Simple mistakes in LLM reasoning might suggest that models are *pattern matching*, not reasoning. But what if people are doing the same thing? In my preprint with @glupyan, we find that LLMs are highly aligned with people *because* they are pattern matching! A thread🧵(1/15):

Show Post

Our lab is growing! We're looking for a graduate student to start next fall (applications due on Dec 1) and a postdoctoral researcher to start soon (ad coming soon; starting date is flexible, probably some time next year). subjectivitylab.org

Show Post

Thanks so much for sharing our work, Chaz! Really glad you found it interesting! Also, my colleagues brought back some coasters from your lab this year, and we all loved them!

Show Post

Preprint: www.biorxiv.org/content/10.6...

Show Post

So, rotation tolerance may have an important caveat: tolerate the transformation, until the image starts looking like something else. Visual anagrams give us a way to probe where object representations remain stable, and where they begin to reorganize around another interpretation. (7/7)

Show Post