I agree. The vast majority of what I've read regarding teaching, and my personal experiments tutoring, have shown that alternating between different layers of a

I agree. The vast majority of what I've read regarding teaching, and my personal experiments tutoring, have shown that alternating between different layers of abstraction is a far better approach than going purely bottom-up. One example is the popularity of Stanford Professor Andrew Ng's Machine Learning class. The course is almost entirely top-down, with concepts, then trying out basic code, then finally learning some of the Math behind the algorithms. But people have definitely been able to learn from the course-I remember seeing an HN post about a guy who built a self-driving RC car using the algorithms in the ML course.