> It's a step forward for "interpretability," which is a field of AI that seeks to explain why neural networks create the outputs they do The problem has been

> It's a step forward for "interpretability," which is a field of AI that seeks to explain why neural networks create the outputs they do The problem has been typically called 'transparency', but ok. (Now trying to find the references to said convention... There is a chance that the use of 'transparency' was promoted from the applications of AI in Decision Support Systems - where the background is that a decision is "transparent" when its originating processes are explicit. But also, "transparent" is opposite of "opaque", and we have been dealing with "black boxes" encoded in the networks.)