They make LLMs play a very abstract game that rewards them points from answering the same as the other, and punishes them from answering differently, and LLMs tend to converge to an answer. From that to "social conventions" there is a long, long stretch. The paper lacks a baseline - wouldn't much simpler (non-LLM) systems also exhibit the same property? Is it that surprising that systems that are clones of each other (because they didn't even try "mixed societies" of different LLMs) agree when you give them points for agreeing?

Maybe I'm missing something but in my view this is pure hype and no substance. And note that I'm far from an LLM skeptic and I wouldn't rule out at all that current LLMs could develop social conventions, but this simulation doesn't really show that convincingly.

As obviously silly as this is, could it actually be useful for getting the observed phenomena acknowledged by people who might just tune out is presented with the underlying math that makes it individually?

How much overlap is there between people who think llms are magic and people who don't approve of applying math to groups of people? And how many in the overlap have positions where their opinions have outsized effects?

Crazy - this is saying agents develop their own biases and culture through their engagement with one another even if the individual agents are unbiased.

I wonder when we will see LLMs being used to test economic theories.

It’s funny how we seem to be on this treadmill of “tech that uses GPUs to crunch data” starting with the Bitcoin thing, moving to NFTs, now LLMs.

Wonder what’s next.

Oh I thought this was going to be about the cult like in group signaling of LLM advocates, but this is a thing imagining language patterns as a society instead of language patterns of a society being a bias that the models have.