It's an interesting idea though, as sophomore as the attempt was. I'd be curious to see images lossy compressed via a neural net. Do you know of any better attempts than this?
Can't find the startup's name though.
I believe it was discussed here a few years back.