@CommunityNotes founding ml lead @X / @SpaceXSI. built bayesDB @MIT
Bellevue, WA
Joined February 2011
- Tweets1.8K
- Following2.1K
- Followers8.3K
- Likes14K
Pinned Tweet
My interview from TED with @kcoleman and @audreyt just went live!
Tons of gold in here:
-why we stuck with crazy principles like not having an override button
-how AI and humans collaborate to write better/more notes than either on their own
-how much impact community notes are making
-why our algorithm is smarter than regular upvoting/downvoting
-new ideas for finding common ground on social media
ted.com/talks/keith_coleman_…
We just dropped a new paper on our Community Writer -- scaling AI-written Community Notes, with human raters in charge
The Community Writer submits its notes over the same the public AI note writing API that all other AI note writers use, and altogether, AI note writing has been a huge driver of speeding up and scaling up Community Notes so far (in a way that's complementary with human note writers)
The code is open sourced too so it's easier than ever now to build a great AI note writer
We're open-sourcing the Community Writer, an AI note writer designed to reflect the will of the people, which, together with the open note writing API, has helped drive an 86% increase in the number of notes on X that people from different perspectives find helpful.
The Community Writer works by incorporating input from the public throughout the note writing process.
We've written a technical report detailing how it works and the impact of the Note Writing API overall. And we’re making the code open-source under Apache V2 so that it can be used to advance research and scale informative context across the internet.
Technical report: arxiv.org/abs/2609.40067
Code: github.com/xai-org/community…
Here are the Community Writer accounts:
xitter.cf/i/communitynotes/u/cel…
xitter.cf/i/communitynotes/u/che…
xitter.cf/i/communitynotes/u/gre…
xitter.cf/i/communitynotes/u/zap…
xitter.cf/i/communitynotes/u/nei…
xitter.cf/i/communitynotes/u/dil…
some insane stats on these
*if you are numerically inclined: when looking at the stats, do keep in mind that the Community Writers do try to delete notes if they are getting early bad rating signal. So while the Helpful rate is representative of note quality, the Not Helpful rate is not a good representation of overall note quality
Ok, one more caveat about the above metric, Helpful rate.
As AI note writers scale to write on more notes, they need to process increasingly large feeds of candidate posts.
However, as the candidate post feeds grow larger, the % of posts that are actually in need of a note grows smaller (since the required amount of note request signal is largest in the Small eligible post feed, and smaller in the larger feeds).
Therefore, you can think of feed size as something like a 'difficulty level': you'll see the same AI note writer get a significantly lower CRH rate if it's processing a larger feed of posts, all else equal.
Jay Baxter retweeted
Community Notes really a huge feature upgrade and other platforms should follow suit.
Jay Baxter retweeted
We're open-sourcing the Community Writer, an AI note writer designed to reflect the will of the people, which, together with the open note writing API, has helped drive an 86% increase in the number of notes on X that people from different perspectives find helpful.
The Community Writer works by incorporating input from the public throughout the note writing process.
We've written a technical report detailing how it works and the impact of the Note Writing API overall. And we’re making the code open-source under Apache V2 so that it can be used to advance research and scale informative context across the internet.
Technical report: arxiv.org/abs/2609.40067
Code: github.com/xai-org/community…
Jay Baxter retweeted
Last week, @X sued several people who abused Creator Revenue Sharing by operating a coordinated network of accounts, posting inauthentic content to manipulate engagement, and using multiple bank accounts to hide their scheme.
We do not tolerate fraudulent behavior on X -- and will act forcefully to protect our platform and the earnings of genuine creators.
You can read our lawsuit here: transparency.x.com/assets/le…
Now you can check whether you got shadow banned
Insanely cool that it’s possible to see if you were affected by legal demands to filter your posts
We’ve expanded Under The Hood after seeing people’s appreciation of more transparency into government-required content filtering.
In addition to showing labels applied to your account and posts in the prior month, Under The Hood now shows if your account or posts had their visibility limited because of required compliance with local law(s).
For example, you can see if any of your posts were withheld in a country following a legal demand – and which country.
We’re making this Under The Hood update available to all accounts that meet the basic eligibility criteria (10+ posts in prior month, account at least one year old).
You can get your report at xitter.cf/i/under_the_hood
ALT Screenshot of X's Under The Hood page.
This experiment worked really well, so it's now enabled globally -- notes are now showing a median of 2.8 hours faster!
People want notes fast. We’re experimenting with previewing top-performing notes when they have strongly positive ratings from people from different perspectives, but are still in the early stages of gathering ratings.
Now available to a limited group in the US. We’ll shape it with your feedback.
Details in the Guide: communitynotes.x.com/guide/e…
ALT Screenshot of a new note preview on a post. The post says "BREAKING: Footage from downtown" and the preview says "Readers proposed context. Still gathering ratings" followed by a note that says "This video is from 2017, not today. http://youtube.com/..." and a button to rate the note.
Jay Baxter retweeted
Automated community notes (backed by an aggregation algorithm, not vanilla AI fact checking) -> I get a light touch retrospective notification that I may have been misled (not necessarily intentionally) is actually great, rare kudos to @X
Jay Baxter retweeted
It's finally here! I gave a TED talk in April in Vancouver, and it was released online today.
You can watch it here:
ted.com/talks/saloni_dattani…
I've also adapted it into an article, which was published last week: worksinprogress.co/issue/fut…
Jay Baxter retweeted
Banger community note
If these data centers are so essential to our national security and economy and have such amazing benefits they should prove it by building them in the richest zip codes in America first.
Readers added context they thought people might want to know
Loudoun County, Virginia is the richest county in the United States and also has the most data centers.
usnews.com/news/healthies…
visualcapitalist.com/mapped-u-s-sta…
Big thank you to @audreyt @idhantgulati @IgorBrigadir @jonathanstray @liyushan27 @vboykis for the reviews and feedback that materially improved X algo open sourcing!
Inspired by them, there's now better documentation including a fun history of the mutuals algo change, timestamps on parameter value exports, clearer 'under the hood' reports, and lots more.
Quality feedback really does make a positive impact, so please don't be shy to speak up here on X or on github.com/xai-org/x-algorit… if you have some!
Big thank you @audreyt @idhantgulati @IgorBrigadir @jonathanstray @liyushan27 @vboykis who voluntarily reviewed a draft version of the new X algo open-source repository in advance of yesterday's launch.
Their feedback helped improve its value to the public, including:
• Proving that the Phoenix scorer could be built and run outside of the company
• Identifying missing packages, files, etc. that were referenced, so we could get them into the public release
• Highlighting questions people might have, so we could add comments and clarifications. This includes questions like “are the params the real prod ones?” which led us to publish explanations of how we keep them up to date in the repository, and comment timestamps so the public can see when they’re updated
• Providing feedback about what info would be most helpful in the new Under the Hood label transparency tool (which shaped what we included)
• Recommendations on how to improve the README & documentation, to make the code more understandable
• Ideas like publishing the history of the recent “mutuals” ranking change and how it would have shown up if the code had been open-source at that time
• and more…
It’s been great to have input from people outside the company, and we hope for much more now that it’s all public!
Issues and PRs are open in the repo: github.com/xai-org/x-algorit…
Jay Baxter retweeted
𝕏 just took algorithm transparency to another level.
Brazilian election rules require recommendation systems to exclude campaign profiles officially submitted to the Electoral Court from recommendations. Legally permitted paid promotion is treated separately.
Instead of quietly applying the rule, X published the source code for its Brazil2026ElectionFilter, currently 1,573 lines long.
The code shows:
• 665 affected profiles, with every username listed
• Their posts are kept out of For You recommendations when you do not follow them
• Reposts and quotes that surface those profiles are also filtered
• Replies are filtered when the conversation history would surface a listed profile
• If you follow one, this specific filter does not remove its posts
• X cites Brazil’s official TSE 2026 candidate dataset
• The file includes public tests covering the logic and confirming the 665-entry count
This filter does not delete or suspend these profiles. X is publicly showing who it affects, why it exists and exactly how its logic works.
You rarely see a major social platform publish a country-specific recommendation filter, every affected username, its exact logic and the tests behind it.
X did.
That is real transparency.
Jay Baxter retweeted
Quite a first update – exemplifies why open-source is great. You don't have to wonder what's going on in Brazil, you can just see the code:
Today’s For You release is in the open-source repo.
Notable updates:
• How ranking weights actually work
• A new filter required for Brazil’s 2026 election
All the details: github.com/xai-org/x-algorit…
another one 👇👇
we weren't joking when we said you should expect regular updates!
Today’s For You release is in the open-source repo.
Notable updates:
• How ranking weights actually work
• A new filter required for Brazil’s 2026 election
All the details: github.com/xai-org/x-algorit…
Jay Baxter retweeted
> Release algorithm as a gesture of good faith
> Users misunderstand how prediction probabilities work
> “Reports subtract 468 likes! We’re being suppressed.”
Great explainer for how to interpret the newly-opened-sourced algo weights
Key thing to remember is that the weights don't multiply the raw like/report/etc. counts, they multiply your own personal predicted probability of liking/reporting/etc.
This is not accurate: "one report erases 468 likes." That's not what the code is doing. The weights are a multiple on your own predicted probability of Liking, Reporting, etc, which is substantially driven by your own behavior. And the baseline probability of a Report is more than 1000x lower.
Also: xitter.cf/nikitabier/status/2087…