BRENNEN

Industries / Technology / Social media

Feeds and ranking

This is the work of choosing which posts each person sees, and in what order, when they open the app.

7

Milestones achieved, from 2006 to 2025.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 2006

    Friends' news in one stream

    On 5 September 2006 Facebook launched News Feed and Mini-Feed, which showed in one running list what a person's friends were doing: relationship changes, groups joined, photos uploaded. Nothing new was shared, but within a day a group called "Students Against Facebook News Feed" was close to 100,000 members, TechCrunch reported, and Mark Zuckerberg answered: "Calm down. Breathe. We hear you."

    LessonGathering posts that were already visible into one stream changes who notices them, and people experience that as a loss of privacy.

    techcrunch.com

    Credit

    Facebook

  2. 2010

    A ranking formula explained

    At Facebook's f8 developer conference in April 2010, engineers Ruchi Sanghvi and Ari Steinberg described EdgeRank, the method that picked the subset of friends' stories shown in News Feed. Each interaction with a post, called an edge, was scored by the viewer's affinity with its creator, a weight for the kind of interaction (a comment counting for more than a like) and its age, and the scores were added up. Developers soon spoke of "News Feed Optimization", writing posts to make the cut.

    techcrunch.com

    Credit

    Ruchi Sanghvi, Ari Steinberg, Facebook

  3. 2016

    Timelines leave time order

    On 10 February 2016 Twitter added a setting, "Show me the best Tweets first", and said it would soon be on by default. On 15 March 2016 Instagram said it would order posts by the likelihood a user would be interested, their relationship with the poster and timeliness, saying the average user missed 70% of their feed (Instagram's figure). Twitter kept a randomly chosen 1% of its users worldwide on the old time-ordered timeline as a control group, and its own researchers later used that group in PNAS to show that in six of seven countries studied, the mainstream political right reached more people through ranking than the mainstream left.

    LessonA control group kept for years is what made the effect of ranking measurable later.

    thenextweb.comtechcrunch.compmc.ncbi.nlm.nih.gov

    Credit

    Twitter; Instagram; Ferenc Huszár and colleagues at Twitter

  4. 2022

    A TikTok-style feed reversed

    In 2022 Instagram tested a full-screen feed with more posts recommended from accounts people did not follow. After a petition to "Make Instagram Instagram again" spread, shared by Kylie Jenner among others, Instagram's head Adam Mosseri told the newsletter Platformer in late July 2022 that Instagram would phase out the full-screen test and cut recommended posts for a time, because "people are frustrated, and the usage data isn't great." The same week Mark Zuckerberg said about 15% of what people saw in Facebook and Instagram feeds was suggested content, with a target of 30% by the end of 2023, Fortune reported.

    LessonPosts from strangers need to clear a higher bar than posts from accounts a person chose; in Mosseri's words, "it should just be great.".

    fortune.com

    Credit

    Instagram, Adam Mosseri

  5. 2023

    Time order tested on Facebook

    Outside academics working with Meta researchers moved consenting Facebook and Instagram users to reverse-chronological feeds for three months of the 2020 US election and published the results in Science in July 2023. Time on both platforms and activity fell substantially. Users saw more political and more untrustworthy content, and on Facebook less uncivil content. Issue polarization, affective polarization (dislike of the other side) and political knowledge did not change significantly.

    LessonChanging the order of a feed changed what people saw and how long they stayed, but over three months it did not measurably change their political attitudes.

    ebi.ac.uk

    Credit

    Andrew Guess, Neil Malhotra, Jennifer Pan, Pablo Barberá and colleagues; Meta

  6. 2023

    Feeds anyone can build

    In May 2023 Bluesky launched custom feeds, which let outside developers build feeds that users install in place of the default, a design it calls algorithmic choice. By July 2023 Bluesky said "the vast majority of feeds on the network are built by independent third-party developers", and tools such as SkyFeed let people build a feed without writing code.

    bsky.social

    Credit

    Bluesky

  7. 2025

    Feeds reranked without platforms

    Tiziano Piccardi, Martin Saveski, Chenyan Jia, Jeffrey Hancock, Jeanne Tsai and Michael Bernstein built a way to reorder participants' X feeds in real time without the platform's help, using a large language model to find posts that expressed antidemocratic attitudes and partisan animosity. In a preregistered 10-day field experiment with 1,256 participants during the 2024 US presidential campaign, showing fewer or more such posts shifted dislike of the other party by more than 2 points on a 100-point scale. Science published it on 27 November 2025.

    LessonOutside researchers can now test a ranking change without a platform's permission.

    ebi.ac.uk

    Credit

    Tiziano Piccardi, Michael Bernstein and colleagues

Watch for

  1. Published code is what runs

    X first published ranking code in March 2023 (the repository twitter/the-algorithm) and published a new repository for its For You feed on 20 January 2026, which ranks posts with a transformer model it calls Phoenix. An update of 13 August 2026 says it added the weights used to combine predicted actions into a post's score and replaced a demonstration model with "the code used to train the models the feed uses", with synthetic data so anyone can run a small training run.

    Achieved whenan independent auditor or outside researchers confirm that the ranking X runs in production matches the published code and weights.

    github.com

    Claimed by

    X, in its xai-org/x-algorithm repository, August 2026

  2. Telling the feed your tastes

    Instagram's "Your Algorithm" shows people the topics it thinks they like and lets them add or remove them. It was tested on Reels from October 2025, reached Explore in April 2026 and was extended to the main feed in June 2026, when Adam Mosseri said that as recommendations took over the main feed, choosing whom to follow "quietly stopped working" as a way to shape it.

    Achieved whenan outside test shows that adding or removing a topic changes what the feed shows, by a measured amount, and that the change lasts.

    socialmediatoday.com

    Claimed by

    Instagram, June 2026

Open

  1. 01

    Check simulators against field results

    A scorecard that runs the open social media simulators, in which AI agents post, follow and read ranked feeds, through the two feed changes that have been measured on real platforms: the switch to time-ordered feeds (Guess and colleagues, 2023) and the down-ranking of hostile partisan posts (Piccardi and colleagues, 2025). For each simulator, the scorecard records whether it reproduces the direction and rough size of each published result: less time spent, more political content, less uncivil content, no change in polarization over three months, and a shift of about 2 points in dislike of the other party.

    The brief
    Credit · unclaimed
  2. 02

    Test Instagram's topic controls

    A controlled test of Instagram's "Your Algorithm" topic controls, using the cloning design that Northeastern University researchers applied to TikTok. The team creates pairs of fresh test accounts with identical follows and viewing histories; then one account in each pair adds or removes a topic and its twin does nothing. Every day both accounts scroll a fixed number of feed items, and a model labels each item's topic on the spot. Only the topic label and the item's position are kept, never a post, name or handle. The test measures how far the topic's share of the feed moves, how fast, and whether it drifts back.

    The brief
    Credit · unclaimed

Open 01 · the brief

Check simulators against field results

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A scorecard that runs the open social media simulators, in which AI agents post, follow and read ranked feeds, through the two feed changes that have been measured on real platforms: the switch to time-ordered feeds (Guess and colleagues, 2023) and the down-ranking of hostile partisan posts (Piccardi and colleagues, 2025). For each simulator, the scorecard records whether it reproduces the direction and rough size of each published result: less time spent, more political content, less uncivil content, no change in polarization over three months, and a shift of about 2 points in dislike of the other party.

Why it is still open

Simulators are now used to test proposed fixes before anyone tries them on people. Maik Larooij and Petter Törnberg tested six interventions, from chronological feeds to bridging algorithms, in a simulated platform in August 2025, but their paper does not check the simulator against field results. The checks that exist cover other questions: LLM-SocioPol, presented at NeurIPS 2026, reproduced the ordering of a large voter-mobilization field experiment but with "substantially larger effect sizes", and Aleksandar Tomašević and colleagues compared 30 simulated runs of a Voat technology forum with the real forum. No study was found that checks a simulator against a measured change in feed ranking.

arxiv.orgneurips.ccarxiv.org

Built with what exists

The open code of Larooij and Törnberg (GitHub cssmodels/prosocialinterventions), CAMEL-AI's OASIS simulator and Google DeepMind's Concordia library; open-weight language models such as Mistral Large 4 (see the Note "A trillion open parameters") to drive the agents; the outcome tables published with the two field studies. Every agent and post is synthetic.

  1. First step

    Run the chronological-feed arm of cssmodels/prosocialinterventions with 10 seeds, record the share of political posts each agent sees, and compare its direction with the 2023 Science result.

  2. Done looks like

    Three simulators, each run through both feed changes with at least 10 random seeds per arm, and a public table giving, for every published field outcome, whether the simulator matched its sign and how far its size differed, with code and seeds so anyone can rerun it.

Open 02 · the brief

Test Instagram's topic controls

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A controlled test of Instagram's "Your Algorithm" topic controls, using the cloning design that Northeastern University researchers applied to TikTok. The team creates pairs of fresh test accounts with identical follows and viewing histories; then one account in each pair adds or removes a topic and its twin does nothing. Every day both accounts scroll a fixed number of feed items, and a model labels each item's topic on the spot. Only the topic label and the item's position are kept, never a post, name or handle. The test measures how far the topic's share of the feed moves, how fast, and whether it drifts back.

Why it is still open

Instagram extended "Your Algorithm" to the main feed in June 2026 and has published no measure of its effect. The closest outside test, by Levi Kaplan, Piotr Sapiezynski, Alan Mislove and colleagues at Northeastern (May 2026), covered TikTok's For You feed only. It found that the feed responds to signals, but that the most effective one, "Not Interested", is "unintuitively buried in the interface", and that many test accounts saw their feeds "dominated by such content again" once they stopped using it.

socialmediatoday.comarxiv.org

Built with what exists

Android emulators or test phones driven by Appium or UI Automator; a vision-language model such as Claude Fable 5.2 (see the Note "Three working devices") to tag each item's topic from a screenshot that is deleted once tagged; the Northeastern study's design as a template.

  1. First step

    Set up two fresh accounts with identical follows on an emulator, scroll 200 items on each, and check that their topic mix matches before either touches a control.

  2. Done looks like

    20 account pairs followed for 30 days across five topics, with each topic's daily share of feed items for treated and twin accounts published with 95% confidence intervals, alongside the code and the written protocol.