Three Eras of the Internet: Curation to Recommendation to Generation
Gustav Söderström
Co-President, Chief Product & CTO, Spotify
14 YEARS AT SPOTIFY
Framework
Three Eras of the Internet
"Each era is as big of a shift as the previous. The recommendation era required us to rethink the entire user experience and business model. Generation will too."
Era 1: Digitize content + users curate
Era 2: Algorithms curate for users
Era 3: AI generates new content at scale
Each shift demands rethinking UX, not just adding features
Product Principle
Fault-Tolerant UI Design
9/10
background hit rate needed
1/10
feed hit rate acceptable
Design must match ML performance: If your algorithm hits 1 in 5, design for 5 items visible simultaneously
Background vs. Discovery: Recommendations in context need 90%+ accuracy; discovery feeds can tolerate 10% hit rate
Escape hatches matter: Users must easily reject bad suggestions and return to known content
The Midjourney principle: Discord chat interface made low accuracy expected—users expect iteration
Critical insightDon't design a simple UI for complex ML. Match your interface complexity to your algorithm's accuracy tier.
The Taste Bubble Problem
Why Algorithmic Discovery Fails at Surprise
The paradox: Recommending something "new" means you have zero signal that the user will like it
Background insertion doesn't work: Suggesting reggaeton in an EDM playlist breaks trust ("Is Spotify broken?")
You need a different paradigm: Feed-like interfaces where low hit rate is expected
The radio nostalgia: People remember radio's ability to flip through content quickly, even though radio was poor on every other metric
Why feeds work for discovery
Users expect high variety and low hit rate. Cost of rejection is one swipe, not two minutes of bad content. Multiple candidates viewed quickly.
The zero-intent use case
When you don't know what you want, Spotify struggled. But radio solved this with a knob. The AI DJ recreates that experience with personalization.
Product Strategy
Organizational Design Trade-offs
Decentralized (Amazon model): Speed wins. Each team ships fast. Tradeoff: UI complexity, multiple search boxes, org chart shipped to users
Spotify's choice: Centralized recommendations, single vertical org because we're unified experience + multiple content types + need to weigh music vs. podcasts vs. audiobooks
The smiling curveExtreme centralized or decentralized both work if they match your strategy. The middle ground loses.
Contrarian
Myths About Product Scale at Spotify
✗Simplicity = fewer featuresINSTEAD →✓ Simplicity means ruthless prioritization. Spotify added content types but kept UX coherent.
✗Feedback from users = truthINSTEAD →✓ Users ask for what they know. They don't ask for the AI DJ until it exists, but they always needed it.
✗ML recommendations = generative AIINSTEAD →✓ Totally different paradigms. Recommendations need accuracy. Generation needs iteration tolerance.
✗Podcast creators want more featuresINSTEAD →✓ They want simpler rights management, easier music licensing, less friction than more tools.