The new stuff in the Twitter algorithm is wild.

"author_is_elon", "author_is_republican", "author_is_democrat", etc., are explicit terms that are special cased.

https://github.com/twitter/the-algorithm/blob/7f90d0ca342b928b479b512ec51ac2c3821f5922/home-mixer/server/src/main/scala/com/twitter/home_mixer/functional_component/decorator/HomeTweetTypePredicates.scala

the-algorithm/home-mixer/server/src/main/scala/com/twitter/home_mixer/functional_component/decorator/HomeTweetTypePredicates.scala at 7f90d0ca342b928b479b512ec51ac2c3821f5922 · twitter/the-algorithm

Source code for Twitter's Recommendation Algorithm - twitter/the-algorithm

GitHub
I'm a bit surprised at how many "well, actually, these are just metrics" replies I'm getting that seem to find the spirit of this comment plausible after (just for example), Elon asked why his engagement was dropping and then fired Twitter's most senior remaining engineer for telling him it's not a bug, which was followed by people seeing a lot more Elon in their feed and people also report seeing way more Republican stuff in their feed.

Like, how do you think ranking works? There are models/weights that determine a tweet's ranks that haven't been released and we know, e.g., Elon tweets, are now extremely highly weighted.

The NYT reported that Elon tweets have 1000x weight and the fatigue filter is disabled, but let's say their sources are wrong and there is no explicit special casing in ranking.

It's still the case that people report seeing way more Elon than before. How do you think this metrics data is being used?

@danluu I'm not sure but it feels like collecting this information and labelling users like this would breach GDPR? Political affiliation comes under special category data - you have to have a damn good reason to collect - let alone be using it in this way. I know GDPR only applies to EU citizens (and the UK for now) but we don't know how they're labelling users with this do we?
@danluu I wouldn’t be at all surprised if he planted these values there just to get to be the good guy and remove them. He already said expect it to be “embarrassing,” and seriously who names a metric “author_is_republican”
@mostlyharmlessz @danluu yeah but this looks like valid code, properly formatted. No way Elon could write anything to this level himself.
@danluu Someone opened a PR removing all that and it got merged. I'm just saying, I never used Twitter.

@danluu

I think rank and file engineers don't always have visibility into the fact that metrics are the product goals.

If you have a metric that wants to keep two groups balanced, then software will be written and models will be trained to achieve that goal. Especially if they are a metric tracked by leadership.

The launches that try to achieve this goal often aren't visible in the code base, because the only visible change is a parameter changing from 0.6 to 0.7.

@danluu Still metrics - though you aren't wrong, it's just that they design the algorithm to not "negatively impact" these groups rather than outright boost Elon.

(probably not good in any case however)

@danluu Even if this is all just for metrics, it's still hilarious that "Elon" is one of the special protected groups.
@danluu feels like we're getting a twitter for government dlc
@danluu Let’s assume for a second that there is absolutely no role for these in the ranking algorithm, they are all just for tracking alone. The question remain (1) why are those tracked at all, given the assumption we just made? For what purpose? And (2) Even if we assume these metrics are designed to be futile (if the assumption is true) what is the meaning of tracking of an individual tweet’s performance? How does it even compare with metrics of groups? WTAF?