Anton

Building AI agents that find your next customers.

[Anton @AntonTamozo]

@ChrisWillx With all respect to hospitality workers, ofc.

[Anton @AntonTamozo]

@ChrisWillx Status, courage, competence and purpose are worth more than income alone.

[Anton @AntonTamozo]

@StepBogoutdinov True, nothing comes without effort.

[Anton @AntonTamozo]

Another X algorithm myth-busting thread.

"1 reply = 27 likes"

Not true. Weights apply to predicted probabilities, not raw engagement counts.

"1 report cancels 468 likes"

Misleading. Reports are far rarer than likes, so they get a larger weight.

"Premium = 4× reach"

No Premium multiplier appears in the published ranking weights.

"Hashtags boost reach"

No hashtag boost appears in the ranking code.

❗️Where the code is silent, call it a hypothesis.

[Anton @AntonTamozo]

@StepBogoutdinov Yes, all findings come from the official Github repository.

[Anton @AntonTamozo]

@elonmusk No wonder. The best one on this planet.

[Anton @AntonTamozo]

@elonmusk Is there enough energy supply in the US for this?

[Anton @AntonTamozo]

7 / The common thread

None of these labels delete your post or notify you. They just switch off part of your audience.

The only way to see them: track the share of impressions from non-followers, not total numbers.

[Anton @AntonTamozo]

Did you know your post has to pass a separate visibility layer to reach For You?

Ranking decides how much reach your post gets.

Visibility filters decide whether it gets shown at all.

So you can have a great post with a strong ranking score and still get buried by a visibility rule.

Most people optimize for ranking and ignore the filters.

Check👇 for the filters X uses and how to avoid triggering them.

[Anton @AntonTamozo]

6 / COMPROMISED

CompromisedUserLabelRule: system decides the owner no longer controls the account.

Example: password leaks, 200 "token giveaway" posts go out at 4am. You clean up by morning. The label stays. Reach is dead until it's lifted.

Before: 2FA via app, not SMS. After: rotate password, revoke all sessions, then contact support.

[Anton @AntonTamozo]

5 / IMPERSONATION

ImpersonationHighPrecisionUserLabelRule. "High precision" = fires rarely, but when it does it's almost always right. Hard to appeal.

The risk isn't you copying someone. It's someone copying your brand and getting verified first — then YOU look like the clone.

Verify early. (Inference, not code.)

[Anton @AntonTamozo]

4 / SPAM (account level)

SpamHighRecallUserLabelRule — same signal, but the label goes on YOU, not the post.

Post label kills one post. Account label cuts everything you publish until it's removed.

How you get there: a week of template threads. One is a post. A pattern is an account.

[Anton @AntonTamozo]

3 / SPAM (post level)

SpamTweetLabelRule / SpamHighRecallDropRule: post read as spam → dropped.

"High recall" = tuned to catch as much spam as possible, false positives accepted. Honest accounts get hit.

Example: 40 replies of "Great post! More in my bio" in an hour. Each one is a candidate.

[Anton @AntonTamozo]

2 / MALICIOUS_URL

MaliciousUrlOonDropRule: a link flagged as suspicious drops the post for non-followers. Text quality doesn't matter.

Example: https://t.co/Z0Biq7rN3C → tracker → landing page. Every hop is a redirect.

Fix: direct link, clean root domain, no shorteners.

[Anton @AntonTamozo]

1 / DO_NOT_AMPLIFY

X's code has DoNotAmplifyOonDropRule. Your post isn't deleted — it's hidden from everyone who doesn't follow you.

Followers see it. Stats look alive. New audience: zero.

Symptom: follower engagement normal, non-follower impressions ~0. That's a label, not "the algorithm changed."

[Anton @AntonTamozo]

@AlexHormozi Well spotted. You can work hard on 10 things and still get nowhere. Pick one thing that works and go all in on it.

[Anton @AntonTamozo]

@ycombinator Thank for sharing ♥️

What was the real bottleneck to making AI agents reliable in banking?

What did you have to change in the product to finally solve it?

[Anton @AntonTamozo]

@AlexHormozi The hardest part isn’t knowing what has the highest return. It’s having the courage to ignore everything else. Most people don’t lack opportunity, they lack conviction.

[Anton @AntonTamozo]

@ChrisWillx Crazy how seriously we take things that really don’t matter in the grand scheme of things.

[Anton @AntonTamozo]

It's "Report post" - spam, abuse, misleading, etc. In the code it's one of five negative actions the model predicts, weighted −234 vs +0.5 for a like. The code comment says why: a report is >1000× rarer than a like, so it needs a big weight just to register. That's a scale correction, not "1 report cancels 468 likes."

[Anton @AntonTamozo]

@ChrisWillx The paradox is that the self you’re trying to improve can become the very thing preventing you from being at peace.