How can I describe why the differences are like this? Is the GLM performing worse than the blackbox?
I guess your model after improvement from blackbox has reduced the number of large errors according to RMSE, but as the MAE suggests, the total error has increased.
Here are some quotes from that page:
Taking the square root of the average squared errors has some interesting implications for RMSE. Since the errors are squared before they are averaged, the RMSE gives a relatively high weight to large errors. This means the RMSE should be more useful when large errors are particularly undesirable.