Social media tells us that data science, machine learning (ML), and artificial intelligence (AI) are big. The author of this article, whose title evokes a well-known self-help book [1], has made important contributions to these areas. But when I saw that the article included a link to a slick video, I was afraid of more hype.

I was almost wrong. While the article appears to be excerpted from Pearl’s recent book [2], it is nonetheless an accessible and thought-provoking review of the state of the art in ML, with indications of where progress can be made.

Pearl actually describes nine tools. His first is the “three-level causal hierarchy,“ which is reminiscent of Bloom’s taxonomy [3], a tool widely used in education, with the big difference being his inclusion of counterfactuals at the top level. He does not see much ability to handle counterfactuals in ML systems, but notes that they are an important reasoning technique for people. He makes the puzzling claim that “only the top echelon of the scientific community can ... formally distinguish ‘mud causes rain’ from ‘rain causes mud.’” The hierarchy of scientists that would imply a “top echelon” is left unclear, but no practicing mathematician would struggle with this task.

The second tool is a structural causal model inference engine; he offers few details here.

Then come the seven good habits. The first can be characterized as constructing a good null hypothesis. The second is an unspecified “backdoor” to deconfounding. Others sound like nonparametric statistics, but there is not enough detail to determine to what extent this is true.

Pearl borrows good ideas from education, philosophy, statistics, and computer science (CS). The real contribution here is the emphasis on computing and computability. It is disappointing that he feels that people in ML need to be reminded about the basics of reasoning.

This article is a good overview, but someone inspired by it would be well advised to do two things: take a rigorous mathematics or CS course that uses first-order logic extensively, and read the technical works of Pearl and others in the bibliography.