Google Deepminds Alpha Zero beats Stockfish

I have not seen another thread discussing the new Deep Mind chess playing program outperforming Stockfish. Considering that it is a totally different AI architecture i would like to hear what you think.
The DeepMind paper is here:
arxiv.org/pdf/1712.01815.pdf

I have created a study of the games here:

I ran the first game through the lichess analysis: lichess.org/0LUhNlLB#85

Very suspicious how stockfish made around 15 acpl, which is higher than the time I put stockfish against Rybka with 10 second clock time: lichess.org/9i2OBdFy/black

Has anyone tried running this themselves?

Edit: nevermind

Very suspicious and hard to believe. But in a way it makes sense, since current top engines are hand-crafted and the neural network could develop better ways of evaluating positions and tree search methods (perhaps?)

This is actually quite scary how the sheer power of deep learning destroyed the years of development and testing network. However, note that this is not the complete story as it may seem:

1) It seems like SF didn't use Syzygy 6-men tablebase, there is nothing about it in the paper. SF is not particularly strong in endgame without it and SF became TCEC-2016 champion using it.
2) This was not the strongest SF version now existing (ASM dev), but a year-old SF 8. No big conceptual difference, though.
3) The most important thing is hardware. SF is well-designed to run on a single core but scales not so well on multi-core machines. (Komodo is better) The participating SF was running on the 64-core machine. And AlphaZero was running on the entirely different hardware. Thus the results are still quite irrelevant for those who just want to analyze with the best program ever on one's computer. It is not definite that AlphaZero can be effectively ported to x86_64 architecture.

Nevertheless, the result is pretty exciting as AlphaZero should still be one of the strongest existing chess entities on the Earth. The fact that it learned to be such from scratch knowing only the rules of chess may have quite large consequences in AI field.

One more of their examples analyzed: lichess.org/1ndiiD4U#0

impressive games, especially game 4 with that queen on h8 ... Took a while for my stockfish here to see that hes lost.

You see, just copy how nature works and you get impressive results :-)

i would be interested if Alpha Zero has an evaluation like other engines or if it does it in winning percentage, like Alpha Go does.

Would also be interesting to see if it understands fortresses like this one (draw):



Current engines dont understand that this is a draw. For example they dont find the first move. Edit: wrong. Stockfish finds it on depth 33. Still (if it takes Bb8) it gives an advantage of 10+ in the end position.

this is also a nice example to see how broken the hash handling of current engines is, because they dont write it to disk. Go into lines until move hundred to get a 0.0 eval, then go back. Around move 50 it will suddenly show +4 again. If you then go back to move 51 it has forgotten that it was 0.0 ...

Should Alpha Zero do this right, i would pay 1000 dollars to buy it.

I am wondering how long it will take the chess software "industry" to follow up on this. From a quick look at their paper it seems as if the techniques used might be within grasp of a good data/computer science student. Their hardware, however, might not be available to most of us. If I understand correctly, they use 700.000 batches of 4096 positions each (which then gives approximately 44 million games). In each position, the computer had 0.04 seconds thinking time (if I understand table S3 correctly) which yields about 112 million seconds = 31.000 hours in total. To get to the 9 hour training time reported in the paper, one needs a lot of parallelism I guess.

By the way, someone is trying to replicate the design in an open source way: github.com/Zeta36/chess-alpha-zero