While recent progress in artificial intelligence (AI) has relied primarily on increasing the size and scale of the models and computing budgets for training, we ask if those trends will continue. Financial incentives are against scaling, and there can be diminishing returns to further investment. These effects may already be slowing growth among the very largest models. Future progress in AI may rely more on ideas for shrinking models and inventive use of existing models than on simply increasing investment in compute resources.

Concerns over risks from generative artificial intelligence systems have increased significantly over the past year, driven in large part by the advent of increasingly capable large language models. But, how do AI developers attempt to control the outputs of these models? This primer outlines four commonly used techniques and explains why this objective is so challenging.

CSET submitted the following comment in response to a Request for Comment (RFC) from the Office of Management and Budget (OMB) about a draft memorandum providing guidance to government agencies regarding the appointment of Chief AI Officers, Risk Management for AI, and other processes following the October 30, 2023 Executive Order on AI.

AI capabilities are evolving quickly and pose novel—and likely significant—risks. In these rapidly changing conditions, how can policymakers effectively anticipate and manage risks from the most advanced and capable AI systems at the frontier of the field? This Roundtable Report summarizes some of the key themes and conclusions of a July 2023 workshop on this topic jointly hosted by CSET and Google DeepMind.

How can policymakers credibly reveal and assess intentions in the field of artificial intelligence? Policymakers can send credible signals of their intent by making pledges or committing to undertaking certain actions for which they will pay a price—political, reputational, or monetary—if they back down or fail to make good on their initial promise or threat. Talk is cheap, but inadvertent escalation is costly to all sides.

Australia, Canada, Japan, the United Kingdom, and the United States emphasize principles of accountability, explainability, fairness, privacy, security, and transparency in their high-level AI policy documents. But while the words are the same, these countries define each of these principles in slightly different ways that could have large impacts on interoperability and the formulation of international norms. This creates, what we call the “Inigo Montoya problem” in trustworthy AI, inspired by "The Princess Bride" movie quote: “You keep using that word. I do not think it means what you think it means.”

This explainer overviews techniques to produce smaller and more efficient language models that require fewer resources to develop and operate. Importantly, information on how to leverage these techniques, and many of the subsequent small models, are openly available online for anyone to use. The combination of both small (i.e., easy to use) and open (i.e., easy to access) could have significant implications for artificial intelligence development.

Artificial intelligence that makes news headlines, such as ChatGPT, typically runs in well-maintained data centers with an abundant supply of compute and power. However, these resources are more limited on many systems in the real world, such as drones, satellites, or ground vehicles. As a result, the AI that can run onboard these devices will often be inferior to state of the art models. That can affect their usability and the need for additional safeguards in high-risk contexts. This issue brief contextualizes these challenges and provides policymakers with recommendations on how to engage with these technologies.

This data brief uses procurement records published by the U.S. Department of Defense and China’s People’s Liberation Army between April and November of 2020 to assess, and, where appropriate, compare what each military is buying when it comes to artificial intelligence. We find that the two militaries are prioritizing similar application areas, especially intelligent and autonomous vehicles and AI applications for intelligence, surveillance and reconnaissance.

Foundation models could eventually introduce several pathways for undermining state security: accidents, inadvertent escalation, unintentional conflict, the proliferation of weapons, and the interference with human diplomacy are just a few on a long list. The Confidence-Building Measures for Artificial Intelligence workshop hosted by the Geopolitics Team at OpenAI and the Berkeley Risk and Security Lab at the University of California brought together a multistakeholder group to think through the tools and strategies to mitigate the potential risks introduced by foundation models to international security.