What did the original 2023 algorithm release reveal?
In March 2023, Twitter open-sourced its recommendation algorithm for the first time — a sprawling set of Scala and Python services. It revealed the pipeline shape still recognizable today (candidate sourcing, a heavy-ranker neural model, heuristics and filters, a visibility-filtering rule engine), the existence of author-reputation signals like TweepCred, special handling tied to specific accounts, and a heavy ranker that predicted multiple engagement types and combined them with weights. It was a genuine transparency milestone — and the system it showed was hand-engineered in a way the 2026 release no longer is.
The 2026 xai-org/x-algorithm release is the second time X opened its algorithm. The
first was March 2023, twitter/the-algorithm — and understanding what it showed (and how
the 2026 system differs) is the most useful history for anyone optimizing today.
What it was
The 2023 release was a large collection of services in Scala and Python — the real production recommendation stack, described in X's own engineering blog. It revealed the overall shape:
| 2023 component | what it did |
|---|---|
| Candidate sourcing | In-network and out-of-network candidate generation (the ancestor of today's Thunder/Phoenix split) |
| The Heavy Ranker | A neural model (in the-algorithm-ml) predicting multiple engagement probabilities |
| Heuristics & filters | Author diversity, visibility filtering, fatigue, dedup |
| VisibilityFilters | A centralized rule engine for filtering, labels, and downranking |
In March 2023, Twitter first open-sourced its recommendation algorithm at twitter/the-algorithm: a large Scala/Python stack covering candidate sourcing, a heavy-ranker model, heuristics, and a visibility-filtering rule engine. This is the historical predecessor to the 2026 xai-org/x-algorithm release.
What it actually revealed
Beyond the architecture, the 2023 code surfaced specifics that drove a thousand threads: author reputation via TweepCred, the engagement weights behind ranking, special-case handling referencing particular accounts, and the visibility filtering library. For the first time, the folklore could be checked against source.
The key difference from 2026
The 2023 system was hand-engineered: explicit weights, named heuristics, rule lists you could read and target. The 2026 release replaced much of that with a Grok-based learned model that predicts engagement per post per viewer. The shape rhymes; the mechanism changed from hand-tuned rules to a learned ranker. That's why 2023-era optimization advice is now partly obsolete.
What the code doesn't say
Everything 2023 is now historical, not current. This release is pinned at its
2023 commit and describes a system X has since replaced. We mark these claims CODE-HISTORICAL: true
of the 2023 code, not evidence about how X ranks in 2026. Don't optimize against 2023 specifics.
The 2023 twitter/the-algorithm release describes a system that has been substantially replaced by the 2026 Grok-based release. Claims about 2023 specifics (weights, TweepCred, visibilitylib) are historical and are not evidence about how X ranks content in 2026.
What to do with this
Use 2023 for understanding, not tactics. It explains where today's architecture came from and why certain myths exist — but the current system is the 2026 pipeline, and that's what xDoctor measures against. History informs; it doesn't instruct.