Need Lc0 help

[zullil]

...
Stockfish seems powerless in these positions. 2 hours with 64 cores does not help.
I can tell by examination of the statistics that the SF answers are not correct.
Does someone have a high end GPU setup and the time to pound these positions for a while?
(BTW, these are from Ed's two move file.)
[d]rnbqkbnr/pppppppp/8/8/P7/8/1PPPPPPP/RNBQKBNR b KQkq -
...

Can you elaborate on how you are determining "correct" here? Why is e5 correct and not Nf6?

I am using statistical results from actual games to compute the "real" centipawn evaluation.
When you only have a hundred games, there is no way to know if the computer or the statistics are correct.
But when you have many thousands of games (filtered to remove weak entrants) then the game outcomes are to be trusted.
The move e5 triumphs. Nf6 is drawish.

Are these statistics from human games? It seems to me that with perfect play this position is almost certainly simply a draw.

[Dann Corbit]

Egad. Nf6.

[Dann Corbit]

Are these statistics from human games? It seems to me that with perfect play this position is almost certainly simply a draw.

See above.

I agree that there is not an obvious advantage.
Since LC0 also switched to Nc6 (so far) I am not sure what to think about this position in particular.

My current hypothesis is that the poor move by white can be punished, but that a very deep counterplay may neutralize that plan.

[Daniel Shawul]

Egad. Nf6.

It stuck with Nf6 after 2 hours using net ID33004

Code: Select all

info depth 22 seldepth 67 time 7225636 nodes 195574102 score cp 16 hashfull 1000 nps 27066 tbhits 0 pv g8f6 d2d4 e7e6 e2e3 c7c5 g1f3 b8c6 c2c3 b7b6 b1d2 c8b7 f1d3 d7d5 a4a5 b6a5 d1a4 f6d7 d3b5 d8b6 d2b3 a7a6 b5e2 f8d6 b3a5 c6a5 a4a5 b6a5 a1a5 e8g8 e1g1 f8b8 f1e1 d6c7 a5a3 h7h6 b2b3 c7d6 a3a2 b7c6 f3d2 a6a5 c1a3 a5a4 b3a4 a8a4 e1a1 b8a8 e2d1
info string b7b5  (234 ) N:    6847 (+ 0) (P:  1.40%) (Q: -0.56104) (U: 0.61149) (Q+U:  0.05045) (V:  -.----) 
info string g7g5  (378 ) N:    8674 (+ 0) (P:  1.46%) (Q: -0.45504) (U: 0.50548) (Q+U:  0.05044) (V:  -.----) 
info string f7f6  (346 ) N:   16136 (+ 0) (P:  1.96%) (Q: -0.31394) (U: 0.36440) (Q+U:  0.05046) (V:  -.----) 
info string g8h6  (161 ) N:   21457 (+ 0) (P:  2.18%) (Q: -0.25424) (U: 0.30470) (Q+U:  0.05046) (V:  -.----) 
info string b8a6  (34  ) N:   24306 (+ 0) (P:  2.05%) (Q: -0.20198) (U: 0.25244) (Q+U:  0.05046) (V:  -.----) 
info string h7h5  (403 ) N:   26064 (+ 0) (P:  2.15%) (Q: -0.19601) (U: 0.24646) (Q+U:  0.05045) (V:  -.----) 
info string a7a5  (207 ) N:   37727 (+ 0) (P:  2.59%) (Q: -0.15505) (U: 0.20551) (Q+U:  0.05046) (V:  -.----) 
info string f7f5  (351 ) N:   52564 (+ 0) (P:  2.64%) (Q: -0.09974) (U: 0.15020) (Q+U:  0.05046) (V:  -.----) 
info string b7b6  (230 ) N:   69919 (+ 0) (P:  2.80%) (Q: -0.06945) (U: 0.11991) (Q+U:  0.05046) (V:  -.----) 
info string h7h6  (400 ) N:   75524 (+ 0) (P:  2.79%) (Q: -0.06016) (U: 0.11062) (Q+U:  0.05046) (V:  -.----) 
info string d7d6  (288 ) N:   79348 (+ 0) (P:  3.04%) (Q: -0.06419) (U: 0.11464) (Q+U:  0.05046) (V:  -.----) 
info string g7g6  (374 ) N:  101168 (+ 0) (P:  3.05%) (Q: -0.03980) (U: 0.09026) (Q+U:  0.05046) (V:  -.----) 
info string c7c6  (259 ) N:  103117 (+ 0) (P:  2.88%) (Q: -0.03317) (U: 0.08363) (Q+U:  0.05046) (V:  -.----) 
info string a7a6  (204 ) N:  139029 (+ 0) (P:  3.12%) (Q: -0.01663) (U: 0.06709) (Q+U:  0.05046) (V:  -.----) 
info string b8c6  (36  ) N:  158649 (+ 0) (P:  3.51%) (Q: -0.01587) (U: 0.06632) (Q+U:  0.05046) (V:  -.----) 
info string e7e6  (317 ) N:  298971 (+ 0) (P:  4.38%) (Q:  0.00658) (U: 0.04388) (Q+U:  0.05046) (V:  -.----) 
info string c7c5  (264 ) N:  579401 (+ 2) (P: 11.80%) (Q: -0.01053) (U: 0.06099) (Q+U:  0.05046) (V:  -.----) 
info string d7d5  (293 ) N: 5260372 (+ 0) (P: 18.07%) (Q:  0.04017) (U: 0.01028) (Q+U:  0.05046) (V:  -.----) 
info string e7e5  (322 ) N: 6660675 (+ 0) (P: 12.62%) (Q:  0.04438) (U: 0.00567) (Q+U:  0.05005) (V:  -.----) 
info string g8f6  (159 ) N: 181854153 (+286) (P: 15.51%) (Q:  0.05020) (U: 0.00026) (Q+U:  0.05046) (V:  -.----) 
bestmove g8f6 ponder d2d4

Lc0 doesn't generate too many nodes it seems only 195 million nodes after 2 hours analysis.
I think scorpio stores 10x more nodes so I don't expect it to last for 2 hours.

[Dann Corbit]

I will be keen to see what scorpio says, and especially if it likes e5 even for a time.

[Dann Corbit]

when you are done, I would like the logs for study

[ChiefPushesWood]

Dann,
I had a little time so I ran position one. I ran it up to 31M nodes and even forced equal visits between Nf6 and e5 to see if she changed her mind. Leela simply prefers Nf6 just slightly over e5. Nf6 is the leading policy move.... Of interest: Leela's #1 move was e5 until around 4.5M nodes.

Chief

[Daniel Shawul]

Second position search stopped after one hour by itself. It searched 72million nodes only so it must be the effect of smart pruning
as g2g3 is the overwhelming favourite. For the previous one, I did "go infinite" but for this one i did "go movetime 7200000" which enabled
the smart pruning i think

Code: Select all

info depth 26 seldepth 66 time 3172339 nodes 72956906 score cp 54 hashfull 1000 nps 22997 tbhits 0 pv g2g3 c7c6 f1g2 g8f6 e1g1 c8f5 c2c4 e7e6 b1c3 h7h6 f3d2 f8e7 e2e4 d5e4 d2e4 e8g8 c1f4 f6e4 c3e4 a6a5 f1e1 b8d7 e4d6 e7d6 f4d6 f8e8 d1b3 d7f6 d6e5 e8e7 h2h3 a5a4 b3c3 e7d7 b2b4 a4b3 a2b3 a8a1 e1a1 f6h7 a1a7 f7f6 e5f4 h7g5 g1h2 g8h7 f4e3 f5e4 e3g5
info string c1h6  (69  ) N:    1324 (+ 0) (P:  0.84%) (Q: -0.89024) (U: 1.05216) (Q+U:  0.16192) (V:  -.----) 
info string e1d2  (105 ) N:    1833 (+ 0) (P:  0.68%) (Q: -0.45237) (U: 0.61302) (Q+U:  0.16065) (V:  -.----) 
info string g2g4  (378 ) N:    2935 (+ 0) (P:  1.08%) (Q: -0.44944) (U: 0.60988) (Q+U:  0.16045) (V:  -.----) 
info string e2e4  (322 ) N:    4329 (+ 0) (P:  1.13%) (Q: -0.27204) (U: 0.43197) (Q+U:  0.15993) (V:  -.----) 
info string f3h4  (586 ) N:    5926 (+ 0) (P:  1.29%) (Q: -0.20298) (U: 0.36260) (Q+U:  0.15962) (V:  -.----) 
info string h1g1  (177 ) N:    6163 (+ 0) (P:  1.19%) (Q: -0.16195) (U: 0.32132) (Q+U:  0.15937) (V:  -.----) 
info string f3g1  (568 ) N:    6440 (+ 0) (P:  1.20%) (Q: -0.15090) (U: 0.31026) (Q+U:  0.15936) (V:  -.----) 
info string f3d2  (570 ) N:    9066 (+ 0) (P:  1.29%) (Q: -0.07798) (U: 0.23703) (Q+U:  0.15904) (V:  -.----) 
info string f3g5  (590 ) N:    9146 (+ 0) (P:  1.60%) (Q: -0.13144) (U: 0.29070) (Q+U:  0.15926) (V:  -.----) 
info string d1d2  (81  ) N:   11645 (+ 0) (P:  1.45%) (Q: -0.04823) (U: 0.20716) (Q+U:  0.15893) (V:  -.----) 
info string b1a3  (34  ) N:   13173 (+ 0) (P:  1.65%) (Q: -0.04876) (U: 0.20770) (Q+U:  0.15894) (V:  -.----) 
info string b2b4  (234 ) N:   13551 (+ 0) (P:  1.47%) (Q: -0.02178) (U: 0.18060) (Q+U:  0.15881) (V:  -.----) 
info string h2h4  (403 ) N:   19273 (+ 0) (P:  1.74%) (Q:  0.00896) (U: 0.14972) (Q+U:  0.15868) (V:  -.----) 
info string a2a4  (207 ) N:   23177 (+ 0) (P:  2.13%) (Q:  0.00631) (U: 0.15238) (Q+U:  0.15869) (V:  -.----) 
info string c1d2  (57  ) N:   24096 (+ 0) (P:  1.95%) (Q:  0.02397) (U: 0.13463) (Q+U:  0.15860) (V:  -.----) 
info string c1e3  (63  ) N:   24438 (+ 0) (P:  1.92%) (Q:  0.02813) (U: 0.13047) (Q+U:  0.15860) (V:  -.----) 
info string d1d3  (86  ) N:   24862 (+ 0) (P:  1.85%) (Q:  0.03529) (U: 0.12325) (Q+U:  0.15854) (V:  -.----) 
info string f3e5  (588 ) N:   25976 (+ 0) (P:  2.00%) (Q:  0.03053) (U: 0.12806) (Q+U:  0.15859) (V:  -.----) 
info string b1c3  (36  ) N:   27671 (+ 0) (P:  1.93%) (Q:  0.04295) (U: 0.11559) (Q+U:  0.15854) (V:  -.----) 
info string b2b3  (230 ) N:   30298 (+ 0) (P:  1.91%) (Q:  0.05405) (U: 0.10444) (Q+U:  0.15849) (V:  -.----) 
info string h2h3  (400 ) N:   37295 (+ 0) (P:  2.32%) (Q:  0.05533) (U: 0.10315) (Q+U:  0.15848) (V:  -.----) 
info string a2a3  (204 ) N:   49642 (+ 0) (P:  2.47%) (Q:  0.07579) (U: 0.08260) (Q+U:  0.15839) (V:  -.----) 
info string e2e3  (317 ) N:   56762 (+ 0) (P:  2.53%) (Q:  0.08428) (U: 0.07407) (Q+U:  0.15835) (V:  -.----) 
info string c2c3  (259 ) N:   86187 (+ 0) (P:  3.03%) (Q:  0.10001) (U: 0.05828) (Q+U:  0.15828) (V:  -.----) 
info string b1d2  (33  ) N:  107173 (+ 0) (P:  3.27%) (Q:  0.10753) (U: 0.05068) (Q+U:  0.15822) (V:  -.----) 
info string c1g5  (67  ) N:  267250 (+ 0) (P:  8.20%) (Q:  0.10727) (U: 0.05097) (Q+U:  0.15824) (V:  -.----) 
info string c1f4  (65  ) N:  675374 (+ 0) (P: 17.36%) (Q:  0.11393) (U: 0.04268) (Q+U:  0.15661) (V:  -.----) 
info string c2c4  (264 ) N: 1022814 (+ 0) (P: 20.36%) (Q:  0.12161) (U: 0.03305) (Q+U:  0.15467) (V:  -.----) 
info string g2g3  (374 ) N: 70369086 (+262) (P: 10.15%) (Q:  0.15797) (U: 0.00024) (Q+U:  0.15821) (V:  -.----) 
bestmove g2g3 ponder c7c6

[Dann Corbit]

I suspect that an analysis of the plans will show some great tactical move that has a foil.

This analysis is surprisingly pivotal for me.
For two decades, I have known that the top engines get it wrong a lot
But it was easy to get to the right answer, just double the time and run again.
Still wrong? double time, run again. Rinse, repeat to deep enough and the engines would find it.
Usually, before 45 plies or so, the engine would find the answer and I would get the plan I was looking for.
Guided by statistics, I knew that the engines were wrong, and I knew how to delve ahead.

But maybe now, after all these years, it is not so simple.
I have solved at least a million positions where the engine had it wrong and a deeper search would find it.

But now, maybe the engines are so much smarter that it is not so simple anymore.

[Dann Corbit]

Second position search stopped after one hour by itself. It searched 72million nodes only so it must be the effect of smart pruning
as g2g3 is the overwhelming favourite. For the previous one, I did "go infinite" but for this one i did "go movetime 7200000" which enabled
the smart pruning i think

SF also liked g3, but it took a long time to find it,

Code: Select all

rnbqkbnr/1pp1pppp/p7/3p4/3P4/5N2/PPP1PPPP/RNBQKB1R w KQkq - acd 26; ce 40; acs 24; acn 430776; pv c4 e6 Bg5 Be7 Bxe7 Nxe7 e3 O-O Nc3 Nd7 Rc1 dxc4 Bxc4 b5 Be2 Bb7 b4 Nf5 O-O Nb6 Ne5 Nd6 a4 Qg5 Bf3 Bxf3 Nxf3 Qe7 Qc2 c6 Ne5 Rfc8 a5 Nbc4 Nd3 Rd8 Nc5 h6 ; bm c4;
rnbqkbnr/1pp1pppp/p7/3p4/3P4/5N2/PPP1PPPP/RNBQKB1R w KQkq - acd 27; ce 29; acs 50; acn 869025; pv c4 e6 Bg5 Be7 Bxe7 Nxe7 Nc3 O-O e3 Nd7 Rc1 dxc4 Bxc4 b5 Be2 Bb7 b4 Nd5 a3 c6 O-O a5 Bd3 Qe7 Nxd5 exd5 Qc2 axb4 axb4 g6 Rb1 Rfe8 ; bm c4;
rnbqkbnr/1pp1pppp/p7/3p4/3P4/5N2/PPP1PPPP/RNBQKB1R w KQkq - acd 28; ce 38; acs 51; acn 886208; pv c4 e6 Bg5 Be7 Bxe7 Nxe7 e3 O-O Nc3 Nd7 Rc1 dxc4 Bxc4 b5 Be2 Bb7 b4 a5 a3 c6 O-O axb4 axb4 Nd5 Nxd5 exd5 Qd2 Qe7 Ra1 Nb6 g3 Nc4 Qc3 h6 Rfc1 ; bm c4;
rnbqkbnr/1pp1pppp/p7/3p4/3P4/5N2/PPP1PPPP/RNBQKB1R w KQkq - acd 29; ce 27; acs 70; acn 1211000; pv c4 e6 Bg5 Be7 Bxe7 Nxe7 Nc3 O-O e3 Nd7 Rc1 dxc4 Bxc4 b5 Be2 Bb7 b4 a5 a3 c6 O-O axb4 axb4 Nd5 Nxd5 exd5 Ne1 Qe7 Nd3 Ra4 Re1 Rfa8 g3 Nf6 Nc5 Ra2 ; bm c4;
rnbqkbnr/1pp1pppp/p7/3p4/3P4/5N2/PPP1PPPP/RNBQKB1R w KQkq - acd 30; ce 34; acs 78; acn 1344000; pv c4 e6 Bg5 Be7 Bxe7 Nxe7 e3 Nd7 Nc3 O-O Qc2 Nf6 Rc1 b6 cxd5 Nexd5 Nxd5 Qxd5 Bc4 Qa5+ Qd2 Qxd2+ Nxd2 Bb7 Bd3 Rac8 f3 Rfd8 Ke2 g6 g4 Kg7 g5 Nd5 h4 h5 gxh6+ Kxh6 Nc4 f6 Nd2 c5 dxc5 bxc5 Ne4 Nb4 Nxc5 Nxd3 ; bm c4;
rnbqkbnr/1pp1pppp/p7/3p4/3P4/5N2/PPP1PPPP/RNBQKB1R w KQkq - acd 31; ce 35; acs 115; acn 1979000; pv c4 e6 Bg5 Be7 Bxe7 Nxe7 e3 Nd7 Nc3 O-O Qc2 Nf6 Rc1 b6 cxd5 Nexd5 Nxd5 Qxd5 Bc4 Qa5+ Qd2 Qxd2+ Nxd2 Bb7 Bd3 Rac8 f3 Rfd8 Ke2 g6 g4 h5 g5 Nd5 a3 h4 Rhd1 Kg7 Be4 c5 Bxd5 ; bm c4;
rnbqkbnr/1pp1pppp/p7/3p4/3P4/5N2/PPP1PPPP/RNBQKB1R w KQkq - acd 32; ce 29; acs 155; acn 2651000; pv c4 e6 Bg5 Nf6 e3 Be7 Qc2 Nbd7 cxd5 exd5 Nc3 O-O Bd3 h6 Bh4 c5 a4 c4 Bf5 a5 O-O Nb6 Ne5 Bxf5 Qxf5 Qc8 Qc2 Re8 Rae1 Nbd7 f4 Bb4 Nxd7 Qxd7 ; bm c4;
rnbqkbnr/1pp1pppp/p7/3p4/3P4/5N2/PPP1PPPP/RNBQKB1R w KQkq - acd 33; ce 27; acs 218; acn 3720000; pv c4 e6 Bg5 Be7 Bxe7 Nxe7 e3 O-O Nc3 Nd7 Qc2 Nf6 Rc1 b6 cxd5 Nexd5 Nxd5 Qxd5 Bc4 Qa5+ Qd2 Qxd2+ Nxd2 Bb7 f3 Rfd8 Ke2 Rac8 Bd3 a5 h4 c5 dxc5 Nh5 Rhd1 Ng3+ Kf2 Rxd3 Kxg3 Rxc5 Rxc5 bxc5 Kf2 ; bm c4;
rnbqkbnr/1pp1pppp/p7/3p4/3P4/5N2/PPP1PPPP/RNBQKB1R w KQkq - acd 34; ce 28; acs 254; acn 4332000; pv c4 dxc4 e3 Nf6 Bxc4 c5 O-O e6 b3 cxd4 Qxd4 Qxd4 Nxd4 b6 Nd2 Bc5 Bb2 Nbd7 a3 Bb7 Rac1 Rc8 b4 Bd6 Be2 O-O f4 Rfd8 Rfd1 Be7 h3 a5 bxa5 bxa5 Nc4 Bd5 Nxa5 Nc5 Nc4 ; bm c4;
rnbqkbnr/1pp1pppp/p7/3p4/3P4/5N2/PPP1PPPP/RNBQKB1R w KQkq - acd 35; ce 23; acs 384; acn 6521000; pv c4 dxc4 e3 e6 Bxc4 c5 O-O Nf6 dxc5 Qxd1 Rxd1 Bxc5 Be2 Nbd7 b3 b6 Nfd2 Bb7 Bb2 O-O Nc4 Rfd8 Nc3 Be7 f3 Bc6 a4 Rab8 Kf2 a5 Nb5 Nd5 Rac1 Bxb5 axb5 Nc5 Be5 Rbc8 Rb1 ; bm c4;
rnbqkbnr/1pp1pppp/p7/3p4/3P4/5N2/PPP1PPPP/RNBQKB1R w KQkq - acd 36; ce 16; acs 581; acn 9847000; pv c4 dxc4 e3 Nf6 Bxc4 e6 a4 c5 O-O cxd4 exd4 Nc6 Nc3 Be7 d5 Nxd5 Nxd5 exd5 Bxd5 O-O Qb3 Na5 Qa2 Nc6 Rd1 Be6 Be3 Bxd5 Rxd5 Qc8 Bd2 Bf6 Re1 Qc7 h3 Rfd8 Bc3 Bxc3 bxc3 Re8 Rxe8+ Rxe8 Qd2 Rd8 Rxd8+ Nxd8 ; bm c4;
rnbqkbnr/1pp1pppp/p7/3p4/3P4/5N2/PPP1PPPP/RNBQKB1R w KQkq - acd 37; ce 29; acs 681; acn 11537000; pv c4 e6 cxd5 exd5 Nc3 Nf6 Bg5 c6 Qc2 Be6 e3 Nbd7 Bd3 Bd6 Bf4 Qe7 O-O O-O Bxd6 Qxd6 Rfc1 Rfc8 b4 b6 Rab1 c5 dxc5 bxc5 bxc5 Rxc5 Nd4 Rac8 Qb2 Ne5 Be2 Ne4 Nxe4 Rxc1+ Rxc1 Rxc1+ Qxc1 dxe4 ; bm c4;
rnbqkbnr/1pp1pppp/p7/3p4/3P4/5N2/PPP1PPPP/RNBQKB1R w KQkq - acd 38; ce 30; acs 813; acn 13727000; pv c4 e6 Bg5 Be7 Bxe7 Qxe7 Nbd2 Nf6 Qc2 O-O Rc1 h6 e3 c5 dxc5 Qxc5 Be2 Nc6 O-O dxc4 Nxc4 b5 Nce5 Qxc2 Rxc2 Nb4 Rd2 Bb7 a3 Nbd5 Rc1 Rac8 Rdc2 Rxc2 Rxc2 Ne4 Nd4 Rc8 Rxc8+ Bxc8 Nd3 Nd6 Kf1 Nf6 Nb3 Nd7 Nbc5 Nxc5 Nxc5 Kf8 ; bm c4;
rnbqkbnr/1pp1pppp/p7/3p4/3P4/5N2/PPP1PPPP/RNBQKB1R w KQkq - acd 39; ce 21; acs 930; acn 15661000; pv c4 e6 cxd5 exd5 Nc3 Nf6 Bg5 c6 Qc2 Be6 e3 Nbd7 h3 Bd6 Bd3 Qc7 Ne2 O-O O-O h6 Bf4 Rfe8 a4 Ne4 b4 Bf5 Bxd6 Qxd6 Rfc1 g5 Qb2 Qf6 b5 axb5 axb5 Rxa1 Qxa1 c5 dxc5 Qxa1 Rxa1 Ndxc5 ; bm c4;
rnbqkbnr/1pp1pppp/p7/3p4/3P4/5N2/PPP1PPPP/RNBQKB1R w KQkq - acd 40; ce 18; acs 1332; acn 22313000; pv c4 e6 cxd5 exd5 Nc3 Nf6 Bg5 Be7 e3 h6 Bf4 Bf5 Qb3 Ra7 Ne5 c6 Bd3 Bxd3 Nxd3 O-O O-O a5 Rac1 Nbd7 h3 Re8 a4 Nf8 Bh2 Ng6 Rfe1 Bf8 Ne2 Ne4 f3 Nd6 Ne5 Nxe5 dxe5 Nf5 Nd4 Nxd4 exd4 ; bm c4;
rnbqkbnr/1pp1pppp/p7/3p4/3P4/5N2/PPP1PPPP/RNBQKB1R w KQkq - acd 41; ce 16; acs 1903; acn 31759000; pv c4 e6 cxd5 exd5 Nc3 Nf6 Bg5 Be7 e3 h6 Bf4 Bf5 Qb3 Ra7 Ne5 O-O Bd3 Bxd3 Nxd3 c6 O-O Nbd7 Rad1 Re8 f3 Qb6 Qxb6 Nxb6 Rfe1 Raa8 Kf2 a5 Rc1 Bb4 h3 Nc4 Bg3 Rad8 b3 Na3 Nxb4 axb4 Na4 Nd7 Bc7 ; bm c4;
rnbqkbnr/1pp1pppp/p7/3p4/3P4/5N2/PPP1PPPP/RNBQKB1R w KQkq - acd 42; ce 27; acs 2089; acn 34816000; pv c4 e6 cxd5 exd5 Nc3 Nf6 Bg5 Be7 e3 h6 Bf4 Bf5 Qb3 Nc6 Qxb7 Nb4 Rc1 O-O Bxc7 Qc8 Qxc8 Rfxc8 Ba5 Nc6 Bb6 Rcb8 Na4 Nb4 a3 Nc2+ Kd1 Nxa3 Bc7 Ng4 Bxb8 Nxf2+ Ke1 Nxh1 Bf4 Nc2+ Kd2 Nb4 Rc7 Bd8 Nb6 Bxc7 Nxa8 Bxf4 exf4 Nf2 Nc7 Nfd3 Bxd3 Nxd3 ; bm c4;
rnbqkbnr/1pp1pppp/p7/3p4/3P4/5N2/PPP1PPPP/RNBQKB1R w KQkq - acd 43; ce 42; acs 4260; acn 69917000; pv g3 Bf5 c4 e6 Bg2 c6 O-O Nd7 Nh4 Bg4 h3 Bh5 cxd5 cxd5 Nc3 Be7 Qb3 Nb6 Nf3 Bxf3 Bxf3 Nf6 e4 O-O Rd1 Rc8 e5 Nfd7 h4 f6 exf6 Nxf6 a4 Rc4 Qa2 Rc6 Bf4 Bb4 Rac1 Nc4 Kg2 Na5 Be2 Qd7 ; bm g3;
rnbqkbnr/1pp1pppp/p7/3p4/3P4/5N2/PPP1PPPP/RNBQKB1R w KQkq - acd 44; ce 23; acs 7200; acn 115387000; pv g3 g6 ; bm g3;

Realistically, this one does not seem nearly as cut and dried, since the top three moves were about even (Bg5, c4, g3) and the games total was only 240 for this position.
LC0 found g3 faster, if it is best.