Computational Mechanics

What can the field of Computational Mechanics, or more broadly the physics of Information, bring to key questions in AI risk safety?

• How can insights from computational mechanics aid in improving the interpretability and explainability of AI models?

• Can computational mechanics offer a mathematical framework to accurately characterize the cognitive abilities and agentic phenomena of advanced AI models, such as long-horizon planning or situational awareness?

Over the last few decades, computational mechanics has rigorously investigated the ultimate limits of prediction, together with the resources and internal representations required to predict optimally. In just the last few years, AI models trained to predict have developed impressive and surprising general capabilities. During this workshop, we explored how computational mechanics can be adapted to understand current and future AI models, with the aim of anticipating behavior and enabling alignment with human values. The program consisted of a mix of tutorials, talks, and ample time for small-group discussions.