Vicente Raja, Dinesh Valluri, Edward Baggs, Anthony Chemero and Michael L. Anderson (2021), Physics of Life Reviews 39:49–72.
The brief’s citation (Physics of Life Reviews, 2021, five authors) is correct.
Key takeaways
- True but narrow. “Given these commitments, we think FEP is true.” The dispute is over its scope as a principle, not its derivation.
- Every thing, not everything (§3.1). FEP’s own literature excludes the candle flame, which “clearly” is a thing; blankets need boundaries more stable than the states they bound, which clouds and groups lack.
- The partition is placed (§3.2). Brain, bacillus and coupled pendulums get blankets wherever is convenient: for the pendulums the beam is the blanket and position and velocity are active and sensory states. Read strictly, the formalism makes the “thing” the whole system-plus-environment.
- What falls outside (§3.3). Relational properties (taller-than, heading, affordances) and constitutive self-organization: a cell’s membrane is produced by the cell; a beam is not produced by the pendulums. The blanket formalism “has no tools to account for” the difference.
- The trick (§3.5). Blankets supply exactly the hidden/observed/system partition that variational inference needs, so any thing can be modelled “as if it were a variational autoencoder”; FEP was built after variational free energy was in use, not the other way round.
- Active inference (§4). Its generative models (A, B, C matrices) are designed by the experimenter; perception of the cue and control of movement are assumed, so the theory presupposes what it claims to explain.
Standing in the record
Filed on 2026-10-03 at the owner’s request, in the free-energy part of the agency batch. No anthology topic holds it. It discusses variational autoencoders and the reparameterization trick only as analogies for what the blanket does, and carries no instruction for machine-learning practice.
Relation to LIT-526: corrects. “Life as we know it” offers autopoiesis
as one of its four marks of the lifelike and claims its simulated soup shows
it: lesioning blanket states makes the structure disperse, so the system
“appear[s] to actively maintain its structural and dynamical integrity”.
Raja et al. argue that the blanket formalism cannot represent constitutive
self-organization, a boundary the system itself produces, and say of
Friston 2013’s simulations specifically that they “assume the existence of
all the particles from the beginning and, therefore it cannot be said that
their blanket states are the product of the activities of internal states”
(footnote 17). So the claim that the simulation exhibits autopoiesis, in the
sense of a self-produced boundary, is shown not to follow. The correction is
narrow: it is conceptual, it leaves LIT-526’s lemma standing (Raja et al.
grant the derivation), and NOTE-421 had rated that claim weak to moderate
(its C6). They also report Biehl et al.’s finding that some blanket
formalizations, Friston 2013 among them, are mathematically inadequate, but
do not develop it.
Where else it bears:
- The candle flame. Raja et al. accept LIT-526’s claim that a flame cannot have a blanket and turn it round: the flame is plainly a thing, so a principle that excludes it is not a theory of everything. Mossio, Saborido & Moreno (LIT-570) grant the flame organizational closure.
- Bruineberg et al. (LIT-603). Cited (“see also Bruineberg et al. 2020”) for the same conclusion, that blankets set up a boundary rather than find one. The two papers reach it by different routes and do not test each other.
- Coordination dynamics (LIT-601). The coupled pendulums are already “a thing measured by the relative phase between them (Haken et al. 1985; Kelso 1995)”, so a blanket partition adds nothing to their explanation but a Bayesian redescription. That is the HKB explanation read in this batch through Kelso 2021; the cited originals are LIT-588 and LIT-578, both Deferred.
- Closure of constraints (LIT-582). The “constitutive self-organization” the authors say blankets miss is what closure of constraints is built to describe. The paper cites autopoiesis and operational closure (Maturana & Varela, Di Paolo et al. 2017), not Montévil & Mossio.
- Sophisticated Inference (LIT-577). Its T-maze rat is the kind of simulation §4.2 criticises: the generative model is the experimenters’ rationalisation of the task. Raja et al. analyse a different T-maze paper (Hesp et al. 2021) in the same family.