Researchers Build an 'Interpretable' AI That Shows How It Thinks - The New Stack