Joey Rudoler

Social scientists should not use chat interfaces when using LLMs in their research: they are impressively inefficient, and obscure/impose important methodological decisions that require thought.

THREAD🧵

Wrote a new, modern stats curriculum.

Teach about probability and sampling via computational examples / simulations with real data. It's unbelievably helpful for intuition. Everything else follows.

Online and open-source: jrudoler-teaching.github.io/understandin...

Check out my (open source) short course on introductory statistics and data science!

It covers fundamentals of programming and probability, and then uses those to build up an intuition for hypothesis testing through sampling and simulation

jrudoler-teaching.github.io/understandin...

I'm building a course from scratch this summer. Premise is that sampling / simulation are key for building intuition and statistical thinking. Students should learn permutation tests before they learn t-tests. Do folks agree?

What's something you wish you learned in your first stats/data science course that would have helped you down the line?
Like some key intuition or insight that crops up all the time in diverse problems? Something better than learning the formula for a t-test

#MLSky #StatsSky

Community for Rigor | Units We made a network of essential units to help you better understand the principles and practices of scientific rigor.

Have you played with our Confirmation Bias Unit, yet? Head out to buff.ly/sa8UIeY and give it a try, share with your colleagues and students, and give us your feedback.

There are two kinds of statisticians: those who like telling other people they're wrong, and those who don't talk at all

Today, we are publishing the first-ever International AI Safety Report, backed by 30 countries and the OECD, UN, and EU.

It summarises the state of the science on AI capabilities and risks, and how to mitigate those risks. 🧵

Full Report: assets.publishing.service.gov.uk/media/679a0c...

1/21

Announcing a new week-long program for young computational neuroscience/ behavior professors to talk about rigorous science, mentoring, lab management, and networking in a stunning retreat setting. Do great science as a community and have fun doing so.

newly converted @weightsbiases.bsky.social fan here! honestly did not realize how much better it could be than tensorboard. just way more organized, pretty, intuitive, and feature-rich.
shoutout once again to the awesome engineers building various MLOps tools for the rest of us to use <3

And the company sells a pendant you can wear to bring the companion with you wherever you go... honestly feels dystopian.
Does anyone have a more positive take on this? I feel like this has to be regulated...

I have no chat history, no account, no data -- and this is the first message the bot sends!! Who does this help? Seems obviously like farming for engagement from users who want to commiserate and encouraging them to engage in negative thought patterns.

I don't understand the product appeal of AI companions... and honestly the companies producing them sometimes feel predatory (though I believe some people have good intentions to reduce loneliness etc.)
When I open friend.com the bot immediately turns on the anxiety / depression / alcoholism

At Penn you can get a certificate for going to some extra training, but it seems pretty divorced from your effort/performance in the classroom.
Departments could give teaching awards to the most highly rated TAs?
Teaching *well* could be part of PhD degree-granting requirements?

The reality is that this is what most researchers want — teaching is a chore and distracts from research. But it feels like a disservice to everyone and super out of whack with a university’s educational values.

Not a new topic but what are effective ways to incentivize grad students to be better TAs?

I really enjoy teaching but it seems there’s zero incentive to do a good job (beyond personal fulfillment and an altruistic desire to help students). We’re evaluated solely on research!

Couldn’t agree more! I got on it ~2 weeks into grad school when I realized how fundamental those challenges are to the PhD experience. Wanted to invest preemptively. Also not true for every school, but student health insurance is pretty generous - my in-network copay is $10 per session. No brainer.

If you’re in grad school, finding a therapist can be really helpful. The thing you’re doing is hard and it’s harder if you don’t have help managing imposter syndrome, stress, self esteem, and a whole bunch of other things.

Key-value memory in the brain Classical models of memory in psychology and neuroscience rely on similarity-based retrieval of stored patterns, where similarity is a function of retrieval cues and the stored patterns. While parsimo...

Key-value memory is an important concept in modern machine learning (e.g., transformers). Ila Fiete, Kazuki Irie, and I have written a paper showing how key-value memory provides a way of thinking about memory organization in the brain:
arxiv.org/abs/2501.02950

yeah exactly

In the original tree-of-thought paper they just prompt an LLM to rate a candidate thought sequence or vote on the best one. Then they use these heuristics to search the tree with standard algos like BFS and DFS.

Not clear to me if they're minimizing a different objective function at test time. if the "deliberative alignment" paper is any indication they may use an explicit reward model (there it's about safety compliance) to filter chains of thoughts to use in SFT. So still max joint probability of the seq?

YouTube video by Etymology Nerd RETURN TO MONKE 🦍 #conlang #linguistics #language

My absolute favorite content on the internet is this guy Adam Aleksic (@etymologynerd, but not on Bluesky) who posts about linguistics. Awesome combo of silly content (“I made a gorilla language”) and super insightful commentary about language and communication interacting with modern tech/society

To be clear, this study doesn't in any way show that e.g. ChatGPT4o will get worse over time

This study shows progress in the field (newer models improve on previous iterations), not "cognitive decline"

Hold on, that's not at all what this shows -- they don't evaluate how a single model's performance changes over time, they just observe that newer models (trained on more data with newer and more advanced methods) are better than different, older models

Visiting NYC really makes me feel like some sort of crunchy small town guy… but I live in Philly

LLM research today would look so different without @hf.co... making models of scale and diversity so easily accessible to researchers is an amazing engineering feat. Such a catalyst for progress. 🙏

Agreed - curating them is expensive though. How do we incentivize orgs that have the data (or have access to the data) to do it?
In research, my sense it that the incentive for curating these benchmarks is that they allow researchers to claim "my model is SOTA on X" which gets published

More than fair :)

I absolutely love Keynote for academic presentations

Built-in support for Latex equations, easy UI, presenter view, super-customizable templates, looks beautiful, has great recording features

Is it just the collaborative editing that has people using google slides or beamer/overleaf?