If you want to build systems that behave sensibly under uncertainty, active inference is the most concrete map on the table. This is our working sketch of the territory, written for people who intend to build something and want to know exactly what has been demonstrated and what has not.
UNI is a working hypothesis on an attainable path toward General Natural Intelligence: a natural, active-inference approach whose evidence is growing, evidence-classed, and tested in the open. Do not take that claim on faith. Test the build, inspect the gates, and help us find where it fails.
The free energy principle, in one careful paragraph
In Parr, Pezzulo and Friston's synthesis Class E, a self-organising system that persists over time can be described as one that minimises a quantity called variational free energy. Free energy in this sense is not thermodynamic. It is an information-theoretic upper bound on surprise, the negative log probability of the sensory data the system encounters, evaluated against the system's internal generative model of the causes of those data. Minimising this bound is mathematically equivalent to doing approximate Bayesian inference. So a system that keeps itself in a viable set of states can be read as a system that is, in effect, performing inference about the world in order to stay alive.
The framework asks two things of a builder. First, name the generative model, the joint distribution over hidden states and observations that the agent implicitly assumes. Second, name the boundary, the Markov blanket that separates internal states from external states via sensory and active states. Everything else, perception, learning, planning, control, is derived from those two commitments.
Perception as inference
Under active inference, perceiving is not a passive readout. It is the update step of approximate Bayesian inference: given a prior belief over hidden states and a likelihood mapping states to observations, the agent computes an approximate posterior Class E. The Kullback, Leibler divergence between the approximate posterior and the true posterior is one component of the variational free energy that perception is minimising.
In a POMDP formulation (which is what our labs use Class B), this looks concrete. The agent holds a belief over discrete hidden states, receives a discrete observation through a likelihood matrix, and updates that belief. Sensory precision, the confidence weight on incoming observations, is a dial the agent (or the operator) can turn. Turn it up and observations dominate the prior. Turn it down and prior structure dominates. Behavioural regimes are not hand-coded; they emerge from where those dials are set.
Action as inference
Action is treated symmetrically to perception. Instead of picking actions by maximising an external reward, the agent picks the policy whose expected free energy is lowest Class E. Expected free energy has two components. The pragmatic term rewards policies whose predicted observations align with the agent's prior preferences over outcomes. The epistemic term rewards policies expected to reduce uncertainty about hidden states of the world.
This decomposition is why active-inference agents show what looks like curiosity without a bolt-on exploration bonus. When a state is ambiguous, actions that would resolve it become attractive because they reduce expected free energy. When the world is well understood, pragmatic terms dominate and the agent commits. Exploration and exploitation are two weights inside the same objective, not two separate systems.
What UNI has actually built, and where the gates sit
We keep the language narrow on purpose. Here is our current state, with claim language matched to evidence class.
- Implemented and inspectable. Five browser-runnable POMDP labs, each with steerable precision dials and a documented generative model Class B. A pre-registered Cell Lab benchmark whose five falsification criteria were fixed before the runs, with wins and losses recorded together Class B. A public Model Context Protocol server that lets any language model call the labs Class B.
- Grounded in the standard reference. The POMDP formulation follows Parr, Pezzulo and Friston (2022) chapter by chapter, with the mapping documented in the Zenodo preprint (DOI 10.5281/zenodo.19785799) Class E.
- Not yet peer reviewed. The preprint is unrefereed. Cite it as a preprint. Expert Layer 2 review is pending.
- Not clinical, not diagnostic. The behavioural labels the labs surface are candidate computational phenotypes, hypotheses about how precision settings map to behaviour, not clinical diagnoses.
- Falsifiers, live. The Cell Lab already shows losses: neural baselines outperform on memory_leak and cpu_noisy_neighbor, the rule-based baseline outperforms on database_flaky Class F. A single active-inference controller is not universally best. If the benchmark stops surfacing those honest losses, that itself is a signal something has gone wrong with the setup.
Themesis, one honest anchor
A reading order for this pillar cluster
Read the pieces in the order below. The first tightens the vocabulary of the model itself. The second unpacks the divergence that shows up in every free energy expression. The third is the discipline that keeps everything else honest.
Where to go if you want to test us, not just read us
Open a lab and move the precision dials. Read the preprint and check the POMDP formulation against Parr, Pezzulo and Friston. Run the Cell Lab and see the losses for yourself. If you find a place where our vocabulary is overreaching the evidence, our transparency page is where we log corrections. That is the point. Test the build, inspect the gates, and help us find where it fails.