A Compositional Philosophy of Science for Agent Foundations

Researcher Jonas Hallgren has published 'A Compositional Philosophy of Science for Agent Foundations,' a methodological framework for tackling AI safety's most complex challenges. Published on LessWrong and originally from equilibria1.substack.com, the 15-minute read addresses the field's fundamental difficulties: AI safety spans mathematics, philosophy, and governance while lacking established paradigms. Hallgren describes the common researcher experience of 'confusion for 2 years' before patterns emerge, noting that existing research plans from figures like Davidad and Wentworth remain hard to verify or follow. His solution draws from cultural evolution theory, particularly Joseph Heinrich's work in 'The Secret of Our Success,' which posits that human progress is combinatorial rather than driven by individual genius.

The framework treats ideas as modular components and emphasizes connection-making between different fields as the primary engine of progress. Hallgren argues this approach is especially valuable for 'pre-paradigmatic' fields like AI safety where questions and connections aren't fully formed. The philosophy acknowledges the field's complexity—what he calls 'philosophy complete' problems requiring insights across disciplines—while providing a structured way to generate new questions. The post itself demonstrates the method, with 10-15% AI-assisted restructuring via Claude to improve coherence. This represents a meta-approach to research methodology that could influence how AI safety teams organize knowledge and prioritize interdisciplinary work.