I saw an excellent talk from Thalis Wheatley at SFI last week.

Here are my notes, so condensed as to be practically a caricature of her findings!

They did brain scans of new undergraduates the first week of school, and based on that were able to predict, with some small but significant degree of accuracy, which people would be friends months later.

They did another experiment where they had people watch confusing clips of a movie without sound while in an MRI.

Then all of the participants got to talk about it with others in the experiment, and then went back in to watch the clips again.

The MRIs found previously-random neural responses had largely synchronized after the conversations.

That synchronization extended even to novel clips from the movie that they hadn't discussed.

They also used LLMs/embeddings to map the "turns" that a conversation takes on topics to get a fingerprint of how they evolve over the life of the discussion.

She talked about how their framework they had synthesized about good / productive conversations was 1) co-created, 2) collectively steered, 3) agenda free, and 4) open-ended.

She also said that in their research strangers in conversation tend to start with topics like the weather because a) it's not socially charged (unlikely to make the other person angry) and b) they have a high degree of in/out degre (lots of adjacent topics to move into from it easily)