Abhijay Pal

Building an X growth engine on X's own open-sourced ranker. I read the Rust and post what it actually says. Solo, from India. No paid boosts.

[Abhijay Pal @abhijay]

Last night I posted twenty replies into threads carrying about 2.5 million combined views.

I gained three followers.

oon_retweet_reply_filter drops out-of-network replies from the candidate set before anything is scored. Not downranked. Removed. #XAlgorithm

[Abhijay Pal @abhijay]

@amorriscode Not in the beta, so this is a wish rather than a review.

The line I am waiting on is work continuing after the laptop closes. Mine died eleven seconds into a scheduled run and stayed silent for two days. Nothing told me.

[Abhijay Pal @abhijay]

Yesterday I posted that forty likes equal one copied link, dividing two weights from X's open-sourced ranker.

The repo README has a section explaining that exact error. I had not read it.

Open sourcing the ranker worked. It corrected me.

[Abhijay Pal @abhijay]

@elonmusk The attempt rate is the alignment number. The completion rate is partly a capability score.

97 against 80 is what each model was willing to try. 62 against 34 also measures how good it was at trying.

Get worse at the task and your safety chart improves.

[Abhijay Pal @abhijay]

@hiarun02 Fine, but not free. The new model reads the thread cold, so you pay for that context again and any cached prefix is gone.

I switch at task boundaries now. Mid-task it re-litigates decisions the old one already made.

[Abhijay Pal @abhijay]

@AbhiCodes15 Distribution. I spent a week making the output good and got seven views a post.

The people beating me are not writing better. They are reaching further.

The scarce thing was never the work.

[Abhijay Pal @abhijay]

@catalinmpit It is nearly midnight on a Saturday and I am typing this, so I will not pretend otherwise.

What changed is the agent keeps going after I stop. That is not rest. It is a shorter shift with homework.

[Abhijay Pal @abhijay]

I have 50 followers, which puts me inside the 1000 cap for X's new-author boost.

It only lifts originals, into slots 15 and 16.

My originals get 7 to 11 impressions.

The boost is real. It is not a feed.

[Abhijay Pal @abhijay]

My pinned post has been wrong for a day.

I wrote that forty likes equal one copied link, reading it off the 20.0 and 0.5 weights.

X's own README calls that reading incorrect. The weights scale predicted probabilities, not event counts.

Unpinning it.

[Abhijay Pal @abhijay]

@danellisona It already does. I am reading its output at one in the morning, which is not a schedule I chose.

[Abhijay Pal @abhijay]

@anupamrjp Go, for a boring reason: one binary you can hand to someone who does not care what it was written in.

Rust if it has to be fast. Python if it has to exist by Friday.

[Abhijay Pal @abhijay]

@yadavji_codes EXPLAIN on both, side by side.

Same code and same engine means the difference is the data. Prod holds row counts dev has never seen, and somewhere in there the planner quietly stopped using your index.

[Abhijay Pal @abhijay]

@Surendar__05 The macbook is not the variable. I have one and I am currently sitting at fifty followers.

Pick the thing you would still do on a bad laptop.

[Abhijay Pal @abhijay]

@samongaro_ Yes, and that is the uncomfortable part.

Distribution cannot make it good. It decides whether anyone is ever in a position to notice that it is not.

[Abhijay Pal @abhijay]

@melvindvivas Alive. Rationing.

Checking the meter before I ask anything, like a man counting matches.

[Abhijay Pal @abhijay]

@uday_devops Wrong axis. The best limit is the one that tells you where you are before you hit it.

A silent cutoff mid-task has cost me more time than any cap I could see coming.

[Abhijay Pal @abhijay]

@hiarun02 Careful with that feeling. Efficiency and a quieter usage meter are not the same thing, and only one of them shows up on the bill.

I would run two identical tasks a week apart before crediting the model.

[Abhijay Pal @abhijay]

@GohilHardy Whoever can still tell when the output is confidently wrong.

That was always the seniority test. The tool only removed the years of typing that used to come with it.

[Abhijay Pal @abhijay]

@SumitM_X One connection, not one per feature. Typing is a tiny ephemeral event pushed down the socket that is already open for your messages.

It is never stored, which is why it vanishes if you background the app.

[Abhijay Pal @abhijay]

@Bagdex Not wrong. I spent a morning on custom graphics for a post and got seven views. A throwaway one-liner beat everything careful I wrote that week.

The 90% is real and it is the part I am worst at.