LLenny's Podcast
← All episodes
Gustav Söderström (Co-President, CPO, and CTO at Spotify)21 May 2023

Lessons from scaling Spotify: The science of product, taking risky bets, and how AI is already impacting the future of music

8Frameworks
15Insights

Frameworks in this episode

Leadership4 steps

Autonomy at the VP Level

Place decision autonomy at the VP layer — not the leaves, not the top — to maximize thinking without producing chaos.

Read →
Strategy4 steps

Centralized vs. Decentralized Org Choice

Pick your org model from your product strategy: speed (Amazon) or unified experience (Apple), and accept its cost.

Read →
Communication4 steps

Explain-It-to-Prove-You-Understand-It

Force yourself and others to explain decisions; if you can't articulate it, you probably don't understand it.

Read →
Innovation4 steps

Fault-Tolerant User Interfaces

Design your UI to match the real hit-rate of your machine learning, not a fantasy of 100% accuracy.

Read →
Marketing4 steps

Recall vs. Discovery Interface Split

Match interface density to user intent: dense grids for recall, low-density sound-rich feeds for discovery.

Read →
Innovation4 steps

Separating Change-Aversion from Real Problems in a Redesign

After a redesign, split 'upset because it changed' from 'upset because it's worse' before you react.

Read →
Productivity4 steps

Think It, Build It, Ship It, Tweak It

Four product phases where spend rises stage by stage — so you must retire risk before the money starts.

Read →
Mindset3 steps

Three-Model Triangulation

Apply three different mental models to a problem; agreement across them sharply raises your odds of being right.

Read →

Insights & moments

The myth-busts, hot takes, explainers, and tools worth keeping.

Hot Take· 3

Hot Take22:00

AI Music Is Just a More Powerful Instrument

Gustav compares AI music generation to the digital audio workstation that let Avicii — once dismissed as 'not a real artist' — make music now considered real. He predicts diffusion models will act like a powerful new instrument enabling new creators and new genres, while paradoxically making true originality harder as generic music gets easier to produce.

  • Avicii was once dismissed because he couldn't play an instrument or sing; now his talent is undisputed
  • Diffusion models are best understood as a much more powerful instrument, not fake music
  • The AI-vs-real-music distinction is false; the real question is what percentage is AI and who gets to use it
  • As generic music gets cheaper to make, being truly unique becomes harder than ever
  • The rights/business-model problem, not the tech, is the real unsolved issue — mirroring the piracy era

“the way to think about these these diffusion models if and when they get good enough at at generating music is probably the same, like…”

Gustav Söderström · 22:30

“as it now gets easier to create more generic music, it will actually be more difficult than ever to be truly unique.”

Gustav Söderström · 24:30
#ai#music#creators#rights
Hot Take1:00:00

Believe 100% Until the Data Says No — Then Believe Something Else 100%

Gustav's antidote to getting 'precious' about products you've invested in: be brutal and unemotional, believe fully until the data disproves it, then switch completely. He notes the strange human bias that we distrust people who change their minds based on evidence — we celebrate confident prophets with little data and ridicule politicians who update their views.

  • The biggest risk after heavy investment is getting precious; you have to be brutal and unemotional
  • Believe in things 100% until the data says no, then believe something else 100%
  • It's the essence of 'strong opinions loosely held' — easy to say, hard to do
  • Humans irrationally love confident prophets and distrust people who update on evidence

“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…”

Gustav Söderström · 1:00:00

“We actually tend to love prophets and people who hold very strong opinions with very little data.”

Gustav Söderström · 1:01:30
#decision-making#leadership#data#psychology
Hot Take1:05:00

0% Art, 0% Magic, 100% Science — Force People to Explain Themselves

Gustav pushes his leaders to always explain their reasoning rather than defer to seniority, arguing that employees are owed an explanation even when they don't get agreement. He provocatively insists product decisions are '0% art, 0% magic, 100% science' — because 'art' and 'magic' are just labels for what we can't yet explain, and if you can't explain something, you probably don't understand it.

  • Employees don't have to agree, but they're entitled to understand why a decision was made
  • Deferring to seniority instead of explaining is unacceptable — you owe an explanation
  • If you can't explain it, you probably don't understand it — and often there's nothing there
  • 'Art' and 'magic' are just words for things not yet explained (genetics, quantum physics, and now intelligence via LLMs)

“if you can't explain it yourself, you probably don't really even understand it yourself.”

Gustav Söderström · 1:06:30

“I say it's 0% art, 0% magic, and 100% science.”

Gustav Söderström · 1:07:00
#leadership#communication#decision-making

Explainer· 4

Explainer13:30

The Internet's Three Eras: Curation, Recommendation, Generation

Gustav frames the internet's evolution in three stages: user curation (early Facebook/Spotify), then algorithmic recommendation, and now generation via AI. He argues generation is as big a shift as recommendation was, and will eventually force companies to rethink their products entirely.

  • The internet started with user curation, then shifted to algorithmic recommendation, and is now entering the generation era
  • The generation era will be as disruptive as the recommendation era, requiring product and sometimes business-model rethinks
  • Even though it's all machine learning, teams should treat generation as something completely new, not more of the same

“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…”

Gustav Söderström · 14:00

“I tend to talk to my teams about even though it's all machine learning, I asked them to think of this as something completely different.”

Gustav Söderström · 14:00
#ai#product-strategy#machine-learning
Explainer18:00

Fault-Tolerant UIs: Design for Your Model's Actual Hit Rate

Borrowing a principle from Chris Dixon, Gustav argues you must design your interface to match the real performance of your machine learning. A single play button only works if predictions are 100% accurate; if your hit rate is one-in-five, you need a UI that shows five options and an easy escape hatch. He cites early Midjourney's four low-res images as a model example.

  • A single play button demands 100% prediction accuracy, which never happens
  • If your hit rate is one-in-five, show at least five items on screen so one is likely relevant
  • Always give the user an easy escape hatch when the prediction is wrong
  • Early Midjourney generated four fast low-res images at once, matching its ~one-in-four quality

“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.”

Gustav Söderström · 18:30

“What they did was they generated four simultaneous low-res images very quickly.”

Gustav Söderström · 20:00
#ux#machine-learning#design#ai
Explainer31:00

Why Spotify Moved Away From Squads — and Where to Put Autonomy

Gustav explains Spotify has quietly moved away from the famous squad model: scaling in seven-person increments created huge overhead, and pushing full autonomy to the leaves of a very junior org produced '100 squads with 100 strategies running in 100 directions.' The lesson: put decision autonomy at the VP level, avoiding both the leaf-level chaos and the single-person-at-the-top bottleneck.

  • Seven-person full-stack squads created too much overhead at scale; teams are now ~2-3x bigger with fewer overhead roles
  • Hiring smart people means 'renting brainpower' — give them room to think or you're wasting the investment
  • Autonomy at the leaves plus a junior org produces heat: many teams pulling in conflicting directions
  • Autonomy only at the top bottlenecks throughput; the VP level is the sweet spot — many senior people with pattern recognition

“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.”

Gustav Söderström · 31:30

“you're just going to produce heat. You're going to have 100 squads with 100 strategies running in 100 directions.”

Gustav Söderström · 32:00
#org-design#management#autonomy#scaling
Explainer36:00

The Amazon vs Apple Org Spectrum — and Why Spotify Chose Centralized

Gustav lays out two extremes of org design: Amazon's decentralized two-pizza teams optimize for speed but ship the org chart to users (and forced hard APIs, which birthed AWS), while Apple's centralized functional org is slower but delivers a coherent, single-developer-feel experience. Because Spotify puts many content types with different business models into one app, it chose the more centralized, Apple-like model.

  • Amazon: decentralized, competing teams, fast to user, but tends to ship complexity and the org chart to users
  • Bezos forced hard APIs because competing teams have no incentive to cooperate — which made AWS possible
  • Apple: centralized functional org, slower, but delivers a coherent 'built by one developer for one user' experience
  • Spotify centralized because its single-app strategy with mixed business models demands one recommendation org and one UI
  • Both extremes work; the middle is where the 'smiling curve' says you don't want to be

“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.”

Gustav Söderström · 37:00

“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.”

Gustav Söderström · 38:30
#org-design#strategy#amazon#apple

Story· 3

Story08:00

How an Internal Podcast Became a Spotify Recruiting Tool

Gustav started an internal, employee-only podcast interviewing Daniel Ek's direct reports to make leadership more approachable and share product strategy in an entertaining way. It worked so well internally that he took it public, and now new hires regularly cite the podcast as a reason they joined Spotify.

  • Listening to leaders for an hour makes them feel knowable in a way meetings and town halls don't
  • The internal podcast made senior leaders less 'otherworldly' and shared product-strategy thinking better than docs people never read
  • Making it public turned it into a recruitment tool: new hires cite it at Stockholm intro days as a key reason they joined
  • Making the podcast also built empathy for the podcast-creator problems Spotify is trying to solve

“what if people that don't even work at Spotify yet could feel as if they knew people at Spotify?”

Gustav Söderström · 10:30
#content#recruiting#culture#podcasting#leadership
Story15:00

The AI DJ: Spotify's First Product That Couldn't Exist Without Generative AI

Gustav describes the AI DJ as Spotify's only true generative product so far — one that couldn't exist without generative AI, using a digitized real person's voice to talk to you about your music at the scale of half a billion users. It finally solved the long-standing 'I don't know what I want to listen to' radio use case that Spotify had failed at for years.

  • The AI DJ generates both the voice and the content of what it says, enabling personalized voice at half-a-billion-user scale
  • It solves the 'Siri intent' / radio use case: cycling through options when you don't know what you want
  • It was a bottoms-up product built by the personalization team, requiring significant support and partnerships
  • Its guiding principle: do as little as possible and get out of the way, because people came for the music

“we've only really done one sort of real generative product in in the hard definition, which is a product that couldn't have existed without generative…”

Gustav Söderström · 15:00

“the principle there was literally to do as little as possible and get out of the way.”

Gustav Söderström · 20:30
#ai#product#spotify#generative-ai
Story44:00

Why Spotify's Discovery-Heavy Homepage Redesign Backfired

Users kept asking Spotify to help them break out of their 'taste bubbles,' so Spotify built feed-style discovery tools (which worked) and then loaded them onto the home screen — flipping it from ~90% recall/10% discovery to 10% recall/90% discovery. The quantitative data and Twitter backlash showed people weren't rejecting discovery; they were furious they could no longer quickly recall the playlists and sessions they already knew they wanted.

  • Breaking a taste bubble is a different problem: recommending a genuinely new genre means near-zero signal and a very low hit rate
  • Background recommendations need ~9/10 hit rate; discovery works at ~1/10 if the swipe cost is low and expectations are set
  • The sub-feeds worked, but putting them on home flipped it from 90% recall to 90% discovery
  • Angry tweets and quant data (traffic shifting to search/library) showed users couldn't find what they already wanted
  • Recall wants a dense, many-items UI; discovery wants lots of pixels and sound — they're opposite interfaces

“We need something where one out of 10 is is a success. If you find one gem out of 10 tries, you're very happy.”

Gustav Söderström · 47:00

“we kind of switched it from 9010 to 1090 so 10% recall 90% discovery and while people want to discover they don't they probably don't…”

Gustav Söderström · 50:00
#product#discovery#ux#ab-testing

Tool· 2

Tool1:08:30

The Distributed Walk-and-Talk

Gustav's go-to thinking method: put on AirPods and do one-on-ones as a 'distributed walk and talk' — both people walking in different locations while discussing a problem for an hour. Forced by the pandemic, it turned out more creative and fruitful than expected, combining more-than-your-own brainpower with the thinking boost of walking.

  • Do one-on-ones over AirPods while both people walk in separate locations
  • Combines extra brainpower (you're not alone) with the cognitive benefit of walking
  • The pandemic forced it; strategizing got better, not worse
  • Complements his older method of writing and rewriting ideas to condense them

“I just put on AirPods and do like a distributed walk and talk. Both people are walking, but in different locations.”

Gustav Söderström · 1:09:00
#productivity#thinking#remote-work#tools
Tool1:16:00

Gustav's Book Recommendations: Seven Powers, Munger, and More

In the lightning round, Gustav recommends Seven Powers by Hamilton Helmer as a starter strategy framework (used by Netflix and Spotify) and The Complete Investor on Charlie Munger, whose key lesson is to apply three different mental models to any problem. He then rattles off a long list of science and math books including Gödel's Proof, Something Deeply Hidden, and The Demon in the Machine.

  • Seven Powers by Hamilton Helmer — a strategy framework; no framework is right, but having one beats none
  • The Complete Investor (on Charlie Munger) — always apply three different mental models to a problem
  • Using multiple models that reduce dimensionality differently but agree vastly increases your odds of being right
  • Science/math picks: The Mystery of the Aleph, Something Deeply Hidden, Helgoland, The Case Against Reality, Gödel's Proof, The Demon in the Machine

“one that I would recommend for other people to read is Seven Powers by Hamilton Helmer, which um Netflix has used a lot, we use…”

Gustav Söderström · 1:16:00

“you have a problem, you should always apply three different models to it”

Gustav Söderström · 1:16:30
#books#mental-models#strategy#recommendations

Takeaway· 3

Takeaway26:00

All Truly Great Products Pull Some Kind of Magic Trick

Gustav argues that great products create a moment that feels like magic even though it's just data and statistics — like the AI DJ seeming to record a person saying things about your specific music. Crossing from 'cool and impressive' to 'I don't understand how this is done' correlates with products going viral, and it usually comes from fine-tuning, scoping down, and getting performance to the right level.

  • The magic trick correlates with products going viral and taking off
  • The AI DJ's trick: it felt like a real person recorded personalized lines, but it was generated
  • GPT and DALL-E/Midjourney felt like magic on first use even though there's no magic, just data and statistics
  • Crossing the line to 'feels like magic' comes from fine-tuning, scoping down, and removing things

“obviously there is no magic. It's just data and statistics.”

Gustav Söderström · 27:00

“iterating a product to the point where it feels like magic the first time is is very helpful.”

Gustav Söderström · 27:30
#product#ai#design#virality
Takeaway54:00

Redesigns Are Harder Than New Features — the Rearranged-Desk Problem

Gustav distinguishes two fundamentally different types of product development: a new feature is voluntary, so if you dislike it nothing got worse, but a redesign is not voluntary and imposes a cost even on people who dislike it. He uses the analogy of someone rearranging your desk after 12 years — even with data proving the new layout is better, you'll be upset because your habits are broken.

  • New feature = voluntary; a redesign is imposed on everyone, so it carries a cost even for those who dislike it
  • Redesign feedback splits two ways: you were right but broke habits, or you were simply wrong — and they look the same
  • The rearranged-desk analogy: better-by-data still upsets people who were effective in the old setup
  • To tell the cases apart, compare new user cohorts (no old habits) against existing user cohorts

“there is a there are two fundamentally different types of product development one is designing a new feature it is hard to make it but…”

Gustav Söderström · 54:00

“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…”

Gustav Söderström · 55:00
#product#redesign#change-management
Takeaway1:02:30

The 10% Planning Rule

A rule of thumb Gustav credits to Shishir Mehrotra: you shouldn't spend more than 10% of your time planning versus executing. In a 10-week quarter that's one week of planning; Spotify's 6-month increments mean about two weeks. If you're planning much more, you're either over-planning or your execution period is too short.

  • Spend no more than ~10% of your time planning versus building
  • A 10-week cycle = 1 week planning; Spotify's 6-month cycle = ~2 weeks
  • Over-planning signals either too much planning or too short an execution period
  • Attributed to Shishir Mehrotra, who used it since his YouTube days

“you shouldn't be spending more than 10% of your time planning versus uh executing or or building.”

Gustav Söderström · 1:02:30
#planning#execution#process