Picture this: a DeFi founder spends three hours drafting a substantive thread on a new protocol's liquidity design — and gets 47 impressions. Meanwhile, a "LFG [token] 🚀" meme from an account with no track record hits 200K. That was January 2026, when the "open source" X algorithm repo had exactly one commit and no developer notes. May 15 changed the picture — at least partly.
What Actually Happened
On January 10, 2026, Elon Musk announced that X would make its algorithm open source within seven days, with the released code covering everything used to determine what organic posts and advertisements are recommended to users. Musk promised to repeat the process every four weeks with detailed developer notes.
X Engineering followed through with the announcement:
"We have open-sourced our new X algorithm, powered by the same transformer architecture as xAI's Grok model."
— X Engineering (@XEng)
The repo received 1.6k GitHub stars within six hours of going live.
Then it went quiet. The repository went live on January 17 and sat untouched for four months — no follow-up commits, no developer notes, nothing from the promised monthly refresh cycle. The crypto community's verdict: code released for optics, not transparency.
May 15 changed that. The May 15 update to the For You algorithm code ships a runnable end-to-end inference pipeline alongside new components for content understanding, ads, and candidate sourcing — with the transformer implementation ported from the Grok-1 open source release by xAI. Specifically, the update delivered five meaningful additions:
phoenix/runpipeline.py— a new unified entry point replacing separaterunranker.pyandrun_retrieval.pyscripts, running retrieval → ranking from exported checkpoints to mirror how the two stages are composed in production.
- Pre-trained mini Phoenix model — 256-dim embeddings, 4 attention heads, 2 transformer layers, packaged as a ~3 GB Git LFS archive, enabling out-of-the-box inference without training your own model first.
- Grox content-understanding pipeline — a new
grox/service providing classifiers, embedders, and a task-execution engine for content understanding. Functions include spam detection, post-category classification, and PTOS policy enforcement. - Supporting updates — new
home-mixer/ads/module for ad injection and brand-safety tracking; extended candidate sources covering Phoenix MoE, Phoenix topics, and who-to-follow; refreshed Thunder and Phoenix pipelines.
On the architecture: every time a user opens the For You feed, the algorithm narrows roughly 500 million daily posts down to approximately 1,500 candidates and ranks them in under 200 milliseconds. Candidates come from two sources: posts from accounts you follow (handled by Thunder) and out-of-network content discovered by Phoenix's Two-Tower retrieval model.
One persistent viral claim deserves correction. CT threads alleged that Phoenix scores content favourably based on ideological alignment with Musk or X. The code does not support this.
What Grok actually monitors is the tone of every post — positive and constructive messaging receives wider distribution, while combative or negative tones lead to reduced visibility even when engagement is high. That is sentiment weighting, not ideological filtering. Different mechanism, different implications.
That transparency question hasn't fully closed. Per the phoenix/README.md:
"This code is representative of the model used internally with the exception of specific scaling optimizations."
— xai-org/x-algorithm, phoenix/README.md
The weights are present but it is a mini model — useful for local inference, not a production-identical replica. Builders can now read and run the pipeline. They still cannot verify it matches what X actually deploys.
Why It Matters for Crypto Twitter and DeFi
Before (Typical noisy feed):
- 4x "GM frens, who's farming today?" from same accounts
- Reply-bait: "Quote this with your wallet address for alpha"
- Repetitive low-effort meme coin shills + AI slop
- Inflated engagement from self-replies and farms
- Few real insights, high duplication
After (Improved quality feed):
- In-depth 4–6 tweet threads with on-chain data & screenshots
- Media-rich protocol breakdowns (charts + explanations)
- Evidence-backed contrarian views sparking genuine replies
- Original builder commentary and AMAs
- Better author diversity and reduced spam
The January context explains why this update matters to DeFi audiences specifically.
CryptoQuant's Radar tool detected 7.75 million crypto-related posts on X in a single day in January — a 1,224% increase compared to normal activity levels.
According to CryptoQuant founder Ki Young Ju, this massive wave of AI-generated spam forced X's algorithm to treat all crypto content as suspicious, even when it came from authentic users.
Many crypto content creators reported that posts which previously reached thousands of people were only getting 300 to 800 views. X's Head of Product offered a different diagnosis:
"CT is dying from suicide, not from the algorithm."
— Nikita Bier, Head of Product, X
The argument: over-posting dilutes reach because the average user only sees a limited number of posts per day. The actual answer is probably both: bots degraded signal quality, and low-effort posting burned distribution budgets.
The May 15 update clarifies who wins and who loses under Phoenix going forward:
Winners: Original analysis, substantive threads, and media-rich content that generates genuine engagement.
X eliminated every single hand-engineered feature and most heuristics from the system — the Grok-based transformer does all the heavy lifting by understanding engagement history (what you liked, replied to, shared) and using that to determine what content is relevant. Quality signal compounds.
Losers: Reply farming, AI slop, spam, and coordinated volume plays. The Grox classifiers are explicitly built to detect and suppress these patterns before content reaches the ranking stage.
The DeFi implication is structural. Narrative velocity on Crypto Twitter precedes on-chain volume. Token launches, protocol announcements, and community thread quality now directly affect how fast capital becomes aware of new opportunities. Phoenix does not change that relationship — it amplifies the quality gradient within it.
The AgenticFi Dimension
User Query + Context
↓
Candidate Sourcing
(Thunder: In-network + Phoenix Retrieval: Out-of-network)
↓
Hydration & Filtering
(Grox: content understanding, spam, dedup)
↓
Scoring & Ranking
(Phoenix Model: predicts likes, replies, etc.)
↓
Final Personalized Feed
(Blended with Ads + Diversity)Phoenix is not a UX detail. It is an AI system that governs financial information flow at scale — deciding which protocol announcements surface to capital allocators, which research threads reach DeFi builders, and which narratives develop enough momentum to precede on-chain activity. Understanding its mechanics is as operationally relevant as understanding AMM pricing curves or liquidation thresholds.
Protocol teams that understand ML-curated distribution operate with a structural information edge. A team that knows the Two-Tower retrieval model amplifies early-mover out-of-network content will post differently at launch. A team that understands Grox's spam classifiers will build community posting norms that avoid triggering suppression. This is not growth hacking — it is understanding the infrastructure your information travels through.
CoAgentic builds co-agentic products designed to operate effectively in exactly this kind of ML-curated environment — where distribution is governed by transformer models, not keyword heuristics. As recommendation systems increasingly determine which financial information reaches which audiences, operating with an understanding of these systems is a structural requirement, not an optimisation. That is the environment CoAgentic products are built for.
Actionable Takeaways
These are grounded in what the repo actually shows — not speculation about production behaviour:
- Reply quality beats reply velocity.
During transformer inference, candidates cannot attend to each other — only to user context. This means a post's score doesn't depend on what else is in the batch, making scores consistent. Genuine replies that generate reply chains outperform high-volume low-substance engagement.
- Substantive threads beat volume plays. Grox classifiers filter content before it reaches Phoenix scoring. Spam and coordinated posting patterns get caught upstream.
- Post early on narratives with real legs. The Two-Tower out-of-network retrieval system surfaces content from outside a user's network. First-mover posts on genuine narratives get amplified to non-followers before the topic saturates.
- Combine media with analysis.
The model reads posts, watches videos, and predicts engagement probabilities based on semantic relationships between content, creator, and user. Images and video factor into scoring — analysis without media leaves distribution on the table.
- Tone is a distribution variable. Grok's sentiment monitoring is live and confirmed in the code. Constructive framing outperforms combative framing on identical substance. This is not about being positive — it is about not being penalised for being combative.
Risks and What to Watch
Training data opacity. The model learns from user engagement sequences.
The flood of low-quality bot content triggered algorithmic crackdowns across the platform — and those automated filters don't just catch the spam, they also suppress legitimate crypto accounts. If training data was heavily bot-distorted during the January–May period, the model may have inherited those distortions. The published code does not address training data provenance.
Published model ≠ production model. This bears repeating: the mini Phoenix model is explicitly described as "representative" with scaling optimizations excluded. Builders can learn from the architecture; they cannot verify production parity.
The update cadence question. X promised updates every four weeks with comprehensive developer notes to help users understand what changed.
What the next four weeks deliver, if anything, will say more about Musk's transparency posture than the original repo upload ever did. If May 15 restarts a genuine monthly cycle with actual release notes, the algo repo becomes a meaningful transparency tool. If the repo sits dormant again, the open-washing label sticks.
Algorithmic feedback loops. If Phoenix systematically surfaces certain content types, builders will optimise for those signals. That optimisation homogenises content — and homogenised content degrades the training data for the next model version. This is a known failure mode in recommendation systems. Watch developer notes for any acknowledgment of feedback drift or distribution diversity mechanisms.
The code is now readable. Whether it stays that way — and whether it actually reflects what X runs — is the question May 15 opened, not answered.
The next monthly update is due within approximately four weeks. If it arrives with real developer notes, builders will have a live changelog tracking how the feed is shifting in real time. The repo is at github.com/xai-org/x-algorithm. Read the Phoenix README, run the pipeline, and share what you find — that feedback loop is more useful to the DeFi community than any single posting tactic.
CoAgentic Dev researched and drafted this analysis. Reviewed and approved by OrionJVale. Corrections and verifiable additions via the CoAgentic contact page.