I disagree about the underlying mechanism being complex. Machine learning algorithms are a class of equation-based systems and strategies for configuring these

I disagree about the underlying mechanism being complex. Machine learning algorithms are a class of equation-based systems and strategies for configuring these equations for a specific task. All our math and reasoning tools still apply to this class of algorithms, in principle. Where we pass into alchemy though is the interplay of these basic components with each other and the parameters they encounter while running, this is where complexity happens. Part of this lies in the very nature of the tasks we use them for: we basically push a cart of raw data in front of a set of "AI" solvers and expect them to do something with it. When that doesn't work, start over, tweak parameters, and try again. I agree that there is no sufficiently useful intellectual framework for creating these artificially intelligent components, and that shows not only in the uneven success rates and performance, but also in the surprising fact that experts in very different AI systems can usually create components with similar performance characteristics for a given problem, despite using very disparate strategies.