WEBVTT 00:00.000 --> 00:03.920 The following is a conversation with Gustav Sonastrom. 00:03.920 --> 00:07.280 He's the chief research and development officer at Spotify, 00:07.280 --> 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 --> 00:27.920 to enrich the music discovery and listening experience. 00:27.920 --> 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 a digital age. 00:36.000 --> 00:39.280 That's what Gustav and I talk about among many other topics, 00:39.280 --> 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 --> 00:53.200 If you enjoy it, subscribe on YouTube, give five stars on iTunes, 00:53.200 --> 00:58.000 support on Patreon, or simply connect with me on Twitter at Lex Freedman, 00:58.000 --> 01:00.400 spelled F R I D M A N. 01:01.280 --> 01:05.120 And now here's my conversation with Gustav Sonastrom. 01:06.400 --> 01:10.240 Spotify has over 50 million songs in its catalog. 01:10.240 --> 01:13.200 So let me ask the all important question. 01:14.160 --> 01:16.320 I feel like you're the right person to ask. 01:16.320 --> 01:19.520 What is the definitive greatest song of all time? 01:19.520 --> 01:22.880 It varies for me, personally. 01:22.880 --> 01:24.960 So you can't speak definitively for everyone? 01:26.240 --> 01:29.920 I wouldn't believe very much in machine learning if I did, 01:29.920 --> 01:32.640 right? Because everyone had the same taste. 01:32.640 --> 01:36.960 So for you, what is you have to pick? What is the song? 01:36.960 --> 01:39.280 All right. So it's it's pretty easy for me. 01:39.280 --> 01:42.160 There is this song called You're So Cool, 01:42.160 --> 01:45.280 Hans Zimmer, a soundtrack to True Romance. 01:45.280 --> 01:49.840 It was a movie that made a big impression on me, 01:49.840 --> 01:52.720 and it's kind of been following me through my life. 01:52.720 --> 01:55.440 Actually, had to play at my wedding. 01:56.000 --> 01:59.040 I sat with the organist and helped him play it on an organ, 01:59.040 --> 02:01.760 which was a pretty pretty interesting experience. 02:01.760 --> 02:06.720 That is probably my I would say top three movie of all time. 02:06.720 --> 02:08.320 Yeah, this is an incredible movie. 02:08.320 --> 02:11.040 Yeah. And it came out during my formative years. 02:11.040 --> 02:13.520 And as I've discovered in music, 02:13.520 --> 02:16.560 you shape your music taste during those years. 02:16.560 --> 02:18.560 So it definitely affected me quite a bit. 02:18.560 --> 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 --> 02:27.600 I didn't really adopt any characters from the movie, 02:27.600 --> 02:30.400 but it was it was a great story of love. 02:30.960 --> 02:32.160 It's fantastic actors. 02:33.040 --> 02:36.480 And and really, I didn't even know who Hans Zimmer was at the time, 02:36.480 --> 02:37.920 but fantastic music. 02:39.040 --> 02:42.160 And so that song has followed me. 02:42.160 --> 02:44.480 And the movie actually has followed me throughout my life. 02:44.480 --> 02:46.320 That was Quentin Tarantino, actually, 02:46.320 --> 02:48.480 I think directed or produced that. 02:49.200 --> 02:52.000 So it's not Stairway to Heaven or Bohemian Rhapsody. 02:52.000 --> 02:53.600 It's those are those are great. 02:53.600 --> 02:54.800 They're not my personal favorites, 02:54.800 --> 02:57.760 but I've realized that people have different tastes. 02:57.760 --> 03:00.400 And that's that's a big part of what we do. 03:00.400 --> 03:02.560 Well, for me, I don't have to stick with Stairway to Heaven. 03:04.000 --> 03:09.360 So 35,000 years ago, I looked this up on Wikipedia, 03:09.360 --> 03:11.760 flute like instruments started being used in caves 03:11.760 --> 03:15.600 as part of hunting rituals and primitive cultural gatherings, 03:15.600 --> 03:16.240 things like that. 03:16.240 --> 03:17.760 This is the birth of music. 03:18.560 --> 03:22.080 Since then, we had a few folks, Beethoven, Elvis, Beatles, 03:22.640 --> 03:25.520 Justin Bieber, of course, Drake. 03:26.320 --> 03:29.840 So in your view, let's start like high level philosophical. 03:29.840 --> 03:34.640 What is the purpose of music on this planet of ours? 03:35.760 --> 03:38.800 I think music has many different purposes. 03:38.800 --> 03:42.880 I think there's there's certainly a big purpose, 03:42.880 --> 03:45.200 which is the same as much of entertainment, 03:45.200 --> 03:50.880 which is escapism and to be able to live in some sort of 03:50.880 --> 03:52.080 other mental state for a while. 03:52.640 --> 03:54.800 But I also think you have the the opposite of escaping, 03:54.800 --> 03:57.200 which is to help you focus on something you are actually doing. 03:57.760 --> 04:02.560 As I think people use music as a tool to to tune the brain 04:03.120 --> 04:05.680 to the activities that they are actually doing. 04:05.680 --> 04:10.720 And it's kind of like in one sense, 04:10.720 --> 04:12.080 maybe it's the rawest signal. 04:12.080 --> 04:14.720 If you if you think about the brain as neural networks, 04:14.720 --> 04:16.720 it's maybe the most efficient hack we can do 04:16.720 --> 04:19.840 to actually actively tune it into some state that you want to be. 04:20.480 --> 04:21.360 You can do it in other ways. 04:21.360 --> 04:23.840 You can tell stories to put people in a certain mood, 04:23.840 --> 04:25.760 but music is probably very effective 04:25.760 --> 04:27.520 to get to a certain mood very fast. 04:28.640 --> 04:32.000 You know, there's a there's a social component historically 04:32.000 --> 04:34.080 to music where people listen to music together. 04:34.080 --> 04:36.160 I was just thinking about this, 04:36.160 --> 04:39.440 that to me, and you mentioned machine learning, 04:39.440 --> 04:44.400 but to me personally, music is a really private thing. 04:45.360 --> 04:47.280 Like I'm speaking for myself. 04:47.280 --> 04:50.160 I listen to music like almost nobody knows 04:50.160 --> 04:52.080 the kind of things I have in my library, 04:52.720 --> 04:54.480 except people who are really close to me 04:54.480 --> 04:56.880 and they really only know a certain percentage. 04:56.880 --> 04:58.400 There's like some weird stuff 04:58.400 --> 05:01.120 that I'm almost probably embarrassed by, right? 05:01.120 --> 05:02.480 It's called the guilty pleasures, right? 05:02.480 --> 05:04.560 Everyone has the guilty pleasures, yeah. 05:04.560 --> 05:05.760 Hopefully they're not too bad. 05:07.760 --> 05:08.640 For me, it's personal. 05:08.640 --> 05:12.000 Do you think of music as something that's social 05:12.640 --> 05:14.880 or as something that's personal? 05:14.880 --> 05:15.680 Or does it vary? 05:16.720 --> 05:22.720 So I think it's the same answer that you use it for both. 05:22.720 --> 05:24.080 We've thought a lot about this 05:24.880 --> 05:26.960 during these 10 years at Spotify, obviously. 05:27.600 --> 05:31.120 In one sense, as you said, music is incredibly social. 05:31.120 --> 05:32.720 You go to concerts and so forth. 05:33.440 --> 05:37.280 On the other hand, it is your escape 05:37.280 --> 05:41.360 and everyone has these things that are very personal to them. 05:42.080 --> 05:47.040 So what we've found is that when it comes to, 05:47.040 --> 05:50.640 to most people claim that they have a friend or two 05:50.640 --> 05:53.840 that they are heavily inspired by and that they listen to. 05:53.840 --> 05:55.680 So I actually think music is very social, 05:55.680 --> 05:57.440 but in a smaller group setting, 05:57.440 --> 06:03.360 it's an intimate form of, it's an intimate relationship. 06:03.360 --> 06:06.400 It's not something that you necessarily share broadly. 06:06.400 --> 06:08.000 Now at concerts, you can argue you do, 06:08.000 --> 06:10.080 but then you've gathered a lot of people 06:10.080 --> 06:11.840 that you have something in common with. 06:11.840 --> 06:14.880 I think this broadcast sharing of music 06:16.160 --> 06:19.920 is something we tried on social networks and so forth, 06:19.920 --> 06:24.400 but it turns out that people aren't super interested in 06:24.400 --> 06:26.960 what their friends listen to. 06:28.480 --> 06:30.400 They're interested in understanding 06:30.400 --> 06:32.880 if they have something in common perhaps with a friend, 06:32.880 --> 06:35.040 but not just as information. 06:35.680 --> 06:37.360 Right, that's really interesting. 06:37.360 --> 06:40.880 I was just thinking that this morning listening to Spotify, 06:41.600 --> 06:45.360 I really have a pretty intimate relationship with Spotify, 06:45.360 --> 06:47.120 with my playlists. 06:48.560 --> 06:51.760 I've had them for many years now 06:51.760 --> 06:53.360 and they've grown with me together. 06:53.360 --> 06:56.640 There's an intimate relationship you have 06:56.640 --> 06:59.520 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 --> 07:04.640 Can you do the impossible task 07:04.640 --> 07:08.240 and try to give a history of music listening 07:09.280 --> 07:12.720 from your perspective, from before the internet 07:12.720 --> 07:14.160 and after the internet? 07:14.160 --> 07:17.760 And just kind of everything leading up to streaming 07:17.760 --> 07:18.800 with Spotify and so on. 07:18.800 --> 07:20.800 I'll try, it could be a 100 year podcast. 07:20.800 --> 07:22.400 Yeah. 07:22.400 --> 07:24.480 I'll try to do a brief version. 07:24.480 --> 07:27.120 There are some things that I think are very interesting 07:27.120 --> 07:28.080 during the history of music, 07:28.080 --> 07:31.200 which is that before recorded music, 07:31.920 --> 07:33.040 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 --> 07:40.480 Creation and consumption had to happen at the same time, 07:40.480 --> 07:41.520 basically concerts. 07:42.400 --> 07:45.440 And so you either had to get to the nearest village 07:46.080 --> 07:47.280 to listen to music. 07:47.280 --> 07:48.960 And while that was cumbersome 07:48.960 --> 07:51.760 and it severely limited the distribution of music, 07:52.320 --> 07:54.000 it also had some different qualities, 07:54.000 --> 07:56.800 which was that the creator could always interact 07:56.800 --> 07:57.440 with the audience. 07:57.440 --> 07:58.480 It was always live. 07:59.280 --> 08:01.520 And also there was no time cap on the music. 08:01.520 --> 08:03.040 So I think it's not a coincidence 08:03.040 --> 08:05.760 that these early classical works, 08:05.760 --> 08:07.520 they're much longer than the three minutes. 08:08.320 --> 08:10.960 The three minutes came in as a restriction 08:10.960 --> 08:13.520 of the first wax disc that could only contain 08:13.520 --> 08:16.000 a three minute song on one side, right? 08:16.000 --> 08:20.640 So actually the recorded music severely limited the, 08:21.440 --> 08:22.240 or put constraints. 08:22.240 --> 08:22.960 I won't say limit. 08:22.960 --> 08:24.080 I mean, constraints are often good, 08:24.080 --> 08:26.880 but it put very hard constraints on the music format. 08:26.880 --> 08:29.760 So you kind of said, like instead of doing this opus 08:29.760 --> 08:32.320 of like many, you know, tens of minutes or something, 08:33.120 --> 08:34.240 now you get three and a half minutes 08:34.240 --> 08:36.240 because then you're out of wax on this disc. 08:36.880 --> 08:40.000 But in return, you get an amazing distribution. 08:40.000 --> 08:41.920 Your reach will widen, right? 08:41.920 --> 08:43.120 Just on that point real quick, 08:43.120 --> 08:47.360 without the mass scale distribution, 08:47.920 --> 08:52.880 there's a scarcity component where you kind of look forward to it. 08:54.000 --> 08:58.800 We had that, it's like the Netflix versus HBO Game of Thrones. 08:58.800 --> 09:02.720 You like wait for the event because you can't really listen to it. 09:02.720 --> 09:04.640 So you like look forward to it and then it's, 09:05.600 --> 09:07.360 you derive perhaps more pleasure 09:07.360 --> 09:10.560 because it's more rare for you to listen to particular piece. 09:10.560 --> 09:13.040 You think there's value to that scarcity? 09:13.040 --> 09:15.280 Yeah, I think that that is definitely a thing. 09:15.280 --> 09:17.680 And there's always this component of 09:17.680 --> 09:19.680 if you have something in infinite amount, 09:19.680 --> 09:21.600 so will you value it as much? 09:22.560 --> 09:23.120 Probably not. 09:23.680 --> 09:27.040 Humanity is always seeking some, is relative. 09:27.040 --> 09:28.480 So you're always seeking something you didn't have 09:28.480 --> 09:30.160 and when you have it, you don't appreciate it as much. 09:30.160 --> 09:32.240 So I think that's probably true. 09:32.240 --> 09:34.000 But I think that's why concerts exist. 09:34.000 --> 09:35.520 So you can actually have both. 09:36.480 --> 09:39.200 But I think net, if you couldn't listen to music 09:39.200 --> 09:42.320 in your car driving, that'd be worse. 09:42.320 --> 09:45.600 That cost will be bigger than the benefit of the anticipation, 09:45.600 --> 09:46.560 I think, that you would have. 09:47.600 --> 09:50.640 So, yeah, it started with live concerts. 09:50.640 --> 09:57.040 Then it's being able to, you know, the photograph invented, right? 09:57.040 --> 09:59.520 You start to be able to record music. 09:59.520 --> 10:00.160 Exactly. 10:00.160 --> 10:02.160 So then you got this massive distribution 10:02.160 --> 10:04.800 that made it possible to create two things. 10:04.800 --> 10:06.640 I think, first of all, cultural phenomenons. 10:06.640 --> 10:08.160 They probably need distribution. 10:08.160 --> 10:10.800 Distribution to be able to happen. 10:11.600 --> 10:16.480 But it also opened access to, you know, for a new kind of artist. 10:16.480 --> 10:18.160 So you started to have these phenomenons, 10:18.160 --> 10:19.680 like Beatles and Elvis and so forth, 10:19.680 --> 10:22.720 that would really a function of distribution, I think. 10:22.720 --> 10:24.640 Obviously, of talent and innovation, 10:24.640 --> 10:26.080 but there was also taking a component. 10:26.720 --> 10:28.800 And of course, the next big innovation to come along 10:28.800 --> 10:30.640 was radio, broadcast radio. 10:31.520 --> 10:34.480 And I think radio is interesting 10:34.480 --> 10:37.040 because it started not as a music medium. 10:37.040 --> 10:40.400 It started as an information medium for news. 10:40.400 --> 10:44.640 And then radio needed to find something to fill the time with 10:44.640 --> 10:47.920 so that they could honestly play more ads and make more money. 10:47.920 --> 10:49.440 And music was free. 10:49.440 --> 10:52.000 So then you had this massive distribution 10:52.000 --> 10:53.280 where you could program to people. 10:53.280 --> 10:54.960 I think those things, that ecosystem, 10:56.000 --> 11:00.160 is what created the ability for hits. 11:00.160 --> 11:02.640 But it was also a very broadcast medium. 11:02.640 --> 11:05.040 So you would tend to get these massive, massive hits, 11:05.040 --> 11:06.800 but maybe not such a long tail. 11:08.400 --> 11:11.520 In terms of choice of everybody listening to the same stuff. 11:11.520 --> 11:14.880 Yeah. And as you said, I think there are some social benefits to that. 11:15.760 --> 11:19.040 I think, for example, there is a high statistical chance 11:19.040 --> 11:21.440 that if I talk about the latest episode of Game of Thrones, 11:21.440 --> 11:23.600 we have something to talk about just statistically. 11:24.240 --> 11:27.200 In the age of individual choice, maybe some of that goes away. 11:27.200 --> 11:33.680 So I do see the value of shared cultural components. 11:33.680 --> 11:36.080 But I also obviously love personalization. 11:37.440 --> 11:40.080 And so let's catch this up to the internet. 11:40.080 --> 11:42.160 So maybe Napster. 11:42.160 --> 11:44.000 Well, first of all, there's like MP3s. 11:44.000 --> 11:45.680 Exactly. There's tape, CDs. 11:45.680 --> 11:48.320 There was a digitalization of music with a CD, really. 11:48.320 --> 11:51.280 It was physical distribution, but the music became digital. 11:52.080 --> 11:55.440 And so they were files, but basically boxed software, 11:55.440 --> 11:56.880 to use a software analogy. 11:57.760 --> 11:59.840 And then you could start downloading these files. 12:00.880 --> 12:03.440 And I think there are two interesting things that happened 12:03.440 --> 12:05.600 back to music used to be longer 12:05.600 --> 12:08.160 before it was constrained by the distribution medium. 12:08.960 --> 12:10.720 I don't think that was a coincidence. 12:10.720 --> 12:14.560 And then really the only music genre to have developed mostly after 12:15.360 --> 12:18.000 music was a file again on the internet is EDM. 12:18.000 --> 12:21.440 And EDM is often much longer than the traditional music. 12:21.440 --> 12:24.480 I think it's interesting to think about the fact that 12:24.480 --> 12:28.080 music is no longer constrained in minutes per song or something. 12:28.080 --> 12:32.080 It's a legacy of an old distribution technology. 12:32.080 --> 12:34.640 And you see some of this new music that breaks the format, 12:34.640 --> 12:36.880 not so much as I would have expected actually by now, 12:37.680 --> 12:39.200 but it still happens. 12:39.200 --> 12:42.080 So first of all, I don't really know what EDM is. 12:42.080 --> 12:44.560 Electronic dance music, you could say. 12:44.560 --> 12:47.280 Avicii was one of the biggest in this genre. 12:47.840 --> 12:50.640 So the main constraint is of time, 12:50.640 --> 12:53.280 something like three, four, five minutes song. 12:53.280 --> 12:55.440 So you could have songs that were eight minutes, 12:55.440 --> 12:56.640 10 minutes, and so forth. 12:57.200 --> 13:00.320 Because it started as a digital product 13:00.320 --> 13:03.920 that you downloaded, so you didn't have this constraint anymore. 13:04.800 --> 13:06.720 So I think it's something really interesting 13:06.720 --> 13:08.480 that I don't think has fully happened yet. 13:09.280 --> 13:11.440 We're kind of jumping ahead a little bit to where we are, 13:11.440 --> 13:14.960 but I think there's tons of formal innovation in music 13:15.760 --> 13:19.040 that should happen now, that couldn't happen 13:19.040 --> 13:22.000 when you needed to really adhere to the distribution constraints. 13:22.000 --> 13:25.280 If you didn't adhere to that, you would get no distribution. 13:25.280 --> 13:28.640 So Björk, for example, an Icelandic artist, 13:28.640 --> 13:30.880 she made a full iPad app as an album. 13:30.880 --> 13:34.960 That was very expensive, even though the App Store 13:34.960 --> 13:38.080 has great distribution, she gets nowhere near the distribution 13:38.080 --> 13:40.240 versus staying within the three minute format. 13:41.120 --> 13:43.840 So I think now that music is fully digital 13:43.840 --> 13:45.840 inside these streaming services, 13:45.840 --> 13:48.320 there is the opportunity to change the format again 13:48.320 --> 13:51.360 and allow creators to be much more creative 13:51.360 --> 13:54.400 without limiting their distribution ability. 13:54.400 --> 13:56.320 That's interesting that you're right. 13:56.320 --> 13:59.680 You're right, it's surprising that we don't see 13:59.680 --> 14:01.680 that taking advantage more often. 14:01.680 --> 14:05.040 It's almost like the constraints of the distribution 14:05.040 --> 14:08.720 from the 50s and 60s have molded the culture 14:08.720 --> 14:12.000 to where we want the three to five minute song, 14:12.800 --> 14:14.800 that anything else, not just... 14:14.800 --> 14:17.840 So we want the song as consumers and as artists, 14:18.640 --> 14:20.400 like, because I write a lot of music 14:20.400 --> 14:22.720 and I never even thought about writing something 14:22.720 --> 14:25.680 longer than 10 minutes. 14:26.960 --> 14:28.720 It's really interesting that those constraints... 14:28.720 --> 14:30.160 Because all your training data 14:30.160 --> 14:31.920 has been three and a half minutes long, right? 14:31.920 --> 14:32.400 It's right. 14:32.960 --> 14:38.400 Okay, so yeah, digitization of data led to then MP3s. 14:38.400 --> 14:40.400 Yeah, so I think you had this file then 14:41.440 --> 14:43.120 that was distributed physically, 14:43.680 --> 14:46.400 but then you had the components of digital distribution 14:46.400 --> 14:49.680 and then the internet happened and there was this vacuum 14:49.680 --> 14:52.560 where you had a format that could be digitally shipped 14:52.560 --> 14:53.760 but there was no business model. 14:54.320 --> 14:58.000 And then all these pirate networks happened. 14:59.040 --> 15:01.600 Napster and in Sweden, Pirate Bay, 15:01.600 --> 15:02.960 which was one of the biggest. 15:02.960 --> 15:07.120 And I think from a consumer point of view, 15:07.120 --> 15:10.560 which leads up to the inception of Spotify 15:10.560 --> 15:11.680 from a consumer point of view, 15:13.200 --> 15:16.240 consumers for the first time had this access model to music 15:17.040 --> 15:21.200 where they could, without any marginal cost, 15:21.200 --> 15:24.400 they could try different tracks. 15:25.040 --> 15:26.720 You could use music in new ways. 15:26.720 --> 15:27.840 There was no marginal cost. 15:28.240 --> 15:30.080 And that was a fantastic consumer experience. 15:30.080 --> 15:31.840 They have access to all the music ever made. 15:31.840 --> 15:33.280 I think it was fantastic. 15:33.840 --> 15:35.600 But it was also horrible for artists 15:35.600 --> 15:37.280 because there was no business model around it. 15:37.280 --> 15:38.640 So they didn't make any money. 15:38.640 --> 15:43.840 So the user need almost drove the user interface 15:43.840 --> 15:45.280 before there was a business model. 15:45.760 --> 15:47.440 And then there were these download stores 15:48.640 --> 15:50.400 that allowed you to download files. 15:50.400 --> 15:51.920 Which was a solution, 15:51.920 --> 15:53.920 but it didn't solve the access problem. 15:53.920 --> 15:55.760 There was still a marginal cost of 99 cents 15:55.760 --> 15:57.040 to try one more track. 15:57.040 --> 16:00.400 And I think that heavily limits how you listen to music. 16:00.400 --> 16:04.480 The example I always give is in Spotify, 16:04.480 --> 16:07.440 a huge amount of people listen to music while they sleep, 16:07.440 --> 16:09.120 while they go to sleep and while they sleep. 16:10.000 --> 16:12.160 If that costed you 99 cents per three minutes, 16:12.160 --> 16:13.760 you probably wouldn't do that. 16:13.760 --> 16:15.760 And you would be much less adventurous 16:15.760 --> 16:17.920 if there was a real dollar cost to explore music. 16:17.920 --> 16:20.880 So the access model is interesting in that 16:20.880 --> 16:22.320 it changes your music behavior. 16:22.880 --> 16:25.600 You can take much more risk 16:25.600 --> 16:27.120 because there's no marginal cost to it. 16:28.240 --> 16:30.480 Maybe let me linger on piracy for a second. 16:30.480 --> 16:32.880 Because I find, especially coming from Russia, 16:33.840 --> 16:36.080 piracy is something that's very interesting. 16:36.640 --> 16:41.520 To me, not me of course ever, 16:41.520 --> 16:47.040 but I have friends who've partook in piracy 16:47.040 --> 16:52.960 of music, software, TV shows, sporting events. 16:53.680 --> 16:55.760 And usually, to me, what that shows 16:55.760 --> 16:59.040 is not that they can actually pay the money 16:59.600 --> 17:01.040 and they're not trying to save money. 17:01.920 --> 17:04.000 They're choosing the best experience. 17:05.120 --> 17:07.520 So what to me, piracy shows 17:07.520 --> 17:09.920 is a business opportunity in all these domains. 17:09.920 --> 17:12.480 And that's where I think you're right. 17:12.480 --> 17:14.880 Spotify stepped in is basically, 17:14.880 --> 17:17.120 piracy is an experience. 17:17.120 --> 17:19.600 You can explore, find music you like. 17:20.240 --> 17:24.320 And actually, the interface of piracy is horrible. 17:24.960 --> 17:26.800 Because it's, I mean... 17:26.800 --> 17:30.960 Bad metadata, long download times, all kinds of stuff. 17:30.960 --> 17:33.040 And what Spotify does is basically 17:34.880 --> 17:36.160 first, rewards artists, 17:36.160 --> 17:38.560 and second, makes the experience 17:38.560 --> 17:40.080 of exploring music much better. 17:40.080 --> 17:43.920 I mean, the same is true, I think, for movies and so on. 17:43.920 --> 17:46.960 Piracy reveals, in the software space, for example, 17:46.960 --> 17:49.760 I'm a huge user and fan of Adobe products. 17:50.480 --> 17:53.600 And there was much more incentive 17:53.600 --> 17:55.360 to pirate Adobe products 17:55.360 --> 17:57.680 before they went to a monthly subscription plan. 17:58.320 --> 18:01.040 And now, all of the said friends 18:01.600 --> 18:05.040 that used to pirate Adobe products that I know, 18:05.040 --> 18:08.000 now actually pay, gladly, for the monthly subscription. 18:08.000 --> 18:09.280 Yeah, I think you're right. 18:09.280 --> 18:11.520 I think it's a sign of an opportunity 18:11.520 --> 18:12.960 for product development. 18:12.960 --> 18:17.760 And that sometimes there's a product market fit 18:17.760 --> 18:21.360 before there's a business model fit in product development. 18:21.360 --> 18:23.200 I think that's a sign of it. 18:23.200 --> 18:25.760 In Sweden, I think it was a bit of both. 18:25.760 --> 18:29.680 There was a culture where we even had 18:29.680 --> 18:32.000 a political party called the Pirate Party. 18:32.000 --> 18:35.120 And this was during the time when people said that 18:35.120 --> 18:36.640 information should be free. 18:36.640 --> 18:39.360 It was somehow wrong to charge for ones and zeros. 18:39.360 --> 18:42.480 So I think people felt that artists 18:42.480 --> 18:44.720 should probably make some money somehow else 18:44.720 --> 18:46.240 in concerts or something. 18:46.240 --> 18:49.280 So at least in Sweden, it was part really social acceptance 18:49.280 --> 18:51.120 even at the political level. 18:51.120 --> 18:55.120 But that also forced Spotify to compete with free, 18:57.120 --> 18:58.800 which I don't think could have happened 18:58.800 --> 19:00.080 anywhere else in the world. 19:00.080 --> 19:02.960 The music industry needed to be doing bad enough 19:02.960 --> 19:04.400 to take that risk. 19:04.400 --> 19:06.240 And Sweden was like the perfect testing ground. 19:06.240 --> 19:10.960 It had government funded high bandwidth, low latency broadband, 19:10.960 --> 19:12.880 which meant that the product would work. 19:12.880 --> 19:15.520 And it was also, there was no music revenue anyway. 19:15.520 --> 19:16.800 So they were kind of like, 19:16.800 --> 19:19.040 I don't think this is going to work, but why not? 19:20.320 --> 19:22.640 So this product is one that I don't think could have happened 19:22.640 --> 19:24.800 in America, the world's largest music market, for example. 19:24.800 --> 19:27.040 So how do you compete with free? 19:27.040 --> 19:29.840 Because that's an interesting world of the internet 19:29.840 --> 19:32.960 where most people don't like to pay for things. 19:32.960 --> 19:37.120 So Spotify steps in and tries to, yes, compete with free. 19:37.120 --> 19:38.320 How do you do it? 19:38.320 --> 19:39.760 So I think two things. 19:39.760 --> 19:43.360 One is people are starting to pay for things on the internet. 19:43.360 --> 19:46.560 I think one way to think about it was that advertising 19:46.560 --> 19:48.560 was the first business model 19:48.560 --> 19:50.560 because no one would put the credit card on the internet. 19:50.560 --> 19:52.720 Transactional with Amazon was the second. 19:52.720 --> 19:54.320 And maybe subscription is the third. 19:54.320 --> 19:57.520 And if you look offline, subscription is the biggest of those. 19:57.520 --> 19:59.120 So that may still happen. 19:59.120 --> 20:00.320 I think people are starting to pay. 20:00.320 --> 20:03.120 But definitely back then, we needed to compete with free. 20:03.120 --> 20:04.960 And the first thing you need to do is obviously 20:04.960 --> 20:06.960 to lower the price to free. 20:06.960 --> 20:10.160 And then you need to be better somehow. 20:10.160 --> 20:13.760 And the way that Spotify was better was on the user experience, 20:13.760 --> 20:19.760 on the actual performance, the latency of, you know, 20:19.760 --> 20:24.560 even if you had high band with broadband, 20:24.560 --> 20:27.360 it would still take you 30 seconds to a minute 20:27.360 --> 20:28.960 to download one of these tracks. 20:28.960 --> 20:31.360 So the Spotify experience of starting within the 20:31.360 --> 20:34.560 perceptual limit of immediacy, about 250 milliseconds, 20:34.560 --> 20:37.360 meant that the whole trick was, 20:37.360 --> 20:39.760 it felt as if you had downloaded all of Pirate Bay. 20:39.760 --> 20:40.960 It was on your hard drive. 20:40.960 --> 20:43.360 It was that fast, even though it wasn't. 20:43.360 --> 20:44.960 And it was still free. 20:44.960 --> 20:48.960 But somehow you were actually still being a legal citizen. 20:48.960 --> 20:52.960 That was the trick that Spotify managed to pull off. 20:52.960 --> 20:56.560 So I've actually heard you say this or write this. 20:56.560 --> 20:58.960 And I was surprised that I wasn't aware of it, 20:58.960 --> 21:00.560 because I just took it for granted. 21:00.560 --> 21:02.560 You know, whenever an awesome thing comes along, 21:02.560 --> 21:04.560 you just like, oh, of course it has to be this way. 21:04.560 --> 21:06.560 That's exactly right. 21:06.560 --> 21:08.560 That it felt like the entire world's libraries 21:08.560 --> 21:12.560 at my fingertips because of that latency being reduced. 21:12.560 --> 21:16.560 What was the technical challenge in reducing the late? 21:16.560 --> 21:20.560 So there was a group of really, really talented engineers. 21:20.560 --> 21:22.560 One of them called Ludwig Strigius. 21:22.560 --> 21:26.560 He wrote the, actually from Gothenburg, 21:26.560 --> 21:30.560 he wrote the initial, the UTorrent Clients, 21:30.560 --> 21:32.560 which is kind of an interesting backstory to Spotify, 21:32.560 --> 21:36.560 you know, that we have one of the top developers 21:36.560 --> 21:38.560 from BitTorrent Clients as well. 21:38.560 --> 21:42.560 So he wrote UTorrent, the world's smallest BitTorrent Clients. 21:42.560 --> 21:47.560 And then he was acquired very early by Daniel and Martin, 21:47.560 --> 21:50.560 who founded Spotify, and they actually sold the UTorrent Client 21:50.560 --> 21:52.560 to BitTorrent, but kept Ludwig. 21:52.560 --> 21:58.560 So Spotify had a lot of experience within peer to peer networking. 21:58.560 --> 22:02.560 So the original innovation was a distribution innovation, 22:02.560 --> 22:05.560 where Spotify built an end to end media distribution system. 22:05.560 --> 22:08.560 Up until only a few years ago, we actually hosted all the music ourselves. 22:08.560 --> 22:10.560 So we had both the server side and the client, 22:10.560 --> 22:14.560 and that meant that we could do things such as having a peer to peer solution 22:14.560 --> 22:17.560 to use local caching on the client side, 22:17.560 --> 22:20.560 because back then the world was mostly desktop. 22:20.560 --> 22:24.560 But we could also do things like hack the TCP protocols, 22:24.560 --> 22:28.560 things like Nagle's algorithm for kind of exponential back off, 22:28.560 --> 22:32.560 or ramp up and just go full throttle and optimize for latency 22:32.560 --> 22:34.560 at the cost of bandwidth. 22:34.560 --> 22:38.560 And all of this end to end control meant that we could do 22:38.560 --> 22:41.560 an experience that felt like a step change. 22:41.560 --> 22:45.560 These days, we actually are on on GCP. 22:45.560 --> 22:48.560 We don't host our own stuff, and everyone is really fast these days. 22:48.560 --> 22:50.560 So that was the initial competitive advantage, 22:50.560 --> 22:52.560 but then obviously you have to move on over time. 22:52.560 --> 22:55.560 And that was over 10 years ago, right? 22:55.560 --> 22:58.560 That was in 2008, the product was launched in Sweden. 22:58.560 --> 23:00.560 It was in a beta, I think, 2007. 23:00.560 --> 23:02.560 And it was on the desktop, right? 23:02.560 --> 23:03.560 It was desktop only. 23:03.560 --> 23:04.560 There's no phone. 23:04.560 --> 23:05.560 There was no phone. 23:05.560 --> 23:08.560 The iPhone came out in 2008, 23:08.560 --> 23:11.560 but the App Store came out one year later, I think. 23:11.560 --> 23:14.560 So the writing was on the wall, but there was no phone yet. 23:14.560 --> 23:19.560 You've mentioned that people would use Spotify 23:19.560 --> 23:20.560 to discover the songs they like, 23:20.560 --> 23:23.560 and then they would torrent those songs 23:23.560 --> 23:26.560 so they can copy it to their phone. 23:26.560 --> 23:27.560 Just hilarious. 23:27.560 --> 23:28.560 Exactly. 23:28.560 --> 23:29.560 Or not torrent. 23:29.560 --> 23:34.560 Seriously, piracy does seem to be a good guide 23:34.560 --> 23:36.560 for business models. 23:36.560 --> 23:37.560 Video content. 23:37.560 --> 23:40.560 As far as I know, Spotify doesn't have video content. 23:40.560 --> 23:42.560 Well, we do have music videos, 23:42.560 --> 23:44.560 and we do have videos on the service, 23:44.560 --> 23:46.560 but the way we think about ourselves is that 23:46.560 --> 23:48.560 we're an audio service, 23:48.560 --> 23:52.560 and we think that if you look at the amount of time 23:52.560 --> 23:53.560 that people spend on audio, 23:53.560 --> 23:56.560 it's actually very similar to the amount of time 23:56.560 --> 23:57.560 that people spend on video. 23:57.560 --> 24:01.560 So the opportunity should be equally big, 24:01.560 --> 24:02.560 but today it's not at all valued. 24:02.560 --> 24:04.560 Video is valued much higher. 24:04.560 --> 24:07.560 So we think it's basically completely undervalued. 24:07.560 --> 24:09.560 We think of ourselves as an audio service, 24:09.560 --> 24:12.560 but within that audio service, I think video 24:12.560 --> 24:13.560 can make a lot of sense. 24:13.560 --> 24:16.560 I think when you're discovering an artist, 24:16.560 --> 24:17.560 you probably do want to see them 24:17.560 --> 24:18.560 and understand who they are, 24:18.560 --> 24:20.560 to understand their identity. 24:20.560 --> 24:21.560 Do you want to see that video every time? 24:21.560 --> 24:22.560 No. 24:22.560 --> 24:24.560 90% of the time the phone is going to be in your pocket. 24:24.560 --> 24:26.560 For podcasters, you use video. 24:26.560 --> 24:28.560 I think that can make a ton of sense. 24:28.560 --> 24:29.560 So we do have video, 24:29.560 --> 24:31.560 but we're an audio service where, 24:31.560 --> 24:33.560 think of it as we call it internally, 24:33.560 --> 24:34.560 backgroundable video. 24:34.560 --> 24:35.560 Video that is helpful, 24:35.560 --> 24:38.560 but isn't the driver of the narrative. 24:38.560 --> 24:42.560 I think also if you look at YouTube, 24:42.560 --> 24:43.560 the way people, 24:43.560 --> 24:47.560 there's quite a few folks who listen to music on YouTube. 24:47.560 --> 24:48.560 So in some sense, 24:48.560 --> 24:52.560 YouTube is a bit of a competitor to Spotify, 24:52.560 --> 24:55.560 which is very strange to me that people use YouTube 24:55.560 --> 24:56.560 to listen to music. 24:56.560 --> 24:59.560 They play essentially the music videos, right? 24:59.560 --> 25:02.560 But don't watch the videos and put it in their pocket. 25:02.560 --> 25:07.560 Well, I think it's similar to what 25:07.560 --> 25:11.560 strangely maybe it's similar to what we were 25:11.560 --> 25:13.560 for the Piracy Networks, 25:13.560 --> 25:17.560 where YouTube for historical reasons 25:17.560 --> 25:20.560 have a lot of music videos. 25:20.560 --> 25:21.560 So you use, 25:21.560 --> 25:23.560 people use YouTube for a lot of the discovery part 25:23.560 --> 25:24.560 of the process, I think. 25:24.560 --> 25:26.560 But then it's not a really good 25:26.560 --> 25:28.560 sort of quote unquote MP3 player, 25:28.560 --> 25:29.560 because it doesn't even background. 25:29.560 --> 25:31.560 Then you have to keep the app in the foreground. 25:31.560 --> 25:32.560 So the consumption, 25:32.560 --> 25:34.560 it's not a good consumption tool, 25:34.560 --> 25:35.560 but it's a decently good discovery. 25:35.560 --> 25:37.560 I mean, I think YouTube is fantastic products. 25:37.560 --> 25:39.560 And I use it for all kinds of purposes. 25:39.560 --> 25:40.560 That's true. 25:40.560 --> 25:42.560 If I were to admit something, 25:42.560 --> 25:44.560 I do use YouTube a little bit for the discovery, 25:44.560 --> 25:46.560 to assist in the discovery process of songs. 25:46.560 --> 25:48.560 And then if I like it, 25:48.560 --> 25:50.560 I'll add it to Spotify. 25:50.560 --> 25:51.560 But that's okay. 25:51.560 --> 25:52.560 That's okay with us. 25:52.560 --> 25:53.560 Okay. 25:53.560 --> 25:55.560 So sorry, we're jumping around a little bit. 25:55.560 --> 25:58.560 So this kind of incredible, 25:58.560 --> 25:59.560 you look at Napster, 25:59.560 --> 26:01.560 you look at the early days of Spotify. 26:01.560 --> 26:02.560 How do you, 26:02.560 --> 26:04.560 one fascinating point is how do you grow 26:04.560 --> 26:05.560 a user base? 26:05.560 --> 26:07.560 So you're there in Sweden. 26:07.560 --> 26:09.560 You have an idea. 26:09.560 --> 26:12.560 I saw the initial sketches that looked terrible. 26:12.560 --> 26:14.560 How do you grow a user base 26:14.560 --> 26:18.560 from a few folks to millions? 26:18.560 --> 26:21.560 I think there are a bunch of tactical answers. 26:21.560 --> 26:22.560 So first of all, 26:22.560 --> 26:23.560 I think you need a great product. 26:23.560 --> 26:25.560 I don't think you take a bad product 26:25.560 --> 26:29.560 and market it to be successful. 26:29.560 --> 26:30.560 So you need a great product. 26:30.560 --> 26:31.560 Sorry to interrupt, 26:31.560 --> 26:33.560 but it's a totally new way to listen to music. 26:33.560 --> 26:34.560 So it's not just, 26:34.560 --> 26:38.560 did people realize immediately that Spotify is a great product? 26:38.560 --> 26:39.560 I think they did. 26:39.560 --> 26:41.560 So back to the point of piracy, 26:41.560 --> 26:44.560 it was a totally new way to listen to music legally, 26:44.560 --> 26:47.560 but people had been used to the access model in Sweden 26:47.560 --> 26:49.560 and the rest of the world for a long time through piracy. 26:49.560 --> 26:51.560 So one way to think about Spotify, 26:51.560 --> 26:53.560 it was just legal and fast piracy. 26:53.560 --> 26:55.560 And so people have been using it for a long time. 26:55.560 --> 26:57.560 So they weren't alien to it. 26:57.560 --> 27:00.560 They didn't really understand how it could be legal 27:00.560 --> 27:03.560 because it seemed too fast and too good to be true, 27:03.560 --> 27:05.560 which I think is a great product proposition 27:05.560 --> 27:07.560 if you can be too good to be true. 27:07.560 --> 27:09.560 But what I saw again and again 27:09.560 --> 27:10.560 was people showing each other, 27:10.560 --> 27:12.560 clicking the song, showing how fast it started 27:12.560 --> 27:14.560 and say, can you believe this? 27:14.560 --> 27:17.560 So I really think it was about speed. 27:17.560 --> 27:20.560 Then we also had an invite program 27:20.560 --> 27:22.560 that was really meant for scaling 27:22.560 --> 27:24.560 because we hosted our own servers. 27:24.560 --> 27:26.560 We needed to control scaling, 27:26.560 --> 27:28.560 but that built a lot of expectation 27:28.560 --> 27:30.560 and I don't want to say hype 27:30.560 --> 27:34.560 because hype implies that it wasn't true. 27:34.560 --> 27:37.560 Expectations, excitement around the product. 27:37.560 --> 27:40.560 And we replicated that when we launched in the US. 27:40.560 --> 27:42.560 We also built up an invite only program first. 27:42.560 --> 27:44.560 There are lots of tactics, 27:44.560 --> 27:46.560 but I think you need a great product 27:46.560 --> 27:48.560 to solve some problem. 27:48.560 --> 27:51.560 And basically the key innovation, 27:51.560 --> 27:54.560 there was technology, but on a meta level, 27:54.560 --> 27:56.560 the innovation was really the access model 27:56.560 --> 27:58.560 versus the ownership model. 27:58.560 --> 28:00.560 And that was tricky. 28:00.560 --> 28:04.560 A lot of people said that they wanted to own their music. 28:04.560 --> 28:07.560 They would never kind of rent it or borrow it. 28:07.560 --> 28:09.560 But I think the fact that we had a free tier 28:09.560 --> 28:12.560 which meant that you get to keep this music for life as well 28:12.560 --> 28:14.560 helped quite a lot. 28:14.560 --> 28:17.560 So this is an interesting psychological point 28:17.560 --> 28:19.560 that maybe you can speak to. 28:19.560 --> 28:21.560 It was a big shift for me. 28:21.560 --> 28:25.560 It's almost like to go to therapy for this. 28:25.560 --> 28:29.560 I think I would describe my early listening experience 28:29.560 --> 28:31.560 and I think a lot of my friends do, 28:31.560 --> 28:33.560 is basically hoarding music. 28:33.560 --> 28:36.560 As you're slowly one song by one song, 28:36.560 --> 28:37.560 or maybe albums, 28:37.560 --> 28:40.560 gathering a collection of music that you love. 28:40.560 --> 28:42.560 And you own it. 28:42.560 --> 28:44.560 Especially with CDs or tape, 28:44.560 --> 28:46.560 you physically had it. 28:46.560 --> 28:48.560 And what Spotify, 28:48.560 --> 28:50.560 what I had to come to grips with, 28:50.560 --> 28:52.560 and it was kind of liberating actually, 28:52.560 --> 28:55.560 is to throw away all the music. 28:55.560 --> 28:58.560 I've had this therapy session with lots of people. 28:58.560 --> 29:00.560 And I think the mental trick is, 29:00.560 --> 29:02.560 so actually we've seen the user data. 29:02.560 --> 29:03.560 When Spotify started, 29:03.560 --> 29:05.560 a lot of people did the exact same thing. 29:05.560 --> 29:06.560 They started hoarding, 29:06.560 --> 29:09.560 as if the music would disappear. 29:09.560 --> 29:11.560 Almost the equivalent of downloading. 29:11.560 --> 29:14.560 And so we had these playlists that had limits 29:14.560 --> 29:16.560 of a few hundred thousand tracks 29:16.560 --> 29:18.560 and we figured no one would ever. 29:18.560 --> 29:19.560 Well, they do. 29:19.560 --> 29:21.560 It's in hundreds and hundreds of thousands of tracks. 29:21.560 --> 29:23.560 And to this day, 29:23.560 --> 29:25.560 some people want to actually save, 29:25.560 --> 29:26.560 quote unquote, 29:26.560 --> 29:27.560 and play the entire catalog. 29:27.560 --> 29:29.560 But I think that the therapy session 29:29.560 --> 29:31.560 goes something like, 29:31.560 --> 29:34.560 instead of throwing away your music, 29:34.560 --> 29:36.560 if you took your files 29:36.560 --> 29:39.560 and you stored them in a locker at Google, 29:39.560 --> 29:40.560 it'd be a streaming service. 29:40.560 --> 29:41.560 It's just that in that locker, 29:41.560 --> 29:43.560 you have all the world's music now for free. 29:43.560 --> 29:44.560 So instead of giving away your music, 29:44.560 --> 29:46.560 you got all the music. 29:46.560 --> 29:47.560 It's yours. 29:47.560 --> 29:48.560 You could think of it as having a copy 29:48.560 --> 29:50.560 of the world's catalog there forever. 29:50.560 --> 29:53.560 So you actually got more music instead of less. 29:53.560 --> 29:56.560 It's just that you just took that hard disk 29:56.560 --> 29:59.560 and you sent it to someone who stored it for you. 29:59.560 --> 30:01.560 And once you go through that mental journey of, 30:01.560 --> 30:03.560 like, still my files, they're just over there. 30:03.560 --> 30:06.560 And I just have 40 million or 50 million or something now. 30:06.560 --> 30:08.560 Then people are like, okay, that's good. 30:08.560 --> 30:10.560 The problem is, I think, 30:10.560 --> 30:12.560 because you paid us a subscription, 30:12.560 --> 30:14.560 if we hadn't had the free tier, 30:14.560 --> 30:15.560 where you would feel like, 30:15.560 --> 30:16.560 even if I don't want to pay anymore, 30:16.560 --> 30:18.560 I still get to keep them. 30:18.560 --> 30:19.560 You keep your playlists forever. 30:19.560 --> 30:21.560 They don't disappear even though you stop paying. 30:21.560 --> 30:23.560 I think that was really important. 30:23.560 --> 30:25.560 If we would have started as, you know, 30:25.560 --> 30:27.560 you can put in all this time, 30:27.560 --> 30:29.560 but if you stop paying, you lose all your work. 30:29.560 --> 30:31.560 I think that would have been a big challenge 30:31.560 --> 30:33.560 and was the big challenge for a lot of our competitors. 30:33.560 --> 30:35.560 That's another reason why I think the free tier 30:35.560 --> 30:36.560 is really important, 30:36.560 --> 30:38.560 that people need to feel the security 30:38.560 --> 30:40.560 that the work they put in, it will never disappear, 30:40.560 --> 30:42.560 even if they decide not to pay. 30:42.560 --> 30:44.560 I like how you put the work you put in. 30:44.560 --> 30:46.560 I actually stopped even thinking of it that way. 30:46.560 --> 30:49.560 Spotify taught me to just enjoy music 30:49.560 --> 30:53.560 as opposed to what I was doing before, 30:53.560 --> 30:57.560 which is like, in an unhealthy way, hoarding music. 30:57.560 --> 31:00.560 Which I found that because I was doing that, 31:00.560 --> 31:03.560 I was listening to a small selection of songs 31:03.560 --> 31:06.560 way too much to where I was getting sick of them. 31:06.560 --> 31:09.560 Whereas Spotify, the more liberating kind of approach, 31:09.560 --> 31:11.560 I was just enjoying, of course, 31:11.560 --> 31:13.560 I listened to Stairway to Heaven over and over, 31:13.560 --> 31:16.560 but because of the extra variety, 31:16.560 --> 31:18.560 I don't get as sick of them. 31:18.560 --> 31:21.560 There's an interesting statistic I saw, 31:21.560 --> 31:24.560 so Spotify has, maybe you can correct me, 31:24.560 --> 31:27.560 but over 50 million songs, tracks, 31:27.560 --> 31:31.560 and over 3 billion playlists. 31:31.560 --> 31:35.560 So 50 million songs and 3 billion playlists. 31:35.560 --> 31:38.560 60 times more playlists than songs. 31:38.560 --> 31:40.560 What do you make of that? 31:40.560 --> 31:43.560 So the way I think about it is that 31:43.560 --> 31:48.560 from a statistical machine learning point of view, 31:48.560 --> 31:50.560 you have all these, 31:50.560 --> 31:52.560 if you want to think about reinforcement learning, 31:52.560 --> 31:54.560 you have this state space of all the tracks, 31:54.560 --> 31:58.560 and you can take different journeys through this world. 31:58.560 --> 32:02.560 And I think of these as like people 32:02.560 --> 32:05.560 helping themselves and each other, 32:05.560 --> 32:08.560 creating interesting vectors through this space of tracks. 32:08.560 --> 32:11.560 And then it's not so surprising that across 32:11.560 --> 32:14.560 many tens of millions of atomic units, 32:14.560 --> 32:17.560 there will be billions of paths that make sense. 32:17.560 --> 32:20.560 And we're probably pretty quite far away from 32:20.560 --> 32:22.560 having found all of them. 32:22.560 --> 32:25.560 So kind of our job now is, 32:25.560 --> 32:28.560 when Spotify started, it was really a search box 32:28.560 --> 32:30.560 that was for the time pretty powerful, 32:30.560 --> 32:33.560 and then I like to refer to this programming language 32:33.560 --> 32:35.560 called playlisting, where if you, 32:35.560 --> 32:37.560 as you probably were pretty good at music, 32:37.560 --> 32:39.560 you knew your new releases, you knew your back catalog, 32:39.560 --> 32:41.560 you knew you're starting with the heaven, 32:41.560 --> 32:43.560 you could create a soundtrack for yourself 32:43.560 --> 32:45.560 using this playlisting tool that's like metaprogramming language 32:45.560 --> 32:47.560 for music to soundtrack your life. 32:47.560 --> 32:49.560 And people who are good at music, 32:49.560 --> 32:51.560 it's back to how do you scale the product. 32:51.560 --> 32:53.560 For people who are good at music, 32:53.560 --> 32:55.560 that wasn't actually enough. 32:55.560 --> 32:57.560 If you had the catalog and a good search tool, 32:57.560 --> 32:59.560 and you can create your own sessions, 32:59.560 --> 33:02.560 you could create really good a soundtrack for your entire life. 33:02.560 --> 33:05.560 Probably perfectly personalized because you did it yourself. 33:05.560 --> 33:07.560 The problem was, many people aren't that good at music. 33:07.560 --> 33:09.560 They just can't spend the time. 33:09.560 --> 33:11.560 Even if you're very good at music, it's going to be hard to keep up. 33:11.560 --> 33:14.560 So what we did to try to scale this 33:14.560 --> 33:17.560 was to essentially try to build, 33:17.560 --> 33:19.560 you can think of them as agents, 33:19.560 --> 33:21.560 this friend that some people had 33:21.560 --> 33:23.560 that helped them navigate this music catalog. 33:23.560 --> 33:25.560 That's what we're trying to do for you. 33:25.560 --> 33:30.560 But also, there is something like 33:30.560 --> 33:34.560 200 million active users on Spotify. 33:34.560 --> 33:38.560 So from the machine learning perspective, 33:38.560 --> 33:42.560 you have these 200 million people plus. 33:42.560 --> 33:45.560 They're creating, 33:45.560 --> 33:48.560 it's really interesting to think of playlists 33:48.560 --> 33:52.560 as, I don't know if you meant it that way, 33:52.560 --> 33:54.560 but it's almost like a programming language. 33:54.560 --> 33:59.560 It's released a trace of exploration 33:59.560 --> 34:03.560 of those individual agents, the listeners. 34:03.560 --> 34:06.560 And you have all this new tracks coming in. 34:06.560 --> 34:12.560 So it's a fascinating space that is ripe for machine learning. 34:12.560 --> 34:18.560 So how can playlists be used as data 34:18.560 --> 34:21.560 in terms of machine learning 34:21.560 --> 34:24.560 to help Spotify organize the music? 34:24.560 --> 34:27.560 So we found in our data, 34:27.560 --> 34:30.560 not surprising that people who playlisted lots, 34:30.560 --> 34:32.560 they retained much better. 34:32.560 --> 34:34.560 We had a great experience. 34:34.560 --> 34:37.560 And so our first attempt was to playlist for users. 34:37.560 --> 34:40.560 And so we acquired this company called Tunigo of editors 34:40.560 --> 34:43.560 and professional playlisters 34:43.560 --> 34:47.560 and kind of leveraged the maximum of human intelligence 34:47.560 --> 34:50.560 to help build kind of these vectors 34:50.560 --> 34:53.560 through the track space for people. 34:53.560 --> 34:55.560 And that brought in the product. 34:55.560 --> 34:57.560 Then the obvious next, 34:57.560 --> 34:59.560 and we used statistical means 34:59.560 --> 35:02.560 where they could see when they created a playlist, 35:02.560 --> 35:04.560 how did that playlist perform? 35:04.560 --> 35:06.560 They could see skips of the songs, 35:06.560 --> 35:08.560 they could see how the songs perform, 35:08.560 --> 35:10.560 and they manually iterated the playlist 35:10.560 --> 35:13.560 to maximize performance for a large group of people. 35:13.560 --> 35:16.560 But there were never enough editors to playlists for you personally. 35:16.560 --> 35:18.560 So the promise of machine learning 35:18.560 --> 35:20.560 was to go from kind of group personalization, 35:20.560 --> 35:23.560 using editors and tools and statistics 35:23.560 --> 35:25.560 to individualization. 35:25.560 --> 35:27.560 And then what's so interesting 35:27.560 --> 35:29.560 about three billion playlists we have is, 35:29.560 --> 35:31.560 the truth is we lucked out. 35:31.560 --> 35:33.560 This was not a priori strategy, 35:33.560 --> 35:35.560 as is often the case. 35:35.560 --> 35:37.560 It looks really smart in hindsight, 35:37.560 --> 35:39.560 but it was dumb luck. 35:39.560 --> 35:41.560 We looked at these playlists 35:41.560 --> 35:44.560 and we had some people in the company, 35:44.560 --> 35:46.560 a person named Erik Bernhardtson, 35:46.560 --> 35:48.560 who was really good at machine learning 35:48.560 --> 35:51.560 already back then, in like 2007, 2008. 35:51.560 --> 35:54.560 Back then it was mostly collaborative filtering and so forth, 35:54.560 --> 35:57.560 we realized that what this is, 35:57.560 --> 36:00.560 is people are grouping tracks for themselves 36:00.560 --> 36:02.560 that have some semantic meaning to them. 36:02.560 --> 36:04.560 And then they actually label it 36:04.560 --> 36:06.560 with a playlist name as well. 36:06.560 --> 36:09.560 So in a sense, people were grouping tracks 36:09.560 --> 36:11.560 along semantic dimensions and labeling them. 36:11.560 --> 36:14.560 And so could you use that information 36:14.560 --> 36:18.560 to find that latent embedding? 36:18.560 --> 36:21.560 And so we started playing around 36:21.560 --> 36:23.560 with collaborative filtering 36:23.560 --> 36:26.560 and we saw tremendous success with it. 36:26.560 --> 36:30.560 Basically trying to extract some of these dimensions. 36:30.560 --> 36:33.560 And if you think about it, it's not surprising at all. 36:33.560 --> 36:35.560 It would be quite surprising 36:35.560 --> 36:37.560 if playlists were actually random, 36:37.560 --> 36:39.560 if they had no semantic meaning. 36:39.560 --> 36:42.560 For most people, they grouped these tracks for some reason. 36:42.560 --> 36:45.560 So we just happened across this incredible data set 36:45.560 --> 36:48.560 where people had taken these tens of millions of tracks 36:48.560 --> 36:52.560 and grouped them along different semantic vectors. 36:52.560 --> 36:55.560 And the semantics being outside the individual user, 36:55.560 --> 36:57.560 so it's some kind of universal. 36:57.560 --> 36:59.560 There's a universal embedding 36:59.560 --> 37:02.560 that holds across people on this earth. 37:02.560 --> 37:05.560 Yes, I do think that the embeddings you find 37:05.560 --> 37:08.560 are going to be reflective of the people who play listed. 37:08.560 --> 37:11.560 So if you have a lot of indie lovers who play list 37:11.560 --> 37:14.560 your embedding is going to perform better there. 37:14.560 --> 37:17.560 But what we found was that, yes, 37:17.560 --> 37:20.560 there were these latent similarities. 37:20.560 --> 37:22.560 They were very powerful. 37:22.560 --> 37:25.560 And it was interesting because 37:25.560 --> 37:28.560 I think that the people who play listed the most initially 37:28.560 --> 37:31.560 were the so called music aficionados 37:31.560 --> 37:33.560 who were really into music. 37:33.560 --> 37:35.560 And they often had a certain... 37:35.560 --> 37:39.560 Their taste was often geared towards a certain type of music. 37:39.560 --> 37:41.560 And so what surprised us, 37:41.560 --> 37:43.560 if you look at the problem from the outside, 37:43.560 --> 37:46.560 you might expect that the algorithms 37:46.560 --> 37:49.560 would start performing best with mainstreamers first 37:49.560 --> 37:52.560 and now feels like an easier problem to solve mainstream taste 37:52.560 --> 37:54.560 than really particular taste. 37:54.560 --> 37:56.560 It was a complete opposite for us. 37:56.560 --> 37:58.560 The recommendations performed fantastically 37:58.560 --> 38:01.560 for people who saw themselves as having very unique taste. 38:01.560 --> 38:04.560 That's probably because all of them play listed 38:04.560 --> 38:06.560 and they didn't perform so well for mainstreamers. 38:06.560 --> 38:09.560 They actually thought they were a bit too particular 38:09.560 --> 38:11.560 and unorthodox. 38:11.560 --> 38:13.560 So we had a complete opposite of what we expected. 38:13.560 --> 38:15.560 Success within the hardest problem first 38:15.560 --> 38:18.560 and then had to try to scale to more mainstream recommendations. 38:18.560 --> 38:21.560 So you've also acquired 38:21.560 --> 38:23.560 Echo Nest 38:23.560 --> 38:25.560 that analyzes song data. 38:25.560 --> 38:27.560 So 38:27.560 --> 38:29.560 in your view, maybe you can talk about 38:29.560 --> 38:31.560 so what kind of data is there 38:31.560 --> 38:33.560 from a machine learning perspective? 38:33.560 --> 38:35.560 There's a huge amount, 38:35.560 --> 38:37.560 what we're talking about, play listing 38:37.560 --> 38:39.560 and just user data 38:39.560 --> 38:42.560 of what people are listening to, the playlist they're constructing 38:42.560 --> 38:44.560 and so on. 38:44.560 --> 38:47.560 And then there is the actual data within a song 38:47.560 --> 38:49.560 what makes a song, 38:49.560 --> 38:52.560 I don't know, the actual waveforms. 38:52.560 --> 38:55.560 How do you mix the two? 38:55.560 --> 38:57.560 How much values are in each? 38:57.560 --> 38:59.560 To me, it seems like user data 38:59.560 --> 39:02.560 is a romantic notion 39:02.560 --> 39:05.560 that the song itself would contain useful information 39:05.560 --> 39:07.560 but if I were to guess, 39:07.560 --> 39:09.560 user data would be much more powerful. 39:09.560 --> 39:11.560 Playlists would be much more powerful. 39:11.560 --> 39:14.560 Yeah, so we use both. 39:14.560 --> 39:17.560 Our biggest success initially was with 39:17.560 --> 39:20.560 playlist data without understanding 39:20.560 --> 39:22.560 anything about the structure of the song. 39:22.560 --> 39:24.560 But when we acquired the Echo Nest, 39:24.560 --> 39:26.560 they had the inverse problem. 39:26.560 --> 39:28.560 They actually didn't have any play data. 39:28.560 --> 39:30.560 They were a provider of recommendations 39:30.560 --> 39:32.560 but they didn't actually have any play data. 39:32.560 --> 39:35.560 So they looked at the structure of songs 39:35.560 --> 39:37.560 sonically 39:37.560 --> 39:39.560 and they looked at Wikipedia for cultural references 39:39.560 --> 39:41.560 and so forth, right? 39:41.560 --> 39:43.560 And did a lot of NLU and so forth. 39:43.560 --> 39:45.560 They had that skill into the company 39:45.560 --> 39:48.560 and combined our user data 39:48.560 --> 39:52.560 with their content based. 39:52.560 --> 39:54.560 So you can think of it as we were user based 39:54.560 --> 39:56.560 and they were content based in their recommendations. 39:56.560 --> 39:58.560 And we combined those two. 39:58.560 --> 40:00.560 And for some cases where you have a new song 40:00.560 --> 40:02.560 that has no play data, 40:02.560 --> 40:04.560 obviously you have to try to go by 40:04.560 --> 40:06.560 either who the artist is 40:06.560 --> 40:09.560 or the sonic information in the song 40:09.560 --> 40:11.560 or what it's similar to. 40:11.560 --> 40:13.560 And we do a lot in both. 40:13.560 --> 40:15.560 But I would say yes. 40:15.560 --> 40:17.560 The user data captures things that 40:17.560 --> 40:19.560 have to do with culture in the greater society 40:19.560 --> 40:21.560 that you would never see 40:21.560 --> 40:23.560 in the content itself. 40:23.560 --> 40:25.560 But that said, we have seen 40:25.560 --> 40:27.560 we have a research lab in Paris 40:27.560 --> 40:29.560 when we can talk about 40:29.560 --> 40:31.560 more about that on 40:31.560 --> 40:33.560 kind of machine learning on the creator side. 40:33.560 --> 40:35.560 What it can do for creators, not just for the consumers. 40:35.560 --> 40:37.560 But where we looked at 40:37.560 --> 40:39.560 how does the structure of a song actually affect 40:39.560 --> 40:41.560 the listening behavior. 40:41.560 --> 40:43.560 And it turns out that 40:43.560 --> 40:45.560 we can predict things like skips 40:45.560 --> 40:47.560 based on the song itself. 40:47.560 --> 40:49.560 We could say that 40:49.560 --> 40:51.560 maybe you should move that chorus a bit 40:51.560 --> 40:53.560 because your skip is going to go up here. 40:53.560 --> 40:55.560 There is a lot of latent structure in the music 40:55.560 --> 40:57.560 which is not surprising 40:57.560 --> 40:59.560 because it is some sort of mind hack. 40:59.560 --> 41:01.560 So there should be structure. 41:01.560 --> 41:03.560 That's probably what we respond to. 41:03.560 --> 41:05.560 You just blew my mind actually 41:05.560 --> 41:07.560 from the creator perspective. 41:07.560 --> 41:09.560 That probably most creators 41:09.560 --> 41:11.560 are not taken advantage of. 41:11.560 --> 41:13.560 So I have recently 41:13.560 --> 41:15.560 got to interact with a few 41:15.560 --> 41:17.560 folks, YouTubers 41:17.560 --> 41:19.560 who are 41:19.560 --> 41:21.560 obsessed with this idea 41:21.560 --> 41:23.560 of what do I 41:23.560 --> 41:25.560 do to make sure people 41:25.560 --> 41:27.560 keep watching the video. 41:27.560 --> 41:29.560 And then you look at the analytics 41:29.560 --> 41:31.560 of which point do people turn it off 41:31.560 --> 41:33.560 and so on. First of all, 41:33.560 --> 41:35.560 I don't think that's healthy. 41:35.560 --> 41:37.560 It's because you can do it a little too much. 41:37.560 --> 41:39.560 But it is a really 41:39.560 --> 41:41.560 powerful tool for helping the creative process. 41:41.560 --> 41:43.560 You just made me 41:43.560 --> 41:45.560 realize you could do the same thing for 41:45.560 --> 41:47.560 creation of music. 41:47.560 --> 41:49.560 So is that something you've looked 41:49.560 --> 41:51.560 into? 41:51.560 --> 41:53.560 Can you speak 41:53.560 --> 41:55.560 to how much opportunity there is for that kind of thing? 41:55.560 --> 41:57.560 I listened to the podcast 41:57.560 --> 41:59.560 with Zirash and I thought it was 41:59.560 --> 42:01.560 fantastic and reacted to the same thing 42:01.560 --> 42:03.560 where he said he posted something 42:03.560 --> 42:05.560 in the morning, immediately 42:05.560 --> 42:07.560 watched the feedback where the drop off was 42:07.560 --> 42:09.560 and then responded to that in the afternoon. 42:09.560 --> 42:11.560 Which is quite different 42:11.560 --> 42:13.560 from how people make podcasts, for example. 42:13.560 --> 42:15.560 The feedback loop is almost 42:15.560 --> 42:17.560 non existent. So if we 42:17.560 --> 42:19.560 back out at one level, I think 42:19.560 --> 42:21.560 actually both for music 42:21.560 --> 42:23.560 and podcasts, which we also 42:23.560 --> 42:25.560 do at Spotify, I think there's 42:25.560 --> 42:27.560 tremendous opportunity just 42:27.560 --> 42:29.560 for the creation workflow. 42:29.560 --> 42:31.560 I think it's really interesting 42:31.560 --> 42:33.560 speaking to you, because you're 42:33.560 --> 42:35.560 a musician, a developer and a 42:35.560 --> 42:37.560 podcaster, if you think about those 42:37.560 --> 42:39.560 three different roles, 42:39.560 --> 42:41.560 if you make the leap as a musician, 42:41.560 --> 42:43.560 if you think about it 42:43.560 --> 42:45.560 as a software tool chain, really, 42:45.560 --> 42:47.560 your door with the stems, 42:47.560 --> 42:49.560 that's the IDE, right? 42:49.560 --> 42:51.560 That's where you work in source code format 42:51.560 --> 42:53.560 with what you're creating. 42:53.560 --> 42:55.560 Then you sit around and you play with that 42:55.560 --> 42:57.560 and when you're happy you compile that thing into 42:57.560 --> 42:59.560 some sort of, you know, AAC, 42:59.560 --> 43:01.560 or something, you do that because 43:01.560 --> 43:03.560 you get distribution. There are so many run times 43:03.560 --> 43:05.560 for that MP3 across the world in car stairs and stuff. 43:05.560 --> 43:07.560 So you kind of compile this executable and you ship it out 43:07.560 --> 43:09.560 in kind of an old fashioned 43:09.560 --> 43:11.560 box software analogy. 43:11.560 --> 43:13.560 And then you hope for the best, 43:13.560 --> 43:15.560 right? But as 43:15.560 --> 43:17.560 a software developer, 43:17.560 --> 43:19.560 you would never do that. First you go 43:19.560 --> 43:21.560 on GitHub and you collaborate with other creators. 43:21.560 --> 43:23.560 And then, you know, 43:23.560 --> 43:25.560 you think it would be crazy to just ship one version 43:25.560 --> 43:27.560 of your software without doing an A.B. test, 43:27.560 --> 43:29.560 without any feedback loop. 43:29.560 --> 43:31.560 And then, issue tracking. 43:31.560 --> 43:33.560 Exactly. And then you would look at the 43:33.560 --> 43:35.560 feedback loops and try to optimize that thing, right? 43:35.560 --> 43:37.560 So I think if you think 43:37.560 --> 43:39.560 about it as a very specific software tool chain, 43:39.560 --> 43:41.560 it looks 43:41.560 --> 43:43.560 quite arcane. 43:43.560 --> 43:45.560 The tools that a music creator has versus 43:45.560 --> 43:47.560 what a software developer has. 43:47.560 --> 43:49.560 So that's kind of how we think about it. 43:49.560 --> 43:51.560 Why wouldn't 43:51.560 --> 43:53.560 a music creator 43:53.560 --> 43:55.560 have something like GitHub where you could collaborate 43:55.560 --> 43:57.560 much more easily? So we bought 43:57.560 --> 43:59.560 this company called Soundtrap, 43:59.560 --> 44:01.560 which has a kind of 44:01.560 --> 44:03.560 Google Docs for music approach where you can 44:03.560 --> 44:05.560 collaborate with other people on the kind of 44:05.560 --> 44:07.560 source code format with Stems. 44:07.560 --> 44:09.560 And I think introducing things like 44:09.560 --> 44:11.560 AI tools there to help you 44:11.560 --> 44:13.560 as you're creating music, 44:13.560 --> 44:15.560 both in 44:15.560 --> 44:17.560 helping you 44:19.560 --> 44:21.560 put a component to your music, 44:21.560 --> 44:23.560 like drums or something, 44:23.560 --> 44:25.560 help you 44:25.560 --> 44:27.560 master and mix automatically, 44:27.560 --> 44:29.560 help you understand how this track will perform. 44:29.560 --> 44:31.560 Exactly what you would expect as a software 44:31.560 --> 44:33.560 developer. I think it makes a lot of sense. 44:33.560 --> 44:35.560 And I think the same goes for 44:35.560 --> 44:37.560 a podcaster. I think podcasters will expect 44:37.560 --> 44:39.560 to have the same kind of feedback loop that Sirosh 44:39.560 --> 44:41.560 has. Like, 44:41.560 --> 44:43.560 why wouldn't you? Maybe it's not healthy, but 44:43.560 --> 44:45.560 Sorry, I wanted to 44:45.560 --> 44:47.560 criticize the fact because you can overdo it. 44:47.560 --> 44:49.560 Because a lot of the 44:49.560 --> 44:51.560 and we're in a new era 44:51.560 --> 44:53.560 of that. 44:53.560 --> 44:55.560 So you can become 44:55.560 --> 44:57.560 addicted to it and 44:57.560 --> 44:59.560 therefore 44:59.560 --> 45:01.560 what people say you become a slave to the YouTube 45:01.560 --> 45:03.560 algorithm. 45:03.560 --> 45:05.560 It's always a danger 45:05.560 --> 45:07.560 of a new technology as opposed to 45:07.560 --> 45:09.560 say if you're creating a song 45:09.560 --> 45:11.560 becoming too obsessed 45:11.560 --> 45:13.560 about the 45:13.560 --> 45:15.560 intro riff to the song 45:15.560 --> 45:17.560 that keeps people listening versus 45:17.560 --> 45:19.560 actually the entirety of the creation process. It's a balance. 45:19.560 --> 45:21.560 But the fact that 45:21.560 --> 45:23.560 there's zero, I mean you're blowing my mind right now 45:23.560 --> 45:25.560 because you're 45:25.560 --> 45:27.560 completely right that there's no 45:27.560 --> 45:29.560 signal whatsoever. There's no feedback 45:29.560 --> 45:31.560 whatsoever on the creation process 45:31.560 --> 45:33.560 and musical podcasting 45:33.560 --> 45:35.560 almost at all. 45:35.560 --> 45:37.560 And are you saying 45:37.560 --> 45:39.560 that Spotify 45:39.560 --> 45:41.560 is hoping to help create tools 45:41.560 --> 45:43.560 to, not tools, but 45:43.560 --> 45:45.560 No, tools actually. 45:45.560 --> 45:47.560 Actually tools for creators. 45:47.560 --> 45:49.560 Absolutely. 45:49.560 --> 45:51.560 So we have 45:51.560 --> 45:53.560 we've made some acquisitions the last few years around music creation 45:53.560 --> 45:55.560 this company called Soundtrap 45:55.560 --> 45:57.560 which is a digital audio workstation 45:57.560 --> 45:59.560 that is browser based 45:59.560 --> 46:01.560 and their focus was really the Google Docs approach. 46:01.560 --> 46:03.560 We can collaborate with people much more easily 46:03.560 --> 46:05.560 than you could in previous tools. 46:05.560 --> 46:07.560 So we have some of these tools 46:07.560 --> 46:09.560 that we're working with that we want to make accessible 46:09.560 --> 46:11.560 and then we can connect it 46:11.560 --> 46:13.560 with our consumption data. 46:13.560 --> 46:15.560 We can create this feedback loop where 46:15.560 --> 46:17.560 we could help you understand 46:17.560 --> 46:19.560 we could help you create 46:19.560 --> 46:21.560 and help you understand how you will perform. 46:21.560 --> 46:23.560 We also acquired this other company 46:23.560 --> 46:25.560 with a podcasting called Anchor 46:25.560 --> 46:27.560 which is one of the biggest podcasting tools 46:27.560 --> 46:29.560 mobile focused. 46:29.560 --> 46:31.560 So really focused on simple creation 46:31.560 --> 46:33.560 or easy access to creation 46:33.560 --> 46:35.560 but that also gives us this feedback loop 46:35.560 --> 46:37.560 and even before that 46:37.560 --> 46:39.560 we invested in something called 46:39.560 --> 46:41.560 Spotify for Artists 46:41.560 --> 46:43.560 and Spotify for Podcasters 46:43.560 --> 46:45.560 which is an app that you can download 46:45.560 --> 46:47.560 you can verify that you are that creator 46:47.560 --> 46:49.560 and then you get 46:49.560 --> 46:51.560 things that 46:51.560 --> 46:53.560 software developers have had for years. 46:53.560 --> 46:55.560 You can see where 46:55.560 --> 46:57.560 if you look at your podcast for example on Spotify 46:57.560 --> 46:59.560 or a song that you released 46:59.560 --> 47:01.560 you can see how it's performing 47:01.560 --> 47:03.560 which citizen it's performing in 47:03.560 --> 47:05.560 who's listening to it, what's the demographic breakup 47:05.560 --> 47:07.560 so similar in the sense that 47:07.560 --> 47:09.560 you can understand how you're actually doing on the platform. 47:09.560 --> 47:11.560 So we definitely want 47:11.560 --> 47:13.560 to build tools. I think 47:13.560 --> 47:15.560 you also interviewed 47:15.560 --> 47:17.560 the head of research for Adobe 47:17.560 --> 47:19.560 and I think that's 47:19.560 --> 47:21.560 back to Photoshop that you like. 47:21.560 --> 47:23.560 I think that's an interesting analogy as well. 47:23.560 --> 47:25.560 Photoshop I think has been 47:25.560 --> 47:27.560 very innovative in helping 47:27.560 --> 47:29.560 photographers and artists 47:29.560 --> 47:31.560 and I think there should be 47:31.560 --> 47:33.560 the same kind of tools for 47:33.560 --> 47:35.560 music creators where you could get 47:35.560 --> 47:37.560 AI assistants for example as you're creating music 47:37.560 --> 47:39.560 as you can do 47:39.560 --> 47:41.560 with Adobe where you can 47:41.560 --> 47:43.560 I want a sky over here and you can get help creating that sky. 47:43.560 --> 47:45.560 The really fascinating thing is 47:45.560 --> 47:47.560 what Adobe 47:47.560 --> 47:49.560 doesn't have 47:49.560 --> 47:51.560 is a distribution for the content 47:51.560 --> 47:53.560 you create. So you don't have 47:53.560 --> 47:55.560 the data of if I create 47:55.560 --> 47:57.560 if I 47:57.560 --> 47:59.560 you know whatever creation I make 47:59.560 --> 48:01.560 in Photoshop or Premiere 48:01.560 --> 48:03.560 I can't get like immediate feedback 48:03.560 --> 48:05.560 like I can on YouTube for example about 48:05.560 --> 48:07.560 the way people are responding 48:07.560 --> 48:09.560 and if Spotify is creating those tools 48:09.560 --> 48:11.560 that's a really exciting 48:11.560 --> 48:13.560 actually world 48:13.560 --> 48:15.560 but 48:15.560 --> 48:17.560 let's talk a little about podcasts 48:17.560 --> 48:19.560 so I have trouble 48:19.560 --> 48:21.560 talking to one person 48:21.560 --> 48:23.560 so it's a bit terrifying 48:23.560 --> 48:25.560 and kind of hard to fathom 48:25.560 --> 48:27.560 but on average 48:27.560 --> 48:29.560 60 to 100,000 48:29.560 --> 48:31.560 people will listen to this episode 48:31.560 --> 48:33.560 okay so 48:33.560 --> 48:35.560 it's intimidating 48:35.560 --> 48:37.560 so I hosted on Blueberry 48:37.560 --> 48:39.560 I don't know if I'm pronouncing 48:39.560 --> 48:41.560 that correctly actually 48:41.560 --> 48:43.560 it looks like most people listen to it on Apple 48:43.560 --> 48:45.560 Podcast, Cast Box and Pocket 48:45.560 --> 48:47.560 Cast and only about 48:47.560 --> 48:49.560 a thousand 48:49.560 --> 48:51.560 listen on Spotify 48:51.560 --> 48:53.560 just my podcast right 48:53.560 --> 48:55.560 so 48:55.560 --> 48:57.560 where 48:57.560 --> 48:59.560 do you see a time when Spotify 48:59.560 --> 49:01.560 will dominate this so Spotify 49:01.560 --> 49:03.560 is relatively new 49:03.560 --> 49:05.560 in podcasting 49:05.560 --> 49:07.560 so yeah in podcasting 49:07.560 --> 49:09.560 what's the deal with podcasting in Spotify 49:09.560 --> 49:11.560 how serious 49:11.560 --> 49:13.560 is Spotify about podcasting 49:13.560 --> 49:15.560 do you see a time where everybody would listen 49:15.560 --> 49:17.560 to you know probably 49:17.560 --> 49:19.560 a huge amount of people, majority perhaps 49:19.560 --> 49:21.560 listen to music on Spotify 49:21.560 --> 49:23.560 do you see a time 49:23.560 --> 49:25.560 when the same is true for 49:25.560 --> 49:27.560 podcasting 49:27.560 --> 49:29.560 well I certainly hope so that is our mission 49:29.560 --> 49:31.560 our mission as a company is actually to 49:31.560 --> 49:33.560 enable a million creators to live 49:33.560 --> 49:35.560 off of their art and a billion people inspired 49:35.560 --> 49:37.560 by it and what I think is interesting about that mission 49:37.560 --> 49:39.560 is it actually puts the creators 49:39.560 --> 49:41.560 first even though it started as a 49:41.560 --> 49:43.560 consumer focused company and it says 49:43.560 --> 49:45.560 to be able to live off of their art not just 49:45.560 --> 49:47.560 make some money off of their art as well 49:47.560 --> 49:49.560 so it's quite an ambitious 49:49.560 --> 49:51.560 product and 49:51.560 --> 49:53.560 so we think about creators of all kinds 49:53.560 --> 49:55.560 and 49:55.560 --> 49:57.560 we kind of expanded our mission from 49:57.560 --> 49:59.560 being music to being audio 49:59.560 --> 50:01.560 a while back 50:01.560 --> 50:03.560 and that's not 50:03.560 --> 50:05.560 so much because 50:05.560 --> 50:07.560 we think we made that decision 50:07.560 --> 50:09.560 we think that decision 50:09.560 --> 50:11.560 was made for us 50:11.560 --> 50:13.560 we think the world made that decision 50:13.560 --> 50:15.560 whether we like it or not 50:15.560 --> 50:17.560 when you put in your headphones 50:17.560 --> 50:19.560 you're going to make a choice between 50:19.560 --> 50:21.560 music 50:21.560 --> 50:23.560 and a new episode of 50:23.560 --> 50:25.560 your podcast or something else 50:25.560 --> 50:27.560 we're in that world whether we like it or not 50:27.560 --> 50:29.560 that's how radio work 50:29.560 --> 50:31.560 so we decided that 50:31.560 --> 50:33.560 we think it's about audio you can see the 50:33.560 --> 50:35.560 rights of audio books and so forth 50:35.560 --> 50:37.560 we think audio is this great opportunity 50:37.560 --> 50:39.560 so we decided to enter it and obviously 50:39.560 --> 50:41.560 Apple 50:41.560 --> 50:43.560 and Apple podcast is absolutely 50:43.560 --> 50:45.560 dominating in 50:45.560 --> 50:47.560 podcasting and we didn't have 50:47.560 --> 50:49.560 a single podcast only like two years ago 50:49.560 --> 50:51.560 what we did though was 50:51.560 --> 50:53.560 we 50:53.560 --> 50:55.560 we looked at this and said 50:55.560 --> 50:57.560 bring something to this 50:57.560 --> 50:59.560 we want to do this but back to the original 50:59.560 --> 51:01.560 Spotify we had to do something that consumers actually value 51:03.560 --> 51:05.560 to be able to do this and the reason 51:05.560 --> 51:07.560 we've gone from not existing at all 51:07.560 --> 51:09.560 to being the 51:09.560 --> 51:11.560 quite a wide margin the second largest podcast 51:11.560 --> 51:13.560 consumption 51:13.560 --> 51:15.560 still wide gap to iTunes but we're growing 51:15.560 --> 51:17.560 quite fast 51:17.560 --> 51:19.560 I think it's because when we looked at the consumer 51:19.560 --> 51:21.560 problem 51:21.560 --> 51:23.560 people said surprisingly that they wanted their podcasts 51:23.560 --> 51:25.560 and music in the same 51:25.560 --> 51:27.560 in the same application 51:27.560 --> 51:29.560 so what we did was we took a little bit of a different 51:29.560 --> 51:31.560 approach what we said instead of building a separate podcast 51:31.560 --> 51:33.560 app 51:33.560 --> 51:35.560 we thought is there a consumer problem to solve here 51:35.560 --> 51:37.560 because the others are very successful already 51:37.560 --> 51:39.560 and we thought there was in making a more 51:39.560 --> 51:41.560 seamless experience where you can have your podcasts 51:41.560 --> 51:43.560 and your music 51:43.560 --> 51:45.560 in the same application 51:45.560 --> 51:47.560 because we think it's audio to you 51:47.560 --> 51:49.560 and that has been successful and that meant that 51:49.560 --> 51:51.560 we actually had 200 million people 51:51.560 --> 51:53.560 to offer this to instead of starting from zero 51:53.560 --> 51:55.560 so I think we have a good 51:55.560 --> 51:57.560 chance because we're taking a different approach 51:57.560 --> 51:59.560 than the competition and back to the other thing 51:59.560 --> 52:01.560 I mentioned about 52:01.560 --> 52:03.560 creators 52:03.560 --> 52:05.560 because we're looking at the end to end flow 52:05.560 --> 52:07.560 I think there's a tremendous amount of innovation 52:07.560 --> 52:09.560 to do around podcast as a format 52:09.560 --> 52:11.560 when we have creation tools and consumption 52:11.560 --> 52:13.560 I think we could 52:13.560 --> 52:15.560 start improving what podcasting is 52:15.560 --> 52:17.560 I mean podcast is this 52:17.560 --> 52:19.560 this opaque big like one to 52:19.560 --> 52:21.560 your file 52:21.560 --> 52:23.560 that you're streaming 52:23.560 --> 52:25.560 which it really doesn't make that much sense in 2019 52:25.560 --> 52:27.560 that it's not interactive 52:27.560 --> 52:29.560 there's no feedback loops nothing like that 52:29.560 --> 52:31.560 so I think if we're going to win it's going to have to be 52:31.560 --> 52:33.560 because we build a better product 52:33.560 --> 52:35.560 for creators and for consumers 52:35.560 --> 52:37.560 so we'll see but it's certainly our goal 52:37.560 --> 52:39.560 we have a long way to go 52:39.560 --> 52:41.560 well the creators part is really exciting 52:41.560 --> 52:43.560 you got me hooked there 52:43.560 --> 52:45.560 it's the only stats I have 52:45.560 --> 52:47.560 Blueberry just recently added the stats of 52:47.560 --> 52:49.560 it's listen to the end 52:49.560 --> 52:51.560 or not 52:51.560 --> 52:53.560 and that's like a huge improvement 52:53.560 --> 52:55.560 but that's still 52:55.560 --> 52:57.560 nowhere to where you could possibly 52:57.560 --> 52:59.560 go into statistics 52:59.560 --> 53:01.560 just download the Spotify podcasters up and verify 53:01.560 --> 53:03.560 and then you'll know where people dropped out in this episode 53:03.560 --> 53:05.560 oh wow okay 53:05.560 --> 53:07.560 the moment I started talking 53:07.560 --> 53:09.560 I might be depressed by this 53:09.560 --> 53:11.560 but okay so one 53:11.560 --> 53:13.560 one other question 53:13.560 --> 53:15.560 the original 53:15.560 --> 53:17.560 Spotify for music 53:17.560 --> 53:19.560 and I have a question about 53:19.560 --> 53:21.560 podcasting in this line 53:21.560 --> 53:23.560 is the idea of albums 53:23.560 --> 53:25.560 I have 53:25.560 --> 53:27.560 music aficionados 53:27.560 --> 53:29.560 friends who are really 53:29.560 --> 53:31.560 big fans of music 53:31.560 --> 53:33.560 often really enjoy albums 53:33.560 --> 53:35.560 listening to entire albums of 53:35.560 --> 53:37.560 an artist correct me if I'm wrong 53:37.560 --> 53:39.560 but I feel like 53:39.560 --> 53:41.560 Spotify has helped 53:41.560 --> 53:43.560 replace the idea of an album with playlists 53:43.560 --> 53:45.560 so you create your own albums 53:45.560 --> 53:47.560 that's kind of the way 53:47.560 --> 53:49.560 at least I've experienced music 53:49.560 --> 53:51.560 and I've really enjoyed it that way 53:51.560 --> 53:53.560 one of the things that was missing 53:53.560 --> 53:55.560 in podcasting for me 53:55.560 --> 53:57.560 I don't know if it's missing I don't know 53:57.560 --> 53:59.560 it's an open question for me 53:59.560 --> 54:01.560 but the way I listen to podcasts is the way I would listen to albums 54:01.560 --> 54:03.560 so I take 54:03.560 --> 54:05.560 Joe Rogan Experience 54:05.560 --> 54:07.560 and that's an album and I listen 54:07.560 --> 54:09.560 I put that on 54:09.560 --> 54:11.560 and I listen one episode after the next 54:11.560 --> 54:13.560 and there's a sequence and so on 54:13.560 --> 54:15.560 is there room for 54:15.560 --> 54:17.560 doing 54:17.560 --> 54:19.560 what Spotify did for music 54:19.560 --> 54:21.560 but creating 54:21.560 --> 54:23.560 playlists 54:23.560 --> 54:25.560 sort of this kind of playlisting 54:25.560 --> 54:27.560 idea of breaking apart from podcasting 54:27.560 --> 54:29.560 from individual 54:29.560 --> 54:31.560 podcasts and creating kind of 54:31.560 --> 54:33.560 this interplay 54:33.560 --> 54:35.560 or have you thought about that 54:35.560 --> 54:37.560 it's a great question so I think in 54:37.560 --> 54:39.560 music 54:39.560 --> 54:41.560 you're right basically you bought an album 54:41.560 --> 54:43.560 so it was like you bought a small catalog of 54:43.560 --> 54:45.560 10 tracks 54:45.560 --> 54:47.560 again it was actually a lot of consumption 54:47.560 --> 54:49.560 you think it's about what you like 54:49.560 --> 54:51.560 but it's based on the business model 54:51.560 --> 54:53.560 you paid for this 10 track 54:53.560 --> 54:55.560 service and then you listen to that for a while 54:55.560 --> 54:57.560 and then when everything was flat 54:57.560 --> 54:59.560 priced you tended to listen differently 54:59.560 --> 55:01.560 now so I think 55:01.560 --> 55:03.560 the album is still tremendously important that's why we have it 55:03.560 --> 55:05.560 and you can save albums and so forth 55:05.560 --> 55:07.560 and you have a huge amount of people who really listen 55:07.560 --> 55:09.560 according to albums and I like that because 55:09.560 --> 55:11.560 it is a creator format you can tell a longer 55:11.560 --> 55:13.560 story over several tracks 55:13.560 --> 55:15.560 and so some people listen to just one track 55:15.560 --> 55:17.560 some people actually want to hear that whole 55:17.560 --> 55:19.560 story 55:19.560 --> 55:21.560 now in podcast I think 55:21.560 --> 55:23.560 I think it's different 55:23.560 --> 55:25.560 you can argue that podcasts might be more 55:25.560 --> 55:27.560 like shows on Netflix 55:27.560 --> 55:29.560 you have like a full season of Narcos 55:29.560 --> 55:31.560 and you're probably not going to do like one 55:31.560 --> 55:33.560 episode of Narcos and then one of House of Cards 55:33.560 --> 55:35.560 like you know 55:35.560 --> 55:37.560 there's a narrative there 55:37.560 --> 55:39.560 and you love the cast 55:39.560 --> 55:41.560 and you love these characters 55:41.560 --> 55:43.560 so I think people love shows 55:43.560 --> 55:45.560 and I think they will 55:45.560 --> 55:47.560 listen to those shows 55:47.560 --> 55:49.560 I do think you follow a bunch of shows at the same time 55:49.560 --> 55:51.560 so there's certainly an opportunity to bring you the 55:51.560 --> 55:53.560 latest episode of 55:53.560 --> 55:55.560 whatever the 5, 6, 10 things that you're into 55:55.560 --> 55:57.560 but I think 55:57.560 --> 55:59.560 I think people are going to listen 55:59.560 --> 56:01.560 to specific hosts 56:01.560 --> 56:03.560 and love those hosts 56:03.560 --> 56:05.560 for a long time because I think there's something 56:05.560 --> 56:07.560 different with podcasts 56:07.560 --> 56:09.560 where this format 56:09.560 --> 56:11.560 of the 56:11.560 --> 56:13.560 experience of the audience 56:13.560 --> 56:15.560 is actually standing here right between us 56:15.560 --> 56:17.560 whereas if you look at something on TV 56:17.560 --> 56:19.560 the audio actually would come from 56:19.560 --> 56:21.560 you would sit over there and the audio would come to you 56:21.560 --> 56:23.560 from both of us as if you were watching 56:23.560 --> 56:25.560 not as if you were part of the conversation 56:25.560 --> 56:27.560 so my experience of having to listen to podcasts 56:27.560 --> 56:29.560 like yours and Joe Rogan 56:29.560 --> 56:31.560 I feel like I know all of these people 56:31.560 --> 56:33.560 and I have no idea who I am 56:33.560 --> 56:35.560 but I feel like I've listened to so many hours of them 56:35.560 --> 56:37.560 it's very different from me watching 56:37.560 --> 56:39.560 like a TV show or an interview 56:39.560 --> 56:41.560 so I think you kind of 56:41.560 --> 56:43.560 fall in love with people 56:43.560 --> 56:45.560 and experience in a different way 56:45.560 --> 56:47.560 so I think shows and hosts 56:47.560 --> 56:49.560 are going to be very important 56:49.560 --> 56:51.560 I don't think that's going to go away into some sort of thing 56:51.560 --> 56:53.560 where you don't even know who you're listening to 56:53.560 --> 56:55.560 I don't think that's going to happen 56:55.560 --> 56:57.560 what I do think is, I think there's a tremendous 56:57.560 --> 56:59.560 discovery opportunity 56:59.560 --> 57:01.560 in podcasts because 57:01.560 --> 57:03.560 the catalogue is growing quite quickly 57:03.560 --> 57:05.560 and 57:05.560 --> 57:07.560 I think podcasts is only 57:07.560 --> 57:09.560 a few like five, six hundred thousand 57:09.560 --> 57:11.560 shows right now 57:11.560 --> 57:13.560 if you look back to YouTube as another analogy 57:13.560 --> 57:15.560 for creators, no one really knows 57:15.560 --> 57:17.560 if you would lift the lid on 57:17.560 --> 57:19.560 YouTube but it's probably billions 57:19.560 --> 57:21.560 of episodes 57:21.560 --> 57:23.560 and so I think the podcast catalogue would probably grow 57:23.560 --> 57:25.560 tremendously because the creation 57:25.560 --> 57:27.560 tools are getting easier 57:27.560 --> 57:29.560 and you're going to have this 57:29.560 --> 57:31.560 discovery opportunity that I think is really big 57:31.560 --> 57:33.560 so a lot of people tell me that they love their shows 57:33.560 --> 57:35.560 but discovering 57:35.560 --> 57:37.560 podcasts kind of suck 57:37.560 --> 57:39.560 it's really hard to get into new show 57:39.560 --> 57:41.560 they're usually quite long, it's a big time investment 57:41.560 --> 57:43.560 so I think there's plenty of opportunity 57:43.560 --> 57:45.560 in the discovery part 57:45.560 --> 57:47.560 yeah for sure, a hundred percent 57:47.560 --> 57:49.560 and even the dumbest 57:49.560 --> 57:51.560 there's so many low hanging fruit too 57:51.560 --> 57:53.560 for example 57:53.560 --> 57:55.560 just knowing 57:55.560 --> 57:57.560 what episode to listen to first 57:57.560 --> 57:59.560 to try out a podcast 57:59.560 --> 58:01.560 exactly because most podcasts 58:01.560 --> 58:03.560 don't have an order to them 58:03.560 --> 58:05.560 they can be listened to out of order 58:05.560 --> 58:07.560 and 58:07.560 --> 58:09.560 sorry to say 58:09.560 --> 58:11.560 some are better than others 58:11.560 --> 58:13.560 episodes so some episodes of Joe Rogan 58:13.560 --> 58:15.560 are better than others 58:15.560 --> 58:17.560 and it's nice to know 58:17.560 --> 58:19.560 which you should listen to 58:19.560 --> 58:21.560 to try it out and there's as far 58:21.560 --> 58:23.560 as I know almost no information 58:23.560 --> 58:25.560 in terms of 58:25.560 --> 58:27.560 like upvotes 58:27.560 --> 58:29.560 on how good an episode is 58:29.560 --> 58:31.560 so I think part of the problem is 58:31.560 --> 58:33.560 it's kind of like music 58:33.560 --> 58:35.560 there isn't one answer, people use music for different things 58:35.560 --> 58:37.560 and there's actually many different types of music 58:37.560 --> 58:39.560 there's workout music and there's classical piano music 58:39.560 --> 58:41.560 and focus music and 58:41.560 --> 58:43.560 and so forth 58:43.560 --> 58:45.560 I think the same with podcasts, some podcasts are sequential 58:45.560 --> 58:47.560 they're supposed to be listened to 58:47.560 --> 58:49.560 in order 58:49.560 --> 58:51.560 it's actually telling and narrative 58:51.560 --> 58:53.560 podcasts are one topic 58:53.560 --> 58:55.560 kind of like yours but different guests 58:55.560 --> 58:57.560 so you could jump in anywhere 58:57.560 --> 58:59.560 some podcasts actually have completely different topics 58:59.560 --> 59:01.560 and for those podcasts it might be that 59:01.560 --> 59:03.560 we should recommend 59:03.560 --> 59:05.560 one episode because it's about AI 59:05.560 --> 59:07.560 from someone 59:07.560 --> 59:09.560 but then they talk about something that you're not interested in 59:09.560 --> 59:11.560 the rest of the episode 59:11.560 --> 59:13.560 so I think what we're spending a lot of time on now 59:13.560 --> 59:15.560 is just first understanding the domain 59:15.560 --> 59:17.560 and creating kind of the knowledge graph 59:17.560 --> 59:19.560 of 59:19.560 --> 59:21.560 how do these objects relate 59:21.560 --> 59:23.560 and how do people consume and I think we'll find that it's going to be 59:23.560 --> 59:25.560 it's going to be different 59:25.560 --> 59:27.560 I'm excited 59:27.560 --> 59:29.560 Spotify is the first 59:29.560 --> 59:31.560 people I'm aware of that are 59:31.560 --> 59:33.560 trying to do this 59:33.560 --> 59:35.560 for podcasting 59:35.560 --> 59:37.560 podcasting has been like a wild west 59:37.560 --> 59:39.560 until now 59:39.560 --> 59:41.560 we want to be very careful though 59:41.560 --> 59:43.560 because it's been a very good wild west 59:43.560 --> 59:45.560 I think it's this fragile ecosystem 59:45.560 --> 59:47.560 and 59:47.560 --> 59:49.560 we want to make sure that 59:49.560 --> 59:51.560 you don't barge in and say 59:51.560 --> 59:53.560 we're going to internetize this thing 59:53.560 --> 59:55.560 and you have to think about 59:55.560 --> 59:57.560 the creators, you have to understand 59:57.560 --> 59:59.560 how they get distribution today 59:59.560 --> 1:00:01.560 who listens to how they make money today 1:00:01.560 --> 1:00:03.560 try to 1:00:03.560 --> 1:00:05.560 make sure that their business model works, that they understand 1:00:05.560 --> 1:00:07.560 I think it's back to doing something 1:00:07.560 --> 1:00:09.560 improving their products 1:00:09.560 --> 1:00:11.560 like feedback loops and distribution 1:00:13.560 --> 1:00:15.560 jumping back into terms of this 1:00:15.560 --> 1:00:17.560 fascinating world of 1:00:17.560 --> 1:00:19.560 recommender system and listening to music 1:00:19.560 --> 1:00:21.560 and using machine learning to analyze things 1:00:21.560 --> 1:00:23.560 do you think it's better 1:00:23.560 --> 1:00:25.560 to 1:00:25.560 --> 1:00:27.560 what currently 1:00:27.560 --> 1:00:29.560 correct me if I'm wrong but 1:00:29.560 --> 1:00:31.560 Spotify lets people pick what they listen to 1:00:31.560 --> 1:00:33.560 for the most part 1:00:33.560 --> 1:00:35.560 there's a discovery process but you kind of 1:00:35.560 --> 1:00:37.560 organize playlists 1:00:37.560 --> 1:00:39.560 is it better to let people pick 1:00:39.560 --> 1:00:41.560 what they listen to or recommend 1:00:41.560 --> 1:00:43.560 what they should listen to 1:00:43.560 --> 1:00:45.560 something like stations by Spotify 1:00:45.560 --> 1:00:47.560 that I saw that you're playing around with 1:00:47.560 --> 1:00:49.560 maybe you can tell me 1:00:49.560 --> 1:00:51.560 what's the status of that 1:00:51.560 --> 1:00:53.560 this is a Pandora style app 1:00:53.560 --> 1:00:55.560 that just kind of as opposed to you select 1:00:55.560 --> 1:00:57.560 the music you listen to 1:00:57.560 --> 1:00:59.560 it kind of feeds you 1:00:59.560 --> 1:01:01.560 the music you listen to 1:01:01.560 --> 1:01:03.560 what's the status of stations by Spotify 1:01:03.560 --> 1:01:05.560 what's its future 1:01:05.560 --> 1:01:07.560 the store is Spotify as we have grown 1:01:07.560 --> 1:01:09.560 has been that we made it more accessible 1:01:09.560 --> 1:01:11.560 to different audiences 1:01:11.560 --> 1:01:13.560 and 1:01:13.560 --> 1:01:15.560 stations is another one of those where 1:01:15.560 --> 1:01:17.560 the question is 1:01:17.560 --> 1:01:19.560 some people want to be very specific 1:01:19.560 --> 1:01:21.560 they actually want to start with heaven right now 1:01:21.560 --> 1:01:23.560 that needs to be very easy to do 1:01:23.560 --> 1:01:25.560 and some people or even the same person 1:01:25.560 --> 1:01:27.560 at some point might say 1:01:27.560 --> 1:01:29.560 I want to feel upbeat 1:01:29.560 --> 1:01:31.560 or I want to feel happy 1:01:31.560 --> 1:01:33.560 or I want songs to sing in the car 1:01:33.560 --> 1:01:35.560 so they put in the information 1:01:35.560 --> 1:01:37.560 at a very different level 1:01:37.560 --> 1:01:39.560 and then we need to translate that into 1:01:39.560 --> 1:01:41.560 musically 1:01:41.560 --> 1:01:43.560 so stations is a test to 1:01:43.560 --> 1:01:45.560 create like a consumption input vector 1:01:45.560 --> 1:01:47.560 that is much simpler where you can just tune it a little bit 1:01:47.560 --> 1:01:49.560 and see if that increases the overall reach 1:01:49.560 --> 1:01:51.560 but we're trying to kind of serve 1:01:51.560 --> 1:01:53.560 the entire gamut of super advanced 1:01:53.560 --> 1:01:55.560 so called music aficionados 1:01:55.560 --> 1:01:57.560 all the way to 1:01:57.560 --> 1:01:59.560 to people who 1:01:59.560 --> 1:02:01.560 they love listening to music but it's not 1:02:01.560 --> 1:02:03.560 their number one priority in life 1:02:03.560 --> 1:02:05.560 they're not going to sit and follow every new release 1:02:05.560 --> 1:02:07.560 from every new artist 1:02:07.560 --> 1:02:09.560 to influence music at 1:02:09.560 --> 1:02:11.560 at a different level 1:02:11.560 --> 1:02:13.560 so we're trying, you can think of it as different 1:02:13.560 --> 1:02:15.560 products and I think when 1:02:15.560 --> 1:02:17.560 one of the interesting things 1:02:17.560 --> 1:02:19.560 to answer your question on 1:02:19.560 --> 1:02:21.560 if it's better to lift the user choose 1:02:21.560 --> 1:02:23.560 or to play I think the answer is 1:02:23.560 --> 1:02:25.560 the challenge when you 1:02:25.560 --> 1:02:27.560 when machine learning kind of came along 1:02:27.560 --> 1:02:29.560 there was a lot of thinking about 1:02:29.560 --> 1:02:31.560 what does product development mean 1:02:31.560 --> 1:02:33.560 in a machine learning context 1:02:33.560 --> 1:02:35.560 people like Andrew Eng for example 1:02:35.560 --> 1:02:37.560 when he went to Baidu 1:02:37.560 --> 1:02:39.560 he started doing a lot of practical machine learning 1:02:39.560 --> 1:02:41.560 went from academia and he thought a lot about this 1:02:41.560 --> 1:02:43.560 and he had this notion that 1:02:43.560 --> 1:02:45.560 you know product manager, designer 1:02:45.560 --> 1:02:47.560 and they used to work around this wireframe 1:02:47.560 --> 1:02:49.560 kind of describe what the product should look like 1:02:49.560 --> 1:02:51.560 was something to talk about when you're doing 1:02:51.560 --> 1:02:53.560 like a chat bot or a playlist 1:02:53.560 --> 1:02:55.560 what are you going to say like it should be good 1:02:55.560 --> 1:02:57.560 that's not a good product description 1:02:57.560 --> 1:02:59.560 so how do you do that and he came up with this notion 1:02:59.560 --> 1:03:01.560 that 1:03:01.560 --> 1:03:03.560 the test set is the new wireframe 1:03:03.560 --> 1:03:05.560 and the job of the product manager is to source 1:03:05.560 --> 1:03:07.560 a good test set that is representative of what 1:03:07.560 --> 1:03:09.560 like if you say like I want to play this 1:03:09.560 --> 1:03:11.560 that is songsticing in the car 1:03:11.560 --> 1:03:13.560 job of the product manager to go and source 1:03:13.560 --> 1:03:15.560 like a good test set of what that means 1:03:15.560 --> 1:03:17.560 then you can work with engineering to have algorithms 1:03:17.560 --> 1:03:19.560 to try to produce that right 1:03:19.560 --> 1:03:21.560 so we try to think a lot about 1:03:21.560 --> 1:03:23.560 how to structure product development 1:03:23.560 --> 1:03:25.560 for machine learning 1:03:25.560 --> 1:03:27.560 age and what we discovered 1:03:27.560 --> 1:03:29.560 was that a lot of it is actually in the expectation 1:03:29.560 --> 1:03:31.560 and you can go 1:03:31.560 --> 1:03:33.560 you can go two ways so 1:03:35.560 --> 1:03:37.560 let's say that 1:03:37.560 --> 1:03:39.560 if you set the expectation with the user 1:03:39.560 --> 1:03:41.560 that this is a discovery product like Discover Weekly 1:03:41.560 --> 1:03:43.560 you're actually setting 1:03:43.560 --> 1:03:45.560 the expectation that most of what we show you will not be 1:03:45.560 --> 1:03:47.560 relevant when you're in the discovery 1:03:47.560 --> 1:03:49.560 process you're going to accept that 1:03:49.560 --> 1:03:51.560 actually if you find one gem every Monday 1:03:51.560 --> 1:03:53.560 that you totally love 1:03:53.560 --> 1:03:55.560 you're probably going to be happy 1:03:55.560 --> 1:03:57.560 even though the statistical meaning 1:03:57.560 --> 1:03:59.560 1 out of 10 is terrible or 1 out of 20 1:03:59.560 --> 1:04:01.560 is terrible from a user point of view 1:04:01.560 --> 1:04:03.560 because the setting was discovered it's fine 1:04:03.560 --> 1:04:05.560 can I start to interrupt real quick 1:04:05.560 --> 1:04:07.560 I just actually learned about Discover Weekly 1:04:07.560 --> 1:04:09.560 which is a Spotify 1:04:09.560 --> 1:04:11.560 I don't know 1:04:11.560 --> 1:04:13.560 it's a feature of Spotify that shows you 1:04:13.560 --> 1:04:15.560 cool songs to listen 1:04:15.560 --> 1:04:17.560 maybe I can do 1:04:17.560 --> 1:04:19.560 issue tracking I couldn't find it on my Spotify app 1:04:19.560 --> 1:04:21.560 it's in your library 1:04:21.560 --> 1:04:23.560 it's in the library it's in the list of live 1:04:23.560 --> 1:04:25.560 because I was like whoa this is cool I didn't know this existed 1:04:25.560 --> 1:04:27.560 and I tried to find it 1:04:27.560 --> 1:04:29.560 but I would show it to you 1:04:29.560 --> 1:04:31.560 and feed it back to our product teams 1:04:31.560 --> 1:04:33.560 there you go 1:04:33.560 --> 1:04:35.560 so yeah sorry 1:04:35.560 --> 1:04:37.560 just to mention 1:04:37.560 --> 1:04:39.560 the expectation there is 1:04:39.560 --> 1:04:41.560 basically you're going to 1:04:41.560 --> 1:04:43.560 discover new songs 1:04:43.560 --> 1:04:45.560 so then you can be quite adventurous 1:04:45.560 --> 1:04:47.560 in the recommendations you do 1:04:47.560 --> 1:04:49.560 but we have 1:04:49.560 --> 1:04:51.560 another product called Daily Mix 1:04:51.560 --> 1:04:53.560 which kind of implies that these are only going to be your favorites 1:04:53.560 --> 1:04:55.560 so if you have 1:04:55.560 --> 1:04:57.560 9 out of 10 that is good 1:04:57.560 --> 1:04:59.560 and 9 out of 10 that doesn't work for you 1:04:59.560 --> 1:05:01.560 you're going to think it's a horrible product 1:05:01.560 --> 1:05:03.560 so actually a lot of the product development 1:05:03.560 --> 1:05:05.560 we learned over the years is about setting the right expectations 1:05:05.560 --> 1:05:07.560 so for Daily Mix 1:05:07.560 --> 1:05:09.560 you know algorithmically 1:05:09.560 --> 1:05:11.560 we would pick among things that feel very safe 1:05:11.560 --> 1:05:13.560 in your taste space 1:05:13.560 --> 1:05:15.560 or discover weekly we go kind of wild 1:05:15.560 --> 1:05:17.560 because the expectation is 1:05:17.560 --> 1:05:19.560 most of this is not going to 1:05:19.560 --> 1:05:21.560 so a lot of that a lot of to answer your question there 1:05:21.560 --> 1:05:23.560 a lot of 1:05:23.560 --> 1:05:25.560 we have some products where the whole point is 1:05:25.560 --> 1:05:27.560 that the user can click play put the phone in the pocket 1:05:27.560 --> 1:05:29.560 and it should be really good music for like an hour 1:05:29.560 --> 1:05:31.560 we have other products where 1:05:31.560 --> 1:05:33.560 you probably need to say like no 1:05:33.560 --> 1:05:35.560 save, no, no 1:05:35.560 --> 1:05:37.560 and it's very interactive 1:05:37.560 --> 1:05:39.560 that makes sense and then the radio product 1:05:39.560 --> 1:05:41.560 the stations product is one of these like click play 1:05:41.560 --> 1:05:43.560 put in your pocket for hours 1:05:43.560 --> 1:05:45.560 that's really interesting so you're thinking of 1:05:45.560 --> 1:05:47.560 different test sets 1:05:47.560 --> 1:05:49.560 for different users 1:05:49.560 --> 1:05:51.560 and trying to create products that sort of 1:05:51.560 --> 1:05:53.560 optimize 1:05:53.560 --> 1:05:55.560 optimize for those test sets 1:05:55.560 --> 1:05:57.560 that represents a specific set of 1:05:57.560 --> 1:05:59.560 users yes I think 1:05:59.560 --> 1:06:01.560 one thing that I think is interesting 1:06:01.560 --> 1:06:03.560 is 1:06:03.560 --> 1:06:05.560 we invested quite heavily in editorial 1:06:05.560 --> 1:06:07.560 in people creating playlists 1:06:07.560 --> 1:06:09.560 using statistical data 1:06:09.560 --> 1:06:11.560 and that was successful for us and then we also 1:06:11.560 --> 1:06:13.560 invested in machine learning 1:06:13.560 --> 1:06:15.560 and for the longest time 1:06:15.560 --> 1:06:17.560 within Spotify and within the rest of the 1:06:17.560 --> 1:06:19.560 industry there was always this narrative of 1:06:19.560 --> 1:06:21.560 humans versus the machine 1:06:21.560 --> 1:06:23.560 algo versus editorial 1:06:23.560 --> 1:06:25.560 and editors would say like well 1:06:25.560 --> 1:06:27.560 if I had that data if I could see your 1:06:27.560 --> 1:06:29.560 playlisting history and I made a choice 1:06:29.560 --> 1:06:31.560 for you I would have made a better choice 1:06:31.560 --> 1:06:33.560 and they would have because they're 1:06:33.560 --> 1:06:35.560 much smarter than these algorithms 1:06:35.560 --> 1:06:37.560 human is incredibly smart compared to our 1:06:37.560 --> 1:06:39.560 algorithms they can take 1:06:39.560 --> 1:06:41.560 culture into account and so forth 1:06:41.560 --> 1:06:43.560 the problem is that they can't make 200 million 1:06:43.560 --> 1:06:45.560 decisions 1:06:45.560 --> 1:06:47.560 per hour for every user that logs 1:06:47.560 --> 1:06:49.560 in so the algo may be 1:06:49.560 --> 1:06:51.560 not as sophisticated but much more efficient 1:06:51.560 --> 1:06:53.560 so there was this contradiction 1:06:53.560 --> 1:06:55.560 but then a few years ago 1:06:55.560 --> 1:06:57.560 we started 1:06:57.560 --> 1:06:59.560 focusing on this kind of human in the loop 1:06:59.560 --> 1:07:01.560 thinking around machine learning 1:07:01.560 --> 1:07:03.560 and we actually coined an internal 1:07:03.560 --> 1:07:05.560 term for it called algotorial 1:07:05.560 --> 1:07:07.560 the combination of algorithms and editors 1:07:07.560 --> 1:07:09.560 where if we take a concrete 1:07:09.560 --> 1:07:11.560 example you think 1:07:11.560 --> 1:07:13.560 of the editor this 1:07:13.560 --> 1:07:15.560 paid expert that we have 1:07:15.560 --> 1:07:17.560 there's really good at something like 1:07:17.560 --> 1:07:19.560 soul, hip hop 1:07:19.560 --> 1:07:21.560 EDM something right there are two experts 1:07:21.560 --> 1:07:23.560 no one in the industry 1:07:23.560 --> 1:07:25.560 so they have all the cultural knowledge 1:07:25.560 --> 1:07:27.560 you think of them as the product manager 1:07:27.560 --> 1:07:29.560 and you say that 1:07:29.560 --> 1:07:31.560 let's say that you want to create 1:07:31.560 --> 1:07:33.560 you think that there's a product 1:07:33.560 --> 1:07:35.560 need in the world for something like songs to sing 1:07:35.560 --> 1:07:37.560 in the car or songs to sing in the shower 1:07:37.560 --> 1:07:39.560 I'm taking that example because it exists 1:07:39.560 --> 1:07:41.560 people love to scream songs in the car 1:07:41.560 --> 1:07:43.560 when they drive right 1:07:43.560 --> 1:07:45.560 so you want to create that product and you have this product manager 1:07:45.560 --> 1:07:47.560 who's a musical expert 1:07:47.560 --> 1:07:49.560 they create, they come up with a concept 1:07:49.560 --> 1:07:51.560 like I think this is a missing thing in humanity 1:07:51.560 --> 1:07:53.560 like a playlist called songs to sing in the car 1:07:53.560 --> 1:07:55.560 they create 1:07:55.560 --> 1:07:57.560 the framing, the image 1:07:57.560 --> 1:07:59.560 the title and they create a test set 1:07:59.560 --> 1:08:01.560 they create a group of songs 1:08:01.560 --> 1:08:03.560 like a few thousand songs out of the catalog 1:08:03.560 --> 1:08:05.560 that they manually curate 1:08:05.560 --> 1:08:07.560 that are known songs that are great to sing in the car 1:08:07.560 --> 1:08:09.560 and they can take like 1:08:09.560 --> 1:08:11.560 through romance into account they understand things 1:08:11.560 --> 1:08:13.560 that algorithms do not at all 1:08:13.560 --> 1:08:15.560 so they have this huge set of tracks 1:08:15.560 --> 1:08:17.560 then when we deliver that to you 1:08:17.560 --> 1:08:19.560 we look at your taste vectors 1:08:19.560 --> 1:08:21.560 and you get the 20 tracks that are songs to sing in the car 1:08:21.560 --> 1:08:23.560 in your taste 1:08:23.560 --> 1:08:25.560 so you have personalization 1:08:25.560 --> 1:08:27.560 and editorial input 1:08:27.560 --> 1:08:29.560 in the same process 1:08:29.560 --> 1:08:31.560 if that makes sense 1:08:31.560 --> 1:08:33.560 it makes total sense and I have several questions around that 1:08:33.560 --> 1:08:35.560 this is like 1:08:35.560 --> 1:08:37.560 fascinating 1:08:37.560 --> 1:08:39.560 so first it is a little bit surprising 1:08:39.560 --> 1:08:41.560 to me 1:08:41.560 --> 1:08:43.560 that the world expert 1:08:43.560 --> 1:08:45.560 humans are outperforming 1:08:45.560 --> 1:08:47.560 machines 1:08:47.560 --> 1:08:49.560 at specifying 1:08:49.560 --> 1:08:51.560 songs to sing in the car 1:08:51.560 --> 1:08:53.560 so 1:08:53.560 --> 1:08:55.560 maybe you could talk to that a little bit 1:08:55.560 --> 1:08:57.560 I don't know if you can put it into words but 1:08:57.560 --> 1:08:59.560 what is it 1:08:59.560 --> 1:09:01.560 how difficult is this problem 1:09:01.560 --> 1:09:03.560 of 1:09:03.560 --> 1:09:05.560 I guess what I'm trying to ask 1:09:05.560 --> 1:09:07.560 is there how difficult is it to encode 1:09:07.560 --> 1:09:09.560 the cultural references 1:09:09.560 --> 1:09:11.560 the context 1:09:11.560 --> 1:09:13.560 of the song, the artists 1:09:13.560 --> 1:09:15.560 all those things together 1:09:15.560 --> 1:09:17.560 can machine learning really not do that 1:09:17.560 --> 1:09:19.560 I mean I think machine learning is great 1:09:19.560 --> 1:09:21.560 at replicating patterns 1:09:21.560 --> 1:09:23.560 if you have the patterns 1:09:23.560 --> 1:09:25.560 but if you try to write 1:09:25.560 --> 1:09:27.560 a spec of what song is great 1:09:27.560 --> 1:09:29.560 to sing in the car definition is 1:09:29.560 --> 1:09:31.560 is it loud 1:09:31.560 --> 1:09:33.560 does it have many choruses to have been in movies 1:09:33.560 --> 1:09:35.560 it quickly gets incredibly complicated 1:09:35.560 --> 1:09:37.560 right 1:09:37.560 --> 1:09:39.560 and a lot of it may not be 1:09:39.560 --> 1:09:41.560 in the structure of the song or the title 1:09:41.560 --> 1:09:43.560 it could be cultural references because 1:09:43.560 --> 1:09:45.560 it was a hasty one 1:09:45.560 --> 1:09:47.560 so the definition problems 1:09:47.560 --> 1:09:49.560 quickly get 1:09:49.560 --> 1:09:51.560 and I think that was the insight of Andrew Eng 1:09:51.560 --> 1:09:53.560 when he said that job of the product manager 1:09:53.560 --> 1:09:55.560 is to understand these things that 1:09:55.560 --> 1:09:57.560 algorithms don't and then 1:09:57.560 --> 1:09:59.560 define what that looks like 1:09:59.560 --> 1:10:01.560 and then you have something to train towards 1:10:01.560 --> 1:10:03.560 then you have kind of the test set 1:10:03.560 --> 1:10:05.560 but today the editors create this 1:10:05.560 --> 1:10:07.560 pool of tracks and then we personalize 1:10:07.560 --> 1:10:09.560 you could easily imagine that once you have this set 1:10:09.560 --> 1:10:11.560 you could have some automatic exploration 1:10:11.560 --> 1:10:13.560 of the rest of the catalog because then you understand 1:10:13.560 --> 1:10:15.560 what it is 1:10:15.560 --> 1:10:17.560 and then the other side of it when machine learning 1:10:17.560 --> 1:10:19.560 does help is this taste 1:10:19.560 --> 1:10:21.560 vector 1:10:21.560 --> 1:10:23.560 how hard is it to construct 1:10:23.560 --> 1:10:25.560 a vector that represents the things 1:10:25.560 --> 1:10:27.560 an individual human likes 1:10:27.560 --> 1:10:29.560 this human preference 1:10:29.560 --> 1:10:31.560 so you can 1:10:31.560 --> 1:10:33.560 you know music isn't like 1:10:33.560 --> 1:10:35.560 it's not like Amazon 1:10:35.560 --> 1:10:37.560 like things you usually buy 1:10:37.560 --> 1:10:39.560 music seems more amorphous 1:10:39.560 --> 1:10:41.560 like it's this thing 1:10:41.560 --> 1:10:43.560 that's hard to specify like 1:10:43.560 --> 1:10:45.560 what is 1:10:45.560 --> 1:10:47.560 if you look at my playlist what is the music 1:10:47.560 --> 1:10:49.560 that I love it's harder 1:10:49.560 --> 1:10:51.560 it seems to be 1:10:51.560 --> 1:10:53.560 much more difficult to specify concretely 1:10:53.560 --> 1:10:55.560 so how hard is it 1:10:55.560 --> 1:10:57.560 to build a taste vector 1:10:57.560 --> 1:10:59.560 it is very hard in the sense that you need a lot of 1:10:59.560 --> 1:11:01.560 data 1:11:01.560 --> 1:11:03.560 and I think what we found was that 1:11:03.560 --> 1:11:05.560 so it's not a stationary problem 1:11:05.560 --> 1:11:07.560 it changes over time 1:11:07.560 --> 1:11:09.560 and so 1:11:09.560 --> 1:11:11.560 we've gone 1:11:11.560 --> 1:11:13.560 through the journey of if 1:11:13.560 --> 1:11:15.560 you've done a lot of computer vision 1:11:15.560 --> 1:11:17.560 obviously I've done a bunch of computer vision 1:11:17.560 --> 1:11:19.560 in my past and we started 1:11:19.560 --> 1:11:21.560 kind of with the handcrafted heuristics 1:11:21.560 --> 1:11:23.560 for you know 1:11:23.560 --> 1:11:25.560 this is kind of in the music this is this 1:11:25.560 --> 1:11:27.560 and if you consume this you probably like this 1:11:27.560 --> 1:11:29.560 so we have 1:11:29.560 --> 1:11:31.560 we started there and we have some of that still 1:11:31.560 --> 1:11:33.560 then what was interesting about the playlist data 1:11:33.560 --> 1:11:35.560 was that you could find these latent things that 1:11:35.560 --> 1:11:37.560 wouldn't necessarily even make sense to you 1:11:37.560 --> 1:11:39.560 that could 1:11:39.560 --> 1:11:41.560 could even capture maybe cultural references 1:11:41.560 --> 1:11:43.560 because they co occurred things that 1:11:43.560 --> 1:11:45.560 that wouldn't have appeared 1:11:45.560 --> 1:11:47.560 kind of mechanistically either in the content 1:11:47.560 --> 1:11:49.560 or so forth so 1:11:49.560 --> 1:11:51.560 um 1:11:51.560 --> 1:11:53.560 I think that 1:11:53.560 --> 1:11:55.560 um 1:11:55.560 --> 1:11:57.560 I think the core assumption 1:11:57.560 --> 1:11:59.560 is that 1:11:59.560 --> 1:12:01.560 there are patterns 1:12:01.560 --> 1:12:03.560 in almost everything 1:12:03.560 --> 1:12:05.560 and if there are patterns 1:12:05.560 --> 1:12:07.560 these embedding techniques are getting better and better 1:12:07.560 --> 1:12:09.560 now as everyone else 1:12:09.560 --> 1:12:11.560 we're also using 1:12:11.560 --> 1:12:13.560 kind of deep embeddings where you can encode binary 1:12:13.560 --> 1:12:15.560 values and so forth 1:12:15.560 --> 1:12:17.560 and what I think 1:12:17.560 --> 1:12:19.560 is interesting is this process 1:12:19.560 --> 1:12:21.560 to try to find things that 1:12:21.560 --> 1:12:23.560 um 1:12:23.560 --> 1:12:25.560 necessarily you wouldn't actually have 1:12:25.560 --> 1:12:27.560 guest so it is very 1:12:27.560 --> 1:12:29.560 hard in an engineering 1:12:29.560 --> 1:12:31.560 sense to find the right dimensions 1:12:31.560 --> 1:12:33.560 it's an incredible scalability 1:12:33.560 --> 1:12:35.560 problem to do for hundreds of millions 1:12:35.560 --> 1:12:37.560 of users and to update it every day 1:12:37.560 --> 1:12:39.560 but in 1:12:39.560 --> 1:12:41.560 theory 1:12:41.560 --> 1:12:43.560 in theory embeddings isn't that 1:12:43.560 --> 1:12:45.560 complicated the fact 1:12:45.560 --> 1:12:47.560 that you try to find some principle components 1:12:47.560 --> 1:12:49.560 or something like that dimensionality reduction 1:12:49.560 --> 1:12:51.560 and so forth so the theory I guess is easy 1:12:51.560 --> 1:12:53.560 because it's very very hard 1:12:53.560 --> 1:12:55.560 and it's a huge 1:12:55.560 --> 1:12:57.560 engineering challenge but fortunately we have 1:12:57.560 --> 1:12:59.560 some amazing both research and 1:12:59.560 --> 1:13:01.560 engineering teams in this space 1:13:01.560 --> 1:13:03.560 yeah I guess the 1:13:03.560 --> 1:13:05.560 question is all 1:13:05.560 --> 1:13:07.560 I mean it's similar I deal with the autonomous 1:13:07.560 --> 1:13:09.560 vehicle space is the question is 1:13:09.560 --> 1:13:11.560 how hard is driving 1:13:11.560 --> 1:13:13.560 and here is 1:13:13.560 --> 1:13:15.560 basically the question is of 1:13:15.560 --> 1:13:17.560 edge cases 1:13:17.560 --> 1:13:19.560 so embedding 1:13:19.560 --> 1:13:21.560 probably works 1:13:21.560 --> 1:13:23.560 not probably but 1:13:23.560 --> 1:13:25.560 I would imagine works well 1:13:25.560 --> 1:13:27.560 in a lot of cases 1:13:27.560 --> 1:13:29.560 so there's a bunch of questions that arise then 1:13:29.560 --> 1:13:31.560 so do song 1:13:31.560 --> 1:13:33.560 preferences does your taste vector 1:13:33.560 --> 1:13:35.560 depend on context 1:13:35.560 --> 1:13:37.560 like mood 1:13:37.560 --> 1:13:39.560 right so there's 1:13:39.560 --> 1:13:41.560 different moods and 1:13:41.560 --> 1:13:43.560 absolutely so how does that 1:13:43.560 --> 1:13:45.560 take in 1:13:45.560 --> 1:13:47.560 is it is it possible to take that 1:13:47.560 --> 1:13:49.560 consideration or do you just leave 1:13:49.560 --> 1:13:51.560 that as a interface 1:13:51.560 --> 1:13:53.560 problem that allows the user to just control 1:13:53.560 --> 1:13:55.560 it so when I'm looking for a work 1:13:55.560 --> 1:13:57.560 out music I kind of specify it 1:13:57.560 --> 1:13:59.560 by choosing certain playlists 1:13:59.560 --> 1:14:01.560 doing certain search yeah 1:14:01.560 --> 1:14:03.560 so that's a great point and back to the product 1:14:03.560 --> 1:14:05.560 development you could try 1:14:05.560 --> 1:14:07.560 to spend a few years trying to predict 1:14:07.560 --> 1:14:09.560 which mood you're in automatically when you open 1:14:09.560 --> 1:14:11.560 Spotify or you create a tab 1:14:11.560 --> 1:14:13.560 which is happy and sad right and you're going to be 1:14:13.560 --> 1:14:15.560 right 100% of the time with one click 1:14:15.560 --> 1:14:17.560 now probably much better to let 1:14:17.560 --> 1:14:19.560 the user tell you if they're happy or sad 1:14:19.560 --> 1:14:21.560 or if they want to work out on the other hand 1:14:21.560 --> 1:14:23.560 if your user interface become 2000 1:14:23.560 --> 1:14:25.560 tabs you're introducing so much friction 1:14:25.560 --> 1:14:27.560 so no one will use the product so then you have to 1:14:27.560 --> 1:14:29.560 get better so it's this 1:14:29.560 --> 1:14:31.560 thing where I think it maybe was 1:14:31.560 --> 1:14:33.560 I remember who coined it but it's 1:14:33.560 --> 1:14:35.560 called full tolerant UIs right to build a UI 1:14:35.560 --> 1:14:37.560 that is tolerant to being wrong 1:14:37.560 --> 1:14:39.560 and then you can be much 1:14:39.560 --> 1:14:41.560 less right in your 1:14:41.560 --> 1:14:43.560 in your algorithms so 1:14:43.560 --> 1:14:45.560 we had to learn a lot of that 1:14:45.560 --> 1:14:47.560 building the right UI that fits 1:14:47.560 --> 1:14:49.560 where the machine learning is 1:14:49.560 --> 1:14:51.560 and a great 1:14:51.560 --> 1:14:53.560 discovery there which was 1:14:53.560 --> 1:14:55.560 by the teams during 1:14:55.560 --> 1:14:57.560 one of our hack days was this 1:14:57.560 --> 1:14:59.560 thing of taking discovery packaging 1:14:59.560 --> 1:15:01.560 into a playlist and saying that 1:15:01.560 --> 1:15:03.560 these are new tracks 1:15:03.560 --> 1:15:05.560 that we think you might like 1:15:05.560 --> 1:15:07.560 based on this and setting the right expectation 1:15:07.560 --> 1:15:09.560 made it a great product 1:15:09.560 --> 1:15:11.560 so I think we have this benefit that for example 1:15:11.560 --> 1:15:13.560 Tesla doesn't have 1:15:13.560 --> 1:15:15.560 that we can 1:15:15.560 --> 1:15:17.560 we can change the expectation we can build 1:15:17.560 --> 1:15:19.560 a full tolerant setting it's very hard to be 1:15:19.560 --> 1:15:21.560 full tolerant when you're driving at 1:15:21.560 --> 1:15:23.560 100 miles per hour or something 1:15:23.560 --> 1:15:25.560 and we have the luxury of 1:15:25.560 --> 1:15:27.560 being able to say that 1:15:27.560 --> 1:15:29.560 of being wrong if we have the right 1:15:29.560 --> 1:15:31.560 UI which gives us 1:15:31.560 --> 1:15:33.560 different abilities to take more risk 1:15:33.560 --> 1:15:35.560 so I actually think the self driving 1:15:35.560 --> 1:15:37.560 problem is much harder 1:15:37.560 --> 1:15:39.560 oh yeah for sure 1:15:39.560 --> 1:15:41.560 it's much less fun 1:15:41.560 --> 1:15:43.560 because 1:15:43.560 --> 1:15:45.560 people die exactly 1:15:45.560 --> 1:15:47.560 and in Spotify 1:15:47.560 --> 1:15:49.560 it's 1:15:49.560 --> 1:15:51.560 such a more fun problem because 1:15:51.560 --> 1:15:53.560 failure will 1:15:53.560 --> 1:15:55.560 I mean failure is beautiful in a way 1:15:55.560 --> 1:15:57.560 at least exploration so it's a really fun 1:15:57.560 --> 1:15:59.560 reinforcement learning problem and the worst 1:15:59.560 --> 1:16:01.560 case scenario is you get these WTF tweets 1:16:01.560 --> 1:16:03.560 like how did I get this 1:16:03.560 --> 1:16:05.560 which is a lot better than the self 1:16:05.560 --> 1:16:07.560 driving failure 1:16:07.560 --> 1:16:09.560 so 1:16:09.560 --> 1:16:11.560 what's the feedback that a user 1:16:11.560 --> 1:16:13.560 what's the signal 1:16:13.560 --> 1:16:15.560 that a user provides 1:16:15.560 --> 1:16:17.560 into the system so 1:16:17.560 --> 1:16:19.560 you mentioned skipping 1:16:19.560 --> 1:16:21.560 what is like the strongest signal 1:16:21.560 --> 1:16:23.560 you didn't mention clicking 1:16:23.560 --> 1:16:25.560 like 1:16:25.560 --> 1:16:27.560 so we have a few signals that are important 1:16:27.560 --> 1:16:29.560 obviously playing 1:16:29.560 --> 1:16:31.560 playing through 1:16:31.560 --> 1:16:33.560 so one of the benefits of music actually 1:16:33.560 --> 1:16:35.560 even compared to podcasts or 1:16:35.560 --> 1:16:37.560 movies is the object itself 1:16:37.560 --> 1:16:39.560 is really only about three minutes 1:16:39.560 --> 1:16:41.560 so you get a lot of chances to recommend 1:16:41.560 --> 1:16:43.560 and the feedback loop is 1:16:43.560 --> 1:16:45.560 every three minutes instead of every two hours 1:16:45.560 --> 1:16:47.560 or something so you actually get 1:16:47.560 --> 1:16:49.560 kind of noisy but 1:16:49.560 --> 1:16:51.560 quite fast feedback and so you can see 1:16:51.560 --> 1:16:53.560 if people played through or if the 1:16:53.560 --> 1:16:55.560 which is the inverse of skip really 1:16:55.560 --> 1:16:57.560 that's an important signal on the other hand 1:16:57.560 --> 1:16:59.560 much of the consumption happens 1:16:59.560 --> 1:17:01.560 when your phone is in your pocket maybe you're running 1:17:01.560 --> 1:17:03.560 or driving or you're playing on a speaker 1:17:03.560 --> 1:17:05.560 and you're not skipping doesn't mean that you love that song 1:17:05.560 --> 1:17:07.560 it might be that it wasn't bad enough 1:17:07.560 --> 1:17:09.560 that you would walk up and skip 1:17:09.560 --> 1:17:11.560 so it's a noisy signal 1:17:11.560 --> 1:17:13.560 then we have the equivalent of the like 1:17:13.560 --> 1:17:15.560 which is you save it to your library 1:17:15.560 --> 1:17:17.560 that's a pretty strong signal of affection 1:17:17.560 --> 1:17:19.560 and then 1:17:19.560 --> 1:17:21.560 we have the more explicit signal 1:17:21.560 --> 1:17:23.560 of play listing like you took the time 1:17:23.560 --> 1:17:25.560 to create a playlist you put it in there 1:17:25.560 --> 1:17:27.560 there's a very little small chance that 1:17:27.560 --> 1:17:29.560 if you took all that trouble this is not 1:17:29.560 --> 1:17:31.560 a really important track to you 1:17:31.560 --> 1:17:33.560 but also 1:17:33.560 --> 1:17:35.560 what are the tracks it relates to so we have 1:17:35.560 --> 1:17:37.560 we have the play listing we have the like and then we have 1:17:37.560 --> 1:17:39.560 the listening or skip 1:17:39.560 --> 1:17:41.560 and you have to 1:17:41.560 --> 1:17:43.560 have very different approaches to all of them because 1:17:43.560 --> 1:17:45.560 at different levels of noise 1:17:45.560 --> 1:17:47.560 one is very voluminous but noisy 1:17:47.560 --> 1:17:49.560 and the other is rare but 1:17:49.560 --> 1:17:51.560 you can probably trust it 1:17:51.560 --> 1:17:53.560 yeah it's interesting because 1:17:53.560 --> 1:17:55.560 I think between those signals captures 1:17:55.560 --> 1:17:57.560 all the information you'd want to capture 1:17:57.560 --> 1:17:59.560 I mean there's a feeling 1:17:59.560 --> 1:18:01.560 for me that there's sometimes that I'll 1:18:01.560 --> 1:18:03.560 hear a song that's like yes this is 1:18:03.560 --> 1:18:05.560 you know this was the right song for the moment 1:18:05.560 --> 1:18:07.560 but there's really no way to express 1:18:07.560 --> 1:18:09.560 that fact except by 1:18:09.560 --> 1:18:11.560 listening through it all the way 1:18:11.560 --> 1:18:13.560 and maybe playing it again at that time 1:18:13.560 --> 1:18:15.560 or something but there's 1:18:15.560 --> 1:18:17.560 no need for a button that says 1:18:17.560 --> 1:18:19.560 this was the best song could have 1:18:19.560 --> 1:18:21.560 heard at this moment well we're playing around 1:18:21.560 --> 1:18:23.560 with that with kind of the thumbs up 1:18:23.560 --> 1:18:25.560 concept saying like I really like this 1:18:25.560 --> 1:18:27.560 just kind of talking to the algorithm 1:18:27.560 --> 1:18:29.560 it's unclear if that's the best way 1:18:29.560 --> 1:18:31.560 for humans to interact maybe it is 1:18:31.560 --> 1:18:33.560 maybe they should think of Spotify 1:18:33.560 --> 1:18:35.560 as a person an agent sitting there 1:18:35.560 --> 1:18:37.560 trying to serve you and you can say like 1:18:37.560 --> 1:18:39.560 that's Spotify, good Spotify 1:18:39.560 --> 1:18:41.560 right now the analogy we've had is more 1:18:41.560 --> 1:18:43.560 you shouldn't think of us 1:18:43.560 --> 1:18:45.560 we should be investable and the feedback 1:18:45.560 --> 1:18:47.560 is if you save it 1:18:47.560 --> 1:18:49.560 you work for yourself you do a playlist 1:18:49.560 --> 1:18:51.560 because you think is great and we can learn from that 1:18:51.560 --> 1:18:53.560 it's kind of back to Tesla 1:18:53.560 --> 1:18:55.560 how they kind of have this shadow mode 1:18:55.560 --> 1:18:57.560 we sit in what you drive 1:18:57.560 --> 1:18:59.560 we kind of took the same analogy 1:18:59.560 --> 1:19:01.560 we sit in what you playlist 1:19:01.560 --> 1:19:03.560 and then maybe we can offer you an autopilot 1:19:03.560 --> 1:19:05.560 we can take over for a while or something like that 1:19:05.560 --> 1:19:07.560 and then back off if you say like 1:19:07.560 --> 1:19:09.560 that's not good enough 1:19:09.560 --> 1:19:11.560 but I think it's interesting to figure out 1:19:11.560 --> 1:19:13.560 what your mental model is 1:19:13.560 --> 1:19:15.560 if Spotify is an AI that you talk to 1:19:15.560 --> 1:19:17.560 which I think might be a bit too abstract 1:19:17.560 --> 1:19:19.560 for many consumers 1:19:19.560 --> 1:19:21.560 or if you still think of it as 1:19:21.560 --> 1:19:23.560 it's my music app 1:19:23.560 --> 1:19:25.560 it's more helpful 1:19:25.560 --> 1:19:27.560 and depends on the device it's running on 1:19:27.560 --> 1:19:29.560 which brings us to 1:19:29.560 --> 1:19:31.560 smart speakers 1:19:31.560 --> 1:19:33.560 so I have a lot of the Spotify listening I do 1:19:33.560 --> 1:19:35.560 is on 1:19:35.560 --> 1:19:37.560 things on devices I can talk to 1:19:37.560 --> 1:19:39.560 whether it's from Amazon, Google 1:19:39.560 --> 1:19:41.560 or Apple 1:19:41.560 --> 1:19:43.560 what's the role of Spotify in those devices 1:19:43.560 --> 1:19:45.560 how do you think of it differently than 1:19:45.560 --> 1:19:47.560 on the phone or on the desktop 1:19:47.560 --> 1:19:49.560 there are a few things 1:19:49.560 --> 1:19:51.560 to say about the first of all 1:19:51.560 --> 1:19:53.560 it's incredibly exciting they're growing like crazy 1:19:53.560 --> 1:19:55.560 especially here in the US 1:19:59.560 --> 1:20:01.560 and it's solving 1:20:01.560 --> 1:20:03.560 a consumer need that I think is 1:20:05.560 --> 1:20:07.560 you can think of it as 1:20:07.560 --> 1:20:09.560 just remote interactivity 1:20:09.560 --> 1:20:11.560 you can control this thing from 1:20:11.560 --> 1:20:13.560 across the room and it may 1:20:13.560 --> 1:20:15.560 feel like a small thing but it turns out 1:20:15.560 --> 1:20:17.560 that friction matters to consumers 1:20:17.560 --> 1:20:19.560 being able to say play, pause 1:20:19.560 --> 1:20:21.560 and so forth from across the room 1:20:21.560 --> 1:20:23.560 is very powerful so basically 1:20:23.560 --> 1:20:25.560 you made the living room interactive 1:20:25.560 --> 1:20:27.560 now 1:20:27.560 --> 1:20:29.560 and 1:20:29.560 --> 1:20:31.560 what we see in our data is that 1:20:31.560 --> 1:20:33.560 the number one use case for these speakers 1:20:33.560 --> 1:20:35.560 is music, music and podcast 1:20:35.560 --> 1:20:37.560 so 1:20:37.560 --> 1:20:39.560 fortunately for us it's been important 1:20:39.560 --> 1:20:41.560 to these companies to have those use case 1:20:41.560 --> 1:20:43.560 covered so they want to Spotify on this 1:20:43.560 --> 1:20:45.560 we have very good relationships with 1:20:45.560 --> 1:20:47.560 them 1:20:47.560 --> 1:20:49.560 and we're seeing tremendous 1:20:49.560 --> 1:20:51.560 success with them 1:20:51.560 --> 1:20:53.560 what I think is interesting 1:20:53.560 --> 1:20:55.560 about them is 1:20:55.560 --> 1:20:57.560 it's already working 1:20:57.560 --> 1:20:59.560 we kind of had this 1:20:59.560 --> 1:21:01.560 epiphany 1:21:01.560 --> 1:21:03.560 many years ago back when we started 1:21:03.560 --> 1:21:05.560 using Sonos if you went through all the 1:21:05.560 --> 1:21:07.560 trouble of setting up your Sonos system 1:21:07.560 --> 1:21:09.560 you had this magical experience where you had 1:21:09.560 --> 1:21:11.560 all the music ever made in your living room 1:21:11.560 --> 1:21:13.560 and we 1:21:13.560 --> 1:21:15.560 made this assumption that 1:21:15.560 --> 1:21:17.560 at home everyone used to have a CD player at home 1:21:17.560 --> 1:21:19.560 but they never managed to get their files 1:21:19.560 --> 1:21:21.560 working in the home having this network attached 1:21:21.560 --> 1:21:23.560 storage was too cumbersome for most consumers 1:21:23.560 --> 1:21:25.560 so we made the assumption that 1:21:25.560 --> 1:21:27.560 the home would skip from the CD all the way to 1:21:27.560 --> 1:21:29.560 streaming box 1:21:29.560 --> 1:21:31.560 where you would buy the steering wheel without 1:21:31.560 --> 1:21:33.560 all the music built in that took longer than we 1:21:33.560 --> 1:21:35.560 thought but with the voice speakers that was the 1:21:35.560 --> 1:21:37.560 unlocking that made kind of the connected speaker 1:21:37.560 --> 1:21:39.560 happen in the home 1:21:39.560 --> 1:21:41.560 so it really 1:21:41.560 --> 1:21:43.560 it really exploded and 1:21:43.560 --> 1:21:45.560 we saw this engagement that we 1:21:45.560 --> 1:21:47.560 predicted would happen 1:21:47.560 --> 1:21:49.560 what I think is interesting though is where it's going 1:21:49.560 --> 1:21:51.560 from now 1:21:51.560 --> 1:21:53.560 right now you think of them as voice speakers 1:21:53.560 --> 1:21:55.560 but I think if you look at 1:21:55.560 --> 1:21:57.560 Google I.O. for example 1:21:57.560 --> 1:21:59.560 they just added a camera 1:21:59.560 --> 1:22:01.560 to it where when the alarm goes off 1:22:01.560 --> 1:22:03.560 instead of saying 1:22:03.560 --> 1:22:05.560 hey Google stop you can just 1:22:05.560 --> 1:22:07.560 wave your hand 1:22:07.560 --> 1:22:09.560 so I think they're going to think more of it as 1:22:09.560 --> 1:22:11.560 an agent 1:22:11.560 --> 1:22:13.560 as an assistant truly an assistant 1:22:13.560 --> 1:22:15.560 and an assistant that can see you 1:22:15.560 --> 1:22:17.560 is going to be much more effective than a blind 1:22:17.560 --> 1:22:19.560 assistant so I think these things will 1:22:19.560 --> 1:22:21.560 morph and we won't necessarily think of them as 1:22:21.560 --> 1:22:23.560 quote unquote voice speakers anymore 1:22:23.560 --> 1:22:25.560 just as 1:22:25.560 --> 1:22:27.560 interactive 1:22:27.560 --> 1:22:29.560 access to the internet in the home 1:22:29.560 --> 1:22:31.560 but I still think that 1:22:31.560 --> 1:22:33.560 the biggest use case for those will be 1:22:33.560 --> 1:22:35.560 will be audio so 1:22:35.560 --> 1:22:37.560 for that reason we're investing heavily in it 1:22:37.560 --> 1:22:39.560 and we built our own NLU stack 1:22:39.560 --> 1:22:41.560 to be able to 1:22:41.560 --> 1:22:43.560 the challenge here is 1:22:43.560 --> 1:22:45.560 how do you innovate in that world 1:22:45.560 --> 1:22:47.560 it lowers friction for consumers 1:22:47.560 --> 1:22:49.560 but it's also much more constrained 1:22:49.560 --> 1:22:51.560 you have no pixels to play with 1:22:51.560 --> 1:22:53.560 in an audio only world 1:22:53.560 --> 1:22:55.560 it's really the vocabulary that is the interface 1:22:55.560 --> 1:22:57.560 so we started investing 1:22:57.560 --> 1:22:59.560 and playing around quite a lot with that 1:22:59.560 --> 1:23:01.560 trying to understand what the future will be 1:23:01.560 --> 1:23:03.560 of you speaking and gesturing 1:23:03.560 --> 1:23:05.560 and waving at your music 1:23:05.560 --> 1:23:07.560 and actually you're actually nudging closer 1:23:07.560 --> 1:23:09.560 to autonomous vehicle space 1:23:09.560 --> 1:23:11.560 because from everything I've seen 1:23:11.560 --> 1:23:13.560 the level of frustration people experience 1:23:13.560 --> 1:23:15.560 upon failure of natural language 1:23:15.560 --> 1:23:17.560 understanding is much higher 1:23:17.560 --> 1:23:19.560 than failure in other context 1:23:19.560 --> 1:23:21.560 people get frustrated really fast 1:23:21.560 --> 1:23:23.560 so if you screw that 1:23:23.560 --> 1:23:25.560 experience up even just a little bit 1:23:25.560 --> 1:23:27.560 they give up really quickly 1:23:27.560 --> 1:23:29.560 and I think you see that in the data 1:23:29.560 --> 1:23:31.560 while it's tremendously successful 1:23:31.560 --> 1:23:33.560 the most common interactions 1:23:33.560 --> 1:23:35.560 are play, pause 1:23:35.560 --> 1:23:37.560 and you know next 1:23:37.560 --> 1:23:39.560 the things where if you compare it to taking up your phone 1:23:39.560 --> 1:23:41.560 unlocking it, bringing up the app and skipping 1:23:41.560 --> 1:23:43.560 clicking skip 1:23:43.560 --> 1:23:45.560 it was much lower friction 1:23:45.560 --> 1:23:47.560 but then 1:23:47.560 --> 1:23:49.560 for longer more complicated things 1:23:49.560 --> 1:23:51.560 can you find me that song about people still bring up their phone 1:23:51.560 --> 1:23:53.560 and search and then play it on their speaker 1:23:53.560 --> 1:23:55.560 so we tried again to build a fault tolerant UI 1:23:55.560 --> 1:23:57.560 where for the more 1:23:57.560 --> 1:23:59.560 complicated things you can still pick up your phone 1:23:59.560 --> 1:24:01.560 have powerful full keyboard search 1:24:01.560 --> 1:24:03.560 and then try to optimize 1:24:03.560 --> 1:24:05.560 for where there is actually lower friction 1:24:05.560 --> 1:24:07.560 and try to, it's kind of like the 1:24:07.560 --> 1:24:09.560 test autopilot thing 1:24:09.560 --> 1:24:11.560 you have to be at the level where 1:24:11.560 --> 1:24:13.560 you're helpful if you're too smart 1:24:13.560 --> 1:24:15.560 and just in the way people are going to get frustrated 1:24:15.560 --> 1:24:17.560 and first of all 1:24:17.560 --> 1:24:19.560 I'm not obsessed with where it happens 1:24:19.560 --> 1:24:21.560 it's just a good song but let me mention that as a use case 1:24:21.560 --> 1:24:23.560 because it's an interesting one 1:24:23.560 --> 1:24:25.560 I've literally told 1:24:25.560 --> 1:24:27.560 I don't want to say the name of the speaker 1:24:27.560 --> 1:24:29.560 because when people are listening to it 1:24:29.560 --> 1:24:31.560 it'll make their speaker go off 1:24:31.560 --> 1:24:33.560 but I talk to the speaker 1:24:33.560 --> 1:24:35.560 and I say play 1:24:35.560 --> 1:24:37.560 Stairway to Heaven 1:24:37.560 --> 1:24:39.560 and every time 1:24:39.560 --> 1:24:41.560 but a large percentage of the time plays the wrong 1:24:41.560 --> 1:24:43.560 Stairway to Heaven 1:24:43.560 --> 1:24:45.560 it plays some cover of the 1:24:45.560 --> 1:24:47.560 and 1:24:47.560 --> 1:24:49.560 that part of the experience 1:24:49.560 --> 1:24:51.560 I actually wonder from a business perspective 1:24:51.560 --> 1:24:53.560 the Spotify control 1:24:53.560 --> 1:24:55.560 that entire experience 1:24:55.560 --> 1:24:57.560 or no 1:24:57.560 --> 1:24:59.560 it seems like the NLU 1:24:59.560 --> 1:25:01.560 stuff is controlled by the speaker 1:25:01.560 --> 1:25:03.560 and then Spotify stays 1:25:03.560 --> 1:25:05.560 at a layer below that 1:25:05.560 --> 1:25:07.560 it's a good and complicated question 1:25:07.560 --> 1:25:09.560 some of which is dependent 1:25:09.560 --> 1:25:11.560 on the 1:25:11.560 --> 1:25:13.560 partners so it's hard to comment on the specifics 1:25:13.560 --> 1:25:15.560 but 1:25:15.560 --> 1:25:17.560 the question is the right one 1:25:17.560 --> 1:25:19.560 the challenge is 1:25:19.560 --> 1:25:21.560 if you can't use any of the personalization 1:25:21.560 --> 1:25:23.560 we know which Stairway to Heaven 1:25:23.560 --> 1:25:25.560 and the truth is maybe for one person 1:25:25.560 --> 1:25:27.560 it is exactly the cover that they want 1:25:27.560 --> 1:25:29.560 to be very frustrated 1:25:29.560 --> 1:25:31.560 I think 1:25:31.560 --> 1:25:33.560 we default to the right version 1:25:33.560 --> 1:25:35.560 but you actually want to be able to do the cover 1:25:35.560 --> 1:25:37.560 for the person that just played the cover 50 times 1:25:37.560 --> 1:25:39.560 or Spotify is just going to seem stupid 1:25:39.560 --> 1:25:41.560 so you want to be able to leverage the personalization 1:25:41.560 --> 1:25:43.560 but you have this stack 1:25:43.560 --> 1:25:45.560 where you have the 1:25:45.560 --> 1:25:47.560 the ASR and this thing called the end best list 1:25:47.560 --> 1:25:49.560 or the end best guesses 1:25:49.560 --> 1:25:51.560 here and then the personalization comes in at the end 1:25:51.560 --> 1:25:53.560 you actually want the personalization to be here 1:25:53.560 --> 1:25:55.560 when you're guessing about what they actually meant 1:25:55.560 --> 1:25:57.560 we're working with these partners 1:25:57.560 --> 1:25:59.560 and it's a complicated 1:25:59.560 --> 1:26:01.560 it's a complicated thing 1:26:01.560 --> 1:26:03.560 where you want to 1:26:03.560 --> 1:26:05.560 so first of all you want to be very 1:26:05.560 --> 1:26:07.560 careful with your users data 1:26:07.560 --> 1:26:09.560 you don't want to share your users data without the permission 1:26:09.560 --> 1:26:11.560 but you want to share some data so that their experience gets better 1:26:11.560 --> 1:26:13.560 so that these partners 1:26:13.560 --> 1:26:15.560 can understand enough but not too much and so forth 1:26:15.560 --> 1:26:17.560 so 1:26:17.560 --> 1:26:19.560 it's really the trick is that 1:26:19.560 --> 1:26:21.560 it's like a business driven relationship 1:26:21.560 --> 1:26:23.560 where you're doing product development across companies 1:26:23.560 --> 1:26:25.560 together 1:26:25.560 --> 1:26:27.560 which is really complicated 1:26:27.560 --> 1:26:29.560 but this is exactly why we built our own NLU 1:26:29.560 --> 1:26:31.560 so that we actually 1:26:31.560 --> 1:26:33.560 can make personalized guesses 1:26:33.560 --> 1:26:35.560 because this is the biggest frustration 1:26:35.560 --> 1:26:37.560 from a user point of view they don't understand 1:26:37.560 --> 1:26:39.560 about ASRs and end best lists 1:26:39.560 --> 1:26:41.560 and business deals they're like how hard can it be 1:26:41.560 --> 1:26:43.560 I've told this thing 50 times 1:26:43.560 --> 1:26:45.560 this version and still it plays the wrong thing 1:26:45.560 --> 1:26:47.560 it can't be hard 1:26:47.560 --> 1:26:49.560 so we try to take that user approach 1:26:49.560 --> 1:26:51.560 if the user is not going to understand 1:26:51.560 --> 1:26:53.560 the implications of business we have to 1:26:53.560 --> 1:26:55.560 solve it 1:26:55.560 --> 1:26:57.560 let's talk about sort of a complicated subject 1:26:57.560 --> 1:26:59.560 that I myself 1:26:59.560 --> 1:27:01.560 I'm quite 1:27:01.560 --> 1:27:03.560 torn about 1:27:03.560 --> 1:27:05.560 the idea sort of of 1:27:05.560 --> 1:27:07.560 paying 1:27:07.560 --> 1:27:09.560 artists 1:27:09.560 --> 1:27:11.560 I saw as of August 31 1:27:11.560 --> 1:27:13.560 2018 1:27:13.560 --> 1:27:15.560 over 11 billion dollars 1:27:15.560 --> 1:27:17.560 were paid to rights holders 1:27:17.560 --> 1:27:19.560 and further distributed 1:27:19.560 --> 1:27:21.560 to artists from Spotify 1:27:21.560 --> 1:27:23.560 so a lot of money is being paid to artists 1:27:23.560 --> 1:27:25.560 first of all 1:27:25.560 --> 1:27:27.560 the whole time as a consumer for me 1:27:27.560 --> 1:27:29.560 when I look at Spotify 1:27:29.560 --> 1:27:31.560 I'm not sure I'm remembering 1:27:31.560 --> 1:27:33.560 correctly but I think you said exactly how I feel 1:27:33.560 --> 1:27:35.560 which is this is too good to be true 1:27:35.560 --> 1:27:37.560 like 1:27:37.560 --> 1:27:39.560 when I started using Spotify 1:27:39.560 --> 1:27:41.560 I assumed you guys would go bankrupt 1:27:41.560 --> 1:27:43.560 in like a month 1:27:43.560 --> 1:27:45.560 it's like this is too good 1:27:45.560 --> 1:27:47.560 a lot of people did 1:27:47.560 --> 1:27:49.560 this is amazing 1:27:49.560 --> 1:27:51.560 so one question I have 1:27:51.560 --> 1:27:53.560 is sort of the bigger question 1:27:53.560 --> 1:27:55.560 how do you make money in this complicated world 1:27:55.560 --> 1:27:57.560 how do you deal with the relationship 1:27:57.560 --> 1:27:59.560 with record labels 1:27:59.560 --> 1:28:01.560 who 1:28:01.560 --> 1:28:03.560 are 1:28:03.560 --> 1:28:05.560 complicated 1:28:05.560 --> 1:28:07.560 these big 1:28:07.560 --> 1:28:09.560 you essentially have the task 1:28:09.560 --> 1:28:11.560 of hurting 1:28:11.560 --> 1:28:13.560 cats but like rich 1:28:13.560 --> 1:28:15.560 and powerful cats 1:28:15.560 --> 1:28:17.560 and also 1:28:17.560 --> 1:28:19.560 have the task of paying artists enough 1:28:19.560 --> 1:28:21.560 and paying those labels enough 1:28:21.560 --> 1:28:23.560 and still making money in the internet 1:28:23.560 --> 1:28:25.560 space where people are not willing to pay 1:28:25.560 --> 1:28:27.560 hundreds of dollars a month 1:28:27.560 --> 1:28:29.560 so 1:28:29.560 --> 1:28:31.560 how do you navigate the space 1:28:31.560 --> 1:28:33.560 that's a beautiful description, hurting rich cats 1:28:33.560 --> 1:28:35.560 I've never heard that before 1:28:35.560 --> 1:28:37.560 it is 1:28:37.560 --> 1:28:39.560 very complicated and I think 1:28:39.560 --> 1:28:41.560 certainly 1:28:41.560 --> 1:28:43.560 actually betting against Spotify has been statistically 1:28:43.560 --> 1:28:45.560 a very smart thing to do 1:28:45.560 --> 1:28:47.560 just looking at the 1:28:47.560 --> 1:28:49.560 line of roadkill in music streaming services 1:28:51.560 --> 1:28:53.560 it's kind of 1:28:53.560 --> 1:28:55.560 I think if I understood the complexity 1:28:55.560 --> 1:28:57.560 when I joined Spotify 1:28:57.560 --> 1:28:59.560 fortunately I didn't know enough 1:28:59.560 --> 1:29:01.560 about 1:29:01.560 --> 1:29:03.560 the music industry to understand the complexities 1:29:03.560 --> 1:29:05.560 because then I would have made a more rational guess 1:29:05.560 --> 1:29:07.560 that it wouldn't work 1:29:07.560 --> 1:29:09.560 so ignorance is bliss 1:29:09.560 --> 1:29:11.560 but I think 1:29:11.560 --> 1:29:13.560 there have been a few distinct challenges 1:29:13.560 --> 1:29:15.560 I think as I said 1:29:15.560 --> 1:29:17.560 one of the things that made it work at all 1:29:17.560 --> 1:29:19.560 was that Sweden and the Nordics 1:29:19.560 --> 1:29:21.560 was a lost market 1:29:21.560 --> 1:29:23.560 so there was 1:29:23.560 --> 1:29:25.560 no risk for labels to try this 1:29:25.560 --> 1:29:27.560 I don't think it would have worked 1:29:27.560 --> 1:29:29.560 if the market was 1:29:29.560 --> 1:29:31.560 healthy 1:29:31.560 --> 1:29:33.560 so that was the initial condition 1:29:33.560 --> 1:29:35.560 then we had this tremendous 1:29:35.560 --> 1:29:37.560 challenge with the model itself 1:29:37.560 --> 1:29:39.560 so now 1:29:39.560 --> 1:29:41.560 we're pirating but for the people who bought 1:29:41.560 --> 1:29:43.560 a download or a CD 1:29:43.560 --> 1:29:45.560 the artist would get 1:29:45.560 --> 1:29:47.560 all the revenue for all the future plays 1:29:47.560 --> 1:29:49.560 then right so you got it all up front 1:29:49.560 --> 1:29:51.560 whereas the streaming model was like 1:29:51.560 --> 1:29:53.560 almost nothing they won almost nothing they too 1:29:53.560 --> 1:29:55.560 and then at some point this curve 1:29:55.560 --> 1:29:57.560 of incremental revenue 1:29:57.560 --> 1:29:59.560 would intersect with your day one payment 1:29:59.560 --> 1:30:01.560 and that took a long time to play out 1:30:01.560 --> 1:30:03.560 before 1:30:03.560 --> 1:30:05.560 the music labels 1:30:05.560 --> 1:30:07.560 they understood that but on the artist side 1:30:07.560 --> 1:30:09.560 it took a lot of time to understand 1:30:09.560 --> 1:30:11.560 that actually if I have a big hit that is going to be played 1:30:11.560 --> 1:30:13.560 for many years this is a much better model 1:30:13.560 --> 1:30:15.560 because I get paid based 1:30:15.560 --> 1:30:17.560 on how much people use the product 1:30:17.560 --> 1:30:19.560 not how much they thought they would use it day one 1:30:19.560 --> 1:30:21.560 or so forth 1:30:21.560 --> 1:30:23.560 so it was a complicated model to get across 1:30:23.560 --> 1:30:25.560 but time helped with that 1:30:25.560 --> 1:30:27.560 and now 1:30:27.560 --> 1:30:29.560 the revenues to the music industry actually are bigger 1:30:29.560 --> 1:30:31.560 again 1:30:31.560 --> 1:30:33.560 it's gone through this incredible dip and now they're back up 1:30:33.560 --> 1:30:35.560 and so we're very proud 1:30:35.560 --> 1:30:37.560 of having been 1:30:37.560 --> 1:30:39.560 a part of that 1:30:39.560 --> 1:30:41.560 so there have been distinct problems 1:30:41.560 --> 1:30:43.560 I think when it comes to the 1:30:43.560 --> 1:30:45.560 labels 1:30:45.560 --> 1:30:47.560 we have taken the painful approach 1:30:47.560 --> 1:30:49.560 some of our competition at the time 1:30:49.560 --> 1:30:51.560 they kind of 1:30:51.560 --> 1:30:53.560 looked at other companies and said 1:30:53.560 --> 1:30:55.560 if we just ignore the rights 1:30:55.560 --> 1:30:57.560 we get really big really fast 1:30:57.560 --> 1:30:59.560 we're going to be too big for the 1:30:59.560 --> 1:31:01.560 labels to fail 1:31:01.560 --> 1:31:03.560 they're not going to kill us 1:31:03.560 --> 1:31:05.560 we're going to take that approach 1:31:05.560 --> 1:31:07.560 we went legal from day one 1:31:07.560 --> 1:31:09.560 and we negotiated 1:31:09.560 --> 1:31:11.560 and negotiated and negotiated 1:31:11.560 --> 1:31:13.560 it was very slow, very frustrating 1:31:13.560 --> 1:31:15.560 we were angry at seeing other companies 1:31:15.560 --> 1:31:17.560 taking shortcuts and seeming to get away with it 1:31:17.560 --> 1:31:19.560 it was this game theory thing 1:31:19.560 --> 1:31:21.560 where over many rounds of playing the game 1:31:21.560 --> 1:31:23.560 this would be the right strategy 1:31:23.560 --> 1:31:25.560 and even though 1:31:25.560 --> 1:31:27.560 clearly there's a lot of frustrations 1:31:27.560 --> 1:31:29.560 at times during renegotiations 1:31:29.560 --> 1:31:31.560 there is this weird trust 1:31:31.560 --> 1:31:33.560 honest and fair 1:31:33.560 --> 1:31:35.560 we've never screwed them, they've never screwed us 1:31:35.560 --> 1:31:37.560 it's ten years but 1:31:37.560 --> 1:31:39.560 there's this trust in like 1:31:39.560 --> 1:31:41.560 they know that if music doesn't get really big 1:31:41.560 --> 1:31:43.560 if lots of people do not want to listen 1:31:43.560 --> 1:31:45.560 to music and want to pay for it 1:31:45.560 --> 1:31:47.560 Spotify has no business model 1:31:47.560 --> 1:31:49.560 so we actually are incredibly aligned 1:31:49.560 --> 1:31:51.560 right? 1:31:51.560 --> 1:31:53.560 other companies have other business models 1:31:53.560 --> 1:31:55.560 where even if they made new music 1:31:55.560 --> 1:31:57.560 no money for music 1:31:57.560 --> 1:31:59.560 they'd still be profitable companies 1:31:59.560 --> 1:32:01.560 and I think the industry sees that 1:32:01.560 --> 1:32:03.560 we are actually aligned 1:32:03.560 --> 1:32:05.560 business wise 1:32:05.560 --> 1:32:07.560 so there is this trust 1:32:07.560 --> 1:32:09.560 that allows us to do 1:32:09.560 --> 1:32:11.560 product development even if it's scary 1:32:11.560 --> 1:32:13.560 you know 1:32:13.560 --> 1:32:15.560 taking risks 1:32:15.560 --> 1:32:17.560 the free model itself was an incredible risk 1:32:17.560 --> 1:32:19.560 for the music industry to take 1:32:19.560 --> 1:32:21.560 that they should get credit for 1:32:21.560 --> 1:32:23.560 now some of it was that they had nothing to lose in Sweden 1:32:23.560 --> 1:32:25.560 but frankly a lot of the labels also took risk 1:32:25.560 --> 1:32:27.560 and so I think we built up that trust 1:32:27.560 --> 1:32:29.560 with 1:32:29.560 --> 1:32:31.560 I think hurting a cat sounds a bit 1:32:31.560 --> 1:32:33.560 what's the word 1:32:33.560 --> 1:32:35.560 it sounds like 1:32:35.560 --> 1:32:37.560 dismissive? no every cat matter 1:32:37.560 --> 1:32:39.560 they are all beautiful and very important 1:32:39.560 --> 1:32:41.560 exactly they've taken a lot of risks 1:32:41.560 --> 1:32:43.560 and certainly it's been frustrating on both sides 1:32:45.560 --> 1:32:47.560 it's really like playing 1:32:47.560 --> 1:32:49.560 it's game theory if you play the 1:32:49.560 --> 1:32:51.560 if you play the game many times 1:32:51.560 --> 1:32:53.560 then you can have the statistical outcome 1:32:53.560 --> 1:32:55.560 that you bet on and it feels very painful 1:32:55.560 --> 1:32:57.560 when you're in the middle of that 1:32:57.560 --> 1:32:59.560 I mean there's risk there's trust 1:32:59.560 --> 1:33:01.560 there's relationships from 1:33:01.560 --> 1:33:03.560 just having read the biography 1:33:03.560 --> 1:33:05.560 of Steve Jobs 1:33:05.560 --> 1:33:07.560 similar kind of relationship were discussed 1:33:07.560 --> 1:33:09.560 in iTunes the idea of selling 1:33:09.560 --> 1:33:11.560 a song for a dollar was very uncomfortable 1:33:11.560 --> 1:33:13.560 for labels and 1:33:13.560 --> 1:33:15.560 and there was no 1:33:15.560 --> 1:33:17.560 it was the same kind of thing it was trust 1:33:17.560 --> 1:33:19.560 it was game theory 1:33:19.560 --> 1:33:21.560 as a lot of relationships that had to be built 1:33:21.560 --> 1:33:23.560 and it's really 1:33:23.560 --> 1:33:25.560 terrifyingly difficult 1:33:25.560 --> 1:33:27.560 process 1:33:27.560 --> 1:33:29.560 that Apple could go through a little bit 1:33:29.560 --> 1:33:31.560 because they could afford for that process 1:33:31.560 --> 1:33:33.560 to fail 1:33:33.560 --> 1:33:35.560 for Spotify it seems terrifying because 1:33:35.560 --> 1:33:37.560 you can't 1:33:37.560 --> 1:33:39.560 initially 1:33:39.560 --> 1:33:41.560 I think a lot of it comes down to 1:33:41.560 --> 1:33:43.560 honestly Daniel and his tenacity 1:33:43.560 --> 1:33:45.560 in negotiating which seems like an impossible 1:33:45.560 --> 1:33:47.560 discipline task 1:33:47.560 --> 1:33:49.560 because 1:33:49.560 --> 1:33:51.560 he was completely unknown and so forth 1:33:51.560 --> 1:33:53.560 and also the reason that 1:33:53.560 --> 1:33:55.560 it worked 1:33:55.560 --> 1:33:57.560 but I think 1:34:01.560 --> 1:34:03.560 I think game theory is probably the best way to think about it 1:34:03.560 --> 1:34:05.560 you could straight go straight for this like Nash 1:34:05.560 --> 1:34:07.560 equilibrium that someone is going to defect 1:34:07.560 --> 1:34:09.560 or you play many times 1:34:09.560 --> 1:34:11.560 and you try to actually go for the top left 1:34:11.560 --> 1:34:13.560 the corporations sell 1:34:13.560 --> 1:34:15.560 is there any magical 1:34:15.560 --> 1:34:17.560 reason why Spotify 1:34:17.560 --> 1:34:19.560 seems to have won 1:34:19.560 --> 1:34:21.560 so a lot of people have tried to do 1:34:21.560 --> 1:34:23.560 what Spotify tried to do 1:34:23.560 --> 1:34:25.560 and Spotify has come out 1:34:25.560 --> 1:34:27.560 well so the answer is that 1:34:27.560 --> 1:34:29.560 there's no magical reason because I don't believe in magic 1:34:29.560 --> 1:34:31.560 but I think there are 1:34:31.560 --> 1:34:33.560 there are reasons 1:34:33.560 --> 1:34:35.560 and I think some of them are that 1:34:35.560 --> 1:34:37.560 people have 1:34:37.560 --> 1:34:39.560 misunderstood 1:34:39.560 --> 1:34:41.560 a lot of what we actually do 1:34:41.560 --> 1:34:43.560 the actual 1:34:43.560 --> 1:34:45.560 Spotify model is very complicated 1:34:45.560 --> 1:34:47.560 they've looked at the premium model 1:34:47.560 --> 1:34:49.560 like you can charge 9.99 1:34:49.560 --> 1:34:51.560 for music and people are going to pay 1:34:51.560 --> 1:34:53.560 but that's not what happened 1:34:53.560 --> 1:34:55.560 actually when we launched the original mobile product 1:34:55.560 --> 1:34:57.560 everyone said they would never pay 1:34:57.560 --> 1:34:59.560 what happened was they started on the free product 1:34:59.560 --> 1:35:01.560 and then their engagement 1:35:01.560 --> 1:35:03.560 grew so much that eventually 1:35:03.560 --> 1:35:05.560 they said maybe it is worth 1:35:05.560 --> 1:35:07.560 9.99 right it's 1:35:07.560 --> 1:35:09.560 your propensity to pay gross with your engagement 1:35:09.560 --> 1:35:11.560 so we had this super complicated 1:35:11.560 --> 1:35:13.560 business model where you operate 1:35:13.560 --> 1:35:15.560 two different business model advertising and premium 1:35:15.560 --> 1:35:17.560 products at the same time 1:35:17.560 --> 1:35:19.560 and I think that is hard to replicate 1:35:19.560 --> 1:35:21.560 I struggle to think of other companies 1:35:21.560 --> 1:35:23.560 that run large scale advertising 1:35:23.560 --> 1:35:25.560 and subscription products at the same time 1:35:25.560 --> 1:35:27.560 so I think the business model is actually 1:35:27.560 --> 1:35:29.560 much more complicated than people think it is 1:35:29.560 --> 1:35:31.560 and so 1:35:31.560 --> 1:35:33.560 some people went after just the premium part 1:35:33.560 --> 1:35:35.560 without the free part and ran into a wall 1:35:35.560 --> 1:35:37.560 where no one wanted to pay 1:35:37.560 --> 1:35:39.560 some people went after just 1:35:39.560 --> 1:35:41.560 music should be free just ads 1:35:41.560 --> 1:35:43.560 which doesn't give you enough revenue 1:35:43.560 --> 1:35:45.560 for the music industry 1:35:45.560 --> 1:35:47.560 so I think that combination is 1:35:47.560 --> 1:35:49.560 kind of opaque from the outside 1:35:49.560 --> 1:35:51.560 so maybe I shouldn't say it here and reveal the secret 1:35:51.560 --> 1:35:53.560 but that turns out to be hard 1:35:53.560 --> 1:35:55.560 to replicate 1:35:55.560 --> 1:35:57.560 than you would think 1:35:57.560 --> 1:35:59.560 so there is a lot of brilliant business strategy here 1:35:59.560 --> 1:36:01.560 brilliance or luck 1:36:01.560 --> 1:36:03.560 probably more luck but it doesn't really matter 1:36:03.560 --> 1:36:05.560 it looks brilliant in retrospect 1:36:05.560 --> 1:36:07.560 so let's call it brilliant 1:36:07.560 --> 1:36:09.560 yeah when the books are written it will be brilliant 1:36:09.560 --> 1:36:11.560 you have 1:36:11.560 --> 1:36:13.560 mentioned that your philosophy is to 1:36:13.560 --> 1:36:15.560 embrace change 1:36:15.560 --> 1:36:17.560 so how will the music streaming 1:36:17.560 --> 1:36:19.560 and music listening 1:36:19.560 --> 1:36:21.560 world change over the next 1:36:21.560 --> 1:36:23.560 10 years, 20 years 1:36:23.560 --> 1:36:25.560 you look out into the far future 1:36:25.560 --> 1:36:27.560 what do you think? 1:36:27.560 --> 1:36:29.560 I think that music 1:36:29.560 --> 1:36:31.560 and for that matter audio 1:36:31.560 --> 1:36:33.560 podcast audio books 1:36:33.560 --> 1:36:35.560 I think it's one of the few 1:36:35.560 --> 1:36:37.560 core human needs 1:36:37.560 --> 1:36:39.560 I think there is no good reason to me 1:36:39.560 --> 1:36:41.560 why it shouldn't be at the scale 1:36:41.560 --> 1:36:43.560 of something like messaging or social networking 1:36:43.560 --> 1:36:45.560 I don't think it's a niche thing 1:36:45.560 --> 1:36:47.560 to listen to music or news or something 1:36:47.560 --> 1:36:49.560 so I think scale is obviously one of the things 1:36:49.560 --> 1:36:51.560 that I really hope for 1:36:51.560 --> 1:36:53.560 I hope that it's going to be 1:36:53.560 --> 1:36:55.560 billions of users, I hope eventually 1:36:55.560 --> 1:36:57.560 everyone in the world gets access to 1:36:57.560 --> 1:36:59.560 all the world's music ever made 1:36:59.560 --> 1:37:01.560 so obviously I think it's going to be a much bigger business 1:37:01.560 --> 1:37:03.560 otherwise we wouldn't be betting this big 1:37:05.560 --> 1:37:07.560 now if you look more at how 1:37:07.560 --> 1:37:09.560 it is consumed 1:37:09.560 --> 1:37:11.560 what I'm hoping is back to this 1:37:11.560 --> 1:37:13.560 analogy of the 1:37:13.560 --> 1:37:15.560 software tool chain 1:37:15.560 --> 1:37:17.560 where I think 1:37:17.560 --> 1:37:19.560 sometimes internally 1:37:19.560 --> 1:37:21.560 I make this analogy to 1:37:21.560 --> 1:37:23.560 text messaging 1:37:23.560 --> 1:37:25.560 text messaging was also based on 1:37:25.560 --> 1:37:27.560 standards 1:37:27.560 --> 1:37:29.560 in the area of mobile carriers 1:37:29.560 --> 1:37:31.560 you had the SMS, the 140 character 1:37:31.560 --> 1:37:33.560 120 carat SMS 1:37:33.560 --> 1:37:35.560 and it was great 1:37:35.560 --> 1:37:37.560 because everyone agreed on the standard 1:37:37.560 --> 1:37:39.560 so as a consumer you got a lot of distributions 1:37:39.560 --> 1:37:41.560 and interoperability but it was a very constrained format 1:37:41.560 --> 1:37:43.560 and when the industry 1:37:43.560 --> 1:37:45.560 wanted to add pictures to that format 1:37:45.560 --> 1:37:47.560 to do the MMS, I looked it up 1:37:47.560 --> 1:37:49.560 and I think it took from the late 80s to early 2000 1:37:49.560 --> 1:37:51.560 this is like a 15, 20 year product cycle 1:37:51.560 --> 1:37:53.560 to bring pictures into that 1:37:53.560 --> 1:37:55.560 now once that entire 1:37:55.560 --> 1:37:57.560 value chain 1:37:57.560 --> 1:37:59.560 of creation and consumption got wrapped 1:37:59.560 --> 1:38:01.560 in one software stack 1:38:01.560 --> 1:38:03.560 within something like Snapchat or WhatsApp 1:38:03.560 --> 1:38:05.560 the first week 1:38:05.560 --> 1:38:07.560 they added disappearing messages 1:38:07.560 --> 1:38:09.560 then two weeks later they added stories 1:38:09.560 --> 1:38:11.560 the pace of innovation when you're on one software stack 1:38:11.560 --> 1:38:13.560 you can 1:38:13.560 --> 1:38:15.560 affect both creation and consumption 1:38:15.560 --> 1:38:17.560 I think it's going to be rapid 1:38:17.560 --> 1:38:19.560 so with these streaming services 1:38:19.560 --> 1:38:21.560 we now for the first time in history 1:38:21.560 --> 1:38:23.560 have enough I hope people 1:38:23.560 --> 1:38:25.560 on one of these services 1:38:25.560 --> 1:38:27.560 actually whether it's Spotify or Amazon 1:38:27.560 --> 1:38:29.560 or Apple or YouTube 1:38:29.560 --> 1:38:31.560 and hopefully enough creators 1:38:31.560 --> 1:38:33.560 can do the format again 1:38:33.560 --> 1:38:35.560 and that excites me 1:38:35.560 --> 1:38:37.560 I think being able to change these constraints from 100 years 1:38:37.560 --> 1:38:39.560 that could really 1:38:39.560 --> 1:38:41.560 do something interesting 1:38:41.560 --> 1:38:43.560 I really hope it's not just going to be there 1:38:43.560 --> 1:38:45.560 iteration on the same thing 1:38:45.560 --> 1:38:47.560 for the next 10 to 20 years as well 1:38:47.560 --> 1:38:49.560 yeah changing the creation 1:38:49.560 --> 1:38:51.560 of music or creation of audio 1:38:51.560 --> 1:38:53.560 or creation of podcasts 1:38:53.560 --> 1:38:55.560 is a really fascinating possibility 1:38:55.560 --> 1:38:57.560 I myself don't understand 1:38:57.560 --> 1:38:59.560 what it is about podcasts that's so intimate 1:38:59.560 --> 1:39:01.560 it just is I listen to a lot of podcasts 1:39:01.560 --> 1:39:03.560 I think it touches 1:39:03.560 --> 1:39:05.560 on a human on a deep 1:39:05.560 --> 1:39:07.560 human need for connection 1:39:07.560 --> 1:39:09.560 that people do feel like they're 1:39:09.560 --> 1:39:11.560 connected 1:39:11.560 --> 1:39:13.560 to when they listen I don't understand 1:39:13.560 --> 1:39:15.560 what the psychology that is 1:39:15.560 --> 1:39:17.560 but in this world is becoming 1:39:17.560 --> 1:39:19.560 more and more disconnected 1:39:19.560 --> 1:39:21.560 it feels like 1:39:21.560 --> 1:39:23.560 this is fulfilling a certain kind of 1:39:23.560 --> 1:39:25.560 need and 1:39:25.560 --> 1:39:27.560 empowering the creator as opposed 1:39:27.560 --> 1:39:29.560 to just the listener 1:39:29.560 --> 1:39:31.560 is really interesting 1:39:31.560 --> 1:39:33.560 I'm really excited that you're working on this 1:39:33.560 --> 1:39:35.560 yeah I think one of the things that is inspiring 1:39:35.560 --> 1:39:37.560 for our teams to work on podcasts 1:39:37.560 --> 1:39:39.560 is exactly that 1:39:39.560 --> 1:39:41.560 whether you think like I probably do 1:39:41.560 --> 1:39:43.560 that it's something biological about 1:39:43.560 --> 1:39:45.560 perceiving to be in the middle of the conversation 1:39:45.560 --> 1:39:47.560 that makes you listen in a different way 1:39:47.560 --> 1:39:49.560 it doesn't really matter people seem to perceive it 1:39:49.560 --> 1:39:51.560 differently and 1:39:51.560 --> 1:39:53.560 there was this narrative for a long time that 1:39:53.560 --> 1:39:55.560 you know if you look at video 1:39:55.560 --> 1:39:57.560 it's something kind of in the foreground it got shorter 1:39:57.560 --> 1:39:59.560 and shorter and shorter because of financial 1:39:59.560 --> 1:40:01.560 pressures and monetization and so forth 1:40:01.560 --> 1:40:03.560 and eventually at the end there's almost like 1:40:03.560 --> 1:40:05.560 20 seconds clip 1:40:05.560 --> 1:40:07.560 people just screaming something 1:40:07.560 --> 1:40:09.560 and 1:40:09.560 --> 1:40:11.560 I'm really I feel really good about 1:40:11.560 --> 1:40:13.560 the fact that 1:40:13.560 --> 1:40:15.560 you could have interpreted that as people have 1:40:15.560 --> 1:40:17.560 no attention span anymore 1:40:17.560 --> 1:40:19.560 they don't want to listen to things they're not interested 1:40:19.560 --> 1:40:21.560 in deeper stories 1:40:21.560 --> 1:40:23.560 like you know people are getting dumber 1:40:23.560 --> 1:40:25.560 but then podcast came along and it's almost like no 1:40:25.560 --> 1:40:27.560 no the need still existed 1:40:27.560 --> 1:40:29.560 once but maybe maybe it was 1:40:29.560 --> 1:40:31.560 the fact that you're not prepared to look at your 1:40:31.560 --> 1:40:33.560 phone like this for two hours 1:40:33.560 --> 1:40:35.560 but if you can drive at the same time it seems like 1:40:35.560 --> 1:40:37.560 people really want to dig deeper 1:40:37.560 --> 1:40:39.560 and they want to hear like the more complicated version 1:40:39.560 --> 1:40:41.560 so to me that is very inspiring 1:40:41.560 --> 1:40:43.560 that podcast is actually long form 1:40:43.560 --> 1:40:45.560 it gives me a lot of hope for 1:40:45.560 --> 1:40:47.560 for humanity that people seem really interested 1:40:47.560 --> 1:40:49.560 in hearing deeper more complicated 1:40:49.560 --> 1:40:51.560 conversations this is 1:40:51.560 --> 1:40:53.560 I don't understand it 1:40:53.560 --> 1:40:55.560 it's fascinating so the majority 1:40:55.560 --> 1:40:57.560 for this podcast listen to the whole thing 1:40:57.560 --> 1:40:59.560 this whole conversation 1:40:59.560 --> 1:41:01.560 we've been talking for an hour and 45 minutes 1:41:01.560 --> 1:41:03.560 and somebody will 1:41:03.560 --> 1:41:05.560 I mean most people will be listening 1:41:05.560 --> 1:41:07.560 to these words I'm speaking right now 1:41:07.560 --> 1:41:09.560 you wouldn't have thought that 10 years ago 1:41:09.560 --> 1:41:11.560 where the world seemed to go 1:41:11.560 --> 1:41:13.560 that's very positive I think 1:41:13.560 --> 1:41:15.560 that's really exciting and empowering the creator 1:41:15.560 --> 1:41:17.560 there's as really exciting 1:41:17.560 --> 1:41:19.560 last question 1:41:19.560 --> 1:41:21.560 do you also have a passion for 1:41:21.560 --> 1:41:23.560 just mobile in general 1:41:23.560 --> 1:41:25.560 how do you see the smartphone world 1:41:27.560 --> 1:41:29.560 the digital space 1:41:29.560 --> 1:41:31.560 of 1:41:31.560 --> 1:41:33.560 of smartphones 1:41:33.560 --> 1:41:35.560 and just everything that's on the move 1:41:35.560 --> 1:41:37.560 whether it's 1:41:37.560 --> 1:41:39.560 internet of things and so on 1:41:39.560 --> 1:41:41.560 changing over the next 10 years 1:41:41.560 --> 1:41:43.560 and so on 1:41:43.560 --> 1:41:45.560 I think that one way to think about it is that 1:41:45.560 --> 1:41:47.560 computing 1:41:47.560 --> 1:41:49.560 moving out of these 1:41:49.560 --> 1:41:51.560 multi purpose devices 1:41:51.560 --> 1:41:53.560 the computer we had in the phone 1:41:53.560 --> 1:41:55.560 into specific 1:41:55.560 --> 1:41:57.560 specific purpose devices 1:41:57.560 --> 1:41:59.560 and it will be ambient that 1:41:59.560 --> 1:42:01.560 at least in my home 1:42:01.560 --> 1:42:03.560 you just shout something at someone 1:42:03.560 --> 1:42:05.560 and there's always one of these speakers close enough 1:42:05.560 --> 1:42:07.560 and so 1:42:07.560 --> 1:42:09.560 you start behaving differently 1:42:09.560 --> 1:42:11.560 it's as if you have the internet ambience 1:42:11.560 --> 1:42:13.560 ambiently around you and you can ask it 1:42:13.560 --> 1:42:15.560 things 1:42:15.560 --> 1:42:17.560 so I think computing 1:42:17.560 --> 1:42:19.560 will kind of get more integrated 1:42:19.560 --> 1:42:21.560 and we won't necessarily think of it 1:42:21.560 --> 1:42:23.560 as 1:42:23.560 --> 1:42:25.560 connected to a device in the same way 1:42:25.560 --> 1:42:27.560 that we do today 1:42:27.560 --> 1:42:29.560 I don't know the path to that maybe 1:42:29.560 --> 1:42:31.560 we used to have these 1:42:31.560 --> 1:42:33.560 desktop computers 1:42:33.560 --> 1:42:35.560 and then we partially replaced that with 1:42:35.560 --> 1:42:37.560 the laptops and left 1:42:37.560 --> 1:42:39.560 desktop at home and at work and then 1:42:39.560 --> 1:42:41.560 we got these phones and we started leaving 1:42:41.560 --> 1:42:43.560 the laptop at home for a while and maybe 1:42:43.560 --> 1:42:45.560 for stretches of time 1:42:45.560 --> 1:42:47.560 you're going to start using the watch and you can leave 1:42:47.560 --> 1:42:49.560 your phone at home like for a run 1:42:49.560 --> 1:42:51.560 or something and we're on this 1:42:51.560 --> 1:42:53.560 progressive path where 1:42:53.560 --> 1:42:55.560 I think 1:42:55.560 --> 1:42:57.560 what is happening with 1:42:57.560 --> 1:42:59.560 voice is that 1:42:59.560 --> 1:43:01.560 you have an 1:43:01.560 --> 1:43:03.560 interaction paradigm that doesn't 1:43:03.560 --> 1:43:05.560 require 1:43:05.560 --> 1:43:07.560 as large physical devices so I definitely 1:43:07.560 --> 1:43:09.560 think there's a future where you can have 1:43:09.560 --> 1:43:11.560 your Airpods and 1:43:11.560 --> 1:43:13.560 your watch and you can do 1:43:13.560 --> 1:43:15.560 a lot of computing 1:43:15.560 --> 1:43:17.560 and I don't think 1:43:17.560 --> 1:43:19.560 it's going to be this binary 1:43:19.560 --> 1:43:21.560 thing, I think it's going to be like many of us 1:43:21.560 --> 1:43:23.560 still have a laptop, we just use it less 1:43:23.560 --> 1:43:25.560 and so you shift your consumption over 1:43:25.560 --> 1:43:27.560 and 1:43:27.560 --> 1:43:29.560 I don't know about 1:43:29.560 --> 1:43:31.560 AR glasses and so forth 1:43:31.560 --> 1:43:33.560 I'm excited about it, I spent a lot of time in that area 1:43:33.560 --> 1:43:35.560 but I still think it's quite far away 1:43:35.560 --> 1:43:37.560 AR, VR, all of that 1:43:37.560 --> 1:43:39.560 Yeah, VR is happening and 1:43:39.560 --> 1:43:41.560 working, I think the recent 1:43:41.560 --> 1:43:43.560 Oculus Quest is quite impressive 1:43:43.560 --> 1:43:45.560 I think AR is further away 1:43:45.560 --> 1:43:47.560 at least that type of AR 1:43:47.560 --> 1:43:49.560 but I do think 1:43:51.560 --> 1:43:53.560 your phone or watch or glasses understanding 1:43:53.560 --> 1:43:55.560 where you are and maybe what you're looking at 1:43:55.560 --> 1:43:57.560 and being able to give you audio cues about that or you can say 1:43:57.560 --> 1:43:59.560 like what is this and it tells you what it is 1:43:59.560 --> 1:44:01.560 that I think 1:44:01.560 --> 1:44:03.560 might happen, you use 1:44:03.560 --> 1:44:05.560 your watch or your glasses as 1:44:05.560 --> 1:44:07.560 a mouse pointer on reality 1:44:07.560 --> 1:44:09.560 I think it might be a while before 1:44:09.560 --> 1:44:11.560 I might be wrong, I hope I'm wrong but I think it might be a while 1:44:11.560 --> 1:44:13.560 before we walk around with these big lab glasses 1:44:13.560 --> 1:44:15.560 that project things 1:44:15.560 --> 1:44:17.560 I agree with you, it's actually really 1:44:17.560 --> 1:44:19.560 difficult when you have to 1:44:19.560 --> 1:44:21.560 understand the physical world 1:44:21.560 --> 1:44:23.560 enough to project 1:44:23.560 --> 1:44:25.560 onto it 1:44:25.560 --> 1:44:27.560 I lied about the last question 1:44:27.560 --> 1:44:29.560 because I just thought 1:44:29.560 --> 1:44:31.560 of audio 1:44:31.560 --> 1:44:33.560 and my favorite topic which is the movie 1:44:33.560 --> 1:44:37.560 Her, do you think 1:44:37.560 --> 1:44:39.560 whether it's part of Spotify or not 1:44:39.560 --> 1:44:41.560 will have 1:44:41.560 --> 1:44:43.560 I don't know if you've seen the movie Her 1:44:43.560 --> 1:44:45.560 absolutely 1:44:45.560 --> 1:44:47.560 and there 1:44:47.560 --> 1:44:49.560 audio is 1:44:49.560 --> 1:44:51.560 the primary form of interaction 1:44:51.560 --> 1:44:53.560 and the connection with another entity 1:44:53.560 --> 1:44:55.560 that you can actually have a 1:44:55.560 --> 1:44:57.560 relationship with or fall in love with 1:44:57.560 --> 1:44:59.560 based on voice alone 1:44:59.560 --> 1:45:01.560 audio alone 1:45:01.560 --> 1:45:03.560 do you think that's possible first of all 1:45:03.560 --> 1:45:05.560 based on audio alone to fall in love with somebody 1:45:05.560 --> 1:45:07.560 somebody or well yeah let's go 1:45:07.560 --> 1:45:09.560 with somebody, just have a relationship 1:45:09.560 --> 1:45:11.560 based on audio alone 1:45:11.560 --> 1:45:13.560 and second question to that 1:45:13.560 --> 1:45:15.560 can we create an artificial intelligence system 1:45:15.560 --> 1:45:17.560 that 1:45:17.560 --> 1:45:19.560 allows one to fall in love 1:45:19.560 --> 1:45:21.560 with it and her, him 1:45:21.560 --> 1:45:23.560 with you 1:45:23.560 --> 1:45:25.560 so this is my personal answer 1:45:25.560 --> 1:45:27.560 speaking for me as a person 1:45:27.560 --> 1:45:29.560 the answer is quite unequivocally 1:45:29.560 --> 1:45:31.560 yes 1:45:31.560 --> 1:45:33.560 on both 1:45:33.560 --> 1:45:35.560 well we just said about podcasts 1:45:35.560 --> 1:45:37.560 and the feeling of being in the middle of a conversation 1:45:37.560 --> 1:45:39.560 if you could have an 1:45:39.560 --> 1:45:41.560 assistant where 1:45:41.560 --> 1:45:43.560 and we just said that feels like a very personal 1:45:43.560 --> 1:45:45.560 setting so if you walk around with these 1:45:45.560 --> 1:45:47.560 headphones and this thing, you're speaking with this 1:45:47.560 --> 1:45:49.560 thing all of the time that feels like it's in your brain 1:45:49.560 --> 1:45:51.560 I think it's 1:45:51.560 --> 1:45:53.560 it's gonna be much easier to fall in love with 1:45:53.560 --> 1:45:55.560 than something that would be on your screen 1:45:55.560 --> 1:45:57.560 I think that's entirely possible 1:45:57.560 --> 1:45:59.560 and then from the, you can probably answer this better than me 1:45:59.560 --> 1:46:01.560 but from the concept of 1:46:01.560 --> 1:46:03.560 if it's going to be possible 1:46:03.560 --> 1:46:05.560 to build a machine 1:46:05.560 --> 1:46:07.560 that can achieve that 1:46:07.560 --> 1:46:09.560 I think whether you 1:46:09.560 --> 1:46:11.560 think of it as, if you can fake it 1:46:11.560 --> 1:46:13.560 the philosophical zombie that simulates it enough 1:46:13.560 --> 1:46:15.560 or it somehow actually is 1:46:15.560 --> 1:46:17.560 I think there's 1:46:17.560 --> 1:46:19.560 it's only a question, if you ask me about time 1:46:19.560 --> 1:46:21.560 I'd have to give an answer but if you say I've given 1:46:21.560 --> 1:46:23.560 some half infinite time 1:46:23.560 --> 1:46:25.560 absolutely, I think it's just 1:46:25.560 --> 1:46:27.560 atoms and arrangement 1:46:27.560 --> 1:46:29.560 of information 1:46:29.560 --> 1:46:31.560 well, I personally think that love 1:46:31.560 --> 1:46:33.560 is a lot simpler than people think 1:46:33.560 --> 1:46:35.560 so we started with true romance 1:46:35.560 --> 1:46:37.560 and ended in love 1:46:37.560 --> 1:46:39.560 I don't see a better place to end 1:46:39.560 --> 1:46:40.560 beautiful 1:46:40.560 --> 1:46:41.560 Gustav, thanks so much for talking today 1:46:41.560 --> 1:46:55.560 thank you so much, it was a lot of fun