It's RL so that means it's going to be great on tasks they created for training but not so much on others.

Impressive but the problem with RL is that it requires knowledge of the future.