从策展到生成：Spotify 的 Gustav Söderström 谈 AI 与产品演变

欢迎收听 Lenny 的播客，在这里我会采访世界级的产品领导者和增长专家，从他们打造和发展当今最成功产品的宝贵经验中学习。今天的嘉宾是 Gustav Söderström。Gustav 是一位产品传奇人物，现任 Spotify 的联席总裁、首席产品官兼首席技术官，负责 Spotify 的全球产品和技术战略，并管理公司的产品、设计、数据和工程团队。自从我推出这个播客以来，Gustav 就一直在我梦想嘉宾名单上，我很高兴我们终于实现了这个愿望。在我们的对话中，我们将深入探讨 Gustav 在承担重大风险方面的经验，以及当这些风险未能如愿时该如何应对；Spotify 如何从小队模式转型，以及他们现在如何构建团队；AI 已经如何影响产品，以及 AI 生成音乐的未来；为什么所有伟大的产品都需要施展某种魔术；《继承之战》对瑞典商业文化的刻画有多准确；以及他那个关于“穿着裤子”的搞笑比喻。在听完赞助商的简短介绍后，请欣赏这期与 Gustav Söderström 的节目。

Welcome to Lenny's podcast where I interview world-class product leaders and growth experts to learn from their hard-won experiences building and growing today's most successful products. Today my guest is Gustav Söderström. Gustav is a product legend and he's now the co-president, chief product, and chief technology officer at Spotify where he's responsible for Spotify's global product and technology strategy and oversees the product, design, data, and engineering teams at the company. I've had Gustav on my wish list of dream guests to have on this podcast since the day I launched the podcast and I'm so happy we made it happen. In our conversation we dig into what Gustav has learned about taking big bets and what to do when they don't work out, how Spotify moved away from squads and how they structure their teams now, how AI is already impacting the product, and also the future of music generated by AI, also why all great products need to pull some kind of magic trick, how accurately Succession represents Swedish business culture, and his hilarious analogy of being in your pants. Enjoy this episode with Gustav Söderström after a short word from our sponsors.

本期节目由 Microsoft Clarity 赞助播出。Clarity 是一款免费且易于使用的工具，可以捕捉真实用户如何使用你的网站。你可以观看实时会话回放，了解用户在哪些环节顺畅通过，在哪些环节遇到困难。你可以查看即时热图，了解用户与页面哪些部分互动，忽略哪些内容。你还可以通过非常酷的挫败感指标（如愤怒点击、死点击等）精准定位用户的问题所在。如果你收听这个播客，你就会知道我们经常谈论了解用户的重要性。通过观察用户真实的产品体验，你可以发现产品机会、转化提升点，并找到你设想的使用方式与实际使用方式之间的巨大差距。Microsoft Clarity 通过一套简单但功能强大的功能让这一切成为可能。你会惊讶于 Clarity 的易用性，而且它永远完全免费。你永远不会遇到流量限制，也不会被迫升级到付费版本。它同时适用于应用和网站。别再猜测了，使用 Clarity 吧。请访问 clarity.microsoft.com 了解 Clarity。

This episode is brought to you by Microsoft Clarity, a free, easy-to-use tool that captures how real people are actually using your site. You can watch live session replays to discover where users are breezing through your flow and where they struggle. You can view instant heat maps to see what parts of your page users are engaging with and what content they're ignoring. You can also pinpoint what's bothering your users with really cool frustration metrics like rage clicks and dead clicks and much more. If you listen to this podcast, you know how often we talk about the importance of knowing your users. And by seeing how your users truly experience your product, you can identify product opportunities, conversion wins, and find big gaps between how you imagine people using your product and how they actually use it. Microsoft Clarity makes it all possible with a simple, yet incredibly powerful set of features. You'll be blown away by how easy Clarity is to use and it's completely free forever. You will never run into traffic limits or be forced upgrade to a paid version. It also works across both apps and websites. Stop guessing, get Clarity. Check out Clarity at clarity.microsoft.com.

本期节目由 Eppo 赞助播出。Eppo 是下一代 AB 测试平台，由 Airbnb 校友为现代增长团队打造。DraftKings、Zapier、ClickUp、Twitch 和 Cameo 等公司都依赖 Eppo 来支持他们的实验。无论你在哪里工作，运行实验都越来越重要，但市场上没有能与现代增长团队技术栈集成的商业工具。这导致浪费时间构建内部工具，或者尝试通过笨拙的营销工具运行自己的实验。当我在 Airbnb 时，我最喜欢在那里工作的一点就是我们的实验平台，我可以按设备类型、国家、用户阶段对数据进行切片和切块。Eppo 能做到这一切，甚至更多，它能快速交付结果，避免烦人的长时间分析周期，并帮助你轻松找到所发现问题的根本原因。Eppo 让你超越基本的点击率指标，转而使用你的北极星指标，如激活、留存、订阅和支付。Eppo 支持前端、后端、电子邮件营销，甚至机器学习客户端的测试。请访问 geteppo.com 了解 Eppo。网址是 geteppo.com，让你的实验速度提升 10 倍。

This episode is brought to you by Eppo. Eppo is a next-generation AB testing platform built by Airbnb alums for modern growth teams. Companies like DraftKings, Zapier, ClickUp, Twitch, and Cameo rely on Eppo to power their experiments. Wherever you work, running experiments is increasingly essential, but there are no commercial tools that integrate with a modern growth team stack. This leads to wasted time building internal tools or trying to run your own experiments through a clunky marketing tool. When I was at Airbnb, one of the things that I loved most about working there was our experimentation platform where I was able to slice and dice data by device types, country, user stage. Eppo does all that and more delivering results quickly, avoiding annoying prolonged analytic cycles, and helping you easily get to the root cause of any issue you discover. Eppo lets you go beyond basic click-through metrics and instead use your north star metrics like activation, retention, subscription, and payments. Eppo supports test on front end, on back end, email marketing, even machine learning clients. Check out Eppo at geteppo.com. That's geteppo.com and 10x your experiment velocity.

Gustav，欢迎来到播客。

Gustav, welcome to the podcast.

谢谢你邀请我，Lenny。很高兴来到这里。

Thanks for having me, Lenny. Pleasure to be here.

这是我的荣幸。到目前为止，你在 Spotify 已经工作了超过 14 年，这在科技界是罕见的成就。在 Spotify 期间，你担任过许多不同的职务。你能先给我们介绍一下这些不同的角色，以及你在 Spotify 这些年都做了些什么，然后说说你现在的工作吗？你现在负责什么？

It's my pleasure to have you on. So at this point you've been at Spotify for over 14 years, which is a rare feat in the tech world. And you've held a lot of different roles while you've been at Spotify. Can you just start off by giving us a sense of what these various roles and what you've done over the years at Spotify and then just what do you What do you have to do these days? What are you responsible for now?

我是在 2009 年初、2008 年底加入 Spotify 的。当时我的工作，我之前是创业者，在早期功能手机和智能手机领域创办过几家公司。所以我在那方面有很多知识。我把一家移动领域的公司卖给了 Yahoo，在那里工作了一段时间，然后回到了瑞典。后来通过一个共同的朋友，我认识了 Daniel Ek，Spotify 的 CEO 和联合创始人。他们已经开发出了桌面产品，也就是免费的流媒体桌面产品。那产品很棒，我也试用过。但他们需要有人来弄清楚移动端该怎么做。因为我在那个领域有创业经验，所以我得到了那份工作。所以我的工作是负责 Spotify 的移动业务，弄清楚移动端的产品形态，这很有挑战性，因为显然 Spotify 桌面版是一个免费的按需流媒体应用，而在当时，尤其是在边缘网络下，你根本无法实时流媒体播放。性能达不到要求。而且你也不能用广告模式来支撑它。所以这是一次产品加商业模式的创新，非常有趣。我就是这样开始的。几年后，我接管了 Spotify 的所有产品开发。又过了几年，我实际上也承担了技术责任，也就是 Spotify 的 CTO 角色。最近，我的正式头衔是和 Alex Nordstrom 一起担任 Spotify 的联席总裁。我们每人负责公司的一半。我负责产品和技术方面，他负责商业和内容方面。这就是最简略的版本。除了承担更多责任，比如接管技术部门，从职位头衔上看，工作一直差不多。我一直向 Daniel 汇报。但因为 Spotify 发展得太快了，每 6 到 12 个月就像在一家新公司工作。最初是瑞典和北欧的挑战。

So I came into Spotify in early 2009, late 2008. And my job then, I had been an entrepreneur, started some of my own companies in the back then very very early sort of feature phone smartphone space. So I had a bunch of knowledge there. I had sold a company to to Yahoo in the mobile space. I worked there for a while. I came back to Sweden. And then I met through a mutual friend Daniel Ek, the the CEO and co-founder of Spotify. And they had built the desktop product already, the free streaming desktop product. And it was amazing and I could try it. But they needed someone to figure out what to do with mobile. And because I had been an entrepreneur in that space I got that job. So my job was to to head up mobile for Spotify and figure out what the mobile offering would be, which was a challenge because obviously Spotify desktop was a free on-demand streaming application and back then specifically with edge networks you couldn't really stream at all in real time. The performance wasn't there. And also you could you couldn't fund that with an ads model. So it was a product and business model innovation that was a lot of fun. So that's how I started. Then after a few years I took on all of product development for Spotify. Then a few years later I actually took on the technology responsibility, sort of the CTO role for Spotify as well. And recently my official title is co-president of Spotify together with Alex Nordstrom. So we kind of run half of the company each. I run the product and technology side and he runs sort of the business and content side. So that's the super fast version. Aside from getting more responsibilities like taking on the technology department it has been sort of the same job by title. I've always reported to Daniel. But because Spotify has grown so much, every 6 to 12 months it's been like starting at a new company. First it was sort of a Swedish Nordic challenge.

互联网最初是从用户策展开始的。你选取一些好东西，比如人、书或音乐，将其数字化，放到网上，然后让用户来策展。这就是 Facebook、Spotify 等产品的模式。过了一段时间，世界从策展转向了推荐，不再是人工来做这些工作，而是由算法来完成。这是一个巨大的变化，要求我们和其他公司真正重新思考整个用户体验，有时甚至是商业模式。我认为我们现在正在进入的阶段，是从策展到推荐再到生成。我怀疑这将是一个同样巨大的转变，你最终将不得不重新思考你的产品。我们必须为推荐优先的时代重新思考用户界面和体验。那么在生成式领域，这意味着什么呢？目前还没有人真正知道。

The internet sort of started with curation of the user curation. So you you took something some good like people or books or music and you digitize it and you put it online and then you ask users to curate it. And that was your Facebook, Spotify and so forth. And then after a while the world switched from curation to recommendation where instead of people doing that work you had algorithms. And that was a big change that required us and others to actually rethink the entire user experience and and sometimes the the business model as well. And I think what we're entering now is we're going from your curation to recommendation to generation. And I suspect it will be as big of a shift that you will eventually have to rethink your products. We have to rethink the user interface and the experience for recommendation first era. And so what what does that mean in the generative area? No no one really knows yet.

然后是一个欧洲挑战，接着进入美国市场，然后我们上市了。所以实际上，尽管头衔和职位大体相同，但我感觉像是在很多工作之间跳来跳去。

And then it was a European challenge, and then it was getting into the US, and then we became a public company. So it's sort of as if I had jumped around between a lot of jobs, actually, even though it was largely the same title and role.

你的故事让我想到那句经典的话：“小心你擅长的事情”，因为你最终会承担越来越多。显然，这些年来你被赋予了越来越多的责任，所以很明显事情进展顺利，你做得很好。

Your story makes me think of the classic 'be careful what you're good at' because you end up taking on more and more. And clearly you've been given more and more responsibility over the years, and so clearly things are going well and you're doing well.

稍微换个话题。你现在在我的播客上。其实你自己也有一个播客，是关于 Spotify 产品故事的限定系列，我听了并且很喜欢。现在实时听到你的声音有点超现实，因为我最近一直在听那个，为这次对话做准备。两个问题。第一，是什么让你决定在有一份全职工作和很多事要忙的情况下，还推出自己的播客？而且据我所知，你播客的制作质量非常高。第二，从那个经历中，你在最终构建的产品方面学到了什么，以及如何与播客创作者产生共鸣？

Shifting a little bit. So you're on my podcast currently. You actually have your own podcast, which was kind of this limited series on the product story of Spotify, which I listened to and loved. And it's kind of surreal to listen to your voice in real time, because I've been listening to that recently in preparation for this conversation. Two questions. Just what made you decide to launch your own podcast knowing you had a full-time job and a lot going on, and the production value for your podcast was very high for what I could tell. And then two, just what did you learn from that experience in terms of the product you ended up building and just like empathizing with the podcast creator side?

我这么做有很多不同的原因。其中一个，而且不是小原因，我想和你一样，我热爱写作，内心有一个秘密的创作者梦想。嗯。很久以前我写过博客，而且我在内部写很多东西。在这样一家公司工作，你不可能对外写太多。但我热爱写作、演讲和展示。所以肯定有这个原因。然后，很大一部分是为了从产品角度与我们的主要用户之一——播客创作者——产生共鸣。不幸的是，我不是一个伟大的音乐家。我尝试演奏乐器等等，但我没有唱片。我唱得不好。但我决定做一个播客。这让我学到了很多关于成为创作者是什么感觉。比如，制作不同风格的播客。例如，我们想做一个制作成本较高的、带音乐的播客。然后你马上会遇到一堆问题。Spotify 实际上在解决这些问题上很有优势，但从版权角度来看，在播客中使用音乐真的很难。所以你就能理解播客创作者遇到的所有这些问题。你就能更好地解决它们。但最大的好处，也是我做公开播客的真正原因，是我之前通过某种技巧做了一个内部播客，我们可以把播客只限员工访问。我试图在内部想办法围绕 Spotify 建立更多文化，帮助新员工和现有员工定义我们是谁、我们犯过的错误、我们取得的成功，以及我们如何思考战略，特别是产品战略，因为我们在外部以技术和小队等闻名，但在产品战略方面不太出名。因为我比 Google Docs 更喜欢讲故事，所以我决定做一个内部播客。我到处采访了丹尼尔的直接下属，比如首席营销官、首席人力资源官、首席财务官等等。就问他们一堆事情。目的是让他们对员工来说更平易近人，因为我觉得听播客，你知道，即使是那些我不知道是谁的人，因为我从未见过他们，我也觉得我认识他们。我觉得我知道他们怎么想，而且我更喜欢他们了。所以秘密的想法是，如果你能比通过偶尔的会议或全体大会更好地了解你的领导者呢？所以我在内部做了这个，因为我是产品人，我们最终谈了很多关于产品战略的事情。内部员工真的很喜欢。所以下一次的问题是，如果还没在 Spotify 工作的人也能感觉认识 Spotify 的人呢？那太好了，因为大多数公司的领导者都非常不透明，看起来像是某种不真实的外星生物，当你在商业报纸上看到他们时。那么如果你听过他们讲一个小时的话呢？所以这就是大致的想法。所以结合了招聘工具、分享更多我们如何思考产品战略，而且我觉得这很有趣。我采访了很多聪明有趣的人，既有外部的也有内部的。

There were a bunch of different reasons why I did that. One is, and not a small one, I think like you, I love writing and I have this secret creator dream in me. Yeah. I used to write blog posts a long time ago, and I write internally a lot. You can't write that much externally when you work at a company like this. But I love writing and talking and presenting. So there was certainly that. And then no small part was to, from a product point of view, to empathize with one of our main constituents, the podcast creator. I'm unfortunately not a great musician. I try to play instruments and so forth, but I don't have any records. I don't sing very well. But I decided to make a podcast. And that taught me a huge amount about what it's like to be a creator. How you know, creating different styles of podcast. For example, we wanted to do a sort of higher production cost podcast with music. And then right away you run into a bunch of problems. As Spotify is actually pretty well positioned to solve, but still like it's really hard to have music in a podcast from a rights perspective. So you get to understand all these problems that podcasters have. And you can be better at solving them. But the biggest benefit and the real reason for doing the public podcast was that I had actually done an internal podcast through sort of a hack where we could gate the podcast only to employees. And I tried to figure out internally how to build more culture around Spotify and sort of help define for new employees and existing employees who we are, the mistakes we did, the successes we had. And how we think about strategy, specifically in product strategy, because we were quite well known externally for technology and the squads and all of these things. Not so much for product strategy. And because I love storytelling more than Google Docs, I decided to do an internal podcast. And I went around and I interviewed actually Daniel's direct reports. So the CMO, the CHRO, and the CFO, and so forth. And just asked them about a bunch of stuff. And the idea was to make them more approachable for employees, because I felt listening to podcasts, you know, even these people that have no idea who I am because I've never met them, I feel like I know them. I feel like I know how they think, and I just like them much more. So the secret idea was, what if you could get to know your leaders much better than you do through occasional meetings or town halls? So I did that internally, and because I'm a product person, we ended up talking a lot about product strategy. And people internally really liked that. So next time the question was, what if people that don't even work at Spotify yet could feel as if they knew people at Spotify? That'd be great, because most leaders in most companies are very opaque and appear as some sort of otherworldly creatures that aren't really real, I think, when you see them in like business papers or something. So what if you have heard them talk for an hour or so? So that was the general idea. So a combination of recruitment tool, sharing more about how we think about product strategy, and just because I think it was a lot of fun. I got to interview a bunch of smart and interesting people both externally and internally.

回顾来看，它达到了你预期的效果吗？

Did it have the effect that you were hoping for, looking back?

我觉得达到了。播客表现不错，而且我们没有给它任何特别的推广。我必须像其他人一样竞争，这也让你对这个问题产生很多共鸣：好吧，现在你有了一个产品，用户获取怎么办？你如何真正让人们去听？所以它确实达到了我想要的，因为我们有所谓的“入职日”，特别是在过去几年我们大量招聘的时候，我们实际上会把人飞到斯德哥尔摩参加某些入职培训，了解 Spotify。领导层会上台谈论他们做什么、他们的部门和战略等等。而且经常有人来告诉我，哦，你知道，我听了这个播客或这一集，这是我加入的关键原因之一，有时甚至是我加入的原因。所以这有点轶事性质，但至少有几十个人这么说过。所以这似乎有效。

I think it did. The podcast did well, and no, we did not give it our own sort of promotion. I had to compete as everyone else, which also gives you a lot of empathy for the problem of like, okay, now you have a product, what about user acquisition? How do you actually get people to listen to it? So it did achieve what I wanted in the sense that we have this thing called intro days, where especially in the past few years when we hired a lot, we actually fly people to Stockholm for certain onboarding sessions to learn about Spotify. And the leadership is on stage talking about what they do and their departments and strategy and so forth. And it's very common that people come and tell me that, oh, you know, I listened to this podcast or this episode, and it's at least one of the key reasons why I joined, or sometimes the reason why I joined. So it's sort of anecdotal, but it may be in the many tens of people at least have said it. So that seems to work.

这真的很有趣。再次，这在播客中出现了几次，就是内容在招聘、文化建设等不同方面的力量，听起来内部是原始目标，就是在内部围绕战略建立这种公司文化。那是原始目标。让高层领导更平易近人，从而缩短距离，然后也以娱乐的方式分享更多思考，而不是仅仅通过人们最终不会阅读的文档。我喜欢这个。

That's really interesting. Just again, and this comes up a few times in the podcast, is just the power of content in all these different ways for hiring, for culture building, and it sounds like internally it was the original goal is just internally build this company culture around strategy. That that was the original goal. Make senior leadership more approachable and so reduce the distance, and then also share more of the thinking in an entertaining way rather than just through docs that people end up not reading. I love that.

所以，正如我所说，我在听这个播客，真正有趣的是，我觉得第四集完全是关于 AI 的，而且我觉得那是你在 Spotify 内部利用机器学习和 AI 的第一次尝试，我觉得那导致了 Discover Weekly 和其他一些工具。那是几年前的事了，但现在听它很有趣，因为 AI 再次成为一个大问题。所以，我很好奇在产品团队中，你建议产品经理和产品团队如何在产品思维中考虑 AI，以及在日常工作中如何考虑？

So, I was listening to it as I said, and what was really interesting is I think episode four was actually all about AI, and I think your first kind of attempts at leveraging machine learning and AI within Spotify, and I think that's what led to Discover Weekly and a few other tools. And that was like years ago, but it's interesting listening to it now where AI is again, like, you know, a huge deal. And so, I'm curious very tactically on the product team, what you advise product managers and product teams on how to think about AI in their product thinking and also just in their day-to-day work.

我可以举几个例子。

I can give a few examples there.

我不认为我们比其他人更复杂，但至少我们在传统机器学习方面已经做了很长时间。在播客里，我谈到了互联网发展的几个阶段。一种思考方式是，互联网始于用户策展。你把一些好东西——人、书或音乐——数字化，放到网上，然后让用户来策展。那就是你的 Facebook、Spotify 等等。过了一段时间，世界从策展转向推荐，不再是人们做那些工作，而是由算法来做。那是一个巨大的变化，要求我们和其他人真正重新思考整个用户体验，有时还有商业模式。我认为我们现在正在进入的阶段是从策展到推荐再到生成。我怀疑这将是一个同样巨大的转变，你最终将不得不重新思考你的产品。所以这是一个视角。我倾向于跟我的团队说：尽管这都是机器学习，但我让他们把这看作是完全不同的东西。推荐时代是一种机器学习，生成时代是另一种。所以不要把它看作只是更多同样的东西，而要把它看作真正全新的东西。

I don't know that we're more sophisticated than anyone else, but it's what we've been doing at least in traditional machine learning for quite a long time. In the podcast, I talk about the journey of the internet in stages. One way to think about it is that the internet started with curation by users. You took something good—people, books, or music—digitized it, put it online, and asked users to curate it. That was your Facebook, Spotify, and so forth. Then after a while, the world switched from curation to recommendation, where instead of people doing that work, you had algorithms. That was a big change that required us and others to rethink the entire user experience and sometimes the business model as well. I think what we're entering now is going from curation to recommendation to generation. I suspect it will be as big a shift that you will eventually have to rethink your products. So that's one lens. I tend to talk to my teams about this: even though it's all machine learning, I ask them to think of this as something completely different. The recommendation era was one type of machine learning. The generation era is a different type. So don't think of it as just more of the same; think of it as something actually completely new.

我们学到的东西——嗯，有几件事。如果你看看这个新时代的大语言模型和扩散模型等等，有两类应用。正如我所说，在推荐时代，我们必须重新思考用户界面和体验。那么在生成时代这意味着什么？还没有人真正知道。像往常一样，有一堆迭代改进。所以，你知道，我们用这些大语言模型来改进我们的推荐。你可以有更大的向量，它们可以有更多的文化知识，你可以用它来对尚未有人听过的播客进行安全分类，等等。所以有很多明显的改进，我们正在做这些。但到目前为止，我们只真正做了一个严格意义上的生成式产品，即没有生成式 AI 就不可能存在的产品，那就是 AI DJ。这是一个我们思考了很久的概念。AI DJ 是：你按下一个按钮，一个人——一个数字化的人，有一个真人叫 X，我们把 X 数字化了。所以他现在是一个 AI——出现并跟你谈论你喜欢的音乐，并推荐音乐。你可以听，如果你不喜欢，你可以把他叫回来，他说，好吧，现在让我们听一些也许是几年前夏天的东西，或者这里有一些昨天在某个节目的最后一集里流行的新东西。所以那个产品没有生成式 AI 就不可能存在，既要生成声音，也要生成声音所说的内容。所以你可以有规模达到五亿人的个性化声音。我们有很多很多年都看到了这个用例。有时人们称之为收音机用例；我们内部称之为 Siri 意图用例，即当你完全不知道想听什么的时候。Spotify 在这方面并不好。Spotify 在你至少大致知道你想要什么的时候很好——如果是锻炼或晚餐，我们有很多选择。但如果你完全不知道，打开 Spotify 然后盯着它看是很困难的。人们曾经渴望地说，这是收音机擅长的一件事。说实话，收音机相当糟糕。它完全不个性化，不是按需的。你中途进来。它在很多方面其实很糟糕。但人们仍然常说它有一些好东西，我认为那个东西就是：你有一个旋钮，可以在不同情境之间切换。就像，不，无聊，无聊，无聊，无聊。好吧，这个不错。Spotify 从来没有那种模式：我不知道我想要什么，但我想循环浏览直到找到我喜欢的东西。我认为有了 AI DJ，我们实际上解决了这个用例。所以 X 出现并说，我要给你推荐一些你可以听的东西。如果你喜欢，你可以继续听，但如果你不喜欢，你可以再把他叫回来，然后换一种风格。出于某种原因，我们尝试了很多次，很长时间，但只是开始播放一首随机歌曲，没有任何关于为什么你会听到这个的上下文，从来没有成功过。所以那是我们第一次涉足以前不可能存在的产品。

And what we learned—well, a few things. If you look at this new era of large language models and diffusion models and so forth, there are two types of applications. As I said for the recommendation era, we had to rethink the user interface and the experience. So what does that mean in the generative era? No one really knows yet. There are, as usual, a bunch of iterative improvements. So, you know, we use these large language models to improve our recommendations. You can have bigger vectors, they can have more cultural knowledge, you can use it for safety classification on podcasts that no one has listened to yet, and so forth. So there's lots of obvious improvements and we're doing those. But so far, we've only really done one sort of real generative product in the hard definition, which is a product that couldn't have existed without generative AI, and that is the AI DJ. That's a concept we've been thinking about for a very long time. The AI DJ is: you press a button, a person—a digitized person, there's a real person named X, and we digitized X. So he's now an AI—comes on and talks to you about music that you like and suggests music. You can listen to it, and if you don't like it, you can kind of call him back, and he says, okay, now let's listen to something maybe from a few summers ago, or here's some new stuff that were trending yesterday in the last episode of something or other. So that product couldn't have existed without generative AI, both generating the voice and generating what the voice says. So you can have individualized personalized voice at the scale of half a billion people. We had the use case we had seen for many, many years. Sometimes people call it the radio use case; we call it the Siri intent use case internally, when you actually don't know what you want to listen to at all. Spotify wasn't that good. Spotify was good when you knew at least roughly, you knew the use case of what you wanted—if it was a workout or dinner, we had lots of options for all of those. But if you really didn't know at all, it was hard to open Spotify and just stare at it. People used to say longingly that this was the one thing that radio was good at. Radio was quite bad, to be honest. It's not personalized to you at all. It's not on demand. You come in in the middle of things. It's actually terrible in many ways. But people still often say that there was something good about it, and I think that something was the fact that you had a knob and you could just switch between contexts. It's like, no, boring, boring, boring, boring. Okay, this is good. Spotify never had that mode of 'I don't know what I want, but I want to sort of cycle through things until I find something that I like.' I think with AI DJ, that's actually the use case we managed to solve. So X comes on and says, I'm going to suggest something to you that you can listen to. If you like it, you can keep listening, but if you don't like it, you kind of bring him back again and you change genre. For one reason or another, we tried to solve that many times for a long time, but just starting to play a random song without any context as to why you would hear this just never worked. So that was our first sort of foray into a product that couldn't exist before.

关于你问的这方面的原则，我们学到了几个非常明确的原则。我非常喜欢的一个原则，根本不是我的原则——我认为它直接来自 Chris Dixon——就是容错用户界面的原则。我无法说清在早期机器学习时代，当我们说要从策展转向推荐时，我多少次看到一个设计草图是一个大的播放按钮。因为显然，这是你能做的最简单的用户界面。但如果你不了解你的机器学习的性能，你就无法为它设计。如果你的机器学习质量，如果你要有一个单一的播放按钮，那它必须达到 100% 或零预测误差。但事实从来不是这样，对吧？所以假设你有五分之一的中奖率。五分之四都是没用的。那么你需要一个用户界面，可能至少在屏幕上同时显示五个东西，这样你就有五分之一的概率在屏幕上看到相关的东西。所以你需要了解你的机器学习的性能来为它设计。它需要容错。而且通常你需要为用户提供一个逃生舱。如果你做了一个预测，但如果你错了，用户需要非常容易地说，不，你错了。我想去我的音乐库或去这里或那里。所以我们有容错用户界面的原则，以及一个与你算法当前性能相对应的用户界面。我认为这对生成式机器学习也是如此。我认为一个非常清晰的例子实际上是 Midjourney。你想想早期 Midjourney 在 Discord 频道里的用户界面，实际上生成一张图片非常非常慢。

And I think to your question of principles around that, there are a few pretty distinct principles that we've learned. One that I really like, that is not my principle at all—I think it is straight from Chris Dixon—is the principle of fault tolerant user interfaces. I can't say how many times during the early machine learning era, when we said we're moving from curation to recommendation, I saw a design sketch that was a single big play button. Because clearly, that is the simplest user interface you can do. But if you don't understand the performance of your machine learning, you can't design for it. The quality of your machine learning, if you're going to have a single play button, needs to be literally 100% or zero prediction error. And that's never the case, right? So let's say that you have a one in five hit rate. Four out of five things are duds. Then you need a UI that probably at least shows five things at the same time on screen, so you have a one in five chance of something being relevant on screen. So you need to understand the performance of your machine learning to design for it. It needs to be fault tolerant. And often you need an escape hatch for the user. If you make a prediction, but if you're wrong, it needs to be super easy for the user to say, no, you're wrong. I want to go to my library or to this or to that. So we have that principle of having a fault tolerant user interface and a user interface that corresponds to the current performance of your algorithms. And I think that is going to be true for generative machine learning as well. I think a very clear example actually is Midjourney. You think about the early Midjourney user interface inside the Discord channel, actually generating an image was very very slow.

生成高质量图像需要很长时间，他们本可以做一个银色按钮的东西，你输入提示词，等几分钟，得到一张图像，我觉得四次里有一次会是坏的。所以，你会四次里失望三次，而且每次要等一分钟。所以，四分钟后，你会觉得这是个烂产品。他们做的是快速生成四张同时的低分辨率图像。你可以说，显然他们的性能大概是四分之一。这就是为什么他们显示四张而不是六张。所以，四分之一通常还不错。你点击那一张，要么继续迭代，要么放大。这也是我认为人们在构建 UI 时理解生成式 AI 性能的一个例子。所以，你知道，这让我很受启发。

It took a long time to generate high quality image and they could have built silver button thing where you put in a prompt, you wait for minutes, you get an image, and I think one out of four times is going to be bad. So, you would have been disappointed three out of four times, and it's a minute each. So, like 4 minutes later, you'd be this is a shitty product. What they did was they generated four simultaneous low-res images very quickly. And you could say like So, So, apparently their performance was probably one in four. That's why they four showed four and not six. And so, one in four was obviously was usually pretty good. You click that one and either continue to iterate or scale it up. So, that's also an example of I think people understanding where the performance of generative AI was when they built the UI. So, that's something that, you know, I would be inspired by.

具体到 AI DJ，另一个原则是尽量避免炫耀技术的冲动，不要让这个语音角色说个不停。你必须记住，人们来是为了音乐。

And for the AI DJ specifically, another principle is to try to avoid this urge of just wanting to show off the technology and have this voice act talk and talk and talk and talk. You have to remember that people came there for the music.

所以，AI DJ 的原则来自团队，顺便说一句，这实际上是一个自下而上的产品。它需要很多支持。我们实际上需要大公司等等才能构建它。但这个想法是由团队自下而上构建的。所以，那里的原则就是尽可能少做，然后让开。我认为这非常有帮助。你知道，它不会告诉你天气如何，新闻里发生了什么，也不会没完没了地谈论这个乐队。它试图让你进入音乐，我认为这就是它有效的原因，因为它对我们来说效果非常好。

So, the principle for the AI DJ coming from the team, by the way, this was a bottoms-up product, actually. It required a lot of support. We actually required big companies and so forth to be able to build it. But the idea has been built by teams bottom-up. So, the principle there was literally to do as little as possible and get out of the way. And I think that was really helpful. You know, it's not telling you what the weather is and what happened in the news and going on and on and on about this band. It is trying to get you to the music, and I think that's why it's working because it is working very well for us.

我喜欢推荐和生成之间的这种区别。这引出了一个问题，我猜你看到了这种趋势，人们使用艺术家的目录自动生成音乐。比如一两周前出现的 Drake 和 The Weeknd 的事情。你觉得这会走向何方，艺术家如何适应这个音乐可以自动生成的世界，你知道，这个播放按钮就像所有东西都是生成的，而不仅仅是歌曲之间的 DJ？

I love this distinction between recommendation and generation. And this kind of begs the question of there's this trend that I imagine you're seeing of people auto-generating music using artists', you know, catalog. Like there's this Drake and The Weeknd thing that came out a week or two ago. Where do you think this ends up going, and how do you think artists adjust to this world where music can just be auto-generated, you know, this play button is like all of it is generated versus just like the DJ in between the songs?

首先，最大的警告是，这还非常早期。没有人知道这会如何发展，或者法律环境等等。但我认为它会产生很大影响。我认为如果我们谈两件事，一是它对音乐能做什么，二是权利状况。如果权利持有人得到补偿等等。所以，我们先单独谈第一件事。我认为一个有趣的例子是，在我成长的时候，Avicii 出现了。想想很有趣，因为 Avicii 并不被现有音乐行业认为是真正的艺术家，因为他不会真正演奏乐器，也不会唱歌。他只是坐在电脑前，用 DAW（数字音频工作站）。所以，它并不被认为是真正的音乐。而现在我们都认为它是真正的音乐，他拥有巨大的真正音乐才华。所以，我认为现在我们可能处于人们说这不是真正的音乐，是某种虚假的阶段。我认为思考这些扩散模型的方式，如果它们能足够好地生成音乐，可能是一样的，就像一种乐器。它只是一种更强大的乐器，我们可能会看到一种新型创作者，他们不精通乐器，他们无法组装完整的管弦乐队，无法实现他们脑海中的东西。而他们现在可以生成非常新的东西。我还认为，顺便说一句，AI 音乐和真实音乐之间的这种区别是不存在的。当然，非常有才华的真正音乐家正在使用 AI 来变得更好，帮助创造新想法。所以，这种区别并不真正存在。一切都将是 AI。问题在于百分比。这使问题更难，因为你不能谈论它是否应该存在。你必须谈论应该存在多少百分比，以及谁可以使用它。但我认为思考它的方式可能是作为一种乐器，可以帮助创造大量艺术。我认为这对你来说不是新闻，你可能经常使用这些东西，但我认为如果你不使用这些生成模型，会有一种看法，你告诉它创作一首热门歌曲，你就会得到它。那不是它的工作方式。实际上，这些模型所做的，因为它们听了大量音乐，它们非常擅长做听起来与现有音乐非常相似的东西。实际上，原创非常难。从某个角度看，随着现在创作更通用音乐变得更容易，真正独特将比以往任何时候都更难。所以，我仍然认为在创造真正独特的东西方面会有巨大的技巧。我的希望是，DAW 和技术飞跃带来的结果是，你得到了一个全新的流派，比如 EDM，你无法用管弦乐队或现场演奏来真正制作它。也许我们会看到这些技术带来全新的音乐风格。我认为那会非常令人兴奋。所以，这是积极的一面。但然后你有权利问题，我对此有很多同理心。Spotify 特别以前见过这种情况。所以，我们有过一次不同的技术转变，那就是在线下载音乐、盗版和点对点的技术转变。所以，首先是点对点的技术转变，对消费者来说很兴奋。更多消费者开始比以往任何时候都听更多音乐。我认为我们现在就处于生成式 AI 的这个阶段。有一种新技术，但它也需要一种新的商业模式，创作者和行业才能真正参与并从中受益。如果这显然是自私的说法，因为我们是创新这种商业模式的重要部分，但我仍然认为这是必要的。我希望我和我们都能成为其中的一部分。所以，我认为我们已经看到了第一部分，技术转变。这里可能会有很多讨论和混乱，我对此有很多同理心。但我认为我们还没有看到第二部分。什么样的模式可以使这成为利益？盗版之后实际发生的是音乐行业变得比以往任何时候都大。不仅仅是同样大，而是比以往任何时候都大。我认为这项技术也可能发生这种情况。但我们正处于开始阶段。

First big caveat this, this is just super early. No one knows anything, you know, about how this is going to play out or the legal landscape and so forth. But I think it's going to have a lot of impact. And I think if we talk about two things, one is what it could do for music. The other is the rights situation. And if rights holders are getting compensated and so forth. So, we talk about the first thing in isolation. I think an interesting example is right about when I grew up, Avicii came along. And it's interesting to think about because Avicii was not really considered by the existing music industry as a real artist cuz he couldn't really play an instrument and he couldn't sing. And he was just sitting with this computer in this DAW, digital audio workstation. And so, it wasn't really considered real music. And I think now all of us consider it very real music and that he had tremendous real musical talent. So, I think right now we're probably in the phase where people say this isn't real music and it's somehow fake. I think the way to think about these diffusion models if and when they get good enough at generating music is probably the same, like an instrument. It's just a much more powerful instrument, and we'll probably see a new type of creator that wasn't proficient at an instrument and they couldn't assemble a full orchestra and do the thing that they had in their head. And they can now generate very new things. I also think, by the way, that there is this distinction between AI music and real music that doesn't exist. For sure, very talented real musicians are using AI to get better and to help create new ideas. So, that distinction doesn't really exist. It's all going to be AI. The question is what percentage. Which makes the problem harder cuz you can't talk about if it should exist or not. You have to talk about what percentage should exist and who gets to use it or not. But I think the way to think about it is probably as an instrument that could help create a huge amount of art. And I think this is not news to you who probably use these things a lot, but I think if you don't use these generative models, there is the perception that you tell it to create a hit and you will get that. That's not how it works. Actually, what these models do is because they've been listening to a lot of music, they are very good at doing something that sounds very similar to what already exists. Actually being original is very hard. And from one point of view, as it now gets easier to create more generic music, it will actually be more difficult than ever to be truly unique. So, I still think there will be tremendous skill in creating something truly unique. And my hope would be that what happened with the DAW and that technology jump was you got a whole new genre like EDM that you couldn't really produce it with an orchestra or live. And maybe we'll see completely new music styles with these technologies. I think that would be very exciting. So, that's on the positive side. But then you have the rights issue, which I have a lot of empathy for. And Spotify specifically has seen this before. So, we had a different technology shift like this, which was the technology shift online downloads of music and piracy and peer-to-peer. So, first it was the technology shift in peer-to-peer, and it was exciting for consumers. More consumers started listening to more music than ever. And I think that's where we are now with generative AI. There's a new technology, but it also required a new business model before creators and industry could actually participate and benefit from this. And if that's obviously self-serving to say because we were a big part of innovating that business model, but I still think that's what's necessary. And I hope that's what I and we could be part of. So, I think we've seen that first part, the technology shift. And there would probably be a lot of discussion and chaos here, which I have a lot of empathy for. But I think we haven't seen the second part yet. What is a model where this could be a benefit? What would actually happen after piracy is that the music industry got bigger than ever. Not just as big, but bigger than ever. And I think that could happen with this technology as well. But we're right in the beginning.

所以，沿着同样的思路，你教的另一件事是，所有真正伟大的产品都必须耍某种魔术。

So, along the same lines, something else you teach is this idea of all truly great products have to pull some kind of magic trick.

这个话题在你的播客里经常出现，我想你在其他地方也提到过。考虑到你在这里谈到的所有内容，感觉在某种意义上，一切都会像魔法一样，因为 AI 已经融入其中了。

This comes up in your podcast a lot, and I think you mentioned this other places. And thinking about all the stuff you're talking about here, it feels like in a sense, everything's going to feel like magic cuz AI is kind of baked into it.

我想当我们做 AI DJ 时，我们做了一个小版本。当人们第一次听到它时，我们能在用户测试中看到那种反应。当他们喜欢……所以，那里的魔法技巧是，他们怎么能录下这个人说这么多不同的话，因为它是在谈论我的音乐。所以，魔法技巧显然不是录了一个人说话，而是生成的。而且那种魔法效果会消退。你现在经常听到它，等等，但那是其中一种魔法技巧。所以，我仍然认为这个概念很重要，它似乎与产品病毒式传播和起飞相关。我认为第一次使用 DALL-E 或 Stable Diffusion 或 Midjourney 时也是一样，它完全像是一个魔法技巧。显然没有魔法。只是数据和统计。但我认为达到那个点，迭代产品到第一次感觉像魔法的程度是非常有帮助的。而且这通常只是让性能达到一定水平、缩小范围、移除东西的问题。我认为有很多微调，让你跨越那条线，从“很酷、令人印象深刻，但不是魔法”到“感觉像魔法”。我不明白这怎么能做到。

I think when we did the AI DJ, we did a small version of that. When people first listened to it, we could see that reaction in user testing. When they like, so, the magic trick there was that how could they record this person saying so many different things because it's talking about my music. So, the magic trick was obviously didn't record a person saying it's generated. And that magic trick wears off. You hear it all the time now and so forth, but it was one of those magic tricks. So, I still think that concept is important and it seems to correlate with products sort of going viral and taking off. And I think it was the same using something like DALL-E or Stable Diffusion or Midjourney the first time, it completely seemed like a magic trick. And obviously there is no magic. It's just data and statistics. But I think getting to that point and iterating a product to the point where it feels like magic the first time is very helpful. And it's often a question of just getting the performance to certain levels, scoping down, removing things. There's a lot of fine-tuning, I think, that makes you cross that line from it's cool and impressive, but not magic to it feels like magic. I don't understand how this could be done.

是的，这让我想起了 GPT 的发布，它最终成为历史上最大、增长最快的产品，就像魔法技巧的缩影。感觉就像真正的魔法。

Yeah, it reminds me of the launch of GPT, which ended up being the biggest, most, fastest-growing product in history, and it's like the epitome of a magic trick. It's like feels like actual magic.

绝对。绝对。对大多数人来说，它仍然非常……实际上，对我们很多人，甚至对研究人员来说，它有点神奇。没有人真正完全理解。所以，我想世界上可能还留有一些魔法。

Absolutely. Absolutely. And to most people, it is still very actually, to a lot of us and even to researchers, it's a little bit magical. No one really understands fully. So, I guess there's maybe some magic left in the world.

绝对。而且我认为很多人担心不理解那里发生了什么。

Absolutely. And I think a lot of people are worried about not understanding what's going on there.

转到你们在 Spotify 构建产品的方式。Spotify 以推广小队和部落的概念而闻名。如果我错了请纠正我，但你们似乎已经放弃了那种方法。

Shifting to the way you all build product at Spotify. So, Spotify is kind of famous for popularizing this idea of squads and tribes. And correct me if I'm wrong, but you guys have kind of moved away from that approach.

是的，没错。

Yeah, that's right.

好的。所以，我很想了解你为什么转变，以及你从那种构建产品的方法中学到了什么。然后你们现在如何组织团队？你们现在做什么？

Okay. So, I'd love to understand just like why you shifted and what you kind of learned from that approach to building product. And then just like how do you organize the teams now? What do you do now?

这是我们早期非常关注的事情。事实证明，我们把这些东西命名为小队、部落等是明智的。这并不真的是……好吧，也许有点刻意品牌化。但我们发明这些名字不是为了品牌。我们发明它们是因为我们认为这是一个好的结构。我们需要给事物命名。而且这是早期互联网时代，所以你可以随意编造。所以，这对我们当时的情况非常好，而且肯定帮助了我们的招聘。它现在对我们来说有点成本，因为人们仍然认为我们是这样组织的。而且在这个规模下，这种组织方式效率不高，或者即使你现在重新开始，因为我们学到了更多。但我认为最大的区别是，小队这个想法具体有两点。它们应该是小而全栈的。一个小队应该大约七个人，应该有前端、后端、移动端、QA、敏捷教练等等。而且它应该非常自主。这就是我们真正转变的地方。所以，首先，随着公司发展，以七名工程师为增量扩展会带来大量开销。所以，显然我们现在的团队往往大得多，每个经理至少是两到三倍。所以，可能有 14 人左右而不是七人。而且更少的开销角色。所以这是其一。随着你学到更多，它看起来更传统，而且随着扩展是合理的。我认为我们挣扎的第二件大事是，当时我加入时，Spotify 的平均年龄是……我是说，我是最老的，这是 14 年前。我认为平均年龄可能不到 30 岁。大多数科技公司都是这样。所以，我们来自瑞典，这是一个与美国不同的文化。我喜欢瑞典文化的很多方面，我认为我们设法保留了最好的部分。但瑞典是一种非常自下而上的自主文化。有一幅著名的画，关于在瑞典和美国如何做决定，我认为在美国只是一个层级，而在瑞典是一种圆圈。你坐在一个圆圈里，没有人在中间。没有领导者等等。有趣。所以，我认为从文化上讲，我们深受这种超级自主的东西启发。我认为自主的想法非常合理，也是正确的，那就是我们过去和现在都在雇佣我们能找到的最聪明的人。我们为此支付高薪。所以，如果你雇佣聪明人，一种思考方式是你在租用脑力。所以，如果你租用所有这些昂贵的脑力，然后不给他们独立思考的空间，那听起来不聪明。那你应该雇佣不那么聪明的人，降低成本之类的。所以，我认为你必须给予大量自主权，才能真正最大化你投资的价值。所以，给予人们大量空间来尽可能发挥他们的才能和能力是非常合理的。但问题是，如果你把自主权放在组织的末端。而且，如果你结合一个非常初级的组织，我们当时就是这样，很有可能你只会产生热量。你会有 100 个小队，有 100 个策略，朝 100 个方向跑。而且，你知道，Spotify 曾经在那个阵营。我的意思是，尽管这样，我们确实取得了一些成就，但我很难说我们这样做是高效的。所以，我们做了几件事。团队结构更传统，团队更大，开销更少。我们一直在专门研究在组织中把自主权放在哪里？因为极端情况是在末端，我们就在那里。另一个极端可能是在顶部。比如说像 Twitter。有一个人。两者都有问题。如果你把它放在末端，你会产生很多热量。如果你把它放在顶部，你需要一个能力很强的人，他一个人有很多能力，但你必然会遇到瓶颈。所有决定都必须经过那里。

This was something that we focused a lot on early. And it turned out to be smart of us to name these things into squads and chapters and so forth. It wasn't really well, maybe it was sort of deliberately branding. But it wasn't for purposes of branding that we made it up. We made it up because we thought it was a good structure to use. And we needed names for things. And this was the early internet era, so you were allowed to like make things up. And so, it was very good for where we were at the time, and it certainly helped us in recruiting. It's become a little bit of a cost to us because people still think that we organize that way. And it's not a very efficient way of being organized at this scale or maybe even if you started over right now because we've learned more. But I think the big difference is the idea with the squad specifically was twofold. They were supposed to be small and sort of full stacks. A squad should be about seven people and it should have, you know, front and back end, mobile, QA, Agile coaches and so forth. And it should be very autonomous was the idea. And that's really what we shifted. So, first of all, as you grow the company, scaling in increments of seven engineers just creates a ton of overhead. So, obviously our teams now tend to be much bigger, maybe two, three times that at least per manager. So, maybe have like 14 or something instead of seven. And just less overhead roles. So that's one. It looks more traditional as you learn more and it's reasonable as you scale. The second big thing I think we struggled with was back then when I joined, the average age at Spotify was I mean, I was the oldest and this was 14 years ago. I think the average age was probably under 30 or something. And it was in most tech companies. And so, we had coming from Sweden, which is a different culture than the US. And I love a lot of things about Swedish culture and I think we managed to keep the best parts. But Sweden is a very sort of bottoms-up autonomous culture. There's this famous drawing of how you make decisions in Sweden and in the US, I think it's just a hierarchy and in Sweden, it's kind of a circle. You sit in a circle, no one is in the middle. There is no leader and so forth. Interesting. So, I think by sort of culture, we were very inspired by this super autonomous thing. And I think the idea with autonomy is very reasonable and the right one, which is we were and we are hiring the smartest people we can find. And we pay high salaries for that. So, if you're hiring smart people, one way to think about it is you're renting brain power. So, if you're renting all of this expensive brain power and then you give them no room to think for themselves, that doesn't sound smart. Then you should actually hire less smart people and like keep your costs down or something. So, I think you have to give a bunch of autonomy to actually maximize the value of the investment you're making. So that's very reasonable that you would give a lot of space for people to use as much of their talent and capacity as possible. But the problem with that is if you put autonomy very far towards the leaves of the organization. And also, if you combine that with having a very junior organization, which we did back then, there's a fair chance that you're just going to produce heat. You're going to have 100 squads with 100 strategies running in 100 directions. And, you know, Spotify has been there in that camp. I mean, we managed to get somewhere for sure in spite of this, but I'd be I'd struggle to say we were like efficient in doing that. So, we've done a few things. The team structure is more traditional, larger teams, less overhead. And we've been specifically working with where in the org do we put the autonomy? Cuz the extremes are at the leaves and we were there. The other extreme may be at the top. Let's say maybe something like Twitter. There's one person. Both have problems. If you have it at the leaves, you're going to produce a lot of heat. If you have it at the top, you need someone with a lot of capacity and he alone has a lot of capacity, but you are by definition going to bottleneck. All decisions have to go through there.

而且丹尼尔，他的性格就不是想自己做所有决定。他想最大化吞吐量，而不是成为瓶颈。所以问题是，如果不在顶层也不在最底层，那该放在哪里？我们发现——我觉得这一点一点都不反直觉，我认为大多数公司都是这样——大概在副总裁这个层级。所以如果你有丹尼尔，然后有高管层，我和其他人，然后就是副总裁层。这是一个很好的组合。与其让公司里只有一个人思考——只有丹尼尔——其他人只是执行，不如在副总裁层，像这样的公司里有几十到几百人拥有很大的自主思考空间。这样你就能获得大量的思想自由，人们朝不同方向思考，但不是 8000 人。而且副总裁层的人数量不少，但通常也很资深。他们有很多模式识别能力。所以我认为这解决了——如果你把它看作一个优化问题，这算是一个很好的优化空间。所以现在 Spotify 的自主权在副总裁层通常很高，然后在这些层级周围较低。

And Daniel, it's just not his personality that he even wants to make all the decisions. He wants to maximize throughput rather than bottleneck it. So the question is, if it's not at the top and not at the very bottom, where do you put it? And what we found, which I don't think is very contrarian at all—I think this is the case in most companies—is around the VP level. So if you have Daniel, then you have the C-level, myself and others, then you have the VP level. That's a good mix. Instead of having one person in the company think—only Daniel—and the rest just do, you have on the VP level in a company like this many tens to maybe hundreds of people that have a lot of autonomy to think. So you get a good amount of freedom of thought and people thinking in different directions, but it's not like 8,000 people. And these people on the VP level are quite a lot of them, but they're also usually quite senior. They have a lot of pattern recognition. So I think that solves for—if you think of it as an optimization problem, it's kind of a good optimization space. So the autonomy level in Spotify now tends to be quite high at the VP level and then lower around those levels.

你说自主权，那实际上是什么意思？比如说，播客产品的副总裁对发生的事情有很大的发言权，而且上面的人参与不多？我不确定，比如上面的人参与度如何？我知道玛雅是产品副总裁，我相信是负责播客产品的。

And when you say autonomy, what does that actually mean? Is it the VP of, say, the podcasting product has a lot of say over what happens, and there's not a ton of—I don't know—like how involved are people above? And I know Maya's the VP of product, I believe, for the podcast product.

没错。我觉得她总有一天会来上播客。嗯，这对她的自主权意味着什么？实际上，这意味着我会让玛雅来定义我们在播客领域的战略——我们如何与众不同，为什么播客创作者愿意来这里？而在另一家公司，我会制定那个战略，或者在另一家公司，丹尼尔会制定那个战略。同样，比如这个想法来自我们个性化团队的一个成员。所以那是他们下的赌注。所以他们有自主权去下这种赌注和定义战略。用户界面也一样。我们有一个体验团队。稍后可以谈谈组织架构。但我给了体验副总裁很大的自主权来定义和提出我们想做的事情。而在其他公司，比如我会自己定义所有这些。

Exactly. Who I think is going to come on the podcast someday. Um, what does that mean in terms of autonomy for her? Practically, it means that I would ask Maya to define a strategy for what we do in podcasting—how are we going to be different, why would a podcaster want to be here? Whereas in another company, I would make that strategy, or in another company, Daniel would make that strategy. Same with the idea, for example, came from one of our personalization team. So that was a bet that they made. So they have autonomy to make those kinds of bets and define strategies. Same with the user interface. We have an experience team. Can talk about the org structure later. But I put a lot of autonomy on the VP of experience to define and suggest what it is that we want to do. And in other companies, I would define all of that myself, for example.

再深入一点。我知道你对团队组织方式有强烈的看法，以及组织如何帮助你优化特定目标。你在这方面有什么想法？关于组织的影响和你优化的目标，你学到了什么？

Just going even a little bit further here. I know you have strong opinions on the way to organize teams and how the organization helps you optimize for specific things. What are your thoughts along those lines, and what have you learned about the impact of organization and what you're optimizing for?

是的，所以我谈的是一个理想化的谱系——也许不是理想化，而是夸张的谱系。实际上不是——没有什么是完全真实的。但你创造极端来阐明观点，对吧？所以在一个谱系上，你有像亚马逊这样的公司，以两个披萨团队闻名。没有依赖。你尽量最小化依赖，以便并行运行。团队之间相互竞争，甚至在同一项目上，等等。但他们直接接触用户。所以好处是，如果你有一个想法，到达用户的时间非常短。这对他们有效。它产生了 Kindle，产生了 Alexa，产生了很多非常新颖的东西。这里有几个有趣的缺点。一个缺点——我非常佩服杰夫·贝索斯能看到这一点——如果团队相互竞争，激励就是隐藏你的结果，隐藏你的代码。这应该会导致一个没有平台杠杆的组织，因为没有人合作。我认为要么他有这个洞察，要么因为他看到了这一点，他不得不这样做，但他以极力推动硬 API 而闻名。比如，如果你不为你的技术创建硬 API，你就出局了。如果你想想，必须是这样，否则没人会做。而硬 API 本质上就是——每个人都知道如何使用这个 API 并连接到这个团队进行交互。没错。你必须向他人暴露你的技术。你必须维护这些 API，而且它们必须非常结构化，否则整个事情会崩溃，因为每个人都应该竞争，因为没有激励。你必须集中强制。有趣的是，尽管理论上他们是最不适合拥有结构化平台的，但我认为因为他们如此强制，他们才做了亚马逊网络服务，因为他们有如此硬定义的 API，因为这个规则，他们更容易把它翻出来暴露给世界其他地方。而如果你看看像谷歌这样的公司，我认为他们在外部化 API 方面更挣扎，也许因为它太友好和软性，所以他们内部不需要那么硬的 API。因为没有竞争。人们可以直接进入彼此的代码。所以这是一个有趣的轶事。但重点是，你在那里更快，但合作会很困难。所以你会看到——也许有点夸张——有时你会看到同一页面上来自不同团队的多个搜索框。顺便说一句，这在 Spotify 也发生过。你会看到现在播放视图中来自不同团队的多个提示条，因为他们在我们处于自主模式时工作。每个人都在跑。所以你得到了速度的好处，但你也得到了把组织架构图和复杂性传递给最终用户的缺点。但显然这对亚马逊来说是正确的选择，因为他们是一家万亿美元的公司。但在另一个谱系上，你有像苹果这样的公司，也是一家万亿美元的公司。所以显然两种模式都有效，你永远不会在 iPhone 上看到来自同一团队的两个搜索框。那是由接近单一个体的东西集中组织的。所以他们处于可能是世界上最大的职能组织中。他们做的事情一样多。如果你想想苹果涉及什么，我的意思是，他们当然做了我们所做的一切。他们有音乐服务、播客服务、有声书，还有无数其他服务。所以并不是他们的问题更简单。然而他们构建的东西感觉更像是由一个开发者为一个用户构建的。所以他们集中化，他们有这种瓶颈功能，一切都必须经过它并决定如何与其他一切配合。

Yeah, so I talk about an idealized spectrum—or maybe not idealized, but exaggerated spectrum. It's not really—nothing is really true. But you create extremes to make a point, right? So on one spectrum, you have something like Amazon, which is known for two-pizza teams. No dependencies. You try to minimize dependencies so you can run in parallel. Teams compete with each other, even on the same project, and so forth. But they have direct access to the user. And so the benefit here is if you have an idea, the time to get to user is very low. And it has worked for them. It's produced Kindle, it produced Alexa, it's produced a lot of very novel things. There are a few interesting downsides here. One downside that I'm extremely impressed with Jeff Bezos for seeing is if you have teams that compete with each other, the incentives are to hide your results, hide your code. And that should make for an organization that gets no platform leverage because no one is cooperating. And I think either he had that insight or because he saw this, he had to do this, but he's well known for pushing extremely hard on hard APIs. Like if you don't create hard APIs to your technology, you're out. And if you think about it, it has to be that way because otherwise no one would do it. And a hard API is essentially—everyone knows how to use this API and connect to this team to interface with. Exactly. You have to expose your technology to others. You have to maintain those APIs, and they have to be very structured because otherwise the whole thing would collapse as everyone's supposed to compete because there are no incentives. You have to centrally force that. And interestingly, even though theoretically then they're the worst position to have a structured platform, I think because they forced it so hard, they were the ones who did Amazon Web Services because they had such hard-defined APIs because of this rule that it was easier for them to turn it inside out and expose it to the rest of the world. Whereas if you look at something like Google, I think they struggled more with externalizing their APIs maybe because it is so friendly and soft, so they didn't need as hard APIs on the inside. Because there was no competition. People could just go into each other's code. So it's an interesting anecdote around it. But the main point is you're faster there, but it's going to be hard to cooperate. And so you will see something like—maybe exaggerating a bit—sometimes you'll see multiple search boxes on the same page from different teams. And this has been true in Spotify, by the way, as well. You've seen like multiple toasters on the now playing view coming up from different teams because they're working when we were in the autonomous mode. Everyone running. And so you get the benefit of speed, but you get the drawback of kind of shipping your org chart and shipping complexity to the end user. But clearly that's been the right choice for Amazon because they're a trillion-dollar company. But then on the other spectrum, you have something like Apple, who's also a trillion-dollar company. So clearly both models work where you would never see two search boxes from the same team popping up on an iPhone. That is centrally organized by something that is close to a single individual. So they are instead in what's probably the world's biggest largest functional org. They're doing as much. If you think about what goes into Apple, I mean, they certainly do everything we do. They have music service, podcast service, audiobooks, and they have a billion other services. So it's not like they have an easier problem. And yet they built something that feels more like it was built by a single developer for a single user. So they centralize and they have this bottlenecking function that everything has to go through and be decided how it fits with everything else.

这样做的优点是用户体验更简单，不会把组织架构图暴露给用户，也不会增加复杂性，但缺点是速度慢。虽然没有确切数据，但我听在 Apple 工作过的人说，某个东西花了 7 年才上市，因为它只能在流程里排队等着。所以你有这些极端情况，我觉得最有趣的例子是，当你双击 iPhone 的电源键时，Apple Pay 会弹出来。这个决定是怎么做出的？你可以想象，所有服务团队都希望双击那个按钮时弹出自己的服务。所以必须有人决定：应该弹出音乐吗？应该弹出 Apple Pay 吗？还是其他什么？所以他们在那里有不同的结构。而在集中与分散的光谱上，因为我们的策略是单一应用，我们尝试加入多种类型的内容，这些内容在后端有非常不同的商业模式，比如收入分成、版税、图书交易等等，但都整合到单一用户体验中。这就是我们的策略。我们认为用户体验和保持简单是最重要的。所以我们选择了更集中的模式，这些不同的垂直业务，比如音乐业务、播客、有声书业务，都必须通过一个统一的推荐组织。那是另一个问题。你知道，向哪个用户推荐哪个内容？应该是书、播客还是音乐？如何权衡它们？而且如果每个人都构建自己的 UI，用户界面很容易变得极其复杂。音乐团队构建了他们的 UI，然后有人在其上添加功能。所以我们选择这样优化。但这是基于我们的策略，我认为两种模式都有效。

And so that has the benefit of the user experience being simpler and not shipping the org chart and increasing complexity, but it also has the drawback of speed. Without having facts on it, I've heard people working at Apple said like, yeah, took 7 years to get that thing to market because it just had to wait in the pipeline. So you have these extremes and I think that the most interesting example to think about is when you double click the power button on an iPhone, that Apple Pay comes up. Like that decision, how did that happen? You can imagine that all the services team would like to pop up when you double click that button. And so someone had to decide should music come up? Should Apple Payments come up? Should something else come up? And so they have a different structure there. And on that spectrum of centralized versus decentralized, because of our strategy, which is we're single application, trying to add or not trying to, we have added multiple types of content with actually very different business models on the back end, you know, rev shares and royalties and book deals and so forth, into a single user experience. That is our strategy. We think the user experience and keeping that simple is the most important thing. So we've chosen more of the centralized model, where these different sort of vertical businesses, if you think about it, the music business, podcast, audiobooks business, they have to go through a single recommendation organization. That's another problem. You know, which one do you recommend to which user? Should be a book or podcast or music? And how do you weigh them against each other? And also the user interface could easily get incredibly complicated if everyone built their own UI. The music team built their UI and then someone had added features on top. So that's how we chose to optimize. But it is based on our strategy and I think both models work.

本期节目由 Eco 赞助。上个月，Eco 用户平均获得 84 美元的返现奖励。怎么做到的？用 Eco，个人理财的未来。Eco 是对错位金融系统的升级，提供一个像银行一样运作的应用，但几乎去掉了所有中间商，帮助即使是最擅长理财的人也能在更短时间内自动优化。如果你支付房租能获得奖励呢？或者在线订餐和购物能获得奖励？甚至每月储蓄也能获得奖励？然后想象一下，如果你因为获得奖励而再次获得奖励呢？使用 Eco，你可以在一些最喜欢的商家消费，自动获得 5% 的返现。此外，Eco 的年化收益率奖励更像是 80 美元，而不是 80 美分。还有 Eco 积分，这是世界上第一个开放奖励系统。你在 Eco 应用里几乎做任何事都能赚取积分。Eco 正在努力让这些积分成为史上最有价值的积分，所以早加入是有好处的。听起来好得难以置信？访问 eco.com/lenny，注册入职，了解为什么这不是真的。Lenny 播客的听众如果参加 Eco 欢迎会，将获得存款超过 1000 美元的独家 4% 年化收益率。了解更多信息，请访问 eco.com/lenny。那就是 eco.com/lenny。

This episode is brought to you by Eco. Last month, Eco users earned an average of $84 in cash back rewards. How? With Eco, the future of personal finance. Eco is the update to a misaligned financial system, providing an app that works just like your bank, but removes almost all of the middleman, helping even the best money optimizers optimize in less time automatically. What if you earned rewards for paying your rent? Or got rewarded for ordering food and shopping online? Or even earn rewards for saving each month? And then imagine if you got rewarded again, just for getting rewarded. With Eco, you can spend at some of your favorite merchants and automatically get 5% cash back. Plus, Eco's APY rewards look more like $80, not 80 cents. And then there are Eco points, the world's first open rewards system. You earn them whenever you do almost anything in the Eco app. Eco is working to make these points the most rewarding points ever, so it pays to be early. Sound too good to be true? Go to eco.com/lenny, sign up for an onboarding, and find out why it isn't. Lenny's podcast listeners who attend an Eco welcome session will get an exclusive 4% APY on deposits over $1,000. Learn more at eco.com/lenny. That's eco.com/lenny.

你举的这两个例子很有意思，Apple 和 Amazon。它们是世界上最大的两家公司，而且处于这个光谱的两个极端。有趣的是，大多数公司都处于中间位置。我想知道，处于极端是否有好处，而这最终变得非常重要。

It's interesting these two examples you gave, Apple and Amazon. They're two of the biggest companies in the world and they're like at the extremes of these two into the spectrum. And it's interesting most companies are somewhere in the middle. I wonder if there's just like a benefit to being in an extreme and that ends up being really important.

我也这么认为。在几乎所有行业，都有微笑曲线的概念，对吧？你要处于微笑曲线的两端。这就是大的商业机会所在，而不是中间。所以在组织模式上可能也是如此。

I think so. In almost all industries, you have this smiling curve concept, right? Where you want to be at the extremes of the smiling curve. And that's what big business opportunities are, but not in the middle. So it's probably true in terms of organizational models as well.

说到极端，我想谈谈下大赌注。你们最近有一个大型发布活动，基本上重新设计了 Spotify 的主信息流，让它更像现在应用的发展方向，比如 TikTok 或 Reels 的感觉，就是流式播放，你开始看到视频，音乐开始播放，有些人喜欢，有些人不喜欢。我很好奇，作为产品负责人，你如何看待长期思考，以及如何应对那些说“这什么鬼变化？我讨厌变化。别改了”的人？你是怎么想的？你听谁的？忽略谁？你怎么知道什么时候该坚持？你如何处理？

Speaking of extremes, I want to talk a bit about taking big bets. So you guys had this big launch event recently where you basically redesigned the whole primary feed of Spotify to make it feel more like where kind of apps are going, like TikTok Reels feel of just, you know, stream and you start hearing videos and music starts playing and some people loved it, some people did not. And I'm curious as a product leader, how you think about thinking long term and dealing with people that are just like, what the hell's change? I hate change. Stop changing things. How do you think about that? Who do you listen to? Who do you ignore? How do you know what to stay the course? How do you approach that?

是的，你太客气了。Twitter 上有很多负面反馈。所以让我深入一些细节，因为我认为这对收听节目的产品人来说是一个有趣的教训，我认为很少有公司会谈论这个。因为你不想谈论那些完全按你预期发展的事情，也不想谈论那些没有按你预期发展的事情。所以我会讲讲我们试图实现什么，以及我们学到了什么。Spotify 主要是一个后台应用，很长一段时间里，我们被认为非常擅长后台音乐和播客推荐，当手机在口袋里，你在听 EDM 播放列表或流行播放列表时。我们非常擅长在后台插入另一首 EDM 曲目或另一首流行曲目。但我们一次又一次从用户那里听到的是，他们说他们被困在品味泡沫里。所以，你知道，我爱我的 Spotify，我喜欢这个，但我现在对 EDM 有点厌倦了，Spotify 没有推荐全新的东西。如果你思考这个问题，它听起来可能和推荐问题相似。这只是另一个推荐问题。但实际上根本不同。因为当你在 EDM 播放列表里推荐另一首 EDM 曲目时，你有很多信号表明用户喜欢 EDM。但如果你要推荐一个全新的流派，根据定义，你一无所知。因为如果你知道，那对他们来说就不是新的。所以你什么都不知道。回到命中率，当你向用户推荐全新的东西时，命中率会非常低。所以帮助人们走出品味泡沫的问题并不像听起来那么容易。我们不能真的拿一些，你知道，一些可能不典型的流派。比如，我是雷鬼顿的超级粉丝。这在瑞典并不常见，如果你看我其他的资料，主要是 EDM，你可能猜不到。Spotify 也猜不到。

Yeah, you're being very kind. There was a lot of negative feedback on Twitter on some of that. So let me actually kind of dig into some detail because I think this is a really interesting lesson for product people listening to this, that I think few companies talk about. Because you don't really want to talk about everything that went exactly as you thought they would and you don't want to talk about the things that didn't go exactly as you thought they would. So I'll go through kind of what we are trying to achieve and what we learned. So Spotify is mainly a background application and for a long time we've been considered very good at background music and podcast recommendation when the phone is in your pocket and you're listening to like an EDM playlist or, you know, a pop playlist or something. We're really good at inserting another EDM track there or another pop track there or something like that in the background. What we hear from users again and again though is that they say that they get trapped in a taste bubble. So, you know, I love my Spotify, I love this, but I am a little bit bored with EDM now and Spotify is not suggesting something completely new. And if you think about that problem, it may sound similar to the recommendation problem. It's just another recommendation problem. But it's actually fundamentally different. Because when you're recommending another EDM track inside the EDM playlist, you have a lot of signal from that user that they like EDM. But if you're going to recommend a completely new genre, by definition, you have no idea. Because if you had no idea, it wasn't new to them. So you can't know anything. So back to hit rate, your hit rate is going to be incredibly low when you suggest something completely new to the user. So this problem of helping people get out of their taste bubble isn't easy as it sounds. And we can't really take some, you know, some genre that maybe isn't typical. So I'm a big fan of reggaeton, for example. It's not typically that common in Sweden and if you would look the rest of my profile, it's kind of EDM heaviness, you probably wouldn't have guessed it. And Spotify wouldn't have guessed it.

所以，如果我在后台听我最喜欢的 EDM 播放列表，或者我的金属乐播放列表，金属乐在瑞典非常流行，我们很难在中间插入一首雷鬼歌曲。你知道，大多数人会认为 Spotify 坏了。他们到底在想什么，对吧？所以这行不通。因此，为了帮助人们打破他们的口味茧房，你需要一些不同的东西。你需要一个命中率可以很低的东西。而且你需要人们预期它很低。所以当我们在后台推荐东西时，我们的命中率至少需要十分之九。也许一个失误是可以的，但如果你有五个失误，你会认为我们弄坏了你的播放列表和你的会话。我们需要一个十分之一就是成功的东西。如果你在十次尝试中找到一颗宝石，你会非常高兴。所以你需要一个完全不同的范式。而且你还需要能够快速浏览很多候选内容，对吧？因为命中率太低了。你不能每项花三分钟。就像，好吧，我不喜欢这个，但离下一个出现还有两分钟。你需要快速地说，不，不，不。所以显而易见的候选方案是这些信息流类型的体验，你可以快速浏览大量内容。你预期命中率会低得多，如果你不喜欢，成本非常低。你只需滑动。这就是为什么当人们想要打破口味茧房，或者当他们来到 Spotify 听一些全新的东西时，通常是因为他们在这些服务上找到了它，比如 TikTok 或 YouTube 之类的，在那里他们接触到大量新内容。所以人们一直在向我们要求这些工具。这就是我们想要解决的问题。

So if I'm listening to my favorite EDM playlist in the background or maybe my metal playlist, metal is very big in Sweden, it's really hard for us to just insert a reggaeton track in the middle of that. You know, most people are going to think Spotify's broken. What the hell are they thinking, right? So that doesn't really work. So in order to help people break out of their taste bubbles, you need something different. You need something where your hit ratio can be very low. And you need people to expect it to be very low. So when we recommend things in the background, our hit ratio needs to be at least nine out of 10. Maybe one dud is okay, but if you get, you know, five duds, you're going to think we broke your playlist and your session. We need something where one out of 10 is a success. If you find one gem out of 10 tries, you're very happy. So you need a completely different paradigm. And you also need to be able to go through many candidates quickly, right? Because the hit rate is so low. You can't take 3 minutes per item. It's like, okay, I didn't like this and it's still like 2 minutes left before the next one comes on. You need to quickly say, no, no, no. So the obvious candidates for this are these feed type experiences, where you can go through lots of content. You're expecting the hit ratio to be much lower and if you don't like it, the cost is very low. You just swipe. And this is the reason why people, when they want to break out of their taste bubbles or when they come into Spotify and listen to something completely new, it is usually because they found it on one of these services, like a TikTok or YouTube or something, where they get exposed to lots of new content. So people were asking us for these tools. And so that's what we wanted to solve for.

所以我们构建了一堆功能，类似信息流的结构，你可以浏览一个包含许多曲目的新流派，或者一个包含许多剧集的播客频道，甚至完整的播放列表。我们实现了这些，并把它们放在叫做子信息流的东西里。所以在当前的体验中，这已经在全球范围内推出，如果你点击播客子信息流，你会得到一个播客剧集的信息流。点击音乐子信息流，你会得到一个播放列表的信息流，你可以快速浏览许多播放列表，如果你不理解名字，你可以快速听到它们听起来的样子，查看几首曲目，了解这是否适合你。如果你去搜索和浏览页面，你可以找到全新的流派，快速浏览。这些功能按我们的预期工作。人们会浏览它们，当他们想找新音乐时会去那里。他们浏览并保存新歌曲。所以它们按我们的预期工作。

And so we built a bunch of features, feed-like structures where you can go through either a short new genre with many tracks, or a podcast channel with many episodes, or even full playlists. And we implemented those and we put them in something called sub-feeds. So in the current experience, and this is rolled out worldwide, if you click the podcast sub-feed, you get a feed of podcast episodes. Click the music sub-feeds, you get a feed of playlists, where you can quickly, you know, you can go through many playlists and if you don't understand the name, you can quickly hear what they sound like and check out a few tracks and understand if this is for you. And if you go to the search and browse page, you can find completely new genres that you can quickly go through. And so those are working as we intended. People go through them, go to them when they want to find new music. They browse through them and they save new songs. So they're working as we intended.

没有按我们预期工作的事情是，当用户一次又一次地要求这个时，我们把这些东西的总和放在了首页上，因为人们非常关注发现，而且我们可以清楚地看到发现与 Spotify 上的留存率等相关性有多强。但我们误判或未能，或者说了解我们自己的首页的是，它现在的工作方式，你可以在 Twitter 评论中看到，如果你去掉愤怒的声音，试着看看他们在说什么，他们说以下内容：我在定量数据中也清楚地看到，如果你看看人们在 Spotify 首页上做什么，当前的首页，几乎 90% 是我们所说的回忆。所以要么是进入你已经参与的会话，要么是进入你知道想要去的特定播放列表，或者至少是一个特定的用例。所以你带着高意图进来，你实际上知道你想要什么，也许只有 10% 的时间是真正的发现，比如我不知道我想要什么。所以如果你想想，那是 90% 的回忆和 10% 的发现。当我们测试那个设计时，子信息流当时有效，现在也有效，但当我们测试它们在首页上的总和时，我们有点把它从 90/10 切换到了 10/90，所以 10% 的回忆和 90% 的发现。虽然人们想要发现，但他们可能不想要 90% 的发现而不是 90% 的回忆。所以如果你再看 Twitter 上的评论，他们说的是，嘿，我再也找不到我的播放列表了，这些东西在哪里？他们并不是真的在抱怨发现，他们在抱怨他们不再得到的东西。我们也可以在定量数据中看到这一点。然后你可以看到流量从首页转移到搜索和资料库，这是一个明显的迹象，表明人们试图找到他们再也找不到的东西。你甚至可以看到人们试图使用这些发现工具，这些工具是为了快速理解新事物而优化的，来做回忆，比如我知道我想要的那个锻炼播放列表在哪里？这实际上对于回忆来说是非常糟糕的 UI。这有点像老虎机，对吧？如果你能到达那个锻炼播放列表，那是非常不可预测的。它是为发现新事物而优化的，而不是为回忆现有事物。当你做回忆时，你想要密集的 UI，屏幕上有许多项目，因为你知道你在找什么，所以你不需要很多空间。当你发现新事物时，你想要很多用户界面，很多像素，而且你可能想要声音，因为你不知道它是什么。

The thing that didn't work as we intended was when users asked us for this again and again, we took sort of the sum of these things and we put it on home because people ask so much about discovery and we can see clearly how correlated discovery is with retention on Spotify and so forth. But what we misjudged or failed to, or rather learned about our own homepage is that the way it works right now, and this is what you can see in the Twitter comments if you remove the angry voices and sort of try to see what they're saying, they're saying the following: I see quite clearly in the quantitative data as well that if you look at what people do on Spotify's homepage, the current one, it is almost 90% what we call recall. So it is either getting to a session that you're already in or a specific playlist that you know you want to get to, or at least a specific use case. So you come in with a high intent, you actually knew what you wanted, and maybe only 10% of the time is it true discovery like I don't know what I want. So if you think about that, it's 90% recall and 10% discovery. When we tested that design, so the sub-feeds were working and are working, but when we tested the sum of them on home, we kind of switched it from 90/10 to 10/90, so 10% recall and 90% discovery. And while people want to discover, they probably don't want 90% discovery instead of 90% recall. So if you then look at the comments on Twitter, what they're saying is like, hey, I can't find my playlist anymore, like where are these things? They're not really complaining about the discovery, they're complaining about the things they don't get anymore. And we could see this in the quant data as well. Then you can see traffic shifting from home into search and into library, which is a clear sign people are trying to find the things they can't find anymore. And you can even see people then trying to use these discovery tools, which are optimized for quickly understanding new things, to do the recall, like where is that workout playlist I know I want? And it's actually very bad UI for recall. It's kind of like a slot machine, right? Very unpredictable if you ever get to that workout playlist. It was optimized for finding new things, not for recall of existing things. When you do recall, you want the dense UI with many items on screen because you know what it is you're looking for, so you don't need a lot of real estate. When you're doing discovery of new things, you want a lot of user interface, a lot of pixels, and you probably want sound because you don't know what it is.

所以我们从 UI 中学到的东西，我认为这里可能有一点产品嫉妒，你总是看其他体验，如果你环顾四周，你可能会认为大多数其他产品，比如你看 YouTube 这样的东西，他们的首页正是那样：它是一个巨大的单一项目发现信息流，只有新项目，人们似乎不会愤怒地发推文说他们有多生气。他们说他们喜欢 YouTube，它是一个大产品。我认为我们发现的是，我们实际上在首页上做得非常好的一件事，那就是支持你同时处于多个会话中。所以你可能正在听两个播客和一个有声书，然后实际上，我只是想找到那个锻炼播放列表。我不记得它的名字，但我知道它是锻炼。我们实际上把那部分做得非常好。我敢说比其他体验好得多，在其他体验中，你实际上必须去某个标签页，进入资料库，开始浏览才能回到你原来的位置。所以也许这是路径依赖的。如果我们有，你知道，因为我们把回忆做得很好，我认为当人们找不到它，当他们不能再做回忆时，他们相当不满。我们真的不想失去那个，因为那是我们做得好的事情之一，而且被低估了。我的收获是，实际上我们比其他体验做得更好，所以我们当然想保留它。所以我们做的是，现在我们只是……

So what we learned about our UI, and I think there's maybe a little bit of product jealousy here, you always look at other experiences, and if you look around, you could be forgiven for thinking that most other products, if you look at something like YouTube for example, their homepage is exactly that: it's a huge single-item discovery feed with only new items, and people don't seem to tweet angrily about how angry they are. They say they love YouTube and it's a big product. And I think what we discovered was that we actually did something really well on our homepage, which was supporting you being inside of multiple sessions at the same time. So you could be in the middle of two podcasts and an audiobook, and then also actually, I just want to get to that workout playlist. I don't remember the name of it, but I know it's workout. We actually did that part really well. I would venture to say much better than the other experiences where you literally have to go to some tab and into library and start browsing to get back to where you were. And so maybe it's path dependent. If we had, you know, because we have done recall pretty well, people got, I think, reasonably upset when they couldn't find it, when they couldn't do the recall anymore. And we really don't want to lose that because it was one of the things we did well and underestimated. And my takeaway is actually we do it better than other experiences, and so we certainly want to keep that. So what we did was now we're just...

更新假设以实现同样的目标，即这些东西是有效的，当人们想要发现时，他们会使用它们，而且它们似乎有效。它们还可以变得更好。你知道，从机器学习的角度来看，你正处于一个爬山式的旅程中。但问题是，你如何确保每当人们觉得自己陷入“我被困在品味泡沫中”的境地时，他们明白这些东西就在那里，而且易于使用。所以现在我们有一个 Home 的版本，我们显然也在测试，其中这些东西非常可用但自愿使用，你仍然可以进行所有的回忆。所以在我看来，这就是我们进行 AB 测试的原因，因为你想对此保持科学态度，并且你想尽可能多地了解你自己的产品和用户。现在我分享了很多经验教训。也许我们应该保密，但我的直觉是这会让产品变得更好。

Updating the hypothesis to achieve the same goal, which is these things are working and when people want to discover, they use them and they seem to work. They can also get better. You know, you're on this hill-climbing journey from a machine learning point of view. But the question is, how do you make sure that whenever people feel that they are in that 'I'm trapped in my taste bubble' situation, they understand that these things are there and they're easy to use. So now we have a version of Home that we're also testing obviously, where these things are very available but voluntary, and you can still do all of the recall. So from my point of view, this is the reason we AB test, because you want to be scientific about it, and you want to learn as much as possible about your own product and your users. And now I'm sharing a lot of the learnings. Maybe we should keep them to ourselves, but my hunch is that it's going to make it a much better product.

当你进行这次重新设计时，你告诉你的团队什么？

What did you tell your teams when you went into this redesign?

因为我做过几次这样的重新设计，我认为产品开发有两种根本不同的类型。一种是设计新功能。它很难做，但人们可以自愿使用。所以你做了 AI DJ，有些人喜欢，那很好。如果你不喜欢，它也不会让你更糟。但当你重新设计时，情况就棘手得多，因为参与重新设计不是自愿的。所以即使对不喜欢它的人来说，也有成本。你在这里遇到了一个非常棘手的问题，即会有两种反馈。一种是你做了某事，而且是对的，但人们因为你的改变而生气。另一种是你做了某事，但不对，人们也生气，但理由充分。那么你如何区分这两者呢？我想当我们和团队讨论这个问题时，我解释过。我认为可以这样类比：你有你的实体桌面。你的电脑在一个地方，你的铅笔在这里，你的笔记本在那里。然后我进来，把所有的东西都重新排列了。而你已经花了，在我们的情况下，可能 12 年，适应了那个布局。即使我有大量定量数据证明我的新布局更好，你也会生气，因为你在旧布局中很高效，而且很难区分这两者。最经典的例子是 Facebook 的信息流，当它变成单一信息流时，人们非常不满，但结果它解决了很多用户问题，你不需要自己跑遍整个 Facebook 去收集事件。所以有一些方法可以理解你是做得更好但打破了人们的习惯，还是做得不好。例如，一个方法是查看没有那种行为的新用户群组与老用户群组，等等。所以我们在做之前和团队一起经历了所有这些。我说这将会很痛苦，可能会有很多推文，因为我们完全做对的可能性很低。因此，这对团队来说并不太难。这很难，你知道，你想回应人们，但正确的做法是倾听、理解、尝试新的假设，真正弄清楚发生了什么。所以我现在已经做了三四次了。一次，三次，可能一次不成功，两次成功。所以有点知道我要面对什么。所以这几乎就像你惩罚自己，非常痛苦，但也是最令人兴奋的事情。我认为任何产品人员都知道，最简单直接的做法是在你现有的基础上迭代。没有风险，你不会被解雇，用户也不会生气。但每个人也都知道，最终，如果你不采用新技术、新范式等，你就会被取代。你必须找到尝试新事物的平衡。这就是当你在软件领域工作时，你有 AB 测试这个工具，并且对它保持科学态度。当你构建硬件时，情况更糟。如果你错了，你就错了，你无法更新。

Because I've done this a few times, like redesigning, I think there are two fundamentally different types of product development. One is designing a new feature. It is hard to make, but it's voluntary for people to use. So you do the AI DJ, some people love it, that's fine. If you don't like it, it didn't make it worse for you. But when you redesign, it is much more tricky because it's not voluntary to participate in the redesign. So there is a cost, even for people who don't like it. You have a very tricky problem here, which is there are going to be two types of feedback. One is you did something and it was right, but people are upset because you changed stuff. The other is you did something and it wasn't right, and people are also upset, but for good reasons. So how do you separate these two? I think I explained this when we talked through this with my teams. I think the analogy to think about is you have your physical desktop. You have your computer in one place, your pencil over here, your notebook over there. And I come in and I just rearrange all of it. And you have spent, in our case maybe 12 years, with that setup. It doesn't matter if I have a lot of quantitative data that my new setup is better. You're going to get upset because you are effective in this old setup, and it's hard to tell those apart. The most classic use case is the Facebook newsfeed, which people were very upset about when it became a single newsfeed, but it turned out to solve a lot of user problems that you didn't have to run around all of Facebook collecting events yourself. So there are some ways of understanding if you made it better but people's habits are broken, or if it's not better. One thing is, for example, to look at new user cohorts that don't have that behavior versus old user cohorts, and so forth. So we went through all of this with the teams before we did it. I said this is going to be painful, probably going to be a lot of tweets, because chances that we get it exactly right are very low. So for that reason, it hasn't been very hard on the team. It is hard, you know, you want to respond to people, but the right way to do it is to listen, understand, try new hypotheses, to really figure out what's going on. So I've done it maybe three or four times now. One, three, maybe one unsuccessfully, two successfully. So kind of knew what I was getting into. So it's almost like you punish yourself, very painful, but also the most exciting things. And I think any product person knows that the easiest and most straightforward thing to do is to iterate around where you are. There's no risk, you're not going to get fired, no users are going to get angry. But everyone also knows that eventually, if you don't adopt new technologies, new paradigms, etc., you're going to get replaced. You have to find this balance of trying new things. That's when you work in software, you have this tool of AB testing and being scientific about it. When you build hardware, it's worse. If you're wrong, you're wrong, you can't update.

我喜欢这个故事，非常感谢你的分享。我想在这样的大型发布中，你实际上无法提前进行 AB 测试，因为新闻季。他们会说：“哦，天哪，看看 Spotify 在做什么。”所以我想你在那里有点受限。你无法真正提前测试这个。

I love this story, so appreciate you sharing it. I imagine also with a big launch like this, you can't actually AB test it ahead of time because of the press season. They're like, 'Oh my god, look what Spotify's doing.' So you're kind of limited there, I imagine. You couldn't really test this ahead of time.

最难的是，如果你尝试完全新的东西，MVP 需要非常大。所以你可以构建一个新的 UI，但如果你没有为单一项目信息流做算法，你就无法判断这是正确的想法但机器学习不佳，还是机器学习不佳或其他。你必须构建很多东西，而且成本相当高。这实际上是最大的痛苦，不是来自外部的反馈，而是你必须在内部承担的成本。你承担了很多成本，你真的希望你是对的。在我们的案例中，首页的更改对我们来说并不难。重要的是，我们能否帮助你打破品味泡沫这个基本假设确实有效。然后你更新获取漏斗到那个体验。但我认为问题是，你需要让这么多东西到位，才能说你是否因为做得不够好而得到假阴性。我认为在这些大规模重写中，最大的机会是每个人都需要更新一切，然后你才能知道你是对还是错。

The hardest thing about this is if you're trying something completely new, the MVP needs to be very big. So you can build a new UI, but if you didn't do algorithms for a single item feed, you can't tell if it was the right idea but with poor machine learning, or you have poor machine learning or the other. You have to build a lot, and it gets quite expensive. That's actually the biggest way it's painful, not really the feedback from the outside, it is the cost you have to take on the inside. You incur a lot of cost and you're really hoping you're right. And in our case, the changes on the homepage aren't that hard for us to do. The important thing is that the underlying hypothesis of can we help you break out of your taste bubble actually works. And then you update the acquisition funnels into that experience. But I think the problem is that you need to get so many things in place to be able to say if you get a false negative just because you didn't do it well enough. That's the biggest chance I think with these big rewrites where everyone has to update everything before you can know if you're right or wrong.

帮助理解什么不工作、什么工作以及你想改变什么的过程是怎样的？我想你正在看一堆数据，一些推文之类的。那种战术上的“哦，糟糕，事情没有按我们预期的方式发展，我们应该这样做”是什么样的？

What was that process like of helping understand what is not working and what is working and what you wanted to change? Like I imagine there's a bunch of data you're looking at, some tweets, things like that. What was kind of like the tactical 'oh shoot, something's not going the way we expected, here's what we should do'?

嗯，我们测试了信息流，但主页信息流我们是在之后推出并测试的。我们测试了几种不同的变体。然后我们得到了数据，我们更多地看了定量数据。我们做了很多用户研究，人们坐下来使用信息流，以理解并建立我们自己的心智模型，了解什么有效，什么无效。然后显然，你当然会看用户反馈。有些用户非常擅长表达什么不工作。其他人则不擅长表达什么不工作。所以这可能很难解析。但当然，这也是一个因素。然后，一旦你做到了，你就有定量数据可以看。然后你坐下来推理你认为对和错，不同的假设是什么，什么有效，什么无效。

Well, we tested feeds, but the home feed we rolled out and tested afterwards. And we tested it out on users with a few different variants of it. And then we got the data back and we looked more at the quantitative data. And we do a lot of user research where people sit and use the feeds to understand and build our own theory of mind of what is working and what is not working. And then obviously, you look at user feedback, of course. And some users are very good at expressing what isn't working. Others are not as good at expressing what isn't working. So it can be hard to parse that. But certainly, that's a factor as well. And so then, once you do that, then you have quantitative data to look at. And then you sit and reason through what you think is right and wrong, what are the different hypotheses, what is working, what is not working.

然后就是不断更新、反复测试，直到你找到、证明或推翻你的假设。尽量以科学的方式去做。另外，我认为当你投入了大量时间在某件事上时，最大的风险就是变得过于珍视它。你必须残酷一点。你必须 100% 相信某件事，直到数据说不行。然后你再 100% 相信另一件事。

And then just update and test again and again until you find, until you prove or sort of disprove your hypothesis. Trying to be as scientific as possible about it. And also, I think the biggest risk also when you've invested so much time in something is you know, getting precious about things. You have to just be brutal. You have to believe in things 100% until the data says no. And then you believe in something else 100%.

这听起来容易，做起来很难。

That sounds easy. It's very hard to do.

确实如此，以至于当你这样做时人们会不高兴，因为出于某种原因，人们不喜欢别人改变主意。这应该是我们希望每个人都做到的。我会喜欢一个说“我看了数据，意识到这其实是对的，现在我相信这个”的政治家。但我们讨厌这样做的政治家。你知道，他们让人觉得不可信，我们会嘲笑他们。所以，我认为这是对任何人来说最大的风险。你必须变得不带感情，只看数据中的证据。然后，你知道，如果你这样做，你就继续前进，最终到达你想去的地方。你解决同样的问题，但你会适应。

It is, to the extent that people get upset when you do it because for some reason people don't like when people change their mind. It is what we should want from everyone. I would love a politician who said, "I've looked at the data and I realized actually this is right and now I believe this." But we hate politicians that do that. You know, they feel untrustworthy and like we ridicule them. So, I think that's the biggest risk with anyone. You just have to be like unemotional. And just look at the proof in the data. And then, you know, if you do that, you just move on and then you get to where you want to be. You solve the same problem, but you adapt.

我真的很喜欢这种哲学。本质上，这就是“坚定观点，灵活持有”的理念，对吧？

I really like that philosophy. Essentially, it's the idea of strong opinions loosely held, right?

没错，正是如此。这听起来很简单，但很难。

Exactly. Exactly what it is. And it sounds so easy, but it's hard.

对，因为正如你所说，人们不喜欢、不尊重改变主意的人。他们会说：“哦，我明白了。他们一直错了，而且他们对自己错了还那么自信。”

Right, cuz to your point, people don't like, don't respect someone changing their mind. They're like, "Oh, I see. They were wrong the whole time and they were so confident about being wrong."

是的，没错。而且不清楚为什么。这应该是我们想要的。但我认为这与人类心理有关。我们实际上倾向于喜欢预言家和那些在数据很少的情况下持有非常坚定观点的人。那些是我们喜欢的人。而那些看了大量数据并真正适应的人，我们不喜欢。不知道为什么。我们有缺陷。有缺陷的生物。

Yeah, exactly. And it's unclear why. It is what we should want. But I think it has something to do with human psychology. We actually tend to love prophets and people who hold very strong opinions with very little data. Those are the people we like. People who look at a lot of data and actually adapt, we don't like. Not sure why. We're flawed. Flawed creatures.

当然。沿着这些思路，你最近有没有改变主意的事情，可能会让你想到“哦，对”之类的？

For sure. Is there something that you recently changed your mind about along these same lines that maybe comes to mind of like, "Oh, yeah."

不，我认为这些关于我们自己的设计系统和首页做得很好、可能比其他人更好的经验，我们不想把它们和洗澡水一起倒掉，或者不管另一个表达是什么。我认为这是目前最大的收获。我实际上非常非常高兴。

No, I think these learnings about what our own design system and homepage does really well, maybe better than others, that we don't sort of want to wash out with the bathwater or whatever the other expression is. I think that's the biggest current learning. I'm actually very, very happy about.

是的，我喜欢了解到我们正在做一些我们不一定真正意识到的事情，而且做得很好，也许我们应该更多地利用这一点。

Yeah, I love learning that we're doing something really well that we didn't really realize necessarily and maybe we should lean into that more.

没错。

Exactly.

换个方向，Shirish Murarka 建议我问你一些事情。我相信他是你们董事会的成员。他建议我问你关于 10% 规划时间的事情。那是怎么回事？

Going in a somewhat different direction, Shirish Murarka suggested I ask you something. He's on your board, I believe. And he suggested I ask you about your 10% planning time. What is that about?

这是 Shirish 在 YouTube 工作以来长期使用的一个概念。大致意思是，你不应该花超过 10% 的时间在规划上，而应该把时间花在执行或构建上。这意味着如果你按季度工作，比如 10 周，你应该花 1 周规划。由于我们以 6 个月为增量工作，所以我们尝试花 2 周规划。我们大致做到了。实际上，当我们谈论组织模型时，要感谢 Airbnb 的 Brian Chesky，我认为他是最早采用这种逆向组织模型的人之一。他比硅谷大多数人更苹果风格。他也以 6 个月为增量工作。所以他在这方面也有很多经验。这就是 10% 规划时间的含义。我认为如果你发现自己规划的时间远超过这个比例，要么是你规划太多，要么是你的执行期对于那么多规划来说太短了。这是一个经验法则，但我发现它有效。

This is a concept that I think Shirish has used for a long time ever since he worked at YouTube. And the idea is that roughly, you shouldn't be spending more than 10% of your time planning versus executing or building. Which means that if you work in quarterly sort of 10 weeks, you should spend 1 week planning. Since we work in sort of 6-month increment, so we try to spend 2 weeks planning. And we're roughly successful. And this is actually, when we talk about org models, give a shout-out to Brian Chesky at Airbnb, who is actually one of the first, I think, to have these more contrarian org models. He's much more Apple-esque than most of Silicon Valley. He also works in 6-month increments. So, he has a lot of experience in that as well. So, that's what the 10% planning time is. And I think if you find yourself planning much more than that, you're either planning too much or your execution period is just too short for that amount of planning. It's a rule of thumb, but I find that it works.

我问了几个产品经理该问你什么，实际上他们是在 Spotify 工作的，我还没告诉你。有人指出你总是给房间带来很多能量和清晰度。他们认为这是你非常擅长的。关于这一点的重要性，或者作为领导者如何做好这一点，你学到了什么？

I asked a few PMs what I should ask you, PMs that work at Spotify, actually, that I haven't told you. And someone pointed out that you always bring a lot of energy and clarity to a room. That's something they see you as really strong at. What have you learned about just importance of that or just how to do that well as a leader?

嗯，听到这个很高兴。我之前不知道。所以，我在想该怎么回答。我认为能量，我不知道。我想我只是对我所做的事情感到兴奋。我一直对技术感到兴奋。我喜欢看到新事物。我的核心驱动力仍然是这个想法，你知道，你看到一些还不存在的东西，我想你会感同身受。你会想：“哇，我想知道那能不能存在。那太酷了。”然后为了让人们去做，你试图分享那种兴奋。所以，我认为我无法对我不兴奋的事情带来很多能量。所以我必须做我真正相信并且感到兴奋的事情。然后，也许能量就会更自然地流露出来。不幸的是，到目前为止，Spotify 一直处于允许大量创新的阶段，我甚至被要求尝试做新的酷东西。也许对于纯粹的优化阶段，我的能量会少一些。关于清晰度，我一直喜欢尝试解释事情。众所周知，理解某事物的最好方法是尝试向别人解释它。所以，我试着到处向那些没有要求的人解释事情。不是为了显得聪明，而是为了看看我是否真的理解了。所以，也许就是这种练习。关于这一点，我确实要求我的下属领导者，并要求他们要求他们的领导者，总是解释自己。当我们谈到自主权等等时，我认为我们不会承诺每个人都必须同意。但我认为我们应该向所有员工做出的承诺是，即使他们不同意，他们也应该有权理解你为什么做出这个决定。我认为不可接受的是说：“不，我们要这样做，因为我更资深。我已经见过很多次了。你不够聪明。”所有这些。我认为你必须解释自己。所以，你欠一个解释。我觉得这很有价值。回到理解某事物的唯一方法是解释它，因为通常结果是，如果你自己不能解释它，你可能自己也没有真正理解它。有时，我认为你可能拥有好的产品直觉，但无法表达出来。但大多数时候，当人们说那里有东西，但他们无法解释时，他们实际上自己也不理解。很多时候，那里实际上什么都没有。而且，如果你作为产品人员能解释它，那么这些知识就被分享了。所以，这对组织来说变得更加有效。

Well, that's great to hear. I didn't know that. So, I'm trying to figure out what to answer. I think that the energy, I don't know. I guess I'm just excited about what I do. I've always been excited about technology. I love seeing new things. My core drive is still this notion of, you know, you see something, which I think you'll empathize with, that doesn't exist yet. And you're like, "Wow, I wonder if that could exist. That would be so cool." And then in order to get people to do it, you try to share that excitement. So, I don't think I can bring a lot of energy for something I'm not excited about. So, I kind of have to work on things I actually believe in and that I'm excited about. And so, maybe then the energy comes more naturally. Unfortunately for me so far, Spotify has been in this phase where a lot of innovation is allowed and I'm even asked to try to do new cool things. Maybe I would have less energy for a pure optimization phase. On the clarity, I've always liked trying to explain things. It's a well-known fact that the best way to understand something is to try to explain it to someone else. So, I try to go around explaining things to people who didn't ask for it. And not to sound smart, but to see if I actually understood it. And so, maybe it's that practice. And on that note, I actually do ask my leaders that work for me and I ask them to ask their leaders to always explain themselves. And I think when we talked a little bit about autonomy and so forth, I don't think we promise everyone that they have to agree. But I think the promise we should make to all employees is that even if they don't agree, they should be entitled to understand why you're making the decision. What I don't think is acceptable is to say, "No, we're going to do it this way because I'm more senior. I've seen this a bunch of times. You're not smart enough." Like all of those things. I think you have to explain yourself. So, you owe an explanation. And I find that valuable. Back to like the only way to understand something is to explain it because it usually turns out that if you can't explain it yourself, you probably don't really even understand it yourself. Sometimes, I think it's possible that you can have product instincts that are good, but you can't express them. But most often, when people say there's something there, you know, but they can't explain it, they actually don't understand themselves. And many times there actually isn't anything there. And also, if you can explain it as a product person, that knowledge is now shared. So, it just becomes much more effective for the organization.

所以我有时会试着稍微挑衅一下别人，当人们问艺术和科学的比例时，我会说这是 0% 的艺术、0% 的魔法、100% 的科学。那是因为我想迫使人们去尝试解释它。我认为我们历史上用艺术和魔法这个词来形容任何我们还无法解释的东西。遗传学在成为科学之前是魔法和艺术。量子物理学在成为科学之前是魔法。而最近，智力和创造力在成为统计学和 LLM 之前是艺术和魔法。所以我试图推动人们说：“你确定你能解释这个吗？”因为那会迫使人们深入思考。所以也许这就是为什么人们觉得我有时能带来清晰。

So, I sometimes try to provoke people a little bit and say, you know, when people ask how much is art versus science, I say it's 0% art, 0% magic, and 100% science. That's because I want to force people to try to explain it. I think we used the word art and magic historically for anything that we couldn't yet explain. Genetics was magic and art until it was science. Quantum physics was magic until it was science. And most recently, intelligence and creativity was art and magic until it was statistics and an LLM. So I try to push people to say, 'Are you sure you can explain this?' Because that forces people to think through. So maybe that's why people think I sometimes bring clarity.

我喜欢这个。顺着这个思路问一个问题。你有没有推荐的解释系统或方法？是像写文档那样写出来？还是用某种风格解释？还是说只要自然就好？

I love that. Question along those lines. Is there a system or an approach to explaining that you recommend? Is it just like write it out in a document? Is it explain in a certain style? Or is it just like however is natural to the person?

我以前什么都写。然后写、重写，让它越来越精炼。那对我很有效。我现在写得没那么多了。现在我倾向于自己边走边在脑子里说。我实际做的是，我发现这对不同的人效果不同。很多人想和别人碰撞想法，那是他们思考的方式。你一遍又一遍地重复同样的事情，然后得到一些反馈。所以我以前写很多。当我想更好地理解一个想法时，我有时也会写。在我人生的某个时刻，我很想写点真正的东西，比如一本书之类的。但我现在越来越多做的是，我和同行或向我汇报的人进行一对一的交流。我戴上 AirPods，进行分布式散步交谈。两个人都在走路，但在不同的地方。你花一个小时讨论某件事。这结果证明非常富有成效。你获得了不孤单的力量，所以你得到了比你自己更多的脑力。我不认为这有很强的进化证据，但肯定有迹象表明你走路时思维更好，不管是因为你给大脑供氧，还是因为其他进化原因，我不确定。但我发现走路、说话、思考，即使不是面对面，只是通过 AirPods，也非常有效。是疫情迫使这样做的。我以为我们在疫情期间会变得不那么有创造力，战略规划会受到影响。我发现恰恰相反。我们比以往任何时候都做得更多，我开始思考为什么。我认为就是这些我们做的散步交谈。

I used to write everything. Then write and rewrite and make it more and more condensed. That worked for me. I don't write as much anymore. Now I tend to walk and talk in my head myself. What I actually do is I find this different for different people. A lot of people want to bounce something with someone else. That's how they think. You kind of repeat the same thing again and again and get some feedback on it. So I used to write a lot. I sometimes do when it's an idea I want to understand better. And at some point in my life, I would love to write something real like a book or something. But what I do increasingly now is I do my one-on-ones with peers or people who report to me. I just put on AirPods and do a distributed walk and talk. Both people are walking, but in different locations. You spend an hour discussing something. That has turned out to be very fruitful. You get the power of not being alone, so you get more brainpower than your own. I don't think there's strong evolutionary proof for this, but there are certainly indications that you're thinking better when you're walking, whether it's because you're oxygenating your brain or for some other evolutionary reason, I'm not sure. But I found that walking, talking, and thinking, even if not in person, just over AirPods, is super effective. It was the pandemic that forced this. I thought we would get less creative and that strategizing would suffer during the pandemic. I found the opposite. We had more of this than ever, and I started thinking about why. I think it's all these walking talks that we did.

你不可能有一个路由器，你想有一天写一本书。你觉得你的书会写什么？

You can't have a router that you want to write a book someday. What do you think your book would be about?

我不知道。从统计上看，它可能会是关于我做了很多的事情，所以它必须是与技术或产品有关的东西。但我会很想写一些虚构的东西。那会很有趣。

I have no idea. Statistically, it's probably going to be about something that I did a lot, so it has to be about something with technology or product or something. But I would love to write something fictional. That would be a lot of fun.

哦，天哪。一上架我就预订。

Oh boy. I'll pre-order as soon as that's up.

另一个我想提的概念，另一位 PM 建议的，他称之为“尿裤子”类比。这有印象吗？聊这个有意思吗？

Another concept I wanted to touch on that another PM suggested, which he called the pee in the pants analogy. Does that ring a bell and is that interesting to talk about?

我不确定这个人指的是哪个场合，但我知道我用过这个类比几次。我不知道这是不是瑞典的类比，因为我以为它更广为人知。这个想法是，你做某事，所以谚语是，在寒冷的天气里尿裤子。一开始感觉真的很温暖、很好，然后过一会儿你开始后悔。这基本上是关于短视的。所以现在我就说“这就像尿裤子”，因为人们知道我的意思。这是短期的事情。

I don't know exactly which occasion this person is referring to, but I know I've used that analogy a few times. I don't know if it's a Swedish analogy because I thought it was more widely known. The idea is that you do something, so the saying is that's like peeing in your pants in cold weather. It feels really warm and nice to begin with, and then after a while you start to regret it. It's about being short-term, basically. So now I just say that's like peeing in the pants instead because people know what I mean. It's a short-term thing.

这是传达那个想法的搞笑方式。肯定是瑞典特色。

That's a hilarious way of communicating that idea. Must be a Swedish thing.

是的，我想瑞典人出于某种原因这样做。显然其他人不这样做。也许因为一年中很多时候都很冷。

Yes, I think Swedish people do it for some reason. Apparently others don't. Maybe because it's cold a lot of times a year.

是的，很可能就是这样。这是寒冷气候中的谚语。在温暖的地方它没用。没人明白你的意思。

Yes, that's probably it. This is a saying in cold climates. In the warm it doesn't help. No one understands what you mean.

说到瑞典，你看《继承之战》吗？

Speaking of Sweden, do you watch Succession?

是的，我看。

Yes, I do.

好的，所以瑞典已经成为这部剧的重要部分，特别是公司试图……我想我不想剧透，但有一个非常重要的角色。

Okay, so Sweden has become a big part of the show, specifically the company trying to... I guess I don't want to spoil, but there's a character that's really important.

是的，没错，那是瑞典人。

Yes, exactly, that is Swedish.

所以我很好奇，你觉得他们描绘瑞典文化和瑞典商业交易的方式怎么样？

And so I'm curious just what do you think of the way they portray the Swedish culture and Swedish business dealings?

作为一个瑞典人，看到这个超级有趣。而且首先，像任何由高大、健壮、英俊的 Alexander Skarsgård 代表的人或国家，都应该相当高兴。所以那很好。然后我认为这是他们在挪威的那一集，不剧透太多。有些元素是真实的。有很多来自瑞典品牌 Fjällräven 的付费品牌定位，我认为那意思是北极狐。这实际上是瑞典非常受欢迎的户外品牌，所以那有点真实。桑拿之类的东西是真实的。所以它是真实的，但被夸大了。实际上，不太真实的是他的谈判风格。瑞典人往往严肃、谨慎，而这个人更像一个玩家。所以从谈判策略的角度来看，他不是典型的瑞典商人，我认为。

It's super fun to see this as a Swede. And first and foremost, like anyone or any country that gets represented by super tall, well-built, great-looking Alexander Skarsgård should probably be pretty happy. So that's good. Then I think this is the episode where they are in Norway without giving away too much. There are elements that are authentic. There's a lot of paid brand positioning from a Swedish brand name Fjällräven, which I think means arctic fox. Which is actually a very popular outdoor brand in Sweden, so that's kind of authentic. The sauna things and so forth are authentic. So it's real, but it's exaggerated. Actually, the thing that isn't very authentic is his negotiation style. Swedish people tend to be serious, cautious, and this guy is more of a player. So he's not the typical Swedish businessman from a negotiation tactic point of view, I think.

是的，这不会让我想到你描述的那种方式，在瑞典人们围坐成一圈，没有人坐在中心。

Yeah, it doesn't make me think of the way you described it where in Sweden people sit in a circle and no one's in the center.

不，没错。他非常处于中心。然后当人们去桑拿时，他们就像念咒语一样，桑拿，桑拿。就像那样。

No, exactly. He's very much in the center. And then when people go saunas, they're just like a chant, sauna, sauna. Like that.

没错。最后一集。这是一部很棒的剧。我喜欢它。这一季太疯狂了。我很好奇它会走向何方。

Exactly. The last episode. It is a great show. I love it. This season is insane. I'm so curious where it all goes.

也许在非常激动人心的快速问答之前，最后一个问题。Spotify 目前对我来说是最大的播客平台，我认为在全球范围内也是。我喜欢使用它。它运行得很好。我很好奇 Spotify 接下来会怎样，特别是 Spotify 播客。

Maybe just the last question before very exciting lightning round. Spotify is at this point the biggest podcasting platform for me specifically and I think globally. And I love using it. It works great. I'm curious just what's next for Spotify and specifically Spotify podcasting.

这有两个方面。一个是对 Spotify 创作者，一个是对 Spotify 听众。对于 Spotify 创作者，有两件事。一个是，这就是我们在 Stream On 上谈到的。我们主要谈的是音乐发现，但同样的问题，对播客来说甚至更难。所以我们仍然非常专注于帮助播客创作者找到更多受众。

There are two sides to it. It's for Spotify creators and for Spotify listeners. For Spotify creators, there are two things. One is, and this is what we talked about at Stream On. We talked about it mostly for music discovery, but the same problem, and even harder for podcast. So we're still focused very heavily on helping podcast creators find more audience.

嗯，就像我说的，在播客领域打破你的习惯和你的信息茧房是一个更大的问题，因为要找到一个新的播客需要很大的投入。所以我认为这是我们能够而且应该做好的事情。所以我们在这方面持续大量投入。嗯，而且正如我所说，随着我们推出更多功能，你会看到更多。创作者的另一大需求是变现。你知道，如今你可以通过多种方式变现，比如 DAI 和 Spotify SAI 等等，但我们正在努力扩展和改进，因为行业开始成熟，我认为这是创作者最大的需求之一，也是我们能为他们做的最重要的事情之一，帮助他们更好地变现。实际上，免费和付费都有。我们也有付费播客。所以这是创作者方面。在消费者方面，我不想分享太多。我们已经展示了我们在发现方面的大量投入。我想保留一些秘密，等到它们推出时再说，但我们确实在用户体验本身投入很多。我认为它离理想状态还差得很远。我可以分享的一点是，我们正在大力投资于跨设备、车载等场景的普及和播放，这些我们在音乐方面已经做得很好。但我认为收听体验可以变得更加无缝。嗯，我认为搜索可以更好，播客的数据，嗯，我不想说太多，但看看 AI 和生成式技术，有很多可以做的。

Uh this is like I said, it's even a bigger problem to break out of your habits and your bubbles in podcasting cuz it's such a big investment to find a new podcast. And so that is something I think we could and should do really well. So we keep investing a lot there. Um and as I said, you'll see more as we roll out more features now. The other big need for creators is monetization. And you know, you can monetize today in many ways with DAI and Spotify SAI and so forth, but we're working hard to expand that and make it better cuz the industry is starting to mature and I think this is one of the biggest needs and the biggest things we could do for creators to help them monetize better. Actually both free and paid. We also have paid podcasts. So that's on the creator side. On the consumer side, I don't want to share too much. We've shown that we're investing a lot in discovery. I want to keep some secrets for when they roll out, but we are investing a lot in the user experience itself. I think it's far from optimal yet what it could be. One thing that I can share that we're investing a lot in is just the ubiquity and playback across different devices and in cars and all these things that we've done well for music. But I think the listening experience can get a lot more seamless. Uh I think search can get better, the data about podcasts and well, I don't want to say too much, but looking at AI and generative technology, there's a lot that can be done.

好的，那我就知足了。至此，我们进入了非常激动人心的快问快答环节。我有六个问题要问你，Gustav。准备好了吗？

All right, well I'll take what I can get. With that, we've reached our very exciting lightning round. I've got six questions for you, Gustav. Are you ready?

我想我准备好了。来吧。

I think I am. Let's do it.

好的，我们拭目以待。你向别人推荐最多的两三本书是什么？

Okay, let's find out. What are two or three books that you've recommended most to other people?

好的，这就是为什么我试图把七本塞进两三本里。如果我们从产品方面开始，我认为有一本众所周知的，但我还是会推荐给别人读的书是 Hamilton Helmer 的《七力》，Netflix 用了很多，我们也用了很多。如果你刚开始创业，有一个战略框架是很好的。没有哪个战略框架是绝对正确的，但有一个总比没有好。另一本在思维模型和框架方面，我认为是 Charlie Munger 的《穷查理宝典》。所以，是的，它是关于投资的，但实际上是他使用的一系列思维模型，我认为关键的收获是，当你遇到一个问题时，你应该总是应用三个不同的模型，因为模型的作用是简化并降低维度。你知道，世界可能有无限的维度，它简化到也许三四个。这样做的风险是，你恰好去掉了一个非常重要的维度，比如，也许是大流行病之类的。但如果你使用三个具有不同维度、以不同方式简化的模型，统计上如果它们得出相同的结论，即使你应用的第二个模型也会大大增加你正确的几率。所以那是一本值得读的好书。然后，如果我们跳出产品领域，我对科学和数学非常感兴趣。所以快速说几本。《阿列夫之迷》，一本很棒的书。Sean Carroll 的《深藏之物》，关于量子力学的埃弗雷特诠释。Carlo Rovelli 的《海格兰》，关于量子力学的关系诠释。David Deutsch 的《无穷的开始》和《真实世界的脉络》。Donald Hoffman 的《反对现实》，关于进化与真理，进化并不优化于看到真理，而只是适应度。Gödel 的《证明》，我认为是一本关于他的不完备定理的精彩书籍，在任何公理系统中，都会有无法被证明的真命题，这想想很奇怪。然后，也许我最喜欢的之一是 Paul Davies 的《机器中的恶魔》，我认为这本书不太为人所知，它讲的是信息实际上就是熵，以及信息引擎的概念，你可以仅用信息来驱动某些东西，而排出物也是信息。

Okay, this is why I tried to squeeze in seven into two and three. So if we start with the on product, I think it's well known, but one that I would recommend for other people to read is Seven Powers by Hamilton Helmer, which Netflix has used a lot, we use a lot. It's just if you're starting out, great to have a strategy framework. No strategy framework is right, but having one is better than none. Another in sort of the space of mental models and frameworks, I think is The Complete Investor by Charlie Munger. So it's yes, it's about investment, but really it's a bunch of mental models that he uses and I think the key takeaway is you have a problem, you should always apply three different models to it because what models do is they simplify and reduce dimensionality. You know, the world has probably infinite dimensions and it reduces to maybe three or four. And the risk with that is you happen to get rid of a really important dimension like, you know, maybe pandemic diseases or something. But if you use three models that have different dimensions and were reduced in different ways, statistically and it comes to the same conclusion, even the second model you apply vastly increases your chances that you're right. So that was a good book to read. Then I think if we go outside of product, I'm very interested in just science and mathematics. So a few quick ones. The Mystery of the Aleph, an amazing book. Something Deeply Hidden by Sean Carroll on the Everettian interpretation of quantum mechanics. Helgoland by Carlo Rovelli on the relational interpretation of quantum mechanics. The Beginning of Infinity and Fabric of Reality by David Deutsch. The Case Against Reality by Donald Hoffman on sort of evolution versus truth and that evolution doesn't optimize for seeing the truth, just for fitness. Gödel's Proof, I think is an amazing book on his incompleteness theorem that in any axiomatic systems, there will be true statements that can never be proven, which is a weird thing to think about. And then maybe one of my favorites is The Demon in the Machine by Paul Davies that I think is lesser known on how information is really just entropy and this concept of information engines that you can power something by just information and the exhaust is also information.

这可不是一个简短的列表。

That was not a quick list.

不，但，不，我只是想说，你创下了最多书籍的记录，但这也表明你是如何变得如此有洞察力和智慧的，就是，你知道，读这样的书。所以我认为如果人们想要达到你现在的位置，我认为有一个教训。

No, but no, I was just going to say you've set the record for the most number of books, but it also shows how you've become so insightful and wise is just, you know, reading books like these. And so I think if people are looking to, you know, get to a place that you're at now, I think there's a lesson.

我保证，其他的问题我会简短得多。

I'll keep the others much shorter, I promise.

没关系，我们有时间。

And it's all good. We got time.

好的，下一个问题。你最近最喜欢的电影或电视剧是什么？

Okay, next question. What's your favorite recent movie or TV show?

我们谈到了《继承之战》，它是最近的最爱。所以我就提前选一个不是最近的，但绝对是挚爱的，那就是《奔腾年代》，我想它是在 FX 上播出的。如果你在科技行业工作过，这是一部很棒的剧。它从 80 年代的硅原开始，一直延续到今天。很棒的剧。《奔腾年代》。

So we talked about Succession and it is a recent favorite. So I'll just previously take something that isn't recent, but is an absolute favorite, which is Halt and Catch Fire, which I think is on FX. Amazing show if you ever worked in technology. It kind of starts out in the Silicon Prairie in the '80s and follows up to present day. Amazing show. Halt and Catch Fire.

是的，我看过一些。实际上我中途弃了，但这是一个很好的提醒，让我再去看看。得回去看。我打算回去看。

Yeah, I watched some of it. I actually fell off of it, but I'm going to It's a good reminder to go check it out. Got to go back. I'm going to go back.

你最近喜欢问的采访问题是什么？

What's a favorite recent interview question you like to ask?

我不问这个问题，但我最喜欢的问题是 Lex Fridman 的简短结尾问题，通常是像“那么，这一切的意义是什么？”我喜欢这个问题。这是一个很难回答的问题。

I don't ask it, but my favorite question is Lex Fridman's small ending question that is usually something like, "So, what's the meaning of it all?" I like that. It's a tough question to get.

我很想问你，但，不，别问。

I'm so tempted to ask you, but no, don't.

好的。我们继续吧。那将是另一期播客。那将是我们第二次尝试这个。

Okay. Let's move on. That'll be another pod. That'll be our second take at this.

你最近发现并喜欢的一些产品是什么？

What are some favorite products you've recently discovered that you love?

显而易见的是 ChatGPT GPT-4，就是随便玩玩，尝试为自己创建能帮你做不同事情的机器人等等。但我认为这对每个人来说可能都是如此。另一个真正喜欢的是你写过和谈过的产品，那就是多邻国，我认为从产品角度来看，它的执行力和他们所做的都非常令人印象深刻。它在我家也被疯狂使用。我们有一个家庭账户，每个人，你知道，每天都在使用并竞争。所以，我既对这个产品印象深刻，也经常使用这个产品。

The obvious one is ChatGPT GPT-4 and just playing around with that, trying to create bots for yourself that do different things for you and so forth. But I don't think that's probably true for everyone. The other really favorite is something you've written about and talked about, which is Duolingo, which I think is both very impressive from a product point of view, the execution and what they've done. It is also insanely used in my family. We have a family account, and everyone is, you know, using it and competing every day. So, I'm both impressed by the product and also use the product quite a lot.

你家里的人都在学什么语言？

What languages are folks learning within your family?

在我家，现在是西班牙语。

In my family, it's Spanish right now.

进展如何？

How's it going?

嗯，不错。

Um bien.

你得到了一颗金星。

You got a gold star.

我有几千点经验值。我还不够好。所以。我不知道那算不算好。

I have like a few thousand XP. I'm not that good yet. So. I don't know if that's good.

听起来很不错。

That sounds pretty good.

下一个问题，你在产品开发过程中做了哪些相对较小的改变，却对团队的执行能力产生了巨大影响？

Next question, what's something relatively minor you've changed in your product development process that's had a tremendous impact on your team's ability to execute?

我不确定我做过什么小事却产生了巨大影响。通常需要更大的改变才能带来大的影响。

I'm not sure I've done anything minor that had a tremendous impact. Usually it takes something bigger to get big impact.

我想也许有一件事我一直试图去做，回到清晰度等等，就是我提到的这个推动所谓苏格拉底式辩论的事情，其理念显然是让最好的想法胜出，而不是最资深的想法。并且推动这种让人们解释自己的观念，而不是说“我觉得那里有点东西，我有种感觉”之类的。显然，正如你所说，这产生了一些影响，因为人们确实那样评价我。所以这可能是最重要的事情。

I think maybe one thing that I've tried to do, back to clarity and so forth, is this thing I mentioned about pushing a lot for what I call Socratic debate, where the idea is obviously that the best idea wins, not the most senior idea. And trying to push for this notion of having people explain themselves, not saying like, 'I think there's something there, I have a feeling,' or something like that. And apparently, as you said, that has had some impact, because people apparently say that about me. So that's probably the biggest thing.

最后一个问题，Spotify 产品团队有什么有趣的仪式吗？是桑拿吗？

Final question, what is one fun ritual of the Spotify product team, and is it saunas?

Spotify 现在太大了，所以我们不……实际上这很本地化。Spotify 的不同部分有不同的产品仪式。很多年前，大概 12 年前，我偶然创造了一个仪式，当时我们讨论一个产品处于哪个阶段。我们需要一些定义。所以，我有点即兴地说：“嗯，你知道，有四个阶段：思考、构建、发布、调整。在思考阶段，你应该保持低成本，不要花太多钱。在构建阶段，你会开始花很多钱。所以你必须已经在思考阶段降低了风险，确保你是对的。然后是发布阶段，接着进入调整阶段。”这并没有经过深思熟虑。但有趣的是，我仍然听到人们这么说，有时甚至来自其他公司，比如“哦，我们处于思考阶段”或“我们处于调整阶段”。所以它有点深入人心了。我不知道这是否很好，但它确实留下来了。

So, Spotify is so big now that we don't... It's quite local, actually. Different parts of Spotify have different product rituals. I accidentally sort of created one ritual many years ago, maybe 12 years ago, when we talked about which phase a product is in. And we needed some definition. So, I think sort of off the cuff, I said like, 'Well, you know, it's four phases: think it, build it, ship it, tweak it. And in the think it phase, you should keep it cheap, not a lot of money spent. In the build it phase, you're going to start spending a lot of money. So then you must have reduced the risk in the think it phase that you're right. And then you have the ship it phase, and then you go over and tweak it.' And it was something that wasn't that thought through. But it's funny because I still hear it, sometimes even from other companies. Like, 'Oh, we're in the think it phase,' or 'We're in the tweak it phase.' So it kind of stuck. I don't know if it's very good, but it's stuck.

这很朗朗上口。我觉得任何能留在人们脑海里的东西都是成功的。Gustav，非常感谢你来做客。我们连续两期都是瑞典人。Gustav 以 F Alstromer 的身份上过播客，他也是一个很棒的人。同样很棒的人。我非常嫉妒那些能和你一起工作或为你工作的人。再次感谢你的到来。最后两个问题：如果人们想了解更多，也许联系你、问一些问题，他们可以在网上哪里找到你？

It is catchy. I think that anything getting stuck in people's head is a success. Gustav, thank you so much for being here. We are two for two for Swedish people. Gustav with an F Alstromer was on the podcast, who is also an amazing person. Also an amazing person. I feel very jealous of people that get to work with you and for you. Thank you again for being here. Two final questions: where can folks find you online if they want to learn more, maybe reach out, ask some questions?

这是 @gustavs。

This is @gustavs.

好的，再说一遍。

Okay. Say it again.

@gustavs。

@gustavs.

太棒了。最后一个问题，听众怎样才能对你有帮助？

Awesome. And then final question is just how can listeners be useful to you?

直接联系我就好。我确实会阅读反馈，我会试着忽略那些愤怒的评论，理解他们真正的想法和原因，你知道，他们为什么不满，或者哪里出了问题。

Just reach out. I do read feedback, and I try to remove the angry comments and understand what they're actually thinking and why, you know, why they're upset or what's not working.

那么联系你，你更推荐在推特上愤怒地@你，还是发邮件到你分享的那个邮箱地址？

And then the reaching out, would you recommend an angry tweet at you or more of an email to that email address you shared?

嗯，@gustavs 是推特账号。所以直接发推@我就行。你也可以友好一点，没关系。

Well, the @gustavs is the Twitter handle. So just tweet at me. And you can be nice as well. It's okay.

太棒了。Gustav，非常感谢你来做客。

Amazing. Gustav, thank you so much for being here.

谢谢你邀请我，Lenny。这是我的荣幸。大家再见。

Thank you for having me, Lenny. It's been a pleasure. Bye, everyone.

非常感谢你的收听。如果你觉得这期有价值，你可以在 Apple Podcasts、Spotify 或你最喜欢的播客应用上订阅本节目。另外，请考虑给我们评分或留下评论，这真的能帮助其他听众找到这个播客。你可以在 lennyspodcast.com 找到所有过往节目或了解更多关于本节目的信息。下期再见。

Thank you so much for listening. If you found this valuable, you can subscribe to the show on Apple Podcasts, Spotify, or your favorite podcast app. Also, please consider giving us a rating or leaving a review, as that really helps other listeners find the podcast. You can find all past episodes or learn more about the show at lennyspodcast.com. See you in the next episode.