Perturbing LLM Attractors, Intentionally: The Thermodynamics of Human-AI Interaction

This working essay proposes a thermodynamic framework for understanding human-AI interaction, arguing that iterative prompting of large language models (LLMs) parallels energy-driven phase transitions in physical systems. Just as sustained energy input causes systems like Rayleigh-Bénard convection cells to reorganize into increasingly sophisticated patterns, continuous human intention—expressed through repeated prompting—drives LLMs toward deeper organizational states of knowledge. Using the example of repeatedly prompting with "even more" to achieve philosophical depth, I demonstrate how LLMs contain vast latent knowledge structures that only manifest through sustained perturbation. This approach aligns with emerging paradigms like Nielsen's "Intent-Based Outcome Specification" and Karpathy's "vibe coding," where human intention shapes AI outputs rather than commanding specific operations. I argue for recognizing LLMs as dynamic state-spaces whose responses are relative to human aesthetics and purposes, requiring new interpretive habits that treat AI interaction as a continuous loop of projection and interpretation rather than single-query exchanges. This essay was originally published as a Substack post and presents an informal theoretical framework for researchers and practitioners interested in human-AI interaction, prompt engineering, and the philosophical implications of large language models. https://pourdavood.substack.com/p/perturbing-llm-attractors-intentionally.