From LLM Inference Limits to the Physical Fast Network of Natural Intelligence: A Falsifiable Hypothesis and Research Program

Abstract

The rapid scaling of large language models (LLMs) and sparse mixture-of-experts (MoE) architectures reframes a foundational question about intelligence: must intelligent systems rely on large-scale, serial, and explicit computation, or does natural intelligence employ a faster, lower-level response mechanism than neuronal networks and explicit cognition? We introduce the Physical Fast Network Theory (PFNT) as a falsifiable, computational model. PFNT does not assume that consciousness is quantum, nor that a Planck-scale structure exists in the brain. Its central, testable claim is that biological intelligent systems may possess a hidden, fast, sparse, and dynamically routed information-processing network whose response time scale falls below what current macroscopic neuroimaging can directly resolve, with neurons and neural networks acting as an amplified, biological organization of that lower-level physical response. PFNT inherits ideas from brain-as-mixture-of-experts, dynamic neural routing, the freeenergy principle, neural Darwinism, and physical intelligence, but advances the question from how the brain computes to why the brain completes state selection on such a rapid physical time scale. We formulate a hierarchical model from fundamental physical interaction to hidden fast network, biological response, neurons, brain-MoE, intuition/emotion, and explicit consciousness, and propose to estimate the existence probability of such a network through early neural signals, latent-variable modeling, Bayesian model comparison, and accumulation of evidence. We explicitly state that there is currently no evidence of a Planck-scale computational network in the human brain, nor evidence that post-mortem brain tissue preserves usable quantum information. PFNT should therefore be regarded as a hypothesis and research program, not an established physical theory. The central question is not to search for quantum consciousness, but to determine whether a faster, dynamically sparse, behavior-predictive response network exists prior to conventional explicit neural computation. We emphasize the distinction between falsifiability and direct measurability. PFNT is falsifiable in the sense that specific predictions can be tested and potentially refuted, even though the hypothesized fast network is inherently not directly observable. The modelcomparison framework converts this into a well-posed question: whether a latent fast variable improves out-of-sample prediction over a classical baseline. The claim is therefore open to disconfirmation, and we give explicit falsification conditions in Section 7.3.