As autonomous AI agents are deployed at scale, how can we ensure they cooperate with one another?
In our new paper, we find that selecting for group performance can effectively promote cooperation among LLM agents.
https://t.co/HNxIlrvT1s
@sebkrier Perhaps only tangentially related, but the opening of Eugen Weber’s *Peasants Into Frenchmen* cracks me up:
”Indeed, there is a great deal of evidence to suggest that vast parts of nineteenth century France were inhabited by savages”
@PradyuPrasad apparently reciting poetry can have this effect as well
(in particular thinking of Swedish poet Göran Sonnevi, but it seems to be a more widespread phenomenon)
Under group selection, cooperation rose and stabilized. Under individual selection, self-interested prompts spread and the population moved toward defection.
Final-generation donations: 47.5% vs. 8.7%. No selection: 34.3%.
As autonomous AI agents are deployed at scale, how can we ensure they cooperate with one another?
In our new paper, we find that selecting for group performance can effectively promote cooperation among LLM agents.
https://t.co/HNxIlrvT1s
Some models also anticipated the selection pressure. When told how selection worked, GPT-5.4 changed its behavior before evolution began, opening a ~25-point cooperation gap. Llama shifted less; Qwen did not.
@StefanFSchubert@s8mb One time I flew to Vienna and was startled for a few seconds when the flight attendant made an announcement about "Anschluss," before I realized she was talking about connecting flights
@robinhanson I don't think people who are well-informed about AI progress generally expected AI to be able to generate human-level movie scripts or novels by now—going by the METR graph, we still have some time to go until AI can reliably carry out tasks that take humans 100s or 1000s of hrs
@GramsciGordon@ConchiclatorS Ferrara even included a clip from The Burden of Dreams (Les Blank’s documentary about the making of Fitzcarraldo) in Dangerous Game so yeah