On variability

An exploration of variability in complex systems, drawing parallels between safety engineering's concept of 'normal variability' in human performance and the non-determinism inherent in AI coding tools. In safety science, the old view treats variability as dangerous while the new view sees it as adaptive capacity. LLMs are inherently non-deterministic due to temperature-based sampling, and this randomness was found to improve output quality. The author argues that variability is not a bug to be eliminated from AI tools but an essential ingredient for performing complex cognitive tasks, and predicts LLMs will never become fully deterministic.