You can read the free energy principle in one sitting, if you follow one thread from a hot kettle to a POMDP. Here is that thread, kept short and cited.
1. The thermodynamic intuition
Left alone, a hot kettle equilibrates with the room. The second law says isolated systems drift toward high-entropy macrostates (Class E: expert citation, standard statistical mechanics). A living cell does the opposite. It stays far from thermal equilibrium, holding temperature, pH, and osmolarity inside a narrow viable set. Something is doing work against the drift.
Friston's move is to read that "something" as inference. A system that persists must, in effect, look like it is predicting the states it needs to occupy and acting to keep observations near those predictions (Class E: Friston 2010, "The free-energy principle: a unified brain theory?", Nature Reviews Neuroscience 11, 127 to 138).
2. Surprise, and why we cannot compute it
Formally, self-organizing systems minimize surprise: the negative log probability of an observation under a generative model, minus log p(o). Surprise is not computable in general, because it requires marginalizing over all hidden causes (Class E: Parr, Pezzulo, Friston, Active Inference, MIT Press, 2022, chapter 2).
Variational inference is the standard workaround. Introduce an approximate posterior q(s) over hidden states s, and compute an upper bound on surprise called variational free energy, F. Minimize F, and you have squeezed surprise from above (Class E: Parr et al. 2022, chapter 2). The bound decomposes cleanly into a KL divergence between q(s) and the true posterior p(s given o), plus the log evidence. When the KL term collapses toward zero, q approximates p.
3. The Markov blanket, and where a "self" lives
To talk about a system that has states, you need a boundary. The Markov blanket of a set of internal states is the set of sensory and active states that mediate every statistical dependency between internal states and the outside world (Class E: Parr et al. 2022, chapter 3). Internal states track external states only through the blanket. In this reading, a "self" is not mystical: it is the internal states of a Markov blanket that persist.
This is not a claim about consciousness or agency. It is a claim about a statistical structure that can be written down, tested, and broken (Class F: falsifier present, the blanket assumption can fail in coupled dynamical systems and often does).
4. Perception and action, one objective
Under the free energy principle, perception updates q(s) to reduce F given fixed observations. Action changes the observations themselves to reduce F given fixed beliefs. Two directions, one loss. Expected free energy, the forward-looking cousin used for policy selection, splits into a pragmatic term (how well a policy is expected to reach preferred outcomes) and an epistemic term (how much a policy is expected to reduce uncertainty) (Class E: Parr et al. 2022, chapter 4).
That decomposition is the practical payoff. It is why an active-inference agent balances information-seeking against goal-seeking without a bolted-on curiosity term.
5. From principle to POMDP
To run experiments, the principle is instantiated as a discrete-time partially observable Markov decision process (POMDP). Hidden states, observations, transitions, and preferences are matrices. Variational free energy and expected free energy become tractable to compute (Class C: this is exactly how the labs on this site are configured, per the site's science page and the linked preprint). You can then move dials for sensory precision, transition precision, and policy temperature, and watch behavior change.
This is where UNI sits: on the attainable path toward General Natural Intelligence, natural not artificial. A working hypothesis with growing, evidence-classed evidence, tested in the open. Do not take the claim on faith. Test the build, inspect the gates, and help us find where it fails.
6. Prep for math-hungry readers
If the variational bound and the KL divergence are new, the fastest bridge is the vocabulary of statistical mechanics that surrounds them. Themesis offers a self-paced primer on the terms this literature keeps re-using: T3, Top Ten Terms in Statistical Mechanics for AI. We recommend it as prep for math-hungry learners headed to the UNI workshop (link, not endorsement, not paraphrase).
Keep reading
Fence: this post cites primary literature and describes UNI's public labs. It is not a clinical, diagnostic, or therapeutic tool. The UNI preprint is not yet peer reviewed. Claims are tagged by evidence class. If a claim outside a tag reads stronger than the evidence, that is a bug, tell us.