean, most people will be listening to these words I'm speaking right 1:41:06.580 --> 1:41:06.580 now. 1:41:06.580 --> 1:41:07.080 It's crazy. 1:41:07.080 --> 1:41:10.740 You wouldn't have thought that 10 years ago with where the world seemed to go. 1:41:10.740 --> 1:41:12.100 That's very positive, I think. 1:41:12.100 --> 1:41:13.300 That's really exciting. 1:41:13.300 --> 1:41:17.060 And empowering the creator there is really exciting. 1:41:17.700 --> 1:41:18.740 Last question. 1:41:18.740 --> 1:41:22.660 You also have a passion for just mobile in general. 1:41:22.660 --> 1:41:32.660 How do you see the smartphone world, the digital space of smartphones and just everything that's 1:41:32.660 --> 1:41:39.780 on the move, whether it's Internet of Things and so on, changing over the next 10 years 1:41:39.780 --> 1:41:40.500 and so on? 1:41:41.460 --> 1:41:47.460 I think that one way to think about it is that computing might be moving out of these 1:41:47.460 --> 1:41:55.140 multipurpose devices, the computer we had and the phone, into specific purpose devices. 1:41:55.140 --> 1:42:01.060 And it will be ambient that at least in my home, you just shout something at someone 1:42:01.060 --> 1:42:03.380 and there's always one of these speakers close enough. 1:42:03.380 --> 1:42:06.980 And so you start behaving differently. 1:42:06.980 --> 1:42:11.460 It's as if you have the Internet ambient, ambiently around you and you can ask it things. 1:42:11.460 --> 1:42:15.780 So I think computing will kind of get more integrated and we won't necessarily think 1:42:15.780 --> 1:42:21.060 of it as connected to a device in the same way that we do today. 1:42:21.700 --> 1:42:22.900 I don't know the path to that. 1:42:22.900 --> 1:42:29.860 Maybe we used to have these desktop computers and then we partially replaced that with the 1:42:30.340 --> 1:42:32.740 laptops and left the desktop at home when I work. 1:42:32.740 --> 1:42:37.380 And then we got these phones and we started leaving the mobile phones. 1:42:37.380 --> 1:42:41.540 We had the desktop at home when I work and then we got these phones and we started leaving 1:42:41.540 --> 1:42:42.820 the laptop at home for a while. 1:42:42.820 --> 1:42:47.460 And maybe for stretches of time you're going to start using the watch and you can leave 1:42:47.460 --> 1:42:50.020 your phone at home for a run or something. 1:42:50.580 --> 1:42:58.420 And we're on this progressive path where I think what is happening with voice is that 1:43:00.740 --> 1:43:06.820 you have an interaction paradigm that doesn't require as large physical devices. 1:43:06.820 --> 1:43:12.820 So I definitely think there's a future where you can have your AirPods and your watch and 1:43:12.820 --> 1:43:14.980 you can do a lot of computing. 1:43:15.860 --> 1:43:20.020 And I don't think it's going to be this binary thing. 1:43:20.020 --> 1:43:23.380 I think it's going to be like many of us still have a laptop, we just use it less. 1:43:23.940 --> 1:43:25.940 And so you shift your consumption over. 1:43:26.820 --> 1:43:31.940 And I don't know about AR glasses and so forth. 1:43:31.940 --> 1:43:32.740 I'm excited about it. 1:43:32.740 --> 1:43:35.700 I spent a lot of time in that area, but I still think it's quite far away. 1:43:35.700 --> 1:43:37.540 AR, VR, all of that. 1:43:37.540 --> 1:43:39.780 Yeah, VR is happening and working. 1:43:39.780 --> 1:43:43.940 I think the recent Oculus Quest is quite impressive. 1:43:43.940 --> 1:43:45.300 I think AR is further away. 1:43:45.300 --> 1:43:46.580 At least that type of AR. 1:43:48.100 --> 1:43:54.660 But I do think your phone or watch or glasses understanding where you are and maybe what 1:43:54.660 --> 1:43:56.980 you're looking at and being able to give you audio cues about that. 1:43:56.980 --> 1:43:58.580 Or you can say like, what is this? 1:43:58.580 --> 1:43:59.700 And it tells you what it is. 1:44:00.980 --> 1:44:02.340 That I think might happen. 1:44:02.340 --> 1:44:08.020 You use your watch or your glasses as a mouse pointer on reality. 1:44:08.020 --> 1:44:09.460 I think it might be a while before... 1:44:09.460 --> 1:44:10.180 I might be wrong. 1:44:10.180 --> 1:44:10.820 I hope I'm wrong. 1:44:10.820 --> 1:44:14.820 I think it might be a while before we walk around with these big lab glasses that project 1:44:14.820 --> 1:44:15.620 things. 1:44:15.620 --> 1:44:16.260 I agree with you. 1:44:16.820 --> 1:44:22.260 It's actually really difficult when you have to understand the physical world enough to 1:44:23.060 --> 1:44:23.940 project onto it. 1:44:25.300 --> 1:44:26.740 I lied about the last question. 1:44:26.740 --> 1:44:32.660 Go ahead, because I just thought of audio and my favorite topic, which is the movie 1:44:32.660 --> 1:44:41.140 Her, do you think, whether it's part of Spotify or not, we'll have, I don't know if you've 1:44:41.140 --> 1:44:42.180 seen the movie Her. 1:44:42.180 --> 1:44:42.660 Absolutely. 1:44:45.060 --> 1:44:53.300 And there, audio is the primary form of interaction and the connection with another entity that 1:44:53.300 --> 1:44:59.300 you can actually have a relationship with, that you fall in love with based on voice 1:44:59.300 --> 1:45:00.740 alone, audio alone. 1:45:00.740 --> 1:45:04.820 Do you think that's possible, first of all, based on audio alone to fall in love with 1:45:04.820 --> 1:45:05.380 somebody? 1:45:05.380 --> 1:45:06.580 Somebody or... 1:45:06.580 --> 1:45:08.020 Well, yeah, let's go with somebody. 1:45:08.020 --> 1:45:11.700 Just have a relationship based on audio alone. 1:45:11.700 --> 1:45:18.500 And second question to that, can we create an artificial intelligence system that allows 1:45:18.500 --> 1:45:21.940 one to fall in love with it and her, him with you? 1:45:21.940 --> 1:45:29.940 So this is my personal answer, speaking for me as a person, the answer is quite unequivocally 1:45:29.940 --> 1:45:32.020 yes on both. 1:45:32.820 --> 1:45:36.580 I think what we just said about podcasts and the feeling of being in the middle of a 1:45:36.580 --> 1:45:42.660 conversation, if you could have an assistant where, and we just said that feels like a 1:45:42.660 --> 1:45:43.940 very personal setting. 1:45:43.940 --> 1:45:47.380 So if you walk around with these headphones and this thing, you're speaking with this 1:45:47.380 --> 1:45:49.940 thing all of the time that feels like it's in your brain. 1:45:49.940 --> 1:45:53.700 I think it's going to be much easier to fall in love with than something that would be 1:45:53.700 --> 1:45:54.740 on your screen. 1:45:54.740 --> 1:45:56.340 I think that's entirely possible. 1:45:56.340 --> 1:46:00.500 And then from the, you can probably answer this better than me, but from the concept 1:46:00.500 --> 1:46:07.060 of if it's going to be possible to build a machine that can achieve that, I think whether 1:46:07.060 --> 1:46:12.740 you think of it as, if you can fake it, the philosophical zombie that assimilates it enough 1:46:12.740 --> 1:46:17.700 or it somehow actually is, I think there's, it's only a question. 1:46:17.700 --> 1:46:20.500 It's only a question if you ask me about time, I'd have a different answer. 1:46:20.500 --> 1:46:24.580 But if you say I've given some half infinite time, absolutely. 1:46:24.580 --> 1:46:28.260 I think it's just atoms and arrangement of information. 1:46:29.620 --> 1:46:33.220 Well, I personally think that love is a lot simpler than people think. 1:46:33.780 --> 1:46:37.780 So we started with true romance and ended in love. 1:46:37.780 --> 1:46:39.780 I don't see a better place to end. 1:46:39.780 --> 1:46:40.340 Beautiful. 1:46:40.340 --> 1:46:41.860 Gustav, thanks so much for talking today. 1:46:41.860 --> 1:46:42.420 Thank you so much. 1:46:42.420 --> 1:46:43.140 It was a lot of fun. 1:46:43.140 --> 1:46:49.300 It was fun.