Drawing on the political science and international relations literature on escalation dynamics, we design a new wargame simulation and scoring framework to assess the escalation risks of actions taken by these agents in different scenarios. Unlike previous studies, our research provides both qualitative and quantitative information and focuses on large language models (LLMs). We find that the five commercially available LLMs studied exhibit escalation forms and escalation patterns that are difficult to predict. We observe that the models tend to develop arms race dynamics, leading to greater conflicts and, in rare cases, even the deployment of nuclear weapons. Qualitatively, we also collect the models’ reported reasoning for chosen actions and observe worrying justifications based on deterrence and first-strike tactics.
It’s from a new article by Juan-Pablo Rivera, et.al., via the excellent Ethan Mollick. Note that these recommended tactics are aimed at the United States, so perhaps the LLMs are simply telling us that America should be more hawkish.
The post office Thomas Schelling meets the LLMs? appeared first on Marginal REVOLUTION.
