We argue that pre-trained Transformers are best modeled not as learning machines, but as frozen spin glasses. We propose that "intelligence" in large language models (LLMs) is not a sequential computational process, but a critical phenomenon occurring within a static, high-dimensional topology sculpted during training. In this work, we reinterpret in-context learning as a trajectory through a potential energy landscape governed by an effective Hamiltonian. The prompt acts as an external symmetry-breaking field, tilting this landscape so that previously latent minima become global attractors. We further introduce an effective Helmholtz decomposition of the inference dynamics, separating a conservative gradient force that pulls the model toward memorized facts, from a non-conservative solenoidal force driven by asymmetric attention that enables multi-step reasoning and analogical "orbits" in representation space. Finally, we propose a thermodynamic observable, Semantic Susceptibility ₒ₄₌, which spikes at the onset of hallucination and functions as an early warning indicator of phase transitions in semantic behavior. The Frozen Manifold: Re-frames LLM inference away from active computation and toward structured relaxation within a disordered energy landscape. Helmholtz Decomposition of Cognition: Formally separates model behavior into a conservative gradient flow (F₆ₑ₀₃) governing literal memorization, and a divergence-free solenoidal flow (Fₒ₎₋) governing creative synthesis and cyclic reasoning. Semantic Phase Diagram: Maps the boundaries between Logic Mode (Type I), Creative Mode (Type II), and Spin-Liquid/Hallucination Mode based on prompt constraint strength and entropy. The Melting Point Test: Demonstrates semantic hysteresis, proving that LLMs possess "supercooled" states where they remain trapped in hallucination-prone metastable basins even after context noise is reduced.
Nicolas walejewski (Fri,) studied this question.