Basically means that machine learning models can be asked to perform tasks and then use external tools to try and accomplish these tasks. The feedback on whether the task was accomplished successfully is then used to tune the model so it performs better in the future.

This article from Google shows how this tech is being applied to make robots that can interact with the real world in a meaningful way.

It means chatbots have gotten better at fooling credulous nerds into thinking they’re HAL 9000.