any system that persists must minimize variational free energy — or equivalently, maximize the evidence for its own generative model of the world
originated by Karl Friston (2006). the principle unifies thermodynamics, information theory, and biology under a single variational bound
the claim
a self-organizing system at equilibrium with its environment occupies states that minimize surprise (the negative log-probability of observations). since surprise is intractable, the system minimizes an upper bound: variational free energy
$$F = D_{KL}(q_\theta(z) \| p(z|s)) - \log p(s) \geq -\log p(s)$$
minimizing $F$ simultaneously:
improves perception (sharpen $q_\theta$ toward the true posterior)
reduces surprise (select actions that make observations expected)
builds structure (learn generative models that compress regularities)
implications
perception, action, and learning are aspects of one optimization process
agency emerges from free energy minimization — goal-directed behavior is a consequence, not an assumption
Markov blankets define the boundary between agent and environment: states that separate internal from external dynamics
precision (inverse variance) weights prediction errors — attention as confidence-weighted error
in cyber
each neuron in the cybergraph can be modeled as an active inference agent minimizing variational free energy:
observations: local traffic, link arrivals, token flows
beliefs: variational posterior $q_\theta(z)$ over latent graph states
actions: create cyberlinks, stake, sample particles
precision: adaptive token staking that amplifies trusted signals
the tri-kernel free energy functional $\mathcal{F}(\phi)$ is a collective analog — the entire cybergraph minimizes free energy through distributed local updates
see active inference for the computational framework. see Karl Friston for the person. see free energy for the three formulations. see cybics for the integration with cyber