[motohagiography]

In the context of ML, there is a great blog post from years ago that compares the ROC curve to an indifference curve in economics, which describes how without a sense of what the costs and conseqeunces of the precision in the ROC curve are, you can't really optimize a model well. It's a question of internal vs. external consistency.

There is a heuristic I learned related to a bunch of others, but it's being able to do a fast assessment of the returns on precision. I talk about it to clients in conceptual terms of doing depth-first searches vs. breadth-first. Technologists tend to be depth first and very deep in a relatively narrow context, where dynamics people tend to be breadth first across a wider set of contexts.

The penny pinching metaphor the author uses reduces to an admonition to look at the "bigger picture." Personally I dislike that trope, but a general set of cognitive tools for depth-first thinkers to pop up conceptual levels, particularly into ones where they can apply their skills to navigate the uncertainties of breadth-first thinking with its dynamics of multiple simultaneous equivalents, it would be really useful to summarize and teach.

There is the issue where the details of precision in fact have nonlinear effects and they are worth drilling down into, but that's part of the analysis.