Lightnews — Scholar-powered news

luna @l4.pm · 4h

whatever i'm a/b testing on those charts is not embedding-based due to various reasons. hopefully i can get around to making one and testing it out soon

luna @l4.pm · 4h

the current post ranking/scoring model is very heavily weighted by authorship and likership, not really content itself. plan is to average out embedding of 100 likes and then find distance from that "positive preference embedding", same for negative (through Show Less on feeds)

1

luna @l4.pm · 4h

i plan to go into more detail on a future blogpost, but the stamina model is an approach to intentionally try to shuffle the posts given by the feed to prevent posts from the same author from tiring the user from reading them, which is more of a workaround to a bad scoring model (bottom chart)

1

luna @l4.pm · 6h

ran multiple a/b tests, ~300 cast votes across multiple different models for nagare

the answer is: everything kinda sucks (v1 on both charts is the original code, what i want to move away from), it's brutal

two charts showing ranking for two models i've been making, top chart is "stamina model a/b test" with the ranking being: v1, v2, v3. the bottom chart is titled "scoring model a/b test" with ranking being: v1e-LR10-SA, v1, v1e-LR100-SA, v1e-L50-R5-S70

1 2

luna @l4.pm · 10h

MUSIC. KETTAMA 🗣️ music!

steelcitydancediscs.bandcamp.com/album/archan...

Archangel [+Free Sticker Pack], by KETTAMA

15 track album

steelcitydancediscs.bandcamp.com

1

luna @l4.pm · 10h

reddit on atproto... theyre making the meme real...

Jay 🦋 @jay.bsky.team · 11h

New reddit alternative built on AT

nooki @nooki.me · 19h

(1/4) I'm soft launching nooki.me to get some early traction and feedback on what to improve. Excited to see how this platform grows!

5

luna @l4.pm · 1d

stop all cryptographic signatures, drink water and touch grass instead of signing

luna @l4.pm · 1d

"i present to the court this signed record that said in the For You Terms Of Service that they can only show pictures of hot women kissing, and as such i show on Slide 13 that For You has not, in fact, shown hot women kissing. i wish for the court to analyze this evidence with due care"

1

luna @l4.pm · 1d

i have an answer to passive interaction feed privacy issue:

put a privacy policy in the feed generator record itself. nothing can go wrong with this!

every change to the privacy policy in realtime, auditable by cid, rendered through markdown

1 1

luna @l4.pm · 1d

YOU on atproto 🫵

6

luna @l4.pm · 1d

the oracledb agents will never catch me alive for i have already created hundreds of sqlite database files in my system while they can only get only 1 running before paying the license fee for the next one

7

luna @l4.pm · 1d

i do not accept any sqlite slander. i can shove 8k concurrent users (~100K a day) in a single webservice powered mainly by 7 sqlite instances in a trench coat. i am simply built different

1 1 20

luna @l4.pm · 1d

Instagram But Atproto. Forums but Atproto. Tiktok on atproto. Its literally git on atproto. It's a cookies on atproto, it's Twitter on Atproto, it's the GFW On atproto. it's signal on atproto. its OTRv3 over atproto . It's literally ice cream on atproto you cant miss it

4 9 60

luna @l4.pm · 1d

"is turing completeness bad" a discussion that will forever haunt us

Reposted by luna

evalka @evalka.bsky.social · 5d

Another essay, this time on what I've learnt from making many worlds for VRChat: their designs impact people's behavior in ways far stronger than I ever expected.

VR is an escape to a better world, here's how world design could help keep it better!

evalkafox.wordpress.com/2025/10/04/c...

Creating for the Human: How World Design Sparks Meaningful Connections in Virtual Reality

Despite a toxic exterior, social spaces in virtual reality are capable of many beautiful things. Through creating worlds for VRChat, I learnt that the design of these spaces has a far stronger impa…

evalkafox.wordpress.com

2 19 48

luna @l4.pm · 2d

if something catastrophic happens and the all of main net atproto is down, cold start in this order for maximum results: plc, relay, dataplane, appview

glad to help!

luna @l4.pm · 2d

went through a wild process to get my homelab postgres out of fedora because it broke and into docker which involved:
- spin up temp docker instance with modified locale
- sed -i 's/en_GB\.UTF-8/C.UTF-8/g' full_backup_american_version.sql
- cold start of my intranet bsky appview stack

1

luna @l4.pm · 2d

i think i can do heavy random sampling from the firehose to bring down the compute/data requirements but idk how small i can make it before the sentiment analysis i'm doing becomes useless

only one way to find out (one more side project)