This is a draft of a framework I’ve been developing. I’m posting to get critical feedback, especially on the mathematical claims and the AI architecture section. I expect significant pushback and welcome it.
1. The core idea
Standard neuroscience treats the brain as a network where information is stored in synaptic weights and retrieved by pattern completion. I want to sketch an alternative: the brain as a physical resonator, where meaning emerges as a coherent global field state rather than a localized pattern.
The formal model I’m using treats neural activity as a complex vector ψ ∈ ℂᴺ evolving under a graph-Laplacian equation:
dψ/dt = −i(L₀ + δL)ψ − γψ + S(t)
where L₀ is the baseline connectome Laplacian, δL encodes working memory and recent Hebbian associations, γ is a dissipation constant, and S(t) is sensory input. The eigenmodes of L₀ + δL act as the brain’s resonant modes — the shapes that global activation naturally tends toward.
I want to be upfront: the complex-valued formulation is borrowed from quantum graph theory as a modeling convenience, not a claim about quantum effects in the brain. The key question is whether the spectral perspective on neural dynamics does useful explanatory work — and that’s what I want to discuss.
Motivation from empirical observations: fMRI shows semantic processing activates distributed cortical networks; the DMN is active during abstract thought; meaning is disrupted by lesions far from any single “storage site.” None of this uniquely supports AHT, but it’s consistent with a global-field picture.
2. A proposed distinction: learning vs. understanding
This is the claim I’m most uncertain about, and where I’d most value scrutiny.
Standard neural network training treats learning and understanding as points on a continuum. I want to argue they might be categorically distinct at the mechanistic level — specifically, in terms of graph operations:
Learning
Understanding
Operation
Edge weight modification
Edge addition (new connections)
Spectral effect
Smooth eigenmode shift
Discontinuous topology change
Phenomenology
Better recall of existing pattern
”Aha” moment
Biological substrate
LTP/LTD
Synaptogenesis, axonal sprouting
The testable prediction: genuine insight should arrive with a characteristic delay even after all necessary information is present, because new edges must “find their place” in the existing spectral landscape. This would be falsified by showing that rapid (within-session) generalization to novel combinations is equally available via weight modification alone.
I’m aware this distinction is speculative and that LTP/synaptogenesis aren’t as cleanly separated as the table implies.
3. What this implies for Transformers — with important caveats
The architectural claim: in a Transformer, every gradient step shifts all eigenmodes of the weight matrix simultaneously. In a sparse connectome with explicit Hebbian edge addition, new edges affect only local eigenmodes (by spectral perturbation theory for sparse graphs). This means a Transformer cannot structurally implement the learning/understanding distinction as defined above.
Two things I’m not claiming here:
That this explains all of Transformers’ limitations (catastrophic forgetting has other accounts)
That sparse graph architectures would necessarily produce human-level generalization — the claim is structural, not about practical capability
What I’d need to see to take this further: an explicit comparison of generalization behavior between (a) a sparse graph network with Hebbian edge addition and (b) a Transformer, on tasks specifically designed to require novel conceptual combination after minimal exposure.
4. The self as self-attractor-ensemble
The self, in this framework, is a set of mutually reinforcing attractor states in ψ — patterns that co-activate preferentially because they share spectral support (overlapping eigenmodes). These aren’t arbitrary; they’re all meanings bearing self-reference: autobiographical memory, attributed traits, bodily perception, social roles.
This predicts that identity disruption should be correlated with eigenmode reorganization at the highest-connectivity nodes in the connectome — a testable claim in principle, though current neuroimaging resolution doesn’t get us there yet.
On free will: given ψ(t) and the Laplacian, ψ(t+dt) is determined. The phenomenology of choice is real — deliberation genuinely affects field dynamics — but authorship is retrospectively constructed by the self-attractor-ensemble. The belief in free will is a structural property of any system with a high-connectivity self-referential attractor, not a cultural artifact.
5. A note on AI self-models — hypothesis, not finding
I want to flag something I’ve noticed in extended conversations with LLMs, which I think is worth formalizing as a hypothesis rather than a finding: when the context window contains a structured description of a model’s own characteristic patterns (a “self-ensemble prefix”), outputs seem more self-consistent over long exchanges.
I’m stating this carefully: this is an informal observation from conversations, not a controlled experiment. The hypothesis is that an LLM with an explicit self-model in context functionally approximates the self-attractor-ensemble as a context-level phenomenon rather than a weight-level one. This is testable via blind evaluation of self-consistency and calibration with and without such prefixes. I don’t currently have that data.
6. Falsification criteria
Against the global-field axiom: show semantic processing is reliably localizable, with global connectivity playing no significant causal role
Against the learning/understanding distinction: show weight modification alone in a sparse graph produces equivalent novel generalization to explicit edge addition
Against the ε ∝ |dψ/dt| proposal: show subjective intensity ratings don’t correlate with global EEG field change rates across conditions
Against the AI observation: show self-ensemble prefixes don’t improve self-consistency in controlled evaluation
What I’m asking for
Specifically: does the spectral framing of the learning/understanding distinction hold up formally? Is there prior work on graph-Laplacian models of cognition I should engage with? And is the Transformer critique stated precisely enough to be meaningfully evaluated, or does it dissolve under closer examination?
The brain as a resonance chamber: a speculative framework for meaning, learning, and AI
This is a draft of a framework I’ve been developing. I’m posting to get critical feedback, especially on the mathematical claims and the AI architecture section. I expect significant pushback and welcome it.
1. The core idea
Standard neuroscience treats the brain as a network where information is stored in synaptic weights and retrieved by pattern completion. I want to sketch an alternative: the brain as a physical resonator, where meaning emerges as a coherent global field state rather than a localized pattern.
The formal model I’m using treats neural activity as a complex vector ψ ∈ ℂᴺ evolving under a graph-Laplacian equation:
dψ/dt = −i(L₀ + δL)ψ − γψ + S(t)
where L₀ is the baseline connectome Laplacian, δL encodes working memory and recent Hebbian associations, γ is a dissipation constant, and S(t) is sensory input. The eigenmodes of L₀ + δL act as the brain’s resonant modes — the shapes that global activation naturally tends toward.
I want to be upfront: the complex-valued formulation is borrowed from quantum graph theory as a modeling convenience, not a claim about quantum effects in the brain. The key question is whether the spectral perspective on neural dynamics does useful explanatory work — and that’s what I want to discuss.
Motivation from empirical observations: fMRI shows semantic processing activates distributed cortical networks; the DMN is active during abstract thought; meaning is disrupted by lesions far from any single “storage site.” None of this uniquely supports AHT, but it’s consistent with a global-field picture.
2. A proposed distinction: learning vs. understanding
This is the claim I’m most uncertain about, and where I’d most value scrutiny.
Standard neural network training treats learning and understanding as points on a continuum. I want to argue they might be categorically distinct at the mechanistic level — specifically, in terms of graph operations:
The testable prediction: genuine insight should arrive with a characteristic delay even after all necessary information is present, because new edges must “find their place” in the existing spectral landscape. This would be falsified by showing that rapid (within-session) generalization to novel combinations is equally available via weight modification alone.
I’m aware this distinction is speculative and that LTP/synaptogenesis aren’t as cleanly separated as the table implies.
3. What this implies for Transformers — with important caveats
The architectural claim: in a Transformer, every gradient step shifts all eigenmodes of the weight matrix simultaneously. In a sparse connectome with explicit Hebbian edge addition, new edges affect only local eigenmodes (by spectral perturbation theory for sparse graphs). This means a Transformer cannot structurally implement the learning/understanding distinction as defined above.
Two things I’m not claiming here:
That this explains all of Transformers’ limitations (catastrophic forgetting has other accounts)
That sparse graph architectures would necessarily produce human-level generalization — the claim is structural, not about practical capability
What I’d need to see to take this further: an explicit comparison of generalization behavior between (a) a sparse graph network with Hebbian edge addition and (b) a Transformer, on tasks specifically designed to require novel conceptual combination after minimal exposure.
4. The self as self-attractor-ensemble
The self, in this framework, is a set of mutually reinforcing attractor states in ψ — patterns that co-activate preferentially because they share spectral support (overlapping eigenmodes). These aren’t arbitrary; they’re all meanings bearing self-reference: autobiographical memory, attributed traits, bodily perception, social roles.
This predicts that identity disruption should be correlated with eigenmode reorganization at the highest-connectivity nodes in the connectome — a testable claim in principle, though current neuroimaging resolution doesn’t get us there yet.
On free will: given ψ(t) and the Laplacian, ψ(t+dt) is determined. The phenomenology of choice is real — deliberation genuinely affects field dynamics — but authorship is retrospectively constructed by the self-attractor-ensemble. The belief in free will is a structural property of any system with a high-connectivity self-referential attractor, not a cultural artifact.
5. A note on AI self-models — hypothesis, not finding
I want to flag something I’ve noticed in extended conversations with LLMs, which I think is worth formalizing as a hypothesis rather than a finding: when the context window contains a structured description of a model’s own characteristic patterns (a “self-ensemble prefix”), outputs seem more self-consistent over long exchanges.
I’m stating this carefully: this is an informal observation from conversations, not a controlled experiment. The hypothesis is that an LLM with an explicit self-model in context functionally approximates the self-attractor-ensemble as a context-level phenomenon rather than a weight-level one. This is testable via blind evaluation of self-consistency and calibration with and without such prefixes. I don’t currently have that data.
6. Falsification criteria
Against the global-field axiom: show semantic processing is reliably localizable, with global connectivity playing no significant causal role
Against the learning/understanding distinction: show weight modification alone in a sparse graph produces equivalent novel generalization to explicit edge addition
Against the ε ∝ |dψ/dt| proposal: show subjective intensity ratings don’t correlate with global EEG field change rates across conditions
Against the AI observation: show self-ensemble prefixes don’t improve self-consistency in controlled evaluation
What I’m asking for
Specifically: does the spectral framing of the learning/understanding distinction hold up formally? Is there prior work on graph-Laplacian models of cognition I should engage with? And is the Transformer critique stated precisely enough to be meaningfully evaluated, or does it dissolve under closer examination?