“Almost always the second best algorithm for any shallow ML task”