Andrea Baronchelli

New paper! When AI agents interact, the group can settle on outcomes none of its members would choose alone. And group size can determine which outcome wins. So a model that passes every test on its own can still go wrong at scale.

pnas.org/doi/10.1073/pn…

New preprint: “Indirect tipping: a social attack surface in AI agent populations.”

What if the easiest way to move an AI swarm from A→B is not direct, but A→C→B? We show that stepping-stone equilibria can lower tipping thresholds and unlock otherwise inaccessible transitions.

A pleasure speaking with Frankfurter Allgemeine Zeitung about the social life of AI agents. We discussed how pops of agents develop shared languages and influence one another, with collective behaviour that can't always be inferred from individual agents.
faz.net/aktuell/wissen…

Twenty years ago we published “Sharp transition towards shared vocabularies in multi-agent systems.” It feels unexpectedly current.

New post on AI agents, conventions and what they might enable: baronca.substack.com/p/sharp-transi…

A FAQ on AI populations: what happens when 10, 100 or 1,000 agents interact? What is collective misalignment? Does bigger mean worse?

Size matters.

baronca.substack.com/p/size-matters…