How has the X algorithm evolved from 2023 to 2026?

The biggest shift is from a hand-engineered system to a learned one. The 2023 stack combined many bespoke services, a heavy-ranker neural model with published engagement weights, named heuristics like TweepCred for author reputation, and a rule-engine visibility filter. The 2026 release rebuilds ranking around Grok — a single transformer-based model family handles retrieval (Phoenix) and content understanding (Grox), predicting engagement contextually instead of applying a fixed weight table. The pipeline shape persists (source, rank, filter), but the brains changed from rules-and-weights to a learned model. That's why old optimization playbooks partly stopped working.

Reading the 2023 and 2026 releases side by side, the evolution is clear and consequential. The architecture rhymes, but the core intelligence was replaced — and that's the single most important thing for understanding why growth tactics shifted.

The throughline: shape persists

Both eras share the fundamental pipeline: gather candidates (in-network and out-of-network), rank them with a model, apply filters and heuristics, blend and serve. If you learned the 2023 pipeline, the 2026 one is recognizable.

CODE-HISTORICALc54bec0verified 2026-06-12
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.
twitter/the-algorithm — repository root README (2023 release)HISTORICAL: describes the 2023 system X has since replaced

What changed: rules-and-weights → learned model

dimension20232026
Ranking brainHeavy Ranker neural net + published engagement weightsGrok-based learned ranker predicting engagement contextually
RetrievalBespoke in/out-network servicesThunder (in-network) + Phoenix (Grok retrieval)
Content understandingFeature engineering, heuristicsGrox classifiers (Grok reading text, images, video)
Author reputationTweepCred and related heuristicsLearned signals; no published TweepCred equivalent
Safety/visibilityvisibilitylib rule enginePTOS classifier (two-stage Grok safety taxonomy)
WeightsPublished, fixed, gameableWithheld, applied to learned predictions
OFFICIAL-STATEDverified 2026-06-12
In January 2026, X Engineering open-sourced its production For You feed algorithm at xai-org/x-algorithm, built on the same transformer architecture as xAI's Grok model.
X Engineering (@Engineering) — announcement post, January 2026X's own characterization of the release as the production system

Why TweepCred fading matters

The 2023 system had explicit author-reputation heuristics — TweepCred, a PageRank-style score. The 2026 release doesn't publish an equivalent named score; reputation effects appear to be absorbed into learned signals rather than a single readable number. This is emblematic of the whole shift: named, inspectable heuristics gave way to a model whose judgments are harder to reduce to one constant.

CODE-HISTORICALc54bec0verified 2026-06-12
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.
twitter/the-algorithm — comparison of the 2023 release against the 2026 xai-org/x-algorithm releaseHISTORICAL framing claim; 2023 facts do not transfer to 2026

What it means for optimizing

The 2023 playbook — target the highest-weighted action, manage your TweepCred, avoid specific filter triggers — assumed a system you could reverse-engineer from constants. The 2026 learned ranker doesn't offer that surface. The durable move is to optimize for genuine engagement signals (which both eras reward) rather than era-specific tricks, because the learned model sees through the tricks.

What the code doesn't say

▲ What the code doesn't say

The exact migration path and what carried over internally. We can compare two open releases at two points in time; we can't see the years of change between them, or which 2023 components secretly persist. This is a before/after comparison of public snapshots, not a complete changelog.

CODE-HISTORICALc54bec0verified 2026-06-12
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.
twitter/the-algorithm — comparison of the 2023 release against the 2026 xai-org/x-algorithm releaseHISTORICAL framing claim; 2023 facts do not transfer to 2026

What to do with this

Anchor on the present system — the 2026 pipeline — and treat 2023 as the explanation for why the folklore exists. xDoctor measures against the current learned system, which is the only one your posts actually face today.

← The historical record