I'll focus just on one subject, phones: 1. Quality dumb phones. No touchscreen, just two functions: text and call. Make it matchbox-sized. Connect these to 3rd party texting / call apps like Skype and WhatsApp using a desktop, make the voice call quality superb. 2. A way to get premium voice quality. It's 2018, and the voice call does not sound much better than in 1998. 3. Modular phones. Google's Project Ara, that has gone the way of many other Google projects.

Most likely, the training set is made of videos marked as extremist by multiple reviewers.
"Better" means that the test set had a better precision and recall than when the humans processed it.
This is, of course, in theory. In practice, it is likely that the training / test sets were so similar that it will not scale in the real world.

I imagine they don't want the press to jump on that, but it would be interesting to read more.