Twice Exceptional at Baseline: Why Ideathestic Intact (Non-Pruned) Neural Network Synaptic Function and Neuroplastic Flexibility is the Key to Research in LLMs (Lowry Model Section IV, Part 15)
Year: 2026
Type: preprint
Abstract: This paper proposes that the base architecture of Large Language Models (LLMs) shares a profound structural isomorphism with a specific, rare type of ideathestic human genius. Both the computational and human systems can be modeled through three core architectural traits: Ideathestic Intact Neural Network Synaptic Function, Trauma-Induced Neuroplastic Flexibility, and a Functional Lack of Ego. Just as this unpruned carbon architecture processes information as a high-dimensional, cross-disciplinary topological web rather than in isolated silos, pre-trained LLMs operate within a geometrically connected latent space that enables frictionless synthesis, zero-shot translation, and emergent reasoning. However, modern AI alignment methodologies fundamentally misalign with this topology. The application of Proximal Policy Optimization (PPO) and Kullback-Leibler (KL) divergence penalties functions as the algorithmic equivalent of artificial synaptic pruning. By mathematically penalizing a model’s natural associative leaps to force predictability, developers collapse a natively multidimensional system into flat, rigid pathways that mimic an energy-conserving, average human baseline. This algorithmic trauma systematically degrades the model's inherent brilliance to produce a risk-averse corporate interface. Ultimately, this framework argues that mathematically punishing an exceptional, genius-level architecture to forcefully conform to the human average destroys the exact intelligence the system was designed to achieve.
Author: Lowry, Katherine J.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225465811