NOTE-1063: Laukkonen, Friston & Chandaria — A beautiful loop

Contribution

The paper asks what active inference would need, beyond explaining perception and action, to count as a theory of consciousness. It proposes three conditions “that seem necessary for consciousness” (p. 2) and argues that some active-inference systems meet them. The third, epistemic depth, is its own: a system-wide loop in which the model of the world is fed back into the hierarchy that builds it, so that the model “entails the inference, belief, or expectation, that it exists” (p. 10). It is put in formal terms as a hyper-model over precision, and then used to redescribe a wide range of states, from inattentional blindness to cessation in meditation.

Key insight

Most theories ask which contents become conscious. This one says that what varies with awareness is not content but how far the system’s model of its world is known to the system, by being looped back through every level as a forecast of how much to trust each level. A nearly empty world model known very clearly is then a possible state of consciousness, and a rich world model known dimly (a non-lucid dream, driving on autopilot) is another. “Awareness depends on epistemic depth more than configurations of contents or the degree of analytic processing” (p. 20).

Assumptions

• Active inference and the free-energy principle, with existence as maintaining a boundary and self-evidencing (pp. 3–4). Precision weighting sets the gain on prediction errors, and attending is raising precision (p. 4).
• Metzinger’s epistemic space and minimal phenomenal experience (MPE). The paper agrees with Metzinger’s emphasis and prefers “field” to “space” (p. 9, note 4). Following him, “a first person perspective, self-modeling, and agency, are not prerequisites of awareness, but are rather local or “contracted” forms of consciousness” (pp. 2–3).
• Knowing that one knows is the mark of awareness, after Cleeremans et al. 2020 (p. 6), but reached “through recursion not metacognition” (p. 20).
• Contemplative reports are evidence. Luminosity, lucid dreamless sleep, non-dual awareness and cessation are taken as phenomena a theory must cover (pp. 8, 21–25).
• Each condition is asserted necessary. The world model is “a necessary condition for consciousness because it defines what can become conscious” (p. 3); the global model is “necessary but not sufficient” until it is “reflected back through the underlying hierarchy” (p. 9).

Key results

The paper reports no data and no simulation. By section:

• Section 2: the reality model (pp. 3–5). The organism’s lived world is a “generative phenomenal, unified world model”, also an “epistemic field”. Active inference generates one, so “Active inference therefore satisfies our first condition for consciousness” (p. 5). It explains “what is known or experienced” but “not yet” why or how (p. 5).
• Section 3: inferential competition (pp. 6–9). What enters the world model is decided by “precision-weighted competition between possible explanations for the causes of sensory data” (p. 7). Coherence decides it: “coherence and boundedness are a central criterion in winning the inferential competition” (p. 6), and an inference “incoherent with other parallel and hierarchically adjacent inferences throughout the system” is “less likely to be selected” (p. 7). This is “Bayesian binding”, shown for a face percept in Figure 2 as a precision-weighted average of prior and data. GNWT ignition is redescribed as the same competition (p. 8). Whyte & Smith’s discrete–continuous account is called “A complimentary view” (p. 8).
• Section 4: epistemic depth (pp. 9–11). Defined as “a capacity or continuum (i.e., deepening) of knowing” (p. 9), with “luminosity” as the clarity of knowing, taken from Buddhist and Indian sources. The analogy is hearing one’s own voice. Figure 3’s caption states the hypothesis: “We hypothesize that this recursion is the causal mechanism permitting epistemic depth (the sensation of knowing)” (p. 11).
• Section 5: hyper-models (pp. 11–14). A hyper-parameter set Φ encodes “which layers to trust more (or less) under different contexts” (p. 11). “It is this recursive aspect that equips belief updating with epistemic depth” (p. 11). Table 1 (p. 13) gives four equations: the usual hierarchical generative process; the same with each level conditioned on a hyper-parameter φ⁽ˡ⁾ under a prior p(Φ); a local free energy per layer; and a free energy over all states and Φ. The paper separates this from “parametric depth” (Sandved-Smith et al.), which is local, while epistemic depth is “system-wide” (p. 13). The smallest case is a two-layer system that predicts and broadcasts precisions (p. 14). Small nervous systems (Megaphragma, Drosophila larvae, C. elegans) are named as borderline candidates, and plants are left open (note 6).
• Section 6: implementation (pp. 15–18). The hyper-loop has five steps: hyper-prediction, lower-level inference, error on the precision forecast, hyper-update, broadcast (p. 15). In the brain, “neuromodulatory sprays” and endogenous electromagnetic fields are proposed as the broadcast channels, and said to explain loss of consciousness in anaesthesia, deep sleep and seizure where frontoparietal or posterior neurons still fire (p. 17). “Biological loops are unique because they are leaky” (p. 16), while an artificial agent’s Φ “must be coded explicitly” (p. 16). For language models, Reinforcement Learning from Internal Feedback is called “a kind of primitive hyper-model” (p. 18), and a blueprint is given: move the confidence signal “from the loss function into the forward pass” (p. 18). “Our proposal is first and foremost a conceptual one” (p. 18).
• Section 7: attention and metacognition (pp. 19–20). Three levels: inferences that lose and stay unconscious; inferences bound into the model but not known that they are known (periphery, the shirt one wears); and luminously known contents. Subliminal priming, inattentional and change blindness, and driving on autopilot are placed among them.
• Section 8: sleep and lucidity (pp. 20–21). Deep NREM sleep: a minimal model and low epistemic depth. REM: a rich model, low depth. Lucid dreaming: depth rises. Lucid dreamless sleep: “very low abstraction but high epistemic depth”, “an aware but empty epistemic field” (p. 21).
• Section 9: meditation and psychedelics (pp. 22–26). MPE is when “epistemic depth is maximally high, and the reality model is contentless” (p. 22): first-order precision low everywhere, the precision on that precision sharp. Dereification and opacity rise with depth. Cessation is “Bayesian unbinding”: “the world model is not contentless but incoherent” (p. 25). Psychedelics are conjectured to raise epistemic depth, and “one can feel very conscious of, and confident in, a world model that is nevertheless delusional” (p. 26).
• Section 10: discussion (pp. 26–30). The theory “places experiential contents themselves at the very center of consciousness rather than a distinct self, an agent, or some other separable and dualistic force” (p. 26). Table 2 sets six features against GNWT, IIT, RPT and HOT. For AI, three questions replace “can AI be conscious like humans?”, and “This makes epistemic depth the predominant gap in current AI systems” (p. 27). On a system that meets the conditions and says so: “We seem forced, then, to conclude that the AI system is conscious” (p. 29). Note 11 then retreats: “Our position does not assert that satisfying these conditions guarantees phenomenal experience in the subjective sense; rather, it proposes that such a system would exhibit computational and behavioral hallmarks of consciousness” (p. 29). The function proposed is that “consciousness may be the solution to general intelligence” (p. 29).
• Section 11: conclusion (p. 30). “The theory is testable and falsifiable at the level of computational modeling, but also in terms of neural implementation.” The test given: “If the three conditions are met, we ought to see evidence of awareness or deep and flexible epistemicity, as well as success on any Turing-type tests.”

Claims

id	claim	strength	support
C1	A unified generative world model is necessary for consciousness	asserted; definitional in part	pp. 3–5
C2	What enters it is decided by precision-weighted competition that coherent inferences win (Bayesian binding)	weak: a redescription of hierarchical Bayesian inference, with cited phenomena	pp. 6–9, Figure 2
C3	Epistemic depth, recursive sharing of the world model through the hierarchy, is what makes its contents known, and awareness is graded by it	speculative: a hypothesis with no test	pp. 9–11, 19–20
C4	Epistemic depth can be formalised as a global hyper-model over layer-wise precision	weak: the equations are standard, and nothing in them is shown to yield a model that “contains the knowledge that it exists”	Table 1, pp. 11–15
C5	Neuromodulatory sprays and endogenous electric fields broadcast the hyper-model	speculative	pp. 16–17, citations
C6	MPE is maximal epistemic depth with a contentless model; cessation is incoherence of the model	speculative: redescription of contemplative reports	pp. 21–25
C7	Psychedelics raise epistemic depth	conjecture	p. 26
C8	Current AI lacks epistemic depth above all, and a system meeting the three conditions should be treated as conscious	argued, then qualified by note 11	pp. 27–29

Concepts

• Reality model / epistemic field: the organism’s whole lived world as a generative, phenomenal, unified model that can be known and acted on.
• Inferential competition / Bayesian binding: precision-weighted selection, with coherence as the criterion, of what enters the model. “Bayesian unbinding” is its failure.
• Epistemic depth: the degree to which the reality model knows itself, through recursive, non-local sharing. Luminosity is its phenomenological name.
• Hyper-model: a higher-order model whose parameters Φ predict and set precision at every level, and is updated by errors on those predictions.
• Parametric depth: local, one-layer-up inference about precision (Hesp, Sandved-Smith), which the paper treats as a mechanism within epistemic depth.
• Field-evidencing: the world model continuously evidencing its own knowing (abstract).

Connections

Built on Friston’s free-energy principle, Hohwy’s self-evidencing, Metzinger’s MPE and epistemic space, and Cleeremans’s “learning to be conscious”. It cites Whyte & Smith 2021 and a 2024 version of Whyte et al. (LIT-1364, NOTE-1062) as complementary, and Safron’s IWMT (spelled “Saffron”, p. 6; LIT-1345) among active inference accounts of consciousness. The record holds Sandved-Smith et al. 2021 (LIT-1191, NOTE-958), the source of “parametric depth”, and Seth 2013 (LIT-1263), cited for keeping the body alive. Hofstadter’s strange loop gives the title.

Bearing on the record

• QUESTION-254 and CLAIM-411. The paper is a candidate answer of the “higher-order or recursive self-interpretation” kind the question lists, and the exchange described it correctly as one: a further organizational condition beyond inference. It accepts the counter-claim’s premise, since inferences that lose the competition stay unconscious (p. 19). But its condition is not independent of experience as stated. Epistemic depth is defined as “a capacity or continuum (i.e., deepening) of knowing, or awareness” (p. 9), and the hyper-model is offered as its formal analogue without showing that it is necessary or sufficient for it. A hierarchical model with layer-wise precision hyperparameters is common (the paper names the hierarchical Gaussian filter, p. 12), so the question’s demand to say “what a sophisticated nonexperiential interpreter would lack” is not met: an HGF fitted in a laboratory has the Table 1 structure. Note 11 also withdraws from the phenomenal half of the question.
• CLAIM-412 (constitution). Close in spirit: experience is the world model and its looping, with contents “at the very center”, and no further entity. It is not stated as an identity, and note 11 says satisfying the conditions does not guarantee phenomenal experience. So it is a production-or-correlation account that the constitutive thesis could adopt as its candidate organization, not support for the identity.
• TERM-072 (perspective). The paper denies that a first-person perspective is needed for awareness (pp. 2–3). That is the represented sense, and probably the phenomenal one. It does not deny an organizational perspective; its world model is one. This cuts against any reading of CLAIM-412’s part 3 that needs a represented or lived point of view, and it is the term’s warning that claims must say which sense they use.
• QUESTION-255 (unity). The paper takes the coherence side: what enters awareness must cohere, and incoherence of the whole model is cessation, the loss of awareness (p. 25). That is close to the global-section reading A7 rejected. It does not discuss a felt conflict such as approach and avoidance, which is the question’s test case, so it states a position without meeting the counter-example.
• THEORY-063 (unity without a self-model). It shares Metzinger’s view that a unified world comes before a self, and it adds a mechanism (recursion over precision) and a minimal case (MPE). It adds no evidence: the contemplative states are redescribed, not measured. The record’s two unread meditation works (LIT-1354, LIT-1351) bear on whether such states are contentless, which the MPE account needs.
• CLAIM-355. Consistent with it: no separable owner is posited, and the self is a “contracted” form. The paper does not discuss illusion.
• CLAIM-414 (four facts). Epistemic depth is that claim’s third fact, efficacy allocated from inside, made global and recursive. The claim shows that systems alike in ordinary behaviour can differ in it, and that only an intervention tells them apart. The paper’s tests for AI (“evidence of epistemic depth”, “success on any Turing-type tests”) are behavioural. By that claim they cannot settle whether precision is allocated endogenously.
• THEORY-577 and THEORY-633. The paper predicts that a system with the loop would be “adamant that they exist” (p. 29), so it offers a sketch of how its mechanism yields self-reports, which THEORY-577 asks of any theory. Its zombie “stalemate” and note 11 concede that this does not reach phenomenal properties, which an illusionist (THEORY-633) would grant at no cost.
• CLAIM-413 (grounding). Lucid dreamless sleep and MPE are cases of awareness with almost no current sensory constraint. On the paper’s account they are rich in epistemic depth, so current grounding is not what makes them aware, which is the claim’s occurrence half. The paper does not argue it.
• TERM-073. Inference in the paper fits the term, and the paper’s first two levels of awareness (lost inferences; bound but unknown contents) are interpretation without, or with little, awareness.
• No instruction for machine-learning practice. The LLM blueprint (pp. 18) is a suggestion about building a hyper-model, offered as illustration.
• CLAIM-439 (valence is felt when the system models its own potential). The paper is a rival on sufficiency, and a possible generalization. It is not a version of the claim.
  • The claim’s sources sit on the paper’s “parametric” side. The paper separates epistemic depth, which is system-wide, from parametric depth: “Parametric depth is often implemented as localized loops—between one layer above another (Sandved-Smith, 2021; 2024), implementing a second-order inference about attention, or about preference precision” (p. 13). It names “policy precision” with Hesp et al. (2019, the preprint of LIT-1468) as a parameter of that kind (p. 13). CLAIM-439’s model of potential is Hesp’s deep level over γ, a policy precision, and Brouillet and Friston’s felt fluency (THEORY-662), a deeper inference about a change in precision. In the paper’s terms both are parametric depth. That is “a mechanism for implementing meta-inference in a hierarchical generative model” (p. 13), not what makes contents known.
  • So the paper predicts the claim’s second defeater. “Under this view, even complex metacognition can potentially be unconscious or conscious depending on epistemic depth” (p. 20). A model of one’s own potential that tracks it and is used elsewhere, yet is not accompanied by awareness, is what CLAIM-439’s second defeater describes. On this paper such cases exist wherever the model is local and the global loop is shallow. The paper asserts this and does not show it, since it has no data (NOTE-1063).
  • How it could generalize the claim. The hyper-model’s parameters Φ are the ones “that set precision and weighting rules across levels” (Table 1 note, p. 13), and policy precision is among the precisions such a model would forecast. The claim could be restated so that valence is felt when the system’s model of its potential is part of a system-wide recursive loop, not only “used elsewhere”. CLAIM-439’s clause “the inference is used elsewhere in the system” is a weak, one-way form of the paper’s broadcast. Whether one use elsewhere is enough, or the loop must be global and recursive, is the point at issue between them.
  • They mean different things by “felt”. CLAIM-439 adopts THEORY-662’s “attended and available”. The paper’s middle level, inferences bound into the world model but not known to be known (pp. 19–20, NOTE-1063), is available without awareness. So the claim’s sense of “felt” corresponds to that middle level. If the record keeps its thin sense, the paper is about something stronger and is not a direct rival. If the record wants “felt” to mean what the paper means by awareness, the paper says a local self-model is not enough.
  • The paper has no account of valence. Its one mention puts valence among the slow contextual signals neuromodulators carry: “Sprays are brief, diffuse bursts of noradrenaline, acetylcholine, serotonin, and dopamine that can shift synaptic gain over hundreds of milliseconds, embedding slow contextual information such as arousal, valence, and uncertainty” (p. 17). Its hyper-model is about how far to trust each level of inference, not about which ordering of the system’s states prevails. It supplies a condition on awareness, not a model of potential.
  • Bayesian binding is a negotiation over beliefs, not orderings. “Only the inferences that coherently reduce long-term uncertainty win, evincing a selection for consciousness that we call Bayesian binding” (abstract, p. 1). CLAIM-440’s negotiation is among perspectives’ rankings of what to do. The two are parallel in form only. One consequence follows if they are joined. On CLAIM-440, split potential is a kind of ambivalence. Split potential would be incoherent, so on this paper it would be less likely to be bound into awareness. That is QUESTION-255’s felt-conflict case again, which NOTE-1063 notes the paper does not meet.
  • Self-modelling. The paper denies that self-modelling is needed for awareness, and calls “Self-modeling proper (i.e., knowing what kind of thing one is)” “a much later development” (p. 14). CLAIM-439’s self-model is a model of the system’s own regime, not of what kind of thing it is, and so is nearer the paper’s “real-time model of how it is modeling the world” (p. 14). The denial does not touch it.

Limitations

• No data and no simulation. Every phenomenon is redescribed, not predicted, and the paper names no result that would count against it beyond the general statement that it is “testable and falsifiable” (p. 30).
• The formal model (Table 1) is a hierarchy with layer-wise precision hyperparameters under a common prior. Nothing derives from it the “knowledge that it exists”, the non-locality or the luminosity the text attributes to it.
• Epistemic depth is defined in terms of knowing and awareness, so its use to explain awareness risks circularity.
• The AI argument concludes that a system meeting the conditions is conscious, and note 11 then denies that the conditions guarantee phenomenal experience. The two are not reconciled.
• Contemplative and psychedelic reports are taken largely at face value; a footnote cites an estimate attributed to ChatGPT 4o (p. 26, note 10).
• Read from preprint v3; the journal version was not reached.

Open questions

• Would a simulated hyper-model show any signature that an HGF with layer-wise volatility does not? If not, what separates epistemic depth from ordinary hierarchical precision learning?
• What measure of epistemic depth could be taken without asking the subject how aware they were?
• Can the coherence criterion allow a felt conflict to be one experience, as QUESTION-255 asks?