Interesting paper by MIT and Google. You can easily get the gist from the introduction and results table.
tl;dr – by using AI to learn the structure of the data (is it dispersed/narrow, regular/random etc) they got AI to decide what indexing algorithm to use (with the idea that the AI could index parts of the data in different ways if it thought that was best). Results were 50%+ speed boost and ~90% reduction in index size in some cases! Apparently the GPU speed increases and dedicated Tensorflow chips make the AI part really quick to compute now.
Considering that AI can now become a world-champion chess player from scratch in a matter of hours, I would think that this is something that’s going to become quite relevant!