Elliot Murphy

Final version is out:

authors.elsevier.com/c/1mEoa5bD-s...

"Neural cellular automata: Applications to biology and beyond classical AI"

@bhartl.bsky.social
Léo Pio-Lopez

The language network responds robustly to sentences across diverse tasks A network of left frontal and temporal brain areas supports language comprehension and production, implementing computations related to word retrieval and combinatorial linguistic processing. Here, we...

A left frontal-temporal network selectively supports language comprehension and production. Are computations in this language network driven primarily by bottom-up input, or by top-down task demands?
🧵👇

www.biorxiv.org/content/10.6...

Interesting question. We only talk about meta-linguistic cognition in one paragraph, kind of tangential to our focus but overall still not convinced these systems have anything remotely like it.

The paper title is a bit ambiguous here. All of our recent papers focus on the issue of modelling the competence of an ordinary language user. Only a couple of our later prompts in the final part of the paper go into exotic territory like “draw me a fancy tree” etc.

The Discussion section includes a long engagement with many recent arguments in the literature about the supposed abilities of LLMs.

Our results support recent hypotheses concerning the ability of language models to represent ‘horizontal’ linguistic information, but their significantly reduced ability to represent ‘vertical’ types of hierarchical compositional syntax-semantics (constituency structure).

These results help emphasize an apparent lack of meta-linguistic understanding (contra various claims). For LLMs, language simply *is* the system it is trying to master, whereas for humans language is exploited as a powerful cognitive and inferential tool.

It was unable to generate a Jabberwocky structure that accurately represented a clear syntactic violation. It was unable to accurately assess the output of applying two distinct syntactic violations to a sentence. It was unable to represent semantically impossible entities.

ChatGPT's o3 fails to cleanly dissociate conceptual content from structural configuration – a basic requirement of compositional syntax. It provides surreal meanings instead of truly ungrammatical sentences.

We tested dependencies, phrase structure constituency, Escher sentences, center-embedding, semantic paradoxes, the generation of multiple syntactic violations, acceptability assessments, scope, Jabberwocky, and semantically 'impossible' objects.

New paper just published with @evelinaleivada.bsky.social @garymarcus.bsky.social, Vittoria Dentella, Raquel Montero and Fritz Günther

Fundamental Principles of Linguistic Structure Are Not Represented by ChatGPT

bioling.psychopen.eu/index.php/bi...

How it started (quote post)
How it's going (linked article)

www.theguardian.com/technology/2...

My rough attempt at sorting people on this spectrum of epiphenomenal to causal...

My rough attempt at sorting people on this spectrum of epiphenomenal to causal.

**How distributed is the brain-wide network that is recruited
for cognition?**
That goes to the top of the list!
#neuroskyence
doi.org/10.1038/s415...

How children map causal verbs to different causes across development - Nature Human Behaviour This study shows that, by age 4, children understand lexical causatives to refer to direct causes and periphrastic causatives to indirect causes in causal chains. Understanding causation by absence develops later in older children.

By age 4, children understand lexical causatives to refer to direct causes and periphrastic causatives to indirect causes in causal chains. Understanding causation by absence develops later in older children. @tobigerstenberg.bsky.social

Extension Condition "violations" and Merge optimality constraints We analyze, using the mathematical formulation of Merge within the Strong Minimalist Thesis framework, a set of linguistic phenomena, including head-to-head movement, phrasal affixes and syntactic cli...

Extension Condition "violations" and Merge optimality constraints

arxiv.org/abs/2511.225...

I missed an opportunity to mess with AI on meta-Lynchian levels not yet seen before

Another highly educational conversation with ChatGPT

New Correspondence with @davidpoeppel.bsky.social in Nat Rev Neurosci. www.nature.com/articles/s41...

Here, we critique a recent paper by Rosas et al. We argue that "Bottom-up" and "Top-down" neuroscience have various meanings in the literature.

PDF: rdcu.be/eSKYI

I have a feeling we’re going to see Waymo of this

Figure 1: The important dimensions of CAILs across research and education; clockwise from 12 o’clock: Con- ceptual Clarity is the idea that terms should refer. Critical Thinking is deep engagement with the relationships between statements about the world. Decoloniality is the process of de-centring and addressing dominant harmful views and practices. Respecting Expertise is the epistemic compact between professionals and society. Slow Science is a disposition towards preferring psychologically, techno-socially, and epistemically healthy practices. The lines between dimensions represent how they are interwoven both directly and indirectly.

New preprint! @marentierra.bsky.social @irisvanrooij.bsky.social & I have been working on what CAIL means to showcase & propagate the idea of thinking very differently to tech industry norms on "artificial intelligence"

Towards Critical Artificial Intelligence Literacies doi.org/10.5281/zeno...

1/

🥳

Whole-Brain Millisecond-Scale Effective Connectivity Atlas of Auditory and Visual Naming Neurobiological models suggest that speech relies on interactions among distant cortical regions interconnected by white matter pathways. Prior studies have shown that speech-related functional coacti...

An incredible resource — open dataset of 104 intracranial EEG subjects who did both visual and auditory naming 🧠 😲

Whole-Brain Millisecond-Scale Effective Connectivity Atlas of Auditory and Visual Naming

www.biorxiv.org/content/10.1...

Language, Thought, and Dead-Metaphor Machines A commentary on recent commentaries, and thoughts on reactive science and philosophy.

I began writing a quick take yesterday and, as happens, I just kept writing.

This is my take on the virtues and risks of using cognitive science and philosophy of mind to respond to - particularly industry-driven - developments in LLMs.

vincentcarchidi.substack.com/p/language-t...

Causal inference and the metaphysics of causation - Synthese The techniques of causal inference are widely used throughout the non-experimental sciences to derive causal conclusions from probabilistic premises. This poses a philosophical question. What in the n...

New paper out in Synthese—a metaphysical theory of causation designed to explain statistical causal inference. Basically a summary of a 100,000-word book (Causation: Science, Statistics and Metaphysics) that'll go into production with CUP soon. Watch this space link.springer.com/article/10.1...

𝗪𝗵𝗮𝘁 𝗮𝗿𝗲 𝗯𝗿𝗮𝗶𝗻 𝗻𝗲𝘁𝘄𝗼𝗿𝗸𝘀?
Do they even exist...?
Join @lucinauddin.bsky.social who will present, followed by discussion in the Neuroscience & Philosophy Salon.
Open to all.
Date: Dec 9, 12pm EST-US
Register: umd.zoom.us/meeting/regi... (you need a zoom account which is free)
#neuroskyence

Brilliant eye-tracking study on statistics/structure 👀

"We conclude that linguistic knowledge guides moment-by-moment reading in two dissociable ways: Forward reading is driven by the word’s surprisal, and backward reading reflects syntactic structure building operations."

osf.io/preprints/ps...

CAIL = prerequisite knowledge for a critical perspective, such as to tell apart nonsense hype from true theoretical computer scientific claims. For example, the idea that human-like systems are a sensible or possible goal is the result of circular reasoning and anthropomorphism. olivia.science/ai

🧵