Spotify Co-President Gustav Söderström on their future with Generative AI

00:47:50
https://www.youtube.com/watch?v=jV4MinWX39o

Ringkasan

TLDRIn an insightful discussion, Gustav Sodor, Spotify's co-president and Chief Technology Officer, shares insights on AI's transformative role in music creation and consumption. He emphasizes that AI should be seen as a tool that enhances creativity rather than a threat to human musicians. The conversation also delves into Spotify's mission to provide personalized music experiences through features like Discover Weekly and the AI DJ. They highlight the platform's focus on balancing algorithm-driven recommendations with user empowerment and feedback. Furthermore, Sodor addresses the challenges of podcast discoverability and Spotify's strategy to integrate various audio formats into a unified user experience, ultimately aiming to create a more engaging and enjoyable listening environment.

Takeaways

  • 🎧 Spotify is blending music, podcasts, and audiobooks into a unified platform.
  • 🤖 AI is a tool that enhances creativity rather than replacing musicians.
  • 🔍 Discover Weekly acts as a 'friend' for personalized music discovery.
  • 📈 Spotify aims to improve podcast recommendations for better discoverability.
  • ⚖️ The platform is addressing copyright concerns while fostering creator support.

Garis waktu

  • 00:00:00 - 00:05:00

    The host welcomes Gustav Sodor, Spotify's co-president and Chief Product Officer, emphasizing their appreciation for Spotify and its role in their daily music experience. Gustav expresses excitement about discussing Spotify's functions and innovations, particularly in relation to AI technology.

  • 00:05:00 - 00:10:00

    Gustav shares his thoughts on AI-generated music, highlighting its potential as a tool to amplify creativity rather than replace artists. He compares AI tools to past musical innovations, saying technology has progressively allowed more people to engage in music creation.

  • 00:10:00 - 00:15:00

    When asked about the inclusion of fully AI-generated songs on Spotify, Gustav states that as long as creators follow copyright laws, Spotify aims to support them. He also emphasizes the importance of developing a business model that compensates creators fairly amidst the rise of AI-generated content.

  • 00:15:00 - 00:20:00

    The discussion moves to the nature of AI in the music creation process, with a focus on blending AI with human creativity. Gustav argues that AI can enhance artistic expression while maintaining the need for human connection in music to build identities and emotional ties.

  • 00:20:00 - 00:25:00

    As the conversation shifts to AI-driven recommendation systems, Gustav elaborates on Spotify's goal of becoming a more integrated and ambient platform that understands user contexts. He highlights Spotify's aim to evolve AI recommendations into a more interactive experience, knowing users on a personal level.

  • 00:25:00 - 00:30:00

    The dialogue addresses the balance between algorithmic recommendations and user agency on Spotify, with Gustav acknowledging the different user types and emphasizing that Spotify aims to cater to all. He notes the importance of user feedback to refine this balance.

  • 00:30:00 - 00:35:00

    The host brings up concerns from users about algorithmic recommendations feeling like a 'bubble.' Gustav responds by stressing the need for Spotify to provide diverse content and to help users discover music that falls outside their usual preferences, enhancing enjoyment.

  • 00:35:00 - 00:40:00

    With a shift to podcasts, Gustav discusses Spotify's strategy of consolidating music, podcasts, and audiobooks into one application for a seamless experience. He underscores the need for efficient distribution and user engagement, reflecting the growing interest in audio content.

  • 00:40:00 - 00:47:50

    The final part of the conversation delves into the challenges of podcast discoverability on the Spotify platform. Gustav acknowledges the struggle for new podcasts to gain traction and the investment needed to retain listeners. He reflects on the challenges and expectations for improving the discoverability of various shows.

Tampilkan lebih banyak

Peta Pikiran

Video Tanya Jawab

  • What is Spotify's approach to AI-generated music?

    Spotify views AI as a tool to amplify creativity, believing it allows more people to create music rather than replacing human musicians.

  • How does Spotify handle copyright concerns with AI-generated content?

    Spotify aims to support creators legally and fairly, ensuring proper compensation for those using AI in their music.

  • Will Spotify allow users to tweak their recommendations?

    Yes, Spotify is working on features to allow users more control over their recommendations, including free-text inputs.

  • What is the significance of Discover Weekly for Spotify users?

    Discover Weekly is seen as a 'friend' by users, providing personalized music recommendations based on their listening habits.

  • How does Spotify plan to improve podcast discoverability?

    Spotify is investing in better algorithms and foreground feeds to help users discover podcasts relevant to their interests.

Lihat lebih banyak ringkasan video

Dapatkan akses instan ke ringkasan video YouTube gratis yang didukung oleh AI!
Teks
en
Gulir Otomatis:
  • 00:00:00
    we have a great show for you today
  • 00:00:01
    because we're sitting here in four World
  • 00:00:03
    Trade Center spotify's New York City
  • 00:00:05
    headquarters with the company's
  • 00:00:07
    co-president Chief product officer and
  • 00:00:10
    chief technology officer yes all that in
  • 00:00:12
    one Gustav Sodor is here Gustav great to
  • 00:00:15
    see you welcome to Big technology thank
  • 00:00:17
    you for having me Alex it's a pleasure
  • 00:00:18
    to be here very to be here I mean we're
  • 00:00:20
    in a beautiful studio in your office I
  • 00:00:22
    I've been looking around I just can't
  • 00:00:24
    believe how amazing the studio is and
  • 00:00:26
    also it's cool for me to be sitting here
  • 00:00:28
    with you because I'm using your app
  • 00:00:30
    every day and Spotify is the place where
  • 00:00:33
    I touch some of the most I wouldn't even
  • 00:00:36
    call it possessions cuz I'm subscribed
  • 00:00:37
    to it but one of the most beloved
  • 00:00:39
    experiences that I have which is music
  • 00:00:42
    and so many of us use Spotify all the
  • 00:00:45
    time but we hear from you guys rarely so
  • 00:00:48
    I do appreciate the opportunity to speak
  • 00:00:50
    with you me too I try I appreciate that
  • 00:00:52
    I'm very glad to hear that and I'd love
  • 00:00:54
    to share as much as I can about how
  • 00:00:56
    Spotify actually works it's sort of a
  • 00:00:57
    passion of mine to try to expl things
  • 00:01:01
    and and how they work so I I actually
  • 00:01:03
    love these podcasts in some ways an app
  • 00:01:06
    will determine how people experience a
  • 00:01:08
    format but in some ways a moment in time
  • 00:01:11
    will determine how an app sort of has to
  • 00:01:14
    deal with the content within it yeah and
  • 00:01:18
    Spotify is going through both of those
  • 00:01:19
    both of those regard uh artificial
  • 00:01:22
    intelligence I don't know if you've
  • 00:01:24
    heard ofso in fact I'm sure you've heard
  • 00:01:26
    of it's one of our favorite things to
  • 00:01:29
    use on big technology podcast Rono and I
  • 00:01:32
    we do the show on Friday uh we built a a
  • 00:01:35
    theme song with sunno and played it and
  • 00:01:37
    it was a good time um and I'm curious
  • 00:01:40
    from your perspective running product
  • 00:01:42
    that's Spotify how do you feel about AI
  • 00:01:45
    music AI generated music because the
  • 00:01:48
    songs they're not amazing but they're
  • 00:01:50
    good there have been some big hits um do
  • 00:01:53
    you view this as an opportunity a threat
  • 00:01:55
    do you want it on your platform so the
  • 00:01:57
    way I think about I'm a technologist so
  • 00:01:59
    obviously I'm excited about the
  • 00:02:01
    technology itself and I love AI I think
  • 00:02:04
    it's a super impressive product it works
  • 00:02:06
    amazingly well and it's philosophically
  • 00:02:09
    it's very interesting that something we
  • 00:02:13
    thought was impossible just a few years
  • 00:02:14
    ago that a a machine could sound like
  • 00:02:17
    something a human did can be creative
  • 00:02:19
    legitimately incredible you prompt it
  • 00:02:21
    and out comes as great sounding song it
  • 00:02:24
    is incredible so I think that technology
  • 00:02:26
    is amazing now my interest is to think
  • 00:02:28
    of these Technologies as tools so if you
  • 00:02:31
    if you think about music it's going
  • 00:02:34
    through a journey of more capable tools
  • 00:02:37
    if if you go way back if you were a
  • 00:02:40
    musical genius um like a bck or someone
  • 00:02:43
    you literally needed access to an
  • 00:02:45
    orchestra to be able to realize that
  • 00:02:47
    genius even if you could play multiple
  • 00:02:49
    instruments yourself you couldn't play
  • 00:02:50
    them at the same time so you actually
  • 00:02:52
    needed like an orchestra and then we got
  • 00:02:54
    to record a music and you could record
  • 00:02:56
    one instrument at a time so you got more
  • 00:02:57
    and more independent and then somewhere
  • 00:02:59
    around the 80s the synthesizer came
  • 00:03:02
    along and made that and meant that you
  • 00:03:04
    didn't have to be able to play all the
  • 00:03:05
    instruments yourself you could you could
  • 00:03:07
    sort of quote unquote fake the drums
  • 00:03:09
    using the synthesizer and the guitar and
  • 00:03:11
    so forth so I think there's been this
  • 00:03:12
    progression of more more powerful tools
  • 00:03:14
    that enabled more and more creativity
  • 00:03:17
    and then somewhere in the '90s the the
  • 00:03:19
    do the digital audio workstation came
  • 00:03:22
    along and and being a swed very proud of
  • 00:03:24
    this someone like avichi came along and
  • 00:03:26
    and what is interesting with avichi is
  • 00:03:28
    he was not very Prof I at anyone
  • 00:03:30
    instrument or a singer so in a previous
  • 00:03:33
    world he would not have been considered
  • 00:03:36
    a very creative person because he
  • 00:03:38
    couldn't realize that with access to
  • 00:03:40
    this tool the digital audio workstation
  • 00:03:42
    turns out he was one of the most
  • 00:03:44
    creative people we had that we are very
  • 00:03:46
    very proud of so so for for him the
  • 00:03:49
    digital audio workstation was Steve dos
  • 00:03:51
    would say a bicycle for the mine it
  • 00:03:53
    meant that he could he get more
  • 00:03:55
    productive and he could he could express
  • 00:03:56
    his his genius and the big question with
  • 00:03:59
    this next round of tools is the same is
  • 00:04:02
    it amplifying creativity or is it
  • 00:04:04
    replacing people and I I think it's it's
  • 00:04:07
    amplifying creativity it is giving more
  • 00:04:10
    and more people the access to be
  • 00:04:12
    creative you need even less um motor
  • 00:04:14
    skills on a piano or something you need
  • 00:04:17
    less technical skills than a digital
  • 00:04:18
    audio workstation so I think of them as
  • 00:04:21
    tools and and I think there's this
  • 00:04:23
    interesting question on what is AI music
  • 00:04:27
    I think people say AI music and they
  • 00:04:29
    mean something that was prompted with
  • 00:04:31
    like not too much of a prompt and not
  • 00:04:33
    too much work so like 100% AI but the
  • 00:04:36
    truth is that much of Music being made
  • 00:04:38
    today made today is a combination I
  • 00:04:41
    think many of the big artists are using
  • 00:04:42
    AI for parts of their songs or parts of
  • 00:04:44
    the track or the drums Etc so I think
  • 00:04:46
    there's actually a scale between zero Ai
  • 00:04:49
    and 100% Ai and I think we're on this
  • 00:04:52
    this progression where it's actually
  • 00:04:53
    going to be very difficult to say what
  • 00:04:55
    is an AI song does it have to be 100 99%
  • 00:04:58
    70% 50% % but but the real question is
  • 00:05:01
    do you welcome this stuff on your
  • 00:05:03
    platform let's say somebody does prompt
  • 00:05:05
    100% AI uh spotifi could fill up with
  • 00:05:08
    songs that are AI prompted it's very
  • 00:05:10
    easy to create these songs and then
  • 00:05:12
    upload them to the internet how do you
  • 00:05:14
    feel about those do you want them so
  • 00:05:16
    there there two questions there one is
  • 00:05:18
    what are what is Spotify about we are a
  • 00:05:20
    tool for for creators and if creators
  • 00:05:22
    want to use AI to enhance their music as
  • 00:05:25
    long as we follow the legislation and
  • 00:05:26
    copyright laws we want them to be able
  • 00:05:28
    to monetize their music and payouts
  • 00:05:30
    right so for us um we are trying to
  • 00:05:34
    support creators and and uh the music
  • 00:05:37
    catalog has grown tremendously since we
  • 00:05:39
    started from tens of millions of tracks
  • 00:05:41
    hundreds of millions of tracks and I
  • 00:05:43
    think it's going to keep expanding but
  • 00:05:45
    what I think is important for for us to
  • 00:05:48
    figure out that I think is is our job in
  • 00:05:50
    the rest of the music industry is if you
  • 00:05:52
    go back to the years of piracy there was
  • 00:05:55
    this technology called peer-to-peer and
  • 00:05:57
    file sharing that was amazing you worked
  • 00:05:58
    on that early on exact that exactly we
  • 00:06:00
    actually Incorporated that technology
  • 00:06:02
    into to Spotify but before Spotify the
  • 00:06:05
    technology sort of preceded the business
  • 00:06:07
    model so it was great for consumers they
  • 00:06:09
    could now get all of this music for free
  • 00:06:11
    but it didn't work for creators and I
  • 00:06:14
    think we're in the same period of time
  • 00:06:16
    now where the technology has preceded
  • 00:06:19
    the business model so I think the
  • 00:06:20
    technology is great I do think we need
  • 00:06:23
    to find a way for for the creators who
  • 00:06:26
    have participated in this to be
  • 00:06:28
    reimbursed so that's something that we
  • 00:06:30
    are thinking about and the rest of the
  • 00:06:32
    industry is thinking about if we can
  • 00:06:34
    find the business model I think we could
  • 00:06:37
    unlock a tremendous amount so so there's
  • 00:06:40
    a separate question which is then these
  • 00:06:42
    models would the way they were trained
  • 00:06:44
    will that be considered legal or not
  • 00:06:46
    which is a legal question that is being
  • 00:06:48
    decided uh on some some time period for
  • 00:06:51
    example in the US these companies are
  • 00:06:52
    now sued so I think that question will
  • 00:06:55
    be decided about legislation but let's
  • 00:06:56
    assume that there is one of these models
  • 00:06:59
    whether it has to be retrained on other
  • 00:07:01
    data or not is that an interesting tool
  • 00:07:03
    for us if it was trained legally yes if
  • 00:07:06
    creators can participate in it so first
  • 00:07:08
    of all it's good to hear that you're
  • 00:07:10
    already thinking about issues of
  • 00:07:12
    compensating creators musicians because
  • 00:07:15
    you know I write text in addition to
  • 00:07:17
    podcasting and I know that models have
  • 00:07:19
    trained on my text previously I'm not
  • 00:07:21
    going to see a dime on that um it's a
  • 00:07:23
    little different right with music but
  • 00:07:25
    yeah if you can Channel different
  • 00:07:26
    musicians there should be I think some
  • 00:07:29
    renumeration um but I'm going to just
  • 00:07:32
    ask one last time on this point then
  • 00:07:33
    we're going to move on um So Meta for
  • 00:07:36
    instance they have ai generators the
  • 00:07:39
    feeds have I won't say filled but
  • 00:07:41
    there's lots of AI generated images
  • 00:07:43
    they're engaging meta seems to be okay
  • 00:07:46
    with this it doesn't ban it and now some
  • 00:07:49
    of the top content on a meta platform is
  • 00:07:52
    shrimp Jesus which sort of combines like
  • 00:07:54
    two of people's great loves which is God
  • 00:07:57
    Jesus and seafood and I've seen that
  • 00:08:00
    yeah it's massive these type of images
  • 00:08:02
    are massive on meta so from a Spotify
  • 00:08:06
    perspective if these songs generated by
  • 00:08:09
    AI music generators become engaging and
  • 00:08:11
    let's say they follow the rules is that
  • 00:08:13
    good for Spotify well I think like this
  • 00:08:17
    if creators are using this uh these
  • 00:08:19
    Technologies they are creating music in
  • 00:08:21
    a legal way that we reimburse and people
  • 00:08:23
    listen to them and they are successful
  • 00:08:25
    we should let people listen to them I
  • 00:08:27
    think what is different though I don't
  • 00:08:28
    think is our job
  • 00:08:30
    to generate that music instead of the
  • 00:08:32
    creators right that's a that's a key
  • 00:08:33
    difference are we as a platform for
  • 00:08:35
    creators and then we can have a
  • 00:08:37
    discussion on which tools are they
  • 00:08:38
    allowed to use like they could use the
  • 00:08:40
    or the workstation but not llm maybe
  • 00:08:42
    that's not actually we we shouldn't
  • 00:08:43
    decide that for them but there is a
  • 00:08:45
    question should we generate all the
  • 00:08:46
    music ourselves and that's where we're
  • 00:08:48
    saying no we're not going to generate
  • 00:08:50
    that music and other platforms maybe
  • 00:08:51
    will because it's it's it's cheap
  • 00:08:53
    content right so that's the key
  • 00:08:55
    difference of we decided what we want to
  • 00:08:57
    be in this world and it's a platform for
  • 00:08:59
    Crea
  • 00:09:00
    then then there's question which tools
  • 00:09:02
    they are allowed to have which is
  • 00:09:03
    partially a legal question and partially
  • 00:09:06
    up to up to the creators I think okay so
  • 00:09:08
    there's a potential world where one of
  • 00:09:11
    these tools seems to have violated
  • 00:09:12
    copyright and you might ban creators
  • 00:09:14
    from uploading music that have used that
  • 00:09:16
    tool we already taking if if we get we
  • 00:09:19
    have detection systems for if you if you
  • 00:09:20
    are um if it's a derivative work of of
  • 00:09:23
    something that already exists so we have
  • 00:09:24
    systems to take these down uh if you're
  • 00:09:26
    creating something completely new that
  • 00:09:28
    isn't a derivative of anything there
  • 00:09:30
    isn't a there isn't a a copyright
  • 00:09:32
    infringement then the labels tell us so
  • 00:09:35
    so that's the other question on like
  • 00:09:36
    what are these models trained on and
  • 00:09:38
    we're not creating this model so so
  • 00:09:40
    we're watching what happens there and
  • 00:09:41
    we're going to follow the law but I
  • 00:09:43
    think from a high level this should be a
  • 00:09:46
    very exciting tool for creators for for
  • 00:09:50
    musicians for authors for podcasters I I
  • 00:09:53
    think um I think if you look at
  • 00:09:55
    something like notebook LM for example
  • 00:09:57
    was actually created by uh a journalist
  • 00:09:59
    and a writer as a tool so I I think my
  • 00:10:03
    bet is that these are bicycles for the
  • 00:10:04
    mine but sort of bicycles for the mine
  • 00:10:06
    on steroids right and that when those
  • 00:10:08
    shifts happens there is always tension
  • 00:10:10
    between the the the people who didn't
  • 00:10:12
    use these tools who feels like this is a
  • 00:10:14
    little bit like cheating and the people
  • 00:10:15
    are saying like no I want to be creative
  • 00:10:17
    too and it's always a different
  • 00:10:20
    difficult transition period it's just
  • 00:10:22
    the story of technology and by the way
  • 00:10:23
    we're going to get to notebook LM in a
  • 00:10:25
    bit so I definitely want to hear your
  • 00:10:27
    perspective on that but let me ask this
  • 00:10:28
    one so
  • 00:10:29
    first of all what you're describing is
  • 00:10:32
    just sort of like this is what happens
  • 00:10:34
    in tech companies you think you have
  • 00:10:36
    something figured out and then next
  • 00:10:37
    thing you know new innovation you have
  • 00:10:38
    to account for that's kind of what makes
  • 00:10:40
    it exciting what makes it fun that that
  • 00:10:42
    it happens and you already have
  • 00:10:44
    addressed where this is going which is
  • 00:10:47
    do we get to a place where remember you
  • 00:10:51
    started talking about this saying we
  • 00:10:53
    never could have anticipated that this
  • 00:10:54
    is possible and now it's like feels like
  • 00:10:57
    magic prompt and you get a song
  • 00:10:59
    and I called them great earlier they're
  • 00:11:01
    not great but they're good
  • 00:11:02
    enough and this is literally first
  • 00:11:05
    generation of this stuff it's going to
  • 00:11:07
    get better and as you think deeper about
  • 00:11:10
    it do we go to a place where you can
  • 00:11:14
    start to prompt music that is going to
  • 00:11:17
    be better than any song that you might
  • 00:11:19
    listen to that has been created for
  • 00:11:22
    certain moods for instance like let's
  • 00:11:24
    say you're in like a introspective mood
  • 00:11:26
    or in a loving mood or in an angry mood
  • 00:11:29
    and you're just able to prompt it and
  • 00:11:30
    create that song that perfectly touches
  • 00:11:32
    the heart at that moment and I started
  • 00:11:34
    off talking about how this this format
  • 00:11:36
    is belove music is belov it touches the
  • 00:11:39
    heart and if AI can do that does that
  • 00:11:41
    become the future of music so you've
  • 00:11:43
    already said you don't want to play in
  • 00:11:45
    it but is that something that you can
  • 00:11:48
    discount from coming in so I think two
  • 00:11:50
    things um music is used for many
  • 00:11:53
    different things right um and so you
  • 00:11:56
    have for example music that you're using
  • 00:11:58
    to study I think is a good example the
  • 00:12:01
    extreme version of that is people listen
  • 00:12:02
    to White Noise so like would White Noise
  • 00:12:05
    be generated it's actually already
  • 00:12:06
    artificially generated it's one of the
  • 00:12:07
    top podcast formats ony so so there's a
  • 00:12:10
    scale here and I think you're right for
  • 00:12:12
    for certain things maybe create better
  • 00:12:14
    white noise maybe you could create
  • 00:12:16
    better comp uh you know always varying
  • 00:12:19
    ambient music for your studying maybe
  • 00:12:21
    for gaming maybe that music should
  • 00:12:23
    automatically adjust what's happening on
  • 00:12:25
    the screen so I think we're going to see
  • 00:12:27
    lots of AI generated music for those use
  • 00:12:29
    cases but there is another use case
  • 00:12:32
    which I think is very important a lot of
  • 00:12:33
    people use Music to build their identity
  • 00:12:36
    right especially when you're a teenager
  • 00:12:38
    you go to a concert you buy the jacket
  • 00:12:40
    from that concert why why did you buy
  • 00:12:42
    that jacket well it's it's a it's a it's
  • 00:12:44
    like a pin you're identifying with this
  • 00:12:46
    band you're building your own identity
  • 00:12:48
    through this band I don't think that
  • 00:12:50
    will work with AI generated music
  • 00:12:53
    because there is no one behind it so I
  • 00:12:55
    think some music uh and and I'm sure
  • 00:12:57
    this is happening already I'm sure many
  • 00:12:59
    Publishers are generating music for for
  • 00:13:01
    coffee tables and so forth that will
  • 00:13:03
    probably happen um but I do think the
  • 00:13:06
    human need for for having someone to
  • 00:13:09
    believe in an actual artist that you
  • 00:13:11
    care about I don't think Taylor Swift
  • 00:13:13
    will be replaced by an AI not because
  • 00:13:15
    the music couldn't sound similar but
  • 00:13:17
    because the whole point is Taylor Swift
  • 00:13:19
    and belonging to something so I think
  • 00:13:21
    it's not a it's not a binary answer like
  • 00:13:23
    is it's going to happen or not no it's
  • 00:13:24
    going to not going to happen I think
  • 00:13:26
    both both will probably happen you know
  • 00:13:28
    two years years ago I might have fully
  • 00:13:30
    agreed with you that there's always
  • 00:13:31
    going to be that need for the story and
  • 00:13:33
    the human
  • 00:13:34
    connection and now I'm not so sure
  • 00:13:37
    because because I do think that that
  • 00:13:41
    this stuff can be good enough it's
  • 00:13:43
    already proven that it's it it's already
  • 00:13:46
    exceeded some of our greatest
  • 00:13:48
    expectations and um I think we would
  • 00:13:52
    like to think that we want that
  • 00:13:53
    connection with the human but all right
  • 00:13:55
    let's go right into notebook LM but but
  • 00:13:57
    I think one thing to say that that I
  • 00:13:58
    think is interesting is what tends to
  • 00:14:00
    happen in these worlds is that the thing
  • 00:14:03
    that is scarce gets even more valuable
  • 00:14:05
    so one bet would be that true human
  • 00:14:07
    connection gets more valuable than ever
  • 00:14:10
    when a lot of what you talk to in the
  • 00:14:11
    future may be llms that that would be my
  • 00:14:15
    best I'm I'm hoping that's the case
  • 00:14:16
    because part part of the business that
  • 00:14:18
    I'm running is predicated on the idea of
  • 00:14:22
    connecting to a human who can sort of
  • 00:14:23
    dissect and break stuff down is valuable
  • 00:14:26
    so I'm hoping that is the case so but I
  • 00:14:28
    also I'm not as sure as I used to be and
  • 00:14:31
    I think it's wise to not be sure of
  • 00:14:33
    anything right now given the pr of place
  • 00:14:35
    of progress and I think that brings us
  • 00:14:37
    right into notebook LM which I was
  • 00:14:39
    planning to leave for later but you set
  • 00:14:41
    it up perfectly and it's this Google
  • 00:14:43
    product that you can put notes in and
  • 00:14:46
    then it will actually generate this
  • 00:14:49
    podcast uh with two co-hosts that sound
  • 00:14:53
    like ridiculously human yeah they they
  • 00:14:55
    don't they don't sound like robots and
  • 00:14:57
    in fact people have sort of like like uh
  • 00:14:59
    fed them scripts where they like realize
  • 00:15:01
    that they're actually not real people
  • 00:15:03
    and they're AIS and they just have this
  • 00:15:05
    kind of breakdown and it's insanely
  • 00:15:07
    entertaining but the bottom line is and
  • 00:15:09
    they're they're not quite where they
  • 00:15:11
    need to be they're still a little hokey
  • 00:15:12
    I think and just kind of they're like
  • 00:15:15
    the if you listen for a minute you're
  • 00:15:16
    blown away if you listen for five
  • 00:15:17
    minutes you start to cringe but they
  • 00:15:19
    also do a good enough job of breaking
  • 00:15:21
    things down where they can pass and I
  • 00:15:23
    started to see uh them right now showing
  • 00:15:27
    up in the second half of episodes where
  • 00:15:28
    people like we're going to do the
  • 00:15:29
    episode and in the second half we're
  • 00:15:31
    going to give you the AI to listen to uh
  • 00:15:33
    but what happens if they end up being
  • 00:15:35
    the first half and spotify's made a big
  • 00:15:37
    move into podcasts what do you think
  • 00:15:39
    about the rise of these AI podcast hosts
  • 00:15:42
    so I think notebook LM is very
  • 00:15:44
    impressive and uh you know you could
  • 00:15:48
    predict given the the evolution of voice
  • 00:15:50
    quality of these things and
  • 00:15:53
    understanding of a language model that
  • 00:15:54
    this would happen so I'm not at all
  • 00:15:56
    surprised in a sense that you can
  • 00:15:59
    generate audio that is engaging to
  • 00:16:00
    listen to talk audio but what I think
  • 00:16:03
    was the great innovation of um notebook
  • 00:16:05
    LM was that people generated monologues
  • 00:16:09
    and what what humans really respond to
  • 00:16:11
    are dialogues and in retrospect it's
  • 00:16:12
    pretty obvious like almost all podcasts
  • 00:16:14
    are dialogues like if I sat here for one
  • 00:16:16
    hour it's not that interesting so I
  • 00:16:18
    think the big hack was to to go through
  • 00:16:21
    a piece of material and present it as a
  • 00:16:22
    dialogue and prompted the right way
  • 00:16:25
    there was also obviously um you know the
  • 00:16:28
    internal Gemini model at Google that is
  • 00:16:30
    probably very good and the voice models
  • 00:16:31
    got better but I actually think what
  • 00:16:33
    they found was product Market fit for
  • 00:16:35
    the actual audio format and it turned
  • 00:16:37
    out to be the podcast format quite quite
  • 00:16:40
    literally it's pretty crazy I mean
  • 00:16:41
    somebody on threads tagged me and was
  • 00:16:43
    like the male voice sounds like you and
  • 00:16:46
    I listened and I was like not the same
  • 00:16:48
    tone but also the Cadence and the type
  • 00:16:51
    of questions I'm like does that mean
  • 00:16:52
    that I'm just like the blend of of all
  • 00:16:55
    different am I like this like you know
  • 00:16:57
    kind of um the unk a middle of this or
  • 00:17:00
    do they copy my voice I'm hoping it's
  • 00:17:02
    the second one it'll be interesting to
  • 00:17:04
    see if people either get tired of
  • 00:17:07
    hearing the same two people talk about
  • 00:17:08
    everything or the opposite they get used
  • 00:17:10
    to the same two people and would prefer
  • 00:17:12
    to hear the same buil trust I don't know
  • 00:17:15
    I I think um I think humans are very
  • 00:17:18
    quick and prone to sort of
  • 00:17:19
    anthropomorphize and it's it's sort of a
  • 00:17:21
    hack on our human brain so you feel like
  • 00:17:23
    you know these people because you heard
  • 00:17:25
    them talk about so many things now so I
  • 00:17:27
    think it's very interesting it's hard to
  • 00:17:28
    predict where we'll go as as a platform
  • 00:17:31
    we view it the same way of course people
  • 00:17:33
    are uploading these podcasts uh to
  • 00:17:36
    Spotify as well and I don't I don't know
  • 00:17:39
    um from the top of my head how you know
  • 00:17:41
    if anyone has super high engagement but
  • 00:17:43
    certainly people are are listening to
  • 00:17:45
    them so it's the same question does this
  • 00:17:46
    turn into a tool for Creative people um
  • 00:17:50
    who can write stories but don't want to
  • 00:17:51
    have the podcast around it or or just
  • 00:17:53
    have no one interviewing them so they
  • 00:17:54
    just do an interview around their own
  • 00:17:56
    material um I don't think
  • 00:17:59
    I think you're going to run into the
  • 00:18:00
    same problem where if you just ask it to
  • 00:18:02
    talk about something it's not going to
  • 00:18:04
    be very good you need a good source
  • 00:18:06
    material so it's the same question is
  • 00:18:09
    this a tool for Creative people to get
  • 00:18:11
    even more productive and creative or is
  • 00:18:12
    it a replacement of creative people my
  • 00:18:14
    bet is it's another tool it's pretty
  • 00:18:16
    interesting because it sort of broadens
  • 00:18:18
    out the longtail and for those not
  • 00:18:21
    familiar with the industry jargon it's
  • 00:18:22
    basically just that like a lot of
  • 00:18:25
    listening is concentrated in a small
  • 00:18:27
    amount of shows yeah and then there's
  • 00:18:28
    this great long tail right like if you
  • 00:18:30
    think about like a a bar chart as it
  • 00:18:32
    just sweeps out and there's uh lots of
  • 00:18:35
    you know seldomly listen to shows yeah
  • 00:18:38
    and the thing about these podcast
  • 00:18:41
    generators notebook LM in particular is
  • 00:18:43
    you can take it and create a podcast for
  • 00:18:46
    something that's so Niche that you would
  • 00:18:48
    never have a show similar with AI code
  • 00:18:50
    right you can start coding things I
  • 00:18:52
    think you spoke about this in your
  • 00:18:54
    interview with Tom M con yeah uh on
  • 00:18:56
    building one another LinkedIn podcast
  • 00:18:58
    Network show
  • 00:18:59
    where now you'll code things that you
  • 00:19:00
    would never code before because you can
  • 00:19:02
    do it and it's similar it might go the
  • 00:19:04
    same way with podcasts where you can for
  • 00:19:07
    instance when I before I was uh heading
  • 00:19:09
    down to Menlo Park to interview Andrew
  • 00:19:11
    Bosworth I just dumped in all my source
  • 00:19:14
    material and it read me a created a
  • 00:19:17
    podcast about like his current
  • 00:19:18
    statements there was like seven
  • 00:19:20
    interviews that him and Zuck did before
  • 00:19:22
    I showed up there and I was able to get
  • 00:19:24
    the summary that podcast never would
  • 00:19:26
    have actually made sense to produce but
  • 00:19:27
    for me it made sense and maybe that's
  • 00:19:29
    where this goes yeah I love that framing
  • 00:19:31
    like one useful framing I think of these
  • 00:19:33
    techniques is is financial framing like
  • 00:19:37
    the cost of something goes to zero like
  • 00:19:39
    the cost of writing code goes to zero
  • 00:19:40
    cost of doing a podcast goes to zero
  • 00:19:42
    cost of prediction goes to zero what
  • 00:19:45
    happens you know and and usually what
  • 00:19:48
    happens is is the the alternatives to
  • 00:19:50
    that good they get challenged but the
  • 00:19:52
    complement to that good you know you
  • 00:19:53
    have the famous like what if the the the
  • 00:19:57
    uh price of coffee goes to zero then
  • 00:19:59
    then tea is going to be replaced but
  • 00:20:01
    sugar is a complement is going to
  • 00:20:03
    explode so I like that way of of
  • 00:20:05
    thinking about it and and I think what's
  • 00:20:08
    going to happen is exactly what you're
  • 00:20:09
    saying we're going to have enormous
  • 00:20:10
    amounts of content around niches where
  • 00:20:13
    it didn't make sense to produce a
  • 00:20:14
    podcast so one way to think about it is
  • 00:20:16
    just like the cost went to zero so I do
  • 00:20:20
    think that the catalog is going to
  • 00:20:22
    explode and then what does that mean
  • 00:20:24
    well it probably means that the
  • 00:20:25
    recommendation problem becomes even more
  • 00:20:26
    important because now it's even harder
  • 00:20:29
    to keep track of everything that is
  • 00:20:30
    uploaded I also think that if you have
  • 00:20:33
    this like vast sea of the perfect sort
  • 00:20:36
    of discussion around any topic uh so the
  • 00:20:39
    recommendation problem becomes more
  • 00:20:40
    valuable to solve the bigger the the
  • 00:20:43
    catalog is but I also think you're going
  • 00:20:45
    to see the same thing as we see in music
  • 00:20:46
    the superstars will actually also get
  • 00:20:49
    bigger this is what I find Fascinating
  • 00:20:51
    People say like our you know Netflix
  • 00:20:53
    winning or YouTube well the truth is
  • 00:20:54
    both the tale is getting bigger but the
  • 00:20:57
    shows are getting bigger and they're
  • 00:20:58
    saying saying are the indis winning or
  • 00:20:59
    Taylor Swift well both indis are winning
  • 00:21:02
    but Taylor Swift is bigger than ever I
  • 00:21:04
    tend to see like these both things
  • 00:21:05
    happening at the same time which is why
  • 00:21:07
    I'm hesitant to like say like that is
  • 00:21:09
    going to happen right but not this yep
  • 00:21:12
    okay let's talk about AI recommendation
  • 00:21:14
    uh it's a big part of spotify
  • 00:21:17
    and we're going to just start at the end
  • 00:21:20
    for this conversation because your
  • 00:21:22
    vision eventually is so right now like
  • 00:21:24
    we'll go into Spotify there'll be some
  • 00:21:26
    algorithmic recommendation there'll be
  • 00:21:27
    some stuff that we listen to
  • 00:21:29
    your vision if I have it right is
  • 00:21:31
    eventually you want Spotify to be sort
  • 00:21:33
    of this ambient friend for us that knows
  • 00:21:35
    this context of the situations we in
  • 00:21:37
    maybe AR we're just talking about ory
  • 00:21:39
    glasses before we start uh recording but
  • 00:21:41
    maybe they know the context of where we
  • 00:21:43
    are and can chime in and give us you
  • 00:21:45
    know an example of type of some music
  • 00:21:48
    that we might want to listen to is that
  • 00:21:50
    right why would we why would why would
  • 00:21:51
    you be pursuing that well I I do think
  • 00:21:55
    of so when we um started Spotify I was
  • 00:21:58
    not part of funding Spotify joined in
  • 00:22:00
    2008 late 2008 2009 Spotify was fun in
  • 00:22:03
    2006 but it's pretty early on and um
  • 00:22:07
    it's interesting that this was before
  • 00:22:09
    machine learning became a thing and so
  • 00:22:12
    Spotify was quite focused on social
  • 00:22:14
    features for purposes of recommendation
  • 00:22:16
    we needed social features because that's
  • 00:22:18
    how most people discover music through a
  • 00:22:20
    friend so we wanted to to connect to
  • 00:22:21
    people and then AI came came along or
  • 00:22:24
    what was called machine learning back
  • 00:22:26
    then and we realized that through all
  • 00:22:28
    the playlisting data we had uh which is
  • 00:22:31
    basically One Way think about the
  • 00:22:32
    playlisting data is almost as labeling
  • 00:22:35
    for for the user they creating a set for
  • 00:22:38
    themselves for Spotify they were saying
  • 00:22:39
    like these tracks go well together these
  • 00:22:41
    tracks go well together so we got a lot
  • 00:22:43
    of of labeled data basically and we said
  • 00:22:46
    internally now some people have a
  • 00:22:49
    musical friend that happens to know
  • 00:22:50
    their taste and so forth but most people
  • 00:22:52
    don't so now we can build this friend
  • 00:22:54
    for for everyone that was the AI but the
  • 00:22:58
    interesting thing is like that thing of
  • 00:23:00
    like building a friend for everyone that
  • 00:23:01
    can give music recommendations like this
  • 00:23:03
    discover weekly it was always an analogy
  • 00:23:05
    people did not think of discover week
  • 00:23:07
    thought of as a set as a service and so
  • 00:23:10
    forth I think what's happening now with
  • 00:23:11
    AI is that the analogy is actually
  • 00:23:14
    becoming reality and so you can see you
  • 00:23:17
    can see us moving a little bit in that
  • 00:23:19
    direction you have the AJ that starts to
  • 00:23:21
    give Spotify voice that talks to you um
  • 00:23:24
    and I think what is going to happen with
  • 00:23:26
    these llms is at least for some Brands
  • 00:23:28
    you will start having literal
  • 00:23:30
    relationships with them and I would love
  • 00:23:33
    if it is the case that you think of
  • 00:23:34
    Spotify as actually a friend not an
  • 00:23:36
    analogy anymore but reality this is a
  • 00:23:38
    person that this is a a thing that knows
  • 00:23:40
    me well this is a musical intelligence a
  • 00:23:42
    podcast intelligence a book intelligence
  • 00:23:44
    and actually like hearing it you know
  • 00:23:48
    tell me about new things and suggest
  • 00:23:50
    things I'm interested in so I think
  • 00:23:52
    that's that is where we're moving I
  • 00:23:54
    think other brands are moving there as
  • 00:23:55
    well I think if you if you look at some
  • 00:23:57
    someone like
  • 00:23:59
    dualingo they've actually only
  • 00:24:01
    communicated through four characters all
  • 00:24:03
    along when you get a push notive it's
  • 00:24:04
    not from dualingo it's from l or SAR or
  • 00:24:07
    something they really they uh they give
  • 00:24:08
    me a hard time if I'm away for a couple
  • 00:24:10
    hours it's like and that was also kind
  • 00:24:12
    of an analogy but now with AI you can
  • 00:24:14
    actually talk to these characters so I
  • 00:24:16
    think this is a journey many companies
  • 00:24:17
    are on and it's interesting to to to
  • 00:24:20
    play that out means that part of what
  • 00:24:22
    was called branding before it's like
  • 00:24:23
    what personality do you want your
  • 00:24:25
    company to have not as an analogy but
  • 00:24:28
    literally
  • 00:24:28
    what personality should Spotify have I
  • 00:24:31
    think that's fascinating time to work in
  • 00:24:34
    in Tech and it's something we're
  • 00:24:35
    thinking a lot about and I think that
  • 00:24:37
    you might be underrating how much people
  • 00:24:38
    view discover weekly as a friend now for
  • 00:24:41
    folks who don't use Spotify discover
  • 00:24:42
    weekly will basically take into account
  • 00:24:44
    you're listening and your preferences
  • 00:24:46
    and give you a playlist of what 30 songs
  • 00:24:48
    on a Monday morning and they're just new
  • 00:24:51
    songs for you to discover and people
  • 00:24:53
    will be like uh discover weekly really
  • 00:24:55
    got me this week or discover weekly infc
  • 00:24:59
    some pain on me this week or what
  • 00:25:00
    happened I thought we had a close
  • 00:25:02
    relationship and now you don't owe me at
  • 00:25:04
    all and you also have so you have this
  • 00:25:06
    AI DJ it's you can find it in the app um
  • 00:25:11
    it's okay I think I there's definit I'm
  • 00:25:13
    curious the feedback I've heard is
  • 00:25:15
    people were excited about it initially
  • 00:25:17
    and have grab have moved away from it
  • 00:25:19
    and what is so now I'm sitting in front
  • 00:25:21
    of the you know person running product
  • 00:25:23
    at Spotify what is actually happening
  • 00:25:24
    with this AI DJ is the experience there
  • 00:25:26
    and are people using it yeah so the
  • 00:25:28
    numbers they not moving away from it
  • 00:25:30
    it's actually very successful so my
  • 00:25:31
    friends are just pretty snobby music
  • 00:25:32
    listeners well for the people that use
  • 00:25:34
    it it's actually U their biggest set
  • 00:25:37
    it's bigger than their discover weekly
  • 00:25:39
    usage so it's quite a quite a binary
  • 00:25:41
    experience I think it's a for people who
  • 00:25:44
    don't know what to want to listen to and
  • 00:25:45
    just want to put something on it's
  • 00:25:47
    working very very well um what I would
  • 00:25:50
    say though is when we launched um thej
  • 00:25:53
    the big innovation there was that we
  • 00:25:55
    managed to basically digiti a voice of a
  • 00:25:58
    real person to make it sound very
  • 00:26:01
    believable but the things that it said
  • 00:26:03
    around the music were were were like to
  • 00:26:06
    some extent juristic and kind of
  • 00:26:08
    repetitive after a while uh so what
  • 00:26:10
    we've done since then is we've invested
  • 00:26:11
    quite a lot in um this is quite recent
  • 00:26:14
    that is rolling out in llms that
  • 00:26:16
    actually tell interesting stories about
  • 00:26:17
    the music and we see very strong effects
  • 00:26:20
    on this on the retention of the
  • 00:26:23
    application so whereas the thing used to
  • 00:26:25
    say here's this and this song from this
  • 00:26:26
    and that I think you like it now we can
  • 00:26:28
    say things like um this artist was just
  • 00:26:31
    in Copenhagen or has played here and
  • 00:26:33
    here last week you're starting to you're
  • 00:26:35
    starting to get interesting stories
  • 00:26:37
    we're starting to feel more personal the
  • 00:26:39
    other thing that I think is missing that
  • 00:26:41
    I hope we can do someday is it can talk
  • 00:26:44
    to you and you can talk back by skipping
  • 00:26:47
    but obviously in the in the age of like
  • 00:26:49
    talking to machines you would like to be
  • 00:26:50
    able to just talk to it and say like no
  • 00:26:53
    this was not very good my Discover
  • 00:26:54
    weekly this week was not what I wanted
  • 00:26:56
    and give actual feedback and that is
  • 00:26:58
    technically very possible now with these
  • 00:27:00
    llms so so that's what I'm hoping will
  • 00:27:03
    happen this should not be a one-way
  • 00:27:04
    relationship which Spotify has been for
  • 00:27:06
    technical reasons it should turn into a
  • 00:27:09
    two-way relationship okay I have
  • 00:27:11
    questions about that coming up and to
  • 00:27:13
    introduce that segment I want to talk to
  • 00:27:15
    you a little bit about how much we
  • 00:27:18
    should allow the algorithms to dictate
  • 00:27:21
    what our music experience and podcast
  • 00:27:23
    experience is going to be versus how
  • 00:27:25
    much should be uh dictated by us how
  • 00:27:29
    much agency should we have over our own
  • 00:27:32
    choices um Kyle cha New Yorker reporter
  • 00:27:36
    recently wrote about how he's leaving
  • 00:27:38
    Spotify I'm just going to put the
  • 00:27:40
    argument out there and hear what you
  • 00:27:41
    think and I'll just read it straight
  • 00:27:43
    from the story he goes through Spotify I
  • 00:27:45
    can browse many decades of published
  • 00:27:47
    music more or less instantly I can
  • 00:27:49
    freely sample the uh work of new
  • 00:27:52
    musicians it has become aggravatingly
  • 00:27:55
    difficult to find what I want to listen
  • 00:27:57
    to with the recent product update he
  • 00:27:59
    says it became clearer than ever what
  • 00:28:02
    the app has been pushing me to do listen
  • 00:28:04
    to what it suggests not choose my music
  • 00:28:07
    on my own what do you think about that
  • 00:28:09
    argument well I think this is an
  • 00:28:11
    individual feedback but I think
  • 00:28:12
    generally you have very different types
  • 00:28:15
    of users so I'm I'm going to get I'm
  • 00:28:17
    going to get this person back on Spotify
  • 00:28:19
    100% I think there is a there's an
  • 00:28:22
    interesting trade-off here that is that
  • 00:28:23
    is real so people want less friction um
  • 00:28:28
    they want to spend less time searching
  • 00:28:30
    you want to make things as as easy as
  • 00:28:31
    possible Right but there is this end of
  • 00:28:33
    the line where you you sit there and you
  • 00:28:35
    just receive you're kind of force-fed
  • 00:28:37
    and you don't give any signal back maybe
  • 00:28:39
    a few clicks and so forth um and that's
  • 00:28:41
    something that that we want to avoid I
  • 00:28:43
    think this is where the industry is
  • 00:28:45
    going it's going more towards
  • 00:28:47
    distruction content and sort of just
  • 00:28:49
    sitting and receiving and it's a little
  • 00:28:51
    bit of a distopian um end of the line
  • 00:28:54
    there so what is interesting with
  • 00:28:56
    Spotify which we re pising is that it
  • 00:28:58
    was actually a platform where you
  • 00:29:00
    invested quite a lot in your own
  • 00:29:02
    playlisting right and the there's a
  • 00:29:05
    trade-off here between if we you could
  • 00:29:07
    have a vision is we should be so good at
  • 00:29:09
    machine learning that you should never
  • 00:29:11
    playlist again that would be the goal um
  • 00:29:15
    because then you've done the user a
  • 00:29:16
    great service supposedly but then you
  • 00:29:18
    also receive no signal and the user does
  • 00:29:20
    no investment so we're actually
  • 00:29:21
    reemphasizing playlisting quite a lot
  • 00:29:24
    okay your own investment and and you
  • 00:29:26
    know over the years we we've gone more
  • 00:29:28
    towards um machine learning and
  • 00:29:30
    algorithms because it works people
  • 00:29:32
    listen more and they they appreciate the
  • 00:29:34
    service more um but we need to cater to
  • 00:29:37
    everyone including this reporter so the
  • 00:29:39
    Spotify user base is divided into many
  • 00:29:41
    different kinds of people you have the
  • 00:29:43
    the sort of the track listeners only
  • 00:29:46
    listen to playlist you have the hardcore
  • 00:29:48
    album listeners it's like I just want to
  • 00:29:50
    listen to an album the way the Creator
  • 00:29:52
    thought about it I don't want all the
  • 00:29:54
    songs in between um you have like the
  • 00:29:56
    artist radio listeners only listen to to
  • 00:29:59
    one one type of artist and it's it's
  • 00:30:01
    actually a big challenge to build a
  • 00:30:02
    service that serves everyone when people
  • 00:30:06
    are very different uh so we we try our
  • 00:30:08
    best to make sure that the sort of Music
  • 00:30:11
    Aion AOS who want their library to be
  • 00:30:14
    album album album can have their service
  • 00:30:17
    and then but then you have the other
  • 00:30:18
    people who just want like I just want my
  • 00:30:21
    daily mix to play in my air I don't you
  • 00:30:23
    know I just want to collect tracks They
  • 00:30:25
    also need to be successful so we're
  • 00:30:27
    we're trying to build and cater for both
  • 00:30:30
    you can never Place everyone 100% but
  • 00:30:33
    we're trying to be statistical about it
  • 00:30:35
    uh to make sure that it is U it is uh
  • 00:30:39
    vastly better for the majority of people
  • 00:30:42
    but we our goal is to cater to everyone
  • 00:30:43
    and I do think there's a real Point
  • 00:30:45
    around going to zero user investment
  • 00:30:48
    seems good in the short term but I don't
  • 00:30:50
    think it's good in the long term because
  • 00:30:51
    you actually lose signal from that user
  • 00:30:53
    and at the end I I think they feel less
  • 00:30:55
    participatory in the experience even if
  • 00:30:58
    the engagement looks high if you've done
  • 00:31:00
    no feedback I don't know how much you
  • 00:31:02
    feel this is actually your service
  • 00:31:04
    definitely and look I'll confirm that
  • 00:31:06
    Spotify does listen to user feedback I
  • 00:31:08
    sent a tweet out uh a couple years ago
  • 00:31:11
    talking about how like some of sometimes
  • 00:31:13
    I'm baffled by the Spotify product
  • 00:31:15
    decisions and I mean maybe it was
  • 00:31:17
    because I was a reporter but someone
  • 00:31:18
    from your team reached out and I talked
  • 00:31:20
    about how I wanted to see recently
  • 00:31:22
    played like oftentimes I'll be listening
  • 00:31:24
    to something and then I'll go away from
  • 00:31:26
    it and I can't find in the app and then
  • 00:31:28
    a couple months later there's a recently
  • 00:31:30
    played button in the app there are some
  • 00:31:32
    great updates coming for you as well on
  • 00:31:34
    that topic because this is a big user
  • 00:31:36
    need maybe it takes a little bit longer
  • 00:31:38
    than we want but obviously our goal is
  • 00:31:40
    to is to listen to us feedback and try
  • 00:31:42
    but we get very sometimes really
  • 00:31:44
    completely opposing you to feedback
  • 00:31:46
    that's the tricky thing who who do you
  • 00:31:48
    listen to the most the people who want
  • 00:31:49
    this desperately or hate this
  • 00:31:51
    desperately and and there's a lot of
  • 00:31:53
    both types of feedback so it's product
  • 00:31:55
    development at this scale is sort of a
  • 00:31:57
    statistical experience but you still
  • 00:31:59
    have to have a bit of an opinion if you
  • 00:32:01
    only treat statistics the application is
  • 00:32:04
    going to be very weird at the end of the
  • 00:32:06
    day so you have to combine some sort of
  • 00:32:08
    vision and conviction but you have to be
  • 00:32:10
    still very data driven I think an
  • 00:32:13
    interesting example of user investment
  • 00:32:16
    and
  • 00:32:17
    AI that that we launched recently is
  • 00:32:19
    something called um AI
  • 00:32:21
    playlisting uh so this is I think a good
  • 00:32:24
    example of like the first time you can
  • 00:32:26
    talk to Spotify so the DJ talk to you
  • 00:32:28
    and it's getting better but it doesn't
  • 00:32:30
    listen it listens to clicks maybe but
  • 00:32:32
    with AI playlisting um we built this
  • 00:32:35
    experience where you can you can prompt
  • 00:32:37
    what is an llm with what kind of
  • 00:32:39
    playlist so we have an llm and the llms
  • 00:32:42
    have a set of World Knowledge about
  • 00:32:43
    music but then we have the music catalog
  • 00:32:45
    and we have your listening history so
  • 00:32:46
    this is an llm that understands your
  • 00:32:48
    particular taste and you can ask it for
  • 00:32:50
    a playlist with you know big uh big
  • 00:32:54
    drops and EDM for driving fast at night
  • 00:32:56
    or something and then it will try to do
  • 00:32:58
    that and then you can say like no um a
  • 00:33:01
    bit more upbeat or not that artist and
  • 00:33:03
    so forth and and this I think is a good
  • 00:33:06
    mix of using AI but not to force video
  • 00:33:09
    stuff it's actually very high signal you
  • 00:33:11
    are literally telling us what you want
  • 00:33:13
    right and then when we say here it is
  • 00:33:14
    you say that one yes no no yes and then
  • 00:33:18
    you can reprompt so so it's back to I
  • 00:33:20
    think it should be a two-way
  • 00:33:21
    conversation and I think the first wave
  • 00:33:23
    of machine learning allowed us to do the
  • 00:33:25
    oneway push uh the the next wave
  • 00:33:28
    generative allows us to actually listen
  • 00:33:30
    to you even in clear text so
  • 00:33:32
    communicating with Spotify just through
  • 00:33:33
    skip buttons is a pretty narrow signal
  • 00:33:35
    so it's kind of hard for us to
  • 00:33:37
    understand like when you skip it was it
  • 00:33:39
    because you hated it or because you
  • 00:33:40
    liked it but it was too many times now
  • 00:33:42
    you can actually say like I really don't
  • 00:33:44
    like this G like remove it so I was
  • 00:33:46
    dming with Kyle last night as like hey
  • 00:33:48
    I'm gonna meet with Gustav what should I
  • 00:33:50
    ask him and one of the things he said is
  • 00:33:52
    uh should Spotify users be able to tweak
  • 00:33:54
    their recommendations and your answer
  • 00:33:56
    here is resounding yes
  • 00:33:58
    absolutely absolutely we are working on
  • 00:33:59
    these things both the obvious things
  • 00:34:01
    where you can say like I didn't like
  • 00:34:03
    this particular thing but I think the
  • 00:34:05
    free text element is very interesting if
  • 00:34:07
    you could talk to it You' probably it
  • 00:34:09
    would learn much more but you would
  • 00:34:11
    probably also get more trust definitely
  • 00:34:14
    let me ask you one broader question
  • 00:34:15
    about this because I I'll I won't stick
  • 00:34:19
    on Kyle's stuff for the entire uh
  • 00:34:21
    conversation but I thought it was really
  • 00:34:24
    interesting and he wrote a book called
  • 00:34:25
    filter World main argument he's been on
  • 00:34:27
    the show I I'll link it in the show
  • 00:34:29
    notes the main argument is that Al our
  • 00:34:31
    world mediated by algorithms has become
  • 00:34:33
    too bland and you know effectively that
  • 00:34:35
    the algorithm have flattened out you
  • 00:34:38
    know what used to be a more vibrant yeah
  • 00:34:41
    experience with things like music have
  • 00:34:42
    do you see that at all I think this is a
  • 00:34:45
    really interesting argument there there
  • 00:34:46
    are two ways I want to address that uh
  • 00:34:49
    one is for Spotify specifically we've
  • 00:34:53
    seen the feedback that people feel like
  • 00:34:55
    it's great for the kind of stuff I
  • 00:34:57
    already listen to but I feel like I'm in
  • 00:34:59
    a bubble I'm getting more of the same
  • 00:35:01
    I'm not getting new stuff this is sort
  • 00:35:03
    of a Spotify specific challenge because
  • 00:35:05
    most of the time your phone is in the
  • 00:35:07
    pocket and you're listening and when
  • 00:35:09
    you're listening you're listening to a
  • 00:35:11
    session let's say you're listening to
  • 00:35:12
    indie folk or something then it's quite
  • 00:35:14
    easy for us to say here's another indie
  • 00:35:16
    folk song and and you're going to say oh
  • 00:35:18
    that's that's a good recommendation but
  • 00:35:19
    if we start playing Metallica there
  • 00:35:21
    you're going to be like what is this so
  • 00:35:23
    most of the recommendation sort of
  • 00:35:25
    inventory we have is kind of constrained
  • 00:35:28
    naturally to watch or they're listening
  • 00:35:29
    to because we can't put in very random
  • 00:35:31
    things you would say this is a bad
  • 00:35:33
    recommendation so this is a challenge
  • 00:35:35
    for us when you know when we want to
  • 00:35:37
    show you something completely new the
  • 00:35:39
    favorite example is I love reaton but
  • 00:35:42
    you wouldn't have seen that from my
  • 00:35:43
    listening history how do we solve that
  • 00:35:45
    problem so we started investing about
  • 00:35:47
    two years ago in in other types of of
  • 00:35:50
    foreground recommendation so sort of
  • 00:35:52
    like the feeds that you see on social
  • 00:35:54
    media but you can you can literally say
  • 00:35:57
    like okay I'm bored I want to go wide
  • 00:36:00
    then you can go into these um foreground
  • 00:36:03
    feeds of Music where you can swipe
  • 00:36:05
    through many tracks and they're very
  • 00:36:06
    efficient the hit rate is going to be
  • 00:36:07
    low because now we're in a territory
  • 00:36:09
    where the whole point is we don't know
  • 00:36:11
    that you like this so our hit rate is
  • 00:36:13
    going to be low then I think you need a
  • 00:36:14
    very efficient UI to evaluate lots of
  • 00:36:17
    content right because the hit rate may
  • 00:36:19
    be one in 20 you're not going to listen
  • 00:36:21
    to 20 songs that's over an hour of music
  • 00:36:23
    you need to go quick so we try to solve
  • 00:36:26
    that problem for for when like Alex is
  • 00:36:28
    bored and he wants to Branch out as soon
  • 00:36:30
    as we see that signal we didn't have
  • 00:36:31
    tools for that before so so we built
  • 00:36:34
    that so that's part of the answer
  • 00:36:36
    Spotify being an audio service made it a
  • 00:36:37
    bit harder to go explore so now we have
  • 00:36:40
    these foreground feeds we have music
  • 00:36:41
    videos not in the US yet but but in much
  • 00:36:44
    of the rest of the world we have music
  • 00:36:45
    videos that very helpful when you're
  • 00:36:46
    evaluating new music but the more
  • 00:36:49
    philosophical part of this answer is did
  • 00:36:52
    the algorithms sort of flatten out
  • 00:36:53
    because they are to some extent trying
  • 00:36:55
    to find statistical patterns and average
  • 00:36:58
    and I think if you look at
  • 00:37:00
    recommendation technology I don't think
  • 00:37:02
    this is widely known yet but these deep
  • 00:37:04
    learning based systems they had
  • 00:37:06
    flattened out in terms of if you added
  • 00:37:08
    more use data or more parameters they
  • 00:37:10
    did not get better like the llms there
  • 00:37:12
    were no scaling laws it's just like it
  • 00:37:15
    is what it is and you could move it 0 2%
  • 00:37:18
    there's something that has happened
  • 00:37:19
    there recently recently which is called
  • 00:37:20
    generative recommendations where you
  • 00:37:23
    actually use a sort of a large language
  • 00:37:25
    model instead of these old deep learning
  • 00:37:27
    models mod and you basically think of um
  • 00:37:31
    user actions as a language so you have a
  • 00:37:34
    sequence for user they they click this
  • 00:37:35
    they listen to that they click this they
  • 00:37:36
    listen to that and then just if you turn
  • 00:37:39
    that into tokens just as you can turn a
  • 00:37:42
    language into tokens you can just as you
  • 00:37:44
    can try to predict the missing word in a
  • 00:37:46
    sentence you can try to predict the
  • 00:37:47
    missing action in a sequence and it
  • 00:37:50
    turns out that these generative
  • 00:37:51
    recommendations they do scale with more
  • 00:37:53
    use of data and more parameters just
  • 00:37:55
    like the llms so this is is a
  • 00:37:57
    long-winded way of saying I think he's
  • 00:38:00
    right that the recommendations did
  • 00:38:01
    flatten out it's also true that people
  • 00:38:03
    are changing recommendations stacks and
  • 00:38:05
    it now is unclear why they couldn't
  • 00:38:08
    continuously get better so I'm hoping
  • 00:38:10
    that the recommendations do get more
  • 00:38:11
    intelligence because intelligent because
  • 00:38:13
    now it's not just a statistical average
  • 00:38:16
    they can look at your specific user
  • 00:38:17
    history going years back and they could
  • 00:38:20
    potentially understand that it's
  • 00:38:21
    actually uh you know Christmas again and
  • 00:38:24
    last year at Christmas you did this I'm
  • 00:38:26
    hoping it gets more intelligent and one
  • 00:38:29
    last question about recommendations or
  • 00:38:30
    maybe I have two but one important one
  • 00:38:32
    that comes from Ronan Roy who's on the
  • 00:38:34
    Friday show with us he would like there
  • 00:38:37
    to be a parent mode on Spotify where if
  • 00:38:40
    you have kids you can be like I'm on
  • 00:38:41
    child mode and then recommend kid music
  • 00:38:44
    and then parent mode you know and don't
  • 00:38:47
    uh blur my recommendations what do you
  • 00:38:48
    think about that so so we have a a bunch
  • 00:38:52
    of different solutions uh for this
  • 00:38:54
    obviously there's a family plan so
  • 00:38:56
    hopefully your kid can have their own
  • 00:38:58
    account and then it doesn't that cost
  • 00:38:59
    more the recommendation exactly what are
  • 00:39:00
    you going to do for you three-year-old
  • 00:39:02
    exactly there's the other thing is you
  • 00:39:04
    can create a playlist for your kid and
  • 00:39:05
    then if you click the settings you can
  • 00:39:08
    say do not include in my recommendations
  • 00:39:10
    and then it actually doesn't destroy
  • 00:39:12
    your recommendations at all uh so so
  • 00:39:16
    there are those Solutions we're also
  • 00:39:18
    trying to understand that all of this is
  • 00:39:20
    kids music so while this is part of your
  • 00:39:22
    taste profile we should not play this in
  • 00:39:25
    your other sets because this is probably
  • 00:39:27
    something you're doing for sort of a use
  • 00:39:29
    case so you probably want a kids music
  • 00:39:31
    playlist in there but you don't want
  • 00:39:32
    that music to affect your your other
  • 00:39:35
    sets there's an algorithmic component
  • 00:39:37
    there's a there's a subscription plan
  • 00:39:39
    component and then it's back to like
  • 00:39:41
    more user control you can actually
  • 00:39:43
    already say that this playlist should
  • 00:39:44
    not be considered my taste and so we're
  • 00:39:47
    going to build more of those controls
  • 00:39:49
    okay R will be happy to hear that yeah
  • 00:39:51
    uh okay really last question about
  • 00:39:53
    recommendations then we're going to go
  • 00:39:54
    into podcasts and some other formats
  • 00:39:57
    um I don't know if you have seen this
  • 00:40:01
    YouTuber his name is Fontana he did this
  • 00:40:05
    thing about the shabzi being song being
  • 00:40:07
    the song of the summer explaining why
  • 00:40:10
    and he made an observation there that
  • 00:40:13
    was interesting to me talking about how
  • 00:40:17
    we used to hear music on the radio often
  • 00:40:21
    and that was the music that was played
  • 00:40:23
    there was music that would often be
  • 00:40:25
    played when we with other people with
  • 00:40:27
    friends
  • 00:40:27
    having a good time and it led to more
  • 00:40:30
    you know dance songs rock Al anthems and
  • 00:40:32
    stuff like this and today we're like
  • 00:40:35
    mostly accessing music via streaming
  • 00:40:37
    platforms and he says those are much
  • 00:40:39
    more individualized recommendations
  • 00:40:42
    which has kind of shifted the way that
  • 00:40:44
    music is made and even the hits in music
  • 00:40:47
    what do you think about that
  • 00:40:48
    argument so there is a philosophical
  • 00:40:51
    question there which has been researched
  • 00:40:53
    a few times which is do you have an
  • 00:40:54
    innate taste in your brain and our job
  • 00:40:56
    is to search for that and find it or do
  • 00:40:58
    what we play actually affect what you
  • 00:41:01
    like and there are all these experiments
  • 00:41:03
    in colleges where you know you play like
  • 00:41:04
    different songs to different groups and
  • 00:41:06
    then you see what they like and it seems
  • 00:41:08
    like it's a bit a bit a bit of both you
  • 00:41:10
    have some sort of innate taste but
  • 00:41:12
    you're also affected by what you he to
  • 00:41:13
    this argument like the the radio can
  • 00:41:15
    change your your taste uh so so I think
  • 00:41:19
    there's um there's two to that argument
  • 00:41:21
    what I think is interesting about um our
  • 00:41:24
    music listening is that when we survey
  • 00:41:27
    you users and we ask them what
  • 00:41:29
    percentage of your listening is with
  • 00:41:31
    others it's a huge
  • 00:41:33
    percentage double digit percentage so
  • 00:41:36
    music is actually a very social activity
  • 00:41:39
    still and in some cases we see this we
  • 00:41:42
    have this feature called Jam that is is
  • 00:41:44
    taking off like a rocket for us it's
  • 00:41:46
    doing very well and and jam is
  • 00:41:48
    essentially we can detect when two
  • 00:41:50
    phones are close to each other it justs
  • 00:41:52
    like hey do you want to join Alex's jam
  • 00:41:54
    and now we have a joint queue so had a
  • 00:41:56
    party the way you party WR now with
  • 00:41:58
    Spotify is you don't go and like
  • 00:42:00
    interrupt you just bring up your phone
  • 00:42:01
    you join the queue and then you can
  • 00:42:02
    queue things up right and so uh we have
  • 00:42:06
    a lot of of U joint listening and people
  • 00:42:08
    are listening like I said I don't want
  • 00:42:11
    to say the exact percentage but it's
  • 00:42:12
    double digit percentage of listening
  • 00:42:14
    happening in groups it just looks to
  • 00:42:16
    individual as individual listening to us
  • 00:42:19
    so I think it's actually happening more
  • 00:42:21
    than maybe people think it's not 100%
  • 00:42:23
    individual listening but because we
  • 00:42:25
    don't see them as group listenings
  • 00:42:27
    we're still treating them as individual
  • 00:42:29
    listen so now that we're getting more
  • 00:42:31
    data on what is good group music that
  • 00:42:34
    becomes a different category so I I
  • 00:42:37
    think the radio use case is happening
  • 00:42:39
    you're hearing songs at parties and with
  • 00:42:42
    others and when you're writing in the
  • 00:42:43
    car and so forth it just looks to these
  • 00:42:45
    Services as lonely listening but it's
  • 00:42:48
    actually quite social right okay let's
  • 00:42:50
    take a quick break and come back to talk
  • 00:42:52
    about podcast audio books and see how
  • 00:42:55
    many random questions I can get to
  • 00:42:56
    before time is out we'll be back right
  • 00:42:58
    after this and we're back here on big
  • 00:43:00
    technology podcast with G Gustav Sodom
  • 00:43:03
    he's the chief product officer Chief
  • 00:43:04
    technology officer and co-president of
  • 00:43:08
    Spotify so Spotify is investing heavily
  • 00:43:11
    in podcasts um this has been going on
  • 00:43:14
    for a long time first through largely
  • 00:43:17
    through an original strategy and now
  • 00:43:19
    less so um also audio books you can find
  • 00:43:22
    my book always day one on Spotify if
  • 00:43:24
    you're a premium listener which I'm
  • 00:43:25
    happy about because people can listen to
  • 00:43:28
    to the book what has gone into the
  • 00:43:30
    decision to just bring all these formats
  • 00:43:33
    together in one app and um I mean are
  • 00:43:37
    they good businesses for you spot uh uh
  • 00:43:39
    podcasts and audiobooks yes if we start
  • 00:43:42
    with the first one how do we come to
  • 00:43:44
    this decision uh what happened is that
  • 00:43:48
    we saw internally actually at Spotify a
  • 00:43:51
    lot of our developers sort of hacking
  • 00:43:53
    Spotify into or hacking podcasts using
  • 00:43:56
    RSS into the Spotify experience and we
  • 00:43:58
    saw it again and again at hawis and
  • 00:44:01
    first we thought like maybe it's a it's
  • 00:44:02
    a it's a niche random need we saw it
  • 00:44:05
    again and again and so then we just it's
  • 00:44:07
    like user feedback user research you
  • 00:44:09
    know Spotify is still like many
  • 00:44:10
    thousands of employees so it's it's not
  • 00:44:12
    a very representative sample of society
  • 00:44:14
    but it is some sample of society so if
  • 00:44:16
    you see the same user need many times
  • 00:44:17
    you should take it seriously so we
  • 00:44:19
    started looking at that and then we
  • 00:44:21
    looked at podcast that we saw had a lot
  • 00:44:23
    of potential and was growing but we
  • 00:44:25
    didn't think anyone was doing something
  • 00:44:27
    interesting with it so we decided to to
  • 00:44:29
    then uh just approach it because we saw
  • 00:44:32
    the US need internally we saw the market
  • 00:44:34
    growing we sized it and then we saw that
  • 00:44:35
    there was no one really investing in it
  • 00:44:37
    Apple hadn't invested in it and they had
  • 00:44:39
    like 98% of the market so that's how we
  • 00:44:43
    came to it and then the question is yeah
  • 00:44:45
    that Apple podcast app needs work okay
  • 00:44:47
    but sorry go ahead but we were grateful
  • 00:44:49
    for that uh so then the question is why
  • 00:44:52
    in in the same application why not as a
  • 00:44:56
    separate application
  • 00:44:57
    and uh that's there there are two views
  • 00:45:00
    of that one is it's a strategic decision
  • 00:45:03
    the
  • 00:45:04
    the the biggest barrier to something new
  • 00:45:07
    right now unfortunately isn't
  • 00:45:08
    necessarily the quality of the
  • 00:45:10
    application it's the user acquisition
  • 00:45:12
    cost distribution is everything
  • 00:45:13
    distribution is still everything and and
  • 00:45:16
    actually at the beginning of the iPhone
  • 00:45:17
    era there was a lot of organic
  • 00:45:19
    distribution people went to the App
  • 00:45:21
    Store every day it's like no one goes
  • 00:45:22
    there anymore so you almost have to pay
  • 00:45:25
    for revenues so user acquisition cost is
  • 00:45:27
    probably the biggest inhibitor to most
  • 00:45:29
    business plans so if we built a separate
  • 00:45:31
    app we would have to reacquire our own
  • 00:45:33
    users again and that would make it very
  • 00:45:35
    expensive and we have seen all of these
  • 00:45:37
    big big companies the American tech
  • 00:45:39
    companies launching app after app and
  • 00:45:41
    basically nothing worked then we look at
  • 00:45:43
    China which is a different strategy of
  • 00:45:45
    the super apps where they double down on
  • 00:45:47
    their in on their own distribution and
  • 00:45:50
    so you can think of like pcast
  • 00:45:51
    pre-installed so that was the Strategic
  • 00:45:53
    angle for what this made sense but I
  • 00:45:56
    actually have
  • 00:45:57
    a user angle on this where I think it is
  • 00:45:59
    the better experience so I think in
  • 00:46:03
    2024 the user should not adapt the
  • 00:46:06
    software to the content I think in 2024
  • 00:46:10
    the software should adapt to the content
  • 00:46:12
    so if you play a piece of music there
  • 00:46:14
    should be skip buttons If you play a
  • 00:46:16
    podcast it's not rocket science to
  • 00:46:17
    change the skip buttons to 15sec scrub
  • 00:46:19
    and if you play an audiobook to to
  • 00:46:21
    change them to Chapters like come on
  • 00:46:22
    it's 2024 why do you have to switch apps
  • 00:46:25
    for that right right so we we actually
  • 00:46:27
    both believe that it was strategically
  • 00:46:29
    the best for us because then we we could
  • 00:46:30
    double down our own distribution but we
  • 00:46:32
    also think this long term is the right
  • 00:46:34
    user experience it is the easiest for
  • 00:46:36
    the user now we have these beautiful
  • 00:46:37
    connections between the audiobook and
  • 00:46:40
    the author being interviewed in a
  • 00:46:41
    podcast on the same thing where it's
  • 00:46:42
    seamless instead of like no you should
  • 00:46:44
    switch the app and go somewhere else so
  • 00:46:47
    so that's the reason that we do it in
  • 00:46:48
    the same application and talk a little
  • 00:46:50
    bit about discoverability because that's
  • 00:46:52
    the biggest issue for podcasts I mean if
  • 00:46:55
    I and as a company that's an expert in
  • 00:46:58
    recommendations which we've spent like
  • 00:47:00
    most of the show talking about that
  • 00:47:02
    should be something that you get done
  • 00:47:03
    pretty well but for instance like if I'm
  • 00:47:05
    listening to Tech shows you know and and
  • 00:47:07
    I'm not listening to Big technology
  • 00:47:08
    podcast I probably want to see that um
  • 00:47:11
    there's a show called Big technology
  • 00:47:13
    podcast out there and from what I've
  • 00:47:14
    heard discoverability like both from um
  • 00:47:17
    product people and from podcast
  • 00:47:20
    producers has been the biggest issue uh
  • 00:47:22
    probably because there's like a huge
  • 00:47:23
    investment that goes into listening to
  • 00:47:25
    even that for first five minutes of a
  • 00:47:28
    show I mean that's like 2 minutes longer
  • 00:47:30
    than your average song to try out a new
  • 00:47:32
    show and most of them most I mean I
  • 00:47:34
    actually changed my show that we could
  • 00:47:36
    do our like really like you know
  • 00:47:39
    information Rich uh intro which you just
  • 00:47:41
    experienced and then take a break take a
  • 00:47:43
    break and come back in because if people
  • 00:47:46
    are going to try it out I want them to
  • 00:47:47
    know what they're getting versus like
  • 00:47:49
    the typical long wind
Tags
  • Spotify
  • AI in music
  • Music recommendations
  • Podcasting
  • User experience
  • Copyright
  • Discover Weekly
  • Generative music
  • Content discoverability
  • Technology and music