Stockfish 070324 a Disaster

[Eduard]

The SF developers changed the parameters for small net threshold yesterday, and they are now not in evaluate.cpp but in evaluate.h, current: "constexpr inline int SmallNetThreshold = 1136, PsqtOnlyThreshold = 2656;"

[Eduard]

After compiling and installing the new Stockfish from yesterday, the PC shut down for a few minutes. Tested with cleaned memory on GUI ChessBase 16, tested with 16 threads on Ryzen 3900X, hash 4 GB, all 3456men Syzygy.

Here are some well known test positions. Test it yourself and have fun with Stockfish!

[d]5b2/p2k1p2/P3pP1p/n2pP1p1/1p1P2P1/1P1KBN2/7P/8 w - - 0 1

Analysis by Stockfish dev-20240313-nogit:

1.Bxg5 hxg5 2.Nxg5 Ke8 3.Nf3 Nxb3 4.h4 Nc1+ 5.Kd2 Na2 6.Kc2 b3+ 7.Kb2 Nb4 8.Kxb3 Nd3 9.Ka4 Nf4 10.Kb5 Ba3 11.Kc6 Bc1 12.Kb7 Kd7 13.Ng5 Be3 14.Nxf7 Bxd4 15.h5 Ke8 16.Ng5 Be3 17.h6 Kd7 18.Nf3 Ng6 19.h7 Nh8 20.Ne1 Bd4 21.Nd3 Be3 22.Kb8 Bg1 23.Nb4 Be3 24.Nc2 Bc5 25.g5 Bg1 26.Ne1 Bb6 27.Ka8 Bd4 28.Nd3 Be3
Weiß steht etwas besser: +/= (0.60 ++) Tiefe: 33/69 00:00:45 96854kN, tb=88549
1.Nxg5
Weiß steht etwas besser: +/= (0.64 ++) Tiefe: 33/76 00:00:48 108MN, tb=92602
[---]
1.Nxg5
Weiß steht besser: +/- (1.06 ++) Tiefe: 33/76 00:00:48 110MN, tb=93371
1.Nxg5 hxg5 2.Bxg5 Nxb3 3.h4 Bc5 4.dxc5 Nxc5+ 5.Kc2 b3+ 6.Kb1 d4 7.Kb2 Kc6 8.h5 Nd3+ 9.Kxb3 Nxe5 10.h6 Ng6 11.h7 Kb5 12.Bh4 Kxa6 13.Kc4 e5 14.Kc5 Kb7 15.Bg3 d3 16.Be1 Kc7 17.Kd5 Kd7 18.Bb4 Ke8 19.Kd6 e4 20.Bd2 a5 21.Kd5 Kd7 22.Kxe4 Ke6
Weiß steht besser: +/- (1.24) Tiefe: 33/76 00:00:49 114MN, tb=94721


[d]1B1r4/rp2npkp/2b1pbp1/1qp5/nPN1R3/1P1P1QP1/2P2PBP/5R1K w - - 0 1

Analysis by Stockfish dev-20240313-nogit:

1.Bxa7 Bxe4 2.dxe4 e5 3.bxc5 Nxc5 4.Qc3 Ne6 5.Bb6 Rd7 6.Be3 Rd8 7.Qb2 h5 8.h4 Nc8 9.Kh2 Qd7 10.Ra1 Qe7 11.Bh3 Nd6 12.Nd2 Nb5 13.b4 Qd6 14.Ra5 Nec7 15.Bf1 Qc6 16.Bxb5 Nxb5 17.c4 Nd4 18.Rd5 Rxd5 19.exd5 Qd7 20.Bxd4 exd4 21.Ne4 Qe7 22.Nxf6 Qxf6 23.Kg2 d3 24.Qxf6+ Kxf6
Weiß steht deutlich besser: +- (1.67) Tiefe: 33/56 00:00:28 288MN, tb=1799
1.Bxa7 Bxe4 2.dxe4 e5 3.bxc5 Nxc5 4.Qc3 Ne6 5.Bb6 Rd7 6.Be3 Rd8 7.Qb2 h5 8.h4 Nc8 9.Kh2 Qd7 10.Ra1 Qe7 11.Bh3 Nd6 12.Nd2 Nb5 13.b4 Qd6 14.Ra5 Nec7 15.Bf1 Qc6 16.Bxb5 Nxb5 17.c4 Nd4 18.Rd5 Rxd5 19.exd5 Qd7 20.Bxd4 exd4 21.Ne4 Qe7 22.Nxf6 Qxf6 23.Kg2 d3 24.Qxf6+ Kxf6
Weiß steht deutlich besser: +- (1.80 ++) Tiefe: 34/68 00:00:32 333MN, tb=2120
1.Qxf6+
Weiß steht deutlich besser: +- (1.91 ++) Tiefe: 34/68 00:00:35 369MN, tb=2160
[---]
1.Qxf6+
Weiß steht klar auf Gewinn: +- (4.54 ++) Tiefe: 34/68 00:00:39 404MN, tb=2609


[d]1r3rk1/3bbppp/1qn2P2/p2pP1P1/3P4/2PB1N2/6K1/qNBQ1R2 w - - 0 1

Analysis by Stockfish dev-20240313-nogit:

1.Rh1 Bf5 2.Bxf5 Qbxb1 3.Rxh7 Qxc1 4.Rxg7+ Kh8 5.Rh7+ Kg8 6.Rg7+
Die Stellung ist ausgeglichen: = (0.00) Tiefe: 40/11 00:05:21 4938MN, tb=603886
1.Bxh7+
Die Stellung ist ausgeglichen: = (0.02 ++) Tiefe: 41/23 00:05:23 4962MN, tb=604661
[---]
1.Bxh7+
Weiß hat entscheidenden Vorteil: +- (3.00 ++) Tiefe: 41/24 00:05:30 5027MN, tb=610830
1.Bxh7+ Kxh7 2.Rh1+ Kg8 3.Rh8+ Kxh8 4.Qh1+ Kg8 5.g6 fxg6 6.Ng5 Qbb2+ 7.Bxb2 Qxb2+ 8.Kg3 Qf2+ 9.Kxf2 Rxf6+ 10.exf6 Bxf6 11.Qxd5+ Kh8 12.Nf7+ Kh7 13.Qxd7 Rf8 14.Ke2 Ne7 15.Nd6 Bh4 16.Qg4 Nf5 17.Ne4 a4 18.Kd3 a3 19.Nxa3 Kg8 20.Nc4
Weiß steht auf Gewinn: +- (3.74) Tiefe: 41/57 00:05:32 5051MN, tb=620221

LIVE:



[d]2r2r1k/qb3p1p/p2p1Np1/np1Pp1P1/2P5/3RP2P/5R2/4Q1K1 w - - 0 1

Analysis by Stockfish dev-20240313-nogit:

1.Qxa5 Rxc4 2.Kh2 Rfc8 3.Rf3 Qc5 4.Qe1 Qb4 5.Qxb4 Rxb4 6.e4 Rbc4 7.Ra3 Kg7 8.Kg2 h6 9.h4 Rc2+ 10.Kg1 hxg5 11.hxg5 Kf8 12.Ra1 Rc1+ 13.Rxc1 Rxc1+ 14.Kf2 a5
Schwarz hat entscheidenden Vorteil: -+ (-3.09) Tiefe: 34/55 00:04:43 3577MN, tb=26079

Not solved in 5 Minutes.


[d]4q1kr/p6p/1prQPppB/4n3/4P3/2P5/PP2B2P/R5K1 w - - 0 1

Analysis by Stockfish dev-20240313-nogit:

1.Qa3 Rxe6 2.Qxa7 Qe7 3.Qa8+ Qe8 4.Qxe8+ Rxe8 5.a4 Kf7 6.Be3 Rb8 7.b4 Rhc8 8.Bd4 Nc6 9.Bf2 Ne5 10.Rf1 g5 11.Bd4 Ke7 12.Rf5 Ra8 13.Bb5 h5 14.c4 Nxc4 15.Bxf6+ Ke6 16.Bxg5 Ra7 17.Bc6 Ne3 18.Rf6+ Ke5 19.Rg6 Nc2 20.Bf6+ Kf4 21.Bg5+
Die Stellung ist ausgeglichen: = (0.00) Tiefe: 61/41 00:05:35 2967MN, tb=532244

Not solved in 5 Minutes.

[chrisw]

Stockfish dev:

https://www.youtube.com/watch?v=mDhNQPt8An0

[Cumnor]

Stockfish dev:

https://www.youtube.com/watch?v=mDhNQPt8An0

[Eduard]

The SF developers like to celebrate their progress and are dazzled by the few Elos that some testers like Stefan Pohl have achieved. Supposedly +3 Elo better than SF 16.1. Test it you self!

Here are the same 5 positions like above with Stockfish 16.1 (I restarted the PC):

5b2/p2k1p2/P3pP1p/n2pP1p1/1p1P2P1/1P1KBN2/7P/8 w - - 0 1

Analysis by Stockfish 16.1-avx2:

1.Nxg5 hxg5 2.Bxg5 Nxb3 3.h4 Bc5 4.Be3 Bxd4 5.Bxd4 Nxd4 6.Kxd4 Ke8 7.Kd3 Kf8 8.g5 Kg8 9.h5 Kh7 10.Kd4 Kg8 11.g6 fxg6 12.hxg6 Kf8 13.Kd3 Kg8 14.g7 Kh7 15.Kd4 Kg8 16.Kc5 b3 17.Kd6 b2 18.Ke7 b1Q 19.f7+ Kxg7 20.f8Q+ Kg6 21.Kxe6 Qe4 22.Qe8+ Kh6 23.Qc6 Qc4 24.Qxd5 Qc8+ 25.Ke7 Qc7+ 26.Ke8 Qc8+ 27.Qd8
Weiß hat entscheidenden Vorteil: +- (2.85 ++) Tiefe: 33/54 00:00:22 123MN, tb=84094


1B1r4/rp2npkp/2b1pbp1/1qp5/nPN1R3/1P1P1QP1/2P2PBP/5R1K w - - 0 1

Analysis by Stockfish 16.1-avx2:

1.Qxf6+ Kxf6 2.Be5+ Kg5 3.Bg7 Bxe4 4.f4+ Kh5 5.Bxe4 g5 6.Ne5 Qc6 7.g4+ Kh4 8.Bf6 h6 9.fxg5 Ng8 10.Bxd8 Ra8 11.Bxc6 Rxd8 12.Rg1 Kxg5 13.Nxf7+ Kf6 14.Nxd8 Nc3 15.b5 b6 16.Re1 e5 17.Rf1+ Ke7 18.Nf7 Ke6 19.g5 hxg5 20.Nxg5+ Kd6 21.Ne4+ Nxe4 22.Bxe4 Ke6 23.Bf5+ Kf6 24.Bg4+ Ke7
Weiß steht klar auf Gewinn: +- (5.98) Tiefe: 28/67 00:00:06 95527kN, tb=3450


1r3rk1/3bbppp/1qn2P2/p2pP1P1/3P4/2PB1N2/6K1/qNBQ1R2 w - - 0 1

Analysis by Stockfish 16.1-avx2:

1.Bxh7+ Kxh7 2.Rh1+ Kg8 3.Rh8+ Kxh8 4.Qh1+ Kg8 5.g6 fxg6 6.Ng5 Qbb2+ 7.Bxb2 Qxb2+ 8.Kg3 Qf2+ 9.Kxf2 Rxf6+ 10.exf6 Bxf6 11.Qxd5+ Kh8 12.Nf7+ Kh7 13.Qxd7 Nd8 14.Nxd8 Rxd8 15.Qh3+ Kg8 16.Ke3 a4 17.Kd3 Rb8 18.Qe6+ Kh7 19.Na3 Rb3 20.Qh3+ Kg8 21.Nc2 Rb2 22.Qd7 Kh7 23.Qxa4 Rb1 24.Kc4 Rf1 25.Qe8 Bxd4
Weiß steht klar auf Gewinn: +- (4.44) Tiefe: 29/66 00:00:05 55522kN, tb=16934


2r2r1k/qb3p1p/p2p1Np1/np1Pp1P1/2P5/3RP2P/5R2/4Q1K1 w - - 0 1

Analysis by Stockfish 16.1-avx2:

1.cxb5 axb5 2.Kh2 Nc4 3.e4 h5
Schwarz steht deutlich besser: -+ (-1.68) Tiefe: 10/14 00:00:00
1.Qxa5 Rxc4 2.Rc3 Rxc3 3.Qxc3 Rc8 4.Qd3 Rc1+ 5.Kh2 Qc5 6.Qe4 Qc4
Schwarz steht deutlich besser: -+ (-1.63) Tiefe: 11/14 00:00:00
1.Rf4 h5 2.Qh4 Rxc4 3.Rxc4 Nxc4 4.Nxh5 gxh5 5.Qxh5+ Kg8
Schwarz steht besser: -/+ (-0.94) Tiefe: 12/21 00:00:00
1.Rf4 h5 2.Qh4 Rfd8 3.Nxh5 Kg8 4.Nf6+ Kf8
Die Stellung ist ausgeglichen: = (0.00) Tiefe: 13/22 00:00:00 732kN


4q1kr/p6p/1prQPppB/4n3/4P3/2P5/PP2B2P/R5K1 w - - 0 1

Analysis by Stockfish 16.1-avx2:

1.Qa3 Rxe6 2.Qxa7 Qf7 3.Qb8+ Qe8 4.Qxe8+ Rxe8 5.a4 Kf7 6.Be3 Rb8 7.b4 Rhc8 8.Rf1 Ke7 9.Bd4 h5 10.Bd1 Rc7 11.Bb3 g5 12.Bd5 g4 13.c4 Nf3+ 14.Rxf3 gxf3 15.Kf2 Rbc8 16.Bxb6 Rxc4 17.Bxc4 Rxc4 18.a5 Kd7 19.b5 Rxe4 20.a6 Kc8 21.Ba7 Re2+ 22.Kxf3 Rxh2
Die Stellung ist ausgeglichen: = (0.04) Tiefe: 32/52 00:00:05 52729kN, tb=2945
1.Qxe5
Die Stellung ist ausgeglichen: = (0.07 ++) Tiefe: 33/52 00:00:06 57931kN, tb=4153
1.Qxe5
Die Stellung ist ausgeglichen: = (0.09 ++) Tiefe: 33/52 00:00:06 58587kN, tb=4235

Good and fast!

Instead of questioning themselves whether what they are doing is really better, they prefer to block me on Discord, and some of them even come here to the forum to make fun of me. They don’t even shy away from sharing fake content (another topic). Sad.

[Eduard]

Next Dimension 3072 Network is here to test: nn-8747f5ce199f.nnue
https://tests.stockfishchess.org/nns

The new 3072 network (see above) seems very strong. It is currently being tested on Fishtest against the master network with dimension 2560. In Bullet 60s + 0.6s with UHO Openings it currently looks like this
-> Picture https://up.picr.de/47254550cq.png

I have been using the large network for a long time because it plays better in balanced openings and especially in complex closed positions. I included the network in the developers' latest engine update from yesterday. I only changed the threshold for the small network (see Changes.txt). In analysis mode it is currently the best Stockfish dev I have tested after SF 16.1. I implemented Slow Mover, but no longer Minimum Thinking Time. Yesterday the developers implemented a new TM for Sudden Death (at less than 1s). It is possible that the Minimum Thinking Time option would now be harmful. Read about here:
https://github.com/official-stockfish/S ... b48b0e739e

Download Big-SF 150324, Win 64-Bit (avx2, avx512, bmi2, sse41, source code):
https://pixeldrain.com/u/onhHKCZY

[dkappe]

How do you know that the 3072 net is better in balanced positions than the 2560 net?

[Eduard]

Own experience. There are many examples. Some very strong correspondence chess players told me that they prefer to use dimension 3072. My own experience on Lichess is that Dimension 3072 make fewer repetitions against weaker engines than Dimension 2560. This refers to positions close to a eval of 0.00. Even if such positions objectively end in a draw, weaker opponents always make mistakes. With Dimension 3072, games run longer in balanced positions and as result I win more games against weaker engines.

[Ciekce]

translation: I made it up

what is this weird "dimension" terminology you like to use

[Whiskers]

translation: I made it up

what is this weird "dimension" terminology you like to use

You won't be laughing anymore when I implement dimension T-4 into Willow; the way it works is quite simple, every neuron in the hidden layer is itself a neural network, and every neuron in that neural network is ALSO a small neural network! This greatly improves analysis performance in 1 year + 1 week time controls :truetrue: