Reverse Philosophical Algorithmics: A Falsifiable Model of Recursive Decline from Ibn Khaldun to Model Collapse

This article introduces the reading method of Reverse Philosophical Algorithmics (RPA) that employs the cyclical theory of civilization developed by Ibn Khaldun in the Muqaddimah to read the practical reasoning of contemporary artificial intelligences. Typically, criticism of AI proceeds in the reverse direction, namely, from a given philosophical framework, ethical or otherwise, to the instance of the technology to which it is applied. RPA turns this sequence around and starts from a set of structural peculiarities intrinsic to machine learning, such as attention mechanisms, optimization of loss-functions, and the phenomenon of model collapse, to posit conceptual parallels with Khaldun’s notions of asabiyyah, cyclical time, and a labor-based understanding of value that ground his theory of urban life (umran) and its historical vicissitudes. Large language models learn from the textual content of human asabiyyah, making their internal dynamics a kind of a reenactment of the process of civilization development described by Khaldun. The article formally identifies conditions for the application of the Khaldunian framework to study the collapse of AI-driven societies of meaning. By analyzing the disjuncture between endogenous genesis and exogenous optimization in Khaldun’s theories and attention mechanisms and loss-function optimization in large language models, this article demonstrates how algorithmic instabilities known as model collapse formalize Khaldun’s intuition of decadence. It concludes by suggesting governance principles for building open asabiyyah based on the dynamics of Khaldun’s cycle, thus providing a set of falsifiable hypotheses for testing their applicability to AI-civilization without appealing to the philosophical AI governance literature. RPA is presented as a generalizable method to develop homologies between theories of decline by reading their purported manifestations in the design of learning algorithms. The application to Khaldun’s historical sociology is therefore posited as the beginning of the analysis, not its end.