Prediction: Any AI problem that you can simulate and sample endlessly many training samples for can be solved with today's algorithms such as deep and reinforcement learning.

Playing games like chess, go or computer games where you have knowledge of all necessary inputs and can sample by self playing.

That is partly my point. The techniques sound exciting but may not be applicable to many hard real world problems.

Robot hand was mostly about meta-learning to adjust to different similar envs. Very hard to get the details right, so they made the real be just another sim (probably the most complex) and it learned to adapt. So, prepending “closely ” on “simulate” might improve your prediction.

I guess I should have clarified. Treated reality as just another sim. Just never encountered during training time.