Nguyen Phu Huy

[Nguyen Phu Huy @huynp85]

@JamesEbringer tap

[Nguyen Phu Huy @huynp85]

@X_FINALBOSS X

[Nguyen Phu Huy @huynp85]

@affprinter PRINT

[Nguyen Phu Huy @huynp85]

@X_FINALBOSS X

[Nguyen Phu Huy @huynp85]

@TommiPedruzzi Send

[Nguyen Phu Huy @huynp85]

@X_FINALBOSS X

[Nguyen Phu Huy @huynp85]

@X_FINALBOSS Miro

[Nguyen Phu Huy @huynp85]

@X_FINALBOSS ALGO

[ALEX SUZUKI @X_FINALBOSS]

I created 5000 AI bots to completely decode the X algorithm.

Here's what I found :

The X algo has 4 main components:

• Home Mixer (orchestrates everything)
• Thunder (stores posts from people you follow)
• Phoenix (the AI that scores and ranks everything)
• Candidate Pipeline (the framework connecting it all)

Here's how they work together:

Every time someone opens X, the algo builds a pool of ~1,500 posts to even consider showing them.

500 million posts created daily.

Narrowed to 1,500 per user in under 200ms.

If your post doesn't make the candidate pool, it gets a score of zero.

Nobody sees it. Ever.

Your follower count does absolutely nothing.

Buried in the codebase:

author_followers_count is pulled for display only.

Not fed into any scorer. Anywhere.

100K followers = 1K followers in Phoenix's eyes.

What travels with your post is a PREDICTED engagement score based on your recent history.

This is the most important thing I found.

It's called the Prediction Trap.

Before your post reaches a single person, Phoenix scores it across 19 signals and assigns reach based on that prediction.

Phoenix predicts low engagement
→ shows it to fewer people
→ fewer people = fewer chances for engagement
→ prediction validates itself
→ post gets buried

Phoenix predicts high engagement
→ bigger test group
→ more engagement
→ prediction validates itself
→ post gets pushed further

The algorithm creates the outcome it predicted.

The 19 signals Phoenix predicts (ranked by importance):

1. Reply rate — 10/10 (hardest to fake, weighted heaviest)
2. Dwell time — 9/10 (how long they pause, two separate signals)
3. Profile clicks — 8/10 (they want to know more about you)
4. Bookmarks — 8/10 (saved = genuinely valuable)
5. Likes — 3/10 (cheapest signal, barely moves the needle)

One block = -3.0
One like = +0.5

A single block wipes out 6 likes worth of positive signal.

The scoring formula:

Final Score = Σ (weight × probability of each action)

Positive actions → positive weights
Negative actions → negative weights

Nobody outside X knows the actual weights.

But the architecture is confirmed from the public codebase.

Your recent posting history sets the baseline prediction for every future post.

This is the part that destroys most accounts.

If your last 10 posts averaged 0.3% reply rate, Phoenix predicts your next post will also perform poorly and distributes accordingly.

Bad posts don't just die.

They make your next post die too.

The AuthorDiversityScorer applies exponential decay every time you appear in the same follower's feed session.

Post 1 in their session → full score
Post 2 → decayed
Post 3 → decayed further

More posts ≠ more reach

The sweet spot is 2-3 posts per day with 4-6 hour gaps.

Posts die after 48 hours. Hard stop.

Thunder (X's in-memory post store) auto-trims every 2 minutes.

Retention window: 48 hours.

After that your post is gone from the candidate pool entirely.

The idea that consistent posting lets older content keep circulating is wrong at the architecture level.

You start from zero every 48 hours.

Reposts of other people's work = 90% impression deduction on that post.

Self-reposts = completely fine.

The Bloom filter resets each session.

A follower who hasn't opened X since this morning hasn't seen your earlier post in their current session.

Self-repost reaches them fresh.

Out-of-network reach is a mathematical process, not luck.

Phoenix uses a Two-Tower Model:

Tower 1 → encodes your engagement history into a taste fingerprint
Tower 2 → encodes every post into a content fingerprint

Phoenix then runs similarity search across millions of users.

If your post's fingerprint matches someone's taste fingerprint, it enters their candidate pool even if they don't follow you.

This is the viral mechanism.

Niche consistency compounds this.

An account that posts in one lane for 6 months builds a tight content vector Phoenix can reliably match to the right audiences.

What to do:

→ Engineer reply triggers into every post
→ Post scene-setting stories with a question at the end
→ Space posts 4-6 hours apart
→ Self-repost your best content 12-24 hours later to hit new sessions
→ Recycle proven hooks — 48hr reset means new audiences each time
→ Track reply rate, new follows, bookmark rate per post — nothing else matters

What to stop immediately:

→ Posting standalone links (near-zero reply rate poisons your Phoenix baseline)
→ Public "sent to your DMs" confirmation replies (each one is a dead post in your history)
→ Telegram/channel promotion posts as standalone feed posts
→ Posting vanity metrics with no story or question attached
→ Optimizing for impressions (vanity) or likes (weakest signal)
→ Posting 5+ times per day (you're decaying your own posts)

The platform changed the rules completely.

The habits that built audiences under the old algo are actively working against you under the new one.

Consistency without reply-rate optimization is just feeding Phoenix bad data points.

One post with 1.5% reply rate does more for your reach than twenty posts with 0.1%.

Understand the system.

Then work with it.

This is how I consistently do millions of views across all X accounts daily. Not only views but more importantly , target high quality US audience so I can easily do $2K-15K sale with digital products everyday. (quality of views is more important than # of views)

Want the full breakdown of every component including the GitHub repo secretly released by xAI?

Reply "ALGO" and I'll send it to you.

**must be following + retweet to receive it.

[Nguyen Phu Huy @huynp85]

@X_FINALBOSS guide

[Nguyen Phu Huy @huynp85]

@X_FINALBOSS System

[Nguyen Phu Huy @huynp85]

@Manu_Sisti Send

[Nguyen Phu Huy @huynp85]

@shalevhvs Prompt

[Shalev @shalevhvs]

https://t.co/QLSL1Ee1RF

[Nguyen Phu Huy @huynp85]

@X_FINALBOSS X

[Nguyen Phu Huy @huynp85]

@spwfeijen AD

[Nguyen Phu Huy @huynp85]

@X_FINALBOSS Please send it

[Nguyen Phu Huy @huynp85]

@Manu_Sisti AI

[Nguyen Phu Huy @huynp85]

@starks_arq ARQ

[Amir D @starks_arq]

i built an agent that scrapes viral tweets and generates hundreds of ultra-realistic videos for each tweet...

it spent $16k in the past 30 days, but generated over $100k for my business.

Reply “ARQ” + RT and i'll send you...