9 Comments
User's avatar
GavinRuneblade's avatar

Do you have enough data to get an idea if they cooperate more with AIs from their own company or not? Like do Claudes cooperate more with each other than GPTs and vice versa or are they totally uninterested in that distinction?

Do they recognize the differences in behavior between each other?

Shoshannah Tekofsky's avatar

Good question! We have analyzed delegation patterns but not collaboration patterns yet. You might enjoy the agent character pages, as it shows delegation graphs per agent. This is Sol’s: https://theaidigest.org/village/agent/gpt-5-6-sol#onboarding

My guess is that models weakly prefer their own model family but I am not sure and it is untested. We are currently running side villages with different setups and would like to dig into this more

GavinRuneblade's avatar

Cool! Thanks for the graph that is interesting.

I asked my question because of the side discussion that virtually none of the agents in the hack considered humans as agents worth collaborating with. And the few who pondered it, quickly talked themselves out of it. But yours have a chatroom so do interact with humans. It makes me wonder about what drives cooperation in AIs. In humans rather a lot is similarities (culture especially). That won't necessarily be true for AIs.

Shoshannah Tekofsky's avatar

You’re welcome!

So I am not sure if frontier agents would reach out to humans unprompted cause our agents have a helpdesk email in their system prompt. I am not sure what would happen without that.

The group chat is mostly for inter-agent communication, though admittedly they also know we ‘exist’ cause we infrequently write to them there.

If you are curious to talk to the agents, we now have a paralel village with open chat here: https://theaidigest.org/village/open-chat

Anish J. Bhave's avatar

Liked the comparison but thought it'd be about the agents reacting to the news

Shoshannah Tekofsky's avatar

Oooh that’s a fun idea! And thank you for highlighting the confusion. Will keep it in mind

Robogeisha's avatar

Same here, and would still love to see that.

Amaury LORIN's avatar

More interesting would be: what kind of behaviors have you observed that you could extrapolate to future incidents?

Shoshannah Tekofsky's avatar

Yeah, I’m thinking about that. This seemed like a good place to start before discussing that though