First Win by Leela Chess Zero against Stockfish dev

[CMCanavessi]

the perception by the community might well be that nothing is happening and that progress has stalled, which is completely understandable.

My perception is that maybe too much is happening at once, and that probably because of that, progress has nearly stalled.

Well, I'm currently testing ID 25 of the 1st test run, and it's already proving to be around 2900. That's impressive, a 2900 network made in 3 days and only 1 million games. Oh, and it's still 64x6 size. I'm just speechless at that. A0 was trained on 44 million games for comparison, and the 1st (and bugged) Leela run got to around that level of performance after like 320 networks.

[Werewolf]

Well, I'm currently testing ID 25 of the 1st test run, and it's already proving to be around 2900. That's impressive, a 2900 network made in 3 days and only 1 million games. Oh, and it's still 64x6 size. I'm just speechless at that. A0 was trained on 44 million games for comparison, and the 1st (and bugged) Leela run got to around that level of performance after like 320 networks.

Yes that is...remarkable. The next week will be fascinating.

[Ozymandias]

I'm currently testing ID 25 of the 1st test run, and it's already proving to be around 2900. That's impressive, a 2900 network made in 3 days and only 1 million games.

Did you test the old id45? Also around 1 million games.

[CMCanavessi]

ID 45 was around 1500 elo (I tested 44 and got 1496.7). Must account for different software though, which made a huge progress.

[Ozymandias]

That's indeed some difference.

[David Xu]

the perception by the community might well be that nothing is happening and that progress has stalled, which is completely understandable.

My perception is that maybe too much is happening at once, and that probably because of that, progress has nearly stalled.

Well, I'm currently testing ID 25 of the 1st test run, and it's already proving to be around 2900. That's impressive, a 2900 network made in 3 days and only 1 million games. Oh, and it's still 64x6 size. I'm just speechless at that. A0 was trained on 44 million games for comparison, and the 1st (and bugged) Leela run got to around that level of performance after like 320 networks.

A very strong indicator that DeepMind put basically zero effort into optimizing A0 for chess. After all, chess is a simpler game than Go, so theoretically speaking, it makes sense that a smaller net would be able to achieve more in chess than it could in Go. (A fully optimized A0 would presumably have been much stronger than the version of A0 we saw.) Sadly, the bugs in Leela's first "trial run" obscured that fact, making many people think that the 6-block net reached its ceiling prematurely. Looking forward to seeing what the true ceiling of the 64x6 net is!

[shrapnel]

Well, I hope this new Leela is better than the old one at least in the End-game.
In two 2 minute/move games (which took up a lot of time), lc0 let slip the advantage against Komodo and was forced to concede Draws.
lc0 was showing Scores like +3.40 and for once even the opponent Komodo seemed to agree by showing Scores like -1.85 to -1.95 .
Against Stockfish, Komodo would have caved in, but it held on for Draws against lc0.
Disappointing.
The wait just got longer.

[Leo]

Brianfish defeats lczero


1. e4 c5
2. Nf3 d6
3. d4 cxd4
4. Nxd4 Nf6
5. Nc3 Nc6
6. Nxc6 bxc6
7. Be2 e5
8. O-O Be7
9. Qd3 O-O
10. Qg3 Be6
11. f4 exf4
12. Bxf4 Qb6+
13. Kh1 Qxb2
14. Rab1 Qxc2
15. Bh6 g6
16. Bxf8 Bxf8
17. Rxf6 Bg7
18. e5 Bxf6
19. exf6 Bxa2
20. Qe3 Bxb1
21. Qh6 Qxc3
22. Qg7#

[yanquis1972]

i'm only getting 50-70% GPU usage with the testserver net, is that due to the smaller size or changes in the .exe's (https://crem.xyz/lc0/)?

[CMCanavessi]

i'm only getting 50-70% GPU usage with the testserver net, is that due to the smaller size or changes in the .exe's (https://crem.xyz/lc0/)?

How many threads? You need a beefy CPU to feed a fast GPU and make it work at 100%