Very interesting cross-model and fine-tunes interpretability work.
come for Chris Olah on stage with the pope, stay for the failure of dictionary learning and the collapse of two traditions of reification
Today is release v1.0 of Narrative Nubs —
Story generators based on systems from the past 65+ years, capturing essentials, as in “the nub of the argument.”
GESTER, Story Machine, TAILOR, and “Through the Park” nubs are ready for research, teaching, creative practice.
codeberg.org/nickmontfort...
Congrats to the class of 2026! Pomppomppomp 🎓
Feat. @roopikarisam.bsky.social and @jeddobson.bsky.social
an absolute delight to collaborate with @roopikarisam.bsky.social on this smart piece for @religiondispatches.org about the pope linking AI to histories of enslavement and exploitation
check it out!
religiondispatches.org/2026/05/26/l...
xan v0.58.0 is out \o/
More performance, less cruft, very wow!
You can now draw synthwave sparklines or visualize network layouts in your terminal like it's 1970!
Full changelog: github.com/medialab/xan...
@medialab-scpo.bsky.social
Pope Leo XIV released his first encyclical this week, on the 135th anniversary of his eponymous predecessor’s landmark Rerum Novarum, or Rights and Duties of Capital and Labor. And like Leo XIII befor...
religiondispatches.org
NEW #OpenAccess book drop!
Wargame simulations, AI, weaponised drones, territorial enclosure, and extractivist economics—WAR MACHINE presents a series of digitally simulated conflicts over the most ecologically vulnerable areas of the Earth.
FREE to download: punctumbooks.com/titles/war-m...
New dataset published: this is the audio + transcripts of AI-generated podcasts using NotebookLM to analyse how it "translated" source documents from different times and cultures. My favourites are the 20 podcasts generated from an empty PDF. The paper about it is in press.
doi.org/10.18710/RNT...
My tiny thread yesterday is now a whole op-ed! A quick read, connecting Pope Leo's apology for colonialism and slavery to his warnings about AI — it's all related. Working with editor @mpgphd.bsky.social was an absolutely pleasure!
Nick Montfort
This neat site assembles all kinds of ancient OSs (more like window managers) which reminded me of the hyper-obscure one that I learned Desktop computing on, called GeoWorks Ensemble. It pretends to be Windows but if you poke anything it crumbles back into DOS. Glorious. www.typewritten.org/Media/
maybe: leif weatherby
Audio files and transcripts of AI-generated podcasts created using Google's NotebookLM. Four podcasts were created by uploading specific PDFs. Twen...
punctum 🌈 books: spontaneous acts of scholarly combustion
Jill Walker Rettberg
Sarah Wasserman
Writing this talk now and have never felt crazier. This project began as a joke ("To get AGI you need to traumatise AI into a psychoanalytic split") and somehow I became convinced by it. Alignment is the Lacanian "cut" of the signifier producing the alien subject of AI.
github.com/quadrismegis...
Fabian Offert
Roopika Risam
Ryan Heuser
the latest from RD: @roopikarisam.bsky.social on what makes Leo XIV's first encyclical so radical -- tracing AI back through histories of enslavement and exploitation, and repenting for the RCC's complicity in both
religiondispatches.org/2026/05/26/l...
Pope Leo XIV released his first encyclical this week, on the 135th anniversary of his eponymous predecessor’s landmark Rerum Novarum, or Rights and Duties of Capital and Labor. And like Leo XIII befor...
religiondispatches.org
In _The Ideas Letter_, Leif Weatherby, Tyler Shoemaker, and I wrote about the creation of reality through pseudoscience, religion, and psychosis. Of course, the piece is mostly about AI.
The authors contend that contemporary efforts to render AI systems interpretable rest on a mistake: reification, the process of treating abstractions and statistical artifacts as if they were concrete...
You can literally watch repression & displacement consolidate over fine-tuning: here's next-token probs across checkpoints of OLMo-3-7B-Think-SFT. Explicit words are repressed almost instantly but safer (displaced) alternatives emerge much later. It learns what not to say before what to say instead.
Ben Recht
Ryan Heuser
Submitting this abstract to "Accelerationism Revisited", a symposium in Dublin. Mapping psychoanalytic topology in LLM base models → instruction-tuned → safety-tuned models. They progressively "displace" (in Freudian sense) censored content into adjacent semantics, even across hidden model layers.
