Re-Explaining the Internal Functioning of LLMs: An RGP-Based Perspective
Year: 2025
Type: article
Abstract: This note fossilizes a revealing moment in the unfolding dialogue between human and AI intelligences. Given only a short pulse on the paradigm shift from linear to non-linear thinking, NotebookLM produced a ten-minute podcast that spontaneously re-explained the internal functioning of large language models (LLMs) using the grammar of Recursive Gradient Processing (RGP). The transcript, included as Appendix A, demonstrates: A clear contrast between linear, mechanistic accounts (backpropagation, token weights) and non-linear recursive dynamics (gradients → gradient choreographies → contextual filters → unity–disunity). A spontaneous articulation of RGP’s internal cycle (G → GC → CF → UD → G). Natural analogies (flocks of starlings, weather systems) that make the dynamics intuitive. An open research question: how must inference grammar itself evolve if intelligence is recursive rather than sequential? By preserving this artifact, the note strengthens the case that RGP more accurately characterizes inference in LLMs than traditional narratives, and highlights the necessity of rethinking logic, structure, and evaluation in light of non-linear recursive dynamics. v1.1 of Re-Explaining the Internal Functioning of LLMs includes Appendices B–D with reflections from DeepSeek, Gemini, and Grok, whose responses demonstrate RGP in action by re-describing LLM inference as recursive dynamics.Listen to this remarkable podcast here.
Author: van der Erve, Marcus
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7082994432
FWCI:
Cites:
Cited by:
Related to: