iiinova — where your digital self keeps evolving

Reposted by Michael Castelle

Anna Weichselbraun @annaw.bsky.social · 25/08/2026

@mcastelle.bsky.social offers a brief intellectual history of how a widespread disinterest for language as social action in STS and sociological theory has been challenged by the rise of LLMs. sociologica.unibo.it/article/view...

sociologica.unibo.it

Is Language Social Action? A Commentary on Language Agents | Sociologica

153

Reposted by Michael Castelle

Zeerak Talat زیرک طلعت (they/them) @zeerak.bsky.social · 18/06/2026

Call for papers for special issue on Ethics in NLP in the journal Computational Linguistics. Discussions around ethics in #NLP and #ComputationalLinguistics is often limited by the expectations for what should be published in *CL conferences. Link to call: www.aclweb.org/portal/conte...

1109

Michael Castelle @mcastelle.bsky.social · 28/03/2026

Same source as thinking "step by step" (monter peu à peu) btw

010

Michael Castelle @mcastelle.bsky.social · 15/03/2026

It may not be part of the original critique, but (as fascinating as obligatory non-zero temperature is) the requisite "stochastic" element of the training process (e.g. batch SGD) sometimes seems underexplored

110

Michael Castelle @mcastelle.bsky.social · 20/01/2026

Whether these outputs constitute "societies" will be an interesting debate of course, should social scientists care to have it; at the very least, they should be aware that the term is presently escaping their orbit.

000

Michael Castelle @mcastelle.bsky.social · 20/01/2026

Was a bit upset that Google managed to cite Bakhtin in a publishable paper on LLMs before me (and *cough* all of the humanities), but then saw that the lead author was a UChicago Sociology PhD student, nice one.

110

Michael Castelle @mcastelle.bsky.social · 30/09/2025

it won't unless you really want to describe the linear/nonlinear vector trajectories of high-dimensional (and often multimodal!) input tokens as "symbolic"

010

Michael Castelle @mcastelle.bsky.social · 25/09/2025

when they say Stewart Brand prefigured the Internet they definitely got that part right

010

Michael Castelle @mcastelle.bsky.social · 25/09/2025

The history of computer science is full of “theorems” that, once sufficiently popular in a given subfield, appear to obstruct conceptual progress for about a decade

020

Michael Castelle @mcastelle.bsky.social · 15/09/2025

> be a human in an oral society without writing > all you hear are phonemes > you don't care, it's all abstracted away (Sapir 1933) > you live for a world of pure ideas, chain of concepts, reasoning streams: la pensée sauvage (Lévi-Strauss 1962) > phonemes don't exist.

060

Michael Castelle @mcastelle.bsky.social · 13/09/2025

Very back-of-the-envelope math but going by recent official stats, Chicago appears to have over 10x crashes-per-registered-vehicle than London

022

Michael Castelle @mcastelle.bsky.social · 05/09/2025

We’ve also been training on a mix of code and natural language (cleaned/augmented or not) since GPT-3 which should have been extremely weird to everyone. But I agree that humanists are well placed to contrast the long history of “reasoning” with whatever it is that currently goes under that name

010

Michael Castelle @mcastelle.bsky.social · 05/09/2025

I was disappointed less in the humanists themselves than in the computers they had to use

010

Michael Castelle @mcastelle.bsky.social · 05/09/2025

My position (as a disappointed bystander to much of the initial project) is that it is only now that “Digital Humanities” can truly begin

110

Michael Castelle @mcastelle.bsky.social · 05/09/2025

I guess I’m personally on your side here but for whatever reason, people who love counting seem to be slightly uncomfortable with having counting taken away from them (such as, for example, when you do gradient descent on hundreds of random token sequences simultaneously and repeat ad infinitum)

110

Michael Castelle @mcastelle.bsky.social · 05/09/2025

Yes, and I believe that sense-making machines made from layers of linear and non-linear rotations/rectifications of high-dimensional vectors are quite intellectually (and semiotically) inimical to the many cultural descendants of symbol-centric logic/probability/computation/tabulation

000

Michael Castelle @mcastelle.bsky.social · 05/09/2025

If what you mean by “statistical” is “learns from data somehow”, sure. Otherwise, disagree. Calling softmax() occasionally is pretty distant from actual probability theory; and there’s nothing continuous about 1990s “statistical NLP” vs. today’s NLP.

110

Michael Castelle @mcastelle.bsky.social · 04/09/2025

It's worth questioning how much of the Sutton & Barto-style reinforcement learning framework actually remains in the present LLM fine-tuning context IMO (not saying I have the answer to this question... but my intuition is that it's less than one might think!)

000

Michael Castelle @mcastelle.bsky.social · 04/09/2025

A lot of existing "digital humanities" is also (overtly or covertly) indebted to the symbolic/tabular/statistical forms of computing that contemporary large-scale connectionist AI deeply problematizes and/or doesn't much care about.

180

Michael Castelle @mcastelle.bsky.social · 06/08/2025

Don't worry, such content has been reclassified at the system prompt level as purely metaphorical x.com/AmandaAskell...

000

Michael Castelle @mcastelle.bsky.social · 28/07/2025

As Lévi-Strauss once fantasized, logical properties are now perfumes. Best of luck to those attempting to reverse-engineer the subtle art of the blind idiot bricoleur-perfumers in question.

020

Michael Castelle @mcastelle.bsky.social · 24/07/2025

They are, but they don't seem to actually give access to historians of computing so (outside of Cortada's books) they constitute more of a scholarly black hole than not

010

Michael Castelle @mcastelle.bsky.social · 23/07/2025

Even my "aggressive timelines" for weird interdisciplinary consilience at the hands of deep learning would not have predicted that in 2025 you could jam this much linguistic anthropology into an ACL paper (and get selected for an oral presentation). Congrats to all involved!

141

Michael Castelle @mcastelle.bsky.social · 17/07/2025

Just checked and my council has actually been installing a couple hundred new hangars throughout 2024-2025, including an upcoming one around the corner 😎

010

Michael Castelle @mcastelle.bsky.social · 16/07/2025

FWIW it’s a nice idea but in practice most of these have very long waiting lists, I got an email 5 years after applying for a spot on my block. (Possibly fixable with more enforcement of semi-regular use, idk)

110

Michael Castelle @mcastelle.bsky.social · 27/06/2025

(P.S. one of the greatest outsider sociology books of all time, saying www.amazon.com/Playing-at-W... )

020

Michael Castelle @mcastelle.bsky.social · 27/06/2025

doesn't all that stuff (including the Northern IL/Southern WI wargaming variants in the US) derive from Kriegspiel / von Clausewitz? Already Lutheran.

130

Michael Castelle @mcastelle.bsky.social · 27/06/2025

Not sure Princeton Univ Press will accept my candidate title "Trust in 14,000-Dimensional Vectors"

010

Michael Castelle @mcastelle.bsky.social · 26/06/2025

My assumption is that this kind of extraction is only possible when model developers don't make a serious/extensive attempt to thoroughly deduplicate the pretraining data (though you can still expect it for e.g. passages in novels that are widely quoted elsewhere).

010

Michael Castelle @mcastelle.bsky.social · 10/06/2025

Who will inform the Bayesian cognitive scientists that their discipline is currently running its periodic systemwide update to model the mind on the fancy new computer

010

Michael Castelle @mcastelle.bsky.social · 28/05/2025

don't worry, ReLU is one of the worst-sounding distortion filters out there

010

Michael Castelle @mcastelle.bsky.social · 28/05/2025

Gadamer has entered the chat (to no effect whatsoever)

030

Michael Castelle @mcastelle.bsky.social · 20/05/2025

Yeesh, you millenarian Borgesians. We're merely becoming Stanisław Lem's Perfect Vacuum

010

Michael Castelle @mcastelle.bsky.social · 14/11/2024

Very nice critique re: "democratization" in AI research from @arjunsubgraph.bsky.social and @zeerak.bsky.social arxiv.org/abs/2406.11598

091

Michael Castelle @mcastelle.bsky.social · 28/02/2024

So you're telling me all of the billions of weight parameters underlying a seemingly uninterpretable generative model of language and culture can be ternary

052

Reposted by Michael Castelle

Nick Seaver @nickseaver.website · 22/12/2023

bringing it over from the old place: ‘tis the season to read “Father Christmas Executed” (1952) by Claude Lévi-Strauss, for your yuletide structuralism fix: pds12.egloos.com/pds/200901/1...

Father Christmas was hanged yesterday afternoon from the railings of Dijon Cathedral and burnt publicly in the precinct. This spectacular execution took place in the presence of several hundred Sunday school children. It was a decision made with the agreement of the clergy who had condemned Father Christmas as a usurper and heretic. He was accused of ʻpaganizing' the Christ mas festival and installing himself like a cuckoo in the nest, claiming more and more space for himself. Above all he was blamed for infiltrating all the state schools from which the crib has been scrupulously banished.

13411

Michael Castelle @mcastelle.bsky.social · 22/12/2023

Erik Borra at Univ. of Amsterdam is doing something like this serving up LLMs via his "Prompt Compass" server for local/affiliated digital methods researchers. Simple interface and (more) reproducible outputs. He might be willing to give suggestions, tell him I and/or @tedunderwood.me sent you

120

Michael Castelle @mcastelle.bsky.social · 07/12/2023

It's recently had an expansion (was closed for a bit and reopened in mid-2021) so could be worth checking out, have not been back myself yet!

010

Michael Castelle @mcastelle.bsky.social · 07/12/2023

Probably too late here but since nobody mentioned it: Musée de la Chasse et de la Nature in Le Marais (sort of the Parisian version of Museum of Jurassic Technology)

110

Reposted by Michael Castelle

Valentin Goujon @vgoujon.bsky.social · 01/12/2023

Join us for EASST-4S 2024 in Amsterdam! Take a look at our panel "Rebooting the STS programme for AI: Emerging controversies and methods for studying 21st-century AI" (P228), with members of the Shaping AI project (N. Chartier-Edwards, J. Roberge, E. Grenier and @mcastelle.bsky.social). Abstracts 👇

easst4s2024.net

Open Panels - EASST-4S 2024

Conference programme

022