Do interest-graph-based recommendation algorithms outperform social-graph-based ones for content discovery?

Kevin Systrom · Feb 22, 2023

Social networks need to become less social; the assumption that your friends are into the same things you're into is the greatest mistake social networks have made in the last twenty years.

The foundational thesis that your friends share your interests is simply not true.

Scope: about the last twenty years of social networking

0:00 20VC: Instagram Founders Kevin Systrom and Mike Krieger on Why Social Networks Should Be Less Social & The Next Wave of Social | Why San Francisco Will Return with a Vengeance and The Future For Remote Work | Let's Get Personal: Relationships to Money, Be

Kevin Systrom · Feb 22, 2023

Social networks need to become less social: learned affinity from user behavior should replace the social graph as the primary content filter, though the graph will not be abandoned entirely

Friend-graph filtering doesn't scale — over time you get your angry uncle posting about Trump — while behavioral association can serve users their actual interests, and algorithms can be built to work for the user rather than for the company

Scope: not a complete annihilation of the social network; acknowledges risks: filter bubbles, over-concentrated distribution, untrue content

28:54 20VC: Instagram Founders Kevin Systrom and Mike Krieger on Why Social Networks Should Be Less Social & The Next Wave of Social | Why San Francisco Will Return with a Vengeance and The Future For Remote Work | Let's Get Personal: Relationships to Money, Be

Victor Riparbelli · Jan 15, 2025

TikTok is an excellent product because it builds a recommendation graph from interests rather than social connections, and tests new content on small audiences to discover what's actually good

His own feed is ~90% informative educational content, and TikTok's usage numbers show the interest-graph plus small-cohort testing model clearly works

Scope: acknowledges this may be an unpopular opinion; based partly on his personal feed

42:12 20VC: Why Scaling Laws Will Not Continue | OpenAI vs Anthropic vs X.ai: Who Wins and Why | How Far Will Model Providers Go Into the Application Layer | The End State for Models: Many Specialised or Few Generalised with Victor Riparbelli @ Synthesia

Casey Winters · Sep 22, 2021

Pinterest's early growth stalled because its core loops broke — Facebook's Open Graph changes killed new-user acquisition, and the friend-based content graph stopped driving engagement as the audience expanded beyond its original core

New users who joined via a friend only saw that friend's pins, which often didn't match their tastes, so they concluded the service wasn't for them

29:50 Social graph based growth breaks down as audience scale diverges from the original core and external platform rules change

Mike Krieger · Feb 22, 2023

The right architecture is to invert discovery and conversation: algorithms should drive discovery while social connections drive the conversation about what's discovered

The desire to talk about content with people you care about doesn't go away — he instantly shares housing policy articles to a DM group and Warriors stories to a friend, and beta users share out to Twitter and WhatsApp to start conversations — so social strengthens relationships without needing to gate discovery

31:45 Hybrid algorithmic discovery paired with social conversation is the right architecture