The Future of Audio

Sonal Choksi: What happens when we give creators of all kinds more tools, both for expression and for monetization and engagement? And how does this play out in particular in the world of audio and podcasting?

[Intro music fades]

Sonal Choksi: In this episode from October 2020, A16Z general partner Connie Chan and Spotify's chief R&D officer Gustav Schroederstrom join host Sonal Choksi to discuss the past, present, and future of audio. They dig into everything from what the past and radio can tell us about the future, what audio can and will borrow from mediums like video or platforms like TikTok, the role for more interactivity and increased use of tools like machine learning and AI, and more.

[Music transitions]

Sonal Choksi: Hi, everyone. Welcome to the A16Z podcast. I'm Sonal, and today we are talking about one of my many, but actually probably most favorite topics, the future of audio. Our special guest is Gustav Schöderstrom, the chief R&D officer of Spotify, which is the world's most popular audio streaming subscription service. As a reminder, none of the following should be taken as investment advice. Please see A16Z.com slash disclosures for more information.

Sonal Choksi: Also joining this episode is A16Z general partner Connie Chan, who covers consumer, writes a lot about tech trends and product in China and beyond, alternative monetization models and more. She and I have actually done a couple of podcasts on podcasting. One, a podcast about podcasting with Nick Kwa and the other on how we at A16Z podcast. You can find both of those episodes as well as other resources on the topic at A16Z.com slash podcasting.

Sonal Choksi: Note also that Spotify actually got into podcasting in 2015. We were actually included as one of their launch partners for that among select others since they were huge fans of the pod. We still are, so it's still true. Thank you.

Sonal Choksi: Anyway, in this episode, we actually go beyond podcasting to talk about the broader category of audio, past, present, and future. So we chat about the parallels and differences in audio and video, including referencing an episode I recently did with Eugene Wei on TikTok, which you can also catch in this feed. We discuss the trend of interactivity as well as augmented audio and where we are right now, what's possible, what are the challenges. We talk about where podcasting and music converge and diverge, both on user experience and design, as well as technically in machine learning. And finally, we go deep on recommender systems, the idea of hearing like an algorithm and where subscription models come into machine learning.

Sonal Choksi: But we also talk throughout this episode about the trade-offs of full-stack approaches, regardless of what kind of company you are, and the topic of super apps as well. And we're also really talking about how innovation happens in practice, whether it's having an opinionated point of view about the future or listening to users, disrupting oneself and how to change an organization and much more.

Sonal Choksi: But we begin, however, with a super quick debate on how much things have or haven't changed in the podcasting world, at least since we did our last podcasting episode over a year and a half ago.

Connie Chan: I actually personally think that audio hasn't changed that much yet. A lot of things are still — I don't know if broken is the right word — but just problems that are not solved yet. Discovery is still difficult. Search is still difficult. It's really like a one-way listening experience. You aren't interacting with other listeners. You aren't interacting with the creators. Creators still have to rely on very old business models for monetization that ultimately don't work for a lot of long-tail creators. A lot of those big problems still exist. But I do have this optimistic feel that we're on the cusp of change that's going to come to the broader audio market.

Gustav Schöderstrom: You're right. Those things actually haven't changed very much. I was thinking of the fact that the content landscape in podcasting has super exploded in the last year, two years alone. Spotify itself has led a number of content acquisitions, which is such an interesting evolution.

Sonal Choksi: Yes, it's both very much the same, but very much more of the same, right? So like the forklifting of your time into your AirPods, that just keeps increasing.

Gustav Schöderstrom: Right. There's certainly been shifts in listening behavior due to COVID. A lot of listening was in the car, that shifted to speakers in the home. So overall, there's much more listening. And to your point, certainly we've invested aggressively in content and exclusives. The creator side of this landscape has changed in a direction that we wanted to change. But I would also agree that we're on the cusp, on the consumer experience. And what's so interesting about audio is it feels like you have this cheat sheet which is what happened in video. We just haven't done monetization in a 21st century way yet. We have no interactivity. You can really just look at the other media industries and see what's missing in a sense.

Sonal Choksi: So Edison Research, which publishes a lot of the leading work in studying podcasting behavior, they argued a few things last year. One, that one of the major inflection points in podcasting, interestingly, came through Spotify because of the streaming. And that brought in a new generation of users. Two, the other argument they made — and this is, of course, pre a lot of the content acquisitions — is that for a new generation, the medium of audio is really not that different than video. That in fact, for a lot of people, their default podcast player is often a video app or just turning off the visuals and listening. And so I'm curious for your guys' thoughts on where audio and video, which is another big trend, do and don't intersect, both from a trend perspective and a product development perspective. And then we can dig in deeper on other aspects.

Gustav Schöderstrom: I mean, video is really just the combination of using your ears and your eyes. It's the audio plus the visual, which means the stakes are actually higher for audio because I can't have like a 20 second gap of silence in a podcast and expect you to be okay with it. But in a video, you can go quiet and there might just be some visual distraction and you don't have to be on as much every second. And so it's still a different medium, but I do think that the stakes in audio are higher.

Gustav Schöderstrom: So I think that when you talk about audio, it's different things depending on the type of audio, actually. So you have foreground audio, which is more similar to video. It is the main activity you're doing. You're really concentrating. It requires most of your attention. Then you have background audio, like you're listening to music and you're actually paying attention to something completely different. You're working out or you're studying or something, right? So there are these different modes of audio that don't really exist in video. Video is mostly all your attention or you're doing something else, right?

Gustav Schöderstrom: This is also the benefit of audio. That's why it's so much engagement because you have both foreground moments and background moments. But even in the foreground moments, when you're paying full attention, you can still do other things. You can drive, you can do dishes, you can walk around the house, right? So it is this other mode that video doesn't cover. That's why we think it is almost as much engagement as foreground video, but it's not nearly valued the same yet. And that's not because it's less valuable. We think that's because it's undervalued.

Gustav Schöderstrom: You can think about it the other way as well. You have some video that actually works quite well as audio that you can background, that you watch every now and then. Joe Rogan, for example, it certainly has video, right? And that actually does help the user experience, but it is what we call backgroundable video or foregroundable audio, if you want to call it that.

Connie Chan: I just wanted to comment, Gustav, on your point about the modes. That's a phrase that I use when I think about describing people's behaviors. And I actually describe it less as foreground and background and more as passive versus active mode. And so I really believe strongly that audio has different modes. Sometimes you're just in hanging out in chill mode. Sometimes I'm in passive mode, which means I just want to listen to other people. Other times I'm in active mode, which means I want to talk or super active mode, which means I want to lead a discussion. So I just think it's really interesting to think in terms of modes.

Sonal Choksi: I'd love to hear your initial thoughts on just the medium differences between audio and video. What do you make of the differences and similarities between TikTok and what we can and can't learn from TikTok when it comes to product in audio? Do you guys have any thoughts on that? Connie, you've written so many posts about TikTok since very early on.

Connie Chan: Yeah, like TikTok's an extreme example. If you don't look at the screen and you just listen, none of the videos make sense. You'll miss the punchline, like the whole video.

Gustav Schöderstrom: Exactly. The value proposition is in the visual for TikTok. So I think there are at least two similarities. What they do really well is they take, to Connie's point, commodity music that if you just listen to it in the background, you miss the whole point. But then they let their users unikify that commodity music, right? By adding uniqueness to it with their video.

Sonal Choksi: I think you just made up a word, by the way, unikify. Keep going.

Gustav Schöderstrom: And I think that's a great pattern, right? You have something that is commodity. You can use your user base to turn that into something that is non-commodity. It's this engine that takes these clips and creates unique content around it. So I think that's a really interesting pattern that you could probably copy to other businesses that has commodity content. Let your audience do something with it to make it unique.

Gustav Schöderstrom: The other analogy that I see to audio is specifically music. If you think about Eugene Wei's post on seeing an algorithm, what he said was that the medium itself is built to be understood by an algorithm. If you're presented with one item at a time, you either consume or you swipe. So it's built for the algorithm to understand what you're paying attention to versus, for example, a scrolling feed where the algorithm has no idea which item your eyes are actually looking at.

Sonal Choksi: Isolating the specific variables so that the product developer knows what is working or not working, essentially, for the user.

Gustav Schöderstrom: Exactly. And if you think about music, actually, it's the exact same thing. You present one audio track at a time, you either listen to it or you skip. So in that sense, you can say it's a similar sort of UI, but in audio. The tricky part is actually just the length of the song versus the length of the TikTok video, because you get to a very quick decision if you like that TikTok video or not, literally within like two, three seconds. For a song, as many of you know, the first couple seconds of a song doesn't sound anything like the chorus or the ending. So you just have to go further into the song before you really gauge if someone truly likes it or not. But to me, that's the only difference.

Connie Chan: Yeah. In TikTok, you have more evaluations per minute because there are shorter clips, but it's also more direct.

Gustav Schöderstrom: But it is interesting that you mentioned this because this is what is happening in the label industry. It is super clear that the intro matters more and more. So you do have the TikTok effect in music. Songs used to start slow, they don't anymore because people speak within the first 10 seconds.

Sonal Choksi: Oh, that's so fascinating. So the TikTok effect where people are now creating different kind of music.

Gustav Schöderstrom: I would say one more thing on TikTok. So while there are some similarities between evaluating audio one track at a time and evaluating video one track at a time, there is a big difference, which is TikTok has your full attention. If you're full screen and you're paying full attention, then it's a pretty good signal. But if you're washing dishes and listening on a speaker, you get very poor signals. So it depends on the context and you have to take that into account when you look at the signal.

Sonal Choksi: I'd love to probe briefly on this part, which is you both have talked a lot. Connie, you in particular have written so much about how mobile is literally the thing that made a lot of China's apps work the way they do because everything was mobile first. And we talked about mobile leapfrogging in our posts from what now five years ago. Wow, that's been a long time. So where does that come in when you think about innovation in audio? And then Gustav, I'd love your thoughts on this as well, because when you said that in the pandemic, a lot of the listening behavior has shifted to home speakers. I'm curious how that changes your views given initially mobile default interface.

Connie Chan: So if I just break down what a phone is and the different components of it. You have the touchscreen, which means whatever you're doing on the phone, you can have more interactivity, ideally. But you also have camera and GPS. And the camera is the unlock for TikTok. And the microphone could be the unlock for a bunch of audio platforms. Because now it means that I don't just have to be listening. I'm not just leveraging the speaker on the phone, but I'm leveraging the microphone and I'm giving back. The microphone in particular for audio and video, I think, is dramatic.

Gustav Schöderstrom: Yeah, that is one of the sensors that is super interesting and under leveraged for audio, I would say. So one of the benefits of being a streaming service is that we understand the consumption situation. We understand if you're listening on a speaker, but putting on an Apple watch or a phone, we understand if you're in your car, for example, because the phone is connected and so forth. So we actually think that's a very important signal. And we try to think of them as different jobs to be done. And what we want to try to understand is the situation that you're in. And it's always a combination of your play history, your time, and your taste, but the device is actually a really good signal.

Gustav Schöderstrom: So there are two levels. One is the UI and the hardware that you can leverage. And that changes when you go from a phone to a connected speaker, for example. You have much less control. You actually still do have a feedback channel in terms of a microphone, as Connie mentioned, but you have less UI, right? So we're thinking about multimodal consumption quite a lot, where you have some devices that are really good for input on your body, but they're not that good for output. You actually want the sound in your speakers. That's why we built this remote control protocol so that you don't have to interact in the same place that you're listening. You can interact on one device and so forth.

Gustav Schöderstrom: The other way to think about it is on the content level. So one of the things that happened during COVID when a lot of consumption shifted from the car to the home was that we have this very successful playlist called the Daily Drive, where we mix music and talk and create literally your daily drive. Now people stop driving, right? So then we try to pivot and we create the same job to be done, but not while driving. It's different. So these are the two levels, the content level and the pure UX interactivity level.

Sonal Choksi: Okay, so we can shift into discovery and recommendations in a bit. But before we close this thread, what do you guys think of this trend and phrase augmented audio, which means different things to different people, but the idea that you can actually, to your point, Connie, much like video has many layers, you can actually bring more and more layers into audio as well. Do you guys have any quick thoughts on that?

Connie Chan: Oh, so many. But that really just leads me to the belief that audio today is still this more sit back experience. It very much like a one-way consumption experience, the same way that we consume television or the same way that we consume movies. And like more YouTube live streaming, that kind of format hasn't really arrived in mainstream in audio yet. And so even just capturing the comments, the feedback to podcasts, that kind of content is not well harnessed today. So there's so many more layers around the listener feedback or interacting with other listeners or interacting with the creator. A lot of fun should be added on and layered on into audio that right now at least doesn't exist.

Sonal Choksi: It doesn't have to even necessarily be fun. As a creator, I found the news when you guys rolled out your polls feature to be quite interesting because we just had the debates here in the United States. And I literally was like, I wonder if a lot of the political news shows should do their own polling as part of their audio experience. I mean, it's not just fun, it's instant feedback.

Gustav Schöderstrom: Yeah, I agree. We started with polls, which is both a safe and constructive way to bring feedback. You mentioned the consumers or the listeners talking to each other. You mentioned the creator talking to the listener. We try to focus on the creator and what tools does the creator want? And actually, not just for having fun, but to your point, Sonal, to be a better creator? What information do you want from your fans? And what would make it easier for a creator to produce another episode, for example? And so we started with Pulse, which is one way to get clear answers on questions you have. And we want to continue in this way, focusing not really on listener-to-listener conversations. I mean, you have Instagram, Facebook, Twitter. There's lots of places to go and talk to other users, but there aren't a lot of places to have good conversations with the creators.

Connie Chan: And I think if you focus on creators, there's also a huge opportunity to expand the funnel of creators. If you look at trends in video, lots of the top trending YouTube videos are actually reaction videos where people are watching a video and showcasing a reaction. And TikTok is all about remixing. There's a lot of great audio content out there today that if you talk about augmented audio, you could take a podcast and then have another person share their thoughts directly, just like a sports broadcaster even commenting directly on what's happening in the audio, whether it's music or even another podcast.

Gustav Schöderstrom: Yeah, you have these two extremes, like the old world broadcast one-way media. And then on the other extreme, I would put gaming, where the interactivity is the experience. You're not being broadcast at anything, you're actually creating it. And then you have this thing in between. And I think audio needs to move towards interactivity. And like I said, there is basically a cheat sheet where you can look at other types of media. And as soon as you add a feedback loop, the creator gets a chance to improve. So I think that's vital.

Sonal Choksi: Tell me more about some of your thinking behind polls. When you guys design a product, do you actually have an opinionated philosophy that this is how we think people are going to use it? Or are you just giving them the bare minimum and then unlocking your community to let loose? A simplified way of asking that is also, is it a Steve Jobs point of view or a Bezos point of view?

Gustav Schöderstrom: That's a great question, a great way to put it. And it's a tough question to answer. It's definitely not a Steve Jobs point of view in the sense that we know how people are going to use it. But we try to be slightly more opinionated. We don't have the complete bottoms up or throw stuff at the wall. I think it's due to our history. So when we've developed products in music, it usually involved, once you came up with the idea, you had a three-year roadmap to go and license that idea from four majors. And if you licensed the wrong thing, you lost four years. So you needed to be right and you needed to be more short because the cost of being wrong used to be so high for us.

Gustav Schöderstrom: And I don't know if it's good or bad. I think if we had grown up in a world where the cost of being wrong was just the engineering time put into it or something, you can just pull it back. Maybe we would be different, but we have a pretty specific culture where we actually do plan quite a lot more. I wouldn't say Steve Jobs, for sure. And Daniel himself actually talks all the time about distributing decisions, but it is more opinionated. And then for Pulse, we're lucky enough to have Gimlet and all these studios in-house with lots of fantastic creators. So we get to test this internally and we use them as an internal inspiration. And sometimes they are the product owners because they represent the user needs.

Sonal Choksi: That's fantastic. Connie, more thoughts on interactivity. I feel like you live in this world and you talk so much about China apps and what's possible when it comes to interactive audio.

Connie Chan: So another interesting thing about creators that comes from looking at what's working in China is not just giving them feedback on what the audience wants to hear next or what the audience is thinking, but also separating your average listener from your super listener, the person who really wants to even pay you directly for your work and helping you identify who your real true fans are, right? If you think about the creator economy, very clear trend that's already been in Asia for a while now.

Connie Chan: So something like a QQ Music, which is the main music app that people are using in China. If you have someone who's hosting a radio show or a listen together type of group chat, there's the option to basically be part of their paid fan club. And then if you're part of their paid fan club, you get a different badge on your own profile. You get access to exclusive virtual gifts that you can send that post. So everyone knows that you're part of that paid fan club. You can get a different announcement when you enter the room, different kinds of bonus check-in tasks. There's a bunch of new features that get unlocked if you're part of this creator's fan club. And ultimately what that allows the creator to do is monetize better than just a traditional advertising route. Because in addition to receiving normal virtual gifts from their listeners, from anyone who drops in and participates, you also are cultivating your small following of super fans who really, really love you.

Sonal Choksi: I love that you're pointing that out because it's basically making this link that these tools and features are not just about getting more information or data, but actually their paths to monetization as well, which is super interesting.

Connie Chan: Well, it helps you create your own empire in a different way. Like one feature I love is this battle feature where you can almost battle another radio station at the same time and almost compare how many gifts each of you are able to aggregate in a certain period of time. It's like duets with an audio challenge. It's really focused on how to help creators motivate their community and build that core fan base.

Gustav Schöderstrom: So one of the things I think is really interesting with these things that you mentioned, they're dependent on actually having a logged in service so that the creator can understand their audience. That wasn't really possible over the previous protocols. You got download numbers, but you couldn't really understand your audience and who is your super fan, what they look like and who they are and where they live and so forth. RSS protocol doesn't actually support feedback to the creator. It's a one-way broadcast protocol. But because we're now full stack, we can start doing these things that have happened in other industries.

Gustav Schöderstrom: And the thing that happened in video and in many of these other things — like you take text messaging for example — it used to be standardized and innovating on that text messaging protocol needed a ton of carriers to sit in different forums and agree, right? So the benefit was ubiquity and reach, but innovation was really slow. And then at some point something like Snapchat happened that verticalized the whole thing, and WhatsApp and so forth, and innovation just ran away. One day you had disappearing messages, the next day you had stories, the third day you had lenses because it didn't really have to wait. So I'm really excited about that happening to audio.

Sonal Choksi: Yeah, this is what we mean when we say very early inning of audio.

Connie Chan: Exactly, but there was a technical foundation that needed to exist that does exist in China to your point. They're all vertical.

Sonal Choksi: I'm very obsessed with and a student of the history of innovation. And to me, this is the classic arc from when you go from a utility layer to a value add layer. And of course, there's a lot of debates around what platforms should and shouldn't have control over. And that's something that's playing out a lot with crypto and a lot of other discussions. That said, I think the point you're making, Gustav, which makes it less academic and more interesting to users is it really comes down to you are giving me something I can't get right now.

Gustav Schöderstrom: Yeah. If you have one app that can give you a vertical solution, basically give you everything you want, that app's true understanding of you is very strong and its ability to personalize things towards you is higher. Your ability to create a profile that you then are proud to share with other people or that you want to build upon, whether it's earning different levels or different points, that also increases.

Connie Chan: I mean, I love what Gustav is saying about how things are more vertical. There's a lot of benefits when you take the super app mentality. And a super app is basically a product or a platform that focuses on all the different needs a particular customer wants versus giving a single feature solution. Recognizing that, oh, this person loves listening to these kinds of music, but this person also probably loves listening to all these other things. So why not let's offer this all in one package? We now better understand that listener and we can solve more of their problem.

Gustav Schöderstrom: So we were actually quite inspired by the super apps of China. When we thought about podcasting, the obvious solution, if you're going to build a podcasting app, if you come from a pure design angle, is to build a standalone app. But the trade-off then is distribution. And so we looked at it more from a super app point of view. Then we realized that what users actually wanted was all of their audio, which they used to have on radio, music and talk and so forth mixed. We had a zero user base in podcasting, so we'd be starting from scratch. We had hundreds of millions of music users, and that's an advantage in itself. But more importantly, we understood these users. They were logged in, and so we could just augment their moments.

Gustav Schöderstrom: And one of the interesting things we found was that it turns out that your music listening is actually very predictive of your podcast listening. You can probably guess a person's age range from their music listening alone, right?

Sonal Choksi: Yes, you can, for sure.

Gustav Schöderstrom: So you're saying people's music listening predicted their podcast taste?

Sonal Choksi: Yeah. When you want to cold start a podcast listener, it turns out that your music listening is actually a really good signal for that, for which podcast to recommend.

Sonal Choksi: That is incredible to me. I just think people's music listening is so much more visceral and less intellectual that I'm just so shocked by that fact.

Gustav Schöderstrom: I would not say it was obvious to me either, but it's a very clear result. It also supports the idea of the audience. You should think of them as one person, right? And try to serve them in the different needs they have.

Sonal Choksi: Yes. Think of the customer as one person.

Sonal Choksi: Right. What you're basically both really saying is when you think of the super app mindset, it's a cohesive identity of a user's needs. And in fact, if I were to visualize it, I think of that classic Da Vinci Renaissance man where you have this person at the center and then you have multiple spokes of interest radiating around them. And then you think of each of these moments in their day, it could be time, it could be interest, it could be need, it could be whatever job to be done to use a Clayton Christensen framework and that you've referenced a few times, Gustav. But what you're both also essentially saying is that a super app, once you have one, is built in distribution. And so you'd be silly not to use that base and do a cold start.

Gustav Schöderstrom: Yeah, it's much easier to say, let's put a competing team over there and let evolution take care of it. They build their own app and they compete. But it's at the cost of the user to do it that way. And so the first thing we did was we figured out that instead of having the apps be as different as possible, you actually wanted to have them be the same thing. And you can say that radio has always done this. People have been mixing these mediums. So it didn't seem that far-fetched, but it wasn't clear.

Gustav Schöderstrom: And if you optimize for ease of implementation, you have small things such as just the fact that the UI has to change from skipping a whole song when you're listening to music to all of a sudden skipping 15 seconds back and forth and scrubbing within a podcast. That's a big challenge to solve dynamically in the same UI. It would have been much easier to just maximize the two different hypotheses.

Sonal Choksi: So basically what I'm hearing is even something as seemingly mundane to the user as the ability to scrub forward 15, 10 seconds, which I do all the time in my podcast. If you're in music, you can just skip an entire song forward. And even that kind of trade-off is actually really complex when you're doing it in the same UI. That's super fascinating.

Gustav Schöderstrom: Exactly. So the UI has to be much more dynamic. I mean, even how you show a track versus an album cover, right?

Gustav Schöderstrom: Or a podcast episode versus the podcast cover. Like, it's a very different thing. It's not easy to pull off. And it gets harder and harder the bigger the company is. Because it requires real changes that are top-down, that have to come from leadership. It's a change in your org structure. It's a change in your release cycle. It's a massive change that's very hard to pull off. It was painful. We needed to, quote-unquote, force. It's not like people didn't want to do it, but you needed to get people to work with each other instead of putting in a different team. And it certainly needed global prioritization from Daniel down. And we have this system to prioritize things globally called Bets Board in Spotify, which was very helpful to get these things through the company. And I don't think if we had that global prioritization tool, we could get this through the company. It's very hard to do. But this is the benefit of software, right? And this is one of the benefits of being full stack. We can actually try to solve these problems and actually improve the consumer experience.

Sonal Choksi: So let me ask you guys a quick question, especially given Spotify worked within the existing UI to blend from music to podcasting. Where do you stand on the definition of podcast, music, audio? I always talk about how audio is a huge category. Like, I honestly think trying to homogenize audio is like trying to homogenize text. It's like a word is the same thing as a book, is the same thing as an article, as a blog post, as a tweet. That's ridiculous. However, Connie, you made the argument in our podcast about podcasting with Nick Kwa, how podcasting and music, and I agreed with you as well then, that there's a big difference between the spoken word and the sung word. And so I'd love to hear your guys' thoughts on where are we today?

Connie Chan: Radio is the integration of both talk and music. They live very symbiotically together. And if you look at most podcasts, they have a music introduction already. There are sound effects in a bunch of them too. So this combination, or this belief that normal talking can be improved with music, or music can be improved with talking breaks, has been here forever. But even then, where does and doesn't the blending of music and podcasting actually work, and where does it fall apart?

Gustav Schöderstrom: Right. So we had this intuition that people wanted their music and their podcast in the same app. And that certainly turned out to work. But there's a category where they're actually related. It is the same session, right? So this is the thing that we just released. So now we are going to let creators do this new type of session where they can mix talk with licensed music in a seamless session. So you see these two user needs. If you take the Clay Christensen approach, you see podcasters really wanting to use and talk about music, but they can't because the creators do not get paid for some burnt-in song in a podcast. And then you see the music creators that would like to talk about the music. So you have both of these sides at the same time. And it's been really hard to solve it, especially if they were two different apps. But now it feels very natural that you should be able to have this new type of show.

So you've seen us play around with things like Daily Drive, for example, for a long time, where we mix talk and music. And we've seen a lot of success. People love hearing their news and then their new music in the same session instead of, especially when they're driving, trying to switch to the music session and hear the new releases as well. But so what we were thinking now is we want to enable anyone to do that. And on the consumer side, it is neither a podcast or a playlist. It's just the best of podcasts and the best of playlisting. But it is neither because podcasting has the problem that you actually aren't allowed to feature music in it. And playlisting has the problem that you actually can't comment between the tracks. So we created this new format where you can do some talk, then you can add a Spotify track in there. Then you can do some more talking. And so the user can then listen to the talk part as if it was a podcast. They can listen to the track, they can skip the track, but they can also save the track if they like it. One of the things that radio always missed. So it's a new format, but hopefully it's not new in the bad sense that you have to learn anything new. It should be just like listening, just that it works the way you kind of always want it to work.

Sonal Choksi: What would you call this new format?

Connie Chan: I think very broadly of, again, I mentioned how audio is heterogeneous as text. So it's ridiculous to use one word for everything, but it is a new kind of audio experience. It's not a podcast. It's not a music or a song. I think of this as going back to radio. For me, this is the new radio station.

Gustav Schöderstrom: Yeah.

Connie Chan: This is the new way you can listen together. In a sense, a very obvious innovation, but also an innovation that requires tons and tons of licensing work over many years and a big investment in podcasting and creative tools and so forth. I'm smiling because it's going to open the door for a whole batch of brand new creators, people who don't want to host the podcast and talk the whole way through, but now can use music as their passion, as their content, as the thing they're kind of anchoring their talk around. And then this also brings about curation, social discovery. I mean, I can even think of several A16Z colleagues myself that I think would be really good.

Sonal Choksi: That's what I'm hoping for. I'm hoping for you, Connie.

Gustav Schöderstrom: I think she means Anish because Anish is a side DJ.

Connie Chan: My stuff will all be probably Chinese music.

Sonal Choksi: We want that tip.

Connie Chan: Yeah. But the point is, it really opens the door to new batches of creators. And it brings in social discovery. And it brings in the idea of curation. It's back to kind of the Spotify playlist, but with more color, right? And with more storytelling.

Gustav Schöderstrom: Augmenting, I might even argue.

Connie Chan: And the interaction that you can have with the listener, right? In Asia, you can have people order different songs and pay to try and see what's already on the playlist and change that playlist even in real time. So the kind of interaction you can build on top of this is also exciting.

Gustav Schöderstrom: You spoke about augmenting there, and I think that's a great point. So we spoke about TikTok, and I mentioned this pattern of taking sort of commodity licensed music and letting your users make it unique. So one way to think about this is it's a similar pattern. And we've had tremendous success by letting our users work with the music catalog and playlist it. They created billions and billions of playlists that have helped them. It has helped other users, but it has also helped our algorithms to learn, right? So you can think of this as a similar pattern where you take the commodity catalog, but you let any creator through Anchor work with it and make it more unique and unikify it, right?

Sonal Choksi: I love it. Unikify again.

Gustav Schöderstrom: Well, the other interesting point is when Eugene and I talked about TikTok on this podcast, He did bring up that one of the big unlocks, as minor as it might seem for the remix culture as well, was the ability to quickly license combined with the creator tools, combined with the distribution, so that you do then get, quote, this creativity network effects flywheel, which sort of then reinforces.

Connie Chan: Yeah, it's a big way that people are interacting with music on the QQ Music app. When you tap into radio stations or listen together, you see all these different hosts and you can listen to them live. When you're listening together with other people, you can choose different topics or categories like friendship, music, emotions, talk shows. And the interactions that you already see happening on these radio stations or listen together. There's a chat that's usually going on while people are listening to music. There are different leaderboards for these different creators. You can have different tasks that the creator asks you to do. You can order songs. You can see what's next on the playlist. You can gift the creator and thank them for curating this kind of music. And you can even subscribe to their fan club, right? Like if they always have great music choices, you can make sure that you're always able to know when they release something new or when they go on. So it does unlock a brand new batch of creators that today don't live on YouTube. Today, they're not podcasters, but they have a lot of things to say and they love music. So a lot more people will be able to participate, be creators themselves, build a following and eventually monetize.

Gustav Schöderstrom: I agree. The increased participation of new types of creators is really interesting because there are all of these creators who clearly want to talk about music. And there are all of these artists who they've always wanted to be on radio. Like they want to be featured by someone, but business models is often a problem. No one has been able to solve that. Both parties actually get paid for that. We solved what I think is a harder part, actually, of licensing all the music in the world and paying rentals to all organizations. We already solved that. So it feels like a very natural product for us to play with.

Connie Chan: When I was growing up, I used to listen to radio shows. I used to listen to Delilah. And she would have stories in between. And then she would have audience people call in. And then she'd have a nice, soft music to go with that story. Exactly. It was fantastic. And then you probably recorded the tracks, right? Because you really wanted the music. And that's how I discovered music too, right? And that's how she could also resurface music from the past. Rather than having us listen to only stuff that was released in the last 18 months, let's resurface some of these oldies. And this is potentially a great way to do that.

Sonal Choksi: What's really fascinating to me about this is it's almost like a vector to social, because there's nothing more inherently social than music listening and music sharing, as you're noting, of playlists, music curating. And to your earlier points about it unlocking creators, one of my favorite podcasts actually is Song Exploder by Rishi Krish Herway. I actually think I heard about this podcast from Eugene actually like a year ago and now it's going to be a Netflix show. And he really deconstructs these songs on air. But imagine all the people, like all the kids, all the adults who just lie around listening to music, talking music with their friends, bonding over music. So to me, what's really fascinating here is there is a social vector, both socially and parasocially with acquaintances and strangers when you think about then connecting with fellow fans of those playlists and other people. So I think there's actually a really interesting vector to all that too. Because TikTok is not a social network, but this theoretically could be.

Gustav Schöderstrom: So this is an interesting point. We think about Spotify more like YouTube and TikTok than Facebook and Twitter. It's actually not about following your friends, but I think you're right. I think there are so many creators out there who would love to tell a story about a specific piece of music, right? Their own story, some story or something. We'll see how it gets used. I'm hoping obviously that many artists would like to tell their story of their own album that they released, for example.

Connie Chan: Yeah, that's amazing. There are many different things that could happen. Even in that great example where the artist is telling the story, the artist doesn't have to sign up and say, okay, I'm going to start a brand new podcast. That is such a big responsibility and commitment to take on. And now you kind of have these kind of-

Sonal Choksi: A Trojan horse is starting a podcast.

Connie Chan: This really lowers the bar of commitment for creating a show. And you can try it with no real consequence and get that distribution too.

Sonal Choksi: Okay. So now let's then talk about how do you solve, this is like the big elephant in the room and potentially the big exciting thing in the room, recommendation and discovery. How do you then think about that side of this, both in the context of Spotify shows and also beyond? We open this conversation about what has and hasn't changed. This has been a broken problem quote in podcasting. It might not be as broken in music. We've talked about TikTok. We've talked about the parallels and differences between video. Let's bring it all back together around this theme and topic of recommendation and discovery.

Gustav Schöderstrom: For music, there is a commitment of more than two or three seconds to figure out if you like a song, right? So the bar for who you trust as your source for who's giving you that recommendation is higher. And so you either have to have a system that builds trust, showing that their algorithm has given you enough hits. Like TikTok can't be wrong five times in a row. Stakes are really high. So you either have an algorithm that is so good, that knows enough about you already, that the majority of the time when they give you something, and you like it. Or you have a creator that also has that same kind of hit rate that you realize, hey, most of the stuff that that person likes, I also like. That is also a great way to kind of get that discovery element. It's all about giving the user this end trust that they're willing to test your recommendation because say 80, 90% of the time, you're going to be right. So I think you're completely right. That was a success with user playlists. There are literally many billions of different curations of the Spotify catalog. So you literally have something for everyone. And either they find that playlist or you can use machine learning to learn from that to be able to serve users.

Then you have the UI elements themselves. And I think that's different between music and podcast. Music is easier in a sense because it is three minute items and you can skip through. And what we see in music is that it's like the investment of how much time you spend versus finding one gem. So it is actually okay if even most of the songs theoretically are not that good. If they're easy to skip through, and like the seventh song is like your dream song because that can make your entire week or maybe month, right? So I try to think about, I think Chris Dixon said this, a fault-tolerant UI. If your machine learning is perfect, you only need to show one item. If your machine learning is one out of 10, you probably need to show 10 items because then there's always one jam on the screen. You have to adapt your user interface to your kind of level of recommendation. And so these playlist formats, we try to think of as kind of a GTD, get things done. Can you quickly go through and like, yep, that was perfect. Save that to my library. It's like a productivity flow in the discovery moment, which is very different from the consumption moment when you may be on a speaker and then it's not okay that you have three bad songs in a row, but it's okay if the fourth one is good. Does that make sense?

Sonal Choksi: That goes back to modes, actually, thinking about the mode the user is in.

Connie Chan: Yeah. I also think if there are good mechanisms in there for the creators to have potential financial payoff from participating, the creators are actually going to be incented to have discovery. That incentive is actually built in because you cannot have thousands of concurrent Spotify shows all showcasing the same music. No one's going to want to listen to that. And so all these creators are naturally going to be incented to showcase you something brand new because what they're really being valued for is their ability to curate and then match that with the storytelling. Let me give you a concrete example. When I go to the gym and someone is trying to do a workout and they're talking through and they have music spliced in between, or just think about a yoga class. They want that variety of music. They don't want you to be listening to the same thing time and time again. And now even that gym workout, that yoga class could exist as a Spotify show where they're making you do pushups and counting down. And then there's music right there in the background. You have to really think what this can unlock.

Gustav Schöderstrom: I'm definitely hoping for that yoga and pushup workout to happen. You have to make it happen.

Sonal Choksi: Okay, Connie, so either you make a yoga show or you do like a Chinese song playlist.

Connie Chan: No, but the point is like there's so much context that can now be wrapped around recommendations. Like even the time of day, what are the right kinds of shows that work for the morning? What are the right kind of shows that you want to wind down to? Those creators will have the incentive to naturally pick what they think makes sense for you.

Gustav Schöderstrom: Exactly. So I think there are two things that are really interesting here. So one is, when we think about machine learning overall and recommendations from a product point of view, and this is completely borrowed from Andrew Ng, by the way, so it's nothing that we came up with, that we try to use is, if you think about what algorithms do really well, they tend to scale really well. They tend to be able to personalize at an okay level to hundreds of millions of people. Humans don't do that really well. Humans are incredibly smart and creative, though, but they don't scale so well. So one way to think about this that I think Andrean coined was to let the editor, for example, or the creator, if we're talking a Spotify show, but an editorial playlist, this algoritorial principle that we use.

Sonal Choksi: Algorithm plus editorial.

Gustav Schöderstrom: Exactly. Algorithm plus editorial that we call algoritorial. You literally think of the editor as the product owner. This is the product person that has the idea and the hypothesis, and they come up with what the job to be done is or what the hypothesis is or what the use case is. So for example, you take something like songs to sing in the car. No machine came up with that idea. It was a human who sat and said like, I think there a user need here. People want to scream their lungs out when they're driving to work. So how do you teach a machine this? The algorithm doesn't understand what songs to sing in the car means. Is that like a bit of 80s music? Is it a bit of movie music? But for a human, it's super clear. Like, this is a song to sing in the car. This is not. So what the editor does is they literally create like a playlist of a few thousand tracks. And then the algorithm can understand it and they can personalize it to 300 million people and scale it, right? So the job of a product owner is to create this data example, this data wireframe. I think it's very useful. That loop has been very useful for us.

Sonal Choksi: So basically bonding the best of human creativity with the best of algorithmic scaling in order to deliver on the personalization and recommendations to a mass of users.

Gustav Schöderstrom: Exactly. Humans have to come up with the ideas. They have to show the ML system what that idea actually looks like for the ML system to understand it because the ML systems are great at scaling, but not great at coming up with new ideas.

Sonal Choksi: Can you give me a little bit more color on some of the challenges here? I'd love to hear about how you have to think about solving them, what's hard about auditorial, but then more specifically about how you had to negotiate that when you transitioned from music to podcasting, and then now in blending the two. I want to hear a little bit more color about it, basically.

Gustav Schöderstrom: So in music, we have really two sources, traditionally, of recommendation information. One big source is the playlists. The other is editors. But then we have the third way, obviously, which is the engagement from the users, listen and skips and so forth. Those are the signals in music. But music is different because the items are three minutes long. Like we spoke about, it's more like TikTok. Then you go to podcast and it's like maybe one and a half hour and then you get one skip. It doesn't fit at all with like, let's just feed the machine, right? It's very low signal. So we had to think about it completely differently. But not only is it much further between the skips, we don't have anything equivalent to a billion playlist. So we had to go back and start working with quote unquote more old tech, like knowledge graphs. You have other advantages in podcasts, which is there's actually information in the audio. You have other signals, you have show notes and you have the transcripts on the shows. So we started working with those technologies instead to get some understanding. So actually, these two stacks are quite different. We certainly could leverage a lot of learnings, but they're not the same thing because they're such different objects. Especially because podcasts are usually multiple people on a podcast. There's oftentimes a host and a guest. You actually don't know who people are following sometimes. You don't know. If there's like a Joe Rogan talking to Elon Musk, you don't know if it's because I like Elon Musk or if I like Joe Rogan. That's quite different than music where there's a bunch of artists, any song they put out, I'm going to like, I'll take a listen to. It's like a cult of personality show because you're following the host in that case. In this case, you're following the artist.

Sonal Choksi: But one thing that I think is really interesting when talking about the knowledge graph is the mood graph. I always talk about coin the phrase when I assigned an op-ed on it a number of years ago at Wired because I actually think we're missing a huge opportunity in optimizing things. Frankly, my playlists are all organized by mood and emotion. They're not organized by any other criteria.

Gustav Schöderstrom: That's a great point. And in music, that is one of our biggest vectors, like one of the biggest sections of editorial playlists are the mood playlists. You're completely right.

Sonal Choksi: Ooh, that's great.

Connie Chan: It's interesting you bring up a knowledge graph, Gustav, because it's tough to know, is it a book author? They're just listening to every single podcast they're on. Is it a content thing? It's so complex and multidimensional.

Gustav Schöderstrom: Exactly. And the answer, as far as we can see, is it's all of the above. There's personality cult. There is you following a certain guest around all the podcasts that they visit, there's interest. It's just going to computing. I don't care who's talking right. So you really need this knowledge graph with all of those dimensions. And then you need to be able to let the user kind of traverse along these different dimensions. And then you can lead them to some discovery. You remember this debate around music. Everyone had a music friend that influenced them. And for a while, early Spotify, we invested heavily in social to try to replicate that. But it turned out that most of your friends on Facebook, they don't inspire you so much musically. If you average them, it's just a US billboard. So we take the same approach in podcasting. I mean, we have a core belief that if Spotify can make you discover something that you wouldn't otherwise have discovered, it will be more important in your life. So we really try to make sure that we measure and understand how many discoveries we generate for you.

Sonal Choksi: It's almost like a new metric of return on discovery. Instead of return on investment or return on energy, if I think about every app, what is my return on discovery or ROD on that particular platform? I'll borrow that from you.

Gustav Schöderstrom: But another difference from these things is that we are, revenue-wise, mostly a subscription service. So in machine learning, in the practical world, there's been a lot of deep learning and so forth. But in the academic world for a long time, there's been a lot of focus and discovery and exciting results around reinforcement learning. But AlphaGo and all these things.

Sonal Choksi: Yeah, we've actually talked about it on this podcast quite a bit too.

Gustav Schöderstrom: And not to go through it, but the main idea is just you look for some long-term reward and you backpropagate it through time. Instead of looking at what is the most likely next click. And so I think if you have a service that is free only and you have an average engagement, same every day, it's going to be really hard to backpropagate signal. It's going to be noisy. But if you have an event four months down the line, that is, I went from just consuming ads to paying $120 per year, you have this massive amount of sort of gradient that you can backprop through time.

Sonal Choksi: Oh, I love this.

Gustav Schöderstrom: And the thing that is different between, for example, YouTube or TikTok is, every month, all the paying users, hundreds of millions of them, they go and they evaluate. It's like, should I still pay? And they vote with their wallet, regardless of how much they actually consume. So we have a different signal that is not just engagement and consumption and attention. We can see, do you keep paying? And obviously, as you know, it's not really possible to do the real reinforcement learning. You basically need a perfect simulator of the world, but you can approximate it quite well. And so that's something that is happening in the rest of the industry as well, slowly, you need enough signal for that to really be valuable. So that's something I'm excited about in the recommendation space.

Sonal Choksi: What you're basically saying, I talk about this quite often on the podcast about how subscription models change so much. But what you're saying, which is so fascinating to me, is that it's also a way to get much better signal into your system. You're also basically saying you're essentially weighting higher people with more skin in the game, which is exactly how you want to design something.

Gustav Schöderstrom: Exactly. Everyone has saves and likes, but you can think of like paying $10 as a super big like.

Sonal Choksi: Yes, exactly. You're weighting it higher. And you have that data because people are logged in and they're streaming. One of my favorite books is James Kars' Finite and Infinite Games. And he just died, actually. Rest in peace, James Kars. But the idea, what you're saying is you're playing a repeated game with your users, which then gives them an even better game board to play on versus a transactional game only.

Gustav Schöderstrom: That's exactly it, which is a big problem that is important to solve, I think. And you can try to understand what the user actually values long-term versus just in the moment.

Connie Chan: Yeah, and subscriptions are a fantastic business model. But also, I can see how that would allow new revenue streams for these creators. And I'm not just talking about the people who create the music, but I'm talking also about the people who are going to create and deliver a brand new experience that lives on top of the music. If those people can find some kind of financial payoff from participating, that's a brand new revenue stream. And then think about the possibilities, the kind of interaction you have with that listener at that moment is another area you can charge for.

Sonal Choksi: I also love that while we've talked so much about putting the power back for creators, it really does actually most empower the listener. Just one quick question, Gustav, how do you think about the tension between data and all the data you're getting and all the signals and where it goes too far? Like, is there a risk that sometimes in listening to your users, you're missing out on what they don't tell you? And how do you think about that as a head of R&D at a company where you're not just abstract R&D, you're actually building product?

Gustav Schöderstrom: Yeah, I think that's a fantastic question and really hard to answer. It is an age-old problem. I think one way to think about it is to simplify a little bit. Algorithms, they kind of look in the rearview mirror and draw a straight line into the future. And so that's great for a while, but product development, usually good product development is based on some sort of ideally contrarian hypothesis. And your machine learning is not going to come up with a contrarian hypothesis, right? So you need some mechanism for that to happen. And so we try to think of this in different ways. I mentioned the algatorial, where the editor actually has the ability to say like, no, I believe in something different. So we try to build in this mechanism where humans can go in and, you know, they have the steering wheel, they can take a left turn or something, and then the algorithms follow. And, you know, there are incentives to not do it. It's always going to be safer to keep going straight for a while more. Why take risk? All of these things, right?

Gustav Schöderstrom: But back to playing infinite games. If you play the game many times, think about it as game theory, you have to end up in a place where the optimal thing is to try new things every now and then, to try to cover as much space as possible. And as I said, we have a culture of being quite specific in the hypothesis we have. And we try to think about it, as do many companies, sort of a portfolio. I want to have some things that are quite contrarian and have a pretty high chance of failing, whereas I want a bunch of things that are obvious. But that balance — no one has the perfect solution, but everyone at some scale has to start thinking about it.

Gustav Schöderstrom: And so we found a few mechanisms that were useful for product development. One was to take the concept of simple prioritization and the Kanban board all the way to the C-suite. Everyone thinks they're good at prioritizing, but they're not. And I bet that in most companies, the C-suite is the worst at prioritizing. They actually want to do everything. And so we have something like five to seven things that the company needs to do. And Daniel owns that. But the one rule is two things cannot have the same priority. It reminds me of the Steve Jobs bio anecdote where at one of their off-sites, they put a whole list of things and he literally crossed everything off the list and they only did the first four.

Sonal Choksi: What you're describing though is not just siphoning off what to do versus not to do, but what to order the priority from the top so that the managers don't have this friction and they don't waste in terms of building things. And that's the trick.

Gustav Schöderstrom: Yes, I agree. And the other thing that I think is fascinating about that is that when you say that Daniel kind of owns that too, when you are disrupting yourself, so to speak, like when you went from music to podcasting, putting that higher up on the bets board in his office is like, "Hey, no complaints, guys, this is it." So that's exactly what happened. Podcast was the number one company bet for two years and everyone in the company knew it. And so what happens if you don't have that? You push that decision to managers and you create conflict in the org. The truth is Daniel can't have any idea in a company of thousands of people what is going to clash with what resources.

Sonal Choksi: Of course.

Gustav Schöderstrom: The only thing he can do is like, when you clash, this is the priority. I love that as a management thing. It's so simple. Everyone thinks it's complicated. It's actually very simple. The discussion is hard. Actually prioritizing is very hard.

Sonal Choksi: Okay. So we started with talking about where podcasting has been. We've gone through what's shifted, parallels and differences between video and music. We've talked about the trend of interactivity in augmenting audio in different ways. We've talked about recommendations and hearing like an algorithm and even an editor. What do you guys think is sort of the future of a lot of these? Like, where do you think the future is kind of going?

Sonal Choksi: My guess is that if we use the cheat sheet of other media, I think audio is going to increase on the creator side, just like the other mediums. I think it's going to increase in numbers of creators. The market for audio is bigger than I think people realize, or as Connie said earlier too, we're still in the very early innings. So my obsession is this two-word phrase that I use all the time: world building. And to me, one of the missed opportunities in audio for a long time — and, Gustav, you painted this range from gaming models all the way to music models to different things — I actually think we're starting to increasingly see more game-like behavior in audio. And I'm so excited for that kind of world building, but it's a very different kind of world building because audio has an immersiveness that's very different than the visual-based world building of other worlds.

Sonal Choksi: And so I'm super excited for what we can do. I mean, I already think about our expanding podcast network as a form of world building. And when you mentioned Spotify shows, that to me is another form of world building because you're essentially bridging different worlds and creating new experiences. So to me, that's actually the thing that I'm most excited about.

Gustav Schöderstrom: So I think that's a great way to think about it. And you think of the music world, the podcast world, and now you can think of this new world where you can mix them and then you can have other worlds. The thing that I think is going to happen is you look at something like audio and it's so easy to create — it's even easier to create than video. So as we both make it even easier and lower the friction for everyone, we let creators make more money and we add these new formats. What I'm hoping is that that market is going to grow as well, just like we've seen the market for creators grow in other medias.

Connie Chan: I think audio will be further optimized in the sense that you can almost peel apart the different nuggets of a podcast. You can take certain segments now, you can take a commentary around it now, and you're going to be able to do new things when you break apart a song, when you break apart a podcast, and you can see what that will unlock. TikTok is breaking apart a song and kind of getting to a specific five, ten second slice of it, right? Snip it. And then this idea of now taking something that used to be one piece of content and chunking it down to different things now gives you new building blocks to build new kinds of shows, new kinds of interactions, which means things will get much more participatory.

Connie Chan: More people can become creators. More people can probably become listeners. More listeners will find each other. Listeners will become stronger fans of their creators. So I think there's a very hopeful, very optimistic future where now technology actually can help everyone win.

Sonal Choksi: That's fantastic. I love that. Gustav, Connie, thank you so much. Thank you for joining the a16z podcast. This was super fun.

Gustav Schöderstrom: Super fun. I wish we could all talk for hours. Take care everyone. Bye.

Connie Chan: I should put in a plug for my Spotify show.

Sonal Choksi: Do it.

Connie Chan: The China Song Show, Connie. It's going to be huge. Bye guys. Have a really good day or evening for you. Take care everybody. Thank you.