Zachary K Stine

Today's offering from the #NaturalPhilosophy Symposium: Kevin Zollman @kevinzollman.com on Refutation and Models of Social Organization. Commentary by Henry Farrell @himself.bsky.social.

www.youtube.com/watch?v=ybOr...

Fintan Mallory, Formats of Representation in Large Language Models - PhilPapers This paper argues for a pluralist approach to representation in large language models. There are two parts to this pluralism, the first is that we should recognise more than one vehicle ...

New paper on formats of representation in LLMs. I defend pluralism both about the vehicles of representation in ANNs (e.g. neurons, polytopes, embeddings) and about the formats of representation (e.g. nominal, analogue, structural). Hope it’s useful to someone. philpapers.org/rec/MALFOR-2

As we put it here, “the shape is the meaning”: arxiv.org/abs/2509.00248

What problem is explainability/interpretability research trying to solve in ML, and do you have a favorite paper articulating what that problem is?

The biggest limitation to this is that all of this hermeneutics stuff takes place in the math/comp domain without translating it back into the linguistic/cultural. But that’s where we are looking next. I suspect we will have to learn how to read networks/geometries/etc as text.

5/5

I’m skeptical that performance measures/benchmarks will be able to help us reason about model hermeneutics unless we can make their own interpretive lenses visible. Our framework offers a way to do this by contrasting a benchmark, as a modeling decision, against alternatives.

4/5

This is what we attempt to do in the paper. We start with a very simple theory of semantic models based on the distributional hypothesis and Saussure. From that we develop a theory of model semantics in which models themselves function as signifiers.

3/5

But to contrast a model against alternatives requires a theory of what exactly they measure that is sufficiently rigorous so as to motivate a measure of the semantic differences between models (without reducing semantics to a narrow task).

2/5

I’ve been preparing a presentation on this paper and trying to more concisely get the main point across. I’d say it comes down to this:

To make the interpretive lens of a particular modeling decision visible, you have to contrast it against a population of alternative decisions.

1/5

if attention mechanisms in transformers create graphlike transformations across text then it is actually an incredible pun that groups of sentences are called "para-graphs"

"Helen de Cruz (1978–2025): an academic and personal tribute" *Religious Studies* Published online 2025:1-4.
doi.org/10.1017/S003...
@helendecruz.net

Extremely good stuff from Leif here

To Understand AI, Watch How It Evolves | Quanta Magazine Naomi Saphra thinks that most research into language models focuses too much on the finished product. She’s mining the history of their training for insights into why these systems work the way they d...

I really enjoyed talking to @nsaphra.bsky.social about her thoughts on what much language model interpretability research misses. My latest in @quantamagazine.bsky.social:

i think regardless of who has authored a paper, we (the scientific community) can and should stop criticizing the work on lack of perceived novelty. we can do better than perpetuating harmful myths about individual scientific papers having to present a discovery no one has ever thought of before.

That is wild!

Hey thanks! Definitely going to submit elsewhere. But it’s hard because this is all very specific to a subfield of a subfield that has like 3 publication venues. It seems like the natural home for this work but I’m second-guessing that.

“Oh, you’re bringing up epistemology? Name every philosopher of science.”

But I’m glad to know that The Structure of Scientific Revolutions lays out a case of how supervised learning can easily mask cultural complexity in computational humanities work. Shame we’re still falling short on such an old and obvious problem.

Sadly, we limited ourselves to C. S. Peirce, Terrence Deacon (by way of Eduardo Kohn) and a bunch of current phil sci that is actually about ML.

Jean Baudrillard, Umberto Eco, Thomas Kuhn, Paul Feyerabend, Ian Hacking, Bruno Latour, Donna Haraway, and the anthology Models as Mediators edited by Mary Morgan and Margaret Morrison.

Forgive one last point to vent about: The reviewer’s only other criticism was a lack of novelty. The obsession with novelty has made a lot of ML work insufferable so it’s unfortunate to see it here. But here’s who is apparently already making our points about how ML is used for humanities research:

Miyagawa Shuntei's 1898 painting, "Playing Go (Japanese Chess)"

How to quantify the impact of AI on long-run cultural evolution? Published today, I give it a go!

400+ years of strategic dynamics in the game of Go (Baduk/Weiqi), from feudalism to AlphaGo!

Well, this got rejected today as a short position paper from CHR so I suppose I can share it now: arxiv.org/pdf/2508.00095

Here’s a hopefully coherent ramble about what we’re trying to make clear:

Yes please.

I don’t why I’ve felt like there’s a need to understand brains in cogsci because of course you’re right, and I’ve certainly been exposed to some great work in the field that has nothing to do with brains. I think I just like to gatekeep myself out of things 🙃

look dig around in some papers and see if our stuff might fit. I don’t know anything about real brains though. Really, I need to focus on journals generally instead of conferences. It’s always such a dice roll on reviewers.

Your paper has been at the top of my to read pile and I am looking forward to finally getting to it soon! I really do think the rejection is overall reasonable, despite the points we’re making being important for the field. Re cog sci venues, I think I would be a bit of an imposter but I will …

P.S. LLMs will make it even easier to hide our ignorance about what we’re actually doing and what claims we can actually make.

Empirical work can’t justify itself. You need an outer layer of theory in which to do the justification that is non-empirical. Gödel 101. I love the enthusiasm for computing, but maybe look into its foundations and the uncertainty at its core. No model is a view from nowhere. 18/18

Our second reviewer (we only had two) voted strong reject. Their biggest problem with the paper seemed to be the lack of empirical work, and they claimed CHR only wants empirical stuff. If that’s correct, it’s a real blight on an otherwise cool conference. 17/