Stafford Beer's VSM offers a cybernetic blueprint for AI safety and multi-scale agency

Jonas Hallgren's LessWrong post explores the Viable System Model (VSM) by cybernetician Stafford Beer as a framework for multi-scale hierarchical agency in AI safety. Hallgren translates Beer's five-layer model — originally designed for organizational management — into modern active inference terms. The core principle is Ashby's law of requisite variety: a system (e.g., an AI or a company) must be as internally complex as the environment it tries to regulate. Hallgren illustrates this with a football team example — blocking a team requires a matching opposing team, not static obstacles.

The post argues that dividing complex tasks into simpler sub-goals reduces error rates and enables scalable control. Hallgren admits his interpretation is approximate but believes VSM offers a rigorous foundation for designing AI systems that can self-regulate across multiple scales. He plans future posts expanding on each of Beer's five layers. The work bridges classical cybernetics (which influenced early AI) with contemporary active inference, potentially providing a principled way to build safe, hierarchical AI agents that avoid catastrophic failures.