Deep learning has a fractal-basin limit to predictability

University of Sydney researchers reported in Nature Machine Intelligence that fractal basin geometry can explain why neural networks reach different outcomes after tiny disturbances, even in fully deterministic settings. The parameter regions that enabled the strongest learning were also harder to predict and reproduce; lowering the learning rate improved repeatability but reduced learning effectiveness. The team introduced tools for measuring this sensitivity and points to architectural redesign as a future way to reduce unstable conditions. The result does not mean AI is inevitably uncontrollable; it means reliable evaluation must report distributions across runs and the cost of reproducibility, not only a single best score.