GNI vs AGI: A Precise Distinction
The words we use for a build constrain the build. 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. This post says exactly what we mean by "General Natural Intelligence" and why the word "natural" is doing real work.
The short version
General Natural Intelligence, GNI, is our name for a research direction, not a finished product. It names systems that maintain themselves by minimizing variational free energy through both perception and action, in the sense of Parr, Pezzulo and Friston (2022) (Class E). The generative model is exposed. The Markov blanket is drawn. The dials are legible. You can steer the agent, watch inference change, and read the number you are minimizing.
AGI, as the term is used publicly, points somewhere else: a scaled statistical function approximator asked to behave as if it had a mind. We do not build that, and we do not want to be read as building that. Calling our work "AI" or "AGI" collapses a distinction the code actually enforces, so we do not use those words for our work.
Why "natural" is load-bearing
"Natural" here is a technical qualifier, not a marketing flourish. It picks out three commitments we keep in the running code (Class C):
- An explicit generative model. A POMDP with named hidden states, observation likelihoods, and transition dynamics. Not a bag of parameters. You can inspect it, and you can print the posterior.
- Variational free-energy minimization as the objective. The agent updates beliefs to reduce KL divergence between its recognition density and the true posterior, and selects policies that minimize expected free energy over future observations, per Parr, Pezzulo and Friston (2022) (Class E).
- A Markov blanket that is drawn on purpose. Internal states, sensory states, active states, and external states are separated in the implementation. That separation is what lets a system "maintain itself" mean something checkable, not metaphorical.
Strip any one of these out and you may still have a useful engineering artifact. You do not have what we mean by natural. The active-inference literature uses "natural" the same way biology does: a system whose organization emerges from its own upkeep, not from a training loss defined by an outside optimizer.
How the distinction cashes out in the build
In the Precision Lab, three dials (sensory precision, transition precision, policy temperature) map to distinct behavioral regimes in a small maze POMDP (Class C). You can watch a single change to sensory precision flip the agent from a cautious explorer to a committed executor. That is not a black box you interrogate after the fact. It is the equation exposed as a slider.
In the Cell Lab, a UNI active-inference controller is run against random, rule-based, and neural baselines on a hidden 216-state service cell under seven disturbance families (Class C). The falsification criteria were written before the runs. Losses are reported alongside wins. A single controller is not universally best, and the benchmark is designed to make that visible. If our approach were a general claim about universal superiority, the Cell Lab would embarrass it, and we published the Cell Lab anyway.
The MCP server exposes the whole surface (list_labs, describe_dial, run_episode, read_state, and twelve more) so any language model can drive the agent and read its posterior over hidden states (Class C). If the internals were not natural in the sense above, there would be no coherent posterior to read.
What the distinction does not claim
GNI, as we use it, is a research direction with growing evidence, not an achievement. We do not claim consciousness, sentience, or a solved problem. We do not claim that active inference is the correct theory of mind. We report which claims survive their pre-registered falsifiers and which do not. The preprint (DOI 10.5281/zenodo.19785799) is unrefereed, and we label it that way (Class C).
Do not take the claim on faith. Test the build, inspect the gates, and help us find where it fails.