Torch 2 is here

[pohl4711]

I received a copy of Torch 2 just some hours ago from chesscom (thanks to A.Grant!) for testing.
Testrun started. If all works as expected, the result of Torch 2 should go online on my website on monday.

Stay tuned!

[Draude]

Do you know if Torch 2 has a dimension 3072 network? It is very good in other strong engines like Big-SF.

[AndrewGrant]

Do you know if Torch 2 has a dimension 3072 network? It is very good in other strong engines like Big-SF.

I like to look at things from a different perspective, in a very literal sense, so I don't see Torch as mirrored-halfka x 2560, but rather as N x 22528. So in some sense you could say that Torch has a dimension of 22528. Which makes us 733% more powerful at extremely balanced positions strictly in low sample size tests, compared to 3072 users.

[sarona]

+1

Brilliant explanation, Andrew!

[Jouni]

Torch needs only +5 elo to be number 1. in CCRL .

[Chessqueen]

Torch needs only +5 elo to be number 1. in CCRL .

By the time that Torch get 5 elo, Stockfish will get another 5 or more elo, the back wheel on the car will always follow the front, unless the front get rotated to the back

[Uri Blass]

Torch needs only +5 elo to be number 1. in CCRL .

Torch needs more than it.

https://computerchess.org.uk/ccrl/404FRC/

1 Stockfish 16.1 4065 +13 −13 72.9% −165.1 51.8% 2530
2 Torch v2 4038 +14 −14 68.9% −126.4 56.6% 2073

https://computerchess.org.uk/ccrl/Chess324/

1 Stockfish 16.1 3804 +5 −5 73.5% −137.1 49.9% 2800
100.0%
2 Torch v2 3780 +5 −5 69.8% −111.4 51.7% 2800

[CRoberson]

Do you know if Torch 2 has a dimension 3072 network? It is very good in other strong engines like Big-SF.

I like to look at things from a different perspective, in a very literal sense, so I don't see Torch as mirrored-halfka x 2560, but rather as N x 22528. So in some sense you could say that Torch has a dimension of 22528. Which makes us 733% more powerful at extremely balanced positions strictly in low sample size tests, compared to 3072 users.

Neural networks don't work that way. More neurons and dendrites aren't linearly mapped to greater ability.

[AndrewGrant]

Torch needs only +5 elo to be number 1. in CCRL .

Torch needs more than it.

https://computerchess.org.uk/ccrl/404FRC/

1 Stockfish 16.1 4065 +13 −13 72.9% −165.1 51.8% 2530
2 Torch v2 4038 +14 −14 68.9% −126.4 56.6% 2073

https://computerchess.org.uk/ccrl/Chess324/

1 Stockfish 16.1 3804 +5 −5 73.5% −137.1 49.9% 2800
100.0%
2 Torch v2 3780 +5 −5 69.8% −111.4 51.7% 2800

CCRL's other lists are extremely compressed. 5 elo on those lists is right.

[AndrewGrant]

Do you know if Torch 2 has a dimension 3072 network? It is very good in other strong engines like Big-SF.

I like to look at things from a different perspective, in a very literal sense, so I don't see Torch as mirrored-halfka x 2560, but rather as N x 22528. So in some sense you could say that Torch has a dimension of 22528. Which makes us 733% more powerful at extremely balanced positions strictly in low sample size tests, compared to 3072 users.

Neural networks don't work that way. More neurons and dendrites aren't linearly mapped to greater ability.

I'm exceptionally aware that Neural Networks don't work like this.