It looks like the big 3 in TCEC are LC0, SF and Allie now!

[Laskos]

24 games between 256 threads (192 cores ) vs Leela! Leela is leading in TCEC!!

It looks like 10 cores, 100 cores or 1000 cores SF doesnt matter much against Leela , Leela on good GPU will still kick positional weakness of Stockfish!

Update of 100 games against Lco vs SF!

SF marginally won with +13 -12 =75 score (50.5 vs 49.5).
Facts. (Excluding margin of error in 100 games)

1. It was the same score as Season 14 final ( Leela T30 vs 44 cores Stockfish)

2. Stockfish has x4 speed hardware.

3. x4 speed boost Stockfish +30 elo against T40 Leela (there were not much elo improvement in both SF and Leela sides, although different nets were used ).

4. Beyond 40 threads, doubling Stockfish is not so great against Leela. (x4 speed= 30 elo means, doubling speed gives Stockfish +7.5 elo).

What time control was that last one at TCEC? Doesn't look like tournament TC if the match ended so quickly.

[Nay Lin Tun]

Time control was 4x shorter.
This one is 30 min +10 secs.

[Laskos]

Time control was 4x shorter.
This one is 30 min +10 secs.

Ah, so don't compare it to SuFi TC conditions. I am prepared to bet on Leela against RedFish at tournament TC. Corroborated from data I saw, Leela T40 scales better than SF from rapid to LTC on large hardware. Things like CPUCT have to be adjusted to LTC, folks operating Leela seem to know that.

[Nay Lin Tun]

And now Stockfish vs Leelenstein in ccc9 final.

My bet is like 105 vs 95 Leelenstein win.

[Laskos]

And now Stockfish vs Leelenstein in ccc9 final.

My bet is like 105 vs 95 Leelenstein win.

I do not know how this Leeelenstein behaves. Is it a hybrid or just a SL helped net?

[Nay Lin Tun]

https://www.patreon.com/jjosh, Jjosh website.

Leela developer/tester/enthusiast. Gauging interest on current projects. Working on supervised learning now with "Leelenstein", combined with some other lc0 branch things.

Leelenstein started out mostly CCRL games, with some 40b and T30 games. Currently uses CCRL games + more T30 games + SGDR + GGT + SE, but started with the same weights as the previous to retain some information from those other games.
(There have been several different SL runs with some nice nets picked from each, and some slightly different datasets.)

Trade Penalty is also recommended (and used at CCCC in current tournament).

LS did well against SF as well.

[Laskos]

https://www.patreon.com/jjosh, Jjosh website.

Leela developer/tester/enthusiast. Gauging interest on current projects. Working on supervised learning now with "Leelenstein", combined with some other lc0 branch things.

Leelenstein started out mostly CCRL games, with some 40b and T30 games. Currently uses CCRL games + more T30 games + SGDR + GGT + SE, but started with the same weights as the previous to retain some information from those other games.
(There have been several different SL runs with some nice nets picked from each, and some slightly different datasets.)

Trade Penalty is also recommended (and used at CCCC in current tournament).

LS did well against SF as well.

Yes, I see, it's as far as I understand not a hybrid performing well against weaker engines and not so well against top dogs. Might be a very good net overall (does it have its own engine or just different settings of the same?). Seeing these results, it seems the favorite against SF. Nice work.

[Damir]

This Jjosh guy can not be trusted. I registered and supported his website by donating..... and he sent me download link to Leelenstein 1.0 that is not working

[Dicaste]

This Jjosh guy can not be trusted. I registered and supported his website by donating..... and he sent me download link to Leelenstein 1.0 that is not working

Before you made some accusations for love of god make sure you set-up engine correctly. It's working fine you just couldn't do it.

[Damir]

Yes it is working fine my ass….