AlphaZero vs Stockfish

[Rodolfo Leoni]

So with a fraction of the processing power Stockfish is almost as good as Alpha Zero. Interesting...

Or better...

[Milos]

It is rather funny to see how the notion had Stockfish would be better if it was using a book seems to prevale. Imagine how strong a Chess player I would be, if I could make Kasparov do the moves for me...

Wouldn't help much if Kasparov would only make first 10 moves, you'd almost equally suck at it.

[Jouni]

From fishcooking:

SF8 (1/1 minute and limited hash) against SFlatest (40/40 minutes)

Score of sf8_1in1 vs sf_master: 2 - 26 - 72 [0.380] 100
Elo difference: -85.04 +/- 34.16

And book is additional +50 Elo.

[Uri Blass]

From fishcooking:

SF8 (1/1 minute and limited hash) against SFlatest (40/40 minutes)

Score of sf8_1in1 vs sf_master: 2 - 26 - 72 [0.380] 100
Elo difference: -85.04 +/- 34.16

And book is additional +50 Elo.

This is at 1 minute per move on 1 core.

The difference is going to be clearly smaller with 64 cores and
I see no evidence that book at that time control is additional +50 elo.
Maybe book does not give much at long time control because the program can often find better moves by itself.

[Milos]

This is at 1 minute per move on 1 core.

The difference is going to be clearly smaller with 64 cores and
I see no evidence that book at that time control is additional +50 elo.
Maybe book does not give much at long time control because the program can often find better moves by itself.

Smaller yes, but not 1/5 of it, maybe 1/2 of it.
How I see it on 64 cores (we also don't know if these are cores or threads, since they don't say which machine they used, if it is cores Numa was not used, Large Pages were also not used):
SFdev vs SF8 at least 30Elo, hash up to 10ELO, EGTBs up to 10 ELO, large pages 5-10Elo, better SMP implementation of SFdev compared to SF8 for large number of cores (>32) 5-10Elo, Cerebellum book 30Elo against conventional engine, but against A0 most probably more, coz A0 is well trained for only small amount of openings, and good book would help SF a lot to get much more draws with black.
If you add all this together it is easy 100Elo if not even more.

[Leo]

From fishcooking:

SF8 (1/1 minute and limited hash) against SFlatest (40/40 minutes)

Score of sf8_1in1 vs sf_master: 2 - 26 - 72 [0.380] 100
Elo difference: -85.04 +/- 34.16

And book is additional +50 Elo.

Deep Mind Google weakened SF terribly. Yes, they have a new impressive self learning machine, but if its so great lets see it play SF when its at full power. Then it would be impressive. Not now with flawed conditions.

[Leo]

This is at 1 minute per move on 1 core.

The difference is going to be clearly smaller with 64 cores and
I see no evidence that book at that time control is additional +50 elo.
Maybe book does not give much at long time control because the program can often find better moves by itself.

Smaller yes, but not 1/5 of it, maybe 1/2 of it.
How I see it on 64 cores (we also don't know if these are cores or threads, since they don't say which machine they used, if it is cores Numa was not used, Large Pages were also not used):
SFdev vs SF8 at least 30Elo, hash up to 10ELO, EGTBs up to 10 ELO, large pages 5-10Elo, better SMP implementation of SFdev compared to SF8 for large number of cores (>32) 5-10Elo, Cerebellum book 30Elo against conventional engine, but against A0 most probably more, coz A0 is well trained for only small amount of openings, and good book would help SF a lot to get much more draws with black.
If you add all this together it is easy 100Elo if not even more.

I am beginning to think the Alpha team was afraid the results would have been a lot less impressive if they had beefed things up for SF.

[FWCC]

Alpha Zero can prob give a GM two minor odds and do favorably.

[Rodolfo Leoni]

This is at 1 minute per move on 1 core.

The difference is going to be clearly smaller with 64 cores and
I see no evidence that book at that time control is additional +50 elo.
Maybe book does not give much at long time control because the program can often find better moves by itself.

Smaller yes, but not 1/5 of it, maybe 1/2 of it.
How I see it on 64 cores (we also don't know if these are cores or threads, since they don't say which machine they used, if it is cores Numa was not used, Large Pages were also not used):
SFdev vs SF8 at least 30Elo, hash up to 10ELO, EGTBs up to 10 ELO, large pages 5-10Elo, better SMP implementation of SFdev compared to SF8 for large number of cores (>32) 5-10Elo, Cerebellum book 30Elo against conventional engine, but against A0 most probably more, coz A0 is well trained for only small amount of openings, and good book would help SF a lot to get much more draws with black.
If you add all this together it is easy 100Elo if not even more.

I am beginning to think the Alpha team was afraid the results would have been a lot less impressive if they had beefed things up for SF.

The graph here show difference between SF8 and latest SF dev is more than 100 ELO.

https://nextchessmove.com/dev-builds

[Tobber]

This is at 1 minute per move on 1 core.

The difference is going to be clearly smaller with 64 cores and
I see no evidence that book at that time control is additional +50 elo.
Maybe book does not give much at long time control because the program can often find better moves by itself.

Smaller yes, but not 1/5 of it, maybe 1/2 of it.
How I see it on 64 cores (we also don't know if these are cores or threads, since they don't say which machine they used, if it is cores Numa was not used, Large Pages were also not used):
SFdev vs SF8 at least 30Elo, hash up to 10ELO, EGTBs up to 10 ELO, large pages 5-10Elo, better SMP implementation of SFdev compared to SF8 for large number of cores (>32) 5-10Elo, Cerebellum book 30Elo against conventional engine, but against A0 most probably more, coz A0 is well trained for only small amount of openings, and good book would help SF a lot to get much more draws with black.
If you add all this together it is easy 100Elo if not even more.

I am beginning to think the Alpha team was afraid the results would have been a lot less impressive if they had beefed things up for SF.

The graph here show difference between SF8 and latest SF dev is more than 100 ELO.

https://nextchessmove.com/dev-builds

No it doesn't, read again.

/John