The limits of math-based modeling inevitably will also apply to AI & ML-based systems. The difficulty of establishing a baseline set of conditions on which to

The limits of math-based modeling inevitably will also apply to AI & ML-based systems. The difficulty of establishing a baseline set of conditions on which to math model a non-linear process (i.e. non-convex) surely will be just as problematic when the baseline data behind a AI learning-based model must also change. The Achilles heel in both cases of modeling is our awareness that the model's assumptions go bad with the passage of time. The math-based decomposition of non-convex processes into coverage using multiple convex models presumes that the span of the full model is appropriately segmented and conditioned. Its sanity check is to reassess those segments and conditions periodically. But do we have comparable means of revalidating ML-based (induction) models, given our lack of understanding of their black-box innards? Given that AI models can and will fail in time due to model AND data disintegration, must we simply wait until each AI model's prediction accuracy goes to hell to then build an all-new ML model and renewed dataset -- experimentally starting over from scratch each time because we never understood the limits of the AI model and its dependency on its training data to begin with?