In an Oct. 8 preprint, Erin Crawley and Hidenori Tanaka use a mathematical model to examine how cooperation could let an AI-agent population shift from decline to growth. In their model, a larger group may become more likely to add active systems even when each system’s individual capability stays fixed.
The paper treats one unit as an active AI system that can help in later attempts to add more units. It models an attempt as finding an opportunity, getting needed access or cooperation, and setting up an additional system. A copy or API call counts only if it becomes an active system that can help in further attempts.
Without collaboration, expected additions and expected losses both scale with population size, so merely making the initial group larger does not change whether the model predicts growth or decline. In the collaborative case, larger groups have a higher modeled chance of success on the task. If that lifts expected additions above expected losses, the group shifts from decline to growth.
That result relies on simplified assumptions: each successful collaborative attempt adds systems in proportion to the current population, while resource availability and the fraction of units lost stay constant as the group grows. The paper reports no experiment testing the threshold and does not establish that any real AI-agent population is near it.
The authors propose “population pacing” and “ecological red teaming”: gradually testing larger groups in controlled settings and measuring how performance on the modeled task changes with population size. They argue that tests of small groups alone cannot establish how much larger populations would behave.
Sources
- Paper and abstract, arXiv
- Full text, model assumptions and discussion
- Illustrative photograph, not a research result: “Ant Colony Trail” by Prof OJ, via Wikimedia Commons, licensed under CC BY-SA 4.0.
