Life beyond the free energy principle: how to survive without invariance (www.dialecticalsystems.eu)

This piece argues that Karl Friston’s free energy principle (FEP) — which reframes existence as maintaining invariant statistical properties and minimizing surprisal via variational free energy — overreaches when presented as a universal ontology. Technically, the FEP rests on assuming an invariant joint probability distribution over a system’s variables, from which notions like surprisal, Markov blankets, and “agents” are derived. Critics point out this move is often tautological: stability framed as “avoid surprises” only applies if you already posit the probability structure to be conserved. Mathematical ingredients invoked (e.g., Langevin stochastic differential equations to capture fixed-frequency fluctuations) show how broad the formalism can be, but they don’t explain why particular decompositions (agent vs environment) or model choices should be privileged. Empirical examples such as rapid membrane turnover in cells (or slime mould) further expose difficulties in identifying stable individuating features for living systems. For the AI/ML community, the takeaways are practical and conceptual. FEP remains a powerful modeling framework for active inference and embodied cognition, but its claims shouldn’t be yoked to metaphysical conclusions about intentionality or life. Modelers should treat Markov blankets and variational approximations as modeling choices, not derived truths, and be cautious when using free-energy arguments to justify agentive or normative behavior in agents. In short: free energy minimization is a useful one‑way to describe stability, not a universal explanation that obviates the need for explicit, mechanistic models in AI and cognitive science.