A Beautiful Loop
Current reading notes on the active-inference account of consciousness and the recursive loop that makes a model include itself.
These are the durable reading notes for Laukkonen, Friston, and Chandaria's active-inference account of consciousness.
What matters for the site is the way the book turns consciousness into a loop: not just a model of the world, and not just a model of the self, but a model whose activity includes evidence for its own modeling.
The loop, again
Consciousness is the inconvenient enigma. Most fields of inquiry hit it at the edge of their explanatory reach and quietly look away. The active-inference research program keeps coming back to it.
Laukkonen, Friston, and Chandaria propose that active inference becomes a theory of consciousness, not just for consciousness, when three conditions are met:
- A generative world model: the epistemic field, or the space of contents that can be known.
- Inferential competition: the mechanism by which a particular inference wins entry into the world model.
- Epistemic depth: the property that the field is shared recurrently and hierarchically, so the model contains evidence for its own existence.
The third condition is the part that does the heavy lifting. Without it the system has perception; with it the system has the strange property of knowing it is perceiving.
Binding
The binding problem in its classical form asks how the brain stitches a red blur, a moving edge, and a smell into the percept apple.
The active-inference reframing is sharper: which inference wins entry into the world model, and on what authority?
Multiple candidate inferences are always in play. Most lose. The winning inference is the one that, conditional on the rest of the generative hierarchy, reduces long-term expected free energy most effectively. The competition is not judged by a separate module; it is implicit in the dynamics.
That matters because it tells us what to look for. Consciousness, on this view, is not a substance you add. It is a regime you stabilize. The same hardware can run in or out of it.
Depth
Epistemic depth is not merely "the system has a self-model." Lots of systems have self-models in a thin sense.
Epistemic depth is the property that the world model contains knowledge that it exists. It is the recursive coupling of the model to itself, distributed across the hierarchy and continuously self-evidencing.
The authors call this field-evidencing. The world model knows itself non-locally: the knowing is not located in any one node of the inference hierarchy, but in the aggregate dynamic of the hierarchy as a whole.
This is the move that distinguishes consciousness of from metaconsciousness about. You can have the latter without the former. The former, on their view, is constitutive.
The hyper-model
The technical move is a hyper-model: a parameter set that controls precision-weighting across the entire generative hierarchy. Not the contents of inference, but the rules by which inference is weighted.
What I find interesting beyond the theory itself is the shape of the argument. They are not adding a new module called consciousness. They are arguing that a particular recursive coupling of components already in the active-inference framework gives you the phenomena.
The loop is beautiful because it does not need a homunculus.
I argued something adjacent in Reaching for the intangible: that the search for "human-level" AI is a category error precisely because the phenomenon being approximated is not a discrete capability. It is a recursive coupling. Different framing, similar shape.
Neighborhood