Indeterminism in Large Language Models: An Unintentional Step Toward Open-Ended Intelligenceu

The paper challenges the prevailing paradigm that views Large Language Models (LLMs) as deterministic computational systems where unpredictability is an error. It proposes a new paradigm, drawn from artificial life and enactive cognitive science, which reframes the intrinsic, emergent indeterminism in LLMs not as a flaw, but as a foundational feature. This new perspective considers properties like irrepeatability ("mortality") and open-endedness as essential for developing more lifelike and genuinely intelligent artificial systems, shifting the goal from perfect control to embracing constitutive unpredictability.

This work identifies a critical anomaly—the persistent indeterminism in LLMs despite efforts to ensure deterministic outputs. Instead of treating this as a puzzle to be solved within the current framework (Normal Science), the paper elevates it to a central crisis point. It proposes a revolutionary conceptual shift by arguing this indeterminism is not a bug but a feature, linking it to lifelike properties like "mortality" and open-endedness. By proposing a new test for "lifelike complexity" and advocating for a reunion with artificial life theory, the paper actively constructs a new paradigm for understanding and designing future AI.