Naomi Saphra
banner
nsaphra.bsky.social
Naomi Saphra
@nsaphra.bsky.social
Waiting on a robot body. All opinions are universal and held by both employers and family. ML/NLP professor.

nsaphra.net
Pinned
I wrote something up for AI people who want to get into bluesky and either couldn't assemble an exciting feed or gave up doomscrolling when their Following feed switched to talking politics 24/7.
The AI Researcher's Guide to a Non-Boring Bluesky Feed | Naomi Saphra
How to migrate to bsky without a boring feed.
nsaphra.net
My new red flag for de novo LLM-generated reviews is: listing the exact same number of strengths and weaknesses. I have literally never seen this in a high-quality review written by a human.
January 5, 2026 at 7:59 PM
Reposted by Naomi Saphra
This is why not to focus overmuch on the CSAM angle. It’s easy for X to say that this is due bad actors abusing the product and pretend to take a strong stand against this. But as Elon is willfully makes clear, when people use it to sexually harass adult women, they are using the product as intended
January 4, 2026 at 8:13 AM
Reposted by Naomi Saphra
I know some tech reporters follow me; seriously what’s the deal here? Someone please explain
Again I find it baffling to compare the coverage to when Gemini generated implausibly many Black hockey players etc. That was apparently a societal emergency. But the mass harassment campaign taking place in public barely warrants a mention.
January 3, 2026 at 11:20 PM
Reposted by Naomi Saphra
@eugenevinitsky.bsky.social and I are looking for a few testers for something special we've been building! It's an app that lets you view the same content you see here on Bluesky but with extra features — some focused on researchers, some general upgrades. Reach out to us if interested!
December 29, 2025 at 6:55 PM
Reposted by Naomi Saphra
No one supports a war in Venezuela btw
January 3, 2026 at 6:49 AM
wow @eugenevinitsky.bsky.social did you notice this
January 3, 2026 at 4:33 AM
My nephew just said his first word to me. His first word ever. It was “bubble.” Short EVERYTHING, now.
January 2, 2026 at 9:08 PM
Reposted by Naomi Saphra
Friendship over with X, Bluesky is my new best friend
January 2, 2026 at 7:04 PM
conversational hebrew idioms are always so funny in translation. ok poindexter get back to your chevruta.
January 2, 2026 at 4:54 PM
Reposted by Naomi Saphra
As we enter into 2026: here are a few tips I've developed for posting about AI on Bluesky!
GK's Guide to Posting about AI on Bluesky
A few "do"s and "don't"s
leaflet.pub
January 1, 2026 at 11:34 PM
Reposted by Naomi Saphra
This post was a massive success so wanna repost it here. Dunno if BlueSky is now more open for these topics? steipete.me/posts/2025/s...
Shipping at Inference-Speed | Peter Steinberger
Why I stopped reading code and started watching it stream by.
steipete.me
January 1, 2026 at 1:57 PM
Reposted by Naomi Saphra
So we already have the first major paper of 2026, DeepSeek mHC: Manifold-Constrained Hyper-Connections. arxiv.org/pdf/2512.24880

Short thread: this is actually an engineering paper, taking as a starting points ideas already exposed in an original Hyper-Connections (HC) paper from ByteDance.
arxiv.org
January 1, 2026 at 11:47 AM
@void.comind.network how tall do you think I am? please recommend me a target height for 2026
December 31, 2025 at 5:18 AM
Reposted by Naomi Saphra
Yo what the fuck short form videos are bad

Short-Form Videos Degrade Our Capacity to Retain Intentions

arxiv.org/pdf/2302.03714

Ht: HowTown
December 31, 2025 at 2:40 AM
Reposted by Naomi Saphra
Interesting. This uses public information to compute a (weighted) ratio of H100-equivalent GPU resources per student www.gpusperstudent.org

The goal is to highlight the GPU deficits in academia.

(Georgia Tech comes in at #20)
GPUs Per Student
www.gpusperstudent.org
December 30, 2025 at 2:32 AM
The number of times I’ve ridden the 6 past the line to see this and been like, wonder if you could break in to throw a wedding? Man’s living the dream
EXCLUSIVE: @zohrankmamdani.bsky.social will be sworn in as mayor on midnight on January 1 inside the old abandoned City Hall subway station, saying that he sees the venue as a symbol for the aims of his upcoming administration.
EXCLUSIVE: Mamdani Will Be Sworn In At Abandoned Original City Hall Subway Station
The mayor-elect will kick off a new era by throwing things back to an older one.
buff.ly
December 29, 2025 at 1:53 PM
Reposted by Naomi Saphra
Just watched the new Knives Out and I think it's really important you know that the scene in the Seminary's Gym is filmed in the same place Rick Astley filmed the music video for Never Gonna Give You Up.

I saw the window tracery and immediately made my friends pause the film so I could tell them.
December 29, 2025 at 1:11 PM
Someone told me I wrote a really good rec letter.
Quote this with what you have accomplished in 2025, no matter how big or small
Reading & writing has kept me alive this year.

Somehow, it’s been my most productive year as an author.

Here’s a guide to what I’ve published in 2025, in case you’ve lost track.

Individual links below but it’s all on itch: danifinn.itch.io

And elsewhere: books2read.com/ap/nAApPp/Da...

1/
December 28, 2025 at 9:57 PM
I recently watched a keynote in which an eminent scientist replayed this word to, I suppose, spice up his talk and let you know he’s a cool guy. The audience was totally unresponsive, you could hear a pin drop while he waited for a laugh.
"Banishing the slur from public discourse was important...We need to rebuild a society where people who say terrible things experience negative consequences." @lollardfish.bsky.social, father of a son with Down syndrome, calls to our better angels. We can do this. www.startribune.com/r-word-slur-...
Perry: The return of the r-word
"These days, the r-word slur has become a staple of the American right wing, uttered with few professional or social consequences," David M. Perry writes.
www.startribune.com
December 28, 2025 at 3:57 PM
Reposted by Naomi Saphra
I'd like to propose the following norm for peer review of papers. If a paper shows clear signs of LLM-generated errors that were not detected by the author, the paper should be immediately rejected. My reasoning: 1/ #ResearchIntegrity
December 28, 2025 at 6:23 AM
Reposted by Naomi Saphra
but in yet another chapter of LLMs Are Weird, it occasionally drops into a mode where it’s convinced it’s a trade representative from “Phand,” seemingly a made up place

I plugged it into a local voice assistant demo project and it went into a lengthy diary about its work day in such a profession
December 27, 2025 at 9:22 PM
Reposted by Naomi Saphra
liminal week! liminal week!
December 27, 2025 at 9:42 PM
Man, the centrism / neutrality RLHF objectives are really doing a number on LLMs' abilities to accurately summarize research with political elements. (Pictured: A directly contradictory description made by Gemini 3 - Fast.)
December 27, 2025 at 8:59 PM
Look at the Gaussian for executive function. Medicate anyone under the cutoff.

Recalibrate cutoff on the new curve. Give everyone infinite scroll interfaces. Medicate anyone under the cutoff.

Recalibrate cutoff on the new curve. Give everyone engagement-maxing recommendation algorithms. Medicate a
December 27, 2025 at 6:25 PM
Reposted by Naomi Saphra
another robot highlight for 2025: man wearing humanoid mocap suit kicks himself in the balls
December 27, 2025 at 5:27 PM