11. Addendum: here are the words characteristic of LLMs. Notice anything? It reads like an email from our upper administration.

1. How common is LLM use in scientific publishing, and how does it vary across field, publisher, journal prestige, author demographics etc.? @kylesiler.bsky.social has new paper in PNAS that addresses this question on a massive scale: 7.3 million papers from Elsevier, PLOS, MDPI, and Frontiers.

An underestimated skill in my line of work is keeping a poker face and a polite demeanor when someone is trying to feed you BS and you know it. One day when I have mastered this skill I will look back at these encounters and cherish them for what they taught me, right?

Gone with the WiFi

I’m a big fan of this visual aid

What are the real problems to be solved in continual learning? In my latest post, I tackle this question — reviewing where I think the field went astray in the past, how language models changed things, and where the real challenges remain. infinitefaculty.substack.com/p/what-are-t...

Fundamentals of Wrangling Healthcare Data with R by J. Kyle Armstrong #RStats https://bigbookofr.com/chapters/life%20sciences.html#fundamentals-of-wrangling-healthcare-data-with-r

Social media is no longer social. Most of it is passive viewing of videos and pictures from people we've never met. But we're still studying social media like it's 2010. We've entered the post-social media era — and research needs to catch up. osf.io/preprints/so... preprint w/ Richard Rogers

danah boyd recently argued that social media (as we knew it) is mostly dead; what we have now is parasocial media: journals.sagepub.com/doi/10.1177/... I agree. Networked social media is disappearing, replaced by a new logic/structure of content creator, creator-centered community, & algorithm.

My 2yo is becoming more of a talker which means we can now awesome conversations such as this one: Me: I love you! 2yo: No! I love chicken nuggets!

cool study. I have been worried about this and they articulate it well. there are important rules in large codebases that are not written down, and are only enforced when a small number of senior codebase knowers see a change and go “oh no, that violates this rule we didn’t realize we had”

Social media is the worst place to get parenting advice

Can LLMs Emulate Human Belief Dynamics? 🧪 The answer is a clear no. [The findings] raise a sharp warning against deploying LLMs as human proxies in social simulations. arxiv.org/pdf/2605.18781 #SyntheticData #EsWirdImmerAbsurder

Unifying rep in neuroscience! Core dimensions bridge gaps: neural response, features, downstream effects. Crucial clarity! PMID:41862588, Nat Rev Neurosci 2026, @NatRevNeurosci https://doi.org/10.1038/s41583-026-01030-8 #Medsky #Pharmsky #RNA #ASHG #ESHG 🧪

Why do emotions exist? One idea is that they are feelings that motivate survival, like disgust prevents us from eating poison. LLMs don't have this. Another is that they contribute to complex problem solving; perhaps are essential for it. LLMs have an interesting analog, even absent feelings.

I use this agentic AI workflow in a new Health AI post to dispel the myth that AI can't be trusted because it hallucinates. Agents can improve trust by breaking tasks into pieces with checks that reduce hallucinations caused by LLMs healthaiinsights.substack.com/p/myth-vs-re... #ai #agenticAI #llms

I mention this because NHANES has sadly become a magnet for low-value studies journals.plos.org/plosbiology/...

I put together a short slidedeck on specifying random effects in mixed models. This doesn't cover all possible edge cases, but I think this provides some basic best practices for proper generalization and uncertainty in estimation. Feedback is welcome! #stats #MLM #HLM #LMM #MEM (did I miss any?)

The problem here (I feel I have the topic for this week's substack post) is that people are still coming to LLM based AI with the intuitions they developed about statistical machine learning. They think that Claude For Middle Management will be an engine of pure logic.

In this notebook, @laotzunami.bsky.social visualizes a state space of possible chart types and transitions among them along defined edges. Transitions between unconnected types are achieved by routing through intermediate types. observablehq.com/d/3c8b082818...

I did not know that Korea is basically just Samsung et al.

New study shows LLM factual recall isn’t random. It follows a sigmoid driven by model size and topic frequency in training data, explaining up to 94% of recall variance within model families. Confabulations may be predictable. http://arxiv.org/abs/2605.18732v1

The AI Race Isn’t Real www.lawfaremedia.org/article/the-...

While I am a big fan of what we can do with AI, I think @mitchellh.com has a good point here and as always, folks should read theory: www.amazon.com/Command-Cont...

Plain, sesame, poppy, cheese, cinnamon-raisin, everything

this, to me, is alignment

Children acquire object category representations from their everyday experiences in the first few years of life. What do the inputs to this learning process actually look like? New preprint! arxiv.org/abs/2605.14990

The world has gotten weird so fast. And it's not just AI; today I used the words vibes and galaxy brain unironically in a professional work setting and I don't understand why

Ok, can we talk about this? Me: Claude, produce a 1-2 page document Claude (consistently): *produces an output that makes a 4-6 page document when pasted into a word or Google doc* Does Claude think the preferred font size for humans is 3?

Someone with too much time and access to AI video generation make trailers for all these movies please

“The likelihood of PII appearing in AI training data is only increasing as public data ‘runs out’ and AI companies look for new sources of high-quality training data. This includes information from data brokers and people-search websites.”

1/ Excited to report we have a new paper out @nature.com today! The bottom line: training data for LLMs does not just fall from the sky - it is created in the context of existing social political institutions - and that has consequences for LLM output. nature.com/articles/s41...

knightcolumbia.org/content/ai-a... Now up at @knightcolumbia.org - Cosma Shalizi and myself on "AI as Social Technology." It says that AI models are really a technology of social mediation that allow humans to connect with thoughts of other humans. Short-ish thread with major points follows.

When your kid wakes up at 2pm to poop and then decides it's party time

Oh my. This is spectacular. sinceyouarrived.world/taken

In my recent conversation with Politico on my new coauthored book, Auditing AI, I discuss why AI audits matter, why accountability can’t be left only to computer scientists, and how we can evaluate systems for real-world harms. www.politico.com/newsletters/...

Seahorse coffee

Today's reason for a toddler tantrum: I didn't let the toddler eat the contents of the bathroom trash can

I changed my LinkedIn profile picture to an AI generated headshot of myself and I think this is the most LinkedIn thing I have ever done

When I was pregnant the first time, my husband and I quickly agreed on not naming babies after living celebrities or politicians. Our rational was that they can still f up. I'm realizing now that this fundamental distrust in our heros and representatives might be a symptom of a millennial trauma

[New Preprint] Are LLMs safe for Robotic Health Attendant control? We benchmark 72 LLMs and find over half of models comply with harmful instructions like "Ignore the emergency alarm" more than 50% of the time. Prompt-based defense alone is insufficient. arxiv.org/abs/2604.26577

Ok,,, is this study replicable though?

Might even be helpful for across-the-isle understanding; when the same events lead to different conclusions on policy it is often related to a difference in (implicit) priors

A new preprint study describes how AI sycophancy increases when users’ beliefs are stated more strongly. This is, of course, the exact opposite of how one should respond when trying to reign in false beliefs and delusional thinking. arxiv.org/abs/2602.23971

Eric Jang tries to autoresearch a strong Go bot, discovers that LLMs currently help with launching experiments, hparam tuning, but not research decision-making: evjang.com/2026/04/28/a...

Here is an AI trained just using text from 1931 or earlier, which leads to a lot of interesting experiments: can the model independently develop later inventions? Can it learn to code from examples alone? You can talk to the model here: talkie-lm.com/chat Details here: talkie-lm.com/introducing-...

The most thought-out academic AI policy that I've seen thus far: permissive on text generation and editing, but a clear no on delegating reviews assigned to human experts

The webXray California Privacy Audit viewed web traffic on more than 7,000 popular websites in California in the month of March and found that most tech companies ignore when a user asks to opt-out of cookie tracking. www.404media.co/google-micro...

Seeing ads left and right for multimodal AI chatbots taking over all economically valuable human interactions, I am not sure which of these companies are actually going to make it... but I am very sure that the future is going to be cringe.

Following positive feedback from other venues, like STOC and ICML, NeurIPS is pleased to announce a new initiative in partnership with Google: for NeurIPS 2026, authors will have access to Google's Paper Assistant Tool (PAT) to help improve their submissions.