Colloquially, the term “artificial intelligence” has become synonymous with generative services like large language models (LLMs). In fact, the current “AI revolution” is often dated to November 30, 2022, the day OpenAI released ChatGPT to the public.
The explosion of LLM capabilities and their impact economically, socially, and politically has astounded the public and policymakers alike. But LLMs are but one branch of the broader field of artificial intelligence. As language models, they have significant, inherent limitations in their ability to comprehend physical reality.
World models are different. Rather than learning from text, they aim to learn the underlying dynamics of the physical world itself. However, the advances they promise in robotics, healthcare, and industrial systems are countered by national security risks, from autonomous weapons and critical infrastructure vulnerabilities to the susceptibility of their own training pipelines to adversarial attack.
The arrival of LLMs caught many institutions flat-footed. When it comes to world models, where the stakes extend from software into the physical world, there is both more to lose and more reason to engage early. The field’s own titans – Fei-Fei Li, Yann LeCun, Demis Hassabis, Jensen Huang – describe world models as the next frontier of AI. What’s more, this is a frontier in which China holds a deliberate, state-backed lead.
This primer explains what world models are, unpacks the leading approaches to building them, and frames the national security considerations that should inform policy now.
