Lightnews — Scholar-powered news

Naomi Saphra

@nsaphra.bsky.social

9.7K followers 1.5K following 2.4K posts

Waiting on a robot body. All opinions are universal and held by both employers and family. Literally a professor. Recruiting students to start my lab. ML/NLP/they/she.

Posts Media Videos Starter Packs

Pinned

Naomi Saphra @nsaphra.bsky.social · Apr 26

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

23 88 310

Reposted by Naomi Saphra

thebes @vgel.me · 9h

if you're interesting in gaining a better intuition for how llms behave at inference time, you should try logitloom🌱, the open-source tool i made for exploring token trajectory trees (aka looming) on base and instruct models! more info in thread

🌱 vgel.me/logitloom
💻 github.com/vgel/logitloom

4 20 87

Naomi Saphra @nsaphra.bsky.social · 10h

How many flagged images could you look at before you lose the ability to look normally at a child? At a puppy? At any adult woman? At any black person? I would rather give up visual social media entirely than increase the number of human casualties.

2 18

Naomi Saphra @nsaphra.bsky.social · 10h

It is wild that anyone out there is advocating for MORE human labor in image moderation. a friend of mine used to be a 4chan janitor. they’re dead now, but they spent the rest of their life unable to look at a child without feeling sick.

2 11 42

Naomi Saphra @nsaphra.bsky.social · 11h

Yeah. The huggingface incident was about someone collecting a small, public, open access dataset to support others’ data analysis.
Every company whose product is an actual model (and not scientific infrastructure) has quietly scraped far more.

Naomi Saphra @nsaphra.bsky.social · 11h

They want more traumatized human moderators forced to flip through reams of CP and snuff tyvm. Can’t have clankers steal these key human jobs which have a turnover rate of well under a year because it only takes a couple weeks to get intractable PTSD

Naomi Saphra @nsaphra.bsky.social · 11h

One-in-a-million events happen all the time and many will happen to you eventually. Live like an actuary.

Naomi Saphra @nsaphra.bsky.social · 12h

It is very grounding to calculate how many people in your country are dealing with a newly diagnosed physical disability, a child with cancer, a parent with dementia, the death of a spouse. Nothing will make you feel smaller than to look up how common the worst thing that’s ever happened to you is.

1 1 22

Reposted by Naomi Saphra

Alexander Doria @dorialexander.bsky.social · 13h

Ok while I do think many issues with AI are overblown, web agents hitting hard on infra sustainability and knowledge access. Was just browsing the reference database for Disney comics: now login only indefinitely.

2 5 16

Reposted by Naomi Saphra

Greg Durrett @gregdnlp.bsky.social · 17h

Find my students and collaborators at COLM this week!

Tuesday morning: @juand-r.bsky.social and @ramyanamuduri.bsky.social 's papers (find them if you missed it!)

Wednesday pm: @manyawadhwa.bsky.social 's EvalAgent

Thursday am: @anirudhkhatry.bsky.social 's CRUST-Bench oral spotlight + poster

5 7

Reposted by Naomi Saphra

AAUP @aaup.org · 18h

NO LOYALTY OATHS IN HIGHER ED!

Trump's attacks on our universities are an attempt to consolidate power. This loyalty oath directly undermines our right to academic freedom & goes against every democratic principle our country should uphold.

Please click below, sign, & share.

#DefendHigherEd

University Administrations: Reject Trump's "Loyalty Oath" Compacts

The Trump administration is trying to blackmail schools to let him and his unqualified bureaucrats run our schools. They want to dictate what schools teach, who they admit and hire, what researchers s...

actionnetwork.org

1 63 80

Naomi Saphra @nsaphra.bsky.social · 20h

Omg congrats @ajyl.bsky.social

Conference on Language Modeling @colmweb.org · 22h

Outstanding paper 2🏆: Shared Global and Local Geometry of Language Model Embeddings
openreview.net/forum?id=aJD...

Reposted by Naomi Saphra

Jennifer Hu @ COLM (recruiting PhDs and postdocs!) @jennhu.bsky.social · 1d

At #COLM2025 and would love to chat all things cogsci, LMs, & interpretability 🍁🥯 I'm also recruiting!

👉 I'm presenting at two workshops (PragLM, Visions) on Fri

👉 Also check out "Language Models Fail to Introspect About Their Knowledge of Language" (presented by @siyuansong.bsky.social Tue 11-1)

4 21

Reposted by Naomi Saphra

Maria Antoniak @mariaa.bsky.social · 1d

Here’s a #COLM2025 feed!

Pin it 📌 to follow along with the conference this week!

2 15 22

Naomi Saphra @nsaphra.bsky.social · 2d

It’s a drawing my sister made for my bat mitzvah when I was 13, because I liked octopuses!

1 3

Reposted by Naomi Saphra

Quanta Magazine @quantamagazine.bsky.social · 2d

An unexpected challenge at the start of Naomi Saphra’s career shaped her research as a computer scientist. www.quantamagazine.org/to-understan...

6 56

Reposted by Naomi Saphra

thebes @vgel.me · 2d

new blog post! why do LLMs freak out over the seahorse emoji? i put llama-3.3-70b through its paces with the logit lens to find out, and explain what the logit lens (everyone's favorite underrated interpretability tool) is in the process.

link in reply!

8 45 200

Naomi Saphra @nsaphra.bsky.social · 3d

do you think it’s ok to skip your 1500th wedding anniversary??? this is a special moment. congratulations to all bats

1 3

Reposted by Naomi Saphra

Greta Tuckute @gretatuckute.bsky.social · 3d

Check out @mryskina.bsky.social's talk and poster at COLM on Tuesday—we present a method to identify 'semantically consistent' brain regions (responding to concepts across modalities) and show that more semantically consistent brain regions are better predicted by LLMs.

Maria Ryskina @mryskina.bsky.social · 4d

Interested in language models, brains, and concepts? Check out our COLM 2025 🔦 Spotlight paper!

(And if you’re at COLM, come hear about it on Tuesday – sessions Spotlight 2 & Poster 2)!

Paper title: Language models align with brain regions that represent concepts across modalities.
Authors: Maria Ryskina, Greta Tuckute, Alexander Fung, Ashley Malkin, Evelina Fedorenko.
Affiliations: Maria is affiliated with the Vector Institute for AI, but the work was done at MIT. All other authors are affiliated with MIT.
Email address: maria.ryskina@vectorinstitute.ai.

4 14

Naomi Saphra @nsaphra.bsky.social · 3d

congratulations to all my bats

Carl Zimmer @carlzimmer.com · 3d

1500th bat species discovered in Africa www.batcon.org/press/1500th...

1,500th Bat Species Discovered in Africa's Equatorial Guinea

Breakthrough marks global milestone in biodiversity.

www.batcon.org

1 1 16

Naomi Saphra @nsaphra.bsky.social · 3d

COLM is approaching! You should meet our lead author Natalie and ask her about how LMs learn to beat human experts!

Naomi Saphra @nsaphra.bsky.social · Aug 29

How can an imitative model like an LLM outperform the experts it is trained on? Our new COLM paper outlines three types of transcendence and shows that each one relies on a different aspect of data diversity. arxiv.org/abs/2508.17669

1 17

Naomi Saphra @nsaphra.bsky.social · 4d

I think I’ve seen work on clustering ensembles and pruning out underperforming members of each cluster but not sure if it works in modern scales

Naomi Saphra @nsaphra.bsky.social · 4d

My friends and family know I love looking at random variation. But my ENEMIES claim random seed will be UNIMPORTANT at large scales. Fools! I will say YOU ARE WRONG, MY ENEMIES. LOOK.

2 16

Naomi Saphra @nsaphra.bsky.social · 4d

Ensembling is good in data bottlenecked regimes! My takeaway: random variation is only going to get more important at scale—even as it decreases in magnitude. The deterministic benefits of scale eventually decline faster than a model’s randomly selected advantages!

Pre-training under infinite compute

Since compute grows much faster than web text available for language model pre-training, we ask how one should approach pre-training under fixed data and no compute constraints. We first show that exi...

arxiv.org

2 6 41

Naomi Saphra @nsaphra.bsky.social · 4d

A preliminary result that ChatGPT-3.5 was more likely to answer an under specified request directly, rather than ask necessary follow up questions, if you introduced yourself with a female name. We were gonna call it The Mansplaining Effect. Wound up focusing on guardrails in the paper instead!

1 1

Naomi Saphra @nsaphra.bsky.social · 4d

I have posted stuff on here that I probably shouldn’t have, so imagine what I post to the groupchat