Lenny's Podcast · Feb 27

An inside look at X’s Community Notes | Keith Coleman & Jay Baxter

About 95,000 notes were seen roughly 30 billion times in 2024, more than double the year before, and only about 8% of proposed notes are shown. Showing a note cuts likes and reposts 30 to 40% in A-B tests, and outside studies found total reposts fall 50 to 60%. After the acquisition, Musk's 'fork in the road' email made staff click a form to stay, and a tiny team scaled Spaces in two or three weeks, a job that would have taken a year at the old Twitter. Shrinking the company meant faster launches and more experiments.

Tips

Show a note only when people who usually disagree rate it helpful▶ 7:48Notes were seen about 30 billion times in 2024 from about 95,000 notes, more than double the prior year.
Set the quality bar high and accept publishing only about 8% of notes▶ 20:20About 8% of proposed notes show, typically between 7% and 11% over time.
Let anonymous or pseudonymous users contribute instead of requiring real names▶ 1:27:11Pilot feedback showed more good notes could be written than were getting written under real names.
Make the product prove itself at each stage before expanding▶ 1:29:58The pilot with only 1000 contributors already showed neutral, informative notes on controversial topics.
Run a small project with one owner, one senior sponsor and full focus▶ 42:03The original team was about five people: backend, client, ML, design and research.
Replace OKRs and Jira with a single running Google Doc▶ 45:05They once tried Asana but spent more meeting time grooming irrelevant backlog than discussing priorities.
Let people opt in to the team instead of assigning them▶ 56:38A tiny team scaled Spaces in two or three weeks, work that would have taken a year at the old company.
Hold an internal bake-off to pick your core algorithm▶ 47:47A PageRank-type graph algorithm can amplify bias when more users sit on one side.
Remove the admin override and let the system decide▶ 1:11:48
Publish your code and data so outsiders can replicate and audit it▶ 1:14:15The full algorithm can run on one machine with about 500 gigs of RAM in about a day.
Put the context directly on the content with zero clicks▶ 15:32A single note can match thousands of posts, and the median time to a note in the first days of the Israel Hamas conflict was five hours versus two to four days for typical fact checking.
Expect a note to cut engagement and reposts by large amounts▶ 27:2030 to 40% engagement rate drops in A-B tests, a 50 to 60% drop in total reposts, and authors 80% more likely to delete.
Delete features whose maintenance cost outweighs their gains▶ 1:04:01
Ship speed and coverage features before the surge that needs them▶ 1:20:00About 500 notes in the first three days, with a median of five hours from post to note.
Start from a user-facing name and mockup that people already understand▶ 35:30
Leave room for users to invent uses you did not plan▶ 1:42:52

Twitter

What was said about Twitter

Product strategy and PMF6

Teams and execution6

Experimentation and analytics3

Brand and positioning3

Building the company1

Capital and deals1

Retention and engagement1

How Elon Works

An inside look at X’s Community Notes | Keith Coleman & Jay Baxter

Twitter’s ex-Head of Product on Elon, consumer products, culture, more | Kayvon Beykpour

5 Ways to Increase Mobile Game Revenue

More in Social & Dating

Facebook

Instagram

TikTok

LinkedIn

Tinder

Snap

Discord

Snapchat