Aaron Steven White
LightNews LightNews
banner
aaronstevenwhite.io
Aaron Steven White
@aaronstevenwhite.io
2.1K followers 510 following 510 posts
computational semanticist. into modular synths and cocktails. http://aaronstevenwhite.io
Posts Replies Media Videos
aaronstevenwhite.io
What would great grandpappy think!?!
He bears the name of his third-generation great-grandfather, Civil War General William Tecumseh Sherman, as did his father and grandfather before him.
November 20, 2025 at 12:30 PM
aaronstevenwhite.io
lol.
A screenshot of https://www.githubstatus.com/ showing that the only services down is the git services.
November 18, 2025 at 9:08 PM
aaronstevenwhite.io
back on my bullshit
The header for a problem set on Bayesian Model Comparison for a course on Statistical Methods in Linguistics.
November 18, 2025 at 8:15 PM
aaronstevenwhite.io
deadlift day.
The album cover of Surgeon’s Force+Form.
November 16, 2025 at 3:56 PM
aaronstevenwhite.io
A picture of square waves filtered by a low pass filter with high resonance.
November 15, 2025 at 3:39 AM
aaronstevenwhite.io
Coming up in 30 minutes! Join by zoom here: rochester.zoom.us/meeting/regi...
Towards Robust Language Technologies for Everyone Antonios Anastasopoulos George Mason University Tuesday, November 11, 2025 12:30 p.m.–2 p.m. Humanities Center, Conference Room D Language technologies, despite all the incredible recent progress, do not yet robustly work for everyone. In this talk I will first summarize some of my group's recent work on addressing challenges we are still facing in the real world, such as handling language varieties (dialects), minority languages from bilingual communities, and catering to the needs of indigenous and endangered language communities. I will discuss solutions around data curation, modeling, and models' cultural adaptation, as well as some exciting preliminary results on incorporating human knowledge into LLMs in a way that reduces data requirements. Antonios (Antonis) Anastasopoulos is an Assistant Professor in Computer Science at George Mason University. He received his PhD in Computer Science from the University of Notre Dame and then completed a postdoc at Languages Technologies Institute at Carnegie Mellon University. His research is on natural language processing with a focus on low-resource settings, endangered languages, and cross-lingual learning, and has been funded by the National Science Foundation (including by a CAREER award), the National Endowment for the Humanities, the DoD, Google, Amazon, Meta, and Microsoft Research.
November 11, 2025 at 5:02 PM
aaronstevenwhite.io
GPT-1...said it better?
A generation from GPT-1 with the seed text: “Our ultimate strength does not come from equipment,” he said. “ GPT-1 continues: we have no such power because we have no mind . i know what it means , she said . the two of them spoke in the same language in which they had spoken only minutes before . the two of them spoke at the same time . they spoke in one language with different accents , and spoke with the same accents at the same time .
October 30, 2025 at 11:21 PM
aaronstevenwhite.io
This language model is very 2018.
“Our ultimate strength does not come from equipment,” he said. “It comes from the men and women of the rank and file. That’s true. It comes from you people, incredible people, good-looking people. Too many good-looking people. I don’t like good-looking people. I never liked good-looking people. I’ll be honest with you, never admitted that before.”
October 29, 2025 at 1:34 AM
aaronstevenwhite.io
full Anathem or bust.
A screenshot of the Wikipedia page for Neal Stephenson's Anathem, with the plot summary: "Anathem is set on the fictional planet of Arbre. Thousands of years before the events in the novel, the planet's intellectuals entered concents (monastic communities) to protect their activities from the collapse of society. The avout (intellectuals separated from Sæcular society) are banned from possessing or operating most advanced technology and are supervised by the Inquisition, which answers to the outside world. The avout are normally allowed to communicate with people outside the walls of the concent only once every year, decade, century, or millennium, depending on the particular vows they have taken."
October 21, 2025 at 4:53 PM
aaronstevenwhite.io
@reilly-coglab.com's giving this week's LS2 talk in about 20 minutes! Join by zoom here: rochester.zoom.us/meeting/regi...
Executive Decline v. Vocab Growth: In Aging, Who Wins? Jamie Reilly (Temple University) Aging is associated with a paradoxical combination of vocabulary growth and declining executive functions. It is unclear how the interaction between vocabulary knowledge and reduced processing manifests in a naturalistic language task such as conversation. We analyzed >1500 unscripted conversations (>8 million words) between speakers of different ages (18-66) using a newly developed software application, ConversationAlign. We specifically focused on concrete versus abstract word retrieval and lexical alignment as functions of age differences between conversation partners. Aging was robustly associated with abstract word dropout, and age differences between interlocutors were associated with growing misalignment on word concreteness. We will discuss how these observed patterns inform our understanding of patterns of age-based lexical retrieval during naturalistic communication.
October 7, 2025 at 4:11 PM
aaronstevenwhite.io
The only Halloween movie you need.
A movie poster in French for Last Year at Marienbad.
October 7, 2025 at 12:49 PM
aaronstevenwhite.io
Happening in one hour! Join by zoom here: rochester.zoom.us/meeting/regi...
Surgery on LLMs: Editing, Unlearning, and Securing Knowledge Wengpeng Yin Pennsylvania State University LLMs are rapidly being deployed in high-stakes settings, from scientific discovery to legal assistance to personal health, where we can no longer treat them as static black boxes. We need ways to update, correct, and safeguard their knowledge as the world changes and as privacy risks grow. This talk will present a vision for LLM surgery: systematic ways to make models learn, forget, and protect in a controlled manner. We will explore how to edit models so they gain new knowledge without losing old capabilities, how to make them forget harmful or outdated information without collapsing, and how to truly safeguard private data rather than relying on superficial protections. Along the way, I will share surprising failure cases, new evaluation tools, and early solutions that point toward a future where LLMs can safely and reliably evolve with human needs.
September 30, 2025 at 3:39 PM
aaronstevenwhite.io
I mean. Obviously the only way to listen.
The sleeve of Cate Le Bon’s Michelangelo Dying on top of a record player with a pink record playing.
September 29, 2025 at 10:52 PM
aaronstevenwhite.io
A Google AI overview: The question "where was Kris by Bergman filmed?" is ambiguous because there are two famous individuals with that name, and the word "kris" can refer to a Swedish word or a type of sword. If you're referring to the 1946 film Crisis (Kris), it was filmed in Hedemora, Dalarnas län, Sweden. If you mean the 2021 film Bergman Island, which features Ingmar Bergman and was filmed on his home island, it was shot on Fårö, Sweden.
September 25, 2025 at 2:39 AM
aaronstevenwhite.io
Happening in one hour! Join by zoom here: rochester.zoom.us/meeting/regi...
GENERIC PROPOSITIONS: WHAT ROLE DO THEY PLAY IN GLIDING ACTION BERNHARD NICKEL Generic generalizations such as "ravens are black" or "tigers have stripes" have seen an upsurge of interest, due in significant part to the fact that they are also the language of stereotypes ("Germans lack a sense of humor"). This interest is presumably motivated by the thought that generics guide actions. In this talk, I investigate this link more closely. The basic problem is that generics display incidence variability: the truth of a given generic is compatible with an incredibly wide range of frequencies with which the predicated property is present among members of the kind. Yet the most widely accepted model of rational action that takes risk into account is expected utility theory where such frequency information is crucial. I'll look at various solutions and present my own, according to which the truth of a generic proposition is only distally related to any particular instance of action guidance.
September 23, 2025 at 3:18 PM
aaronstevenwhite.io
The only way to send summer off is with a zombie in the yard.
A zombie glass filled with a cocktail sitting in grass with trees in the background.
September 21, 2025 at 10:05 PM
aaronstevenwhite.io
The Fall 2025 UR CLS Language Science Lecture Series starts next week!

Zoom registration: rochester.zoom.us/meeting/regi...
Mailing list: rochester.us10.list-manage.com/subscribe?u=...

All talks are Tuesdays 12:30–2pm.
UNIVERSITY OF ROCHESTER CENTER FOR LANGUAGE SCIENCES LANGUAGE SCIENCE LECTURE SERIES FALL 2025 9/23 BERNHARD NICKEL HARVARD U 9/30 WENGPENG YIN PENN STATE U 10/07 JAMIE REILLY TEMPLE U 11/11 ANTONIOS ANASTASOPOULOS GEORGE MASON U 11/18 MAYA ABTAHIAN U OF ROCHESTER
September 16, 2025 at 4:47 PM
aaronstevenwhite.io
In other unsolicited music recommendations: the new Water From Your Eyes album is so fucking good.
The vinyl sleeve of Water from your Eyes’
September 7, 2025 at 10:40 PM
aaronstevenwhite.io
Been listening obsessively to No Joy’s Bugland since it came out. Definitely in my current top 5 of the year.

Also been revisiting Wait to Pleasure (pictured) and let me tell you: Hare Tarot Lies is just as transcendent as it was in 2013.
The vinyl sleeve of No Joy’s Wait to Pleasure laying on top of the dust cover of a spinning record player.
September 7, 2025 at 10:34 PM
aaronstevenwhite.io
Like, some people just have taste.
Tracklist Cybotron - Earth Little Computer People - Little Computer People (Karl Bartos Remix) Mobile Space Unit - Penalty Dream Klangwerk - Die Kybernauten Scope - The Future (Cybotron Mix) Calagad 13 - Super Jam (Xtndd) Novamen - Bubblegum Middle Men - Defeat The Object Interfunk - Travel To Another Dimension Excel - Don't Stop Now Chaos - Afrogermanic T.Linder - Wolfpack T.Linder feat. Blak Tony - Leatherneck The Martian - Skypainter The Shadow - The Vision 6K - Exponential Hazed - Bellchant Joey Beltram - Orion Sebastian S., Jessie Deep! - Escape Lory D - Sickness Underground Resistance - Riot Oligo - Light Orbit The AM - Sex Worker uNN - Make Me Listen (Electro) EDO8 - Voyager Insync vs Mysteron2 - Science Fact Scan X - Higher
August 23, 2025 at 10:41 PM
aaronstevenwhite.io
Came down to the wire on the shift from "this brand new course is going to suck" to "actually, this brand new course is going to be so sick", but I did in fact make it.
A course schedule: Week Dates Topics Due 1 Aug 25, 27 Introduction; What is a probability? — 2 Sep 1, 3 Labor Day ; Aaron away — 3 Sep 8, 10 Random variables; Discrete distributions — 4 Sep 15, 17 Continuous distributions — 5 Sep 22, 24 Statistical inference — 6 Sep 29, Oct 1 Linear regression PS1 7 Oct 6, 8 Model evaluation PS2 8 Oct 13, 15 Fall Break; Aaron away PS3 9 Oct 20, 22 Generalized linear models — 10 Oct 27, 29 Experimental design;  Midterm Proposal (414) 11 Nov 3, 5 Mixed effects models — 12 Nov 10, 12 Clustering and mixture models PS4 13 Nov 17, 19 Factorization PS5 14 Nov 24, 26 Custom model design; Thanksgiving — 15 Dec 1, 3 ø Presentations (414) — 16 Dec 8 ø Presentations (414) — Dec. 9-12 Ü Oral assessments (414) Paper (414)
August 23, 2025 at 6:13 PM
aaronstevenwhite.io
Couldn’t really justify lugging my modular rack cross country for a one week trip, but obviously I had to bring at least my elektron boxes, right?
An Elektron Digitakt and Digitone on a wood table with a plant in the background.
June 26, 2025 at 6:10 AM
aaronstevenwhite.io
Enviable, deep, blissful ignorance.
A greyhound lying in grass in sunshine.
June 12, 2025 at 10:09 PM
aaronstevenwhite.io
Just over here doing some truly unholy shit.
A VSCode window with the sidebar open, showing a python library with modules referring to category theoretic concepts and deep learning concepts. The file that is open implements a Monad class.
June 12, 2025 at 6:43 PM
aaronstevenwhite.io
A large pile of books underneath a chalkboard with the words “please take” written above them and arrows pointing toward the book pile.
June 11, 2025 at 5:45 PM