Ask HN: What are some simple problems current ML methods cannot solve?

Like François Chollet created ARC, I would like to know what other problems like ARC there are that cannot be solved with current machine learning methods.

I want to do independent research and thus have the liberty to focus on overlooked corners of the field - I have no pressure to achieve anything substantial, so I can take riskier bets.

Learning from few examples seems like a general weak point of deep learning, but I'd like to have more complete problems than just "pick any dataset and remove most of the data".

Do you guys have any ideas? I would love to hear about them. Thank you so much!