In order to perform "machine learning", we must first collect a large dataset and teach the machine to distinguish that data, learning what makes for the right answer in the process, based on our guidance. We teach the machine to memorize and to pattern-match. Humans don't learn like that. Instead, we learn to ask the right questions about the data, and use those to make choices.

If we could skip the costly dataset training and instead use the labeling directives to teach the machines to understand, it could save a lot of money and make machine learning more efficient.