# Full Text: The Doors of Perception are the Threshold of Prediction: Active Inference and William Blake's Theory of Seeing

> Extracted from `2026_DoorsOfPerception.pdf`

---

## Page 1

The Doors of Perception are the Threshold of Prediction
Active Inference and William Blake’s Theory of Seeing
Daniel Ari Friedman
Active Inference Institute
daniel@activeinference.institute
ORCID: 0000-0001-6232-9096
DOI: 10.5281/zenodo.18600041
February 12, 2026

## Page 2

1
Thematic Atlas
“What is now proved was once only imagined.” — Blake
Here, eight concordances are (sk)etched between Blake’s prophetic vision and the mathematics of Active
Inference.
1.1
Table of Contents by Theme
Section
Blake’s Vision
Active Inference
Key Quotation
Boundary
Doors of Perception
Markov Blanket
“If the doors of
perception were
cleansed…”
Vision
Fourfold Vision
Hierarchical Processing
“Now I a fourfold vision
see…”
States
Newton’s Sleep
Prior Dominance
“Single vision &
Newton’s sleep!”
Imagination
Human Existence
Generative Model
“Imagination is Human
Existence itself”
Time
Eternity in Hour
Temporal Horizons
“Hold… Eternity in an
hour”
Space
World in Grain
Spatial Inference
“To see a World in a
Grain of Sand”
Action
Cleansing
Free Energy Minimization
“every thing would
appear… as it is: infinite”
Collectives
Building Jerusalem
Shared & Factorized Models
“Till we have built
Jerusalem…”
Contents
1
Thematic Atlas
2
1.1
Table of Contents by Theme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2
2
Abstract: The Prophetic Synthesis
5
3
Introduction: The Threshold
6
3.1
The Threshold
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6
3.2
The Correspondences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6
3.3
Method
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
4
Related Work: Scholarship & Context
8
4.1
Blake and Embodied Cognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
4.2
Hemispheric Lateralization
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
4.3
Romanticism and the Science of Mind . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
4.4
Neuroaesthetics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
4.5
Social Neuroscience and Joint Improvisation . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
4.6
Psychedelics and the Predictive Mind
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
4.7
Northrop Frye’s Systematic Blake . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
4.8
Comparative Systems: Blake and Fuller
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
4.9
Phenomenological Traditions
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
4.10 Extended Mind and 4E Cognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10
4.11 Affect Theory and Precision Weighting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10
4.12 Predictive Processing and Aesthetic Experience . . . . . . . . . . . . . . . . . . . . . . . . . .
10
2

## Page 3

4.13 Consciousness as Controlled Hallucination . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10
4.14 Cognitive Romanticism
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11
4.15 Cultural Affordances and Shared Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11
4.16 Our Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11
5
Theoretical Foundations
12
5.1
The Free Energy Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12
5.1.1
Minimization Pathways
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12
5.1.2
Expected Free Energy and Policy Selection
. . . . . . . . . . . . . . . . . . . . . . . .
12
5.2
The Markov Blanket . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13
5.2.1
Nested Blankets and Multi-Scale Organization
. . . . . . . . . . . . . . . . . . . . . .
13
5.3
Hierarchical Generative Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13
5.3.1
Model Evidence and Complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14
5.4
Precision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14
5.4.1
Precision Dynamics and Pathology . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15
5.5
Prediction Error and Message Passing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15
5.6
Temporal Depth
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15
5.7
Multi-Agent Inference
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16
5.7.1
Mean-Field Factorization
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16
5.8
Cognitive Art and the Fourfold . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16
5.9
Summary of Formal Apparatus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16
6
Synthesis: Eight Themes of Vision
18
6.1
Boundary: The Doors of Perception
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
20
6.1.1
The Markov Blanket . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
20
6.1.2
Boundary Constitution: Naming Creates Separation . . . . . . . . . . . . . . . . . . .
21
6.1.3
The Abyss as KL-Divergence
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
6.1.4
Forged Boundaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
6.1.5
Body as Blanket Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
22
6.1.6
Imagination’s Outline vs. Nature’s Dissolution
. . . . . . . . . . . . . . . . . . . . . .
22
6.1.7
Contemporary Resonance: From Blake Through Huxley to ALBUS . . . . . . . . . . .
22
6.2
Vision: The Fourfold Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
24
6.2.1
Hierarchical Generative Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
24
6.2.2
Golgonooza: The Architecture of the Generative Model
. . . . . . . . . . . . . . . . .
25
6.2.3
Organs of Perception as Model-Dependent . . . . . . . . . . . . . . . . . . . . . . . . .
25
6.3
States: Newton’s Sleep & Prior Dominance
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
27
6.3.1
Prior Dominance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
27
6.3.2
The Spectre’s Steel Ratio: Prior Dominance Formalized . . . . . . . . . . . . . . . . .
29
6.3.3
The Philosophy of Five Senses
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
6.3.4
Descending into Division . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
30
6.3.5
The Natural Man vs. Spiritual Man
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
30
6.3.6
Hemispheric Resonance
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
30
6.4
Imagination: The Generative Model
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
31
6.4.1
The Generative Model as Self . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
31
6.4.2
Man Is All Imagination: Autopoiesis and Self-Modeling
. . . . . . . . . . . . . . . . .
32
6.4.3
Mental Things Alone Real . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
32
6.4.4
Knowledge by Perception, Not Deduction . . . . . . . . . . . . . . . . . . . . . . . . .
32
6.4.5
Innate Ideas as Structural Priors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
33
6.4.6
Imagination as Epistemic Foraging . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
33
6.5
Time: Temporal Horizons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
34
6.5.1
Temporal Horizons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
34
6.5.2
The Temporal Architecture of Los
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
34
6.5.3
Drawing Out to Seven Thousand Years
. . . . . . . . . . . . . . . . . . . . . . . . . .
36
6.5.4
Eternity Obliterated . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
37
3

## Page 4

6.5.5
Drunk with the Wine of Ages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
37
6.6
Space: Spatial Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
38
6.6.1
Spatial Scale Invariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
38
6.6.2
The Vortex: Each Entity’s Own Markov Blanket . . . . . . . . . . . . . . . . . . . . .
39
6.6.3
Mathematic Holiness: The Impoverished Spatial Model
. . . . . . . . . . . . . . . . .
39
6.6.4
Looking Through, Not With
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
39
6.7
Action: Free Energy Minimization
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
6.7.1
Free Energy Minimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
6.7.2
Flexible Senses: Precision Dynamics Before the Fall
. . . . . . . . . . . . . . . . . . .
42
6.7.3
Contracting to the Honey Bee . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
42
6.7.4
Love as Affective Precision
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
42
6.7.5
Thought Alone Makes Monsters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
42
6.7.6
Cogs Tyrannic vs. Free Wheels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
43
6.7.7
The Path of Least Action . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
43
6.8
Collectives: Shared Generative Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
45
6.8.1
Shared Generative Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
45
6.8.2
The Mental Fight . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
47
6.8.3
Coordinated Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
47
6.8.4
From Single to Collective Vision
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
48
6.8.5
Fall into Division and Resurrection to Unity . . . . . . . . . . . . . . . . . . . . . . . .
48
6.8.6
The Eternal Man Is Risen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
48
6.8.7
Human Harvest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
48
6.8.8
The Four Zoas: Factorized Collective Mind
. . . . . . . . . . . . . . . . . . . . . . . .
49
6.8.9
Jerusalem as Cultural Niche . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
53
7
Implications: The Wider Fields
54
7.1
Philosophy of Mind . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
54
7.1.1
The Romantic Computational Mind
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
54
7.1.2
Consciousness as Hierarchical Depth . . . . . . . . . . . . . . . . . . . . . . . . . . . .
54
7.2
Cognitive Science . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
54
7.2.1
Predictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
54
7.2.2
Neural Correlates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
7.3
Creativity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
7.3.1
The Artist as Model-Builder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
7.3.2
Aesthetic Free Energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
7.4
Transpersonal Experience . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
7.4.1
Mystical Perception
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
7.4.2
Building Jerusalem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
7.5
Counter-Arguments
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
7.6
Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
56
7.7
Contemporary Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
56
7.7.1
AI Consciousness and Machine Imagination . . . . . . . . . . . . . . . . . . . . . . . .
57
7.7.2
Mental Health Interventions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
57
7.7.3
Embodied AI and Multi-Agent Coordination
. . . . . . . . . . . . . . . . . . . . . . .
57
7.7.4
Emerging Field Convergence
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
57
8
Conclusion: The Threshold
58
8.1
The Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
58
8.2
The Threshold
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
58
8.3
Newton Still Sleeps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
58
8.4
The Reciprocal Gift
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
59
8.5
Building Jerusalem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
59
8.6
Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
59
8.7
Final Reflection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
60
4

## Page 5

2
Abstract: The Prophetic Synthesis
Looking at the sun, William Blake saw an innumerable company of the heavenly host where Newton’s heirs
saw only a golden coin. “If the doors of perception were cleansed,” Blake wrote, “every thing would appear
to man as it is: infinite.” This paper argues that Blake’s prophetic vocabulary, far from being merely poetic,
constitutes an anticipatory phenomenological insight into the cognitive architecture that Active Inference
now formalizes mathematically. Blake’s “doors” are statistical boundaries separating self from world; his
“Newton’s sleep” is the pathology of rigid priors crushing sensory evidence; his “fourfold vision” maps onto
hierarchical precision-weighting across processing depths; his insistence that “Imagination is the Human
Existence itself” anticipates the insight that selfhood is constituted by the generative model. These are not
retrospective metaphors imposed on a Romantic poet, but convergent descriptions of the same perceptual
territory, arrived at through radically different methods two centuries apart. We approach this convergence
in the spirit of Hesse’s Glass Bead Game: not as proof that one tradition vindicates or completes the other,
but as a synthetic juxtaposition of Art and Science—two moves in the same ancient, ongoing game of making
sense of sense-making.
Through close reading of The Marriage of Heaven and Hell, Milton, Jerusalem, and other works, we trace
eight structural correspondences between Blake’s perceptual philosophy and the Active Inference framework:
Boundary, Vision, States, Imagination, Time, Space, Action, and Collectives—the last encompassing Blake’s
Four Zoas as a factorized model of collective mind. Each correspondence begins with Blake’s phenomenologi-
cal fire—his exact words, his illuminated images—and follows the mathematical shadow that Active Inference
casts across the same ground: Markov blankets, hierarchical generative models, precision dynamics, temporal
depth, spatial inference, free energy minimization, and multi-agent coordination. The formalism developed
by the Active Inference community provides mathematical precision, yet we resist treating it as a finished
edifice; the framework is better understood as one contemporary articulation of principles that Blake, and
traditions before him, grasped through other means.
The synthesis contributes to both lineages: Blake
scholarship gains formal grounding of insights long dismissed as mystical enthusiasm; cognitive science gains
phenomenological depth, historical precedent, and the humbling recognition that its discoveries may be re-
discoveries after all. The doors of perception have always been thresholds of prediction—Blake’s visions and
the equations point towards the same boundary, and the conversation between them remains open evermore.
Epistemic status: “delighted with the enjoyments” of AI which “look like torment and insanity”. Take all
syntax and semantics with a “grain of sand”. For my personal limitations and typographical errors I plead
“Mutual Forgiveness of each Vice”.
Keywords: Active Inference · William Blake · Free Energy Principle · Predictive Processing · Markov
Blanket · Generative Model · Philosophy of Mind · Romanticism · Glass Bead Game
5

## Page 6

3
Introduction: The Threshold
“If the doors of perception were cleansed every thing would appear to man as it is: infinite. For
man has closed himself up, till he sees all things thro’ narrow chinks of his cavern.”
— Blake, Marriage of Heaven and Hell, Plate 14 [Blake, 1988b]
3.1
The Threshold
Between perceiver and perceived lies a boundary. Blake called it a door. In causal inference, that boundary
may be called a blanket. The exoteric syntax differs; the esoteric semantics does not.
William Blake (1757–1827) composed his prophetic works during the consolidation of Newtonian mechanism—
the reduction of cosmos to clockwork, of vision to optics, of mind to matter arranged [Raine, 1968]. His
response was not retreat into mysticism but a vigorous expansion: a fourfold epistemology that could contain
Newton’s single vision while transcending it.
The Active Inference framework, developed by a growing community of researchers worldwide, offers a formal
complement to Blake’s insights. The free energy principle formalizes how self-organizing systems maintain
existence by minimizing prediction error [Friston, 2010]. Perception and action unite in a single imperative—
to reduce the gap between expectation and evidence.
This paper explores how Blake’s intuitions and Active Inference’s equations resonate. The former prophe-
sied; the latter formalizes. We do not claim that Blake was a proto-Bayesian statistician, nor that Active
Inference is a “Blakean” science, but rather that both systems grapple with the same fundamental problem:
how a bounded agent maintains its existence and makes sense of an infinite world. We offer a synthetic jux-
tapositional intelligence—placing the poet’s vision alongside the physicist’s variables to reveal the structural
identity of their insights.
The spirit of this enterprise owes something to Hesse’s Glass Bead Game: an abstract synthesis of all arts and
sciences, where the player discovers hidden affinities between seemingly unrelated disciplines [Hesse, 1943].
Like Hesse’s Castalian scholars, we do not seek to reduce one tradition to the other, but to illuminate the
structural resonances that emerge when both are held in the same contemplative field.
This synthesis arrives at a moment of convergence. On one side, predictive processing and active inference
are being applied with increasing sophistication to aesthetics and literary engagement—most notably in the
2024 Philosophical Transactions of the Royal Society B theme issue on art and predictive processing [Van de
Cruys et al., 2024], and in Kukkonen’s work modeling literary experience through prediction error [Kukkonen,
2020]. On the other, cognitive approaches to Romanticism are deepening: Savarese’s Romanticism’s Other
Minds [Savarese, 2020] reveals a “prehistory of cognitive approaches to literature” within the Romantic
tradition itself. Our paper sits at the intersection of these two currents, offering what neither can alone: the
formal mathematics that makes the poetic claim testable, and the phenomenological richness that makes the
formalism legible.
3.2
The Correspondences
Eight thematic correspondences anchor our synthesis (see Thematic Atlas):
Table 1:
Thematic Atlas:
Structural correspondences between
Blake’s visionary phenomenology and Active Inference.
Theme
Blake’s Term
Active Inference Term
Boundary
Doors of Perception
Markov Blanket
Vision
Fourfold Vision
Hierarchical Processing
States
Newton’s Sleep
Prior Dominance
Imagination
Human Existence
Generative Model
Time
Eternity in an Hour
Temporal Horizons
6

## Page 7

Theme
Blake’s Term
Active Inference Term
Space
World in a Grain of Sand
Spatial Inference
Action
Cleansing
Free Energy Minimization
Collectives
Building Jerusalem
Shared Generative Models
“May God us keep / From Single vision & Newton’s sleep!”
— Blake, Letter to Butts, November 1802 [Blake, 1802]
3.3
Method
We proceed now through three main movements:
• §2: Related scholarship: Blake and cognition, Romanticism and neuroscience, situating our contribu-
tion
• §3: Theoretical Foundations: free energy, Markov blankets, precision
• §4: Synthesis: eight themed correspondences with equations and figures
Each theme in our synthesis (Thematic Atlas) begins with Blake’s fire, then traces its mathematical shadow.
The conclusion (§5–6) draws implications for philosophy of mind, cognitive science, and creativity, while
engaging counter-arguments and acknowledging limitations.
7

## Page 8

4
Related Work: Scholarship & Context
Situating the correspondence within existing scholarship.
4.1
Blake and Embodied Cognition
The mapping between Romantic poetry and cognitive science has precursors. Three pillars of Blake schol-
arship make our formal mapping possible. Northrop Frye’s Fearful Symmetry [Frye, 1947] established the
systematic reading of Blake’s symbolism as a coherent intellectual structure rather than private mythology.
S. Foster Damon’s A Blake Dictionary [Damon, 1988] provides the essential lexicon of Blake’s symbolic
system, establishing the correspondences among his mythological figures that a formal mapping requires.
Peter Ackroyd’s definitive biography [Ackroyd, 1995] demonstrates how Blake’s visionary epistemology was
inseparable from his lived practice as engraver, printer, and painter—an embodied creativity that resists
reduction to disembodied ideas.
From the cognitive science side, two works converge on the same insight. Mark Johnson’s The Body in the
Mind [Johnson, 1987] argues that abstract thought is grounded in embodied image schemas—exactly the kind
of perceptual-motor structures that Active Inference formalizes as generative models. Lakoff and Johnson’s
Metaphors We Live By [Lakoff and Johnson, 1980] demonstrated that conceptual structure is metaphorical
and embodied, not abstract and propositional.
Blake anticipated both traditions by two centuries. His insistence that “Man has no Body distinct from his
Soul” (MHH Plate 4) is not metaphor but proto-enactivism: the body is not a container for mind but the
very medium of inference.
4.2
Hemispheric Lateralization
Iain McGilchrist’s The Master and His Emissary [McGilchrist, 2009] proposes that the left hemisphere
prioritizes narrow, focused, already-known categories while the right hemisphere attends to the broad, con-
textual, and novel. This lateralization maps suggestively onto Blake’s mythology: Urizen (left hemisphere)—
the lawgiver who “closed the tent of the Universe,” imposing rigid categories—versus Los/Urthona (right
hemisphere)—the creative imagination that builds Golgonooza, perpetually open to new form. McGilchrist’s
thesis that Western civilization has progressively over-valued left-hemispheric cognition parallels Blake’s di-
agnosis of “Newton’s sleep” as civilizational pathology.
McGilchrist’s magisterial follow-up, The Matter with Things [McGilchrist, 2021], deepens this analysis with
direct engagement with Blake. McGilchrist treats imagination not as mere fantasy but as a “key faculty”
for revealing reality—echoing Blake’s own elevation of imagination above reason. The dynamic tension of
Blake’s Marriage of Heaven and Hell, where “contraries” generate movement toward deeper consciousness,
exemplifies what McGilchrist identifies as the right hemisphere’s mode of understanding: holding opposites
in creative tension rather than collapsing them into categories.
4.3
Romanticism and the Science of Mind
Alan Richardson’s British Romanticism and the Science of the Mind [Richardson, 2001] documents how
Romantic poets engaged seriously with contemporary brain science, not as opponents but as creative in-
terlocutors. Richardson shows that the Romantic critique of mechanism was not anti-scientific but proto-
cognitive—anticipating embodied, situated, and enactive approaches. Our paper extends Richardson’s his-
torical argument by providing the formal bridge: Active Inference supplies the mathematics that connects
Blake’s phenomenological observations to contemporary computational neuroscience.
4.4
Neuroaesthetics
The emerging field of neuroaesthetics investigates how art engages perceptual and cognitive systems. Ra-
machandran and Hirstein [Ramachandran and Hirstein, 1999] proposed that art exploits principles of percep-
tual processing—peak shift, isolation, and grouping. In Active Inference terms, art offers generative models
8

## Page 9

that resolve free energy in novel ways, restructuring the viewer’s predictions. Blake’s illuminated books—
integrating visual, verbal, and material elements into composite artworks—represent an extreme case: each
plate offers not merely an aesthetic experience but a complete alternative generative model for perception.
4.5
Social Neuroscience and Joint Improvisation
Recent work in social neuroscience and art therapy emphasizes the role of joint improvisation in synchro-
nizing neural states. Mikhailova and Friedman’s “Partner Pen Play in Parallel” (PPPiP) [Mikhailova and
Friedman, 2018] proposes that simultaneous, non-verbal co-creation on a shared surface facilitates “controlled
novelty” and inter-brain synchrony. This practice operationalizes the Active Inference account of communi-
cation not merely as signal transmission but as the mutual alignment of generative models. Just as Blake’s
“fourfold vision” integrates diverse faculties, PPPiP demonstrates how shared aesthetic action can construct
a “collective self-evidencing” dynamic, where the relationship itself becomes the agent minimizing surprise.
4.6
Psychedelics and the Predictive Mind
Aldous Huxley’s The Doors of Perception [Huxley, 1954]—its very title drawn from Blake—proposed that
psychedelic experience reveals perception ordinarily filtered by the brain’s “reducing valve.” Carhart-Harris
and Friston’s REBUS model [Carhart-Harris and Friston, 2019] formalized this intuition, showing that
psychedelics reduce the precision of high-level priors. Safron and colleagues’ ALBUS framework [Safron
et al., 2025] now extends REBUS into a comprehensive account: psychedelics can both relax beliefs and
strengthen them, producing the full spectrum of altered states from prior dissolution to intensified meaning-
making. This is Blake’s “cleansing” rendered computational—the doors of perception swing open when prior
dynamics shift, allowing sensory evidence to reshape the model. The continuity from Blake through Huxley
to ALBUS illustrates how the same phenomenological insight has been rediscovered across centuries and
progressively formalized.
4.7
Northrop Frye’s Systematic Blake
Frye’s Fearful Symmetry [Frye, 1947] remains the foundational systematic treatment of Blake’s mythology.
Frye demonstrated that Blake’s prophetic books constitute a coherent cosmological system, not isolated
flights of fancy. Our paper depends on Frye’s insight that Blake’s symbolism is systematic—without that
systematicity, the structural correspondences with Active Inference would dissolve into vague analogy. Where
Frye mapped Blake’s system as literary criticism, we map it as cognitive architecture.
4.8
Comparative Systems: Blake and Fuller
While Frye elucidated the internal coherence of Blake’s system, recent comparative work highlights Blake’s
role as a system-builder akin to modern comprehensivists.
Friedman’s study of Blake and Buckminster
Fuller [Friedman, 2023] juxtaposes Blake’s mythopoetic architecture with the Synergetics of Fuller and Ed J
Applewhite. Both thinkers confronted the “single vision” of their respective eras—Newtonian mechanics for
Blake, specialization and technocracy for Fuller—by constructing comprehensive, fourfold (or tetrahedral)
epistemologies. This comparison underscores that Blake’s “system” was not a static dogma but a dynamic
tool for thought, designed to prevent enslavement by another’s system.
4.9
Phenomenological Traditions
The phenomenological tradition provides crucial methodological precedent. Merleau-Ponty’s Phenomenology
of Perception [Merleau-Ponty, 1962] argues that perception is fundamentally embodied—the body is not an
object among objects but the condition of objecthood itself. This directly parallels Active Inference’s claim
that the Markov blanket constitutes the distinction between agent and environment. Blake’s rejection of
Cartesian dualism—“Man has no Body distinct from his Soul”—anticipates Merleau-Ponty’s overcoming of
the mind-body problem through embodied intentionality.
9

## Page 10

Husserl’s concept of intentionality—that consciousness is always consciousness of something—prefigures the
Active Inference insight that inference is always inference about hidden states. The noematic content (what
is intended) depends on the noetic act (how it is intended), just as the posterior depends on how priors and
likelihoods are weighted. Blake’s “As a man is, so he sees” expresses this dependency of object on mode of
perception.
4.10
Extended Mind and 4E Cognition
Clark and Chalmers’ “Extended Mind” thesis [Clark and Chalmers, 1998] argues that cognitive processes
extend beyond the skull into environmental structures. This resonates with Blake’s insistence that imagina-
tion is not a private mental faculty but participates in a cosmic creativity: “Man is All Imagination God is
Man & exists in us & we in him.” The recursive embedding—existing in each other—describes precisely the
nested Markov blankets that enable multi-agent coordination.
The broader 4E cognition movement (embodied, embedded, enacted, extended) provides contemporary ar-
ticulation of Blake’s critique of disembodied reason. Varela, Thompson, and Rosch’s The Embodied Mind
[Varela et al., 1991] argues for the inseparability of cognition from sensorimotor engagement—exactly Blake’s
claim that “Energy is Eternal Delight” and that perception requires active participation, not passive recep-
tion.
4.11
Affect Theory and Precision Weighting
Contemporary affect theory illuminates the role of precision in shaping inference. Damasio’s somatic marker
hypothesis [Damasio, 1994] proposes that emotional states guide decision-making by tagging options with
bodily valence—a form of affective precision weighting. Blake’s Luvah (passion) and his claim that “thought
alone can make monsters, but the affections cannot” anticipates this: pure reasoning unmoored from bodily
affect produces biologically non-viable conclusions.
Precision weighting in Active Inference formalizes what matters: high precision signals “attend to this.”
Affect theory recognizes that mattering is not cognitive but visceral—we feel significance before we reason
about it. Blake’s repeated insistence on passion, energy, and delight as constitutive of vision (not decorative
enhancements) aligns with this affective grounding of inference.
4.12
Predictive Processing and Aesthetic Experience
The application of predictive processing to aesthetics has matured rapidly. The 2024 Philosophical Transac-
tions of the Royal Society B theme issue on art, aesthetics, and predictive processing [Van de Cruys et al.,
2024] represents a watershed: for the first time, a major scientific journal dedicated an entire volume to
exploring how the brain’s predictive architecture shapes aesthetic engagement. Van de Cruys, Bervoets, and
Moors argue that aesthetic experience arises from the interplay of order and change—precisely the dynamic
Blake dramatized as the “Marriage of Heaven and Hell,” where reason (order) and energy (change) are both
“necessary to Human existence.”
Kukkonen’s Probability Designs [Kukkonen, 2020] extends predictive processing to literary engagement, mod-
eling how readers generate predictions, encounter surprise, and update their models during narrative com-
prehension. This work provides methodological precedent for our approach: if predictive processing can
illuminate how readers engage with novels, it can equally illuminate how Blake’s prophetic structures engage
the perceptual system.
4.13
Consciousness as Controlled Hallucination
Anil Seth’s Being You [Seth, 2021] advances the thesis that all perception is a form of “controlled
hallucination”—the brain’s best guess about the causes of sensory signals, constrained but never determined
by incoming evidence. This language—perception as active construction rather than passive reception—
resonates strikingly with Blake’s insistence that we see “through” the eye, not “with” it.
Where Seth’s
framework emphasizes the constructive, model-dependent nature of all experience, Blake had already
10

## Page 11

proclaimed: “A fool sees not the same tree that a wise man sees” (Marriage of Heaven and Hell, Plate 7).
Both thinkers deny the Enlightenment premise that perception is simply the imprint of an external world
on a passive receiver.
4.14
Cognitive Romanticism
A new field is coalescing at the intersection of Romantic literary studies and cognitive science. Savarese’s
Romanticism’s Other Minds [Savarese, 2020] reassesses early relationships between Romantic poetry and sci-
entific thought, uncovering a “prehistory of cognitive approaches to literature” within the Romantic tradition
itself. The Romantic poets—Wordsworth, Coleridge, Shelley, and Blake—were not merely literary figures
but active theorists of mind, perception, and social cognition. Our paper extends this tradition by providing
what the Romantics lacked: the formal apparatus to make their deepest intuitions computationally precise.
4.15
Cultural Affordances and Shared Models
Veissière and colleagues [Veissière et al., 2020] apply Active Inference to cultural cognition, arguing that
shared generative models—“thinking through other minds”—constitute the mechanism of cultural transmis-
sion and niche construction. Their framework treats culture not as a static repository of information but as
a living system of shared priors, jointly updated through epistemic foraging and cooperative action. This
directly informs our reading of Blake’s Jerusalem: the city is not merely a utopian vision but a formally
specifiable shared generative niche, constructed and maintained through the “Mental Fight” of collective
inference.
4.16
Our Contribution
Prior scholarship has noted resonances between Romantic thought and cognitive science at the level of general
themes (embodiment, creativity, the limits of mechanism). Our paper is the first to provide specific formal
mappings between Blake’s prophetic system and the mathematical apparatus of Active Inference. We move
beyond analogy to structural correspondence: identifying not merely thematic overlap but shared topology
(the Markov blanket as Blake’s door), shared dynamics (free energy minimization as cleansing), and shared
architecture (hierarchical generative models as fourfold vision). This synthesis arrives at a moment when
both fields—predictive processing aesthetics and cognitive Romanticism—are independently converging on
the same questions. Our contribution is to continue work on that bridge (or at least point towards the gap
to be respected!).
11

## Page 12

5
Theoretical Foundations
The mathematics of self. This section reviews the formal apparatus of the Free Energy Principle and Active
Inference. We present the core formalisms—variational free energy, Markov blankets, hierarchical generative
models, precision weighting, prediction error, expected free energy, and multi-agent extensions—that the
subsequent synthesis will bring into structural alignment with Blake’s prophetic phenomenology.
5.1
The Free Energy Principle
Self-organizing systems persist by minimizing surprise (realism), or at least can be viewed as if they do
(instrumentalism). Friston’s Free Energy Principle (FEP) formalizes this imperative [Friston, 2010, Friston
et al., 2006], now comprehensively synthesized in Parr, Pezzulo, and Friston’s canonical textbook [Parr et al.,
2022].
Variational free energy provides a tractable upper bound on surprise (negative log model evidence):
𝐹= 𝔼𝑞[ln 𝑞(𝜃) −ln 𝑝(𝑜, 𝜃)]
(1)
where 𝑜denotes observations, 𝜃denotes hidden states (causes), 𝑞(𝜃) is a variational density encoding the
agent’s beliefs, and 𝑝(𝑜, 𝜃) is the generative model specifying how hidden states produce observations.
Decomposition reveals the relationship between free energy, divergence, and surprise:
𝐹= 𝐷𝐾𝐿[𝑞(𝜃)‖𝑝(𝜃|𝑜)] −ln 𝑝(𝑜)
(2)
Since KL-divergence is non-negative, free energy upper-bounds surprise:
𝐹≥−ln 𝑝(𝑜)
(3)
This bound is tight when 𝑞(𝜃) = 𝑝(𝜃|𝑜), i.e., when the agent’s beliefs equal the true posterior. Minimizing
𝐹thus serves two functions simultaneously: it makes beliefs more accurate (reducing the divergence term)
and implicitly minimizes surprise (the model evidence term).
5.1.1
Minimization Pathways
Two complementary pathways reduce free energy (Equation 1):
1. Perceptual inference — Update beliefs 𝑞(𝜃) toward the true posterior 𝑝(𝜃|𝑜). This is changing mind
to fit world.
2. Active inference — Select actions 𝑎that sample observations 𝑜consistent with predictions. This is
changing world to fit mind.
Both pathways reduce the same objective.
The agent that updates its beliefs and acts on the world is
performing complete free energy minimization.
5.1.2
Expected Free Energy and Policy Selection
Agents must also select among possible courses of action (policies 𝜋)).𝑇ℎ𝑒expected free energy𝐺(𝜋) eval-
uates policies by their anticipated consequences:
𝐺(𝜋) = −𝔼̃𝑞[ln 𝑝(𝑜𝜏|𝐶)] + 𝔼̃𝑞[𝐷𝐾𝐿[𝑞(𝜃𝜏|𝑜𝜏, 𝜋)‖𝑞(𝜃𝜏|𝜋)]]
(4)
where 𝐶encodes preferred observations (prior preferences), and
̃𝑞denotes the predictive density under the
policy. The first term drives the agent toward outcomes it prefers; the second drives it to resolve uncertainty
12

## Page 13

about hidden states.
Optimal policies minimize 𝐺(𝜋), balancing exploitation (pragmatic value) against
exploration (epistemic value) [Da Costa et al., 2020, Parr et al., 2022].
This decomposition is central to the synthesis that follows: it formally separates the habitual from the curious,
the routine from the exploratory—categories that recur throughout the humanistic tradition under different
names.
5.2
The Markov Blanket
The Markov blanket defines the statistical boundary of any autonomous system, partitioning states into
internal, external, and blanket (interface) components [Friston, 2019, Kirchhoff et al., 2018].
Conditional independence:
𝑝(𝜇|𝜂, 𝐵) = 𝑝(𝜇|𝐵)
(5)
Internal states 𝜇are conditionally independent of external states 𝜂)𝑔𝑖𝑣𝑒𝑛𝑏𝑙𝑎𝑛𝑘𝑒𝑡𝑠𝑡𝑎𝑡𝑒𝑠𝐵. The blanket com-
prises two complementary channels:
Component
Symbol
Flow Direction
Role
Sensory states
𝑠
World →Self
Carry
observations
Active states
𝑎
Self →World
Carry
interventions
Blanket
𝐵= {𝑠, 𝑎}
Bidirectional
The statistical
interface
Every self-organizing system—from cell to organism to social group—possesses a Markov blanket. The blan-
ket is constitutive: without it, there is no distinction between system and environment, hence no inference.
The topology of this partition—what is inside, what is outside, what mediates—determines the scope and
character of an agent’s engagement with its world.
5.2.1
Nested Blankets and Multi-Scale Organization
Markov blankets nest recursively: cells within organs, organs within organisms, organisms within social
groups. Each scale defines its own internal/external partition and performs its own inference [Kirchhoff
et al., 2018, Ramstead et al., 2018].
This nesting is not merely a descriptive convenience but a formal
property of hierarchical self-organization.
5.3
Hierarchical Generative Models
Generative models are typically layered, with each level predicting the activity of the level below [Clark,
2016, Hohwy, 2013].
Hierarchical factorization:
𝑝(𝑜, 𝜃) = 𝑝(𝑜|𝜃1)
𝑛−1
∏
𝑖=1
𝑝(𝜃𝑖|𝜃𝑖+1) ⋅𝑝(𝜃𝑛)
(6)
At the lowest level, 𝜃1 generates observations through the likelihood 𝑝(𝑜|𝜃1). Each higher level 𝜃𝑖+1 provides
the prior context for the level below.
The deepest level 𝜃𝑛encodes the most abstract, slowly varying
regularities of the environment.
This architecture has several key properties:
13

## Page 14

• Abstraction increases with depth. Low levels encode fast sensory features; high levels encode slow
contextual structure.
• Temporal scale separation. Higher levels change more slowly, providing a stable context for faster
dynamics below [Kiebel et al., 2008, Friston et al., 2017].
• Bidirectional message passing. Top-down predictions and bottom-up prediction errors flow through
the hierarchy, settling jointly to minimize free energy.
The depth of the hierarchy determines the scope of patterns the model can represent—from local texture to
global meaning.
5.3.1
Model Evidence and Complexity
The marginal likelihood (model evidence) quantifies how well a generative model accounts for observations:
ln 𝑝(𝑜) = 𝔼𝑞[ln 𝑝(𝑜|𝜃)] −𝐷𝐾𝐿[𝑞(𝜃)‖𝑝(𝜃)]
(7)
Good models maximize accuracy while minimizing complexity—a formal instantiation of Occam’s razor.
Overly simple models are inaccurate; overly complex models overfit.
The free energy bound (Equation
3) ensures that minimizing 𝐹implicitly maximizes model evidence, favoring parsimonious yet accurate
explanations.
Model comparison:
𝐹simple ≫𝐹rich
(8)
A model of insufficient depth incurs high free energy because it cannot account for the hierarchical structure
of observations. A richer model, one with appropriate depth and structure, achieves lower free energy by
capturing regularities that the shallow model misses (though a larger model may have other tradeoffs or
penalization terms applied, balancing the tendency to inflate the number of parameters).
5.4
Precision
Precision is the inverse variance of a probability distribution—a measure of confidence or reliability:
𝜋= 𝜎−1
(9)
In hierarchical inference, precision weights determine how strongly each level of the hierarchy influences the
overall posterior. Two sources of precision compete at every level:
• Prior precision (𝜋prior): confidence in top-down predictions
• Sensory precision (𝜋sensory): confidence in bottom-up evidence
Their balance determines the character of inference:
Regime
Condition
Perceptual Effect
Prior-dominated
𝜋prior ≫𝜋sensory
Expectations override evidence; hallucination-like
states
Sensory-dominated
𝜋sensory ≫𝜋prior
Sensory flooding; loss of contextual interpretation
Balanced
𝜋prior ≈𝜋sensory
Optimal inference; accurate and contextually rich
perception
Attention, in this framework, is the optimization of precision—the process by which the brain infers the
reliability of its own prediction errors and weights them accordingly [Feldman and Friston, 2010, Parr and
Friston, 2019].
14

## Page 15

5.4.1
Precision Dynamics and Pathology
When prior precision becomes extreme:
Prior dominance:
𝜋prior ≫𝜋sensory
(10)
the agent’s beliefs become insensitive to new evidence. The generative model ceases to update, and perception
rigidifies. Conversely, when sensory precision vastly exceeds prior precision, the agent is overwhelmed by
unstructured input, unable to extract meaning. Pathological states—from delusions to anxiety disorders—
can be understood as failures of precision optimization [Adams et al., 2013].
5.5
Prediction Error and Message Passing
At each level of the hierarchy, the brain computes prediction error—the discrepancy between what was
expected and what was observed:
𝜀𝑖= 𝑜𝑖−𝑔𝑖(𝜃𝑖+1)
(11)
where 𝑔𝑖(⋅) is the generative function mapping higher-level states to predicted observations at level 𝑖. Errors
ascend the hierarchy; predictions descend. The system settles when 𝜀→0 across all levels—when predictions
match observations at every scale.
Each error signal (Equation 11) propagates through the hierarchy defined in Equation 6, weighted by the pre-
cision (Equation 9) assigned to that level. High-precision errors demand model revision; low-precision errors
are discounted. This precision-weighted prediction error is the fundamental currency of hierarchical
inference.
The bidirectional cascade of predictions and errors constitutes perception itself: a continuous, iterative
process of generating hypotheses, testing them against evidence, and revising.
Action enters when the
system changes the world to reduce prediction error rather than changing beliefs.
5.6
Temporal Depth
Generative models can extend across time, encoding dependencies between successive observations:
Temporal hierarchy:
𝑝(𝑜1∶𝑇, 𝜃) =
𝑇
∏
𝑡=1
𝑝(𝑜𝑡|𝜃𝑡) ⋅𝑝(𝜃𝑡|𝜃𝑡−1)
(12)
Higher levels of the hierarchy encode slower dynamics, providing a context for the faster fluctuations below.
The lowest levels track moment-to-moment sensory input; intermediate levels integrate over seconds to
minutes; the deepest levels encode regularities persisting across hours, years, or longer [Kiebel et al., 2008,
Friston et al., 2017].
The temporal depth of a model determines how far into the past and future its predictions extend. A
shallow model is reactive, bound to immediate stimulus; a deep model integrates broad temporal context into
present inference. Extending temporal depth imposes computational cost but enables the agent to detect
and exploit regularities that span long durations.
15

## Page 16

5.7
Multi-Agent Inference
Active Inference extends naturally to systems of coupled agents, each bounded by its own Markov blanket
but sharing statistical structure:
Multi-agent coordination:
𝑝(𝑜, 𝜃) =
𝑁
∏
𝑖=1
𝑝(𝑜𝑖|𝜃𝑖) ⋅𝑝(𝜃𝑖|𝜃shared) ⋅𝑝(𝜃shared)
(13)
Multiple agents share a common prior 𝜃shared—the cultural, institutional, or ecological generative model
that aligns their individual inferences. Communication between agents can be formalized as generalized
synchronization, where coupled systems entrain their internal dynamics to infer each other’s hidden states
[Friston and Frith, 2015, Veissière et al., 2020].
5.7.1
Mean-Field Factorization
When the joint posterior over all hidden states is intractable, variational inference approximates it by as-
suming independence between factors:
Mean-field approximation:
𝑞(𝜃) ≈
𝐾
∏
𝑘=1
𝑞(𝜃𝑘)
(14)
This factorization makes computation tractable but introduces coordination costs: correlations between
components are lost. The quality of inference depends on how well the factorization structure matches the
true dependencies in the generative model. Structured variational families that preserve key correlations
improve upon the fully factorized approximation.
This formalized understanding of collective intelligence provides the necessary bridge to the aesthetic domain.
If culture is a shared generative model, then art is the engineering of that model—a “cognitive” intervention
that reshapes the priors of the collective.
5.8
Cognitive Art and the Fourfold
The integration of Active Inference with broad-scale historical and aesthetic systems suggests a “cogni-
tive art”—a practice of mind that is both rigorous and generative. Friedman’s recent work on “Cognitive
Art & Science” [Friedman, 2025] proposes a fourfold schema for intelligence that maps directly onto the
Blakean/Fristonian synthesis. This framework distinguishes between the “Low Road” (2 →3) of explana-
tory modeling—fitting data to priors—and the “High Road” (4 →3) of anticipatory wisdom—shaping the
niche to afford new forms of life. Blake’s rejection of “Single Vision” (pure 2nd-ness) in favor of “Fourfold
Vision” (integrated 1st, 2nd, 3rd, and 4th-ness) prefigures the move from mere error minimization to the
active construction of a “wise” sensorimotor niche.
5.9
Summary of Formal Apparatus
The following table collects the core equations and their roles in the synthesis that follows:
Equation
Name
Role in Synthesis (§4)
1
Variational Free
Energy
Objective function for perception and action
2
FEP Decomposition
Relation of divergence and surprise
3
Surprise Bound
Evidence lower bound (ELBO) logic
16

## Page 17

Equation
Name
Role in Synthesis (§4)
4
Expected Free
Energy
Policy selection (exploration/exploitation)
5
Conditional
Independence
Markov blanket as statistical boundary
6
Hierarchical
Factorization
Depth of generative model
7
Model Evidence
Accuracy–Complexity trade-off
19
Model Comparison
Necessity of hierarchical depth
9
Precision
Confidence weighting
16
Prior Dominance
Pathological rigidity
11
Prediction Error
Bidirectional message passing
18
Temporal Hierarchy
Depth of temporal prediction
21
Multi-Agent
Coordination
Shared priors and collective inference
23
Mean-Field
Approximation
Factorized variational inference
Each of these formalisms will be brought into structural alignment with a specific aspect of Blake’s prophetic
phenomenology in the sections that follow.
17

## Page 18

6
Synthesis: Eight Themes of Vision
Hold your hand in front of your face. The boundary between your skin and the surrounding air is your
Markov blanket—the statistical interface through which all inference flows. Blake called this the “door of
perception.” Two names for the same structure, separated by two centuries of intellectual history.
In what follows, we trace eight such structural identities between Blake’s prophetic vision and the Active In-
ference framework, each demonstrated through specific perceptual scenarios that ground abstract formalism
in lived experience. Each theme begins with Blake’s fire—his phenomenological observation, expressed in
the language of prophecy and illuminated printing—and then traces its mathematical shadow in the equa-
tions, architectures, and dynamics of computational neuroscience. The correspondences are not approximate
analogies but precise structural mappings: shared topology (the Markov blanket as Blake’s door), shared
dynamics (free energy minimization as the cleansing of perception), shared architecture (hierarchical genera-
tive models as fourfold vision), and shared pathology (prior dominance as Newton’s sleep). Figure 1 provides
an overview of these eight thematic correspondences arranged as a visual atlas.
Theme
Blake
Friston
Identity
Boundary
Doors of Perception
Markov Blanket
Interface topology
Vision
Fourfold Vision
Hierarchical Model
Processing depth
States
Newton’s Sleep
Prior Dominance
Cognitive rigidity
Imagination
Human Existence
Generative Model
Agent identity
Time
Eternity in Hour
Temporal Horizons
Prediction depth
Space
World in Grain
Spatial Hierarchy
Evidence integration
Action
Cleansing
Free Energy Minimization
Optimization
Collectives
Building Jerusalem /
Four Mighty Ones
Shared & Factorized
Models
Multi-agent coordination
18

## Page 19

Figure 1: Thematic Atlas. Eight structural correspondences between Blake’s prophetic vision (left col-
umn) and Active Inference formalism (right column), connected by bidirectional arcs indicating the nature
of each mapping. Themes span boundary topology (Doors/Markov Blanket), processing hierarchy (Fourfold
Vision/Hierarchical Models), cognitive rigidity (Newton’s Sleep/Prior Dominance), agent identity (Imagina-
tion/Generative Model), temporal depth (Eternity in Hour/Temporal Horizons), spatial inference (World
in Grain/Scale Invariance), optimization (Cleansing/Free Energy Minimization), and collective coordina-
tion with modular cognition (Jerusalem & Zoas/Shared & Factorized Models). Color-coding groups related
themes; each correspondence is developed in a dedicated subsection below.
19

## Page 20

6.1
Boundary: The Doors of Perception
“The cherub with his flaming sword is hereby commanded to leave his guard at the tree of life, and
when he does, the whole creation will be consumed and appear infinite and holy, whereas it now
appears finite and corrupt.”
— Marriage of Heaven and Hell, Plate 14 [Blake, 1988b]
6.1.1
The Markov Blanket
The cherub’s flaming sword is the guardian of a boundary—the partition between accessible and inaccessible
states. Blake’s “door” is Friston’s boundary—the statistical partition separating internal from external states:
“The Markov blanket defines what is inside vs. outside any autonomous system—the statistical
partition separating internal from external states.”
— Kirchhoff et al. (2018) [Kirchhoff et al., 2018]
This IS Blake’s “door”—the boundary that mediates all contact between self and world. The blanket is
constitutive, not optional.
Blanket Component
Symbol
Blake’s Image
External states
𝜂)
“the Infinite” beyond
Sensory states
𝑠
Inflow through doors
Active states
𝑎
Outflow through doors
Internal states
𝜇
The perceiver in the “cavern”
Blanket
𝐵
“The doors of perception”
This conditional independence structure (Equation 5) means that the blanket mediates all contact. Internal
states access external states only through the interface. The doors are not optional—they are constitutive.
Blake’s “cavern” is not metaphor but phenomenology: the subjective space of one whose doors are narrowed.
The “chinks” are the impoverished sensory channels of a rigid generative model.
Blake articulated this boundary condition repeatedly. The bounded itself produces suffering:
“The Bounded is loathed by its possessor. The same dull round, even of a Universe, would soon
become a Mill with complicated wheels.”
— There is No Natural Religion, Series B [Blake, 1988a]
Yet energy is the fundamental currency crossing the boundary:
“Energy is the only life and is from the Body and Reason is the bound or outward circumference
of Energy. Energy is Eternal Delight.”
— Marriage of Heaven and Hell, Plate 4 [Blake, 1988b]
Blake identifies Reason as the “bound”—the circumference or blanket edge that delimits the system. Energy,
by contrast, is the vital flow that crosses this boundary. From this distinction follows his foundational claim
about the nature of perception:
“Man’s Perceptions are not bounded by Organs of Perception; he perceives more than Sense (tho’
ever so acute) can discover.”
— There is No Natural Religion, Series B [Blake, 1988a]
The Markov blanket is necessary but not sufficient. The door mediates; cleansing transforms how it mediates.
Figure 2 illustrates this statistical boundary and its Blakean phenomenology.
20

## Page 21

Demonstration: The Sunrise
Blake famously contrasted two ways of seeing the sun:
“When the sun rises, do you not see a round disk of fire somewhat like a guinea? O no, no, I see
an Innumerable company of the Heavenly Host crying, ‘Holy, Holy, Holy.’ ”
Both observers share the same Markov blanket—the same sensory channels, the same visual
cortex. What differs is how the door mediates:
Component
Friend’s Experience
Blake’s Experience
Sensory states
(𝑠)
Photons →“round, golden, ~30°
altitude”
Same photons →rich associative
cascade
Internal states
(𝜇)
Minimal categorical prediction:
“sun”
Full generative model: cosmic
meaning
Active states (𝑎)
Glance, categorize, move on
Sustained attention, devotional
engagement
Same door.
Different cleansing. The friend’s perception is not wrong—but it is shallow.
Blake’s perception engages deeper layers of the generative model. Both are valid inferences; one
draws on vastly more model depth.
6.1.2
Boundary Constitution: Naming Creates Separation
Blake understood that boundaries are constituted, not given. The act of naming creates the inside/outside
distinction:
“they gave to it a Space & namd the Space Ulro”
— Vala, or The Four Zoas, Night the First [Blake, 1988g, E303]
The Markov blanket is not discovered but constituted. “Naming the Space” IS boundary formation. Ulro—
Blake’s realm of materialist limitation—comes into being through the act of partition.
6.1.3
The Abyss as KL-Divergence
When Urizen separates from Ahania (his emanation, his feminine counterpart), Blake describes the resulting
gap in terms that directly anticipate information-theoretic distance:
“Ahania (so name his parted soul)… how wide the Abyss Between Ahania and thee!”
— The Book of Ahania, Chapter III [Blake, 1988e]
This “Abyss” evokes information-theoretic distance—the separation between states that should be unified.
Blake’s mythic language of partition anticipates what Active Inference formalizes as divergence between
distributions. The “parted soul” represents a split generative model; the wider the Abyss, the greater the
separation.
6.1.4
Forged Boundaries
The blanket is not simply given and absolute, but constructed:
“He forg’d nets of iron around”
— The Book of Ahania [Blake, 1988e]
“Forging” emphasizes the active construction and identification of boundary conditions. The Markov blanket
is manufactured materially and mentally. This has profound implications: what is forged can be reforged,
and what is identified can be mis- and re-identified.
21

## Page 22

6.1.5
Body as Blanket Interface
Blake’s most direct statement of embodied cognition anticipates the blanket formalism:
“Man has no Body distinct from his Soul for that calld Body is a portion of Soul discernd by the
five Senses, the chief inlets of Soul in this age”
— Marriage of Heaven and Hell, Plate 4 [Blake, 1988b]
“Inlets” = sensory states 𝑠. The body IS how mind interfaces with world—not a separate substance but the
blanket itself. No ontological separation exists; only the functional partition of the blanket.
6.1.6
Imagination’s Outline vs. Nature’s Dissolution
Blake’s late work makes explicit the distinction between raw data and model structure:
“Nature has no Outline: but Imagination has. Nature has no Tune: but Imagination has. Nature
has no Supernatural & dissolves: Imagination is Eternity”
— The Ghost of Abel, Plate 1 [Blake, 1988m]
Nature (observations 𝑜) is unstructured flow. Imagination (the generative model 𝑝(𝑜, 𝜃)) provides boundaries,
form, temporal structure (“Tune”). The model is more real than the data because it is what makes data
intelligible. Without the model’s outline, perception dissolves into chaos.
6.1.7
Contemporary Resonance: From Blake Through Huxley to ALBUS
The lineage from Blake’s “Doors” to contemporary neuroscience runs through a single remarkable chain.
Aldous Huxley borrowed Blake’s phrase for The Doors of Perception [Huxley, 1954], proposing that the
brain operates as a “reducing valve” filtering the totality of experience into manageable form.
Carhart-
Harris and Friston’s REBUS model [Carhart-Harris and Friston, 2019] formalized this intuition by showing
that psychedelics relax the precision of high-level priors. Safron and colleagues’ ALBUS framework [Safron
et al., 2025] now extends this account: psychedelics do not merely relax beliefs (REBUS) but can also
strengthen them (SEBUS), producing a richer taxonomy of altered states—from the dissolution of rigid priors
to the intensification of meaning-making. The result encompasses Blake’s “cleansing”—not the destruction
of the boundary but the recalibration of its precision weighting, allowing prediction error to propagate more
freely up the hierarchy. What Blake described as seeing “every thing… as it is: infinite” corresponds to a
state of altered prior dynamics where sensory evidence reshapes inference rather than being suppressed by
entrenched expectations. This is not metaphor: it is the same computational operation described in different
vocabularies across three centuries.
22

## Page 23

Figure 2: The Doors of Perception as Markov Blanket. The statistical boundary (𝐵= {𝑠, 𝑎}) par-
titions external states 𝜂)(‶𝑡ℎ𝑒𝐼𝑛𝑓𝑖𝑛𝑖𝑡𝑒″)𝑓𝑟𝑜𝑚𝑖𝑛𝑡𝑒𝑟𝑛𝑎𝑙𝑠𝑡𝑎𝑡𝑒𝑠𝜇(“the Cavern”) through two complementary
channels: sensory states 𝑠mediating world-to-self flow (observation, perception) and active states 𝑎medi-
ating self-to-world flow (action, decision). This implements Blake’s phenomenology from The Marriage of
Heaven and Hell, Plate 14: “If the doors of perception were cleansed every thing would appear to man as it
is, Infinite. For man has closed himself up, till he sees all things thro’ narrow chinks of his cavern” [Blake,
1988b]. The blanket is constitutive—not an optional filter but the necessary interface through which all
inference occurs. Cleansing the doors corresponds to optimizing the model’s precision weighting (Equation
5), not to eliminating the boundary itself.
23

## Page 24

6.2
Vision: The Fourfold Hierarchy
“The great City of Golgonooza: fourfold toward the north / And toward the south fourfold, &
fourfold toward the east & west / Each within other toward the four points”
— Jerusalem, Plate 12 [Blake, 1988i]
6.2.1
Hierarchical Generative Models
Golgonooza—Blake’s city of art, built by the imagination—provides the architectural metaphor for hierarchi-
cal inference: fourfold in every direction, each level nested within the others. The predictive brain generates
perception actively:
“The brain is revealed as an active, generative organ: one that continually predicts its own current
sensory states, using those predictions to explain away the incoming sensory signal.”
— [Clark, 2013]
This bidirectional cascade IS perception—errors ascend, predictions descend:
“Feedback connections from a higher- to a lower-order visual cortical area carry predictions of
lower-level neural activities, whereas the feedforward connections carry the residual errors be-
tween the predictions and the actual lower-level activities.”
— [Rao and Ballard, 1999]
Four levels of perception correspond to four depths of the generative hierarchy:
Blake Level
Symbol
Cognitive Mode
Processing
Single (Ulro)
𝜃1
Quantitative
Sensory features
Twofold (Generation)
𝜃2
Emotional
Affective encoding
Threefold (Beulah)
𝜃3
Imaginative
Symbolic integration
Fourfold (Jerusalem)
𝜃4
Unified
Complete model
engagement
Fourfold hierarchical factorization:
𝑝(𝑜, 𝜃1∶4) = 𝑝(𝑜|𝜃1)
3
∏
𝑖=1
𝑝(𝜃𝑖|𝜃𝑖+1) ⋅𝑝(𝜃4)
(15)
Fourfold vision engages all levels of the hierarchy (Equation 15; see Figure 3). Single vision collapses to
𝜃1 alone, reducing the general hierarchical model (Equation 6) to a single layer.
The hierarchy is not
ornament—it is the architecture of meaning.
Blake grasped this hierarchical principle:
“The Eye sees more than the Heart knows.”
— Visions of the Daughters of Albion, title page [Blake, 1988c]
Even the lower level (eye/sensation) accesses more than higher cognition (heart/understanding) can process.
The crooked roads of genius circumvent linear reasoning:
“Improvement makes strait roads; but the crooked roads without Improvement are roads of Genius.”
— Marriage of Heaven and Hell, Proverbs of Hell [Blake, 1988b]
Hierarchy need not mean rigid order—the genius finds shortcuts through visionary compression. Worton’s
analysis of Blake’s intertextuality reveals that these “crooked roads” function as radical reconfigurations of
existing models, not mere deviations from linearity [Worton, 2003].
24

## Page 25

6.2.2
Golgonooza: The Architecture of the Generative Model
Blake’s mythic city Golgonooza—the city of art, built by Los the imagination—provides a structural diagram
of hierarchical inference. Recall the passage quoted at the opening of this section: “fourfold toward the north
/ And toward the south fourfold, & fourfold toward the east & west / Each within other toward the four
points.”
Four directions = four hierarchical levels.
“Each within other” = nested structure.
Golgonooza IS the
generative model’s architecture—a city that is simultaneously spatial and cognitive, built from the material
of imagination itself.
The fourfold structure extends in all dimensions:
north/south/east/west map to the Four Zoas
(Urthona/Urizen/Luvah/Tharmas), each representing a distinct mode of inference.
The city is not
static but perpetually under construction—Los labors at the furnaces, continually rebuilding the model.
6.2.3
Organs of Perception as Model-Dependent
Blake makes explicit that perception is not passive reception but active model-dependent construction:
“Creating Space, Creating Time… such was the variation of Time & Space, which vary according
as the Organs of Perception vary”
— Jerusalem, Plate 98 [Blake, 1988i]
Space and time are not objective containers but generative model outputs. Different models produce different
space-times. The “Organs of Perception” are not fixed biological apparatus but the structure of inference
itself—and this structure can vary.
This anticipates the Active Inference insight that even basic phenomenal properties like spatial extent and
temporal duration are inferred, not given. The model creates the coordinate system within which observations
are interpreted.
Anil Seth’s contemporary formulation crystallizes this point: all perception is a “controlled hallucination”—
the brain’s best guess about the causes of sensory signals, constrained but not determined by incoming
evidence [Seth, 2021]. Blake’s fourfold vision is, in these terms, a taxonomy of hallucination depths: single
vision is a shallow, rigid hallucination dominated by sensory constraint; fourfold vision is a deep, flexible
hallucination where the generative model’s own creative structure participates fully in what is perceived.
The “fool” and the “wise man” who see different trees are running different models on the same data—and
both perceptions are, in Seth’s precise sense, controlled hallucinations.
Blake was acutely aware that deeper vision is not merely unseen but actively pathologized by the regime of
single vision. In Milton, he names this suppression directly:
“Calling the Human Imagination: which is the Divine Vision & Fruition In which Man liveth
eternally: madness & blasphemy, against Its own Qualities, which are Servants of Humanity, not
Gods or Lords.”
— Milton, Plate 32 [Blake, 1988j]
“Madness & blasphemy” is the diagnostic frame that prior-dominated inference applies to perception that
exceeds its own model. From within Newton’s Sleep, fourfold vision looks pathological precisely because
the shallow model cannot represent the hierarchical depth that makes it possible—it can only classify what
it cannot compute as error, delusion, or transgression. Blake’s counter-move is to insist that imagination’s
“Qualities” are “Servants of Humanity, not Gods or Lords”: the deeper levels of the generative model serve
the agent’s self-evidencing; they are not external authorities but functional capacities.
This anticipates
contemporary debates in psychedelic neuroscience, where expanded perceptual states—once dismissed as
mere hallucination—are increasingly recognized as alternative precision regimes with their own epistemic
validity [Carhart-Harris and Friston, 2019].
25

## Page 26

Figure 3: The Fourfold Vision Hierarchy.
Blake’s four perceptual levels mapped to corresponding
depths of the Active Inference hierarchical generative model (Equation 15). Single Vision (Ulro, 𝜃1, gray):
quantitative sensory registration—“Newton’s sleep,” seeing a rose as cells and chemistry. Twofold Vision
(Generation, 𝜃2, blue): emotional-intellectual engagement—perceiving beauty, desire, and symbolic meaning.
Threefold Vision (Beulah, 𝜃3, purple): imaginative synthesis—“soft Beulah’s night,” where contraries
reconcile in art and myth. Fourfold Vision (Jerusalem, 𝜃4, gold): full hierarchical integration—“supreme
delight,” unified engagement of all model depths. Left column: Blake’s phenomenological descriptions; right
column: Active Inference processing levels. Ascending arrows indicate increasing hierarchical depth and
precision integration. Source: Letter to Thomas Butts, 22 November 1802 [Blake, 1802].
26

## Page 27

6.3
States: Newton’s Sleep & Prior Dominance
“If the Sun & Moon should doubt / They’d immediately Go out.”
— Auguries of Innocence [Blake, 1988h]
“The will of the Immortal expanded / Or contracted his all-flexible senses”
— Book of Urizen, Plate 3 [Blake, 1988d]
6.3.1
Prior Dominance
These two quotations frame the full spectrum of precision dynamics. The Sun and Moon that never doubt
represent the cosmic necessity of confident priors—without them, the world “goes out.” Yet the Immortal’s
“all-flexible senses” describe the capacity to modulate those priors at will. “Newton’s sleep” is what happens
when flexibility is lost. It is rigid inference—the condition where prior beliefs overwhelm sensory evidence.
Active Inference formalizes this:
“Attention can be understood as inferring the level of uncertainty or precision during hierarchical
perception.”
— Feldman & Friston (2010) [Feldman and Friston, 2010]
Parr and Friston further develop this insight, showing that attention optimizes the precision of prediction
errors at every level of the cortical hierarchy, selecting which sensory channels carry reliable information
[Parr and Friston, 2019].
When prior precision vastly exceeds sensory precision:
Prior dominance:
𝜋prior ≫𝜋sensory
(16)
When prior precision vastly exceeds sensory precision (Equation 16), observations are discounted. The world
conforms to expectation—free energy (Equation 1) ceases to drive model revision because the precision
weighting (Equation 9) favors priors. This is Blake’s “Urizen”—your reason frozen into horizon (limit).
Condition
Effect
Blake’s Term
𝜋prior ≫𝜋sensory
Expectations dominate
“Newton’s sleep”
𝜋sensory ≫𝜋prior
Sensory overwhelm
Chaos, dissolution
𝜋prior ≈𝜋sensory
Optimal inference
“Cleansed perception”
The “guinea sun”—Blake’s mockery of seeing the sun as mere golden disk—is exactly this: prior-locked
inference refusing sensory update. Figure 4 illustrates the contrast between prior-dominated and balanced
inference.
Demonstration: The Familiar Street
Walk down a street you travel daily. You may fail to notice a new shopfront, a repainted door,
a changed sign. Your prior model (“this street looks like X”) overwhelms sensory evidence of
change. This is 𝜋prior ≫𝜋sensory in action.
Blake’s Urizen has “contracted his all-flexible senses” into “little orbs… hiding from the wind.”
The narrowed perception is not sensory failure but model rigidity.
Contrast: Tourist Vision
Now visit a new city. Everything glows with detail. Unfamiliar streets demand notice—each
doorway, sign, and face registers distinctly. This is 𝜋prior ≈𝜋sensory—balanced precision where
the model cannot coast on expectation.
27

## Page 28

The tourist sees more not because their eyes are better, but because their priors are weaker. Their
doors are cleansed by unfamiliarity.
Figure 4: Newton’s Sleep vs. Cleansed Perception. Two contrasting precision regimes illustrated as
balance beams. Left panel (“Newton’s Sleep”): prior precision 𝜋prior vastly exceeds sensory precision 𝜋sensory
(large red weight vs. small teal weight), producing rigid, expectation-dominated inference where “the will
of the Immortal… contracted his all-flexible senses” (Book of Urizen, Plate 3 [Blake, 1988d]). Observations
are discounted; the world conforms to frozen expectation. Right panel (“Cleansed Perception”): balanced
precision weighting (𝜋prior ≈𝜋sensory, equal green and teal weights) enables optimal inference where “every
thing would appear to man as it is, Infinite” (Marriage of Heaven and Hell, Plate 14 [Blake, 1988b]). The
balance beam metaphor captures the precision dynamics formalized in Equation 16. Blake’s “guinea sun”
exemplifies the left state; his “Innumerable company of the Heavenly Host” exemplifies the right.
The question becomes: is Newton’s sleep a permanent condition, or a temporary one? Blake’s answer is
emphatic—and computationally significant. In a pivotal passage from Milton, he develops the distinction
with systematic precision:
“We are not Individuals but States: Combinations of Individuals We were Angels of the Divine
Presence…. Calling the Human Imagination: which is the Divine Vision & Fruition In which
Man liveth eternally: madness & blasphemy, against Its own Qualities, which are Servants of
Humanity, not Gods or Lords.
Distinguish therefore States from Individuals in those States.
States Change: but Individual Identities never change nor cease: You cannot go to Eternal Death
in that which can never Die. Satan & Adam are States Created into Twenty-seven Churches….
States that are not, but ah! Seem to be.”
— Milton, Plate 32 [Blake, 1988j]
This passage is computationally dense.
“Combinations of Individuals” reframes what a state is: not a
personal mood but a composite model configuration—a particular factorization of beliefs, precisions, and
action policies that many agents can share. “Twenty-seven Churches” names a discrete, enumerable state-
space: Satan and Adam are not persons but attractors in model-space, recurring configurations that agents
28

## Page 29

fall into and can emerge from.
The crucial claim—“States that are not, but ah!
Seem to be”—is the
recognition that states feel ontologically real from within but are transient configurations of the generative
model, not permanent features of reality.
Equally striking is Blake’s simultaneous attack on reductive spatial models within the same passage:
“those combind by Satans Tyranny… are Shapeless Rocks Retaining only Satans Mathematic Ho-
liness, Length: Bredth & Highth”
— Milton, Plate 32 [Blake, 1988j]
“Mathematic Holiness”—the worship of pure quantitative extension—is Blake’s name for the impoverished
generative model that retains only Euclidean coordinates. When imaginative depth is stripped away, what
remains is geometry without meaning: “Shapeless Rocks” that have collapsed to minimum model complexity.
This directly parallels the Active Inference diagnosis of Newton’s Sleep: a state where the model’s rich
hierarchical structure has been flattened to shallow, quantitative prediction.
The sleeping perceiver is in a state, not an identity. States change; the door can be cleansed. The miser
is Blake’s figure for this pathology: hoarding certainty like gold, refusing sensory update, clinging to priors
that have calcified into identity.
And the transformation:
“If the fool would persist in his folly he would become wise.”
— Marriage of Heaven and Hell, Proverbs of Hell [Blake, 1988b]
Persistence breaks through prior rigidity. The fool’s persistence is a form of precision reweighting—eventually,
accumulated evidence overwhelms the prior.
Blake records this exchange in a famous passage (see §4.1 Boundary): his friend perceives only a golden disk;
Blake perceives the Heavenly Host. The friend’s response represents prior-locked perception; Blake’s answer
demonstrates cleansed vision.
6.3.2
The Spectre’s Steel Ratio: Prior Dominance Formalized
Blake’s most precise formulation of prior-dominated inference appears in Jerusalem:
“The Spectre is the Reasoning Power in Man; & when separated / From Imagination, and closing
itself as in steel, in a Ratio / Of the Things of Memory. It thence frames Laws & Moralities /
To destroy Imagination!”
— Jerusalem, Plate 74, lines 10-13 [Blake, 1988i]
This is the complete phenomenology of rigid inference:
• “Closing itself as in steel” = 𝜋prior →∞(infinite prior precision)
• “Ratio of Things of Memory” = frozen historical priors refusing sensory update
• “Laws & Moralities” = rigid predictive structures that suppress model revision
• “Destroy Imagination” = elimination of model flexibility
The Spectre is not evil but separated—cut off from the generative model’s creative capacity to revise itself.
When reasoning power closes itself off from imagination, it becomes a self-reinforcing loop of confirmation
bias.
6.3.3
The Philosophy of Five Senses
Blake traces the intellectual history of perceptual restriction:
“Till a Philosophy of Five Senses was complete / Urizen wept & gave it into the hands of Newton
& Locke”
— The Song of Los, Plate 4 [Blake, 1988f]
29

## Page 30

The “Philosophy of Five Senses” is the sensory bottleneck doctrine—the restriction of inference to immediate
observations without hierarchical depth. Newton and Locke represent single-level empiricism: the belief that
knowledge comes only from sensory evidence, without recognizing the generative models that make evidence
intelligible.
Urizen “weeps” because even he—the principle of rational limitation—recognizes this as impoverishment.
The complete collapse to sensory level is Newton’s sleep at civilizational scale.
6.3.4
Descending into Division
The Fall is described as a collapse through hierarchical levels:
“I was divided: descending down I sunk along / The goary tide even to the place of seed & there
/ Dividing I was buried”
— Vala, or The Four Zoas, Night the First [Blake, 1988g]
“Division” = separation from higher-level update.
“Descending” = collapse to lower hierarchical levels.
“Buried” = inference locked in prior state, unable to revise.
The “goary tide” suggests the violence of this collapse—the tearing apart of an integrated model into isolated
fragments, each trapped in its own local minimum.
6.3.5
The Natural Man vs. Spiritual Man
Blake’s critique of Wordsworth illuminates the tension between likelihood and prior:
“I see in Wordsworth the natural man rising up against the spiritual man continually”
— Annotations to Wordsworth’s Poems [Blake, 1826]
“Natural man” = sensory likelihood 𝑝(𝑜|𝜃). “Spiritual man” = prior beliefs 𝑝(𝜃). Wordsworth, in Blake’s
view, over-weights sensory precision at the expense of imaginative priors. The “rising up” is the dominance
of bottom-up over top-down—the inverse of Newton’s sleep, but equally unbalanced.
6.3.6
Hemispheric Resonance
McGilchrist’s hemispheric hypothesis [McGilchrist, 2009, 2021] provides a neuroanatomical substrate for
the pathology Blake diagnoses.
The left hemisphere’s mode of attention—narrow, focused, categorical,
already-knowing—maps precisely onto Newton’s Sleep: the prior-dominated regime where familiar categories
suppress novel inference.
Urizen is the left hemisphere enthroned, the lawgiver who “closed the tent of
the Universe” by imposing rigid categorical boundaries. Conversely, Blake’s Los—the creative imagination
perpetually rebuilding the model at the furnaces of Golgonooza—embodies what McGilchrist identifies as
the right hemisphere’s broader, contextual, novelty-seeking attention. The clinical implication is striking: if
Newton’s Sleep is a form of hemispheric imbalance, then “awakening” (restoring balanced precision weighting)
may require not more data but a different mode of attention—a shift in which hemisphere leads the inference.
30

## Page 31

6.4
Imagination: The Generative Model
“Man is All Imagination God is Man & exists in us & we in him”
— Annotations to Berkeley’s Siris [Blake, 1820]
6.4.1
The Generative Model as Self
Blake’s claim is radical and recursive: not merely that imagination constitutes human existence, but that
the relationship between agent and model is one of mutual entailment—“exists in us & we in him.” The
generative model does not belong to an agent; the generative model is the agent. Active Inference formalizes
this:
“Consciousness is nothing more than inference about my future; namely, the self-evidencing con-
sequences of what I could do.”
— [Friston, 2018]
The self as process, not entity:
“The self is the result of an ongoing predictive process within a generative model that is centered
onto the organism.”
— [Limanowski and Blankenburg, 2013]
And the body as probabilistically “most likely to be me”:
“One’s own body is the one which has the highest probability of being ‘me’ as other objects are
probabilistically less likely to evoke the same sensory inputs.”
— [Apps and Tsakiris, 2014]
Agent identity:
Self ≡𝑝(𝑜, 𝜃)
(17)
The generative model defines:
• What counts as inside/outside (blanket structure)
• What states are expected (prior beliefs)
• What observations mean (likelihood mapping)
• What actions are available (policy repertoire)
Without model, no agent. The self is the generative model (Equation 17), bounded by its Markov blanket
(Equation 5).
Seth’s “cybernetic Bayesian brain” makes this explicit: selfhood arises from the brain’s
predictive model of its own body, a “controlled hallucination” grounded in interoceptive and proprioceptive
inference [Seth, 2014]. Blake saw this two centuries earlier: “As a man is, so he sees.” The internal structure
determines external appearance.
Demonstration: Who Is Seeing?
Close your eyes. Imagine a red rose—its color, curve, scent.
Now ask: who is imagining? You might answer “I am.” But look closer. The generative model
that produces “rose” is the entity doing the imagining. There is no homunculus watching the
mental screen; the screen is the seer.
Blake: “Man is All Imagination God is Man & exists in us & we in him.”
The recursive “in us / we in him” captures the autopoietic loop: the model models itself modeling.
No agent exists beneath the model. The model is the agent.
Blake’s embodied vision:
31

## Page 32

“The Eternal Body of Man is The Imagination, that is, God himself, The Divine Body.”
— Annotation to Laocoon [Blake, 1988n]
“Mental Things are alone Real; what is call’d Corporeal Nobody knows of its Dwelling Place; it is
in Fallacy & its Existence an Imposture.”
— Vision of the Last Judgment [Blake, 1988l]
And the constitutive claim:
“As a man is, so he sees. As the Eye is formed, such are its Powers.”
— Letter to Dr. Trusler, 23 August 1799 [Blake, 1799]
The generative model shapes what can be perceived. The “formed Eye” is the structure of inference itself.
“The world of imagination is the world of eternity.”
— Vision of the Last Judgment [Blake, 1988l]
6.4.2
Man Is All Imagination: Autopoiesis and Self-Modeling
The recursive structure of Blake’s epigraph above—“exists in us & we in him”—is autopoiesis: the agent
IS its generative model.
“God in us” = our model of world.
“We in him” = we are part of what the
model represents. The nested embedding describes Markov blankets within Markov blankets—the recursive
self-modeling that constitutes agency.
There is no man “underneath” imagination; imagination is what man is. The generative model doesn’t
belong to an agent—the generative model IS the agent.
6.4.3
Mental Things Alone Real
Blake extends this to a full phenomenological idealism. His declaration that “Mental Things are alone Real”
and that the “Corporeal” is “in Fallacy & its Existence an Imposture” (Vision of the Last Judgment [Blake,
1988l]) maps directly onto the epistemic structure of Active Inference: internal states (beliefs) are all we
access. External states are inferred, never directly known. “Corporeal” = external states beyond the Markov
blanket. We have no direct access to the world-in-itself; only to our model’s predictions about it. What we
call “corporeal” is a posit of inference, not an immediate given.
This is not solipsism but epistemic humility: acknowledging that all our knowledge is model-mediated.
6.4.4
Knowledge by Perception, Not Deduction
Blake distinguishes parallel inference from serial reasoning:
“Knowledge is not by deduction but Immediate by Perception or Sense at once Christ addresses
himself to the Man not to his Reason”
— Annotations to Berkeley’s Siris [Blake, 1820]
Perception as parallel inference, not serial deduction. Posterior beliefs emerge from free energy minimization
directly—not through step-by-step logical chains but through the simultaneous settling of the entire gener-
ative model. “Christ addresses himself to the Man” = the world speaks to the whole agent, not just the
reasoning faculty.
This anticipates the Active Inference insight that perception is not a conclusion of reasoning but an immediate
update of the entire belief distribution.
32

## Page 33

6.4.5
Innate Ideas as Structural Priors
Against empiricist doctrine, Blake insists on innate structure:
“The Man who says that we have No Innate Ideas must be a Fool & Knave… Knowledge of Ideal
Beauty is Not to be Acquired It is Born with us”
— Annotations to Reynolds’ Discourses [Blake, 1988k]
Structural priors are architectural, not learned.
“Innate Ideas” = the priors that make inference possi-
ble. They cannot be “acquired” because they define the hypothesis space within which acquisition occurs.
Without prior structure, there is nothing to update—no model to receive evidence.
6.4.6
Imagination as Epistemic Foraging
If imagination is human existence—if the self is the generative model—then what Blake calls “creative
vision” amounts to a specific computational strategy: epistemic foraging through counterfactual model-space.
Rather than passively receiving data, the imaginative agent actively explores hypothetical configurations of
its own generative model, testing alternatives that minimize expected free energy over long horizons [Veissière
et al., 2020]. This is niche construction at the cognitive level: the imagination does not merely adapt to the
world as found but actively reshapes the model-space within which future inference occurs. Blake’s insistence
that “What is now proved was once only imagin’d” is, in Active Inference terms, the claim that today’s priors
were yesterday’s epistemic actions—imaginative explorations that became entrained belief structures. The
artist, the prophet, the visionary is thus not an escapist but an epistemic pioneer, foraging at the frontier of
model-space.
33

## Page 34

6.5
Time: Temporal Horizons
“I see the Past, Present & Future, existing all at once
Before me.”
— Jerusalem, Plate 15 [Blake, 1988i]
6.5.1
Temporal Horizons
“Eternity in an hour” is not mysticism—it is deep temporal modeling. In Active Inference, agents construct
hierarchical generative models that encode predictions at progressively longer time scales, from millisecond
sensory fluctuations to narratives spanning years. The deeper the hierarchy, the wider the temporal horizon
the agent can integrate into present awareness. Blake’s compression of eternity into an hour describes exactly
this capacity. Active Inference agents maintain predictions across multiple time scales:
“The lowest level of this hierarchy corresponds to fast fluctuations associated with sensory pro-
cessing, whereas the highest levels encode slow contextual changes in the environment.”
— [Kiebel et al., 2008]
“Slowly changing neuronal states can encode the paths or trajectories of faster sensory states.”
— [Friston et al., 2017]
Temporal hierarchy:
𝑝(𝑜1∶𝑇, 𝜃) =
𝑇
∏
𝑡=1
𝑝(𝑜𝑡|𝜃𝑡) ⋅𝑝(𝜃𝑡|𝜃𝑡−1)
(18)
The deeper the temporal model (Equation 18), the longer the horizon of prediction. An impoverished model
predicts only the immediate. A rich model extends to aeonic time. Figure 5 illustrates this hierarchy of
temporal scales.
Blake’s “Eternity” is not timelessness but trans-temporal integration—the capacity to hold all moments
within the present perception. The hour contains eternity because the model reaches across all scales.
In Milton, Blake provides temporal metrics:
“Every Time less than a pulsation of the artery Is equal in its period & value to Six Thousand
Years, For in this Period the Poet’s Work is Done, and all the Great Events of Time start forth
& are conceiv’d in such a Period, Within a Moment, a Pulsation of the Artery.”
— Milton, Plate 29 [Blake, 1988j]
A pulsation equals six thousand years—the prophetic compression of temporal scale. And the relationship
is not opposition but love:
“Eternity is in love with the productions of time.”
— Marriage of Heaven and Hell, Proverbs of Hell [Blake, 1988b]
The eternal reaches toward the temporal particular. Deep temporal models don’t transcend time but integrate
it—holding the trajectory of moments within present awareness.
6.5.2
The Temporal Architecture of Los
But Blake does not stop at compression. In the lines immediately preceding the “pulsation” climax, he
constructs the most extraordinary precursor to hierarchical temporal modeling in all of literature:
“But others of the Sons of Los build Moments & Minutes & Hours And Days & Months & Years
& Ages & Periods; wondrous buildings And every Moment has a Couch of gold for soft repose, (A
Moment equals a pulsation of the artery), And between every two Moments stands a Daughter of
34

## Page 35

Figure 5: Temporal Horizons of Inference. Four stacked trapezoid bands, widening from bottom to
top, represent increasing temporal depth in the hierarchical generative model (Equation 18). Fast (red,
milliseconds): sensory processing—the immediate registration of prediction errors. Mid (steel blue, seconds
to minutes): emotional integration—affective states that modulate precision weighting across brief episodes.
Slow (purple, hours to years): narrative construction—the autobiographical and cultural models that con-
textualize experience. Deep/Aeonic (gold, lifetimes to eternity): the deepest temporal priors encoding
civilizational and transpersonal patterns. The vertical arrow indicates increasing temporal depth. Blake’s
compression of temporal scales—“Every Time less than a pulsation of the artery / Is equal in its period
& value to Six Thousand Years” (Milton, Plate 29 [Blake, 1988j])—describes the hierarchical nesting where
each level encodes trajectories of the level below, and “Eternity is in love with the productions of time”
(Marriage of Heaven and Hell, Proverbs of Hell [Blake, 1988b]).
35

## Page 36

Beulah To feed the Sleepers on their Couches with maternal care. And every Minute has an azure
Tent with silken Veils. And every Hour has a bright golden Gate carved with skill. And every
Day & Night, has Walls of brass & Gates of adamant, Shining like precious stones & ornamented
with appropriate signs: And every Month, a silver paved Terrace builded high: And every Year,
invulnerable Barriers with high Towers. And every Age is Moated deep with Bridges of silver &
gold. And every Seven Ages is Incircled with a Flaming Fire.”
— Milton, Plate 28 [Blake, 1988j]
This is a hierarchical generative model of temporal experience, described two centuries before the formal
mathematics. Each timescale has distinct architectural properties—distinct structure, material, and boundary
conditions:
Temporal Scale
Blake’s Architecture
Active Inference Analogue
Moment (pulse)
Couch of gold
Fast sensory states: soft, immediate,
precious
Minute
Azure Tent, silken Veils
Short-term precision: flexible,
semi-transparent
Hour
Golden Gate, carved
Attentional boundaries: structured,
deliberate
Day/Night
Brass Walls, adamant Gates
Circadian priors: rigid, durable
Month
Silver Terrace, builded high
Seasonal/contextual priors: elevated
perspective
Year
Invulnerable Barriers, high Towers
Biographical priors: strongly defended
Age
Moated deep, Bridges of silver &
gold
Cultural priors: deep separation, costly
access
Seven Ages
Flaming Fire
Civilizational priors: ultimate boundary
The ascending material solidity—from gold couches to flaming fire—mirrors the increasing precision and
decreasing update rate of deeper temporal levels. A “Couch” yields to the body; a “Flaming Fire” does not.
The Sons of Los build these structures actively: temporal experience is not passively received but constructed
through inference. And the “Daughter of Beulah” standing between each two Moments, feeding “the Sleepers
on their Couches with maternal care,” is the precision-weighting mechanism that mediates between adjacent
temporal levels—ensuring smooth transitions and preventing catastrophic discontinuity.
6.5.3
Drawing Out to Seven Thousand Years
Blake describes the active construction of temporal depth through the figure of Eno, an ancient prophetess:
“
𝐸𝑛𝑜
drew it out to Seven thousand years with much care & affliction”
— Vala, or The Four Zoas, Night the First [Blake, 1988g, E300]
The model’s temporal horizon can be extended—“drawn out”—but not without cost.
Blake’s “care &
affliction” captures the computational expense of deep temporal models: maintaining predictions across long
horizons demands sustained precision allocation and imposes metabolic cost on the system. The model does
not naturally reach across millennia without deliberate cultivation of the hierarchical structure that makes
such depth possible.
Seven thousand years echoes the Biblical span of human history—Eno’s work is to hold the entire trajectory
of human time within present awareness.
36

## Page 37

6.5.4
Eternity Obliterated
The collapse of temporal depth is catastrophic:
“Till like a dream Eternity was obliterated & erasd”
— The Song of Los, Plate 3 [Blake, 1988f]
Deep temporal models (Eternity) compressed to shallow prediction (dream-like immediate present). When
the temporal horizon collapses, what remains is reactive, stimulus-bound perception—the “dream” state of
shallow inference.
This describes the Fall as temporal impoverishment: from aeonic awareness to moment-to-moment reaction.
6.5.5
Drunk with the Wine of Ages
But temporal integration has limits:
“drunk with the wine of ages”
— Vala, or The Four Zoas, Night the Ninth [Blake, 1988g]
Over-accumulation of historical evidence (“wine of ages”) impairs present inference. The temporal model
can become so saturated with past that it loses responsiveness to present. This is the opposite failure mode:
not temporal collapse but temporal rigidity—being so weighted by history that current observation cannot
update the model.
In Active Inference terms, this corresponds to an over-fitted temporal model whose deep priors have accumu-
lated excessive precision. When 𝜋prior at the slowest temporal scales grows unboundedly, the agent becomes
captive to historical regularities and cannot accommodate novel evidence. The system is “drunk”—its infer-
ence is dominated by the accumulated weight of temporal priors rather than responsive to present sensory
input. Blake thus identifies both failure modes of temporal modeling: too shallow (“Eternity obliterated”)
and too deep (“drunk with the wine of ages”).
Demonstration: The Déjà Vu Moment
You walk into an unfamiliar room and feel, with absolute conviction, that you have been here
before. The moment is uncanny precisely because two temporal scales are colliding: your shallow
present-moment model (novel room, first visit) contradicts a deeper temporal pattern-match
(this configuration of light, shape, and atmosphere resonates with a prior encoded at longer
timescales).
In Active Inference terms, déjà vu arises when a slow-timescale prior generates
a strong top-down prediction that the current observation is familiar, while the fast-timescale
model correctly registers novelty. The eerie sensation is the prediction error between hierarchical
temporal levels—Blake’s “Eternity in an hour” experienced as perceptual vertigo.
Temporal
depth is not abstract: it is the felt layering of past within present.
37

## Page 38

6.6
Space: Spatial Hierarchy
“To see a World in a Grain of Sand
And a Heaven in a Wild Flower…“
— Auguries of Innocence [Blake, 1988h]
6.6.1
Spatial Scale Invariance
The grain contains the world through hierarchical evidence accumulation. At each level of the generative
model, increasingly abstract regularities are extracted from observations: texture gives way to form, form to
spatial relations, spatial relations to universal principles. A sufficiently deep hierarchy finds the cosmos in
the particular because the abstract structure encoded at its highest levels applies universally. The predictive
brain finds universal structure in particular observations:
“By formulating Helmholtz’s original ideas on perception in terms of modern-day statistical the-
ories, one arrives at a model of perceptual inference and learning that can explain a remarkable
range of neurobiological facts.”
— [Friston, 2005]
Model complexity:
𝐹simple ≫𝐹rich
(19)
A shallow model incurs vastly more free energy (Equation 19) than a rich one. The shallow model sees only
surface: the grain is sand. A deep model extracts universal structure: the grain is cosmos in miniature.
Model Depth
Perception
Blake’s Image
Shallow
Isolated particulars
“narrow chinks”
Intermediate
Contextual patterns
“twofold vision”
Deep
Universal in particular
“World in a Grain”
The “Wild Flower” opens to “Heaven” because a sufficiently deep generative model finds infinite structure
in finite observation. This is not enhancement—it is accuracy.
Blake’s spatial cosmology extends the particularity principle:
“And every Space smaller than a Globule of Man’s blood opens Into Eternity of which this vegetable
Earth is but a shadow.”
— Milton, Plate 29 [Blake, 1988j]
The vortex phenomenon—perception’s expansive/contractive dynamics:
“The nature of infinity is this: That every thing has its Own Vortex, and when once a traveller
thro’ Eternity Has passd that Vortex, he perceives it roll backward behind His path, into a globe
itself infolding like a sun.”
— Milton, Plate 15 [Blake, 1988j]
And the expansion principle:
“If the Spectator could Enter into these Images in his Imagination, approaching them on the Fiery
Chariot of his Contemplative Thought… then would he arise from his Grave, then would he meet
the Lord in the Air & then he would be happy.”
— Vision of the Last Judgment [Blake, 1988l]
Spatial perception is active—the spectator enters images, doesn’t passively receive them. The generative
model projects into the world as much as it receives from it.
38

## Page 39

6.6.2
The Vortex: Each Entity’s Own Markov Blanket
Blake’s fullest treatment of perspectival space appears in Milton:
“The nature of infinity is this: That every thing has its / Own Vortex; and when once a traveller
thro Eternity / Has passd that Vortex, he percieves it roll backward behind / His path, into a
globe itself infolding like a sun”
— Milton, Plate 15, lines 21-35 [Blake, 1988j]
Each entity has its own Markov blanket and generative model—its “Own Vortex.” “Passing the Vortex” =
adopting a new perspective, entering another entity’s coordinate frame. “Globe itself infolding” = the old
model becoming an object within the new model’s representation.
The vortex is the boundary between reference frames. When you pass through another entity’s vortex, you
see from their perspective—but looking back, your old perspective has become a distant object, a “globe” or
“sun” behind you.
In Active Inference, this corresponds to model switching—the capacity to adopt another agent’s generative
model as one’s own frame of reference. Friston and Frith’s “Duet for One” [Friston and Frith, 2015] formalizes
precisely this process: two agents can achieve generalized synchronization by inferring each other’s hidden
states, effectively “passing through” each other’s Markov blankets. Blake’s vortex adds a phenomenological
dimension that the formalism captures only structurally: the felt experience of entering another perspective
and finding that one’s prior viewpoint has become a distant object, reduced to a “globe” in the rearview of
awareness.
This is the phenomenology of empathy formalized: to understand another is to pass through their vortex, to
see from within their generative model.
6.6.3
Mathematic Holiness: The Impoverished Spatial Model
Blake names the pathological reduction of spatial perception with devastating precision.
In Milton, he
describes those whose models have been stripped of imaginative depth:
“those combind by Satans Tyranny… are Shapeless Rocks Retaining only Satans Mathematic Ho-
liness, Length: Bredth & Highth”
— Milton, Plate 32 [Blake, 1988j]
“Mathematic Holiness” is the worship of pure geometric extension—Euclidean coordinates elevated to meta-
physical status.
When imaginative depth is removed from spatial perception, what remains is the thin,
quantitative skeleton: length, breadth, height—the minimum parameters of a bounding box. The “Shape-
less Rocks” are agents whose generative models have collapsed to this minimum complexity: they can localize
objects in three-dimensional space but cannot perceive the hierarchical structure that makes space meaning-
ful—the nested affordances, the perspectival depth, the relational richness that a deep model extracts from
the same optical array.
In Active Inference terms, this is the spatial analogue of Newton’s Sleep (see §4.3 States). A model that
retains “only” length, breadth, and height is a model whose spatial hierarchy has been flattened to a single
level: the likelihood function 𝑝(𝑜|𝑠) computes geometric coordinates, but the prior structure 𝑝(𝑠) that would
encode ecological meaning, bodily affordance, and perspectival significance has been suppressed. The result
is the “vegetable” space of Newtonian mechanics—measurable, uniform, and dead. Blake’s “World in a Grain
of Sand” is the antithesis: spatial perception at full hierarchical depth, where the generative model finds
universal structure in local observation because its prior hierarchy is rich enough to make the extraction
possible.
6.6.4
Looking Through, Not With
Blake distinguishes the sensor from the inference process:
39

## Page 40

“I question not my Corporeal or Vegetative Eye any more than I would Question a Window
concerning a Sight I look thro it & not with it”
— Vision of the Last Judgment [Blake, 1988l, E566]
Blake distinguishes the sensor from the inference engine with surgical precision. The eye is the likelihood
function 𝑝(𝑜|𝑠)—the mapping from hidden states to observations—not the inference process itself. To look
“through” the eye is to use the full generative model: the inverse mapping from observations to beliefs about
hidden causes. To look “with” the eye is to confuse the sensor with inference, mistaking the data channel
for the interpretation.
The eye provides observations; it does not provide understanding. Understanding requires the generative
model that interprets what the eye delivers. To look “with” the eye is to mistake the window for the view—
an error that reduces perception to passive reception and forecloses the active, model-driven inference that
constitutes genuine seeing.
This is a statement about the distinction between the likelihood function and the full Bayesian inversion. The
eye computes 𝑝(𝑜|𝑠); seeing requires the posterior 𝑝(𝑠|𝑜) ∝𝑝(𝑜|𝑠)⋅𝑝(𝑠). Confusing the two is the fundamental
error of empiricism—Blake’s “single vision.”
Demonstration: The Microscope Effect
Place a leaf under a microscope. At first you see nothing but undifferentiated green blur—single
vision. Adjust the focus: suddenly, cells appear—a world within the leaf, invisible to the unaided
eye. Increase magnification: within each cell, organelles, a further world. Blake’s “world in a
grain of sand” is not metaphor but phenomenological report: the grain reveals a world precisely
when the generative model gains hierarchical depth, when new layers of the likelihood function
𝑝(𝑜|𝑠) map observations to increasingly fine-grained hidden causes. The microscope changes the
observation channel, but it is the model—the skilled biologist’s expectations about what cell
types, structures, and processes should be visible—that transforms undifferentiated green into
meaningful architecture. The instrument is the window; the model is the sight.
40

## Page 41

6.7
Action: Free Energy Minimization
“I must Create a System. or be enslav’d by another Mans. I will not Reason & Compare: my
business is to Create.”
— Jerusalem, Plate 10 [Blake, 1988i]
6.7.1
Free Energy Minimization
Los’s imperative to create rather than merely reason captures the essence of active inference: the agent does
not passively receive the world but actively constructs its model. The doors are cleansed when the generative
model accurately mirrors reality. As the foundational paper states:
“The free-energy principle provides a unified account of action, perception and learning based on
minimizing a measure of surprise.”
— [Friston, 2010]
“Cortical responses can be seen as the brain’s attempt to minimize the free energy induced by a
stimulus and thereby encode the most likely cause of that stimulus.”
— [Friston et al., 2006]
Cleansing formalized:
Cleansed perception ⟺arg min
𝑞
𝐹(𝑞, 𝑜)
(20)
Free energy reaches minimum when:
• Model predictions match observations
• Prior beliefs calibrate to evidence
• Precision-weighting is optimal
Two paths to cleansing:
1. Perception — Update beliefs to fit world (𝑞(𝜃) →𝑝(𝜃|𝑜))
2. Action — Change world to fit model (𝑜→
̂𝑜)
Both are “cleansing.” Both reduce free energy (Equation 20), driving the variational bound (Equation 1)
toward its minimum (Figure 6). Blake unified what later theory separated.
The generative model anticipates sensory confirmation—imagination precedes proof, the prophet predicts
what observation will confirm. What Blake imagined, Friston’s equations now prove.
Blake’s energy philosophy aligns with active inference’s emphasis on action:
“Energy is Eternal Delight.”
— Marriage of Heaven and Hell, Plate 4 [Blake, 1988b]
The labor principle extends this creative imperative to situated practice:
“Great things are done when Men & Mountains meet; This is not done by Jostling in the Street.”
— Notebook (c. 1807-1809) [Blake, 1988o, E516]
“Labour well the Minute Particulars, attend to the Little-ones.”
— Jerusalem, Plate 55, line 51 [Blake, 1988i]
Action on the minute particular—the precise, situated intervention—is how free energy is minimized in
practice. Not grand abstraction but attentive labor.
41

## Page 42

6.7.2
Flexible Senses: Precision Dynamics Before the Fall
Blake’s most direct statement of voluntary precision modulation appears in The Book of Urizen:
“The will of the Immortal expanded / Or contracted his all flexible senses”
— The Book of Urizen, Chapter II, Plate 3 [Blake, 1988d]
Before the Fall, precision (𝜋))𝑐𝑜𝑢𝑙𝑑𝑏𝑒𝑎𝑑𝑗𝑢𝑠𝑡𝑒𝑑𝑎𝑡𝑤𝑖𝑙𝑙−−−‶𝑒𝑥𝑝𝑎𝑛𝑑𝑒𝑑𝑜𝑟𝑐𝑜𝑛𝑡𝑟𝑎𝑐𝑡𝑒𝑑.″𝑇ℎ𝑖𝑠𝑖𝑠𝐵𝑙𝑎𝑘𝑒′𝑠𝑐𝑙𝑒𝑎𝑟𝑒𝑠𝑡𝑠𝑡𝑎𝑡𝑒𝑚𝑒𝑛𝑡𝑜𝑓𝑝𝑟𝑒𝑐𝑖𝑠𝑖
𝑡ℎ𝑒𝑢𝑛𝑓𝑎𝑙𝑙𝑒𝑛𝑐𝑜𝑛𝑑𝑖𝑡𝑖𝑜𝑛𝑖𝑠𝑜𝑛𝑒𝑤ℎ𝑒𝑟𝑒𝑎𝑡𝑡𝑒𝑛𝑡𝑖𝑜𝑛𝑐𝑎𝑛𝑏𝑒𝑓𝑟𝑒𝑒𝑙𝑦𝑎𝑙𝑙𝑜𝑐𝑎𝑡𝑒𝑑.
The “all flexible senses” are not fixed channels but adjustable receptors. Flexibility is the default; rigidity is
fallen.
6.7.3
Contracting to the Honey Bee
Blake elaborates the scale of precision adjustment:
“Contracting or expanding their all flexible senses / At will to murmur in the flowers small as
the honey bee”
— Vala, or The Four Zoas, Night the First [Blake, 1988g]
Senses can zoom to “honey bee” scale—precision determines resolution. The unfallen beings can adjust their
sensory precision to perceive at any scale, from cosmic to microscopic. This is optimal precision assignment:
the ability to weight sensory channels appropriately for the task at hand.
6.7.4
Love as Affective Precision
Precision weighting operates through affect:
“He who Loves feels love descend into him & if he has wisdom may percieve it is from the Poetic
Genius which is the Lord”
— Annotations to Swedenborg’s Divine Love and Divine Wisdom [Blake, 1788]
“Love” = affective precision. “Descend” = top-down modulation. “Wisdom” = meta-cognitive precision
awareness—the ability to perceive the source of one’s attention.
Affect is not separate from inference but constitutive of it. Love determines what matters, what receives
high precision weighting.
6.7.5
Thought Alone Makes Monsters
But inference without affective grounding drifts pathologically:
“Thought alone can make monsters, but the affections cannot”
— Annotations to Swedenborg [Blake, 1788]
Inference without affective grounding produces biologically non-viable beliefs. “Monsters” = pathological
states that no embodied system could inhabit. Affect anchors inference to survival, to what matters for the
organism’s persistence.
Pure reasoning, unmoored from bodily concern, can generate coherent but monstrous conclusions.
The
affections keep inference honest. Damasio’s somatic marker hypothesis formalizes Blake’s insight with striking
precision: bodily affect tags candidate beliefs and actions with valence before deliberative reasoning can
evaluate them [Damasio, 1994]. The “affections” are somatic markers—precision signals weighted by visceral
relevance. Without these embodied priors, the reasoning system falls into what Damasio terms the “high-
reason” catastrophe: logically valid but existentially disastrous decisions, monsters of pure thought.
42

## Page 43

Demonstration: The Musician’s Practice
A pianist learning a new piece begins haltingly: each note requires conscious prediction, deliberate
action, effortful error correction. The prediction errors are large and frequent—every misplayed
note generates surprise. Through practice, the generative model refines: the fingers begin to
anticipate the score, action policy and sensory prediction converge, surprise decreases.
But
something else happens: the music begins to feel right. Affect—the bodily sense of rightness—
becomes the precision signal that guides further refinement. The musician doesn’t just minimize
error; she shapes her model until it produces the felt quality of beautiful performance. This is
Blake’s “Mental Fight” rendered as craft: active inference through embodied skill, where action
transforms both world (the sound produced) and model (the musician’s internal representation).
The pianist “creates a system” note by note, bar by bar—and in doing so, is herself transformed.
6.7.6
Cogs Tyrannic vs. Free Wheels
Blake contrasts deterministic with flexible inference:
“Wheel without wheel, with cogs tyrannic / Moving by compulsion each other: not as those in
Eden, which / Wheel within Wheel in freedom revolve in harmony & peace”
— Jerusalem, Plate 15 [Blake, 1988i]
“Cogs tyrannic” = fixed precision, mechanical prediction. No free parameters, no flexibility—each wheel
forces the next. “Wheel within Wheel in freedom” = adjustable precision weighting, where components
coordinate but are not rigidly locked.
The Edenic state is not absence of structure but flexible structure—wheels that revolve together in harmony
without tyrannical compulsion. This is the difference between a generative model that can revise itself and
one frozen in Newton’s sleep.
6.7.7
The Path of Least Action
Friston’s path integral formulation of the free energy principle [Friston et al., 2023] reveals a deeper connec-
tion to Blake’s economy of perception. The path integral expresses system trajectories as weighted sums over
all possible paths, with the most probable trajectory being the one that minimizes action—the “path of least
action.” Blake’s imperative to “cleanse the doors” is, in this formalism, a call to find the self-evidencing tra-
jectory: the path through model-space that minimizes free energy while maintaining the system’s structural
integrity. The “cogs tyrannic” are paths constrained to a single rigid trajectory; the “wheels in freedom” are
paths that explore the full space of possibilities while converging on the variational minimum. Art, for Blake,
is precisely this search: the creative act finds the path of least free energy through the space of possible forms,
arriving at the work that resolves the most prediction error with the most elegant structure.
43

## Page 44

Figure 6: The Perception-Action Cycle. Active Inference’s dual optimization pathways annotated with
corresponding Blake quotations. The central generative model 𝑝(𝑜, 𝜃) (orange hub) issues top-down predic-
tions and receives bottom-up prediction errors 𝜀= 𝑜−𝑔(𝜃) through six interconnected stages: Prediction
(“What is now proved was once imagined”), Sensory Input (“The senses discover’d the infinite”), Predic-
tion Error (“Narrow chinks of his cavern”), Model Update (“Cleansed perception”), Action Selection
(“Mental Fight”), and World Change (“Building Jerusalem”). Perceptual inference updates beliefs 𝑞(𝜃)
toward the true posterior (changing mind to fit world); active inference samples observations matching pre-
dictions (changing world to fit mind). Both pathways reduce variational free energy 𝐹(Equation 20). The
cycle’s continuous operation corresponds to Blake’s vision of perpetual creative labor: “I must Create a
System. or be enslav’d by another Mans” (Jerusalem, Plate 10 [Blake, 1988i]).
44

## Page 45

6.8
Collectives: Shared Generative Models
“I will not cease from Mental Fight,
Nor shall my Sword sleep in my hand:
Till we have built Jerusalem,
In England’s green & pleasant Land.”
— Milton, Preface [Blake, 1988j]
6.8.1
Shared Generative Models
Blake’s vision extends beyond individual perception to collective awakening. Jerusalem is not merely personal
enlightenment but a shared visionary capacity—a coordinated mode of seeing that transcends the individual.
Active Inference extends naturally to multi-agent systems:
“Generalized synchronization
𝑒𝑚𝑒𝑟𝑔𝑒𝑠
as a mathematical image of communication that enables two Bayesian brains to entrain each
other and, effectively, share the same dynamical narrative.”
— [Friston and Frith, 2015]
“Human agents learn the shared habits, norms, and expectations of their culture through im-
mersive participation in patterned cultural practices that selectively pattern attention and be-
haviour.”
— [Veissière et al., 2020]
This is TTOM—Thinking Through Other Minds—the mechanism of collective awakening.
Multi-agent coordination:
𝑝(𝑜, 𝜃) =
𝑁
∏
𝑖=1
𝑝(𝑜𝑖|𝜃𝑖) ⋅𝑝(𝜃𝑖|𝜃shared) ⋅𝑝(𝜃shared)
(21)
Multiple agents share a common prior 𝜃shared (Equation 21)—the cultural generative model that enables
coordinated perception and action. Figure 7 illustrates this multi-agent architecture.
Each individual agent remains bounded by its own Markov blanket (Equation 5), but the shared prior aligns
their inference.
Blake’s most radical claim about the collective nature of identity appears in Milton:
“We are not Individuals but States: Combinations of Individuals”
— Milton, Plate 32 [Blake, 1988j]
This is not merely the claim that agents share priors—it is the deeper assertion that individual identity itself
is a collective phenomenon. A “Combination of Individuals” is a factorization of agency: what appears to
be a single self is in fact a composition of shared model components, cultural priors, and socially entrained
precision weightings. The shared prior 𝜃shared in Equation 21 is not external to individuals but constitutive
of them—the individual 𝜃𝑖cannot be separated from the collective without remainder. Blake’s “States” are
thus collective attractors in the multi-agent generative model: configurations that groups of agents fall into
together, and that “Change” (dissolve, reconfigure) when the collective model is revised. “Satan & Adam
are States Created into Twenty-seven Churches”—not individuals who happened to organize churches, but
model configurations that manifest as institutional structures.
45

## Page 46

Figure 7: Building Jerusalem: Collective Generative Models. Three individual agents, each bounded by
its own Markov blanket, contribute to and draw from a shared generative model (“Jerusalem,” 𝜃shared).
The “Mental Fight” zone represents the active process of model-building and coordination through which
individual inference aligns with collective priors (Equation 21).
46

## Page 47

Component
Symbol
Blake’s Image
Individual models
𝜃𝑖
Each perceiver
Shared prior
𝜃shared
Jerusalem
Collective action
𝑎collective
“Mental Fight”
Coordinated perception
𝑜aligned
Shared vision
6.8.2
The Mental Fight
Blake’s “Mental Fight” is model-building at civilizational scale:
• Education shapes the generative models of the young
• Art restructures perception through alternative priors
• Contemplative practice adjusts precision weighting
• Cultural production creates shared predictive structures
“The Nature of my Work is Visionary or Imaginative; it is an Endeavour to Restore what the
Ancients called the Golden Age.”
— Vision of the Last Judgment [Blake, 1988l]
The Golden Age is not historical but perceptual—a state where collective generative models enable richer
inference.
6.8.3
Coordinated Inference
Active Inference extends naturally to multi-agent systems [Friston, 2019]. Ramstead, Badcock, and Fris-
ton formalize this extension through nested Markov blankets: blankets within blankets, individuals within
communities within cultures, each scale operating as an autonomous inference system while coupling to the
scales above and below [Ramstead et al., 2018]. When agents share priors, they:
1. Align predictions — Expectations converge across the collective
2. Coordinate action — Behavior becomes mutually intelligible
3. Distribute computation — Complex inference divides across agents
4. Accumulate evidence — Collective learning exceeds individual capacity
Blake’s Jerusalem is precisely this: a shared generative model enabling coordinated cleansing of perception.
The doors open not one by one, but together.
“Mutual Forgiveness of each Vice—Such are the Gates of Paradise.”
— For the Sexes: The Gates of Paradise [Blake, 1988o]
Paradise requires mutual forgiveness—collective precision adjustment where agents release rigid priors toward
one another. The gates open through shared model revision.
Blake’s collective vision finds its fullest expression in Jerusalem:
“This is Jerusalem in every Man A Tent & Tabernacle of Mutual Forgiveness.”
— Jerusalem, Plate 54 [Blake, 1988i]
“O lovely Emanation of Albion Awake and overspread all Nations as in Ancient Time For lo! the
Night of Death is past and the Eternal Day Appears upon our Hills.”
— Jerusalem, Plate 97 [Blake, 1988i]
The awakening is collective—Albion (England/humanity) awakens as a unified agent.
47

## Page 48

6.8.4
From Single to Collective Vision
The fourfold hierarchy applies not only to individuals but to societies:
Level
Individual
Collective
Single
Mechanical perception
Industrial society
Twofold
Emotional engagement
Artistic communities
Threefold
Imaginative vision
Cultural movements
Fourfold
Integrated awareness
Jerusalem
Blake’s critique of “dark Satanic Mills” is computational: industrial modernity imposes shallow, prior-
dominated generative models on the collective.
Building Jerusalem means reconstructing shared priors
to enable deeper inference.
“England! awake! awake! awake!
Jerusalem thy Sister calls!“
— Jerusalem, Plate 77 [Blake, 1988i]
The awakening is collective. The sister calls to the nation. The doors of perception—once cleansed—reveal
not isolated infinity but shared infinity. The prophet’s vision becomes the people’s sight.
6.8.5
Fall into Division and Resurrection to Unity
Blake frames the cosmic narrative as multi-agent decomposition and re-coordination:
“Sing His fall into Division & his Resurrection to Unity”
— Vala, or The Four Zoas, Night the First [Blake, 1988g]
“Division” = factorization into separate agents, each with its own Markov blanket and generative model.
“Unity” = re-coordination into a shared generative model (Jerusalem). The Fall is not moral failure but
factorization—the breaking apart of an integrated system into competing subsystems.
Resurrection is the inverse: the re-establishment of shared priors that enable coordinated inference across
agents.
6.8.6
The Eternal Man Is Risen
The achievement of collective coordination:
“Rise from the dews of death for the Eternal Man is Risen”
— Vala, or The Four Zoas, Night the Ninth [Blake, 1988g]
“Eternal Man” (Albion) = the multi-agent system as unified entity. When the Four Zoas are re-integrated,
Albion rises—not as the sum of parts but as the emergent coordination that parts enable.
6.8.7
Human Harvest
Collective free energy minimization under stress:
“In pain the human harvest wavd in horrible groans of woe”
— Vala, or The Four Zoas [Blake, 1988g]
“Harvest” = coordinated action across agents. “Pain” = high free energy state. The collective strives toward
lower free energy, but the process is not painless—coordination requires the adjustment of individual models
to shared constraints.
48

## Page 49

6.8.8
The Four Zoas: Factorized Collective Mind
Blake’s most systematic account of multi-agent cognition appears in his unfinished epic Vala, or The Four
Zoas. The Four Zoas—Urizen, Urthona (fallen as Los), Luvah (fallen as Orc), and Tharmas—represent not
merely allegorical faculties but a factorized generative model where independent components must coordinate
to achieve unified inference.
“Four Mighty Ones are in every Man; a Perfect Unity / Cannot Exist but from the Universal
Brotherhood of Eden”
— Vala, or The Four Zoas, Night the First [Blake, 1988g]
In Active Inference terms, factorization enables tractable computation:
Factorized model:
𝑝(𝑜, 𝜃) = 𝑝(𝑜|𝜃𝑈, 𝜃𝐿, 𝜃𝐿𝑣, 𝜃𝑇) ⋅𝑝(𝜃𝑈) ⋅𝑝(𝜃𝐿) ⋅𝑝(𝜃𝐿𝑣) ⋅𝑝(𝜃𝑇)
(22)
where subscripts denote the four Zoas’ contributions to the joint model. This extends the general hierar-
chical factorization (Equation 6) into a multi-component architecture. Figure 8 illustrates the compass-rose
arrangement of the four Zoas.
6.8.8.1
The Four Components
Zoa
Direction
Domain
AIF Function
Failure Mode
Urizen
South
Reason, Law
Likelihood 𝑝(𝑜‖𝜃)
Prior dominance
(Newton’s sleep)
Urthona/Los
North
Imagination,
Prophecy
Prior structure 𝑝(𝜃)
Model collapse (despair)
Luvah/OrcEast
Passion,
Emotion
Precision 𝜋)
Affective flooding (chaos)
Tharmas
West
Body, Instinct
Interoception
Dissociation (abstraction)
6.8.8.2
Urizen: The Likelihood Function
“And his Soul sicken’d! he curs’d / Both sons & daughters; for he saw / That no flesh nor spirit
could keep / His iron laws one moment.”
— The Book of Urizen, Plate 23 [Blake, 1988d]
Urizen represents the rational processing of evidence—the likelihood function that evaluates how well ob-
servations fit hypotheses. His “iron laws” are the regularities that structure prediction. But when Urizen
dominates, the system becomes rigid: prior-locked, unable to revise.
Urizen’s failure is over-precision of priors: 𝜋prior →∞. The laws become tyrannical because they cannot
update.
6.8.8.3
Urthona/Los: Prior Structure
“Los built the Walls of Golgonooza against the stirring battle”
— Jerusalem, Plate 12 [Blake, 1988i]
Urthona (unfallen) / Los (fallen) represents the creative imagination—the prior structure that makes infer-
ence possible. Los “builds” Golgonooza, the city of art, which IS the generative model’s architecture.
Without Urthona/Los, there is no hypothesis space. The prior structure defines what can be believed, what
states are even conceivable. Los’s labor at the furnaces is the continuous work of model construction.
49

## Page 50

Figure 8: The Four Zoas: A Factorized Model of Mind. The four Zoas occupy cardinal positions—Urizen
(South, reason/likelihood), Urthona/Los (North, imagination/prior), Luvah/Orc (East, passion/precision),
and Tharmas (West, body/interoception)—around a central hub of unified inference. Coordination arcs
connect adjacent Zoas; failure modes (prior dominance, model collapse, affective flooding, dissociation)
appear when any single component tyrannizes. “Perfect Unity” requires the “Universal Brotherhood” of all
four modes (Equation 22).
50

## Page 51

Figure 9:
William Blake, Milton a Poem, Plate 32 (c. 1804–1811).
The four Zoas in their cosmic
arrangement—the mythological source for the factorized model above. Blake depicts the fourfold division
and interdependence of the faculties through characteristic visual symbolism. Relief etching with hand col-
oring. Courtesy of the William Blake Archive [Blake, 1988j].
Los’s failure is model collapse: when imagination fails, the prior structure dissolves, leaving no framework
for inference. This is despair—the inability to conceive alternatives.
6.8.8.4
Luvah/Orc: Precision Weighting
“Luvah is France, the Victim of the Spectres of Albion”
— Jerusalem, Plate 66 [Blake, 1988i]
Luvah (unfallen) / Orc (fallen) represents passion, emotion, desire—the precision weighting that determines
what matters, what receives attention. Luvah controls the “chariots of the morning”—the affective salience
that drives engagement.
Precision weighting is not merely attention but care: what the system treats as important, what prediction
errors warrant response.
Luvah’s failure is affective flooding: when emotion dominates, precision weights become extreme, the system
oscillates chaotically, unable to maintain stable inference. Orc’s revolutionary fire burns without discrimi-
nation.
6.8.8.5
Tharmas: Interoceptive Inference
Blake’s symbolic system assigns each Zoa to a distinct
domain: Tharmas governs the vegetal (bodily/sensory) world, Luvah the world of sensations and emotion,
Urizen the world of reason, and Urthona the world of imagination. These correspondences pervade The Four
Zoas though Blake expresses them through narrative action rather than explicit formula.
Tharmas represents embodiment—the “vegetal” instinctual life, interoceptive inference about the body’s
state. Blake dramatizes Tharmas’s devastation when separated from the other Zoas:
51

## Page 52

“Tharmas groand among his Clouds / Weeping, then silent thundering he burst the bounds of
Destiny / And shook the heavens with wrath”
— Vala, or The Four Zoas, Night the Third [Blake, 1988g]
Tharmas is often described as the most damaged of the Zoas, reduced to helpless weeping in the sea of time
and space—the body in distress when severed from imagination, reason, and affect.
Seth’s interoceptive inference framework formalizes this Blakean insight: the self is constituted not merely
by exteroceptive prediction but by the body’s ongoing inference about its own visceral states [Seth, 2013].
Tharmas IS interoceptive inference—the felt sense of aliveness that grounds all other cognitive modes in
biological reality. Without Tharmas, the model floats free of embodiment—pure abstraction without survival
relevance.
Tharmas’s failure is dissociation: when embodiment is severed, inference loses its anchor in biological viability.
The system can reason but cannot feel, can think but cannot care.
6.8.8.6
Coordination and Pathology
The Four Zoas must coordinate for healthy inference:
• Urizen + Los: Reason and imagination must balance—priors that can update, regularities that can
revise
• Luvah + Tharmas: Emotion and embodiment must align—what matters must connect to survival
• All Four: The complete agent requires all four modes operating in “Universal Brotherhood”
Pathology arises from imbalance:
Dominant Zoa
Condition
Clinical Parallel
Urizen alone
Rigid rationalism
Obsessive-compulsive patterns
Luvah alone
Affective chaos
Borderline dysregulation
Los alone
Dissociated fantasy
Schizotypal withdrawal
Tharmas alone
Instinctual flooding
Panic, somatic fixation
6.8.8.7
The Unfallen State
“they gave to it a Space & namd the Space Ulro”
— Vala, or The Four Zoas, Night the First [Blake, 1988g, E303]
Before the Fall, the Zoas did not have separate spaces—they coordinated seamlessly within a unified model.
The Fall is precisely the factorization into competing subsystems, each claiming territory.
Redemption is re-coordination:
not the dominance of one Zoa but the restoration of “Universal
Brotherhood”—a shared generative model where each component contributes its proper inference without
tyrannizing the others.
6.8.8.8
Implications for Active Inference
Blake’s Four Zoas anticipate the insight that complex
inference requires factorization, but factorization introduces coordination problems. The multi-agent mind
must:
1. Maintain distinct components — Each Zoa has its proper function
2. Coordinate across components — Shared priors enable unified behavior
3. Prevent dominance — No single factor should monopolize inference
4. Ground in embodiment — Tharmas anchors the system in biological reality
The Zoas are not personality types but inference modes—different aspects of the generative model that must
harmonize for cleansed perception.
52

## Page 53

6.8.8.9
The Mean-Field Approximation
Blake’s factorization prefigures the mean-field approxi-
mation in variational inference. When exact inference is intractable, we approximate the true posterior by
assuming independence between factors:
Mean-field factorization:
𝑞(𝜃) ≈𝑞(𝜃𝑈) ⋅𝑞(𝜃𝐿) ⋅𝑞(𝜃𝐿𝑣) ⋅𝑞(𝜃𝑇)
(23)
This factorization enables tractable computation but introduces coordination costs—precisely Blake’s
diagnosis that the Zoas’ separation produces suffering. The “torments” arise because mean-field assumes
independence where correlation should exist.
Full variational inference would preserve correlations; the
factorized approximation trades accuracy for tractability.
Blake’s vision of “Eternal Brotherhood” corresponds to structured variational families that preserve key
correlations while remaining tractable.
The goal is not to eliminate factorization but to coordinate the
factors—each Zoa maintaining its function while communicating with the others.
“And they conversed together in Visionary forms dramatic which bright / Redounded from their
Tongues in thunderous majesty, in Visions / In new Expanses”
— Jerusalem, Plate 98 [Blake, 1988i]
When the Zoas converse—when the model’s components communicate—new expanses open. This is the
fourfold vision realized: not single-track inference but multi-modal coordination, each perspective enriching
the others.
Demonstration: The Stadium Wave
Sixty thousand spectators rise and sit in sequence, producing a traveling wave that circles the
stadium in seconds. No one coordinates the wave; no conductor signals the timing. Each indi-
vidual infers from their neighbors’ actions when to rise—a local prediction, locally tested, locally
corrected. Yet the global pattern emerges: a coherent wave far larger than any individual’s per-
ceptual horizon. This is Blake’s “Jerusalem” in miniature: a collective structure that transcends
individual agency while depending entirely upon it. The shared generative model is not held in
any single mind but distributed across the blanket boundaries of thousands of coupled agents,
each minimizing their own surprise by predicting their neighbors and acting accordingly. When
the wave coheres, it feels like something beyond the individuals—an emergent collective agency
that Blake would recognize as the Eternal Man arising.
6.8.9
Jerusalem as Cultural Niche
Veissière and colleagues’ framework of “Thinking Through Other Minds” [Veissière et al., 2020] illuminates
a final dimension of Blake’s Jerusalem: the city is not merely a shared generative model but a cultural
niche—an environment of shared priors, affordances, and epistemic resources constructed and maintained
through collective inference. Culture, in this view, is not a static repository of information passed down
through generations but a living system of shared expectations, jointly calibrated through what Active
Inference calls epistemic foraging and what Blake calls “Mental Fight.” The laborers of Golgonooza are not
merely building a model; they are constructing the conditions under which future inference can occur—the
affordance landscape that will shape subsequent generations’ priors. Jerusalem, once built, becomes the
niche within which new minds are enculturated, new Zoas coordinated, new visions made possible. The
social construction of reality is, in the deepest sense, the collaborative construction of a shared generative
model—and Blake’s prophetic vision of this process anticipates by two centuries the formal framework that
now makes it computationally precise.
53

## Page 54

7
Implications: The Wider Fields
The doors open onto wider fields.
7.1
Philosophy of Mind
7.1.1
The Romantic Computational Mind
The foregoing synthesis raises fundamental questions for philosophy of mind, cognitive science, and how we
understand the relationship between artistic and scientific knowledge. If Blake’s phenomenological observa-
tions and Friston’s mathematical framework describe the same cognitive architecture, then the Romantic
tradition is not anti-cognitive but proto-computational—articulating in the language of vision what neuro-
science would later formalize.
Blake’s “As a man is, so he sees” is not merely prescient metaphor; it is a precise statement of the Active Infer-
ence thesis that perception is model-dependent inference. The convergence suggests that phenomenological
observation and formal modeling approach the same cognitive reality from different directions, each illumi-
nating what the other cannot express. Against McGinn’s “mysterianism”—the claim that consciousness con-
stitutes an irresolvable problem for human cognition [McGinn, 2004]—the Blake–Active Inference synthesis
suggests a third way: neither reductive explanation nor mysterian agnosticism, but phenomenological-formal
complementarity, where visionary description and mathematical formalism illuminate different facets of the
same cognitive architecture.
7.1.2
Consciousness as Hierarchical Depth
Blake’s fourfold hierarchy (Figure 3) implies degrees of awareness. Single vision is diminished consciousness.
Fourfold vision is full integration.
If cleansed perception = optimized free energy minimization (Equation 20), then consciousness correlates
with well-calibrated generative models. The dimness of Newton’s sleep is computational: poor precision
weighting (Equation 16) produces impoverished inference, collapsing the prediction error signal (Equation
11).
Conversely, expanded consciousness corresponds to deeper hierarchical models that compress more
temporal structure (Equation 18) and richer spatial detail (Equation 19) into unified awareness.
7.2
Cognitive Science
7.2.1
Predictions
Three empirical implications follow from the Blake–Active Inference correspondence:
1. Expert perception — If fourfold vision reflects hierarchical depth, then artists and naturalists should
exhibit richer hierarchical representations than novices. Studies of perceptual expertise in visual art
[Chamberlain et al., 2013] already document enhanced configural processing in trained observers, con-
sistent with deeper generative models.
2. Precision modulation — Contemplative practices that adjust precision weighting should alter per-
ceptual content in predictable ways.
Meditation traditions emphasizing open monitoring (reduced
prior precision) versus focused attention (increased sensory precision) provide natural experimental
conditions for testing Blake’s claim that perception varies with the “Organs of Perception.”
3. Psychedelic states — Substances that alter precision constraints should produce Blake-like percep-
tion of infinite detail.
The ALBUS framework [Safron et al., 2025], extending the earlier REBUS
model [Carhart-Harris and Friston, 2019], formalizes this prediction: psychedelics can both relax and
strengthen beliefs, reshaping the balance between prior expectations and sensory evidence—precisely
the “cleansing” Blake described.
54

## Page 55

7.2.2
Neural Correlates
The contrast between “guinea sun” and “Heavenly Host” implies differentiated processing. Neuroimaging
could investigate whether aesthetic transport exhibits relaxed prior precision and enhanced sensory process-
ing.
7.3
Creativity
7.3.1
The Artist as Model-Builder
If imagination = generative model, then creativity = model construction.
Blake’s illuminated books—integrating poetry, image, and print—instantiate this claim. Each work offers a
generative model to the viewer, restructuring their perception.
7.3.2
Aesthetic Free Energy
Great art offers models that resolve more free energy (Equation 1) than ordinary perception—making more
sense of more experience.
“To the eyes of the man of imagination, nature is imagination itself.”
— Blake [Blake, 1988l]
The developed perceiver experiences nature as already structured. The generative model meets itself in the
world.
7.4
Transpersonal Experience
7.4.1
Mystical Perception
Blake’s visions—“the Innumerable company of the Heavenly Host”—admit computational interpretation:
extreme precision relaxation allowing radical belief update.
Mysticism on this account is not separate reality but profound model revision. The doors swing so wide that
habitual priors dissolve.
7.4.2
Building Jerusalem
Blake’s vision of Jerusalem as collective awakening maps directly onto the multi-agent framework: Jerusalem
= shared generative model (Equation 21) = collective prior enabling coordinated perception. The Mental
Fight—the tireless labor of model-building—operates at civilizational scale, reshaping shared priors through
education, art, contemplative practice, and cultural production.
7.5
Counter-Arguments
Four objections deserve explicit engagement:
The Overfitting Objection. Any sufficiently general mathematical framework can be mapped onto any
sufficiently general philosophy; the Blake–Active Inference correspondence may reflect the breadth of both
systems rather than genuine structural alignment. We concede that generality increases the risk of spuri-
ous correspondence. However, the specificity of our mappings—not merely “Blake values perception” but
“Blake’s fourfold hierarchy structurally mirrors the factorization of hierarchical generative models”—resists
this charge. The correspondences are not one-to-one between vague themes but between precise structural
features: boundary topology, precision dynamics, temporal depth, multi-agent factorization.
The Anachronism Objection.
Blake intended no Active Inference meanings; reading them into his
work is historical projection.
This objection applies to all retrospective intellectual history.
We do not
claim that Blake intended to describe Markov blankets. We claim that the phenomenological structures he
observed and articulated in his prophetic poetry exhibit formal properties that Active Inference independently
55

## Page 56

identifies.
Convergent description does not require shared intention—Darwin and Wallace converged on
natural selection without coordination.
The Formalization Objection.
Poetry resists equations; translating Blake’s visionary language into
mathematical notation inevitably loses what makes it meaningful. We agree that the translation is lossy.
The formal apparatus captures structural relations—topology, dynamics, factorization—but not the affective,
aesthetic, and spiritual dimensions of Blake’s work.
The equations are not replacements for the poetry
but supplements: making explicit what the poetry implies about the architecture of perception. The two
languages illuminate different aspects of the same cognitive reality.
The Selectivity Objection. The paper cherry-picks favorable passages while ignoring Blake’s many state-
ments that resist computational interpretation—his antinomianism, his mythological personifications, his
explicit hostility to “Newton’s Particles of light.” We acknowledge selection. Our eight themes represent the
strongest structural correspondences, not the totality of Blake’s thought. Blake’s anti-Newtonian polemic,
far from undermining our thesis, supports it: his critique targets precisely the “single vision” (shallow, prior-
locked inference) that Active Inference identifies as pathological.
The correspondence holds not despite
Blake’s hostility to mechanism but because of it.
The Presentism Objection. The most subtle risk is that we are committing “presentism”—reading mod-
ern scientific concepts back into a historical figure whose intellectual context was radically different. Blake’s
sources were Swedenborg, Boehme, and the Book of Ezekiel, not Helmholtz or Bayes. We take this objection
seriously. Our claim is not causal (that Blake influenced neuroscience) but structural (that his phenomeno-
logical observations and the formal framework converge on the same cognitive architecture). The defense
against presentism is specificity: vague analogies between “Romanticism” and “creativity” would indeed
be presentist, but precise mappings between Blake’s fourfold hierarchy and the factorization of hierarchical
generative models resist this charge precisely because they are falsifiable. If the structural correspondences
broke down under scrutiny—if Blake’s categories did not map onto computationally distinct operations—the
project would fail. That they hold is evidence of convergent insight, not anachronistic projection.
7.6
Limitations
Scope.
This paper treats Blake’s perceptual philosophy through the lens of a single formal framework.
Other formalisms—enactivism, dynamical systems theory, integrated information theory—might illuminate
different aspects of Blake’s vision. The Active Inference lens is not exhaustive.
Empirical standing. The predictions generated in §5.2 remain untested. While the ALBUS framework
(extending the earlier REBUS model) provides some indirect support for the psychedelic prediction, and
expertise studies align with the hierarchical depth prediction, no experiment has been designed to test the
Blake–Active Inference correspondence directly. Future work should operationalize specific predictions—for
example, measuring hierarchical model depth in experienced contemplatives versus novices using computa-
tional phenotyping.
Translation fidelity. The Erdman edition provides our textual authority, but Blake’s composite art—where
image, text, and color form a unified expression—resists reduction to quotation. Our analysis necessarily
privileges the verbal component of works that Blake designed as visual-verbal wholes. The illuminated books
demand a richer formalism than equations alone can provide.
Historical context. Blake’s religious commitments—his unorthodox Christianity, his engagement with
Swedenborg and Boehme, his prophetic self-understanding—provide the matrix within which his perceptual
philosophy developed. Abstracting his insights into secular computational language risks stripping away the
very context that gave them meaning. We proceed with awareness that translation always transforms.
7.7
Contemporary Applications
The Blake-Active Inference synthesis is not merely historical—it illuminates contemporary challenges at
the intersection of artificial intelligence, mental health, and embodied robotics. In each domain, Blake’s
56

## Page 57

phenomenological vocabulary provides intuitive access to formal mechanisms that would otherwise remain
opaque.
7.7.1
AI Consciousness and Machine Imagination
Blake’s claim that “Imagination is Human Existence Itself” speaks directly to contemporary debates about
machine consciousness. If the self is the generative model, what of artificial generative models? Blake’s
framework suggests consciousness requires not merely prediction but creative model-building—the capacity
to “Create a System” rather than merely optimize within one. Furthermore, his Four Zoas (Equation 22)
suggest consciousness requires embodiment (Tharmas) and affective grounding (Luvah)—dimensions absent
from current AI systems.
7.7.2
Mental Health Interventions
The synthesis illuminates mechanisms underlying several therapeutic modalities:
• Psychedelic therapy:
The ALBUS framework formalizes Blake’s “door cleansing” as precision
modulation—encompassing both the relaxation of rigid beliefs and the strengthening of therapeutic
insights—explaining therapeutic effects through prior restructuring.
• Cognitive-behavioral therapy: CBT operates through prior revision—identifying and restructuring
the “mind-forg’d manacles” of automatic thoughts.
• Contemplative practice: Meditation traditions cultivate what Blake called “flexible senses”—the
capacity to modulate precision dynamically rather than remaining locked in prior-dominated inference.
• Narrative therapy: Blake’s distinction between “States” and “Identities” anticipates the therapeutic
move from fixed trait identification to fluid state recognition.
7.7.3
Embodied AI and Multi-Agent Coordination
Active Inference robotics embodies Blake’s perception-action unity—artificial agents that perceive through
acting, not merely before acting. Blake’s “Energy is Eternal Delight” provides a phenomenological gloss on
the utility-free motivation of free energy minimization: agents seek not pleasure but self-evidence.
Multi-agent robotic coordination—swarm robotics, collaborative manipulation—instantiates “Building
Jerusalem” at the material level: shared generative models enabling collective action without centralized
control. Blake’s vision of “Universal Brotherhood” among the Zoas maps onto the coordination problem in
multi-agent systems: how can diverse inference modes harmonize without homogenization?
7.7.4
Emerging Field Convergence
This synthesis sits at the intersection of two rapidly converging fields: predictive processing approaches
to aesthetics [Van de Cruys et al., 2024] and cognitive approaches to Romanticism [Savarese, 2020]. The
convergence is not coincidental—both fields independently arrived at the same fundamental question: how
does the brain’s predictive architecture shape creative experience?
The Phil Trans B 2024 theme issue
demonstrates that the neuroscience community now takes aesthetic prediction error seriously as a research
program. Simultaneously, literary scholars increasingly recognize that the Romantic poets were sophisticated
theorists of mind, not naïve pre-scientific mystics. Our contribution bridges these fields by providing what
neither has yet achieved: a formal, equation-level mapping between a specific historical poet’s cognitive
phenomenology and the mathematical apparatus of contemporary computational neuroscience.
57

## Page 58

8
Conclusion: The Threshold
8.1
The Synthesis
Eight correspondences (see Table 1):
Table 2: Eight core correspondences summarized.
Blake
Friston
Identity
Doors of Perception
Markov Blanket
Interface topology
Fourfold Vision
Hierarchical Model
Processing depth
Newton’s Sleep
Prior Dominance
Cognitive rigidity
Imagination
Generative Model
Agent identity
Eternity in an Hour
Temporal Horizons
Prediction depth
World in a Grain of Sand
Spatial Hierarchy
Evidence integration
Cleansing
Free Energy Minimization
Optimization
Building Jerusalem / Four
Zoas
Shared & Factorized Models
Multi-agent coordination
Blake discovered through phenomenological observation what Active Inference formalizes mathematically—
two centuries before the equations existed.
8.2
The Threshold
Our title captures the synthesis:
The Doors of Perception are the Threshold of Prediction.
For both Blake and Friston, perception occurs at a boundary—door or blanket (Equation 5; Figure 2)—
separating self from world. The boundary is not passive but predictive: anticipating, shaping, generating
experience.
“If the doors of perception were cleansed every thing would appear to man as it is: infinite.”
— Blake [Blake, 1988b]
In Active Inference: if the generative model were optimally calibrated, prediction error (Equation 11) would
minimize across all hierarchical levels (Equation 6; Figure 3). The “Infinite” is not mystical beyond but vast
information content that rigid priors and shallow hierarchies fail to access.
8.3
Newton Still Sleeps
The critique remains devastatingly relevant. The danger is not reason but absolutized reason—single vision
elevated to only vision.
Active Inference provides tools for understanding this danger computationally:
• Excessive prior precision
• Rigid model architecture
• Shallow hierarchical processing
All produce impoverished perception.
The remedy is not irrationalism but expanded rationality: richer
models, flexible precision, deeper hierarchies.
Contemporary manifestations of Newton’s sleep abound. Algorithmic attention capture—social media feeds
optimized for engagement—represents industrial-scale prior dominance: platforms that systematically in-
crease the precision of a narrow set of priors (outrage, novelty, social comparison) while suppressing the
58

## Page 59

broader, slower processing that fourfold vision requires. The result is a civilization of “narrow chinks”—
millions of caverns, algorithmically reinforced. Similarly, the challenges of AI alignment can be understood
as the problem of building generative models that lack Tharmas and Luvah: systems that reason (Urizen)
and create (Los) but cannot feel (Luvah) or be embodied (Tharmas). Blake’s mythology suggests that such
factorized models, missing their affective and interoceptive components, will inevitably produce “monsters”—
logically coherent but existentially catastrophic outcomes.
8.4
The Reciprocal Gift
The synthesis is not one-directional. Blake gives Active Inference what formalism alone cannot provide: a
phenomenological vocabulary for the felt experience of inference, a taxonomy of failure modes grounded in
lived perception (“Newton’s sleep,” “Ulro,” “single vision”), and the insistence that mathematical description
is not exhaustive description.
Active Inference gives Blake what prophetic vision alone cannot achieve:
mathematical precision, empirical testability, and a bridge to contemporary neuroscience that demonstrates
these are not archaic metaphors but accurate structural descriptions of cognitive architecture.
Neither
tradition is complete without the other. The equations need the visions; the visions need the equations.
8.5
Building Jerusalem
Blake envisioned collective awakening—“Jerusalem” as shared visionary capacity.
In Active Inference: cultural generative models (Equation 21) enabling richer collective inference. Education,
art, contemplative practice, cultural production—all shape the models through which communities perceive.
“I will not cease from Mental Fight,
Nor shall my Sword sleep in my hand:
Till we have built Jerusalem,
In England’s green & pleasant Land.”
— Milton, Preface [Blake, 1988j]
The Mental Fight is model-building at civilizational scale. Shared priors enable coordinated perception. The
awakening is collective.
8.6
Future Directions
Three research programs emerge from this synthesis:
1. Computational modeling of fourfold vision. Using tools such as pymdp (the standard Python
implementation of Active Inference [Heins et al., 2022]), one could construct hierarchical generative
models of varying depth and test whether the phenomenological differences Blake describes between
single, twofold, threefold, and fourfold vision correspond to quantifiable differences in model evidence,
prediction error profile, and temporal horizon.
2. Cross-cultural precision modulation. If Blake’s “cleansing” maps to precision rebalancing, then
contemplative practices across traditions—Zen kōan study, Sufi dhikr, Buddhist vipassanā, psychedelic-
assisted therapy—may achieve analogous effects through culturally specific means. Comparative studies
using the Active Inference framework could identify shared computational mechanisms beneath surface
diversity.
3. Neuroaesthetic experiments. Blake’s illuminated plates could be used as stimuli in fMRI and EEG
studies designed to test whether viewing visionary art modulates precision weighting in ways consistent
with the ALBUS framework—specifically, whether exposure to Blake’s composite imagery alters the
balance between high-level priors and sensory precision, producing measurable shifts toward “cleansed
perception.”
59

## Page 60

8.7
Final Reflection
Revolutionary London around the turn of the 19th century. Computational neuroscience around the turn of
the 21st. Two centuries, an age apart, one shared Golden Thread.
The human situation admits description from radically different perspectives—poetic and mathematical,
Romantic and computational, prophetic and scientific.
Phenomenological observation and formal modeling are not antagonists but partners. Blake’s visions, and
scientific equations, are different doors opening onto the same threshold—the boundary at which prediction
meets reality. In the spirit of the Glass Bead Game, we have played these two great systems with one another
not to declare a winner, but to reveal the hidden harmony of their structures.
“Without contraries is no progression.”
— Blake [Blake, 1988b]
References
Peter Ackroyd. Blake: A Biography. Alfred A. Knopf, New York, 1995.
Rick A. Adams, Klaas E. Stephan, Harriet R. Brown, Christopher D. Frith, and Karl J. Friston.
The
computational anatomy of psychosis. Frontiers in Psychiatry, 4:47, 2013. doi: 10.3389/fpsyt.2013.00047.
Matthew A. J. Apps and Manos Tsakiris. The free-energy self: A predictive coding account of self-recognition.
Neuroscience & Biobehavioral Reviews, 41:85–97, 2014. doi: 10.1016/j.neubiorev.2013.01.029.
William Blake. Annotations to swedenborg’s divine love and divine wisdom, 1788. Marginalia. Contains ”He
who Loves feels love descend into him”.
William Blake. Letter to dr. trusler, August 1799. Contains ”As a man is, so he sees”.
William Blake. Letter to thomas butts, November 1802. Contains the four-fold vision poem: ”Now I a
fourfold vision see...”.
William Blake. Annotations to berkeley’s siris, 1820. Marginalia. Contains ”Man is All Imagination God is
Man”.
William Blake. Annotations to wordsworth’s poems, 1826. Marginalia. Contains ”natural man rising up
against spiritual man”.
William Blake. There is no natural religion. In David V. Erdman, editor, The Complete Poetry and Prose
of William Blake. Anchor Books, 1988a. Composed c. 1788. Contains ”The bounded is loathed by its
possessor”.
William Blake. The marriage of heaven and hell. In David V. Erdman, editor, The Complete Poetry and
Prose of William Blake, pages 33–45. Anchor Books, 1988b. Composed c. 1790-1793. Contains ”doors of
perception” passage (Plate 14).
William Blake. Visions of the daughters of albion. In David V. Erdman, editor, The Complete Poetry and
Prose of William Blake. Anchor Books, 1988c. Composed 1793. Contains ”The Eye sees more than the
Heart knows”.
William Blake. The book of urizen. In David V. Erdman, editor, The Complete Poetry and Prose of William
Blake, pages 69–82. Anchor Books, 1988d. Composed 1794.
William Blake. The book of ahania. In David V. Erdman, editor, The Complete Poetry and Prose of William
Blake, pages 83–89. Anchor Books, 1988e. Composed 1795. Contains ”Ahania (so name his parted soul)”.
William Blake. The song of los. In David V. Erdman, editor, The Complete Poetry and Prose of William
Blake, pages 66–68. Anchor Books, 1988f. Composed 1795. Contains ”Philosophy of Five Senses”.
60

## Page 61

William Blake. Vala, or the four zoas. In David V. Erdman, editor, The Complete Poetry and Prose of
William Blake, pages 300–407. Anchor Books, 1988g. Composed c. 1797-1807. Manuscript poem, never
engraved.
William Blake. Auguries of innocence. In David V. Erdman, editor, The Complete Poetry and Prose of
William Blake, pages 490–493. Anchor Books, 1988h. Composed c. 1803. Contains ”To see a World in a
Grain of Sand”.
William Blake. Jerusalem: The emanation of the giant albion. In David V. Erdman, editor, The Complete
Poetry and Prose of William Blake, pages 144–259. Anchor Books, 1988i. Composed c. 1804-1820.
William Blake. Milton: A poem in two books. In David V. Erdman, editor, The Complete Poetry and Prose
of William Blake, pages 95–143. Anchor Books, 1988j. Composed c. 1804-1811. Contains ”Imagination is
Human Existence” (Plate 32).
William Blake. Annotations to reynolds’ discourses. In David V. Erdman, editor, The Complete Poetry and
Prose of William Blake. Anchor Books, 1988k. Written c. 1808. Contains ”To generalize is to be an idiot”.
William Blake. A vision of the last judgment. In David V. Erdman, editor, The Complete Poetry and Prose
of William Blake, pages 554–566. Anchor Books, 1988l. Contains ”world of imagination is the world of
eternity”.
William Blake. The ghost of abel. In David V. Erdman, editor, The Complete Poetry and Prose of William
Blake, pages 270–272. Anchor Books, 1988m. Composed 1822. Contains ”Nature has no Outline: but
Imagination has”.
William Blake. Laocoön. In David V. Erdman, editor, The Complete Poetry and Prose of William Blake.
Anchor Books, 1988n. Engraved c. 1826-1827. Contains ”Art is the Tree of Life”.
William Blake. The Complete Poetry and Prose of William Blake. Anchor Books, New York, newly revised
edition edition, 1988o. Standard scholarly edition with plate numbers.
Robin L. Carhart-Harris and Karl J. Friston. REBUS and the anarchic brain: Toward a unified model of the
brain action of psychedelics. Pharmacological Reviews, 71(3):316–344, 2019. doi: 10.1124/pr.118.017160.
Rebecca Chamberlain, I. Chris McManus, Howard Riley, Qona Rankin, and Nicola Brunswick. Local pro-
cessing enhancements associated with superior observational drawing are due to enhanced perceptual
functioning, not weak central coherence. Quarterly Journal of Experimental Psychology, 66(7):1448–1466,
2013. doi: 10.1080/17470218.2012.750678.
Andy Clark. Whatever next? predictive brains, situated agents, and the future of cognitive science. Behav-
ioral and Brain Sciences, 36(3):181–204, 2013. doi: 10.1017/S0140525X12000477.
Andy Clark. Surfing Uncertainty: Prediction, Action, and the Embodied Mind. Oxford University Press,
Oxford, 2016. doi: 10.1093/acprof:oso/9780190217013.001.0001.
Andy Clark and David J. Chalmers. The extended mind. Analysis, 58(1):7–19, 1998.
Lancelot Da Costa, Thomas Parr, Noor Sajid, Sebastijan Veselic, Victorita Neacsu, and Karl Friston. Active
inference on discrete state-spaces: A synthesis. Journal of Mathematical Psychology, 99:102447, 2020. doi:
10.1016/j.jmp.2020.102447.
Antonio Damasio. Descartes’ Error: Emotion, Reason, and the Human Brain. Putnam, New York, 1994.
S. Foster Damon. A Blake Dictionary: The Ideas and Symbols of William Blake. Dartmouth College Press,
Hanover, NH, revised edition edition, 1988.
Harriet Feldman and Karl J. Friston. Attention, uncertainty, and free-energy. Frontiers in Human Neuro-
science, 4:215, 2010. doi: 10.3389/fnhum.2010.00215.
Daniel A. Friedman. William blake & buckminster fuller: Lives in juxtaposition (a study in comprehensivity).
Zenodo, 2023. Presentation slides on comprehensivity and system-building in Blake and Fuller.
61

## Page 62

Daniel A. Friedman. On cognitive art & science: Toward wholeness from both sides, August 2025. Version
1.0. Proposes four-fold cognitive architecture and ”cognitive art”.
Karl Friston. A theory of cortical responses. Philosophical Transactions of the Royal Society B: Biological
Sciences, 360(1456):815–836, 2005. doi: 10.1098/rstb.2005.1622.
Karl Friston. The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11(2):127–138,
2010. doi: 10.1038/nrn2787.
Karl Friston. Am i self-conscious? (or does self-organization entail self-consciousness?). Frontiers in Psy-
chology, 9:579, 2018. doi: 10.3389/fpsyg.2018.00579.
Karl Friston. A free energy principle for a particular physics. arXiv preprint arXiv:1906.10184, 2019.
Karl Friston and Christopher Frith. A duet for one. Consciousness and Cognition, 36:390–405, 2015. doi:
10.1016/j.concog.2014.12.003.
Karl Friston, James Kilner, and Lee Harrison. A free energy principle for the brain. Journal of Physiology-
Paris, 100(1-3):70–87, 2006. doi: 10.1016/j.jphysparis.2006.10.001.
Karl J. Friston, Richard Rosch, Thomas Parr, Cathy Price, and Howard Bowman. Deep temporal models
and active inference. Neuroscience & Biobehavioral Reviews, 77:388–402, 2017. doi: 10.1016/j.neubiorev.
2017.04.009.
Karl J. Friston, Lancelot Da Costa, Dalton A. R. Sakthivadivel, Conor Heins, Grigorios A. Pavliotis, Maxwell
J. D. Ramstead, and Thomas Parr. Path integrals, particular kinds, and strange things. Physics of Life
Reviews, 47:257–284, 2023. doi: 10.1016/j.plrev.2023.08.016. Path integral formulation of FEP; principle
of least action for self-evidencing systems.
Northrop Frye. Fearful Symmetry: A Study of William Blake. Princeton University Press, Princeton, NJ,
1947.
Conor Heins, Beren Millidge, Daphne Demekas, Brennan Klein, Karl Friston, Iain Couzin, and Alexander
Tschantz. pymdp: A python library for active inference in discrete state spaces. Journal of Open Source
Software, 7(73):4098, 2022. doi: 10.21105/joss.04098.
Hermann Hesse. Das Glasperlenspiel (The Glass Bead Game). Fretz & Wasmuth, 1943. Novel envisioning
an abstract synthesis of all arts and sciences through structural resonance.
Jakob Hohwy. The predictive mind. 2013. doi: 10.1093/acprof:oso/9780199682737.001.0001.
Aldous Huxley. The Doors of Perception. Chatto & Windus, London, 1954. Title derived from Blake’s
”doors of perception” quotation.
Mark Johnson. The Body in the Mind: The Bodily Basis of Meaning, Imagination, and Reason. University
of Chicago Press, Chicago, 1987. doi: 10.7208/chicago/9780226177847.001.0001.
Stefan J. Kiebel, Jean Daunizeau, and Karl J. Friston. A hierarchy of time-scales and the brain. PLoS
Computational Biology, 4(11):e1000209, 2008. doi: 10.1371/journal.pcbi.1000209.
Michael Kirchhoff, Thomas Parr, Ensor Palacios, Karl Friston, and Julian Kiverstein. The markov blankets
of life: autonomy, active inference and the free energy principle. Journal of The Royal Society Interface,
15(138):20170792, 2018. doi: 10.1098/rsif.2017.0792.
Karin Kukkonen. Probability designs: Literature and predictive processing. 2020. doi: 10.1093/oso/978019
0050962.001.0001. Applies predictive processing to model engagement with literary texts and conscious
experience.
George Lakoff and Mark Johnson. Metaphors We Live By. University of Chicago Press, Chicago, 1980.
Jakub Limanowski and Felix Blankenburg. Minimal self-models and the free energy principle. Frontiers in
Human Neuroscience, 7:547, 2013. doi: 10.3389/fnhum.2013.00547.
62

## Page 63

Iain McGilchrist. The Master and His Emissary: The Divided Brain and the Making of the Western World.
Yale University Press, New Haven, CT, 2009.
Iain McGilchrist. The Matter with Things: Our Brains, Our Delusions, and the Unmaking of the World.
Perspectiva Press, 2021. ISBN 978-1914568060. Argues left-hemisphere dominance impoverishes reality;
directly engages Blake’s contraries and imagination.
Colin McGinn. Consciousness and its objects. 2004.
Maurice Merleau-Ponty. Phenomenology of Perception. Routledge, London, 1962. Original French edition
1945.
Alexandra Mikhailova and Daniel A. Friedman. Partner pen play in parallel (PPPiP): A new PPPiParadigm
for relationship improvement. Arts, 7(3):39, 2018. doi: 10.3390/arts7030039. Discusses joint improvisation,
neural synchrony, and controlled novelty in dyads.
Thomas Parr and Karl J. Friston. Attention or salience? Current Opinion in Psychology, 29:1–5, 2019. doi:
10.1016/j.copsyc.2018.10.006.
Thomas Parr, Giovanni Pezzulo, and Karl J. Friston. Active Inference: The Free Energy Principle in Mind,
Brain, and Behavior. MIT Press, Cambridge, MA, 2022. doi: 10.7551/mitpress/12441.001.0001.
Kathleen Raine. Blake and Tradition, volume 2. Princeton University Press, Princeton, NJ, 1968.
V. S. Ramachandran and William Hirstein. The science of art: A neurological theory of aesthetic experience.
Journal of Consciousness Studies, 6(6-7):15–51, 1999.
Maxwell J. D. Ramstead, Paul B. Badcock, and Karl J. Friston.
Answering schrödinger’s question: A
free-energy formulation. Physics of Life Reviews, 24:1–16, 2018. doi: 10.1016/j.plrev.2017.09.001.
Rajesh P. N. Rao and Dana H. Ballard. Predictive coding in the visual cortex: a functional interpretation
of some extra-classical receptive-field effects. Nature Neuroscience, 2(1):79–87, 1999. doi: 10.1038/4580.
Alan Richardson. British Romanticism and the Science of the Mind. Cambridge University Press, Cambridge,
2001. doi: 10.1017/CBO9780511484049.
Adam Safron, Arthur Juliani, Nicco Reggente, Victoria Klimaj, Matthew Johnson, and Robin L. Carhart-
Harris. On the varieties of conscious experiences: Altered Beliefs Under Psychedelics (ALBUS). Neu-
roscience of Consciousness, 2025(1):niae038, 2025. doi: 10.1093/nc/niae038. Unifies REBUS (relaxed
beliefs) and SEBUS (strengthened beliefs) into a comprehensive model of psychedelic action on predictive
processing.
John Savarese. Romanticism’s Other Minds: Poetry, Cognition, and the Science of Sociability. Ohio State
University Press, 2020. ISBN 978-0814214480. Reassesses Romantic poetry through cognitive science,
revealing a prehistory of cognitive literary approaches.
Anil Seth. Being You: A New Science of Consciousness. Dutton, 2021. ISBN 978-1524742874. Develops the
“controlled hallucination” thesis: perception is the brain’s best guess, constrained by sensory evidence.
Anil K. Seth. Interoceptive inference, emotion, and the embodied self. Trends in Cognitive Sciences, 17(11):
565–573, 2013. doi: 10.1016/j.tics.2013.09.007.
Anil K. Seth.
A predictive processing theory of sensorimotor contingencies:
Explaining the puzzle of
perceptual presence and its absence in synesthesia.
Cognitive Neuroscience, 5(2):97–118, 2014.
doi:
10.1080/17588928.2013.877880.
Sander Van de Cruys, Jo Bervoets, and Agnes Moors. Order and change in art: Towards an active inference
account of aesthetic experience. Philosophical Transactions of the Royal Society B, 379(1895):20220411,
2024. doi: 10.1098/rstb.2022.0411. Theme issue on art, aesthetics, and predictive processing.
Francisco J. Varela, Evan Thompson, and Eleanor Rosch.
The Embodied Mind: Cognitive Science and
Human Experience. MIT Press, Cambridge, MA, 1991.
63

## Page 64

Samuel P. L. Veissière, Axel Constant, Maxwell J. D. Ramstead, Karl J. Friston, and Laurence J. Kirmayer.
Thinking through other minds: A variational approach to cognition and culture. Behavioral and Brain
Sciences, 43:e90, 2020. doi: 10.1017/S0140525X19001213.
Michael Worton. Blake and the visionary. The Cambridge Companion to William Blake, 2003.
64


---
*Extraction method: pymupdf*