Ethereal 12 (3400) loses to God! (Most Amazing Game I've Seen)

Discussion of anything and everything relating to chess playing software and machines.

Moderators: Harvey Williamson, bob, hgm

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Post Reply
jp
Posts: 1404
Joined: Mon Apr 23, 2018 5:54 am

Re: Ethereal 12 (3400) loses to God! (Most Amazing Game I've Seen)

Post by jp » Sun Apr 12, 2020 3:00 pm

M ANSARI wrote:
Sun Apr 12, 2020 5:49 am
I think as a human looking at this position

[d]r4rk1/1p2qpbp/2pp1np1/2n5/2PNPBP1/p1N1Q3/PP2B2P/2KR3R w - - 0 18

White could have taken the bull by the horns and ignore pawn takes with check and played Nf5! ...

r4rk1/1p2qpbp/2pp1np1/2n2N2/2P1PBP1/p1N1Q3/PP2B2P/2KR3R b - - 0 18

I suspect that would be what someone like Magnus or Anand would have played as humans tend to try and give back material once they feel a sacrifice is giving up way too much initiative. Of course against an engine this would be suicide, but against another human ... the idea would be to reverse roles and let the other guy work out the crazy complications. Nf5 seems playable as the black Queen has no good squares and will lose a tempo in a very dynamic position if pawn takes N.
Yes, you are right --- but way underselling the move! As you say, it's an obvious move for humans, but the engines approve too and have it top choice almost instantly -- Komodo 13.2 at depth 13, SF10 at depth 15.

It looks totally winning, so this should be called a blunder by Ethereal (except that the position is so good, the second-best move is probably also winning, with the line on the previous page where White ends up with a Bishop for 3 Pawns and then picks up one or two of those Pawns) and pathetic play by Leela to end in a lost position by move 18.

SF10:

depth=34
+3.41 18.Nf5 axb2+ 19.Kc2 gxf5 20.Bxd6 Qd8 21.Qxc5 Re8 22.Be5 Nd7 23.Rxd7 Qxd7 24.Bxg7 Kxg7 25.Rd1 Qc7 26.gxf5 f6 27.Qe3 Red8 28.Rg1+ Kh8 29.Qh6 b1=B+ 30.Rxb1 Qe7 31.a4 Rg8 32.Qh3 Rab8 33.Qe3 Rbd8 34.e5 Rg5 35.Qf3 Qxe5 36.Rxb7 Rg7 37.Rxg7 Kxg7 38.Qg3+ Qxg3 39.hxg3

Komodo 13.2:

depth=25
+1.62 18.Nf5 gxf5 19.Bxd6 axb2+ 20.Kc2 Qd8 21.Qxc5 Re8 22.e5 Nd7 23.Qb4 Qg5 24.h4 Qh6 25.gxf5 Nxe5 26.Qxb2 Qf6 27.Kb1 Qxf5+ 28.Qc2 Qf2 29.Rhf1 Qb6+ 30.Qb2 Qxb2+ 31.Kxb2 Ng6 32.Bd3 Re3 33.Kb3 Nxh4 34.Bc5 Rh3 35.Rh1 Rxh1 36.Rxh1 Ng6 37.Bb1 Ra5 38.Bb4 Ra8 39.Bf5
+1.22 18.b3 Nfxe4 19.Bf3 Nxc3 20.Qxe7 Nxa2+ 21.Kb1 Nb4 22.Qe3 a2+ 23.Kb2 Ne6 24.Bxd6 a1=Q+ 25.Rxa1 Bxd4+ 26.Qxd4 Nxd4 27.Bxb4 Nxf3 28.Bxf8 Rxf8 29.Raf1 Ne5 30.g5 Rd8 31.Rd1 Rxd1 32.Rxd1 Nf3 33.Rd7 Nxg5 34.h4 Ne4 35.Rxb7 f5 36.Rd7 c5 37.Kc2 h6 38.Ra7

Milos
Posts: 3923
Joined: Wed Nov 25, 2009 12:47 am

Re: Ethereal 12 (3400) loses to God! (Most Amazing Game I've Seen)

Post by Milos » Sun Apr 12, 2020 8:25 pm

jp wrote:
Sun Apr 12, 2020 3:00 pm
M ANSARI wrote:
Sun Apr 12, 2020 5:49 am
I think as a human looking at this position

[d]r4rk1/1p2qpbp/2pp1np1/2n5/2PNPBP1/p1N1Q3/PP2B2P/2KR3R w - - 0 18

White could have taken the bull by the horns and ignore pawn takes with check and played Nf5! ...

r4rk1/1p2qpbp/2pp1np1/2n2N2/2P1PBP1/p1N1Q3/PP2B2P/2KR3R b - - 0 18

I suspect that would be what someone like Magnus or Anand would have played as humans tend to try and give back material once they feel a sacrifice is giving up way too much initiative. Of course against an engine this would be suicide, but against another human ... the idea would be to reverse roles and let the other guy work out the crazy complications. Nf5 seems playable as the black Queen has no good squares and will lose a tempo in a very dynamic position if pawn takes N.
Yes, you are right --- but way underselling the move! As you say, it's an obvious move for humans, but the engines approve too and have it top choice almost instantly -- Komodo 13.2 at depth 13, SF10 at depth 15.

It looks totally winning, so this should be called a blunder by Ethereal (except that the position is so good, the second-best move is probably also winning, with the line on the previous page where White ends up with a Bishop for 3 Pawns and then picks up one or two of those Pawns) and pathetic play by Leela to end in a lost position by move 18.

SF10:

depth=34
+3.41 18.Nf5 axb2+ 19.Kc2 gxf5 20.Bxd6 Qd8 21.Qxc5 Re8 22.Be5 Nd7 23.Rxd7 Qxd7 24.Bxg7 Kxg7 25.Rd1 Qc7 26.gxf5 f6 27.Qe3 Red8 28.Rg1+ Kh8 29.Qh6 b1=B+ 30.Rxb1 Qe7 31.a4 Rg8 32.Qh3 Rab8 33.Qe3 Rbd8 34.e5 Rg5 35.Qf3 Qxe5 36.Rxb7 Rg7 37.Rxg7 Kxg7 38.Qg3+ Qxg3 39.hxg3

Komodo 13.2:

depth=25
+1.62 18.Nf5 gxf5 19.Bxd6 axb2+ 20.Kc2 Qd8 21.Qxc5 Re8 22.e5 Nd7 23.Qb4 Qg5 24.h4 Qh6 25.gxf5 Nxe5 26.Qxb2 Qf6 27.Kb1 Qxf5+ 28.Qc2 Qf2 29.Rhf1 Qb6+ 30.Qb2 Qxb2+ 31.Kxb2 Ng6 32.Bd3 Re3 33.Kb3 Nxh4 34.Bc5 Rh3 35.Rh1 Rxh1 36.Rxh1 Ng6 37.Bb1 Ra5 38.Bb4 Ra8 39.Bf5
+1.22 18.b3 Nfxe4 19.Bf3 Nxc3 20.Qxe7 Nxa2+ 21.Kb1 Nb4 22.Qe3 a2+ 23.Kb2 Ne6 24.Bxd6 a1=Q+ 25.Rxa1 Bxd4+ 26.Qxd4 Nxd4 27.Bxb4 Nxf3 28.Bxf8 Rxf8 29.Raf1 Ne5 30.g5 Rd8 31.Rd1 Rxd1 32.Rxd1 Nf3 33.Rd7 Nxg5 34.h4 Ne4 35.Rxb7 f5 36.Rd7 c5 37.Kc2 h6 38.Ra7
Yes 18.Nf5 is quite a certainly win for white. It is also clear that black played quite crappy opening. The nice thing there is that best play for black involves another queen sac this time with much quicker death.
The question was once white missed that and played 18. b2-b3, was black queen sac justified and who would win in ideal play.
First black queen sac was the best continuation, and not only that, the only continuation that provides any chances for a draw. Therefore, this is the reason I don't find a particular beauty in that sacrifice. First, because it's quite obvious, and second because it doesn't change the possible outcome assuming high enough level of play.

h1a8
Posts: 441
Joined: Fri Jun 04, 2010 5:23 am

Re: Ethereal 12 (3400) loses to God! (Most Amazing Game I've Seen)

Post by h1a8 » Tue Apr 14, 2020 1:51 pm

BrendanJNorman wrote:
Fri Apr 10, 2020 3:34 am
I have been studying and playing chess seriously for 20 years, and in all this time, I have NEVER seen a game like this.

I have been very impressed with Ethereal 12, it is very, very strong and now with a more refined positional style to boot.

But the way Lc0, even on my very weak hardware (laptop with built-in Nividia card) beat it from a position where Ethereal thought it was +4 is the most amazing thing I've EVER seen.

Witness this amazing game.



To do this to a weak engine is one thing...but to Ethereal?

With a BISHOP vs a queen?

What the hell is going on here? Amazing.
What was the hardware (processor and threads used, hash size, gpu)?

What were the time conditions?

AndrewGrant
Posts: 808
Joined: Tue Apr 19, 2016 4:08 am
Location: U.S.A
Full name: Andrew Grant
Contact:

Re: Ethereal 12 (3400) loses to God! (Most Amazing Game I've Seen)

Post by AndrewGrant » Tue Apr 14, 2020 2:03 pm

Man, where are the Ethereal wins threads smh

jp
Posts: 1404
Joined: Mon Apr 23, 2018 5:54 am

Re: Ethereal 12 (3400) loses to God! (Most Amazing Game I've Seen)

Post by jp » Tue Apr 14, 2020 3:44 pm

h1a8 wrote:
Tue Apr 14, 2020 1:51 pm
What was the hardware (processor and threads used, hash size, gpu)?

What were the time conditions?
It was 3min+2s blitz, which no doubt helped create the blunders. The exact hardware was not specified.

AndrewGrant wrote:
Tue Apr 14, 2020 2:03 pm
Man, where are the Ethereal wins threads smh
Maybe you need to create one. :wink:

I guess the Ethereal blunders (as when it did not play Nf5 in a totally winning position) are more instructive for development, though. (If we test Ethereal now, at what depth and node count does it see Nf5 is winning?)

chrisw
Posts: 3809
Joined: Tue Apr 03, 2012 2:28 pm

Re: Ethereal 12 (3400) loses to God! (Most Amazing Game I've Seen)

Post by chrisw » Tue Apr 14, 2020 5:00 pm

BrendanJNorman wrote:
Fri Apr 10, 2020 3:34 am
I have been studying and playing chess seriously for 20 years, and in all this time, I have NEVER seen a game like this.

I have been very impressed with Ethereal 12, it is very, very strong and now with a more refined positional style to boot.

But the way Lc0, even on my very weak hardware (laptop with built-in Nividia card) beat it from a position where Ethereal thought it was +4 is the most amazing thing I've EVER seen.

Witness this amazing game.
Hi Brendan,

I just came across this thread. Yes, very revealing game, plus of course I'm inclined to agree with most of what you've been saying on the thread too.

I think what is going on here is the power of the AZ/Leela move history concept, plus its MCTS search, coordinating together very well. (I'm expecting a mass of flak from people who won't understand what I have to say, but anyway, nothing unusual there, and I'm only going to say it because I guess you're the only one here who is going to understand it).

What's important about this game (not this 'position') is that white, although material ahead, is completely stuck to do anything. It's one of those positions where GM turns up seeing all the kibbitzers around the board, looks at it a bit, and says "what can white do?", shrugs and walks off. Meanwhile the kibbitzers all look a bit bemused. AB programs are a bit like those kibbitzers, AB evaluates each position in the tree, divorced from how it got there, so, as long as at the end of the long AB 'best' line it still has the material up, then that's the evaluation it returns. Big fail.
What is AZ/Leela doing in this game (note use of word game, not position)? Well, it creates lines of play
M1, M2, M3, ..... M61, M61 and so on. And it doesn't evaluate M61 alone, it evaluates the prior seven positions as well (that's the history element presented to the NN inputs) [M61, M60, M59 ..... M54, M53], it evaluates a kind of moving window of eight positions up the line.
Brief digression here, MCTS is not picking out a best line, it picks out an average of results of lines, and the lines it will see (when white piece is nastily pinned on d4) are a mass of fails (white moves something and the position falls apart) and a numbers of 'Okays', where white doesn't lose his piece on d4.
So, what the NN has learnt, is that when white is in the same situation on move M61 as he was on the seven priors back to M53, white is in a GM "What can white do?" situation, discovered by the MCTS search. All that's needed is the averaging search and the NN 'knowledge' that being bogged down for move after move, is not a good situation for white. AZ/Leela will know that "what can white do?" doesn't mean white is going to lose, but it means white as sure as hell isn't going to win - and curiously enough Leela 'score' is mostly in the draw zone, while AB program is +4 or whatever.

So, I guess what is happening is the power of the move history inputs, and the gradual NN learn over time that stuck=not very likely to win. I recollect getting flamed as an idiot for saying long ago that the 'physics' parallel is the AZ/Leela have effectively added a time dimension to the evaluation. AB just measures mass and position, basically, it has two dimensions to work with. NN-MCTS also measures along the time axis, that's the eight history positions presented to the NN. With mass, position and time, you can get many concepts (see Physics) that you can't get with mass and position alone.
That's why you're correct arguing in this thread. The level of knowledge is way beyond the old AB method, it sees something they don't even consider and I'm not at all sure they ever would be able to. Time axis, another dimension. Also the fundament of the knowledge about fortresses that Leela finds so easy, and AB not. Time axis.



To do this to a weak engine is one thing...but to Ethereal?

With a BISHOP vs a queen?

What the hell is going on here? Amazing.

Terje
Posts: 254
Joined: Tue Nov 19, 2019 3:34 am
Location: https://github.com/TerjeKir/weiss
Full name: Terje Kirstihagen

Re: Ethereal 12 (3400) loses to God! (Most Amazing Game I've Seen)

Post by Terje » Tue Apr 14, 2020 5:07 pm

jp wrote:
Tue Apr 14, 2020 3:44 pm
h1a8 wrote:
Tue Apr 14, 2020 1:51 pm
What was the hardware (processor and threads used, hash size, gpu)?

What were the time conditions?
It was 3min+2s blitz, which no doubt helped create the blunders. The exact hardware was not specified.

AndrewGrant wrote:
Tue Apr 14, 2020 2:03 pm
Man, where are the Ethereal wins threads smh
Maybe you need to create one. :wink:

I guess the Ethereal blunders (as when it did not play Nf5 in a totally winning position) are more instructive for development, though. (If we test Ethereal now, at what depth and node count does it see Nf5 is winning?)
Single-threaded Ethereal 12.04 finds Nf5 and sticks to it in 31ms at depth 11 here.

jp
Posts: 1404
Joined: Mon Apr 23, 2018 5:54 am

Re: Ethereal 12 (3400) loses to God! (Most Amazing Game I've Seen)

Post by jp » Tue Apr 14, 2020 7:06 pm

jp wrote:
Tue Apr 14, 2020 3:44 pm
h1a8 wrote:
Tue Apr 14, 2020 1:51 pm
What was the hardware (processor and threads used, hash size, gpu)?
It was 3min+2s blitz, which no doubt helped create the blunders. The exact hardware was not specified.
Semi-correction: the CPU (but not more details) is specified in the pgn and said to give 2,081kN/s.

Terje wrote:
Tue Apr 14, 2020 5:07 pm
jp wrote:
Tue Apr 14, 2020 3:44 pm
I guess the Ethereal blunders (as when it did not play Nf5 in a totally winning position) are more instructive for development, though. (If we test Ethereal now, at what depth and node count does it see Nf5 is winning?)
Single-threaded Ethereal 12.04 finds Nf5 and sticks to it in 31ms at depth 11 here.
Thanks. That's interesting. So Ethereal 12.04 find it and sticks to it at depth 11, Komodo 13.2 at depth 13, SF10 at depth 15 (which it possibly reaches at least as quickly as the other engines reach lower depths).

I wonder why Ethereal did not play it in this game. After the 18th-move blunder, the pgn says "{2.31/22 15}", which I'm guessing is eval 2.31, depth 22.

Terje
Posts: 254
Joined: Tue Nov 19, 2019 3:34 am
Location: https://github.com/TerjeKir/weiss
Full name: Terje Kirstihagen

Re: Ethereal 12 (3400) loses to God! (Most Amazing Game I've Seen)

Post by Terje » Tue Apr 14, 2020 7:20 pm

jp wrote:
Tue Apr 14, 2020 7:06 pm
jp wrote:
Tue Apr 14, 2020 3:44 pm
h1a8 wrote:
Tue Apr 14, 2020 1:51 pm
What was the hardware (processor and threads used, hash size, gpu)?
It was 3min+2s blitz, which no doubt helped create the blunders. The exact hardware was not specified.
Semi-correction: the CPU (but not more details) is specified in the pgn and said to give 2,081kN/s.

Terje wrote:
Tue Apr 14, 2020 5:07 pm
jp wrote:
Tue Apr 14, 2020 3:44 pm
I guess the Ethereal blunders (as when it did not play Nf5 in a totally winning position) are more instructive for development, though. (If we test Ethereal now, at what depth and node count does it see Nf5 is winning?)
Single-threaded Ethereal 12.04 finds Nf5 and sticks to it in 31ms at depth 11 here.
Thanks. That's interesting. So Ethereal 12.04 find it and sticks to it at depth 11, Komodo 13.2 at depth 13, SF10 at depth 15 (which it possibly reaches at least as quickly as the other engines reach lower depths).

I wonder why Ethereal did not play it in this game. After the 18th-move blunder, the pgn says "{2.31/22 15}", which I'm guessing is eval 2.31, depth 22.
There could be a relevant difference between 12.00 and 12.04, though I'm doubtful.

User avatar
Dan Honeycutt
Posts: 5258
Joined: Mon Feb 27, 2006 3:31 pm
Location: Atlanta, Georgia

Re: Ethereal 12 (3400) loses to God! (Most Amazing Game I've Seen)

Post by Dan Honeycutt » Tue Apr 14, 2020 7:29 pm

BrendanJNorman wrote:
Fri Apr 10, 2020 3:34 am
With a BISHOP vs a queen?
A bishop vs a queen but WHAT a bishop. Very entertaining game, Brendan. Thanks for sharing.

Best
Dan H.

Post Reply