Inside X’s recommendation algorithm: incentives for crypto accounts

I went through X’s newly open-sourced algorithm instead of relying on the flood of “algorithm hacks” already circulating.

But the code does show/ confirm 100% why scammers, rage farmers, LARPERS and xenophobes can thrive here: X rewards predicted reaction, not truth or integrity.

If outrage, fear or manipulation drives clicks, replies, shares and dwell without triggering enough blocks, mutes or reports, the system can reward it.

Let's start it here:

X open-sourced the architecture, not the full live machine. So anyone claiming they can now give you guaranteed reach hacks is either confused or selling bullshit.

And there is a problem: a lot of people are reading the repository completely wrong.

There is no universal score attached to your post, and the code does NOT prove that “a reply is worth X likes” or that following some fixed engagement recipe will suddenly increase your reach.

X first has to decide whether your post should even become a candidate for a specific user. Phoenix Retrieval builds a representation from that user’s history and retrieves potentially relevant content before the expensive ranking stage even begins.

Only then does Phoenix predict how that individual user is likely to react.

The exposed system predicts actions including likes, replies, reposts, quotes, clicks, profile clicks, photo expansion, video viewing, shares, DM shares, copied-link shares, dwell, follows, not interested, blocks, mutes and reports.

The ranking code then combines those predicted probabilities using configurable weights.

And this is where some of the viral “X algorithm weights” posts fall apart.

The production-style scorer reads values such as "FavoriteWeight", "ReplyWeight", "RetweetWeight", "DwellWeight", "FollowAuthorWeight", "BlockAuthorWeight", "MuteAuthorWeight" and "ReportWeight" from runtime parameters.

They are not permanent numbers hard-coded into the ranking formula.

Yes, X’s public demo currently contains example weights such as:

Favorite: 1.0
Reply: 0.5
Repost: 0.3
Dwell: 0.2

Those numbers are really in the repository.

But they are in the DEMO pipeline.

Presenting them as the current production weights of X without the live configuration is simply wrong.

There is another part I find much more interesting: X explicitly models negative engagement.

"not_interested", "block_author", "mute_author", "report" and "not_dwelled" appear on the negative side of the ranking calculation.

So “all engagement is good engagement” is not what the published architecture actually says.

Rage bait can still work extremely well because anger creates replies, clicks, shares and dwell.

But there is a point where provoking interaction becomes provoking rejection, and the model is explicitly trying to predict both.

The code also confirms author-diversity attenuation. If several posts from the same author compete for the same feed, subsequent posts can receive a decaying multiplier.

Out-of-network content has another configurable multiplier.

Previously seen and previously served posts can be filtered.

Blocked and muted authors can be filtered.

Muted keywords can be filtered.

Duplicates and repost duplicates can be removed.

And even after ranking, content still passes through additional visibility systems.

So when somebody says “the algorithm killed my post”, that statement tells us almost nothing technically.

The post may never have been retrieved, may have failed a filter, may have received poor engagement predictions, may have lost against stronger candidates, may have been affected by author diversity or OON weighting, or may have failed downstream visibility checks.

➡️ There is a major transparency caveat that should not be buried:

X explicitly says the released Phoenix model is smaller than production. The public checkpoint is frozen, while production Phoenix is larger and continuously trained on real-time engagement.

So yes, X has open-sourced an extremely useful map of its recommendation architecture.

No, it has not given us everything required to reproduce the live For You algorithm.

And perhaps the most uncomfortable finding is much simpler tbh.

Phoenix is fundamentally predicting what you are likely to DO, not whether something is true, intelligent, useful or worth knowing.

There is no "P(truth)" or "P(investigative_quality)" output sitting beside likes, replies, clicks and dwell.

That helps explain why someone can spend hours investigating a scam while a KOL posting a casino promotion, manufactured controversy or memecoin fantasy can generate vastly more distribution.

The latter is often engineered specifically to trigger the behavioural signals recommendation systems can observe.

And Crypto Twitter keeps feeding the machine exactly that data.

If you constantly stop for shitcoins, open casino posts, watch pump videos, visit KOL profiles and spend 20 minutes arguing underneath obvious engagement bait, your moral opinion about that content is invisible to the model unless your behaviour expresses it.

You are still generating behavioural data.

Production Phoenix continuously learns from engagement.

Creator/ scammers learn what gets distributed.

Users react to what gets distributed.

Creators produce more of it.

The system learns again.

That feedback loop is considerably more interesting than another bullshit infographic telling you that commenting within 3 minutes gives you 27 algorithm points.

I put the relevant source code below so you can verify it yourself.

Repository:
https://t.co/7URYpy77ML

Phoenix architecture:
https://t.co/Asm5xwu7G6

Retrieval/ranking demo:
https://t.co/xbBOWYTAPG

Ranking scorer:
https://t.co/5tdqUKECEG

Candidate pipeline:
https://t.co/0kXEWkP8pD

Basically, this is evidence that X has built an incentive system where ragebait, panic, hype and manipulation can outperform truth because the ranking model optimises predicted human reaction, not truth or integrity. You are feeding the machine with your behaviour, and the machine learns what keeps you reacting.

One of the worst incentive structures I’ve seen on a major social platform.

The worst maybe.

The architecture also explains why repetitive KOLs have such an obvious advantage here. If you post the same bullish shit every day to the same audience, Phoenix gets an extremely clean behavioural pattern to learn from. Same topic, same crowd, same reactions, same prediction loop. Easy to model, easy to retrieve, easy to keep distributing.

Now compare that with an account covering scams, security, infrastructure, regulation, market structure and whatever else actually matters. The content is broader, the audience response is less uniform and the behavioural pattern becomes harder to predict. X does not need an explicit “punish diverse accounts” switch. The recommendation architecture already creates an incentive to become repetitive, tribal and painfully predictable.

Which is exactly why so many large accounts eventually turn into caricatures of themselves. The system rewards them for saying the same thing forever.