Something goes wrong with lc0 since yesterday?

[mesilikas]

Ok thanks again! I will download 9044 from testserver and 437 from mainserver to try it on my old dual core pc...

[Laskos]

Ok thanks again! I will download 9044 from testserver and 437 from mainserver to try it on my old dual core pc...

I think ID395 from mainserver is still the best against AB engines, although I haven't' checked lately. Meanwhile the new small 6x64 net ID9050 from testserver achieved 81.0/200 in the gauntlet, by far the best result for any 6x64 nets, putting it close to later large 15x192 nets of mainserver in this gauntlet on my GTX 1060 GPU. Maybe some people can inform the devs that this run with the small net is the best so far, by a wide margin.

[mesilikas]

Super! Ok 9050 instead of 9044. Do you believe that someday 6x64 net will reach a level of 3000 elo on a cpu? Maybe this will be done when Leela reach 3800 on a 1060 gpu. Maybe a dream maybe not...

[Uri Blass]

Super! Ok 9050 instead of 9044. Do you believe that someday 6x64 net will reach a level of 3000 elo on a cpu? Maybe this will be done when Leela reach 3800 on a 1060 gpu. Maybe a dream maybe not...

rating at what time control?

If the 6x64 scales better than A-B engines maybe it can reach 3000 elo even today when the time control is big enough.

[Laskos]

Super! Ok 9050 instead of 9044. Do you believe that someday 6x64 net will reach a level of 3000 elo on a cpu? Maybe this will be done when Leela reach 3800 on a 1060 gpu. Maybe a dream maybe not...

As Uri said, you have to specify time control too, all these Leelas are scaling better than AB engines.
As of now, on my GTX 1060, small net ID9050 with lc0 engine is 3100+ CCRL 40/4' Elo level. I guess on one core it is about 2500+ CCRL 40/4' Elo level. Going to long time controls, add 100-200 Elo points to these ratings. It is unclear to me by now whether on CPU the best 6x64 net from testserver is stronger than ID395 15x192 net from mainserver. On my GPU, bignet ID395 is still stronger.

[mesilikas]

Leela reminds me the first heroic days of computer chess. At that time we were buying Chessmaster 2100 to run it on Amiga... Standard time control 40 moves in 120 minutes and suddenly somewhere in the fortieth fifteenth move chessmaster give away the game with a terrible move. Leela of course plays at a complete different level. She can win every amateur player even on a low end hardware. But i believe that if we put her to face a strong AB engine even if we give her a lot of time at some point her great weakness ( tactics ) will appear. I refer to my own experience with her. Impressive chess and suddenly an oversight. Last week i left her (id 48 from testserver networks if i remember well) to face Rodent ii at 15 15 time control( okay not a lot of time) for 2 games ( of course not a test, just for the feeling...). In first game at some point according to the viewpoint of both engines was winning and lost after a terrible blunder and in second lost a totally drawing endgame ( over optimistic view is her second weakness but at the same time this weakness gives the ability to play attractive beautiful chess). The same with a weaker one engine : Cdrill build 4 (around 1900-2000 elo). Out of 4 games won 3 and lost 1 ( i expected to win all games). Of course what matters most is that we are entering a new era
for computer chess...

[Laskos]

Leela reminds me the first heroic days of computer chess. At that time we were buying Chessmaster 2100 to run it on Amiga... Standard time control 40 moves in 120 minutes and suddenly somewhere in the fortieth fifteenth move chessmaster give away the game with a terrible move. Leela of course plays at a complete different level. She can win every amateur player even on a low end hardware. But i believe that if we put her to face a strong AB engine even if we give her a lot of time at some point her great weakness ( tactics ) will appear. I refer to my own experience with her. Impressive chess and suddenly an oversight. Last week i left her (id 48 from testserver networks if i remember well) to face Rodent ii at 15 15 time control( okay not a lot of time) for 2 games ( of course not a test, just for the feeling...). In first game at some point according to the viewpoint of both engines was winning and lost after a terrible blunder and in second lost a totally drawing endgame ( over optimistic view is her second weakness but at the same time this weakness gives the ability to play attractive beautiful chess). The same with a weaker one engine : Cdrill build 4 (around 1900-2000 elo). Out of 4 games won 3 and lost 1 ( i expected to win all games). Of course what matters most is that we are entering a new era
for computer chess...

Yes, completely agree. Leela from the early stages was an unusual engine. I used to play the leea chess bot online on "easy" mode, so nodes=1 or no search at all. It is wonderful for a patzer like me. She is playing some superb moves, only to miss a fork or a combination. One can learn a lot from it, in a similar way one learns from a careless strong human. I called Leela on "Normal" or "Hard" mode as "drunken GM", seeing games against 2100-2300 rated humans. Completely careless brilliant positional play, often disregarding material, only to be botched by a tactical super-blunder.

Meanwhile this last run with 6x64 net is excellent, ID9065 scored 93.5/200 in my gauntlet against AB engines, the highest by far for any of 6x64 nets, and very close to top 15x192 nets from the main branch. I recommend you using this 6x64 net, on CPU I guess it's better by now than the bignets from the main branch. On my GTX 1060 GPU, this ID9065 is close to 3200 Elo level in CCRL 40/4' conditions. Which is amazing for such a small net size. Maybe some folks can make devs know that their latest 6x64 nets are excellent, of unseen in the past runs levels.

[JJJ]

You might try id 440 now Kai , seems good in selfplay, maybe the best now ?

[Laskos]

You might try id 440 now Kai , seems good in selfplay, maybe the best now ?

Yes, now running, intermediate results show that ID440 is close to ID395 in the gauntlet against AB engines. ID395 scored 103.0/200, let's see what ID440 will score (in about one-two hours). The small net from testserver ID9065 scored 93.5/200, which is already close to the best bignets from the mainserver.

[Werewolf]

You might try id 440 now Kai , seems good in selfplay, maybe the best now ?

Yes, now running, intermediate results show that ID440 is close to ID395 in the gauntlet against AB engines. ID395 scored 103.0/200, let's see what ID440 will score (in about one-two hours). The small net from testserver ID9065 scored 93.5/200, which is already close to the best bignets from the mainserver.

ID 9071 now at 3606 elo self play.
Presumably this has the potential to be a lot stronger than all the others once the net gets bigger.