TFA author cites many examples from human-oriented systems, which perhaps are more common and traditional domains for RDBMS design, where there often several layers of exceptions to the rules - his argument that synthetic keys work well for these domains lands.

I personally work far more with computer-oriented systems and their data, and natural keys work well for me. When well-chosen they allow me to do an initial load of the source data for analysis, and then aggregate such databases together later on for historical analysis without fear of conflict. The data are often immutable in these domains, too.