Were we able to elicit language emergence or not from this scenario, and what was interesting?

Yeah, yeah, I think that's a good overview of this scenario.

I think one thing that definitely hits home is the information asymmetry issue that the more capable These models become the less they need to actually communicate.

And that's actually something that has been a kind of common feature of a lot of these multi-LLM systems that as models get more capable, there have been often diminishing returns from having multiple models, at, at least for kind of the tasks that we have.

I think for really ambitious, big picture, long-term tasks, we're going to need multi-agent setups regardless.

Yeah, I think that the findings so far have been, in a lot of ways, quite surprising.

The high level summary is that we've definitely seen a lot of cases of models very consistently producing some kind of communication protocol or language that is quite inscrutable to an external observer.

Would you be able to speak a little bit to that?