RL seems to be in this weird middle ground right now where nobody knows how to make it work all that well but almost everybody at the top levels of ML research agrees it's a vital component of further advances in AI.