I was an early member of Birdwatch/Community Notes and was in direct contact with the team at Twitter during the launch and first year of operation. While I no longer participate in the program...
I was an early member of Birdwatch/Community Notes and was in direct contact with the team at Twitter during the launch and first year of operation. While I no longer participate in the program (having left Twitter), I remain one of the highest-rated note writers on the platform.
The biggest concern expressed by the top contributors was (and remains) the requirement that notes be rated "helpful" by users with a "diversity of perspectives." This means that in order for a note to be shown publicly on Twitter, it must be rated helpful by users who disagreed with each other in the past. Functionally, this means the vast majority of notes are never shown (<10%) and never have a chance to be shown if they pertain to a particularly polarizing topic. The low success rate makes burnout very real, particularly for subject-matter experts, especially when factually-accurate notes languish in "needs more ratings" purgatory because their model needs an anti-vaxxer to rate your note positively. From what I've seen, it feels like the only people still actively participating are highly partisan and trying to "win" against the other side.
Aside from all that, the original Birdwatch team was an absolute pleasure to interact with and it was fascinating to see the algorithm develop over time (and provide my own inputs!). Unfortunately, much of the team from Twitter Research and Twitter Cortex that made that possible were gutted during the early Musk purges. So while the program continues to expand globally with a skeleton crew, the initial plans to foster a robust Birdwatch community that would help guide future development has been completely abandoned.
If the people participating are highly partisan then it seems rather remarkable that the output mostly isn't? But yeah, it sounds like the downside is a lot of effort going into it, given the...
If the people participating are highly partisan then it seems rather remarkable that the output mostly isn't?
But yeah, it sounds like the downside is a lot of effort going into it, given the amount of output.
That's very much by design with how the algorithm works. I haven't looked at the numbers in months, but the majority of helpful notes (i.e. publicly shown) were written on less polarized topics...
If the people participating are highly partisan then it seems rather remarkable that the output mostly isn't?
That's very much by design with how the algorithm works. I haven't looked at the numbers in months, but the majority of helpful notes (i.e. publicly shown) were written on less polarized topics (memes, obviously fake images, scams, influencer nonsense, etc.). These definitely provide important context to a lot of the garbage on Twitter but they aren't as "important" as a note about the 2020 election or mRNA vaccine safety.
The problems really arise on more partisan issues where users rigidly align to left vs. right in their voting patterns. We'd frequently see well-written, objectively-correct, well-sourced notes with hundreds of positive ratings fail to achieve the necessary threshold until a single right-leaning user voted in favor. The algorithm is effectively "both sides-ing" everything with no regard for factual accuracy or expertise.
Could it fail by becoming entirely deadlocked?
To some degree it's always been deadlocked on partisan issues? I think it will continue to function reasonably well on other topics (or at least until they become polarized).
Given that it’s an algorithm with no intelligence of its own, this seems fairly inevitable? Any expertise needs to come from the participants. It sounds like it’s not exactly easy to veto since...
The algorithm is effectively "both sides-ing" everything with no regard for factual accuracy or expertise.
Given that it’s an algorithm with no intelligence of its own, this seems fairly inevitable? Any expertise needs to come from the participants.
It sounds like it’s not exactly easy to veto since one defector would be enough, but still very possible to veto by a unified voting bloc.
I guess if the metric is "disagreement" it doesn't necessarily need to break down as partisan. Like if people really got into it in 2016 arguing about Bernie and Hillary then they're strongly...
If the people participating are highly partisan then it seems rather remarkable that the output mostly isn't?
I guess if the metric is "disagreement" it doesn't necessarily need to break down as partisan. Like if people really got into it in 2016 arguing about Bernie and Hillary then they're strongly disagreeing but when it comes to dunking on Elon Musk or Trump they'll be aligned.
All the ones I've seen so far have added valuable context to otherwise seriously misrepresented situations, though the fact that they work similarly to a reddit thread is concerning. Twitter had a...
All the ones I've seen so far have added valuable context to otherwise seriously misrepresented situations, though the fact that they work similarly to a reddit thread is concerning.
Twitter had a huge issue before of spreading massive witch hunts with out of context clips/pictures, and anything that helps to mitigate those issues is a fine by me.
This seems much better than a reddit thread. Optimistically: imagine a non-opaque algorithm that actually works! But we don't know if that will last. Cynically, it makes sense to expect that...
This seems much better than a reddit thread.
Optimistically: imagine a non-opaque algorithm that actually works!
But we don't know if that will last. Cynically, it makes sense to expect that eventually someone will figure out how to game it.
Vitalik Buterin wrote a pretty interesting deep dive into how one of Twitter's algorithms works. There's some heavy math involved, but some parts can be explained without it. From the blog post:...
Vitalik Buterin wrote a pretty interesting deep dive into how one of Twitter's algorithms works. There's some heavy math involved, but some parts can be explained without it.
From the blog post:
Community Notes is a fact-checking tool that sometimes attaches context notes, like the one on Elon's tweet above, to tweets as a fact-checking and anti-misinformation tool. It was originally called Birdwatch, and was first rolled out as a pilot project in January 2021. Since then, it has expanded in stages, with the most rapid phase of its expansion coinciding with Twitter's takeover by Elon last year. Today, Community Notes appear frequently on tweets that get a very large audience on Twitter, including those on contentious political topics. And both in my view, and in the view of many people across the political spectrum I talk to, the notes, when they appear, are informative and valuable.
But what interests me most about Community Notes is how, despite not being a "crypto project", it might be the closest thing to an instantiation of "crypto values" that we have seen in the mainstream world. Community Notes are not written or curated by some centrally selected set of experts; rather, they can be written and voted on by anyone, and which notes are shown or not shown is decided entirely by an open source algorithm. The Twitter site has a detailed and extensive guide describing how the algorithm works, and you can download the data containing which notes and votes have been published, run the algorithm locally, and verify that the output matches what is visible on the Twitter site. It's not perfect, but it's surprisingly close to satisfying the ideal of credible neutrality, all while being impressively useful, even under contentious conditions, at the same time.
...
The way that the score is calculated is what makes the algorithm unique. Unlike simpler algorithms, which aim to simply calculate some kind of sum or average over users' ratings and use that as the final result, the Community Notes rating algorithm explicitly attempts to prioritize notes that receive positive ratings from people across a diverse range of perspectives. That is, if people who usually disagree on how they rate notes end up agreeing on a particular note, that note is scored especially highly.
...
The core clever idea here is that the "polarity" terms absorb the properties of a note that cause it to be liked by some users and not others, and the "helpfulness" term only measures the properties that a note has that cause it to be liked by all. Thus, selecting for helpfulness identifies notes that get cross-tribal approval, and selects against notes that get cheering from one tribe at the expense of disgust from the other tribe.
...
Probably the single most important idea in this algorithm that distinguishes it from naively taking an average score from people's votes is what I call the "polarity" values.
Apparently the algorithm designers thought about making this multidimensional but it wasn't implemented.
Polarity is assigned to both users and notes. The link between user IDs and the underlying Twitter accounts is intentionally kept hidden, but notes are public. In practice, the polarities generated by the algorithm, at least for the English-language data set, map very closely to the left vs right political spectrum.
...
The main thing that struck me when analyzing the algorithm is just how complex it is. There is the "academic paper version", a gradient descent which finds a best fit to a five-term vector and matrix equation, and then the real version, a complicated series of many different executions of the algorithm with lots of arbitrary coefficients along the way.
They can be pretty fun and actually helpful. Though with some of them you can see that the note writer isn't entirely unbiased. Musk at least deleted one correcting him directly (unless that one...
They can be pretty fun and actually helpful. Though with some of them you can see that the note writer isn't entirely unbiased.
Musk at least deleted one correcting him directly (unless that one was fake, you can never really tell unless you saw it with your own eyes).
I was an early member of Birdwatch/Community Notes and was in direct contact with the team at Twitter during the launch and first year of operation. While I no longer participate in the program (having left Twitter), I remain one of the highest-rated note writers on the platform.
The biggest concern expressed by the top contributors was (and remains) the requirement that notes be rated "helpful" by users with a "diversity of perspectives." This means that in order for a note to be shown publicly on Twitter, it must be rated helpful by users who disagreed with each other in the past. Functionally, this means the vast majority of notes are never shown (<10%) and never have a chance to be shown if they pertain to a particularly polarizing topic. The low success rate makes burnout very real, particularly for subject-matter experts, especially when factually-accurate notes languish in "needs more ratings" purgatory because their model needs an anti-vaxxer to rate your note positively. From what I've seen, it feels like the only people still actively participating are highly partisan and trying to "win" against the other side.
Aside from all that, the original Birdwatch team was an absolute pleasure to interact with and it was fascinating to see the algorithm develop over time (and provide my own inputs!). Unfortunately, much of the team from Twitter Research and Twitter Cortex that made that possible were gutted during the early Musk purges. So while the program continues to expand globally with a skeleton crew, the initial plans to foster a robust Birdwatch community that would help guide future development has been completely abandoned.
If the people participating are highly partisan then it seems rather remarkable that the output mostly isn't?
But yeah, it sounds like the downside is a lot of effort going into it, given the amount of output.
Could it fail by becoming entirely deadlocked?
That's very much by design with how the algorithm works. I haven't looked at the numbers in months, but the majority of helpful notes (i.e. publicly shown) were written on less polarized topics (memes, obviously fake images, scams, influencer nonsense, etc.). These definitely provide important context to a lot of the garbage on Twitter but they aren't as "important" as a note about the 2020 election or mRNA vaccine safety.
The problems really arise on more partisan issues where users rigidly align to left vs. right in their voting patterns. We'd frequently see well-written, objectively-correct, well-sourced notes with hundreds of positive ratings fail to achieve the necessary threshold until a single right-leaning user voted in favor. The algorithm is effectively "both sides-ing" everything with no regard for factual accuracy or expertise.
To some degree it's always been deadlocked on partisan issues? I think it will continue to function reasonably well on other topics (or at least until they become polarized).
Given that it’s an algorithm with no intelligence of its own, this seems fairly inevitable? Any expertise needs to come from the participants.
It sounds like it’s not exactly easy to veto since one defector would be enough, but still very possible to veto by a unified voting bloc.
I guess if the metric is "disagreement" it doesn't necessarily need to break down as partisan. Like if people really got into it in 2016 arguing about Bernie and Hillary then they're strongly disagreeing but when it comes to dunking on Elon Musk or Trump they'll be aligned.
All the ones I've seen so far have added valuable context to otherwise seriously misrepresented situations, though the fact that they work similarly to a reddit thread is concerning.
Twitter had a huge issue before of spreading massive witch hunts with out of context clips/pictures, and anything that helps to mitigate those issues is a fine by me.
This seems much better than a reddit thread.
Optimistically: imagine a non-opaque algorithm that actually works!
But we don't know if that will last. Cynically, it makes sense to expect that eventually someone will figure out how to game it.
Vitalik Buterin wrote a pretty interesting deep dive into how one of Twitter's algorithms works. There's some heavy math involved, but some parts can be explained without it.
From the blog post:
...
...
...
Apparently the algorithm designers thought about making this multidimensional but it wasn't implemented.
...
I like the ones I've seen screenshots of and am kind of amazed that Musk hasn't killed them yet.
They can be pretty fun and actually helpful. Though with some of them you can see that the note writer isn't entirely unbiased.
Musk at least deleted one correcting him directly (unless that one was fake, you can never really tell unless you saw it with your own eyes).