Torch - Request to testers

[chrisw]

It’s alleged that not only is Torch closed source but also that it won’t be released as a UCI engine (free or otherwise).

That being the case it means the chess.Com developers can test against us, play mass games against us and use those games for training NNs. We can’t do the same in reverse. This gives Torch a huge unreasonable and outrageous advantage in the top engine space.

Chess.Com want every unfair advantage under the sun and then take advantage of all publicity.

I urge testing groups to NOT test and NOT report any results of this giant commercial cheat unless and until Chess.Com makes it available in the usual way. Freeze them out unless and until they get reasonable.

[Modern Times]

I don't think it is 100% confirmed yet that Torch will in fact be given to ratings lists. If I were chess.com I would not do that. To my mind, the last thing they would want for example is for ratings lists to show that it is the #2 or #3 engine (if indeed that were to be the case). For similar reasons, it would not be wise for them to supply to TCEC either. I think that they need to keep it exclusive to chess.com and promote it from there. They don't need the ratings lists, and ratings lists might actually be detrimental to their strategy.

Their strategy will probably evolve. Of course if it is demonstrably and without doubt the number one engine by a decent margin, they may decide to release commercially. Who knows.

You talk about unfair advantages - that is life. Does Stockfish not have an unfair advantage over 99% of other engines with the thousands of CPU cores they have available on fishtest ? It is what it is.

[JacquesRW]

That being the case it means the chess.Com developers can test against us, play mass games against us and use those games for training NNs. We can’t do the same in reverse. This gives Torch a huge unreasonable and outrageous advantage in the top engine space.

You would think if this was the best way to get data then Stockfish would be using Lichess games to train NNs. They don't however, they use data from engines (e.g. Stockfish's self generated data and Leela data).

[chrisw]

By “us” I mean the engines of us engine programmers. Torch can do RL against known opposition, we can’t against it. That’s a huge advantage.

[AndrewGrant]

Very strange post. But everyone is welcome to do whatever they like.

[connor_mcmonigle]

By “us” I mean the engines of us engine programmers. Torch can do RL against known opposition, we can’t against it. That’s a huge advantage.

What? Private engines are hardly a new thing and I'd challenge the notion that this gives some huge advantage in practice. Up until very recently, your engine was private as well, correct?

[Chessqueen]

That being the case it means the chess.Com developers can test against us, play mass games against us and use those games for training NNs. We can’t do the same in reverse. This gives Torch a huge unreasonable and outrageous advantage in the top engine space.

You would think if this was the best way to get data then Stockfish would be using Lichess games to train NNs. They don't however, they use data from engines (e.g. Stockfish's self generated data and Leela data).

We all know that most of us will use the latest Stockfish to see if Torch can beat it or how close it is from being equal to Stockfish

[CornfedForever]

Very strange post.

Not really...when you realize who posted it.

[RubiChess]

I urge testing groups to NOT test and NOT report any results...

Who would have thought? Chris W. is the chessqueen.

[Ras]

What? Private engines are hardly a new thing

Correct - at the moment, I count three private engines in the CCRL: Stoofvlees II, Deep Sjeng 3.6a16, and Onno 0.12.0. There's no reason to treat Torch any differently. It may be worth discussing to remove private engines altogether, but that would be a completely different discussion because it's not about private engines having any alleged and supposedly unfair advantage, but because their test results cannot be verified independently.