Nathan Bartley

I'm hiring a postdoc to start in summer/fall 2027! My lab uses computational social & cognitive science methods to explore how tech shapes knowledge & power: www.crockettlab.org You can submit an application here: apply.interfolio.com/195297. Review of apps starts Nov 20. Please share!

Think spotlighting a junior researcher like this is excellent behavior

New from @kevinren09.bsky.social @manishraghav.bsky.social @nkgarg.bsky.social: A crisp characterization of how (potentially adversarial) distribution shift is a major challenge for AI text detection, and how semi-supervised learning helps fix this.

I am very excited to release WordLevel, an open source Python package for comparing texts at the word level! wordlevel.dev WordLevel makes it easy to compare texts with measures ranging from the Jensen-Shannon divergence to dictionary-based sentiment, and visualize them as word shift graphs.

Maybe worth saying aloud that I do worry about safety risks (hacking, bio, an “out of control” agent). I just also see those risks as deeply entangled with money, politics, personalities, regulation, such that I doubt we can “solve safety” if we don’t first address foundational issues.

As people have predicted, there are quite a few open-source alternatives to Jev, including: - Laya, this was the original according to him: github.com/NandhaKishor... - on Apple Core ML / neural engine: github.com/mizorewww/la... - with Qwen: github.com/jaredpalmer/...

By the way, bad software engineering does regularly kill people, just not in an easily traceable way.

People randomly assigned to a liberal or conservative group chat groups shifted their expressed stance in the direction of the group, demonstrating that social norms alters political expression This effect emerges after controlling for baseline ideology www.sciencedirect.com/science/arti...

Really good overview of how arxiv is thinking about handling the incoming LLM wave and the considerations going into it

The transparency is great — on the research side it’s great for reasoning about how the system works, but AFAICT there are no model weights in this release (understandable, but still). It’d be great to have even outdated weights to make stronger empirical claims about the effect of the feed

Feeling big #ic2s2 fomo this year, seems like a great conference (like always!)

New preprint with @carlbergstrom.com, Kevin Gross, and @mjcrockett.bsky.social. See Carl's thread. Surprise! Even if LLMs work perfectly, they'll make scientists refine their work less deeply, because they raise the opportunity cost of time, pushing researchers to do more, less well.

Very cool! A good reminder that there are other very useful network libraries besides just networkx and igraph

Cape Verde defeats Argentina via a 2-3 loss in extra time

Argentina barely deserved the win

The World Cup is full of magic moments, loving Cabo Verde today

Hoping Bluesky Communities launches before Musk manages to clean up X, which may take a while given the incentives at work, but maybe not. Even people who remain loyal to Musk know it has become a dumpster fire. It’s a big opportunity, but only if Bluesky provides space for many more communities.

this Türkiye win was for Eric Adams 🇹🇷

That was a great game to watch, honestly that momentum change in the second half gave me a lot more faith in the US squad than anything from the first two matches #TURvUSA

While this is a German court’s ruling, I imagine this adds fuel to the US debate on section 230

absurd saves from Rangel with the replays

Rooting for South Korea but my god that was was an incredible double save from Mexico.

Announcing co/core: a social computing experiment from the Graze team

A few PDSes I haven't seen in awhile broke into the top ten in raw # of active repos (June 2026) 👀 Looking at repos on pds.ls suggests a fair amount of bsky spam. Are there any lists of 'spammy' PDSes (or do devs just roll their own)? I know @mackuba.eu tags some PDSes w/ high bot activity.

Very cool, helping people find people!

I'd want the opposite of "mute thread" on @bsky.app as a feature: "subscribe to thread" à la slacks "get reply notifications". Really useful when someone asks a good question and I'm interested in seeing the answers.

So I've been digging into bluesky retention data to see what is preventing growth of the ecosystem. Using likes/posts/follows you can see who is active or dormant on ATProto. The headline is that only 5% of users are retained at 28d and 10% of current users churn. That's not pretty

Tezos has shot up the past couple of weeks with a ton of active repos 👀 presumably for this work tz2at.xyz

Great to see pre-registered studies testing the effect of algorithms on perception!

an interesting counterpart to the X transformer-centered pipeline setup — just make the whole pipeline one big PyTorch model @spacecowboy17.bsky.social @patrickboueri.com engineering.fb.com/2026/05/26/m...

What is Standard.site? Think of it as a file format for publishing longer posts from blogs, news sites…wherever. If you’re using Leaflet, pckt.blog, or Offprint, you’re already using Standard.site! https://standard.site/

a bird in tree is worth two in the bush

The X For You algorithm has been published: github.com/xai-org/x-al... Now we have the answer to how true this statement was: "Grok will literally read every post and watch every video" This was just marketing. There is no Grok reading any posts. 🧵

enough. i’m tired of github actions. it’s time for github consequences

Great resource — I was starting to build a follower graph from scratch and this makes that so much easier!

It’s D-day in sunny LA #trains

Keep it private! We’re adding more ways to connect with other Atmosphere apps. You can now launch Germ’s E2E encrypted messaging from Bluesky profiles. Check out @germnetwork.com for more.

Okay, it’s here! pdscheck.dev is LIVE!!! Thanks to @bad-example.com and @pds.dad for all their work on debug.hose.cam, and thanks to Claude for doing most of the heavy lifting. License is CC0 with an MIT fallback, and source is at tangled.org/chaosgreml.in/pdscheck, PRs OH SO WELCOME!

The prevalence of acadamese in scientific writing is partly why writing overall in science has been devalued. The average scientific paper is boring to read and has a boilerplate style, a void into which LLMs jump

“A structural change in research contribution and assessment is necessary to unshackle research communication from long-form narrative publication.”

TL;DR on AI-assisted or agentic coding: Some things change; some things stay the same. The same: - Code sent for review needs to be *reviewable* - Trust but verify, all the way through your SDLC - Getting far "off main" accrues technical debt - Risk tolerance informs where and how you experiment

Next @atproto.la meetup is coming up on May 9th!

recently I’ve been more intentionally courteous to other drivers on the road (like letting people merge when it’s bumper-to-bumper and thanking those who let me merge), and I feel like my average road-stress levels have plummeted Maybe it’s spurious but would recommend