Elliot Murphy
@elliot-murphy.bsky.social
• Postdoc, Dept of Neurosurgery, UTHealth
• Compositionality in neural and artificial systems
Website: https://elliot-murphy.com/
• Compositionality in neural and artificial systems
Website: https://elliot-murphy.com/
I’ll be presenting new intracranial work this week at #SFN25 🧠
Sunday afternoon poster session (MM1 @ 1pm)
Cortical cascades support rapid semantic inference during reading concrete and abstract words
Sunday afternoon poster session (MM1 @ 1pm)
Cortical cascades support rapid semantic inference during reading concrete and abstract words
November 12, 2025 at 2:40 PM
I’ll be presenting new intracranial work this week at #SFN25 🧠
Sunday afternoon poster session (MM1 @ 1pm)
Cortical cascades support rapid semantic inference during reading concrete and abstract words
Sunday afternoon poster session (MM1 @ 1pm)
Cortical cascades support rapid semantic inference during reading concrete and abstract words
Nice to see that some researchers have definitely not used ChatGPT to write their papers...
www.mdpi.com/2076-3425/15...
www.mdpi.com/2076-3425/15...
September 23, 2025 at 5:02 PM
Nice to see that some researchers have definitely not used ChatGPT to write their papers...
www.mdpi.com/2076-3425/15...
www.mdpi.com/2076-3425/15...
I also think we need to be much more honest, conservative and cautious about our claims here - there are many reasons for scepticism (here is a recent slide from a talk I gave)
August 16, 2025 at 7:07 PM
I also think we need to be much more honest, conservative and cautious about our claims here - there are many reasons for scepticism (here is a recent slide from a talk I gave)
Putting aside compositionality, if GPT-5 still can't draw me 22 bananas I'm not sure what to say!
August 16, 2025 at 7:06 PM
Putting aside compositionality, if GPT-5 still can't draw me 22 bananas I'm not sure what to say!
August 16, 2025 at 6:00 PM
August 16, 2025 at 5:42 PM
A perennial question! But our recent work suggests that DALLE 3 does not reliably capture compositional syntactic representations.
www.sciencedirect.com/science/arti...
www.sciencedirect.com/science/arti...
August 16, 2025 at 5:41 PM
A perennial question! But our recent work suggests that DALLE 3 does not reliably capture compositional syntactic representations.
www.sciencedirect.com/science/arti...
www.sciencedirect.com/science/arti...
Will take a look! On your teddy bear example: we show that DALLE 3 does not successfully represent compositional syntactic meanings.
www.sciencedirect.com/science/arti...
www.sciencedirect.com/science/arti...
August 16, 2025 at 5:37 PM
Will take a look! On your teddy bear example: we show that DALLE 3 does not successfully represent compositional syntactic meanings.
www.sciencedirect.com/science/arti...
www.sciencedirect.com/science/arti...
GPT-5 vs. basic linguistic structure
August 11, 2025 at 10:16 PM
GPT-5 vs. basic linguistic structure
Lastly, in honor of a glaring typo in the later parts of this paper, I present a universal neural rammer (courtesy of DALL-E).
July 14, 2025 at 5:42 PM
Lastly, in honor of a glaring typo in the later parts of this paper, I present a universal neural rammer (courtesy of DALL-E).
ROSE provides a tentative scaffold for a 'Universal Neural Grammar' – a species-specific format for neurally organizing the construction of compositional syntactic structures, which matures in accordance with a genetically determined biological matrix.
July 14, 2025 at 5:42 PM
ROSE provides a tentative scaffold for a 'Universal Neural Grammar' – a species-specific format for neurally organizing the construction of compositional syntactic structures, which matures in accordance with a genetically determined biological matrix.
ROSE offers a 'mesoscopic protectorate' for natural language; an intermediate level of emergent organizational complexity that demands multi-scale modeling. It also offers new constraints on what a unified neurocomputational settlement for natural language syntax might look like.
July 14, 2025 at 5:42 PM
ROSE offers a 'mesoscopic protectorate' for natural language; an intermediate level of emergent organizational complexity that demands multi-scale modeling. It also offers new constraints on what a unified neurocomputational settlement for natural language syntax might look like.
This perspective helps to furnish a more restrictive 'core language network' – dynamics critically involved in executing specific parsing operations (i.e., establishing phrasal categories, tree-structure depth, resolving dependencies, and retrieving proprietary lexico-syntactic representations).
July 14, 2025 at 5:42 PM
This perspective helps to furnish a more restrictive 'core language network' – dynamics critically involved in executing specific parsing operations (i.e., establishing phrasal categories, tree-structure depth, resolving dependencies, and retrieving proprietary lexico-syntactic representations).
I argue that this neural code ("ROSE") provides a possible infrastructure for flexibly implementing distinct types of minimalist grammar parsers for the real-time processing of language. This leads to more refined experimental predictions, and a principled set of relations between parsing demands.
July 14, 2025 at 5:42 PM
I argue that this neural code ("ROSE") provides a possible infrastructure for flexibly implementing distinct types of minimalist grammar parsers for the real-time processing of language. This leads to more refined experimental predictions, and a principled set of relations between parsing demands.
These are proposed to be a delimited set of neural signatures that are hypothesized to constitute a human-specific "neural grammar" that gives rise to our capacity for language.
July 14, 2025 at 5:42 PM
These are proposed to be a delimited set of neural signatures that are hypothesized to constitute a human-specific "neural grammar" that gives rise to our capacity for language.
This paper explores what neural mechanisms could satisfy the demands of the free non-associative commutative magma emerging from contemporary formulations of a Merge-based syntax, where inferences are made over structural (not linear) relations, eg:
[[old men] and women]]
[old [men and women]]
[[old men] and women]]
[old [men and women]]
July 14, 2025 at 5:42 PM
This paper explores what neural mechanisms could satisfy the demands of the free non-associative commutative magma emerging from contemporary formulations of a Merge-based syntax, where inferences are made over structural (not linear) relations, eg:
[[old men] and women]]
[old [men and women]]
[[old men] and women]]
[old [men and women]]
A commentary on Parr, Pezzulo & Friston's (2025) excellent paper, importing some tools from linguistics to help refine their approach.
June 18, 2025 at 6:06 PM
A commentary on Parr, Pezzulo & Friston's (2025) excellent paper, importing some tools from linguistics to help refine their approach.
ChatGPT's o4-mini-high on counting bananas
May 17, 2025 at 8:15 PM
ChatGPT's o4-mini-high on counting bananas
Google’s AI makes up the meaning of fake idioms (or “proverbs”) when you add “meaning” after the end of a random sentence
April 23, 2025 at 1:55 PM
Google’s AI makes up the meaning of fake idioms (or “proverbs”) when you add “meaning” after the end of a random sentence
Mathematical Structure of Syntactic Merge: An Algebraic Model for Generative Linguistics
mitpress.mit.edu/978026255252...
mitpress.mit.edu/978026255252...
April 20, 2025 at 7:29 PM
Mathematical Structure of Syntactic Merge: An Algebraic Model for Generative Linguistics
mitpress.mit.edu/978026255252...
mitpress.mit.edu/978026255252...
With ChatGPT's o1 model (using "advanced reasoning").
March 31, 2025 at 7:46 PM
With ChatGPT's o1 model (using "advanced reasoning").
February 27, 2025 at 7:47 PM
🤔
AGI is… on its way?
AGI is… on its way?
February 22, 2025 at 5:43 PM
🤔
AGI is… on its way?
AGI is… on its way?