How Alphazero defeated Stockfish with much less computational power and 0 training

The 10 games that are published as part of the paper are at the bottom of the page. It's like watching an alien playing chess, it seems so ... foreign.

I've only gone through a few of the games, but on the 3rd one I was wondering things like "surely AG wins this piece, or a few pawns" yet it chose not to take. Obviously knowing that SF is ELO 3200+, makes one concede that it might be a poisoned carrot, but for a program that was only fed rules to be able to decide that, is crazy.

It makes for very entertaining chess, and I think the wonderful people who work on tuning SF and other engines will have a lot to think about.

What's the highest theoretic ELO rating that a computer can get?

A few people mentioned that it'd be interesting to see how AG performs on a home computer. Maybe that'll be the differentiating factor. AG that's handicapped by input resources.

Lastly, does AG constantly learn as it plays? i.e. once a chess model is created, does it get updated with new info on the fly, or would it require more training?

"0 training", isn't that incorrect? The article doesn't mention that. Playing against oneself is still training.