Since machine learning tools have been employed to build numerical potentials for atomic simulations, this field has known an increase interest, from both fundamental and applied points of view. The two ingredients involved in these potentials are the descriptors (or fingerprint) of atomic configurations and the regression method, each ingredient spanning different level of complexity (from symmetry function to bispectrum, for linear regression to deep neural network). A bird'eye of these methods will be given, together with some applied examples and burning questions.