Both Marcel Binz and Eric Schulz spoke in our CCCM seminar series “The Cognitive Science of generative AI” about their work with Centaur - an #LLM based model/modelling tool for modelling human behaviour - their paper is now out in Nature
Discussion
Loading...
@UlrikeHahn @grimalkina would love your take on this one.
@mayintoronto @UlrikeHahn I'm excited to dig in! but generally speaking find this work very interesting and much more depthful than shallower work of the sort that's like, "just use current commercial llm products to generate human faux survey responses". I imagine many are v skeptical on this (here's an example https://osf.io/preprints/psyarxiv/v9w37_v2 ). Paper makes big future-facing claims I'm not convinced of but for me the idea of developing better predictive/interpretable models is a good/needed thing
@mayintoronto Just answering because you asked; I have a bit of an outsider's view on this vs a really informed one so will have to read and consume in depth! Thank you @UlrikeHahn for the talk links too!
@mayintoronto @grimalkina it’s in the thread below and in the older linked threads ;-)
2/ There is also a write-up in the New York Times
3/ and here is. link to both their talks and some discussion - both well worth watching. Marcel Binz thread….
4/ and the Eric Schulz talk and thread
@cogsci @philosophy 5/ This work with Centaur is game changing for the study of human behaviour for multiple reasons: the most important one (imo) being that the LLM based model has much higher predictive accuracy than the popular cognitive models of these tasks that it is compared to. That matters because it means that LLMs can be used to give us a better estimate of HOW MUCH of the behaviour we could be predicting with better models!
@cogsci @philosophy 6/ that is hugely significant *as a tool* regardless of what one thinks of the LLM model itself.
Furthermore, as Marcel Binz talk sets out, they have also used LLMs to identify better *interpretable* cognitive models, ie models that use the same building blocks as traditional cognitive models in the domain
I suspect this work will irrevocably change the cognitive modelling of human behaviour