While I agree with you somewhat, I also find value in an iterative learning approach - one that, in theory, constantly improves over time. My problem with his

While I agree with you somewhat, I also find value in an iterative learning approach - one that, in theory, constantly improves over time. My problem with his approach, which I think you hint at, is that to properly improve you need to make sure you're optimizing the right metric(s), and those goals need to be clear. In Mark's case, I don't think that he has a clear goal he's trying to optimize all the time, so the different iterations on the system seem like they're swirling around a drain: slowly getting better and better, but sometimes just moving orthogonal to the most effective direction.