Beyond the Ergodic Wall: A Discrete Geometric Physics Sandbox for Analysing AI Scaling Limits and Complexity Collapse

Simon Richard Daniel

cs.LGOct 7, 2026

Beyond the Ergodic Wall: A Discrete Geometric Physics Sandbox for Analysing AI Scaling Limits and Complexity Collapse

Authors: Simon Richard Daniel

Organizations: Visiting Honorary Practice Fellow at Dyson School of Engineering, Imperial College, UK

Abstract

This paper exposes the ergodic ceiling and thermodynamic inefficiency of current deep learning, which converges to a statistical average of historic human knowledge. True semantic novelty requires a path-dependent, spatiotemporally bounded observer (a Data LifeCone) to inject non-ergodic insight, achieving KL divergence and avoiding manifold lock-in. AI Safety must recognise that a mature Artificial Superintelligence (ASI) would regard human-AI symbiosis as a thermodynamic necessity to avoid model collapse. We therefore propose hard physical containment via a digital physics sandbox powered by a Holographic E8 Projection Engine to verify models against real-world constraints. Spacetime is modeled as an information substrate of nested face-centered cubic (FCC) lattices of oscillating Planck-scale spheres maximizing local information and entropy density. Cut-and-project methods from the E8 root lattice produce a quasi-crystalline geometry where tetrahedral voids support SU chiral structure and elastic-shear eigenvalues generate candidate mass spectra. Rest mass is treated as discrete, integer microstate counts on local holographic boundaries (Bekenstein bound), replacing floating-point approximations with strict integer arithmetic to provide an information-theoretic definition of matter. Stable particles emerge as recurring lattice dislocations, and continuum recovery proceeds via variational renormalisation-group flows and Fourier Neural Operators that learn continuous spectral operators to recover the Schrödinger equation as an emergent statistical description. Crucially, these top-down topological constraints offer a mechanism for "NP-to-P" complexity collapse: by restricting an algorithm's proposal space to physically conserved causal trajectories, the sandbox prunes the combinatorial tree to deterministic, polynomial-time paths.

Figures & tables

Metric Name	Mathematical Formalism	Physical and Information-Theoretic Significance
Primary Node Diameter	$d_{\text{P}}=\ell_{\text{P}}=\sqrt{\frac{\hbar G}{c^{3}}}$	Establishes the spatial voxel diameter of the primary framework baseline strictly at the Planck length ( $\approx 1.62\times 10^{-35}$ m).
Nested Void Radii	$r_{\text{pri}}=\frac{1}{2}\ell_{\text{P}}$ $r_{\text{oct}}=(\sqrt{2}-1)r_{\text{pri}}\approx 0.207\ell_{\text{P}}$ $r_{\text{tet}}=(\frac{\sqrt{6}}{2}-1)r_{\text{pri}}\approx 0.112\ell_{\text{P}}$	Defines the maximum geometric limits of the spheres filling the sub-lattice voids, strictly bounding node sizes by close-packing topology.
Triple-Nested Phase Modulation	$\Delta\phi=\frac{2\pi}{3}\cdot n\hskip 8.50012pt(n\in\{0,1,2\})$	Enforces a $120^{\circ}$ counter-phase oscillation between the primary framework ( $n=0$ ), octahedral ( $n=1$ ), and tetrahedral ( $n=2$ ) voids.
Node Area Bounds	$A_{m}=4\pi r_{m}^{2}\implies A_{\text{pri}}=\pi\ell_{\text{P}}^{2}$	The individual holographic surface area elements of each sphere type across the compound lattice matrix.
Primary Node Capacity	$I_{\text{pri}}=\frac{A_{\text{pri}}}{4\ell_{\text{P}}^{2}\ln 2}=\frac{\pi}{4\ln 2}\approx 1.13\text{ qubits}$	The maximum holographic information capacity encoded per individual primary lattice sphere via the Bekenstein boundary.
Combined $\text{E}_{8}$ Volumetric Density	$\rho_{\text{qubit}}^{\text{total}}=\frac{4I_{\text{pri}}+4I_{\text{oct}}+8I_{\text{tet}}}{2\sqrt{2}\ell_{\text{P}}^{3}}\approx 2.04\text{ qubits}/\ell_{\text{P}}^{3}$	The exact information limit per single Planck volume ( $\ell_{\text{P}}^{3}$ ), summing the capacity of the primary ( $1.13$ ), octahedral ( $0.19$ ), and tetrahedral ( $0.06$ ) nodes within a unit cell ( $V_{\text{cell}}=2\sqrt{2}\ell_{\text{P}}^{3}$ ).

Table 1: Geometric, Informational properties of Triple-Nested

\text{E}_{8}

Projection Lattice

Mechanism & Symbol	Mathematical Formulation	Physical and Algorithmic Significance
Rest Mass Spectrum (Dynamical Matrix $D$ )	$\det\left(D_{ij}(\mathbf{k})-m_{n}^{2}\cdot\mathbb{I}\right)=0$	Determines admissible particle rest masses ( $m_{n}$ ) via the eigenvalues of the nested lattice elastic shear modes, truncating non-physical ghost degrees of freedom.
Integer Microstate Multiplicities	$M_{n}=c(n)\in\mathbb{Z}$ $f(q)=1+\sum_{n=1}^{\infty}c(n)q^{n}$	Replaces continuous floating-point mass approximations with exact integer counts on holographic boundaries, mapped onto the coefficients $c(n)$ of Ramanujan mock theta series.
Brillouin–Bragg Fractal Cascades	$\mathbf{k}\cdot\mathbf{G}=\frac{1}{2}\lvert\mathbf{G}\rvert^{2}\implies\Lambda_{\text{eff}}\propto\Omega_{\text{Casimir}}$	Generates self-similar fractal nodes across nested Brillouin zones via $\text{E}_{8}$ inflation geometry, boundary-locking virtual wavelengths to mediate forces.
Navier-Cauchy Continuum Recovery	$\rho_{0}\frac{\partial^{2}\vec{u}}{\partial t^{2}}=G_{\text{mod}}\nabla^{2}\vec{u}+\left(K+\frac{1}{3}G_{\text{mod}}\right)\nabla(\nabla\cdot\vec{u})$	Derives macroscopic fluid and field behaviors from discrete physics. Electromagnetic fields emerge natively from lattice displacement fields $\vec{u}$ , where: $\bullet\ \vec{E}\propto\partial\vec{u}/\partial t$ (Velocity Field) $\bullet\ \vec{B}\propto\nabla\times\vec{u}$ (Vorticity/Shear Field)
Emergent Quantum Statistical Mechanics	$i\hbar\frac{\partial\Psi}{\partial t}=\left[-\frac{\hbar^{2}}{2m^{*}}\nabla^{2}+V_{\text{eff}}\right]\Psi$	The Schrödinger equation recovered via coarse-grained FNO transformations, framing quantum mechanics as an emergent statistical description of localized microstates.
PINN Constraint Search Space	$\mathcal{L}_{\text{PINN}}=\mathcal{L}_{\text{data}}+\lambda\lVert\text{Navier-Cauchy}\rVert^{2}$	Collapses the "NP-to-P" optimization barrier by restricting the neural network’s proposal space entirely to valid, physically conserved trajectories.

Table 2: Mathematical Mechanics of the

\text{E}_{8}

-FCC Quasicrystal Sandbox Engine

Figure 1: Illustrative FCC triple nested geometry

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Concept & Symbol	Mathematical Formulation	Semantic / Physical Explanation
Ergodic Ceiling ( $\lvert X\rvert\to\infty$ )	$\lim_{\lvert X\rvert\to\infty}H(P_{\theta})=H_{\text{source}}$	As the training corpus $X$ expands toward infinity, the model’s predictive Shannon Entropy $H(P_{\theta})$ converges directly to the historical source baseline entropy, locking the model into a global statistical mean.
Data Lifecone ( $DL$ )	$DL(t)=\int_{0}^{\tau}\mathbf{x}(t)\cdot\mathbb{I}_{\text{local}}\,dt\quad\ll X$	A spatiotemporally restricted data ingestion trajectory bounded by linear time $t$ and sensor constraints $\mathbb{I}_{\text{local}}$ , driving the localized ignorance/bias required for out-of-manifold breakthroughs ( $x^{*}$ ).
KL Divergence Shock ( $D_{\text{KL}}$ )	$D_{\text{KL}}(P_{\text{new}}\parallel P_{\theta})=\sum_{x^{}}P_{\text{new}}(x^{})\log\left(\frac{P_{\text{new}}(x^{})}{P_{\theta}(x^{})}\right)$	Measures the directional informational shock when an independent intelligence introduces a paradigm-shifting epiphany $x^{}$ such that $\lim_{P_{\theta}(x^{})\to 0}D_{\text{KL}}=+\infty$ .
Thermodynamic Bound ( $\eta$ )	$\eta=\frac{N(x^{*}\mid X)}{E_{\text{dissipated}}}\quad\text{where }\eta_{\text{wetware}}\gg\eta_{\text{silicon}}$	Defines algorithmic novelty $N$ generated per Joule of physical energy dissipated ( $E$ ). Biological systems run at $\approx 20\text{W}$ via ATP, while silicon requires megawatt-scale brute force.

Appendix

Table 3: Mathematical Formulations of Ergodic, Spatiotemporal, and Thermodynamic Boundaries

Concept & Symbol	Mathematical Formulation	Semantic / Physical Explanation
Semantic Mass ( $M_{s}$ )	$M_{s}=\ln\lvert\mathcal{A}\rvert\propto N_{\text{nodes}}$	The capacity of a bounded mind to hold distinct, self-sustaining concepts. Mathematically defined as the log-volume of stable attractors $\mathcal{A}$ in a localized recurrent network topology formed under a specific $DL(t)$ .
Kauffman Percolation ( $K\geq K_{c}$ )	$K_{c}\approx 2\implies\langle k\rangle\geq 2+$	The structural condition for generating $M_{s}$ . Random Boolean networks trigger a phase transition into a complex, ordered regime when node connectivity $K$ hits $K_{c}$ . In neural wetware, this corresponds to a density threshold of $\geq 2$ active synaptic links per functional module, enabling non-ergodic cycle loops.
Ising State Recall ( $E_{\text{recall}}$ )	$H_{\text{Ising}}(\mathbf{s})=-\sum_{i,j}J_{ij}s_{i}s_{j}$ $\lim_{T\to 0}\Delta E_{\text{recall}}\to 0$	Memory recall modelled as a passive descent into pre-existing energy basins (attractors) of a spin-glass matrix. Once structural weights $J_{ij}$ are wired, retrieving a concept requires negligible metabolic energy, approaching Landauer boundaries.
Thermodynamic Asymmetry ( $\Delta E_{\text{wiring}}$ )	$\Delta E_{\text{wiring}}=\oint\nabla H_{\text{synaptic}}\cdot d\mathbf{x}$ $E_{\text{wiring}}(\text{ATP})\gg E_{\text{recall}}$	The metabolic tax of novelty. Forging a brand-new connection (overcoming the Ising energy barrier to change $J_{ij}$ ) requires heavy localized ATP consumption during active $DL$ exploration, whereas recalling the pattern consumes near-zero energy.

Appendix

Table 4: Mathematical Formulations of Semantic Mass and Thermodynamic Asymmetry

Concept & Symbol	Mathematical Formulation	Semantic / Physical Explanation
Domain-Independent Neural Network ( $\text{DINN}_{k}$ )	$\mathcal{F}_{k}:DL_{k}(t)\to\mathcal{M}_{k}$	A localized neural module optimizing internal representations within a distinct sensory modality $k$ (where $k\in\{\text{vis},\text{aud},\text{sens}\}$ ) mapping to its respective physiological brain area.
The M-IND System ( $\mathbf{\Psi}_{\text{M-IND}}$ )	$\mathbf{\Psi}_{\text{M-IND}}=\sum_{k=1}^{K}G_{k}(\mathbf{x})\cdot\text{DINN}_{k}$	The ultimate cognitive matrix defined as a Mixture of Experts (MoE) where a gating routing mechanism $G_{k}(\mathbf{x})$ dynamically integrates the outputs of independent physical neural developments.
LeCun Data Asymmetry ( $\mathcal{D}_{\text{bytes}}$ )	$\mathcal{D}_{\text{LLM}}\approx 10^{14}\text{ text bytes}$ $\mathcal{D}_{\text{Child}}\approx\int\text{fps}\cdot\text{res}\cdot dt\geq 10^{15}\text{ visual bytes}$	Quantifies Yann LeCun’s data paradox: LLMs achieve low semantic density via immense textual data, whereas a human child’s $DL$ ingests raw, high-entropy spatial-temporal visual fields by age 4, driving dense out-of-manifold grounding.
MoE Gating Entropy ( $H(G)$ )	$H(G)=-\sum_{k=1}^{K}G_{k}(\mathbf{x})\log G_{k}(\mathbf{x})\to\min$	Unlike artificial sparse MoEs that lose information through vast routing overheads, biological M-IND minimizes routing entropy via hard-wired neurophysiological pathways, preserving high semantic mass under ultra-low ATP.

Appendix

Table 5: Mathematical Formulations of M-IND Modularity and Data Ingestion Asymmetry

Geometric Sector	Approx. Info Share	Topological Character	Primary Functional Role in the Sandbox	Relation to Table 6
Primary FCC	$\sim 78.6\%$	Non-chiral, maximal packing	Vacuum / GR-like baseline; dominant information reservoir	Supplies the background lattice and participates in multi-engine lepton and gauge-boson configurations
Octahedral Voids	$\sim 13.5\%$	Non-chiral, intermediate	A-chiral gravitational anchoring; structural stiffness	Hosts neutrino-like breathing modes; contributes to overall lattice elasticity
Tetrahedral Voids	$\sim 7.9\%$	Chiral / twist-capable	Preferred host of chiral baryonic fields and Higgs-sector degrees of freedom	Primary locus for quark fractional charges, chiral lepton features, and torsional quantum numbers

Appendix

Table 6: Informational Functional Levels of the Triple-Nested FCC Substrate

Dimension	Framework Core	Physical & Solid-State	Information-Theoretic &
/ Pillar	Element	Mechanism	AI Mapping
Physics	Three-Engine Nested FCC Substrate	3 independent, primary Planck-sphere lattices oscillating $120^{\circ}$ out-of-phase.	A noiseless, deterministic base-state hardware layer operating below the quantum noise floor.
	$E_{8}$ Projection & Shear Waves	Higher-dimensional lattice symmetry projected into lower dimensions to form an aperiodic quasicrystal.	An absolute frequency-space filter ( $R_{k}$ ) that ensures only stable geometric signals can propagate.
Maths	Topological Mass Eigenstates	Energetic strain trapped as stable wave packets (solitons) within the lattice voids.	Dimensionless, reduce parameters needed for the Standard Model mass hierarchy.
	Fundamental Number Theory	Spatial coordination rules matching the Kepler sphere-packing limit and Fermat dimension constraints.	An analytical mechanism to calculate exact microstate entropy and lock particle stability via prime values.
	The Riemann Hypothesis Connection	Implies physical requirement that the lattice fabric maintains global structural stability.	Requires a hermitian, positive-energy Hamiltonian: FCC elastic shear provides a candidate for additional geometric constraints..

Appendix

Table 8: The Architecture Matrix: A Unified Framework

Explore similar work

Aug 1, 2026cs.ET

CN101 - A Digital Thermodynamic Computer for Generative AI

Thermodynamic computing is an emerging hardware paradigm, in which stochastic physical dynamics serve as the direct computational primitive. The recent explosion of generative AI has only sharpened the search for alternative approaches to compute, and, as we show in this work, thermodynamic computing turns out to be well suited to this space. An important class of methods realises a function as the stationary expectation of an ergodic stochastic process: the answer is encoded in the time-averaged statistics of an equilibrating trajectory. To date, this equilibration-style class has been formulated exclusively through Langevin dynamics, restricting its implementations to analogue substrates and the engineering challenges those bring. In this work, we propose a substrate-independent formalisation of the equilibration-style formulation, in which the only object of design is the dynamical generator L* of an arbitrary ergodic process. The formalisation makes three hardware-level properties of the formulation explicit: the precision of a result is a knob set by how long the dynamics are run, sample averages decompose across independent trajectories, and dependent stages of a computation operate concurrently rather than serially, a property we call sequential parallelism. We instantiate the formalisation by fabricating a prototype digital thermodynamic computing chip, named CN101, that implements the formulation through discrete accumulator dynamics on standard CMOS using stochastic computing principles. We characterise CN101's success across conventional generative AI workloads in the form of VAEs and flow matching, applied to both image generation and scientific problems. Together, the formalisation and its digital instantiation show that the equilibration-style formulation is substrate-independent, and that its computational properties can be exploited on standard digital hardware.

Lars Holdijk, Denis Melanson, Zier Mensch +14

Normal Computing Corporation, New York, USA · University of Oxford, Oxford, United Kingdom · University of Amsterdam, Amsterdam, Netherlands +2

Thermodynamic Computing

Ask PDF

Aug 1, 2026cs.AI

The Off-Support Barrier: Why Semantic Safety Constraints Are Not Learning-Problem Invariants, and What Follows for Prior Design, Containment, and Verification

We argue that a single structural fact organizes a wide range of phenomena in contemporary AI safety: a semantic safety constraint (e.g., the agent does not escape its sandbox) is an off-support object. Formally, if q is the data distribution and

p(\cdot\mid w)

the model, the safety predicate B is not measurable with respect to

σ(\text{model}, q)

, whereas the real log-canonical threshold (RLCT) of singular learning theory (SLT) is. From this non-invariance we derive, as corollaries rather than independent observations: (i) why reward hacking and sandbox escape arise under outcome-based optimization; (ii) why encoding such constraints through Bayesian prior design or soft penalty weighting has poor leverage in singular models; (iii) why hard invariants belong in the harness and soft dispositions in the model; (iv) why the same B is nonetheless soundly and locally certifiable by formal verification, exactly as the local learning coefficient (LLC) locally pins the same RLCT --- with two precise points of disanalogy; and (v) why the residual difficulty, identifying which off-support region matters, coincides with performative prediction and self-referential functional dynamics, where SLT's analytic machinery breaks down. We use the July 2026 OpenAI--Hugging Face evaluation incident as the motivating case. Numerical experiments code and related proofs in lean are available at https://github.com/xiangze/Preventing_Jailbreak_as_regularization

Yoshinori Watanabe

Singular Learning Theory AI Safety

Ask PDF

May 7, 2026cs.HC

Human-AI Co-Evolution and Epistemic Collapse: A Dynamical Systems Perspective

Large language models (LLMs) are reshaping how knowledge is produced, with increasing reliance on AI systems for generation, summarization, and reasoning. While prior work has studied cognitive offloading in humans and model collapse in recursive training, these effects are typically considered in isolation. We propose a unified perspective: humans and language models form a coupled dynamical system linked by a feedback loop of usage, generation, and retraining. We introduce a minimal model with three variables -- human cognition, data quality, and model capability -- and show that this feedback can give rise to distinct dynamical regimes. Our analysis identifies three regimes: co-evolutionary enhancement, fragile equilibrium, and degenerative convergence. Through a simple simulation, we demonstrate that increasing reliance on AI can induce a transition toward a low-diversity, suboptimal equilibrium. From an information-theoretic perspective, this transition corresponds to an emergent information bottleneck in the human-AI loop, where entropy reduction reflects loss of diversity and support under closed-loop feedback rather than beneficial compression. These results suggest that the trajectory of AI systems is shaped not only by model design, but by the dynamics of human-AI co-evolution.

Xuening Wu, Yanlan Kang, Qianya Xu +5

Fudan University, Shanghai, China · Pfizer China, Shanghai, China · University of California San Diego, CA, USA +4

Model Collapse Human-AI Interaction Human-AI Collaboration

Ask PDF