In the past, I was very skeptical of NLP efforts. Thought it was mostly full of cool toys.

I was dead wrong.

I am similarly skeptical of RL, in the sense that for most cases you are better of using optimal control techniques, and maybe sometimes a combination of RL and optimal control.

I am aware of AlphaZero and other impressive achievements in certain games. However, I am still left with the feeling that it is very expensive to train an RL model and it is insanely specific to the task at hand.

Are there recent breakthroughs that point at promising generalization in the RL community?