This way one could chain calls, à la sklearn `model().fit(X, y)`: `compiled_clone(Quadratic()).initialize(X, y)` Clearly not a priority but it may be a nice synctactic sugar. Thoughts @Badr-MOUFAD ?