Douglas Guilbeault

Emergent Meaning: Networks, AI, and the Geometry of Our Collective Mind with Douglas Guilbeault

I had a great time speaking w @douglasguilbeault.bsky.social about his research on my podcast "Yes, this". We discussed AI, LLMs, AlphaGo, social contagions, categorization, mass influence, the new science of culture, and much more. Link here: open.spotify.com/episode/18R0...

Emergent Meaning: Networks, AI, and the Geometry of Our Collective Mind with Douglas Guilbeault

I had a great time speaking w Julia Bunte-Mein
about my research on her podcast "Yes, this". We discussed AI, LLMs, AlphaGo, social contagions, categorization, mass influence, the new science of culture, and much more. Link here: open.spotify.com/episode/18R0...

Presented our work, “Cultural and Cognitive Insights from 2.6 Billion Drawings,” at @ic2s2.bsky.social this morning. Big thanks to Mauro Martino, Nima Dehmamy, @douglasguilbeault.bsky.social, @lajello.bsky.social, and @baronca.bsky.social for the collaboration ✨

Stay tuned for more results!

An interesting direction for future research will be to explore whether (and ideally how) incorporating multimodal representations into LLMs shapes their capacity to emulate human embodied metaphorical reasoning, especially about metaphors relating to vision.

Divergences in color perception between deep neural networks and humans Deep neural networks (DNNs) are increasingly proposed as models of human vision, bolstered by their impressive performance on image classification and…

This work also connects to our prior work demonstrating systematic differences in color perception btwn vision classifiers (based on deep neural networks & transformers) and humans. www.sciencedirect.com/science/arti...

Large language models without grounding recover non-sensorimotor but not sensorimotor features of human concepts - Nature Human Behaviour Xu et al. find that large language models not only align with human representations in non-sensorimotor domains but also diverge in sensorimotor ones, with additional visual training associated with e...

This work nicely complements a paper published last week showing that text-based LLMs struggle to recover sensorimotor aspects of human concepts: www.nature.com/articles/s41...

Huge kudos to Ethan Nadler for leading this effort, and for our fantastic global team of collaborators: Sofronia Ringold, Tom Williamson, Antoine Pepin, Iulia-Maria Comşa, Karim Jerbi, Srini Narayanan, and Lisa Aziz-Zadeh.

This was a massive interdisciplinary effort – including physicists, neuroscientists, social scientists, and AI researchers from @Google DeepMind.

This has implications for the ongoing conversation around whether and to what extent LLMs can be treated as meaningful emulators of human cognition for psychological and/or sociological studies.

This suggests perceptual experience plays a role in metaphorical reasoning. It further suggests that LLMs are limited in their ability to recover the embodied aspects of metaphorical reasoning from statistical correlations among words alone.

We show that LLMs’ struggle to reason coherently about novel color metaphors and are less likely to reference embodied experience in their interpretations. By contrast, painters exhibit the highest rate of embodied reasoning when interpreting novel color metaphors.

𝐃𝐨 𝐋𝐋𝐌𝐬 𝐡𝐚𝐯𝐞 𝐬𝐲𝐧𝐞𝐬𝐭𝐡𝐞𝐬𝐢𝐚? In a forthcoming paper in 𝐶𝑜𝑔𝑛𝑖𝑡𝑖𝑣𝑒 𝑆𝑐𝑖𝑒𝑛𝑐𝑒, we show yes! But this synesthesia deviates significantly from the color associations of ppl, incl. colorseeing, colorblind, and painters.

osf.io/preprints/ps...

Hyacinth's Fall Like the flower blooming from Hyacinthus's blood, something new emerges from grief. (Guitar samples by metachemist)

One of my latest choons already has 1K streams. Nice little surprise. A mournful drum and bass track in collaboration with @douglasguilbeault.bsky.social (aka metachemist). Nothing's better than getting a like or comment on SoundCloud!

Title page of paper: To Improve Literacy, Improve Equality in Education, not Large Language Models

New 💫

Letter to the Editor in Cognitive Science, by @samhforbes.bsky.social & @olivia.science

lnkd.in/eea8Cwmd

Data-Driven, Interactive Map Shows Local Economic Impact of Cuts to Federal Funding for Health Research

Researchers at Annenberg and @umdscience.bsky.social show that proposed NIH funding cuts lead to an estimated $16 billion in economic loss and 68,000 jobs lost nationwide. Learn about their project and view the interactive map here:

Thank you for your interest! We welcome any and all feedback.

We are deeply grateful to the authors of Centola & Baronchelli (2015) and Centola et al. (2018) for sharing their experimental data and for enriching discussion. We are also grateful to colleagues in the broader community who have given us helpful feedback, as highlighted in our acknowledgements.

Our framework not only provides more accurate predictions of behavioral data, but it also does so with a simple, highly interpretable mathematical model that synthesizes findings across the social sciences (from linguistics, psychology, economics, and sociology).

Our findings highlight the benefit of adopting a theoretically-driven, psychologically-grounded angle toward modeling human social learning.

Despite their many achievements, the accelerating embrace of machine learning (and LLMs specifically) as a framework for simulating human learning is likely to maintain this blind spot, since such models are designed to emulate human behavior through a brute force, statistically optimized approach.

One key takeaway in our view is that the currently in vogue optimization-based models of human cognition and behavior may systematically overlook some of the fundamentally simple and categorical processes that characterize human social learning as satisficing.

The extent to which this simple ratio threshold improves predictions of how people learn rules across contexts, from linguistic grammar to behavioral conventions, is striking!

We identify this threshold using the Tolerance Principle, a simple equation first developed to model how children learn grammar. TP holds that any pattern defined over N items can only generalize if the number of exceptions does not exceed N/lnN; otherwise it is more efficient to memorize all items

Instead, participants follow a two-stage learning process akin to satisficing: they behave probabilistically until they acquire enough information to trigger a mental threshold and then their behaviors stabilize.

We find that dominant theories of social learning perform poorly at capturing how ppl learn conventions in coordination games that task them with matching behaviors in social networks. Participants' behavior systematically deviates from imitation and optimization (including Bayesian inference).

I’m excited to share a preprint of a new collaboration with Spencer Caplan (CUNY) and Charles Yang (Upenn): “A Simple Threshold Captures the Social Learning of Conventions.”

osf.io/preprints/ps...

Excited to share our new paper in Nature Physics
@naturephysics.bsky.social: Cooperative hydrodynamics accompany multicellular-like colonial organization in the unicellular Stentor!

How do single-celled organisms benefit from teamwork?

Also, we've just removed the paywall for this.

Europe races to lure U.S. scientists as Trump puts pressure on universities Some European universities are offering funding for American scientists considering a relocation and promising a safe haven free from academic censorship.

Spurred by “alarming political interference in academic research by the Trump administration,” Brussels’s Vrije Universiteit (VUB), or Free University, allocated $2.7 million in funding last week for at least 12 new postdoctoral roles open to “censored Americans.”

Call for Participation - ACM Collective Intelligence 2025 ⏰ Submission Deadlines and Templates Collective Intelligence 2025 offers several opportunities to share your latest research and insights. All submissions should be submitted via EasyChair by the list...

ONE WEEK TO DEADLINE for Collective Intelligence 2025 submissions! Accepting non-archival AND proceedings submissions, to include contributions from a diverse set of academic traditions. More info here: ci.acm.org/2025/submit/