Just to clarify this point. As I said above, the problem in the exercise M6.03 is not overfitting but underfitting: the training error increases with more trees… which is counter-intuitive and we are still not sure if this is a bug in scikit-learn or not.
Yes maybe there is a bug 
Changing the exercise notebook is tracked in Edit Exercise M6.03 on Adaboost overfitting · Issue #456 · INRIA/scikit-learn-mooc · GitHub
Removing the quiz is the simplest thing to do right now and is done in Sign in · GitLab