As someone who tried to create Leela before Leela was there, and then again after Leela was there, I ended up copying Leela and I don't see how anyone would not copy Leela. It is quite standard to copy and then say "we used as a basis and created a lot of custom code around it". It is understandable Chessbase would like to offer something close to the state of the art, but I can hardly see the relevance to most players, top or not. Most pros travel with "gamer laptops" that will run Leela quite well, Leela is the gift that keeps on giving both by being in the public domain and by pushing Stockfish, and top 10 pros, well, they may already have invested in monster hardware far above the typical "rent a cloud".
Fat Fritz, the Chessbase's Alphazero
Sort:
fidetrainerNET wrote:
As someone who tried to create Leela before Leela was there, and then again after Leela was there, I ended up copying Leela and I don't see how anyone would not copy Leela. It is quite standard to copy and then say "we used as a basis and created a lot of custom code around it". It is understandable Chessbase would like to offer something close to the state of the art, but I can hardly see the relevance to most players, top or not. Most pros travel with "gamer laptops" that will run Leela quite well, Leela is the gift that keeps on giving both by being in the public domain and by pushing Stockfish, and top 10 pros, well, they may already have invested in monster hardware far above the typical "rent a cloud".
When you say "copy" do you mean to copy its code or to copy it's playing style?
mooskagh wrote:
drmrboss wrote:
In my opinion, if leela is running on 2x 2080Ti GPU, then she can beat Fat Fritz on their cloud service.
From screenshots it seems they have 2x GTX 2080 (not even GTX 2080ti).
I'm sure on equal hardware Leela would win. But they don't seem to object it, they advertise "Fat Fritz" as some source of opening insights, not as the strongest engine.
Well, welcome crem.
The boss of Leela is here, he is the main developer of Leela. For further questions about Leela, ask him.LoL. 
mooskagh escreveu:
drmrboss wrote:
In my opinion, if leela is running on 2x 2080Ti GPU, then she can beat Fat Fritz on their cloud service.
From screenshots it seems they have 2x GTX 2080 (not even GTX 2080ti).
I'm sure on equal hardware Leela would win. But they don't seem to object it, they advertise "Fat Fritz" as some source of opening insights, not as the strongest engine.
Master, of course it's an assumption, but if a duel between A0 and Lc0 were possible on the same hardware, who do you think would win? Given Leela's current strength and development, who would win? In this equation we could put Google's hardware as hardware.
@barretoff, Lc0 of course.
Just to let you know that Leela on 2x2080Ti GPU is much better than Alpha Zero in 2017. ( Leela surpassed A0 level in 2018 Dec).
drmrboss escreveu:
@barretoff, Lc0 of course.
Just to let you know that Leela on 2x2080Ti GPU is much better than Alpha Zero in 2017. ( Leela surpassed A0 level in 2018 Dec).
Wow! What do you mean stronger? I did not know that!!! man, I can not believe that Lc0 2018 had already surpassed A0 2017. With q base you claim it Drmrboss?
barretoff wrote:
drmrboss escreveu:
@barretoff, Lc0 of course.
Just to let you know that Leela on 2x2080Ti GPU is much better than Alpha Zero in 2017. ( Leela surpassed A0 level in 2018 Dec).
Wow! What do you mean stronger? I did not know that!!! man, I can not believe that Lc0 2018 had already surpassed A0 2017. With q base you claim it Drmrboss?
There are games tested in Leela vs SF8 ( similar speed/ conditions, A0 vs SF8 whereas A0 was running 80knps and SF8 running 88mnps) and got similar or better scores than that A0 vs SF 8.
Additionally, leela today can access endgame tablebase and also has tablebase rescoring in her training that A0 didnot have.
Today Leela is approx 30-50 elo stronger than that Leela in 2018 December.
If you see Leela games in chess.com computer tournment Leela vs Stockfish, you might not be much impressed cos Leela cant beat Stockfish very clearly. But today Stockfish 10+ in chess.com hardware is also +150 elo stronger than SF 8 in 2017.
drmrboss escreveu:
barretoff wrote:
drmrboss escreveu:
@barretoff, Lc0 of course.
Just to let you know that Leela on 2x2080Ti GPU is much better than Alpha Zero in 2017. ( Leela surpassed A0 level in 2018 Dec).
Wow! What do you mean stronger? I did not know that!!! man, I can not believe that Lc0 2018 had already surpassed A0 2017. With q base you claim it Drmrboss?
There are games tested in Leela vs SF8 ( similar speed/ conditions, A0 vs SF8 whereas A0 was running 80knps and SF8 running 88mnps) and got similar or better scores than that A0 vs SF 8.
Additionally, leela today can access endgame tablebase and also has tablebase rescoring in her training that A0 didnot have.
Today Leela is approx 30-50 elo stronger than that Leela in 2018 December.
If you see Leela games in chess.com computer tournment Leela vs Stockfish, you might not be much impressed cos Leela cant beat Stockfish very clearly. But today Stockfish 10+ in chess.com hardware is also +120 elo stronger than SF 8 in 2017.
Thanks for the explanation, drmrboss. Alphazero is a myth, even if you explain to me that it is hard to believe that the monster was overcome by Lc0. Even so, I believe in you.
Regarding the strength of Stockfish 10 dev, do you think the fish will never outperform Leela?
Aug 15, 2019
0
#31
This sounds like Deus X 2, the fact that they were proud of winning an obscure engine tournament is bad news. I'll pay attention when it shows up at CCC or TCEC and gives a run for its money to Lc0.
It is hard to say!
Both Stockfish and Leela are very big community projects.
SF 10+ is already +37 elo than that SF 10, released in Nov 2018.
SF had 1.4 billion games and 50, 000+ patches tested within 6 years of fishtest.
Meanwhile Leela games are also training 500, 000 games per day. This is test 60, it is a much bigger neural network than Alpha Zero and previous Leela( 1 million games per day previously in smaller networks)..
At the moment both engines are very close to hitting the ceiling. There are some flaws in both engines but technically very hard to improve for both. My guess is that Leela is still slightly in favour than SF in cccc and Tcec next 1-2 years. (+20 elo approx). Please remember that Tcec games are 100 games, there is margin of error up to +40 elo, that means underdog can still win in those small sample size tournments.
So, you mean both programs are already close to their maximum possible level? Does this mean that the upper limit of chess is already close to being reached?
About the chess.com hardware, very impressive!
CPUs: 2 x Intel Xeon Platinum 8168 @ 2.70 GHz 33 MB L3
Threads: 90 threads with HT on
16384 MB hash
RAM: 256GB DDR4-2666 ECC Registered RDIMM
SSD: 2x Crucial MX300 (1TB) in RAID1
OS: CentOS 7
GPU: 4x RTX 2080ti (44 GB GPU memory)
CPU: 2x Intel Xeon Gold 6154
Cores: 36 physical
RAM: 96 GB
barretoff wrote:
So, you mean both programs are already close to their maximum possible level? Does this mean that the upper limit of chess is already close to being reached?
About the chess.com hardware, very impressive!
CPUs: 2 x Intel Xeon Platinum 8168 @ 2.70 GHz 33 MB L3
Threads: 90 threads with HT on
16384 MB hash
RAM: 256GB DDR4-2666 ECC Registered RDIMM
SSD: 2x Crucial MX300 (1TB) in RAID1
OS: CentOS 7
GPU: 4x RTX 2080ti (44 GB GPU memory)
CPU: 2x Intel Xeon Gold 6154
Cores: 36 physical
RAM: 96 GB
Technologically close to their maximum level, pretty much , Yes.( That is at least what I feel).
Theoretically for upper limit of chess, ( No).
If you know all possible positions of chess game , it is like 10^43, SF can search like a few trillion positions, 10^12, in OTB time control in chess.com hardware. Meaning SF is searching only 0.00000000000000000000000000000001 % of all possible positions in chess. There could be better chess than today engines if you have trillion times resources for engines. Evaluations are also flawed in engines, they cant still tell 100% correctly in every position they searched.
But top engines have been extremely tuned to get an effective performance even though they are searching a tiny fraction of game tree.
drmrboss escreveu:
barretoff wrote:
So, you mean both programs are already close to their maximum possible level? Does this mean that the upper limit of chess is already close to being reached?
About the chess.com hardware, very impressive!
CPUs: 2 x Intel Xeon Platinum 8168 @ 2.70 GHz 33 MB L3
Threads: 90 threads with HT on
16384 MB hash
RAM: 256GB DDR4-2666 ECC Registered RDIMM
SSD: 2x Crucial MX300 (1TB) in RAID1
OS: CentOS 7
GPU: 4x RTX 2080ti (44 GB GPU memory)
CPU: 2x Intel Xeon Gold 6154
Cores: 36 physical
RAM: 96 GB
Technologically close to their maximum level, pretty much , Yes.( That is at least what I feel).
Theoretically for upper limit of chess, ( No).
If you know all possible positions of chess game , it is like 10^43, SF can search like a few trillion positions, 10^12, in OTB time control in chess.com hardware. Meaning SF is searching only 0.00000000000000000000000000000001 % of all possible positions in chess. There could be better chess than today engines if you have trillion times resources for engines. Evaluations are also flawed in engines, they cant still tell 100% correctly in every position they searched.
But top engines have been extremely tuned to get an effective performance even though they are searching a tiny fraction of game tree.
What about Test 60, what is his current status? Already stronger than test40?
barretoff wrote:
What about Test 60, what is his current status? Already stronger than test40?
It's not very close yet. It'll be a while. The bigger nets take longer to train.
It's doing well so far, though. Right now, I'm watching it take on Laser, and it's roughly breaking even. https://www.twitch.tv/edosani
Fat Fritz and Leelestein are very similar, don't you think?
barretoff wrote:
Fat Fritz and Leelestein are very similar, don't you think?
This is a tricky question.
We need to Categorize two main types of engines
1. Traditional engines ( Stockfish, Komodo, Houdini). Their search is minimax and evaluation is human hand crafted tuning. Those group should have similar style.
2. Neural network engines ( Leela, Fat Friz, Leelenstein). Search is based on MCTS guided by NN and evaluations by NN. These group should have similar style.
Here is a single 1 hour game I've made the other day Fritz 17 vs Fat Fritz:
Despite being less deep and having less kilonodes Fat Fritz ruled the match although having the Black.
Mar 23, 2020
0
#40
OK bummer. I installed Fat Fritz in my new Fritz 17 program. An initial dialogue box opened upon this installation the purpose of which was to set Fat Fritz to optimum levels. Ok well and good. However I have since upgraded my Graphics card (one with Cuda technology) which is supposed to help tremendously Fat Fritz's performance! BAM! I could no longer get the Fat Fritz automatic set up dialogue box that I had gotten initially when I first installed my Fritz 17 program. I uninstalled and reinstalled the fritz 17 program but NO JOY! Still no automatic set up program!! WHAT TO DO? I wasted all that money upgrading my GPU???? I wrote to Chessbase about this but NO response! further in the still available engine parameters dialogue box there is NO mention of 'cuda' Technology as there was in regard to the initial set up. Bummer.
I agree I think Chessbase should have provided this information in their article.
Well, business service wont tell those information about their products.
e.g Fritz 16 is extremely weak vs Stockfish( about -300 elo behind) , but they won't provide CCRL rating!
Looking at the chess.com/ccc and Chessbase hardware, which do you think is the best, most robust?
Technically, we dont know the strength of cloud computing power they are running on fat fritz( dexus) but Leela will be running on consumer graphic cards on pc that everyone has. If they are on equal strength vs Leela, Leela would be better. In 2018, their earlier verison of fat fritz (Dexus) vs Leela, Leela was better as well.
However, if software is too good, hardware is less important.
For example, 64 cores Komodo win 2400 cores Jonny program.
In my opinion, if leela is running on 2x 2080Ti GPU, then she can beat Fat Fritz on their cloud service.
The article says a modern 12-core processor was purchased. On the Intel website, the i9-9920x and i9-7920x with 12 cores. And from chess.com is 2 x Intel Xeon Platinum 8168 @ 2.70 GHz.