Why XGBoost Still Wins the Tabular Data Game

Looking at figure 20 of the blog, my feeling is that the response surface produced the decision tree is too jagged and overfit. Maybe I overestimate the importance of smoothness in the function fit?