Jay Baxter
@_jaybaxter_
Followers
2,402
Following
1,525
Media
37
Statuses
689
@CommunityNotes Founding ML Lead / Sr. Staff ML Eng @X . Prev BayesDB @MIT
Joined February 2011
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
高橋一生
• 145062 Tweets
飯豊まりえ
• 119657 Tweets
共同親権
• 97972 Tweets
MARK 200 OUT NOW
• 92783 Tweets
岸辺露伴
• 72464 Tweets
#BRIDGERTON
• 58236 Tweets
Colin
• 54813 Tweets
都知事選
• 50927 Tweets
ジョジョ婚
• 41226 Tweets
石丸市長
• 39191 Tweets
出馬表明
• 32102 Tweets
露伴先生
• 28803 Tweets
#素のまんま
• 27542 Tweets
ビシエド
• 21964 Tweets
Arda Kardeşler
• 21922 Tweets
石丸さん
• 11331 Tweets
Antero
• 10125 Tweets
Just merged a PR from
@avalanchesiqi
--
Reminder that Community Notes is a true community effort and we really do love and read your PRs 🙏
10
34
131
Community Notes don't affect post ranking (and shouldn't, in general -- they often add useful context to correct posts)
But we measured a ~25-35% drop in organic likes & RTs (& smaller drop in QTs, since QTs are sometimes dunks...) in an A/B vs. the same post w/o a note
We may
25
35
183
We are pro free-speech. Community Notes never censors any posts -- notes provide context only (we don't directly affect visibility or ranking of the original post).
Writing notes, even crappy notes, is not "abusing community notes". For notes to show on X, they need to be rated
Hey
@_jaybaxter_
,
the mere fact that you and your Community Notes colleague,
@__bradmiller__
(He/Him), follow both accounts
@_jasonwillz1_
and
@rt7683
, which systematically (!!!) abuse the function of Community Notes instead of limiting the proven abuse, speaks volumes.
Of
29
100
371
46
11
168
Compared to replies containing Politifact links, Community Notes are more neutral & less toxic.
Open source and independent auditability FTW: because the code and data are fully public, this analysis was run by external researchers and I’m hearing about it now for the first time
When people post misinformation and it's tagged as misinfo by a single person, they retreat into echo chambers, but when the post receive a collective tag (e.g., community notes), they are less likely to retreat, finds
@JunsolK
@gxin_ling
@profjamesevans
3
40
156
7
28
98
People agree with claims in a post significantly less if they see a Community Note attached (in our 2022 RCT survey)
Surprisingly often, it really is as simple as: people just didn’t know all the relevant info and are happy to update on it
Most of the kids chanting "from the river to the sea" have absolutely no idea what it means, and when you tell them what it means, they don't like it anymore.
254
766
6K
3
22
134
Want to work on Community Notes?
We're hiring an exceptional ML engineer and SWE (link in bio):
46
27
142
Major algorithm update today-- this is the first time we are running our bridging-based matrix factorization on rating tag data in addition to the top-level helpful ratings.
Result is a quality win + coverage win (Pareto improvement!!)
Multiple big quality updates launching this week. As always, code is open source in GitHub. First up:
We’re improving note quality while simultaneously increasing the number of notes shown on X by about 10%. Today’s update works by identifying notes that contributors from a wide
127
99
901
12
11
106
For whatever reason, there's a surprising amount of disagreement in Community Notes about Messi and Ronaldo -- e.g. whether one scored more goals than the other (usually boiling down to things like does "this year" mean calendar year or soccer season).
Well, it's not very smart
Community Notes is now bridging across topic-specific dimensions 📈📈📈
0
2
42
8
5
89
🚨Birdwatch Paper Alert!🚨
I am thrilled to finally be able to share this paper with the detailed results evaluating how effective Birdwatch's bridging algorithm is!
In the spirit of transparency, Birdwatch makes everything and the kitchen sink public and open source. Today, we’re sharing a paper that details the research, analysis and outcomes that informed its development and helped us understand its performance. 📊
28
135
610
10
11
87
We are adding an extra penalty for raters who rate a note "Helpful" that ends up being found not helpful by diverse raters-- we would like to encourage raters to be more picky about which notes they rate as helpful.
On the flip side, sometimes our best raters get dinged for
Another quality update:
We’re improving the way Rating Impact is calculated, which improves the average quality of proposed notes (as perceived by people from different points of view). This means contributors’ time and energy gets spent reviewing higher quality proposals.
How
136
78
724
4
2
77
This is why Community Notes doesn’t use labels in training
If we did, we’d only be as trustworthy as the labelers
5
8
77
👀
I designed a concept for the Community Notes Chrome extension!
With this extension, you can write notes not only on posts but also on articles outside of X!
What do you think?
44
32
560
8
4
69
Community Notes are getting even more accurate today. Although notes that are found helpful by people who have disagreed in the past are almost always accurate, today's algorithm update catches more of the ones that aren't.
Details well-described below 👇👇
Another big quality update:
We’ve enhanced the open source scoring algorithm’s ability to spot potentially inaccurate or unsupported notes. How it works: The algorithm identifies notes that people from a wide range of viewpoints agree have potential issues with accuracy or
563
612
5K
0
3
55
Want Community Notes on everything?
We get a lot of requests for “Community Notes on everything.” Would you use this? What would you want it to do? 👇
42
26
538
8
1
51
Media matching coverage 📈
This is one of the biggest levers for showing more notes on more posts, with ~0 delay between when new posts are created and when they get a note
Improved image matching will now show notes on ~30% more posts that contain similar or identical images.
We just rolled out the update and will be monitoring for any erroneous image matches.
252
344
2K
5
1
49
Community Notes keeps getting faster 👇
⚡Community Notes got faster, yet again. We just rolled out an optimization to the scorer that shaves 30mins of the time it takes notes to go live.
Under the hood details: we moved some of the computation (e.g. the initial matrix factorization) into a prescoring step, which
335
462
4K
5
3
46
🚄 Community notes will appear on posts over an hour faster 🚄
Community notes already often appear faster than traditional fact checks. Now we'll show them even faster by rearchitecting another piece of our system to use stream instead of batch processing
Community Notes just got faster, again! ⚡
Notes will appear over 1 hour faster. Since information travels quickly on X, this means notes will be seen by many more people at the time they read a post.
This builds upon a recent update, moving the Community Notes scoring system
590
501
5K
2
4
31
Community Notes is now bridging across topic-specific dimensions 📈📈📈
Under the hood update: Introducing Topic Models, which improve Community Notes’ ability to identify notes that are helpful to people from different points of view.
Community Notes shows notes that are found helpful by people who’ve disagreed in their past ratings. Now, it will
681
806
5K
0
2
42
Many of the best notes are on posts that get deleted.
This is mostly a huge win: lots of misleading posts voluntarily getting deleted! +shows goodwill from post authors
In other ways frustrating since not many people will ever see this huge iceberg of the best notes
We also examine the percentage of deleted tweets around this threshold and find that Community Notes increase the probability of a tweet being deleted by 80%.
2
5
24
6
5
40
Twitter Cortex is hiring an engineering manager for our machine learning/data science team here in Cambridge, MA- come save
@brainiaq2000
from managing me!
0
17
38
It's our 3rd birthday!🎂🎂🎂
The first non-empty commit in our repo, from 1/25/2021, had the entire algorithm as python code embedded in markdown, and it was a smoothed ratio 😬... suffice to say we have come a long way 😅
5
2
40
A lot of the best Community Notes go unseen because the original post gets deleted. That is a successful outcome!!
With this change, we are letting note authors know they succeeded when this happens, even though their note will not be widely visible
Contributors consistently say their goal is to keep others well-informed. This can happen when a helpful note appears on a post, and also when an erroneous post gets deleted. Starting today, writers will be notified when a post on which they wrote a note gets deleted.
79
76
816
4
7
37
Launching in India -- massive % increase in total eligible Community Notes contributors 📈📈
6
2
36
Community Notes will start feeling noticeably more real-time 📈📈
Community Notes just got faster ⚡
Starting today, notes will appear an average of 1.5 hours faster, and as much as 3.5 hours faster in some cases. Since information travels quickly on X, this means notes will be seen by many more people at the time they read a post.
How it
810
770
8K
5
1
36
It is pretty remarkable that the Community Notes algorithm works well out of the box across so many different countries. Misinformation is global, (mostly) not localized to specific countries
Welcome new contributors in 🇮🇩Indonesia, 🇲🇾Malaysia, 🇵🇭Philippines, 🇸🇬Singapore, 🇹🇭Thailand, 🇵🇬Papua New Guinea and 🇧🇳Brunei 👋
Amazing to see some of these new contributors already writing notes that have been found helpful enough to show on X.
2K
1K
10K
4
4
33
New research using fine-grained data finds that users repost 61% less often after a post gets noted!
🚨New WP: Community notes reduce the spread of misleading posts on X🚨
👉Based on *time series* data for N=31,758 cascades, we found that exposing users to CN reduced subsequent reposts by 61.4%
👉The effect on overall engagement was smaller & depended on how quickly CN appear
🧵
3
15
52
7
4
33
Model architecture improvement 👇 smarter normalization, weighting, and regularization
Under the hood update: We’re testing a new scoring model to surface additional helpful notes. The update identifies more instances of consensus among users who have historically disagreed.
As always, code and details are open-source in GitHub:
585
668
5K
6
2
30
Getting details like this dialed in is critical to making Community Notes as accurate as possible. Now raters will see other notes on a Tweet before rating a note (previously it was too easy to tap “helpful” on a note without reading possible counterarguments first)
When rating a note, it’s important that contributors have as much helpful information available as possible. Starting today, when contributors tap “Rate it” on a note, they’ll be taken to a page listing all note proposals, instead of just the note they’re looking at.
This way,
1K
1K
12K
5
3
31
Check out this external researcher's analysis of public community notes data!
What other platform allows this level of transparency?
@WHO
@X
@CommunityNotes
@elonmusk
"Misinformation" is mentioned in JAMA Network journal more than 830 times. However, 0 of these articles offer empirical evidence about the effectiveness of countermeasures to address misinformation. Today we published
#1
. But why was it zero? Social media companies
@facebook
2
2
15
2
5
32
Community Notes (Birdwatch) is “one of the most exciting content moderation innovations ever to come out of not just Twitter, but any major platform.”
Buried underneath all of the froth happening on Twitter at the moment is an experiment more exciting than anything I’ve seen tried by a tech giant. I explain what’s going on for
@WIRED
132
351
2K
1
5
32
@emollick
The most visible posts get noted much faster than the average posts, so the coverage of notes is much better on a per-impression basis
Also-- if a note shows up after you engaged with a post, we send you a notification!
The eng team is working on more
Sometimes a note appears on a post after you’ve seen it. To help, Community Notes sends notifications to people who have engaged with a post that later receives a note. We’ve now scaled it up to handle even the most visible and highly-engaged posts, so more of you will be (and
659
558
5K
1
2
31
When we said it got faster we weren't joking! Love seeing the external analysis double checking us.
IMO, for this launch, the more impressive win was in minimum time to CRH, which dropped by a similar amount-- a disproportionately large % speedup on the most visible notes
A month ago It was announced that community notes got faster, and it actually did!
From a median approval time of around 7.5 it went down by 1.5 hours!
*The graph is a 7 day moving average, I tried to account for the fact that the latest data will be revised upwards by using
0
3
41
4
3
31
Awesome use of the open-source Community Notes algorithm in the wild (from a great team)
Love seeing it applied in domains beyond misinfo -- here it's used to construct a live constitution of policies that were rated helpful by people who tended to disagree on other policies
Happy to share the work Energize AI has done with
@OpenAI
and 9 other teams from all around the world.
We're introducing Aligned, a platform for global governance and alignment of frontier models, and eventually superintelligence.
I recommend reading all the reports from all
3
1
25
4
2
31
This also cleans up a ton of crap notes to save contributors' time: we label ~15-20% more notes "Currently Rated Not Helpful"
How? We find notes with very high numbers of "harassment and abuse" tags from raters who've tended to disagree about harassment/abuse in the past, and
3
2
27
@AaronBergman18
Thanks! I will try to write up a good blog/LW-style post about it. There is already our paper () but I think having a lighter reading version with more intuitions explained would be cool
1
0
28
@bigratkiller
@CommunityNotes
That requirement only ever exists when rating a note helpful (you can rate not helpful without clicking the source link). Is there a case when you'd think a note with a phishing link is helpful? Or some other case we haven't thought about?
2
1
27
@mattyglesias
If this claim is true I wonder how much Community Notes played a role in it. Anecdotally, I've heard people from both sides politically tell me that Community Notes helped them see misinfo from their own side for the first time
4
3
28
Getting noted causally reduces the engagement on a post significantly! This is entirely organic due to organic user behavior since we don’t rank posts differently when they are noted
We found a very large causal drop in reposts, quotes, and likes on noted posts over a year ago
And this effect is very strong! The appearance of a Note roughly halves the number of retweets (after treatment). A significant, albeit smaller, effect is observed when focusing on the number of replies or quotes.
3
4
37
5
6
29
One reason I am long Community Notes
Open source AI foundation models will wipe out closed and proprietary AI models for the same reason Wikipedia wiped out generalist commercial encyclopedia: crowd-sourced human contributions to open platforms can cater to a high diversity of interests, cultures, and languages.
264
636
4K
3
1
28
Community Note appears ~2 days faster than traditional fact checks in this example recently highlighted in the news 📈
This article seems intended as a critique, but actually demonstrates how fast Community Notes is. It cites a Note appearing on a viral post after 4hrs. This is extremely fast relative to most historic approaches. Traditional fact checks in this case were published 2-4 days
17
28
250
1
3
28
@TheElipsoid
💯 Desired end state is 100% algorithmic identification of topics. This initial test is an easy way to ship & test the idea quickly.
1
3
27
Internet Archive is super important for snapshotting historical versions of webpages in Notes — e.g. thanks to them, there is a public log of when AP and BBC changed their headlines after getting Noted yesterday
I donated!
Even though Internet Archive has a ton of negative bs about me, it is still a public good that should exist.
Support the Internet Archive!
4K
7K
50K
1
3
27
Reminder that the Community Notes algorithm code and data are totally open source, so you-- yes, you-- can run an analysis just like
@VitalikButerin
did here to understand why any particular Community Note is showing or not.
1
2
26
This is why bridging-based algorithms work a lot better than people expect them to
2
5
26
Interesting tidbit: "Once a false post is identified, a third-party fact-check may take longer than a Community Note."
Past research has found that the median time until a fact check article created is around twice as long as the median time for a Community Note to appear!
X’s crowdsourced fact-checking system, Community Notes, has been touted by Elon Musk as a way to improve information accuracy.
But we found that over 80% of the thousands of debunked posts about the Israel-Hamas conflict we reviewed had no such notes.
15
171
305
3
2
24
One of our most requested features is now fully rolled out!
Sometimes a note appears on a post after you’ve seen it. To help, Community Notes sends notifications to people who have engaged with a post that later receives a note. We’ve now scaled it up to handle even the most visible and highly-engaged posts, so more of you will be (and
659
558
5K
2
1
24
Come work with me on Birdwatch, possibly the most interesting and challenging project I've ever worked on! Feel free to DM me if interested
We’re looking for a Data Scientist for Birdwatch. Want to help create a better informed world by building tools to help prevent the spread of misleading info? Love finding insights in data? Let’s chat. More details:
4
29
56
0
5
24
Loving all the concepts for viewing Community Notes everywhere
Here's my concept for X Browser:
It can be an in-app feature and it could be used to bring Community Notes for the web to Android/iOS as well.
A Chrome extension would be good for desktops but we probably need X Browser to bring them to mobile phones!
What do you think?
37
24
207
3
2
25
Apply to help us build
@CommunityNotes
!
Give power to the people to keep each other informed. Apply 👇
Software:
Machine Learning:
78
85
445
2
5
23
<3 notes about quasi-experimental variation in housing supply
THE CHART YIMBY KIDS DON’T WANT YOU TO SEE:
House prices and new construction are POSITIVELY correlated. I.e. when there is more supply prices are rising and when there is less they are falling. The OPPOSITE of the YIMBY argument. 🏠🤪
113
12
81
1
4
23
The key piece of the algorithm is: in order for a note to labeled helpful and shown as context on a Tweet to all users, it must be rated helpful by a users who typically disagree in their ratings ("bridging" across those groups of users who disagree)
0
1
20
Yes!! Also important that we don’t estimate political preferences explicitly, we estimate agreement patterns in general, which is necessary when disagreements don’t exactly map to traditional politics
This is really neat: Community Notes work because they estimate users' political preferences and upweight notes that are liked by users of all stripes.
I wonder if other parts of our democracy could be improved with mechanisms like this.
9
72
497
1
3
19
Raising average note quality!
✨Note writing limits now consider hit rate, a writer’s ratio of helpful to total notes written. To write many notes in 24h, a writer must have a history of helpful notes and a solid hit rate. This prevents large volumes of proposed notes that aren’t found helpful. Details in
446
374
3K
2
1
20
Birdwatch just launched author and rater helpfulness scores! It's an important step towards better note ranking, and there's lots more yet to come, including factoring in whether the raters of a note come from diverse points of view
As always, we’ve published the code that underpins this update. We welcome ideas, critiques or any observations that might help make us make this update stronger:
2
1
22
0
0
20
Here's an even bigger effect than reduced organic likes+reposts+quotes: tons of post authors *delete* their post entirely after it gets Noted
Hard to beat a 100% drop in views
@_jaybaxter_
@elequidistante
@el_pais
Oh yeah I guess I already almost have that data I store handles of deleted tweetIds as “not found”. There 9406 of them now. And a total of 52286 handles. So about 18% of notes that are currently rated helpful have had their tweets deleted within 2-3 days. But after that there are
0
1
7
0
2
18
Love that you built this with the public data, keep it up!
I do wish it wasn’t called a leaderboard though, since I’ve seen a lot of people being confused about whether these are the best note authors/contributors
2
0
18
This large drop in organic sharing (and agreement with original post) is consistent with what we've found for notes in
🚨New review🚨 Do warning labels on misinformation actually work?
In Current Opinion in Psychology,
@DG_Rand
& I show that the answer is YES: warnings can reduce belief & sharing by ~25-50%!
We also review moderators (eg more visible=bigger effect)
5
32
118
3
5
18
Glad that we are reducing the incentive to spread falsehoods!
Some helpful Community Notes add context to posts that aren't misleading. I'm curious to see whether this will change rating norms at all: will the community pivot to only add context on misleading posts?
Making a slight change to creator monetization:
Any posts that are corrected by
@CommunityNotes
become ineligible for revenue share.
The idea is to maximize the incentive for accuracy over sensationalism.
21K
25K
259K
5
4
17
Excited to work on the algorithms and data that power Birdwatch, e.g. ?
We've got 2 awesome new roles we are looking to fill! One focused on modeling (ML/stats) and one focused on data/infra
ML:
Data:
This expansion comes with a complete overhaul of the way Birdwatch identifies which notes to show on Tweets. Now, notes need to be rated helpful by enough people from different points of view in order to appear on a Tweet. This is how it works:
2
4
37
1
8
18
We continue to see algorithmic wins from work looking at agreement from raters who typically disagree— this time on negative signals (eg harassment/abuse)
We recently launched an update that improves note quality while simultaneously increasing the number of notes shown on X by about 6%. This update also identifies 6% more Not Helpful notes, which in turn limits writing ability of authors regularly proposing such notes.
It works
333
283
2K
3
1
17
It is a hard, messy human problem! Community Notes would not work out of the box without a very well-thought-through product design, and a moderately complex algorithm on top (raters who have disagreed in the past must find a note helpful)
Remember when everyone used to say that we couldn't improve content moderation with technical solutions ("it's a messy human problem!") and then they went and did it with Community Notes
1
3
33
1
3
16
Now earning out low-quality notes contributors faster
We appreciate those of you who take the time to rate bad notes as not helpful and now take more aggressive action based on those. Thank you!
Contributors’ time is valuable and best spent reviewing high quality notes. We just rolled out an update to boost the quality of proposed notes by making it easier for accounts to lose writing ability if their notes are found Not Helpful by people from wide-ranging points of
457
340
3K
5
0
17
Amazing how well bridging algorithms work -- Community Notes keep being found helpful on content from all over the political spectrum
E.g. from the last 2 days:
1) Fox News featured a note on a PETA post:
2) 2020 Election fraud:
NEVER FORGET: From 130,000-280,000 Completed Ballots Were Shipped Across State Lines from NY to PA in 2020 Election – Never Explained (VIDEO) via
@gatewaypundit
480
10K
17K
1
3
16
"Prototype democratic process" - sounds a lot like a job for something like Community Notes' bridging algorithm, or
@UsePolis
!
one thing coming up in the debate about the pause letter i really agree with:
openai should make a great alignment dataset and alignment evals and release those!
bonus points if we can find a prototype democratic process for 'what we align to'.
505
360
5K
0
9
17
Curious to see how many notes on April fool's jokes we'll get... lot of noters vs. NNN wars being waged today
2
1
17
Twitter's Analytics team (which includes sub-teams like Product Data Science and Applied Research) is hiring data scientists, analysts, data engineers, and a product manager! DM me if interested
0
3
14
Saving our raters’ time, and cleaning up the UI!
Starting today, you won't be prompted to rate notes if all notes on a post are already rated "Not Helpful" or explain why a note isn’t needed. Saves contributors’ time so they can focus where ratings matter most.
104
86
845
2
0
16
Really excited about this launch, since our model can now quantify its uncertainty!
As a result, we can confidently label many more notes “Not Helpful”, by using the upper bound of the uncertainty interval to set the threshold
How does it work? This algorithm update effectively adds a confidence score alongside the existing note helpfulness score. Some notes that have low helpfulness scores are likely to maintain low helpfulness scores as new ratings come in, while others might see their helpfulness
84
84
843
1
1
17
Really excited about this expansion.
I’m curious to see how the latent viewpoint spaces of these new expansion areas compare to the existing ones. Any guesses?
Welcome new contributors in 🇭🇰Hong Kong, 🇰🇷South Korea and 🇹🇼Taiwan 👋
If you’d like to contribute to Community Notes, sign up at
257
264
2K
5
1
16
A must read!! Excited to see more like this in the future too, esp. analyzing our new matrix factorization scorer meant to address the partisanship shown in this paper. Should be much easier for researchers now that we release fully runnable algo code!
🚨Paper alert!🚨Birds of a Feather Don't Fact-check Each Other
#CHI2022
On Twitters crowdsourced factcheck program
@birdwatch
, partisanship is king: users mostly flag counterpartisans' tweets as misleading & rate their notes as unhelpful
1/
6
128
458
0
0
16
+1: this account auto-quotes posts that have especially helpful notes (as judged by our raters + algorithm)
2
3
14
Whether a note contains a source is one of the biggest correlates of whether the note will be rated helpful or not- and now they’re required!
We previously didn’t require sources, since occasionally there are situations where there really is no viable external source.
Rather
Notes that cite sources have a much higher likelihood of earning a 'Helpful' status. This is not surprising, as sources make notes more easily verifiable by viewers and raters. Starting today, sources are now required for proposed notes.
We haven’t previously required this, as
206
233
2K
1
2
11
This article claims that
>"numerous X users proposed Community Notes adding context to Musk’s tweet of the Pizzagate meme. As of early Tuesday afternoon, however, none had received enough positive votes from other users to be shown on X."
While that was true at their original
Elon Musk voiced support Tuesday morning for Pizzagate, the long-debunked conspiracy theory that led a man to fire a rifle inside a Washington D.C. restaurant in 2016.
212
236
494
1
2
15
While I wasn’t directly impacted by the layoffs yesterday, I have been forever impacted by many extremely smart and talented friends who were. I salute all of you and hope we work together again. In the meantime, please reach out if I can be helpful connecting you to new roles.
0
0
15
Note definitely needed on these. I think this may be one of the most misunderstood distinctions
I love when community notes does this correctly e.g.
I correct my students when they mean a "one percentage point" change if they write it as a "1%" change -- for example, in describing unemployment rates. I've seen prominent economists not make this distinction. Am I being too pedantic?
212
93
2K
1
0
15
Thank you to our contributors who give us amazing feedback -- we are listening and acting on it!
Thank you to contributors who've shared feedback about cases of prolific producers of notes that aren't often found helpful. Today's update is based on your input:
22
30
493
2
0
14
@ESYudkowsky
One great thing about Community Notes is that you can go look at the Github repo and public data yourself!
See here a changelog of algorithm changes:
And the old Birdwatch paper detailing the bridging-based ranking algorithm:
0
0
14
Dramatic improvement in minimum time for notes to become visible, too:
high visibility posts can get noted within an hour or two now! (even for original non-matched notes)
plot created via public data:
NEWS: Community Notes is now Much faster!
Approval time improved by over an hour from a median of 7.5 hours to 6 hours!
34
40
391
1
1
13
@colin_fraser
Hypothesis: in the training data, + and * may mostly actually refer to the math operators. But - frequently indicates a range, and / can indicate other things like dates?? (or, more likely, the models are just... not great at math)
1
0
14
Notes on media FTW
One confounder: only top writers are currently allowed to write notes that will automatically match on media.
Likely that both notes written by top writers and media notes (controlling for top writer) are more likely to reach consensus
here's something mildly positive and interesting from the Community Notes dataset: Community Notes on media are about twice as likely to reach consensus and be displayed on the app (31% vs 16%), and the median time-to-consensus for media notes is an hour and a half faster
0
2
6
4
1
13
More precise image matching for notes on media!
Under the hood: we added perceptual hashing as a second pass to catch false positives
Under the hood update: We’ve improved image matching. Notes written on images are shown on other posts that include the same image. While not common, there were occasional cases of images incorrectly matching. We improved precision of matching to avoid these. Thank you to the
81
58
821
3
1
13
👀
3
3
13
Congrats to everyone on the BayesDB paper! Proud to have been a part of this
New paper: BayesDB: A probabilistic programming system for querying the probable implications of data
1
3
13
0
1
12
Work done with an amazing team of coauthors:
@stefanjwojcik
@loafieee
@nclarkjudd
@DeliaMocanu
@stephen_ragain
@mbfhunzaker
@kcoleman
with loads of help from the Birdwatch engineering team!
2
3
13
I accidentally discovered this during one of my first projects at Twitter: I was analyzing trajectories of users over time, trying to understand what made people become and stay heavy users.
So I did an SVD of the usage trajectories and found... just a bunch of sine waves 🤦
Interesting! I hope it's not just SVD/PCA on almost random noise because it's likely to have found similar patterns... (ie. you get sine waves when doing PCA on noise)
#Statistics
12
20
389
0
0
13
@CommunityNotes
For those having trouble finding it:
1. Go to your profile
2. Settings and privacy
3. X Premium
4. Extras
5. Custom navigation
6. Toggle on Community Notes
3
2
10
@TRHLofficial
I'm not an expert on our TOS (I only work on Community Notes) but here's the policy:
If you see something that seems to violate the policy, I'd definitely suggest reporting it.
3
0
12
@chris_bail
@BrendanNyhan
@NathanKalmoe
Thank you Chris! Your team’s early work on bridging algorithms (in additions to others’ like
@UsePolis
) was a major inspiration and laid the foundation for the novel bridging algorithm developed and analyzed in our paper 🙏
0
0
11
@protomemetic
@CommunityNotes
We should! Here's a start, from last year:
How does Community Notes avoid bias, pile-ons, ...?
Video transcript:
105
185
1K
2
1
11
Increased transparency on notes image matching
Notes written on images and videos are often shown on hundreds (sometimes thousands) of posts that contain matching media. We just launched an update to make it easier to see the post on which the note was originally written, so people can see the original context:
While not
774
644
5K
1
1
11
@MarkZlochin
The results are from a timespan where we had launched widely in the US but not yet other countries (I'd argue it wasn't really a beta test phase at the point we ran the A/B test)
Although we'd love new data, there is a huge cost to running another A/B test now: users expect to
1
0
10
Curious if prediction markets can accurately predict if posts will get Community Notes:
1
1
11