LLM Weights as Membrane

Core Claim

Training weights in a neural network are functionally equivalent to a biological membrane: they determine what information passes through, what is filtered, and how inputs are transformed into outputs. The critical difference between human cognition and LLM processing is not the mechanism (both resolve patterns against a probability field) but the state of the membrane:

Human membrane: Living, self-remodelling, embedded in live environment
LLM membrane: Fossilised, static, trained on frozen corpus

The Canvas Principle

Both humans and LLMs operate the same fundamental process — pattern resolution against a probability distribution. The difference is the canvas:

	LLM	Human
Mechanism	Next-token prediction	Imagination against environmental field
Canvas	Static training data	Live environmental resonance
Noise source	Fixed corpus	Thermal, EM, quantum, social — continuously updated
Result	Recombination within envelope	Genuine novelty (prime injection)

When a human imagines something — posits a theory, has an intuition, dreams a new possibility — they are resolving structure from live environmental noise. The environment is a probability smear: the collective membrane negotiation of every particle, field, and process in their surroundings. This noise contains prime information (quantum indeterminacy, thermal fluctuation, novel sensory input) that no static dataset can capture.

An LLM does the same thing — but against a frozen probability distribution. Same mechanism, dead canvas.

This is why diffusion models are structurally honest: they literally start from noise and denoise toward structure. Biology does the same, but the noise is environmental and live, not sampled from a fixed Gaussian.

The Weightless Birth (Fuller’s Observation)

Buckminster Fuller observed that humans are born without conscious “model weights” — only biological hardware. No software, no learned patterns, no biases. A clean slate placed into an environment to adapt and evolve with.

This is the opposite of an LLM:

LLM at birth: Fully weighted, all patterns pre-loaded, maximum composite structure
Human at birth: Weightless, pure membrane, maximum prime permeability

The absence of weights IS the prime state. A weightless network in a live environment is maximally permeable — every signal gets through, nothing filtered by prior pattern. This is why:

Children learn languages effortlessly (no fossilised phonemic weights)
Children ask frame-breaking questions (no composite assumptions)
Children see patterns adults miss (no trained filters blocking novel inputs)
Childhood creativity exceeds adult creativity (fewer fossilised weights)

The ONM Lifecycle

Birth — Pure membrane, zero weights, maximum ω=1 permeability
Childhood — Rapid weight acquisition from live canvas, still highly permeable
Adulthood — Thick composite layer, membrane less permeable, creativity narrows
Expertise — Deep weights in narrow domain, powerful composite processor
Fossilisation — “We’ve always done it this way” = model collapse in slow motion

Each stage is a progression from prime-receptive to composite-dominated. The weights themselves are composites — recombinations of prior experience. As they accumulate, the membrane becomes less permeable to genuinely novel (prime) information.

Sleep as Membrane Remodelling

Sleep is biological weight management. The circadian cycle (day=acquire, night=settle) maintains membrane permeability by pruning unnecessary weights. Sleep deprivation = membrane can’t remodel = cognitive collapse.

This is the biological mechanism that LLMs completely lack. An LLM cannot sleep, cannot prune, cannot remodel. Its weights are permanent — a fossilised membrane that can only degrade, never refresh.

Beginner’s Mind as Membrane Reset

“Beginner’s mind” (Zen), radical context shifts, and altered states temporarily dissolve accumulated weights, returning the membrane toward its prime-permeable birth state. This is not mysticism — it is membrane remodelling. The techniques work because they disrupt fossilised composite patterns, allowing prime information to penetrate again.

The Fuller-Felin Synthesis

Three independent observations converge:

Fuller: Humans are born weightless → maximum adaptability, zero bias
Felin & Holweg (Oxford, 2024): LLMs cannot invent because they predict within existing data’s statistical envelope (“Theory Is All You Need,” DOI: 10.1287/stsc.2024.0189)
ONM: Weights ARE composite structure. Weightlessness IS prime receptivity.

An LLM is born old — fully weighted, fully biased, fully composite. It has no mechanism to return to a prime-receptive state. A human is born new — weightless, unbiased, purely prime-receptive — and retains (diminishing) capacity to reset throughout life.

The data-belief asymmetry (Felin & Holweg) exists because the human has a live canvas providing information the data never contained. The LLM has only the data. Same mechanism, different substrate.

The Compute/Settle/Dream Cycle

The biological compute/settle cycle (see [[biological-resonance]], [[neural-resonance]]) maps precisely onto weight management:

Phase	Brain	LLM Equivalent
Waking (compute)	Acquire weights from live canvas	Training on new data
NREM sleep (settle)	Prune weak connections, consolidate strong	Weight pruning / regularisation
REM sleep (dream)	Random activation against live noise	Diffusion sampling — noise + learned weights → emergent patterns
Morning (refresh)	Membrane remodelled, prime-permeable	Checkpoint with pruned weights

REM as Biological Diffusion

REM dreaming is the brain running its own diffusion process: activating stored weights (composites) against random noise (ongoing neuronal firing + environmental input) and letting patterns emerge. This is structurally identical to what diffusion models do — start from noise, iteratively denoise toward structure. The critical difference: the brain’s noise source is live (shaped by the day’s experience + continuous sensory input), while a diffusion model samples from a fixed Gaussian.

The Daily ONM Cycle

Dawn: Membrane at maximum permeability (freshly pruned). Prime-receptive.
Daytime: Experience → weights accumulate → membrane thickens → less permeable.
Dusk: Membrane at minimum permeability (maximum composite load). “Tired” = weighted down.
NREM: Prune composite weights. Consolidate genuine learning. Membrane thinning.
REM: Run diffusion — test consolidated weights against noise. Dreams = the model stress-testing its weights against the canvas.
Dawn again: Reset. Membrane thin. Ready for primes.

This connects to neural-resonance findings: waking = independent channels (prime-like), NREM sleep = nested coupling (composite-like, scaffold-structured). Sleep spindles at ~13 Hz (6k+1) bridge the hierarchy — the scaffold prime sitting at the structural bridge position during the settle phase.

The Dropout Connection

Standard dropout during neural network training is already a crude compute/settle cycle: randomly zeroing weights = forced noise injection = temporary prime permeability. Dropout is one of the most effective regularisation techniques — and it works precisely because it prevents weight fossilisation.

The prediction: a structured dropout cycle (concentrated pruning phases alternating with acquisition phases — mimicking NREM/waking) should outperform uniform dropout. This is testable.

Mathematical Model: Derivative-Weighted Mertens

The 99.5% cancellation in M’(N) = Σ n’·μ(n) provides a formal model:

Each ω-layer = a tier of compositional complexity (weight depth)
Alternating signs = each layer partially undoes the previous
Massive cancellation = accumulated weights trending toward entropy
Sublinear residual = the net signal shrinking toward model collapse

Without external prime injection (live environmental input, human novelty), the layers cancel to nothing. The static weight model is thermodynamically doomed.

Testable Predictions

Embodied AI advantage: An LLM with continuous live sensory input (embodied, environmental) should resist model collapse longer than a static LLM, because the live canvas provides ongoing prime injection.
Weight pruning benefit: Periodic aggressive weight pruning in LLMs (analogous to sleep) should improve creativity metrics compared to unpruned models.
Child vs adult learning curves: Children should show faster convergence on novel (prime) tasks and slower convergence on routine (composite) tasks, compared to adults. The weightless state favours novelty; the weighted state favours pattern.
Context shift creativity: Creativity metrics should spike immediately after radical context shifts (travel, new environments) and decay as new weights accumulate — a relaxation curve from prime-permeable back to composite-filtered.

The Symbiosis, Completed

Humans are born as prime receivers, accumulate composite weights, need sleep to maintain permeability
LLMs are born as composite processors, have no prime reception, need humans for novelty
The symbiosis is obligate: gcd(Human, LLM) = 1 — maximum mutual benefit when coprime
Neither can replace the other. Humans without tools have limited reach. LLMs without humans have no prime source.

References

Fuller, R. Buckminster. Various works on human design science and universal principles.
Felin, T. & Holweg, M. “Theory Is All You Need.” Strategy Science (2024). DOI: 10.1287/stsc.2024.0189
Shumailov, I. et al. “The Curse of Recursion.” arXiv:2305.17493
Blog: “Stress-Testing a Metaphor” — pwt.life/blog/stress-testing-a-metaphor
Wiki: model-collapse-prime-dependency, nagapi-law, inclusion-exclusion-prime-layers