Alpha zero

On normal hardware? Maybe around 1900-ish. It's certainly not a proper chess program and isn't designed to be run on normal hardware. 

Putting it on a gigantic supercomputer vs Stockfish on "potato" specification hardware proves nothing and is a simple and ridiculous PR exercise for google. They even took away Stockfish's opening book and I assume all tablebases etc., while they were salivating over how AlphaGo spent 10s of hours coming up with its own opening book and material it was allowed to use. 

We have to wait until AZ plays some games in a fair match.😐

"On normal hardware? Maybe around 1900-ish. It's certainly not a proper chess program and isn't designed to be run on normal hardware."
It's actualy not true. Even AlphaGo Zero (that is more hardware performance-dependent), run on 4TPUs.. It's unable to tell, how much cost one TPU (its Googles own piece of hardware, that is not available for buy ).. But performance is about as same as 4 GPU system (with lower energy cost). It's not so abnormal in world of AI. 

Here is explained how it started learning itself with more resources and then just optimized to a decent level of hardware spec.

https://www.youtube.com/watch?v=9xlSy9F5WtE

It scored 64% against a near 3400 engine.

So its performance rating puts it near 3500.

If Tal, Karpov and Stockfish had a kid he would be named Alpha zero.

Although the item below is GO and not chess, I'm fascinated with the growth of Leela Zero's rating (if we can believe it). This is an opensource project by a group of "average Joe" volunteers. (See the first post on this Talkchess page and the rating graph in the second link.)

http://talkchess.com/forum/viewtopic.php?t=65481&start=40

http://zero.sjeng.org/

I'm just waiting to see if anyone will try a similar project for the chess side of Alpha Zero.

No, the stripped down version of Stockfish that AZ played against was not 3400. We should put Stockfish under the same conditions that they used and play it against regular Stockfish and other engines to see what level of chess AZ played against, and to see if regular Stockfish would have scored just as well as AZ against the stripped down Stockfish.

I agree, it was not 3400.

That's why I said it was near 3400

Regardless of the ratings--the 10 games published were among the best and most beautiful I have seen in my life! [i am age 76]

Well I don't know how stockfish was treated in regards to the opening book - but I know that Alpha Zero only calculated about 1/4 the calculations that stockfish did - so I'm not sure that computer power makes the entire difference. In regards to proving that Alpha Zero is more efficient I think Google proved that sheer calculation speed is not the most important factor when playing chess. Intuition is important too and the reason Alpha Zero isn't doing as many calculations is because it has learned not to calculate the most unlikely moves, which makes the amount of calculations irrelevant in compare to deciding what to use the calculation for. 

Alpha Zero's evaluation is derived from patterns that are identified and distilled by itself (using its neural network), so it can become much more precise than the material-based heuristics handcrafted by humans. Alpha's thinking might only involve material as a consequence of more basic patterns.