'Deus X' Unveiled as FAT FRITZ PROJECT...

[Joost Buijs]

Wasn’t he disqualified from something for basically just copying Lc0 and passing it off as his own work?

Did I just dream that up? I have a vague memory of that but might have got that wrong.

You dreamed that bit of libelous BS up. TCEC cooked up some seriously odd rules, though, to prevent third party nets from entering with the lc0 net container.

TCEC rules are fine. TCEC is a competition between different chess engines (by different authors) and not between the same engine with slightly different evaluation weights. And what on earth is a 'net container'?, it sounds like you try do downplay the importance of LC0 by just calling it a container.

[dkappe]

What is "Lc0 net container"?

"DeusX" was Lc0 binary, Lc0 neural network architecture and implementation, trained with Lc0 scripts (so the only thing different is NN weights, it's not even "different neural net" as some people say).
I don't know what "Fat Fritz" is, but I expect it's the same.

Whether it's fine to commercialize is debatable, but what I don't like the most is that what "Fat Fritz" really is is hidden from people paying money for it.

What a beautiful word salad. I suppose your going to claim that Ender, Maddex and the Gyal networks are all just weights as well and really just leela.

Your pretzel logic is breathtaking.

[dkappe]

Wasn’t he disqualified from something for basically just copying Lc0 and passing it off as his own work?

Did I just dream that up? I have a vague memory of that but might have got that wrong.

You dreamed that bit of libelous BS up. TCEC cooked up some seriously odd rules, though, to prevent third party nets from entering with the lc0 net container.

TCEC rules are fine. TCEC is a competition between different chess engines (by different authors) and not between the same engine with slightly different evaluation weights. And what on earth is a 'net container'?, it sounds like you try do downplay the importance of LC0 by just calling it a container.

Maddex could run in lc0, but instead it runs in a different net container: Scorpio.

[Joost Buijs]

Wasn’t he disqualified from something for basically just copying Lc0 and passing it off as his own work?

Did I just dream that up? I have a vague memory of that but might have got that wrong.

You dreamed that bit of libelous BS up. TCEC cooked up some seriously odd rules, though, to prevent third party nets from entering with the lc0 net container.

TCEC rules are fine. TCEC is a competition between different chess engines (by different authors) and not between the same engine with slightly different evaluation weights. And what on earth is a 'net container'?, it sounds like you try do downplay the importance of LC0 by just calling it a container.

Maddex could run in lc0, but instead it runs in a different net container: Scorpio.

I don't know what Maddex is, but I assume it is just weights, and weights don't run at all.

Have you ever written a chess engine? I assume not. Writing a chess engine from scratch takes several months, writing a strong engine takes several years. Training a network by supervised learning is something anyone with decent hardware can do in a couple of days when the software is already there. This is of a totally different order.

[crem]

What is "Lc0 net container"?

"DeusX" was Lc0 binary, Lc0 neural network architecture and implementation, trained with Lc0 scripts (so the only thing different is NN weights, it's not even "different neural net" as some people say).
I don't know what "Fat Fritz" is, but I expect it's the same.

Whether it's fine to commercialize is debatable, but what I don't like the most is that what "Fat Fritz" really is is hidden from people paying money for it.

What a beautiful word salad. I suppose your going to claim that Ender, Maddex and the Gyal networks are all just weights as well and really just leela.

Your pretzel logic is breathtaking.

Ender and Gyal use the same neural network architecture and implementation as Lc0, just differently trained. It's Lc0 project that does improvements to the NN architecture (like SE, or policy masking, or whatever). Effort and importance of training network is surely not trivial, but it's not overwhelming enough either to discard importance of reused parts of Lc0 project.

I don't know what Maddex is, but if it's a part of ScorpioNN, that's completely different story. ScorpioNN has its own NN architecture and implementation, it's clearly different entity from Lc0.

[Guenther]

Wasn’t he disqualified from something for basically just copying Lc0 and passing it off as his own work?

Did I just dream that up? I have a vague memory of that but might have got that wrong.

You dreamed that bit of libelous BS up. TCEC cooked up some seriously odd rules, though, to prevent third party nets from entering with the lc0 net container.

TCEC rules are fine. TCEC is a competition between different chess engines (by different authors) and not between the same engine with slightly different evaluation weights. And what on earth is a 'net container'?, it sounds like you try do downplay the importance of LC0 by just calling it a container.

Maddex could run in lc0, but instead it runs in a different net container: Scorpio.

And what has Scorpio from Daniel to do with a further ripoff from LC0?

[dkappe]

I trained Maddex with a combination SL and RL. It’s a vanilla 256x20 leela style net. Same architecture.

[Guenther]

I trained Maddex with a combination SL and RL. It’s a vanilla 256x20 leela style net. Same architecture.

You are still in a loop. We are talking about what you described as a mere container.
If you don't want to answer fine, but don't try to obfuscate each time you create a pseudo answer.

Actually I think it has not much sense to discuss with you or Silver, you showed already the same behaviour
after you were criticized after DeusX for not acknowleding the effort of LC0.

[dkappe]

Have you ever written a chess engine? I assume not. Writing a chess engine from scratch takes several months, writing a strong engine takes several years. Training a network by supervised learning is something anyone with decent hardware can do in a couple of days when the software is already there. This is of a totally different order.

Yes, I have written a few chess engines, back in the dark ages before null move pruning. If you look around, you can find one of them bundled in a blindfold training tool, and the other one is leela_lite which, despite its name, has nothing to do with lc0. It uses T9 nets and a mcts/puct algorithm implemented in less than 100 lines of python to play chess.

Joost, clearly you’ve never trained a net. Otherwise you’d know that the leela SL tools are very limited (policy in particular is weak) and won’t ever get you to 3000. I’m also pretty sure you’ve never built a nn engine either.

[dkappe]

Actually I think it has not much sense to discuss with you or Silver, you showed already the same behaviour
after you were criticized after DeusX for not acknowleding the effort of LC0.

Why should I have to acknowledge anything about DeusX or lc0? DeusX has nothing to do with me, other than that i’ve also trained alternate and unusual nets (and contributed in small ways to the source code of the project) and feel a great deal of sympathy for A Silver whenever the mean girls turn out in force.