Bayesian brain says perception is inference. Active inference says life is inference, and action is how the inferring thing keeps being one. That second sentence is the whole difference, and it is larger than it looks.
What Bayesian brain actually claims
The Bayesian brain hypothesis, associated with Helmholtz, Dayan,
Hinton, Rao and Ballard, and Knill and Pouget, is a story about
perception. The brain holds a generative model of hidden causes
s that produce sensations o, and it updates a
posterior Q(s | o) that approximates the exact posterior
P(s | o). Under this hypothesis the mind's job is to invert
the world, to work backward from what it senses to what is out there
Class E.
This is a beautiful account of perception, and much of the evidence for it (predictive coding in cortex, mismatch negativity, illusions that exploit priors) still stands. But it stops where the body starts. In the Bayesian brain frame, action is an output stage bolted on after the posterior is computed. The math has no principled place for it Class E.
Where active inference picks up
Active inference (Friston 2010, and the book-length treatment by Parr,
Pezzulo, and Friston, 2022) closes that gap by making action part of
the same objective function that governs perception. A creature is
modelled as a Markov blanket, a statistical partition between internal
states and external states, with sensory and active states as the two
one-way channels between them. Both perception and action minimise the
same quantity: variational free energy F, an upper bound
on surprise Class E.
Perception changes internal states to fit sensations. Action changes sensations to fit internal states. One math, two arrows.
The lever for action is a companion quantity, expected free
energy G(π), computed over policies (short action
sequences). Parr, Pezzulo, and Friston show that G
decomposes into a pragmatic term (how well the policy is expected to
reach preferred outcomes) and an epistemic term (how much the policy
is expected to reduce uncertainty about hidden states). Selecting the
policy that minimises G therefore trades off goal-seeking
against information-seeking without an external referee
Class E.
That is the piece Bayesian brain does not supply. Curiosity is not a separate module glued on top of a perceiver. It falls out of the same objective as belief updating.
The scaffolding, in one page
Under a POMDP with hidden states s, observations
o, and a generative model P(o, s | π), an
active-inference agent maintains an approximate posterior
Q(s | π). The free energy
F = E_Q[log Q(s | π) − log P(o, s | π)] is a
KL-divergence-plus-log-evidence expression. Minimising it in the
perceptual limit recovers Bayesian belief updating. Minimising expected
free energy across policies recovers goal-directed, curious action.
Same lens, wider aperture Class E.
In the Precision Lab this is not abstract. The dials for sensory precision, transition precision, and policy temperature are the weights the agent places on evidence, dynamics, and value respectively when it solves the free-energy problem. Turn sensory precision down and the agent stops trusting its senses, so it stops acting on them. Turn policy temperature up and exploratory action collapses into habit. The behaviour you watch is the math you can read off the page Class C.
Why the distinction matters for a natural, testable model of mind
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 the claim on faith. Test the build, inspect the gates, and help us find where it fails. The reason we sit inside active inference and not Bayesian brain is exactly this: we need an account that can carry action, embodiment, and self-maintenance without importing them from elsewhere. Bayesian brain describes an observer. Active inference describes a participant.
The Cell Lab is where that commitment gets checked. A single active-inference controller is placed against random, rule-based, and neural baselines on a hidden 216-state service cell under seven disturbance families, with falsification criteria written before the runs. It loses three of the seven, which is what a scientific model is supposed to be able to do Class C.
None of this proves active inference is the correct theory of mind or of anything else. It says the math has enough surface area to be wrong in specific, recorded ways, and that is more than Bayesian brain as a perception-only claim can offer on its own.