spacecowboy

They meant this one: bsky.app/profile/spac...

The rewind version is just to see what you have already seen in For You.

Nice! Curious how performant the Lua implementation is and how close the non-inverted algorithm is to For You. Anything like linklonk.com/bluesky?

There are so many redis calls and everything is in a single thread so I imagine it could be quite slow unless the limits at each stage are small.

Thanks for the feedback!

Glad you liked it!

They do have an effect. "show more" has the same meaning as a like.

See how show less works here: bsky.app/profile/spac...

You have been in the experiment group today:

2026-02-28 │ random-sample-on
2026-03-01 │ random-sample-on
2026-03-02 │ random-sample-off
2026-03-03 │ random-sample-on

I'll check later which group you are in today.

Interesting! I see some super weird A/B test metrics - there are some very positive metrics in the slice of users who have 499 or fewer likes. The experiment should have been a no-op for them because sampling 500 out of <500 should return all their likes.

Obligatory @gracekind.net

Good post! Sounds related to bsky.app/profile/sung...

Hm, I'm not sure if it has a threshold because I see some posts with only two likes.

Afaik, Popular with Friends does have a minimum like threshold. And that's why it is mostly empty for me since I don't follow many people.

That's why I didn't want to create a minimum like threshold. It would have been easier to show only posts with >10 likes, but then you would have a cold start problem for items.

For You does surface posts with zero likes because it treats a post to be auto-liked by the author.

There is no model :)

When you create a post For You treats it as if you liked it. Reposts are also treated as a like.

What you "show less like this" in Discover only is visible to Discover. For You only sees what you "show less" in For You.

For You sees all your likes though. Consider liking more posts to improve your For You content.

I don't see any records of "show less like this" for your account.

Either something is not working or you used "show less like this" in a different feed (Discover?)

Could you try open For You and press "show less like this" on any post? bsky.app/profile/spac...

One thing to check: do you by any chance have Portuguese only selected in "Content languages" bsky.app/settings/lan... ?

For You has no concept of post language. Looking at your recommendations it is indeed Portuguese heavy: linklonk.com/bluesky?user...

Maybe the Portuguese community is more tightly knit and that makes you establish stronger connections (more likes in common) compared to the English segment.

Yes, I'd suggest "show less" to tune it down. It simply weakens your connection to people who liked that post so you would see less from them at the top of For You.

How are you flagging as "not interested"? Do you use "show less like this" in For You? Or are you muting/blocking the post authors?

You have only 6 likes - including on political posts. This is why you see political posts in For You.

You can see for yourself here: linklonk.com/bluesky?user...

The key to improve your For You feed is to like some more non-political posts.

The "lande" link gives 404

It is an option.

Discover relies on embeddings and clustering. It is tricky. It's hard to tell the meaning of a post by the short text or the image content. Especially with vagueposting.

I'd like to keep things simple.

How old is super old? 4 hours or 2 days?

The purpose of this test is to reduce the recency bias as people have been reporting that it is too reactive.

Yes, weighting older likes could be a good ideas to try as well. I was thinking somehow measuring how much more weight recent likes end up having and reducing that weight somewhat.

That's an option. Or change the most recent 5000 to something else.

Thanks! It did show a voltage increase 1.1V -> 1.4V in the changes. The BIOS is the latest version - updated back in Jan when installed the new CPU.

Your are in the test group today. Tomorrow you will be in the control.

Thanks for the qualitative feedback!

The initial results are not looking too hot: