Mridul S
26 posts


many many years ago, i worked on recsys and personalization systems. i finally looked at the xai algorithm and i do not understand it at all. its the normal candidate generation, pruning, reranking but without the core parts like collaborative filtering / matrix factorization and any type of simclustering like functionality seems missing? i didnt spend that much time looking at it but the best i could tell grok does the feature engineering in a black box (which is where all the magic is) and they arent using any sort of persona graph anymore? its very possible (even likely) that i am just way behind the SOTA after all these years but from where im sitting this seems like a bad design (even if the implementation is perfect)
English

@Dastan_2 @santiagomed Please try again. We rolled a fix. Sorry for this.
English

@santiagomed hi! please could you pass this on to the relevant people in your team so that they can fix this bug
English

@natevanben @turbopuffer Thanks. Reranking helps, but recall loss in first-stage retrieval would still cap overall search quality; missing candidates cannot be recovered downstream.
English

@singhai @turbopuffer That said, we often see that larger top_k values are paired with the use of a reranking model to filter the large top_k set down further. As a result, a small recall loss for first pass retrieval is usually not a problem.
English

tpuf ANN v3, for when you need to index the entire web
100B+ vectors @ 50ms p50 / 200ms p99 latency
blog: tpuf.link/ann-v3
English

The Best Sichuan, just off Broadway & E 21st St, NYC

will ye@will__ye
our founding eng was vibecoding during dinner but claude got stuck so our waitress had to help him debug 😭😭😭 only in sf
English

why is @X search broken? it no longer ranks literal string matches. it seems to have a semantic search instead that doesn't encode string literals high enough ?
@nikitabier
English

Have seen this a ton in my feed and it honestly gets tiring to see the same media 27 different times just because everyone is trying to squeeze their monetization juice from good content without adding anything of value.
X should do automatic attribution on media to the original poster then quote posts that are adding value will still happen but those just reposting with no value add won't have any incentive to do it.
Would help ensure original content creators get a fair payout while having a side benefit to declutter the feed.
@nikitabier we need @grok on it

APOCTOZ@Apoctoz
@elonmusk @signulll @nikitabier Fix the relationship between content creators and content aggregators. Currently, content aggregators get 100% of the monetization and credit. Embedding someone else’s work with a slight rephrasing or superfluous comment currently demonetizes them and almost everyone is doing it.
English

@cybertruck Are there plans to get automated alerts in sentry mode @wmorrill3 ?
English

Delighted by what the interns have wrought all on their own.
Saurin Patel@SaurinPatelX
We make digital widgets on American soil using Canadian Labor :)
English

@allegrajacchia How about discussing main characters for various topics on that popular microblogging platform?
English

@ZuckermanRoy חשבונות ממשלתיים (אפורים) לא מקבלים את הפיצ׳ר למיטב הבנתי
עברית

I can’t believe we have to say this, but this account has only ever been run and operated from the United States. Screenshots are easy to forge, videos are easy to manipulate. Thank you for your attention to this matter.
Nikita Bier@nikitabier
@lporiginalg This is fake news. Location was not available on any gray check account at any point. Furthermore, the DHS has only shown IPs from the United States since account creation.
English




















