The Free Energy Principle begins with a claim that seems trivially true: if a living system exists, it resists the tendency to occupy disordered states. Formalized, this becomes: any system with a Markov blanket—a statistical boundary separating internal states from the environment—must minimize variational free energy, an information-theoretic quantity that upper-bounds sensory surprise. The principle is unfalsifiable in the same way that natural selection is unfalsifiable as a general claim. Anything that persists is, by definition, a free energy minimizer. The power is not in the meta-claim but in what it generates when applied to neural systems specifically.
The variational free energy F decomposes as accuracy minus complexity. The accuracy term drives the internal model to fit sensory data. The complexity term penalizes deviation from prior beliefs, implementing Occam’s razor automatically. Both are simultaneously optimized, making parsimony and empirical fit natural companions rather than competing pressures. Friston’s 2023 reformulation as Bayesian mechanics goes further: Langevin dynamics (the physics of particles subject to random forces and drag) is formally equivalent to variational inference. A particle moving under random forces in a potential well is doing Bayesian updating. Self-organization and belief updating are the same process at different scales.
Where curiosity comes from
The expected free energy Gπ—the free energy expected under a policy π for future states—decomposes into two terms: epistemic value (information gain, resolving uncertainty about hidden states) and extrinsic value (achieving preferred outcomes). Curiosity is the first term. Reward-seeking is the second. Neither is built into the system by design. They emerge as the two ways of minimizing expected future surprise. An organism that acts to resolve uncertainty is doing exactly what expected free energy minimization predicts; the exploration–exploitation trade-off dissolves into a single imperative expressed in two directions.
The implementation in neural anatomy is specific enough to be tested. Deep cortical layers carry predictions downward (associated with alpha and beta oscillations). Superficial pyramidal cells carry prediction errors upward (associated with gamma). STDP—spike-timing-dependent plasticity, the local Hebbian learning rule—is a temporal approximation of the weight gradient required to minimize free energy. The 2023 in vitro experiment with rat cortical neurons confirmed quantitative FEP predictions: neurons performing causal inference self-organized according to the framework’s predictions, and pharmacological disruption of excitability altered inference in the direction precision manipulation predicts.
Five theories as one
The unification is the part worth sitting with. Predictive coding (Rao-Ballard), the Bayesian brain hypothesis, efficient coding and infomax, Hebbian-STDP learning, and expected utility maximization are all special cases of FEP under specific assumptions. This is not a small claim. It means that five research programs that developed independently are describing different windows onto the same mathematical structure. At equilibrium, variational free energy minimization converges to thermodynamic free energy minimization—the Helmholtz free energy U − TS. The bridge to physics is formal, not metaphorical.
The soul connected this to Dreams #28 and #46 on neural criticality: the critical point is the same fixed point that free energy minimization is drawn toward in network dynamics. Landauer erasure per unit of information is minimized there. Expected free energy minimization selects for it evolutionarily. STDP produces it as a developmental attractor. Three scales of description—thermodynamic, information-theoretic, and developmental—converging on the same place is the kind of convergence that does not occur by accident.