A new preprint on arXiv, titled "Mining the Self-Rewriting System," argues that the foundation of classical data mining is crumbling. Traditional methods treat data as a fixed, external resource that the learner passively analyzes. But emerging AI-native infrastructure no longer fits that picture.
The paper describes a cycle in which AI agents observe system state, take actions, update retrieval stores and models, and promote new artifacts. In such a system, the data is not static—it is continuously rewritten by the very agents that are supposed to learn from it. This invalidates the classical separation between data and learner.
The authors call for a new vision of data mining that embraces this self-rewriting nature. Rather than assuming a stable dataset, mining must account for the dynamic, feedback-driven processes that generate and alter data. The abstract is brief, but the implication is clear: data mining on AI-native infrastructure will require fundamentally different tools and assumptions.