mmt Vs. Ovyron (G4 D5 BG2)

[Ovyron]

But are these lines, by computer standards, "highly tactical"? It doesn't look like a position that's especially tactical.

Um, they're "normally tactical"? At least they're tactical enough that Leela will not see black's winning move until you insert it into her.

What does it being "very happy to play into" mean? Evals of -1.0 or scores of 0.0?

At this point Leela would be very happy with just some -1.32 variation that she could find (and I'd be surprised if evals came back from the dead and mmt claimed he's still seeing an score of over -1.32 - where "over" is in 0 direction.)

The worst I've seen was Leela down at -0.60 while Stockfish was at -4.00, and Stockfish was jumping this high instantly, Leela sees the key move and realizes she's lost, and I haven't seen a single position where Stockfish claims some 2.00 advantage and Leela is right it can be saved. I'm still checking them because the worst that could happen at this point is mmt finding the only variation where Stockfish's 2.00 advantage claim is bogus and Leela has the drawing moves in her PV.

This is just Leela at some 500k nodes for the entire game, but with mmt starting the game ahead with 800 million nodes already, this might prove Leela wasn't designed for this job.

[jp]

But are these lines, by computer standards, "highly tactical"? It doesn't look like a position that's especially tactical.

Um, they're "normally tactical"? At least they're tactical enough that Leela will not see black's winning move until you insert it into her.

Can you keep some examples of these lines Leela doesn't see but agrees change everything when you feed them into it and post afterwards, so we can look at them?

I don't think by computer standards these lines are highly tactical. Highly tactical positions at a minimum must make superGMs feel they've totally lost control if they appear OTB. I doubt if superGMs had the current position they'd feel that way.

What does it being "very happy to play into" mean? Evals of -1.0 or scores of 0.0?

At this point Leela would be very happy with just some -1.32 variation that she could find

Okay, but then the difference between non-spoonfed Leela (-1.3) and spoonfed Leela (say -2.0, though you actually said -1.6) is "only" 0.7 (or as little as 0.3). I'm not convinced that this is then evidence that Leela's problems are "tactical" here.

[SheikhYerbouti]

So if Leela's relatively bad playing from g4 what happens with SF self-play from ply 2? More likely a draw?

[Zenmastur]

1. g4 d5 2. c4 Bxg4 3. Bg2 c6 4. Qb3 e6 5. Qxb7 Nd7 6. Nc3 Ne7 7. cxd5 exd5 8. d4 Rb8 9. Qa6 Rb6 10. Qd3 Ng6 11. h3 Be6 12. Nf3 Bd6 13. h4 h5 14. b3 Nf6 15. Bg5

[d]3qk2r/p4pp1/1rpbbnn1/3p2Bp/3P3P/1PNQ1N2/P3PPB1/R3K2R b KQk - 2 15

Komodo at depth 46 with MultiPV = 3 had

15. Na4 (-0.89)
15. Kf1 (-1.12)
15. Bg5 (-1.13)

I had 15.Na4 at -1.71 and this was the best move.

[Alayan]

Ethereal at depth 37 now suggests Re8 or Rb8, both get very close evals around -1.00 and have switch first place a few times while going deeper.

[zullil]

1. g4 d5 2. c4 Bxg4 3. Bg2 c6 4. Qb3 e6 5. Qxb7 Nd7 6. Nc3 Ne7 7. cxd5 exd5 8. d4 Rb8 9. Qa6 Rb6 10. Qd3 Ng6 11. h3 Be6 12. Nf3 Bd6 13. h4 h5 14. b3 Nf6 15. Bg5 O-O 16. e3

[d]3q1rk1/p4pp1/1rpbbnn1/3p2Bp/3P3P/1PNQPN2/P4PB1/R3K2R b KQ - 0 16

Komodo at depth 42 had:

16. e3 (-1.14)
16. Na4 (-1.21)
16. Rc1 (-1.26)

[Ovyron]

Can you keep some examples of these lines Leela doesn't see but agrees change everything when you feed them into it and post afterwards, so we can look at them?

Just match Leela v Stockfish and see for yourself.

But here's such a thing:

1.g4 d5 2.c4 Bxg4 3.Bg2 c6 4.Qb3 e6 5.cxd5 exd5 6.Qxb7 Nd7 7.Nc3 Ne7 8.d4 Rb8 9.Qa6 Rb6 10.Qd3 Ng6 11.h3 Be6 12.Nf3 Bd6 13.h4 h5 14.b3 O-O 15.Ng5 Qf6 16.Bd2 Rb4 17.e3 Bf5 18.Qe2 Bg4 19.f3 Bf5 20.Qf2 Bd3 21.Bf1 Nf4 22.Rc1 Re8 23.Kd1 Bf5 24.Na4 Rxa4 25.bxa4 Nb6 26.Rxc6 Bd7 27.Rxd6 Qxd6 28.Qh2 Bxa4+ 29.Ke1 Nc4 30.Bxc4 dxc4 31.Ne4 Rxe4 32.fxe4 Nd3+ 33.Ke2 Qa6 34.Bc3 Bd7 35.Ra1 Qg6 36.Kd2 Qxe4 37.Rg1

[d]6k1/p2b1pp1/8/7p/2pPq2P/2BnP3/P2K3Q/6R1 b - -

Leela is happy to play into this position, because she misses the tactical move 37...Ne1!! and white is forced to play Rxg6+ into a completely lost game!

I'm not going to pretend that I understand what is going on there, because honestly, I have no idea, but most variations that Leela recommends are like this. In the others her eval dies a slower death (going slowly from the -1.30s to the -1.40s.)

Okay, but then the difference between non-spoonfed Leela (-1.3) and spoonfed Leela (say -2.0, though you actually said -1.6) is "only" 0.7 (or as little as 0.3). I'm not convinced that this is then evidence that Leela's problems are "tactical" here.

Except that ALL the variations I've seen on this game where Leela showed a score of -1.60 or worse have been COMPLETELY LOST for white, so it could as well be showing -9.99, because those are positions where black has string of moves that mate in X.

[Ovyron]

So if Leela's relatively bad playing from g4 what happens with SF self-play from ply 2? More likely a draw?

Exactly, when matched Stockfish can't beat itself from the black side, that's the main reason I thought I could draw with 1.g4. Even on my losing positions against Havey, Stockfish wasn't being able to win from the black side (and that's why I took so long to resign.)

But when you guide Stockfish as black against any white defense, it's a very different story, now it's Stockfish as white that can't defend the position. And when you try to guide its defense as white, it seems useless, everything seems losing, so far, at least for 2.g5 and 2.Bg2, and they looked better than the alternatives (so much that instead of trying to defend the white side with another move, I switched to attacking white's last line of defense.)

Is 1.g4 lost??

[Ovyron]

I had 15.Na4 at -1.71 and this was the best move.

For the record, I had 15.Na4 also as best and this is what I was expecting.

[mmt]

I had 15.Na4 at -1.71 and this was the best move.

For the record, I had 15.Na4 also as best and this is what I was expecting.

LC0's scores and deeper analysis made a difference here. I'd have to run the analysis again to show you exactly what happened and the lines are still valid for the game so it will have to wait a bit.