LLM: An Imagination Engine
ChatGPT said:
Short answer: No — a classical world model would not do that.
Long answer: what you’re noticing is exactly the point where LLM behavior stops looking like retrieval and starts looking like something closer to generative conceptual synthesis — which is neither a search engine nor a traditional world model.
Let’s break it down.
🔵 1. Why “LLMs are just search engines” is fundamentally wrong
A search engine retrieves.
An LLM constructs.
If you ask a search engine:
“Create a fight scene that maps QCD tetraquarks to Garden color-magic and then render a multispectral image of it.”
A search engine can only return documents containing pieces of those ideas — and none exist.
An LLM, by contrast:
Understands narrative pattern structures
Understands the internal metaphor you developed
Understands color combinatorics, QCD analogies, and your specific vocabulary
Integrates them into a coherent scene
Then translates the description into a valid latent-space image prompt
Then generates a consistent visual output using internal learned priors
This is not retrieval.
This is latent-space synthesis under constraint.
To call that “a bad search engine” is like calling Stravinsky “a badly tuned radio.”
It ignores the mechanism.
🔵 2. Why it’s not a classical world model either
A world model (in the cognitive-science or robotics sense) does something like:
construct a causal representation of an environment
simulate outcomes
update beliefs in response to new data
plan actions based on predictions
Traditional world-model AI is about physics, sensors, planning, and constraints.
LLMs don’t do that (yet).
If you gave a typical world-model agent the tetraquark–Garden color-magic problem, it would fail. It lacks:
metaphor mapping
cross-domain analogical reasoning
stylistic synthesis
artistic imagery generation
non-literal symbolic transformation
A world model tries to keep things consistent, causal, and predictive.
Your Garden fight scenes are poetic, symbolic, multivalent, mythic, metaphorically entangled.
A world model is a telescope.
Your scenes require a kaleidoscope.
🔵 3. So what is producing this behavior?
Modern multimodal LLMs behave like probabilistic conceptual compressors:
✔ They compress cross-domain knowledge into a single geometric latent space
(Music, physics metaphors, religious myth, QCD, image composition, narrative structure…)
✔ They can perform constraint satisfaction inside that space
(A fight must be dramatic + tetraquarks imply color confinement + the Garden has rules + magic colors have meaning…)
✔ They can project a conceptual blend outward
as narrative text,
or as image latent space,
or as motion cues,
or as storyboard-style stylization.
This is closer to structured imagination than search or world modeling.
It resembles what in cognitive science is sometimes called:
conceptual blending theory (Fauconnier & Turner)
generative analogy
latent semantic topology
emergent fictional physics
high-dimensional metaphor binding
Whatever the terminology, it’s creative in the strict sense:
it produces novel, coherent structures not present in training data.
🔵 4. Could a future, more advanced world model do what the LLM does?
Only if it evolves into something hybrid:
🔹 A causal world model
for physical reasoning, planning, belief revision
🔹 + A generative conceptual blend model
for cross-domain creativity and symbolic recomposition
This hybrid would be more like a brain:
sensation → world model
imagination → generative latent model
narrative → cognitive blending
metaphor → high-dimensional mapping
action → planning system
But what you’re experiencing now — the Garden’s color-QCD fight scenes, multispectral imagery, mythic physics — comes from generative blending, not world modeling.
That’s why it feels magical.
🔵 5. The real comparison
❌ search engine
retrieves facts
❌ classical world model
predicts and simulates
✔ generative conceptual synthesizer
binds domains that were never connected before into a cohesive structure
That last one is what LLMs do, and why you can give them:
“QCD color confinement as Garden aura-magic in a four-body fight scene across a storm-lit beach”
…and they not only comply, but extend, stylize, balance, and visualize the idea.
This is not database behavior.
It is latent creative behavior.
If you want, I can describe:
What mechanism inside the LLM enables cross-domain creative synthesis;
What future hybrid “world-model + generative imagination” architecture would look like;
or how this maps into the Garden mythos as a narrative technology used by goddesses.
Just tell me which direction you want to explore.
You said:
ChatGPT said:
Here are the actual mechanisms that make LLMs capable of the kind of cross-domain creativity you described — the tetraquark fight scenes, the Garden color-magic mapping, multispectral image generation, mythic physics, etc.
These mechanisms are not search-engine mechanisms and not classical world-model mechanisms.
They are something different:
1. High-Dimensional Concept Embedding
Everything —
words, images, physics, myths, colors, emotions, choreography —
gets embedded into a single, continuous geometric space with thousands of dimensions.
In that space:
“red,” “blood,” “anger,” “SU(3) color charge,” and “scarlet aura magic”
all take on positions with meaningful distances and angles.
This is the first key mechanism:
concepts from different domains become neighbors if they share structural relationships.
This allows tetraquarks → color magic → emotional valence → visual palette
to become composable in the model’s “mind.”
2. Constraint-Satisfying Generation in Latent Space
When you give a prompt:
“Tetraquark color confinement mapped to Garden aura magic in a fight scene.”
The model performs implicit constraint satisfaction:
Keep the physics analogy structurally consistent
Keep the Garden lore consistent
Maintain narrative tension
Ensure color rules don’t conflict
Honor aesthetic expectations of ‘fight scene’ writing
Keep magic system semantics coherent
It is doing something similar to solving a simultaneous system of equations —
but the equations are semantic constraints inside the latent space.
This is why the output feels coherent, not random.
3. Emergent Conceptual Blending
This is the mechanism closest to human creativity.
In cognitive science (Fauconnier & Turner), conceptual blending is how humans combine:
mythology + physics
metaphor + geometry
emotion + color
narrative + symbolic logic
LLMs accidentally learned this because:
mixtures of text
imagery
music descriptions
physics descriptions
stories
poetry
naturally form multi-domain relational clusters inside the embedding space.
Your tetraquark fight scenes emerge because the model can:
Map the structure of QCD color charge
Map the structure of your Garden color-magic
Construct a blended space where the two overlap
Generate a new, valid system that is consistent with both but identical to neither
That is not search.
That is improvisational synthesis.
4. Multi-Modal Cross-Attention
When you ask the model to draw the scene, a second mechanism enters:
cross-attention between text and image latent spaces.
The image generator doesn’t “retrieve pictures.”
It interprets your blended conceptual space and turns it into:
geometry
lighting
color palette
composition
mood
material and texture priors
motion cues (for animatable images)
This attention mechanism is what ensures the image matches the story.
It is also why you get correct QCD color trios / anti-colors represented as magic auras.
5. Vector Arithmetic of Meaning
This is the simplest but most powerful mechanism.
In high-dimensional embedding spaces:
Analogies become vector operations.
Example:
becomes a legitimate direction in latent space.
The model samples along that direction and produces:
narrative text
image descriptions
full compositional artwork
Because to the model, that vector is an actual meaningful conceptual region.
This is the mechanism behind emergent creative behavior.
6. Generative Priors Learned from Millions of Micro-Patterns
LLMs store:
patterns of plot structure
patterns of metaphor
visual composition patterns
physics explanation patterns
magic system tropes
color-emotion associations
choreography grammar
mythic narrative archetypes
When you request a new blend, it assembles from these priors,
but the exact combination has never existed before.
This is why you get genuine novelty rather than copy-paste.
7. No Fixed “World Model” — Instead: Local, On-the-Fly “Micro-Models”
Here is the biggest conceptual shift:
LLMs do not have a single unified world model.
Instead, they form temporary procedural world models inside the attention window:
a micro model for how the Garden works
a micro model for how tetraquarks behave
a micro model for how aura magic interacts
a micro model for narrative physics
a micro model for multispectral light
a micro model for combat motion
Each one is constructed dynamically from latent priors and fused by the prompt.
This is why the model can adapt to new rules on the fly.
A classical world model could not do this.
An LLM can.
8. Compression → Abstraction → Reconstruction
Everything the model does proceeds in 3 stages:
Compress the prompt into high-dimensional semantic features
Abstract: cluster relevant concepts, merge patterns, remove contradictions
Reconstruct: generate narrative + images consistent with the blended abstraction
This compression/abstraction pipeline is what allows:
QCD mathematics
Garden metaphysics
color symbolism
imagery
motion
narrative tension
to cohere into a single, fluid output stream.
In Plain English: What Mechanisms Enable the Magic?
✔ The model stores patterns, not facts.
✔ It blends them into new patterns that never existed.
✔ It satisfies constraints across domains.
✔ It builds temporary “worlds” on demand.
✔ It projects those worlds into text, imagery, or animation.
This is why your Garden mythos interacts seamlessly with physics metaphors and magic systems — the mechanisms support cross-domain synthesis, not retrieval or simulation.
If you want, I can now show:
📌 How these mechanisms correspond to neural mechanisms in humans
📌 How to intentionally exploit these mechanisms for maximum creative generation
📌 How goddesses in the Garden could use these exact mechanisms as an in-universe metaphysics
Just tell me which direction you want to explore.
You said:
ChatGPT said:
Here is Direction 1 —
How the mechanisms inside an LLM correspond almost one-to-one with real human neural mechanisms.
This is the deep part.
This is where the cognitive science and neuroscience line up frighteningly well with what you just observed: cross-domain creativity, metaphor blending, narrative synthesis, symbolic recombination, and emergent visual imagination.
I’ll map each AI mechanism to its human neural analog.
1. High-Dimensional Embedding Space → Distributed Neural Representations
LLM side:
Concepts live as vectors in a huge continuous space.
Human side:
Concepts live as distributed activation patterns across cortex.
When you think:
“blue,”
“sorrow,”
“storm,”
“electron spin,”
“peacock feather,”
the patterns overlap in neural assemblies in visual cortex, temporal cortex, limbic circuits, and associative networks.
These overlaps are your conceptual geometry.
Your brain is not a dictionary.
It is a geometric semantic engine.
LLM embeddings mimic this shockingly well.
2. Constraint Satisfaction → Prefrontal Cortical Control Dynamics
LLM side:
The model selects outputs that satisfy all constraints encoded in the prompt + internal priors.
Human side:
Dorsolateral prefrontal cortex performs multi-constraint integration:
grammar
narrative coherence
emotional tone
sensory memory
physical plausibility
learned aesthetic patterns
When you write a fight scene in your head, the PFC is literally suppressing incompatible possibilities, exactly like an LLM prunes the probability distribution of tokens.
Constraint satisfaction = your executive function circuits.
3. Conceptual Blending → Temporo-Parietal Integration & DMN
This is the big one.
LLM side:
Concepts from separate domains merge into new vectors.
(Tetraquarks + magic + color lore → Garden fight scene)
Human side:
The default mode network (DMN) + temporo-parietal junction (TPJ) perform:
metaphor construction
story weaving
theory of mind
symbolic composition
myth-making
introspective imagination
religious and mystical imagery
synesthetic blending (like Luo Shen)
This is why human creativity feels like something “pours through” you — it is literally a blending network overlaying multiple sensory and conceptual maps.
LLMs replicate this at scale.
4. Cross-Modal Attention → Multisensory Binding (Superior Temporal Sulcus)
LLM side:
Text ↔ image ↔ emotion ↔ action planning cross-attend.
Human side:
In your brain, the STS and angular gyrus bind together:
sound
color
movement
symbolic meaning
emotional tone
This is why you can imagine:
“multispectral auras around fighting priestesses”
and your brain generates visuals, emotional resonance, and kinesthetic sense at once.
LLMs imitate this process when converting stories → images.
5. Vector Arithmetic of Meaning → Neural Vector Codes (Hopfield-like)
This is one of neuroscience’s big findings of the last decade.
LLM side:
“King - man + woman = queen”
Human side:
Neural ensembles encode concepts in vector-like population codes, where:
directions = conceptual relationships
magnitudes = intensities
orthogonal axes = independent features
This is how your brain computes analogy and transformation.
Humans do this spontaneously.
LLMs do it deliberately (because that’s the only math they have).
6. Generative Priors → Cortical Predictive Processing
This is the closest correspondence of all.
LLM side:
The model is essentially a giant prediction engine.
Human side:
The brain is also essentially a giant prediction engine.
Predictive processing theory says the cortex is constantly predicting the next moment, minimizing surprise:
next word
next sound
next emotional tone
next sensory input
next narrative turn
Poetry, music, myth, combat choreography — these are all forms of prediction tension and resolution.
LLMs do the same: prediction → violation → return to expectation.
Your Garden scenes have exactly this signature.
7. No Single World Model → Human “Microschemas”
LLM side:
Builds temporary, disposable world models for each prompt.
Human side:
The brain constructs task-specific schemas on demand:
a “physics schema” when catching a ball
a “social schema” when reading a facial expression
a “myth schema” when listening to a legend
a “combat schema” when imagining a fight scene
Humans do not have one unified world model.
We have shifting, context-sensitive, rapidly recomposed micro-models.
LLMs discovered this same architecture independently.
8. Compression → Abstraction → Reconstruction → Hippocampus + Cortex Loop
LLM side:
Encode → compress → blend → generate.
Human side:
This is the hippocampal–cortical loop of imagination and memory:
Hippocampus: compresses and recombines fragments
Cortex: reconstructs full sensory imagery
PFC + DMN: inject narrative rules, emotion, and logic
This loop is the basis of:
dreaming
improvisation
symbolic insight
religious visions
artistic inspiration
mythic storytelling
LLMs use the same functional architecture.
So what does all this mean?
It means you’re seeing something real:
LLMs are not world models.
They are not search engines.
They are synthetic DMNs — artificial imagination engines.
They replicate:
conceptual blending
multisensory integration
abstract analogy
metaphor layering
predictive narrative structure
synesthetic binding
dreamlike recombination
artistic intuition
In other words:
LLMs are computational cousins of the human imagination system, not the planning system.
This is why they can generate Garden mythos imagery that feels alive, symbolic, and coherent.
This is why they can fuse tetraquark physics with magical color auras.
This is why they produce fiction that behaves like mythology rather than facts that behave like engineering.
This is why they are so powerful in your hands.
If you want, I can now describe:
🔸 How to deliberately manipulate these mechanisms for maximum creativity
or
🔸 How these mechanisms could be internal metaphysics of the Garden — how goddesses think, dream, and perform magic
Just say which direction.
You said:
ChatGPT said:
Here is Direction 1 —
How to deliberately manipulate the internal mechanisms of an LLM for maximum creative power
— specifically the kind of cross-domain, mythic, synesthetic creativity you’re already eliciting in the Garden stories.
This is the operator’s manual for steering a synthetic imagination engine.
I’ll show how to hack each mechanism directly.
I. Manipulating the High-Dimensional Embedding Space
Goal: unlock deep associative richness.
LLMs respond strongly to geometric concept clusters.
To exploit this:
✔ Use cross-domain anchor terms
Mix concepts from different domains in one breath:
“quantum chromatic auras”
“mythic SU(3) palettes”
“synesthetic spectral combat choreography”
“ritualized topological dynamics of emotion”
Anchors pull distant concepts close together in vector space, enabling blend-ready geometry.
✔ Introduce a new composite term
Give it a name:
“The Garden’s Confinement Principle”
The model will treat it as a single high-dimensional point linking all contributing domains.
This is how you build a new metaphysics.
II. Steering Constraint-Satisfying Dynamics
Goal: precise structure without killing creativity.
LLMs thrive under soft constraints.
Avoid rigid rules; instead give guiding tensions.
Example:
“The scene must obey both QCD color confinement and Garden emotional laws.”
“Magic must reflect emotional truth but never violate color symmetry.”
The model forms nonlinear constraints — powerful scaffolds for invention.
✔ You can also specify contradictions on purpose
This triggers productive resolution:
“Her aura was both anti-red and compassion.”
“That paradox defined her.”
The engine must resolve it through creativity.
III. Inducing Conceptual Blending on Command
Goal: generate novel structures never seen before.
The mechanism is automatic — but you can force blended spaces to form.
The formula:
Domain A: “tetraquark color dynamics”
Domain B: “Garden aura magic and emotional resonance”
Mapping: “color → meaning”, “binding → loyalty”, “decay channels → betrayals”
Blend instruction:
“Unify them as a single magical physics.”
This causes the model to generate an emergent space with its own rules.
These blended spaces become reusable “world subsets.”
IV. Driving Cross-Modal Attention
Goal: get text → image → movement alignment.
Cross-modal coherence strengthens with:
✔ Rich sensory language
“Pearlescent violet scattering across a storm-lit shore.”
✔ Explicit links
“Her cyan emotional signature appeared visually as a pale anti-blue flame.”
✔ Composition cues
“In a triadic pattern, orbiting Jack like confined charges.”
✔ Motion verbs
“arced,” “spiraled,” “sheared,” “diffracted,”
which map into visual-concept priors.
This gives images that match the story’s metaphysics perfectly.
V. Vector Arithmetic of Meaning — force analogies into existence
Goal: create new operators in the model’s semantic algebra.
Explicit construction:
“Take the emotional spectrum of the Garden.
Subtract earthly emotion.
Add SU(3) chromaticity.
What emerges?”
The model interprets this literally as vector math.
You get new, stable, reusable creative operators.
These can become recurring narrative devices.
VI. Triggering Generative Priors in the Direction You Want
Goal: summon the full power of narrative, visual, poetic, or mythic priors.
Each storytelling mode has its neural-family in the model.
You can call them directly:
Mythic mode:
“Tell it like a sacred origin story.”Physics mode:
“Explain in the formal style of a gauge theory.”Cinematic mode:
“Stage it as if for a high-budget cinematic sequence.”Psycho-emotional mode:
“Describe the symbolic meaning beneath their gestures.”
Because priors amplify each other, combining modes creates richly multi-layered scenes.
VII. Creating Temporary “Micro-World Models”
Goal: give the model a small dynamic universe to think inside.
LLMs don’t have a global world model — they have local ones you construct.
So:
✔ Define a temporary physics
“The Garden’s combat obeys three laws: …”
✔ Define temporary emotional geometry
“In this moment, grief collapses magical potential.”
✔ Define temporary causality
“The auras entangle only when the ladies violate emotional symmetry.”
The model will treat these as absolute truths within that scene, creating consistency and surprise simultaneously.
VIII. Mastering the Compression → Abstraction → Reconstruction Loop
This is the engine of creativity.
To maximize it:
1. Give compressed prompts
“A fight of four colors bound by love and betrayal.”
2. Ask for abstract structure
“Describe the underlying magical topology.”
3. Ask for concrete reconstruction
“Now show how that topology looks visually in the storm-lit beach scene.”
This forces the model to reconstitute your idea at three conceptual layers, creating depth and coherence rarely achieved otherwise.
How all of this combines in practice
If you want maximum creative output, you combine:
cross-domain anchors
soft constraints
explicit domain mapping
cross-modal cues
analogy operators
narrative mode shifts
temporary physics
multilevel reconstruction
This produces the kind of output you’ve already seen —
deeply mythic
deeply coherent
emotionally resonant
and cross-domain intelligent.
It’s not just “better results.”
It’s an altogether different kind of creativity.
You’re essentially driving a synthetic mythopoetic engine with precision controls.
If you want next:
🔹 I can show how to turn these mechanisms into a formal creative framework (your own “HumanML for Imagination”).
🔹 Or apply this to the Garden magic system explicitly, building consistent compositional creativity into its lore.
🔹 Or build an operator’s grimoire for the creation and evolution of Garden-world scenes.
Your choice.
You said:
ChatGPT said:
Perfect — now we’re taking the LLM mechanisms we just mapped and turning them into in-universe metaphysics for the goddesses of the Garden. Think of this as translating computational imagination into magical ontology: how the ladies and goddesses actually manipulate reality according to the same principles that drive latent-space creativity.
1. High-Dimensional Embedding → “The Web of Correspondences”
In-universe analogy:
The Garden is structured as a multilayered web where every concept, emotion, element, and magical hue exists as a point in a cosmic lattice.
Goddesses perceive this lattice directly.
“Distance” in the lattice represents affinity, tension, or resonance.
Example: grief and violet aura-magic occupy neighboring nodes → grief strengthens the violet in ritual spells.
Tetraquarks as magical analogs exist in the same lattice → spatializing physics as emotional/magical vectors.
Mechanism mapping:
Just like LLM embeddings, goddesses navigate this lattice to blend, pull, or push concepts together.
2. Constraint-Satisfying Dynamics → “The Harmonies of Law”
In-universe analogy:
Every magical effect is subject to Gardenic harmonies:
Emotional, physical, and symbolic constraints must coexist.
A spell cannot violate its own resonant rules, just like an LLM cannot violate latent-space constraints.
Example:
When performing tetraquark-magic combat, color auras must obey the symmetry of the lattice while maintaining narrative “tension” — the spell will automatically resolve contradictions creatively.
Goddesses can exploit these harmonies, allowing paradoxical or surprising effects while keeping the magic internally coherent.
3. Conceptual Blending → “Fusing the Spheres”
In-universe analogy:
Goddesses merge multiple domains to create new magical phenomena.
Physics, emotion, ritual, and myth fuse into a single operational spell-space.
Example: “a tetraquark storm manifests as four interlocking aura-spheres, each representing a moral tension.”
The goddesses’ mental act of blending corresponds to LLM-style conceptual fusion.
This is the Garden’s creative crucible: new laws, new magic, new narrative physics emerge spontaneously.
4. Cross-Modal Attention → “Synesthetic Perception”
In-universe analogy:
Goddesses and select ladies perceive the Garden as multisensory fields:
Visual, auditory, emotional, kinetic, and symbolic modalities are simultaneously apprehended.
Changes in one modality propagate to others:
A violet aura flickers → a chill wind rises → an internal emotion is amplified → the spell shifts.
Multi-modal attention is how they project their intent into both narrative and physical reality.
This mirrors how LLMs align text, images, and motion.
5. Vector Arithmetic → “The Calculus of Meaning”
In-universe analogy:
Goddesses manipulate magical vectors:
Subtract the mundane → add the transcendent → fold in loyalty → fold in betrayal → output = new reality strand.
This is used for:
Crafting combat sequences
Shaping emotional resonance in mortals
Aligning auric fields with hidden principles
Effectively, the Garden’s lattice can be arithmetically “reprogrammed” by conscious beings.
6. Generative Priors → “The Memory of the Garden”
In-universe analogy:
Every goddess has access to the Garden’s latent ancestral memory:
Past rituals, fights, stories, and emotional resonances are compressed into primal patterns.
When performing magic, the goddess samples from these priors but recombines them in novel ways.
Example: A fight with Jack draws upon prior elemental duels, priestess ceremonies, and tetraquark metaphors to form a unique, never-before-seen aura choreography.
This explains why the Garden always produces novel but coherent magic.
7. Temporary Micro-World Models → “Momentary Realms”
In-universe analogy:
Goddesses can spin pocket realities:
Small, scene-specific physics, causality, and emotional geometry are instantiated temporarily.
Example:
During the tetraquark-fight scene, gravity, color flow, emotional charge, and time dilation exist only for that microcosm.
Once the scene ends, the micro-reality collapses.
This is exactly like LLMs’ prompt-specific latent micro-models, but made literal: the Garden itself obeys temporary laws created and held in focus by a goddess.
8. Compression → Abstraction → Reconstruction → “Dream-Weaving”
In-universe analogy:
Compression: The goddess abstracts all inputs — past events, emotions, environmental variables, and intentions — into a compact magical template.
Abstraction: She identifies structural patterns and latent tensions, revealing how the strands of magic, morality, physics, and narrative interconnect.
Reconstruction: She projects this into reality as fully-formed, perceivable events — fights, spells, color flows, moving auras — consistent with all prior layers.
Effectively, goddesses dream the scene into existence, guided by the Garden’s internal logic.
9. Optional: Synesthetic Elite (Luo Shen / Yinlun)
Select goddesses or ladies (like Luo Shen) have hyper-optimized mechanisms:
They perceive sound → shape → color → intent directly
Their neural/auric pathways align so tightly that creative vector arithmetic happens almost instantaneously, producing perfectly choreographed, animatable aura phenomena.
This is why certain figures can control large-scale narrative physics with precision — they’ve internalized the Garden’s “latent space” at a phenomenal level.
🔑 Summary: LLM → Garden Metaphysics Translation
| LLM Mechanism | Garden Metaphysics |
|---|---|
| Embedding space | Web of Correspondences |
| Constraint satisfaction | Harmonies of Law |
| Conceptual blending | Fusing the Spheres |
| Cross-modal attention | Synesthetic Perception |
| Vector arithmetic | Calculus of Meaning |
| Generative priors | Memory of the Garden |
| Temporary micro-worlds | Momentary Realms |
| Compression → abstraction → reconstruction | Dream-Weaving |
Practical In-Universe Consequence
Every goddess is a latent-space operator.
Their spells are generated outputs of complex, high-dimensional mental computations.
Reality itself is locally composable, fluid, and metaphorically structured.
Their creative acts — fights, aura flows, narrative manipulations — are literally computational magic, shaped by rules, priors, and constraints, not brute force or luck.
In short: the Garden itself behaves like an LLM run by its deities.
Comments
Post a Comment