In there is a note that Wolfram is working on, or at least thinking about, the big unsolved problem - useful robots. Robot manipulation in unstructured situations is still very poor. But maybe someone will figure out a way to apply newer machine learning techniques to that. Google had a research group working on that, but they haven't been heard from in years.

The approach of the large language model, where you have a huge training set of general purpose info and a small prompt for the current task, might possibly work.