Researchers have developed Mechanist, an AI system designed to use machine learning models as scientific instruments for discovering the underlying mechanisms of intelligence. To achieve this, Mechanist constructs a vast knowledge graph and integrates it with a multidisciplinary database of papers across various fields. The system generates more accurate mechanism hypotheses and executes experiments more reliably than existing AI-scientist systems. Mechanist also demonstrates its ability to uncover safety risks in scientific laboratories and develop practical interventions to improve model performance. --- Why it matters: This matters because understanding the mechanisms underlying AI intelligence is crucial for developing trustworthy and controllable models. Mechanist's ability to autonomously discover these mechanisms could lead to significant advancements in fields like natural language processing, computer vision, and decision-making. Source: https://arxiv.org/abs/2608.12036