This repository contains the code, computational simulations, and data supporting the research paper:
“Non-Convergent Intelligence in Unbounded Knowledge Spaces: Entropy Dynamics, Learning Incompleteness, and Implications for AI Alignment.”
Published as a preprint on Zenodo: https://doi.org/10.5281/zenodo.18143313
Artificial Intelligence (AI) research typically assumes that learning converges to stable representations or optimal solutions.
This work develops a theoretical and computational framework for non-convergent intelligence, where:
- Knowledge Entropy quantifies representation disorder,
- Dynamic Knowledge Entropy captures entropy drift over time,
- Dynamic Intelligence Ceilings constrain finite learners,
- Learning Incompleteness formalizes inherent limits on complete understanding.
The Principle of Non-Convergent Intelligence (PNCI) is introduced as a foundational result with implications for:
- Machine Learning Theory,
- Continual/Open-Ended Learning,
- AI Alignment and Safety.
This repository includes:
- Simulation code (Python) implementing entropy dynamics,
- Datasets and preprocessing scripts,
- Reproducible experiments for continual learning and open-ended tasks,
- Plots and statistics used in the paper,
- Notebooks illustrating key concepts.