Congrats to LCZ team... nothing taken away from them... However, that devbuild included a bug that was in the Master... I didnt expect good results from this cup because of it
@@dannygjk In the search function, a line of code that handles the static eval initialized to zero, when its value was actually 32000. I'm no coding expert, nor am I an expert in Stockfish coding. Using pure logic, I think this made SF think complex positions were more equal than they really were. Not a huge problem with lower rated engines, but not ideal for an engine very close in its playing strength. What made things worse imo is that the default neural net was trained with the bug in place.
@@dannygjk I suppose the fix is "Stockfish/pull/4309" (google this), a problem that could be caught with clang-tidy, for example. Also last week stockfish devs added "-Wconditional-uninitialized", I guess because of this problem. Anyway, the lesson learned, it is nice to remember that there are people behind all these engines and these competitions are between developers, not machines.
Thank you for the video. Playing it through on Chessbase, I see that a user “Condo” followed this Nc3 variation through to the ending with LCZero as early as 14th May 2021
Afaik there has been a few breakthroughs for Lc0 recently, so it closed the gap a tiny bit. But the fun thing is, SF is trained on Lc0 data, thus if Lc0 improves, SF automatically improves too, tho perhaps not by the same amount.
This is an early version of LC0 running on a "transformer network" architecture, similar to what GPT-3 is based on. It is currently even slower than than the "old" lc0, but gained already significantly in strength.
28. Re4+ is a draw...aparently the correct move was 28. Re1+ and white wouldve continued to have his +0.7 advantage and get a winning endgame really slowly but surely
It was the cup so only 16 round not 100 round like the super final I am sure if you cut the super final anywhere and took only 16 round you could find sequence where Leela was ahead.
This is probably not true for the last several superfinals in which Stockfish prevailed- Stockfish rarely lost in opening matchups against Leela. The rounds were either drawn with each winning the "good" side of the opening and losing the bad side, or the engines drew both sides. When there was a decisive result, it was the Fish winning a game, or both, and Leela not winning a game or only one. You could probably put together 16 rounds where the two drew 15 and Leela won a single round.
Keep telling yourself that Bullet and blitz stockfish might be faster... Probably it calculates faster... But may be Leela has better positional play given enough time
@@soulsbourne lol stfu. Where is the proof of that? Sf is miles stronger in tournments played and leela discord test. Because it lost one elimination round, means jack. A sample of some 15 games does not show anything.
Stockfish has dominated LC0 for a few years but it's because stockfish can maneuver tirelessly with a space advantage until the opponent misses a deep tactical idea. LC0 is not as good at that kind of struggle. In positional and strategic struggles, LC0 is the equal of stockfish, and they are not identical, so LC0 can still win a crucial game once in awhile.
@@soumyaroy5455 I have no idea but it's doing fairly well as it is. I hope that its programmers don't become discouraged when it loses a match to stockfish. After all, Leela is stronger than every other chess program out there, including KomodoDragon, which is also very strong.
LCzero first 1-2 movies as a white were a waste of movies, it looked confused in every game. It looked it takes couple of movies to asset situation and only than start playing. She just wastes initiative as a white. I was wondering is it a bug or it is just slower. Did not watched last tournament.
Bullet and blitz stockfish might be faster... Probably its engine/AI is more efficient and so calculates faster... But may be Leela has better positional play given enough time may be because it thinks deeper ...
@@Jozarovschesschannel usually tests for strength are done using much faster time control, much weaker hardware and much more games, but importantly the elo uncertainty is kept low. It is not feasible to run many games with such good hardware and long time control, thus these events are geared more towards entertainment. The normal TCEC event does have the 100 game format, but honestly that is still SSS.
A very interesting game. Good to see another machine winning once in a while. Thanks Josip and keep up the good work.
I didn't expect this at all tbh! Great job bro as always!
Thanks
C46 Steinitz Defense / Kindergarden Variation
Its only over,when jozarov says its over....
Can you add more videos of leela from the TCEC? Thank You Jozarov!
Sure, it’s a deal
Stockfish 15.1: Am i a joke to you?
Congrats to LCZ team... nothing taken away from them...
However, that devbuild included a bug that was in the Master... I didnt expect good results from this cup because of it
Describe the "bug" please.
@@dannygjk In the search function, a line of code that handles the static eval initialized to zero, when its value was actually 32000.
I'm no coding expert, nor am I an expert in Stockfish coding. Using pure logic, I think this made SF think complex positions were more equal than they really were. Not a huge problem with lower rated engines, but not ideal for an engine very close in its playing strength.
What made things worse imo is that the default neural net was trained with the bug in place.
@@chrisdavis2161 What do you mean, that it was supposed to be initialized to 32000 instead of 0?
@@dannygjk I suppose the fix is "Stockfish/pull/4309" (google this), a problem that could be caught with clang-tidy, for example. Also last week stockfish devs added "-Wconditional-uninitialized", I guess because of this problem. Anyway, the lesson learned, it is nice to remember that there are people behind all these engines and these competitions are between developers, not machines.
my stockfish at home immediately saw the Nd5xc3 as best move and i dont even have a good pc...
It's only +2=13-1 for Leela. Good result, but hardly dethroning. We need 2*50 games Superfinal to decide.
Thank you for the video. Playing it through on Chessbase, I see that a user “Condo” followed this Nc3 variation through to the ending with LCZero as early as 14th May 2021
Oh didn’t know that. In the lichess database the were no games in this line in human chess
@@Jozarovschesschannel A smoking gun has been located! Condo vs. LCZero 05/14/2021
It was about time for Leela's team to overcome Stockfish again. In a year or two, Stockfish will again be on top.
Surely letting the h6 pawn block it's own rooks is better than targeting it?
Afaik there has been a few breakthroughs for Lc0 recently, so it closed the gap a tiny bit. But the fun thing is, SF is trained on Lc0 data, thus if Lc0 improves, SF automatically improves too, tho perhaps not by the same amount.
This is an early version of LC0 running on a "transformer network" architecture, similar to what GPT-3 is based on. It is currently even slower than than the "old" lc0, but gained already significantly in strength.
wait i think the opening messed up stockfish...i believe stockfish had it
I dont understand the same leela or updated one??
28. Re4+ is a draw...aparently the correct move was 28. Re1+ and white wouldve continued to have his +0.7 advantage and get a winning endgame really slowly but surely
It was the cup so only 16 round not 100 round like the super final I am sure if you cut the super final anywhere and took only 16 round you could find sequence where Leela was ahead.
Exactly
You Are right. Never thought about IT Like this.
This is probably not true for the last several superfinals in which Stockfish prevailed- Stockfish rarely lost in opening matchups against Leela. The rounds were either drawn with each winning the "good" side of the opening and losing the bad side, or the engines drew both sides. When there was a decisive result, it was the Fish winning a game, or both, and Leela not winning a game or only one. You could probably put together 16 rounds where the two drew 15 and Leela won a single round.
Wow, didn't expected that Nxc3 🔥🔥🔥
You should show the reverse to show how leela won and where SF deviated.
I will in the next couple of days for sure
Jozarov, maybe SF needs to subscribe to your channel from now on.
Thx 4 posting it.
I like how the eval bar is cheering for Stockfish the whole time.
Eval probably being done by stockfish though.
@@gm2407 Yeah I was implying that Stockfish "commentary" is biased.
Cups is elimination round. Leela got lucky with opening. Sf is miles stronger still
Keep telling yourself that
Bullet and blitz stockfish might be faster... Probably it calculates faster... But may be Leela has better positional play given enough time
@@soulsbourne lol stfu. Where is the proof of that? Sf is miles stronger in tournments played and leela discord test. Because it lost one elimination round, means jack. A sample of some 15 games does not show anything.
Stockfish has dominated LC0 for a few years but it's because stockfish can maneuver tirelessly with a space advantage until the opponent misses a deep tactical idea. LC0 is not as good at that kind of struggle. In positional and strategic struggles, LC0 is the equal of stockfish, and they are not identical, so LC0 can still win a crucial game once in awhile.
How can Leela be improved?
@@soumyaroy5455 I have no idea but it's doing fairly well as it is. I hope that its programmers don't become discouraged when it loses a match to stockfish. After all, Leela is stronger than every other chess program out there, including KomodoDragon, which is also very strong.
Here is hoping you don't have the same 'premature' problem in other areas. The Fish is now winning. 🙂
LCzero first 1-2 movies as a white were a waste of movies, it looked confused in every game. It looked it takes couple of movies to asset situation and only than start playing. She just wastes initiative as a white. I was wondering is it a bug or it is just slower.
Did not watched last tournament.
this was not Stockfish 15.1 it was a 16 dev. thx for the nice video
Bullet and blitz stockfish might be faster... Probably its engine/AI is more efficient and so calculates faster... But may be Leela has better positional play given enough time may be because it thinks deeper ...
awesome game! thanks
Most welcome
It's SSS (small sample size), so this doesn't say much about raw strength, but this is a spectacle none the less, congrats to the Lc0 team.
The finals should be decided with more games, but as you said its still a great effort by the Lc0 team
@@Jozarovschesschannel usually tests for strength are done using much faster time control, much weaker hardware and much more games, but importantly the elo uncertainty is kept low. It is not feasible to run many games with such good hardware and long time control, thus these events are geared more towards entertainment. The normal TCEC event does have the 100 game format, but honestly that is still SSS.
Even 100 is a small sample size. The crucial factor is the opening lines chosen for the match.
Well well we'll.
Are you happy ? 😁
@@Jozarovschesschannel yeah. I am leela fan I must admit. Great game Great Bishop and pawns by leela.
I dont mind SF is dethroned and will be, Lc0 also I like and Dragon 3rd
sf is not dethroned. It lost an elimination round. Sf is faaaar stronger than leela.
I thought deepblue was stonger.