Has Silver written any code for "his" ZeusX?

[Chris Whittington]

According to other chess forums, ZeusX is a Leela clone trained on a Megabase of human games. How that counts, in any real sense of the term as, "developed by" is beyond comprehension.

Here's commentary from other forum ...................

Milos wrote: ↑Tue Jul 31, 2018 2:03 am
Graham Banks wrote: ↑Tue Jul 31, 2018 1:33 am
Nay Lin Tun wrote: ↑Sun Jul 29, 2018 3:58 pm
Graham Banks wrote: ↑Sun Jul 29, 2018 6:30 am

What about Tencent company?
https://www.chess.com/forum/view/genera ... -season-13
http://www.chessdom.com/deus-x-the-nn-c ... rt-silver/
My God, what a load of BS. I would think that submitting main engine and then it's tweaked clone is something that wouldn't be allowed by TCEC rules, but apparently it it going to be.

For once I have to agree, that's pathetic and ridiculous at the same time.
But that was already clear from the first cryptic announcements though.
The most work done was to find a new name...

ASilverheute um 06:08 Uhr
It is a 100% unique NN built entirely from human games, and only human games, from Megabase 2018. No self-play, or anything else. It is 10x128 for any curious.
That was the philosophy behind it.

From other forum:
Armstrongheute um 06:09 Uhr
Sorry for my naive question . Not gone through the video . If Deus in agreement with the open source policies of Leela ?

ASilverheute um 06:09 Uhr
What open source policies? I am not selling it.

Now you can start with dozens of Brainfish clones (with different self constructed 'cerebellum' books) and dozens of LeelaChess clones with
self-trained NNs. Brave new world.

[Chris Whittington]

More from other forum ...

Re: TCEC season 13, 2 NN engines will be participating, Leela and Deus X

• Quote

Post by Milos » Tue Jul 31, 2018 11:17 am

Uri Blass wrote: ↑

Tue Jul 31, 2018 8:34 am

I think that there is a difference because I guess that Brainfish derivatives are similiar in the choice of moves when LeelaChess derivatives with a new net are not similiar in the choice of moves but I am not sure about it and it may be interesting to see the difference in move choice between LC0 and Deus and compare.

What the hell are you talking about???
Do you know how many lines of code of LC0 Albert changed? My guess is less than what a typical cloner of SF does.
He has identical net, identical number of inputs, levels, filters, using exactly the same backend, same search.
What he did is identical to taking SF, changing PST numbers and evaluation bonuses (like what Tsvetkov was doing) and claiming the new engine that is btw. probably hundreds of Elo weaker than LC0. You really think that changing only some numbers randomly in SF eval wouldn't make it play totally different moves?
And after all there is question of legality, since TCEC is public event and that "Deus X" competes on it, if he doesn't publish the source code he is in direct violation of not only GPL licence but also NVIDIA licence.
If you really want to see innovative NN engine look at Scorpio MCTS+NN of Daniel. Totally different net, backend, search, everything.

[Dietrich Kappe]

What does "He has identical net" mean?

[Chris Whittington]

Ask Milos on Talkchess forum, it's his post.

I imagine when he wrote: He has identical net, identical number of inputs, levels, filters, using exactly the same backend, same search.

eg same net structure as lc0, different weights, but ask him

[Dietrich Kappe]

Allegedly it's a 128x10 net. Looking at the history of Leela's 128x10 nets, that was in the middle of the underpromotion bug. ID125 played in S12, I think. I wouldn't want to pick any of those 128x10 nets as the basis for my engine.

Odd, not sure what he means.

[Chris Whittington]

I don't know, but the Discord suggests Albert Silver was asking for and receiving Discord technical help in compiling LC0 only six weeks ago. Since he is not a programmer and six weeks is a very short time, it seems unlikely he did anything other than build another set of weights. 

[gvergh...@gmail.com]

From the interview Albert says (quote)
"As to whether it will be commercial, who can say...
If I come out with something that's really spectacular and that outdoes Alphazero, very possibly.
Right now at the moment, it's a private project."

Since February, the gpu's are crowd-sourced.
From all the knowledge attained from those 22 million games, comes private commercial ventures.

Doesn't make any sense, especially to those who run their hardware 24/7, thinking that they're contributing to a public venture.

Any idea how much a new net will cost ?

[Dietrich Kappe]

Sure. I'm the developer who fixed supervised learning (and added support for FEN headers for my own purposes :-). So I helped him occasionally. I also helped him edit the tensorflow code (it's about 1000 lines of mostly boilerplate python, with some custom code for reading the training data). But he was getting deep into CLR, so he lost me with some of those modifications.

I'd really like to clean that code up with Keras, but life's too short.

So, to say that he didn't modify or write code is probably not correct.

[Chris Whittington]

the entity "ZeusX" is, I read, no more and no less than:
a standard LC0 executable, (C) Leela Authors with an Albert Silver created weights file.

MCTS, UCT, noise at root, everything, the entire executable program, EVERYTHING, is (C) Leela authors, who, judging by their talkchess posts this evening, are distinctly unhappy with this clone entry into TCEC.

[Dietrich Kappe]

It’s hard to explain the rise in the main net of late. Yes, it’s adopted q-pawn opening over the English, but that can’t explain it all.

[Kasuha]

It all boils down to the question what's the essence and what's the boilerplate. As long as we agree that strength of Leela is mainly given by what the network has learned, the code itself and the implementation falls to the boilerplate level along with the libraries the project uses, the interpreter, the system, and the hardware it all runs on. Obviously, Leela authors spent a lot of time writing their code because it's not readily available elsewhere. But if NNs become widespread, I expect graphic cards might be coming with NN learning and MCTS tools straight from card manufacturers eventually.

The Leela vs DeusX comparison is essentially "supervised learning vs self-learning". Assuming DeusX is indeed only trained on human games. There's plenty of patterns for supervised learning available, but new ones are not coming in at too great rate. Self-learning has potential to surpass that, given just enough time.

[Albert Silver]

It is as I stated. I will add that while building a NN is quite doable for many, building a very good one is really hard. The number of stalls, resyars, tweaks and adjustments along the way, would test the patience of even the most resolute. When I see the challenges and difficulties the Leela devs have had getting the net to emulate A0's successs, none are more sympathetic than myself.

[Chris Whittington]

That's very true, building a good neural network is really hard.

But you haven't actually BUILT a neural net, have you? You trained a set of weights using a neural net whose architecture has already been built by others (the hard part).

You did the easy part. Used code written by others, to train a set of network weights, of size (determined by others), which fit into a neural net (built and tested by others) using a set of chess games (played by others) and then run in a program (built by others).

You just processed some data using tools, kits, software, design, architecture all done by other people.

Then you announce "it's really hard and it's my neural net". Nonsense, it's not your neural network at all. Weights != Neural network. Weights = weights.

[Björn Holzhauer]

As someone who is trying to do the same thing, I can assure you that it is an extremely difficult task requiring many new ideas, solutions and a lot of judgement calls. I have great respect for anyone who manages to get a decent quality NN whether by self-play, supervised learning or any other method.

[Deep Blender]

The neural network from Leela is very simple! What's far more complicated is the whole architecture around it, including the Monte-Carlo tree search, but also the actual training. Training a neural network like the one from Leela with a different approach is still a massive effort.

[Chris Whittington]

Well, it's a massive effort for the GPUs, but feeding a few meg human quality games down a tube that's already been prepared for you beforehand and then naming the weights file afterwards doesn't involve a great deal more than pressing the ENTER key. Leela authors did all the creative slog here didn't they?

[Albert Silver]

Yes, it is clear you are deeply versed in the process.

[Dietrich Kappe]

Hehe,

Cris as Condoleeza Rice: "It wasn't a plan, just a list of things to do."

It wasn't a neural network, just a collection of weights.

If it's so easy, why don't you train one up?

[Galaga]

What a meek reply to seriously asked questions, here on the forum!

Chris has to prove nothing, it is your turn to give more detailed information.
Github or a homepage could be places to start.

[Albert Silver]

Not really. Chris's background and understanding of the process is pretty much zilch. If it were as easy as he claims, you'd think there would be a flood of such NNs, no? Is it because no one could be bothered because it was so 'easy', or could it be because he doesn't know what he is talking about? Logic has a clear answer to that question.

[Chris Whittington]

Hahaha, Kappa. Since 1984/5 I've been familiar with the software functionality on Neural Nets via Igor Aleksander at Brunel, who if I remember it right had a little program that ran on a Commodore Pet or something similar. My company's Go programmer was a NN expert (this since about 1986/7. In 2003-6 I wrote NN systems from the ground up, myself, in C, very successfully for Backgammon, rather feebly for RK endings, and very unsuccessfully for Stock Market Predictions. Since November 2017 I figured getting up to speed with the big changes that have taken place in the last few years and wrote from the ground up, NNs to play TicTacToe, then Connect 4, then Othello, then Chess for which I have some new ideas. Right now I am doing Tensorflow courses with view to some development. One son does neuroscience brain-computer research and a son-in-law is in charge of teams of AI guys putting in machine learning applications across various companies. We talk. A lot. About this stuff. I make a point of trying to understand the very complex new neuroscience brain science material, difficult, but I try.

Meanwhile, as far as I can tell, you are something of an oaf. Why do you continually append actual shit, feces icons, after my posts on the Discord? My posts about chess, nets and so on? I would say it is because you act like a little baby. Yet you claim to be 51. Unbelievable. Why do you look me up on Google, discover I used to write chess programs and then lie that you bought such and such chess program, which according to you is "shit" and you want your money back? Idiot and liar and thief actually, isn't it?

Now, let's be quite clear about net technology. There's the ARCHITECTURE of the neuron layers and then there's the WEIGHTS that weigh the connections between the neurons. If you TRAIN the weights, but didn't create the architecture, then you can claim those are "your" weights", but you can't claim it is "your" architecture when some else built and tested the architecture beforehand. And the Leela network architecture is for sure non-trivial. Therefore since the network is the architecture and the weights together, you can't train weights and then say this is "my" network. And no amount of your "shit" feces icons or personal insults or claims you think my old chess program is shit and you want your money back, as a substitute for your being unable to argue properly will make it any different. Someone who trains another set of weights using an already established network architecture, owns only the weights. And weights do not equal network.

[Dietrich Kappe]

Criz,

so there's no excuse for your lack of understanding, unless of course you're more of a code monkey that just cuts and pastes.

Let's see your "collection of weights." I'm really looking forward to testing them.

[Chris Whittington]

Ah Mr Silver!

Perhaps you would be so good as to explain why it is you've been trolling me, quite unprovoked, since you arrived at the Discord? I hadn't had any contact with you, for how long, ten years? Yet you arrive, as a newbie, eager to get into the social group, no? It's a common tactic, try and be "in" by gratuitously attacking an "outsider", one who wasn't even there, btw ...

ASilver 21/5/2108

Haha. I thought you were linking to Chris Whittington. Who is half troll, but a much longer CC resume.

Chris Whittington hadn't been mentioned by anybody, nor was the topic related to him, but you thought you'ld bring his name in, with an insult. Why that? I call it "trying to ingratiate yourself in a group by attacking a solitary outsider." Who wasn't even there.

Then when they give you no reaction on the Discord, you volunteer:

ASilver 21/5/2018

CW is so beloved that when he, and several other prominent CC programmers founded CCC back in 1995 (I joined 1996), he managed to piss them off so bad he got banned. he then came back 10 years later when I was a mod, and after got banned AGAIN (not by me).

still no reaction, so you just play ASilver the wannabe in-group guy:

ASilver 21/5/2018

Internet and trolls were a new thing, this was 1995 mind you. And yes Exa, that is exactly what happened

You are pretty much a wannabe, are you not? I'm an easy guy to single out on your little trip, yes, is that it?

Oh, and by the way, my background and understanding is not too bad, stretches back a way, and I work at keeping up. We are all struggling a little in such a fast moving environment, but you can read about what I do and don't do elsewhere, I'm sure. It's also a good place for pretend experts to feel important for a little while as everyone around them struggles to understand everything. Does it feel good for you, Albert?!

[Deep Blender]

Training a neural network with a given dataset and even a given architecture is more involved than pressing enter. There are still papers appearing which suggest techniques to optimize this very computationally expensive process. There is a lot more to it than pressing enter.

Also, let me repeat what I have written before: The actual architecture of Leela's neural network is not complicated! They are using a simple ResNet architecture, just as DeepMind suggested. This is one of the standard architectures that is being used in deep learning, alongside with a few others. Implementing them within a project is usually not very complicated. Everything around them is usually much more involved.

Your statements are simply wrong.

[Warren D Smith]

To Albert Silver:

Hi, I'd email you directly but do not know address.

It seems plausible to me that lc0 would be stronger if in addition to (0) selfplay games it

also trained on (1) human games and (2) computer-computer games and (3) perfect play 

from huge endgame databases and (4) games like 0&1&2 ENHANCED by using, e.g, stockfish

to detect forced wins (and/or perhaps also forced draws).

  Re (4) what I am trying to say is, there are a subset of game positions that Stockfish 

can play perfectly, and know it is playing perfectly, e.g. it spots a mate in 3.  One could use that 

perfect info to enhance the NN training.  Re 1&2, strong human and computer game databases

probably are more strongly played than leela training games (which as far as I understand

are run at very fast time controls) hence training with them might able able to supply

something leela would otherwise never see.  

   As far as I know you are using (1).  Re (2) see

  http://www.computerchess.org.uk/ccrl/4040/games.html

  http://privat.bahnhof.se/wb432434/welcome.htm

to download about a million computer-computer games.

Warren D. Smith   warren.wds at gmail

[Dietrich Kappe]

Clis got banned from CCC twice? I had no idea. Given the tone there (and now on TalkChess), that’s a real feat.

[Tom Idermark]

Doing _hard_ work does not constitute doing _original_ work. Ask any painter that forgers Michelangelo or Andy Warhol paintings - it’s _really_ difficult. Compare to someone downloading an open source NN program that predicts future NYSE stock prices who adopts it to predict prices on instead, say, the Frankfurt exchange by adjusting dollar to euro rates, different data sources, cleanup, normalisation etc. If the original program is open source, you can do that but you can’t claim it is _original_ work because it is _derived_ work. Readup on GPL v2 or v3.

I’m sure you spent a good amount of time adjusting LC0 to take training data from a large dataset of best human play, cleaning it, normalising it to fit LZ0 input layer requirements and what not - and probably tweaked many hyper parameters (even net configuration) to adjust to a smaller training set. That is not easy (but not that hard for a programmer either - just takes time) - well done! But if you based your code on LZ0 it is NOT original work. To anyone that says that the NN config/hyper parameters and tweaking is the _hard_ / _important_ part I can only say - ”Then write the easy parts yourself and do not take code from LC0 - really, it is OK, just don't take other peoples work and claim its yours.”

Anyway, unless you very carefully isolated those modification you made you are obliged to submit any changes back to the Leila and LZ0 projects (some argue that it is not even enough to isolate your own work unless LZ0 has an LGPL license (the Library GPL)).

That said, I think it would be very interesting to see a NN trained on human play in TCEC and see some (KC?) analysis of play. Just wished it had been done in a less narcissistic fashion. Why isn't the code published on GitHub or anywhere else? And that TCEC intervju gave me the creeps, is this the next self appointed bitcoin tycoon?

PS. I’ve been doing NN design, research, tweaking and hair-tearing, frustrating work for decades. But always been very clear about giving credits were it’s due - alas, we are all standing on the shoulders of those before us. Leechers, be aware!

[Dietrich Kappe]

Tom,

there’s a common misconception about the GPL and open source licenses in general. They generally require that you distribute the modifications only if you distribute the software in source or binary form. If you only modify an open source markdown converter, for example, and distribute pdf’s produced with the modified converter, you are not obligated to distribute the modified source.

Otherwise no quibbles with your post.

[PrinceZappa]

If you're talking the guy giving the interview, that's Cato. Infamous partly for being author of some TCEC opening books, as well as for wearing that chess sweater.  

[Tom Idermark]

Just a bystander but I think he got hammered a bit to hard, we are all adults right? I defend his position (mostly) but perhaps not always the way he is defending it. You seem rational (and influential) so why not sprinkle some water on this fire, it has enough gas/petrol burning already.

Cheers

[Dietrich Kappe]

Well, i defended him on discord, but that got out of control too.

I didn’t understand at the time who Chris Witt. was or why he was so mad at Albert. Now it all makes sense.

I think cooler heads have prevailed and that the project leaders have come to a peaceful and sensible agreement.

[Tom Idermark]

Agree, and I should have been clear that I was anticipating the release of a future commercial product as indicated in the TCEC intervju. I've met both Linus Torvalds and Richard Stallman (many years ago, I admit) on that and other matters so I think I have a reasonable idea of what the idea behind open source is (or was) ... Much of the hard core debate is probably settled now but there was a time when even the output of Bison was discussed ...

[Tom Idermark]

Haha, thats the guy :-)

[Rudolf Posch]

I hope the  Statements by Deus X and Leela Chess Zero authors make the situation clearer. Do we calm down and hope for a fair battle between Leela Chess and Deus X.

Good luck Leela! (see also my Leela blog entry).

Am Dienstag, 31. Juli 2018 11:53:15 UTC+2 schrieb Chris Whittington:

According to other chess forums, ZeusX is a Leela clone trained on a Megabase of human games. How that counts, in any real sense of the term as, "developed by" is beyond comprehension.
..

[Deep Blender]

Out of curiosity, who made such a claim? I couldn't find it, but it is also likely that you are referring to my comments. As it is written too broadly, it is impossible to tell...

[emma...@gmail.com]

Sounds about as original as changing the numerical values of the pst etc in SF.  However much hard work it may have been, or how difficult (or not) it was.

Which does not detract from the value of what has been done or interest in the final result.

Perhaps we will see lots of these project in TCEC now.  Along with lots of SF clones, with new pst that were derived in a 'very clever' and 'difficult' way.

On Tuesday, July 31, 2018 at 10:53:15 AM UTC+1, Chris Whittington wrote:

According to other chess forums, ZeusX is a Leela clone trained on a Megabase of human games. How that counts, in any real sense of the term as, "developed by" is beyond comprehension.

Here's commentary from other forum ...................

Milos wrote: ↑Tue Jul 31, 2018 2:03 am
Graham Banks wrote: ↑Tue Jul 31, 2018 1:33 am
Nay Lin Tun wrote: ↑Sun Jul 29, 2018 3:58 pm
Graham Banks wrote: ↑Sun Jul 29, 2018 6:30 am

What about Tencent company?
https://www.chess.com/forum/view/genera ... -season-13
http://www.chessdom.com/deus-x-the-nn-c ... rt-silver/
My God, what a load of BS. I would think that submitting main engine and then it's tweaked clone is something that wouldn't be allowed by TCEC rules, but apparently it it going to be.

For once I have to agree, that's pathetic and ridiculous at the same time.
But that was already clear from the first cryptic announcements though.
The most work done was to find a new name...

ASilverheute um 06:08 Uhr
It is a 100% unique NN built entirely from human games, and only human games, from Megabase 2018. No self-play, or anything else. It is 10x128 for any curious.
That was the philosophy behind it.

From other forum:
Armstrongheute um 06:09 Uhr
Sorry for my naive question . Not gone through the video . If Deus in agreement with the open source policies of Leela ?

ASilverheute um 06:09 Uhr
What open source policies? I am not selling it.

Now you can start with dozens of Brainfish clones (with different self constructed 'cerebellum' books) and dozens of LeelaChess clones with
self-trained NNs. Brave new world.

[Tom Idermark]

I was absolutely not referring to your post but more on a "feeling" that the forum, in general - no one singled out, does not appreciate (enough) the coding challenge in making a stable and efficient chess playing agent. The TensorFlow network declarative statements constitute a small (but very important) part of the whole coding challenge. Wouldn't be better if anyone that had a good idea improved on what is available and shared that?

[Albert Silver]

Absolutely, except I did not change the code, I built a NN from the code. I used new material and ideas in it and the end-result is the participating entry in TCEC. That's all there is to it. As to the rest, well, CW and I have a long history, so I would take anything said by either on each other with a huge grain of salt.

--
You received this message because you are subscribed to a topic in the Google Groups "LCZero" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/lczero/vGdNYW-Ou58/unsubscribe.
To unsubscribe from this group and all its topics, send an email to lczero+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/827d2f95-fc7e-48e1-b0cc-9a3fef27176b%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Deep Blender]

Comparing Leela's neural network to a pst is a little far fetched. If you play Leela with an initial neural network that was not or barely trained, it plays pretty much the worst possible chess. Now, if you take the exact same Leela code and swap only the neural network with a highly trained one out, you get brilliant results.

Improvements in a pst are barely the main reason why engines like Stockfish are getting better, while the neural network can and often will be the main reason why Leela is improving.

[Tom Idermark]

Great! But please realise that some people (is that me?) finds your previous secretive stance a bit ”disturbing.” This happens all the time - you want to protect your IP (which you can claim) and some of the Leela team members wants to ensure that their IP (which is free) is not ”abused”. All well and understood but please, be VERY clear about what parts is yours and what’s inherited.

I wish you and Deux-X good luck in the TCEC!

[Chris Whittington]

I only ever speak truth. As you know perfectly well. 

As we can see from the timed and dated Discord, you however speak falsified history of no relevance at all, attacking me, for no reason at all, I've not interacted with you in any way for over ten years, your only possible reason is to try to get yourself into an 'IN'-group by attacking an OUTSIDER (who just happened to be me). It's a known bad behaviour, called Poisoning the well (or attempting to poison the well) is a type of informal logical fallacy where irrelevant adverse information about a target is preemptively presented to an audience, with the intention of discrediting or ridiculing something that the target person is about to say.

Pretty dumb, wasn't it Albert? Better would have been apologise for the plagiarism in 2003 or whenever that was, called in some credit on your correct call on the Vas Rybka issue, and I'ld probably never have bothered with drilling down to exactly define the meaning of "my neural net that I developed" and ignored it all. But .

Albert wrote:

Absolutely, except I did not change the code, No code changes, yes, agreed

I built a NN from the code. No, you trained a set of weights which when loaded into an existing net architecture can be referred to in total as a Neural Network, "your weights", Leela architecture

I used new material Yes, you trained on a different set of games (human ones)

and ideas in it What does "ideas" mean? "In" what? You said you made no code changes, LC0 applies the same chessic inputs to the same net architecture by definition. What ideas? I call BS here, sorry 'bout that.

the end-result is the participating entry in TCEC. No, the end result of what you did is a set of weights. Leela + this set of weights is participating in TCEC, somehow. Scorpio would have been a better alternative. I have my suspicions what this is actually about, and I think you know what those are. Time will tell.

[Albert Silver]

Chris,

I don't see any purpose or future in a dialogue between us, so you can have all the last words you want, and be comforted in the knowledge I will not reply. If you ever have anything genuinely constructive to discuss with me, you have my email.

Peace,

Albert

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/77d22084-cdd6-41e5-89ae-ceb77995f8ac%40googlegroups.com.

[Albert Silver]

Actually, the secrecy at first was more along the lines of not knowing if I was even going to play, even after I had been accepted in TCEC. The same day I was told there was a good chance there would be no GPU still. They had no sys admin, it would require technicalities to be resolved, etc. If this were confirmed, no GPU that is, I would withdraw, since the idea was to enjoy a competition, but not crippled. Why bother announcing the author of the NN if it wasn't going to play? While this was being resolved, the rumors started becoming absolutely comical, with allusions to a Chinese firm, a hoax, a publicity stunt, and more. When the GPU was indeed resolved, I just let it roll, as the name would be cleared when all the candidates were announced anyhow. I told a few members in private in Discord, who told me they had already strongly suspected it was me, since my work on an NN was hardly a secret.

[Edward Panek]

I understand the need to argue but can you take this to private email or another forum? Its really distracting and makes some of us uncomfortable reading it.

[Cezary Wagner]

I think talk about supervised and self learning sometimes is strange - Leela is trained with data from games - DeusX is trained with data from games - which data from games is better and which is not supervised training.

Leela is exploring data by playing games ...

Human games :) - What if I play with Leela it is human game or half human game or quarter human game - it is hard to judge I think it is just impossible judge since no such border in reality - maybe exists in mind but not on board.

No idea what is human games and ai games - you can say there is woman and man games too or child and adult ... :)

W dniu środa, 1 sierpnia 2018 14:14:20 UTC+2 użytkownik Kasuha napisał:

It all boils down to the question what's the essence and what's the boilerplate. As long as we agree that strength of Leela is mainly given by what the network has learned, the code itself and the implementation falls to the boilerplate level along with the libraries the project uses, the interpreter, the system, and the hardware it all runs on. Obviously, Leela authors spent a lot of time writing their code because it's not readily available elsewhere. But if NNs become widespread, I expect graphic cards might be coming with NN learning and MCTS tools straight from card manufacturers eventually.

The Leela vs DeusX comparison is essentially "supervised learning vs self-learning". Assuming DeusX is indeed only trained on human games. There's plenty of patterns for supervised learning available, but new ones are not coming in at too great rate. Self-learning has potential to surpass that, given just enough time.

[Cezary Wagner]

I have no idea what is problem if someone do DeusX with secret NN.

Maybe should be license to only do Open NN on Leela if someone not want Closed NN on Leela but there is no such limitation now.

[Tom Idermark]

Wait a minute, if everybody had that same view of everything then we (humans) would never made it through the out of Africa challenge. I encourage you to instead listen to the debate - read, ponder and learn. Nobody is biting you, right? What I would like to know from Albert is:

1) How much of the source code is LC0? Percentage wise.

2) What is your FIDE rating (as of yesterday)?

3) What database was used to train the NN?

4) If any changes was made to the NN, what?

5) What is the name of your girlfriend?

I think that is enough for now :-)

[Chris Whittington]

Albert,

That would be fine if you weren't also engaged in gratuitous "Poisoning the Well" activity (against me again, of course) on other forums where I don't even post, and haven't done for years.

Albert Silver writes on Talkchess, yesterday afternoon, 

Thanks.
I found some information on the leela forum site.

If you mean Chris Whittington's fountain of wisdom, I would question the use of the word 'information'.

following your earlier Poisoning the Well" activity last May on the Discord (already documented, and actually much more serious).

I don't take "peace offers" from you when you simultaneously are spreading poison around the internet, thank you.

Apologise properly for the plagiarism of 2003 (or whenever), for making gratuitous, unprovoked insults and continued Well Poisoning, and admit your "ownership" is weights only, with Leela Authors owing the NN Architecture, and I'll consider it.

You must have been feeling very frightened the big bad wolf was going to come and call you out on your planned BS, I guess? Is just gratuitous insults, personal references, with what in mind? Bigging up Albert it would seem. Your claim is also well debunked on Talkchess, since I last looked:

Post by IQ » Thu Aug 02, 2018 10:04 am

I am late to the party and never would have expected to fully agree with Milos. 

The snowflakes from the lc0 team have been conned. Albert Silver pulled a fast-one on them and now tries to lull them into compliance with all his sweet talk. The video that was released is absolutely outrageous and shows Albert Silvers true colors. He did not write the tools, nor the search, and did not participate in the work and research that went on when lc0 was still leela zero, and even before that when it originated from leela go. And now he wants to be seen as some kind of AI-expert? Did coding an engine come down to changing a couple of numbers in a scrip written by somebody else, setting a couple of filters in megabase 2018 and letting the computer run for a couple of hours? And then having the audacity to secretly enter a competition, stealing the main teams efficient implementaion of search, NN query, and uci code? How does he think DeusX would do without search? The same logic would allow anybody to enter SF clones with their own individual clop tuned values. And the myth that different NNs play sooooooooooooooooo differently is just a myth. In fact the competition should be even more furious, but as TCEC gains from the increased attention they have no incentive to adhere to their own rules. 

This is like Hamilton stealing Vettels Ferrari for the upcoming GP, calling Vettel from his cell-phone while driving of the Ferrari lot requesting some tuning tipps and telling him its all just a big misunderstanding. The lc0 team is much to nice to realize that this is a well planned publicity stunt. Hats off to Albert Silver for getting away with this! The added fineprint in the chessdom page hardly makes a dent and whenever Albert Silver tells this story at his villians anonymous group they all break into laughter.

The same evil laugh I hear at the end of the video when he says:
""As to whether it will be commercial, who can say... If I come out with something that's really spectacular and that outdoes Alphazero, very possibly. Right now at the moment, it's a private project."

Last edited by IQ on Thu Aug 02, 2018 10:15 am, edited 1 time in total.

I can only agree with the above.

My opinion? That video was a pre-planned opening shot for ChessBase to distribute Leela only with their own mega-hyped-super-duper-Albert-net (or weights as they are better known), calling it by a new name "ZeusX", "developed by ASilver" bla-di-bla and generally spreading as much BS as possible to convince the public to part with some money whilst simultaneously being fed BS hype articles as to "ownership, development, new NN, human style" and all the other BS we'll be reading. The video was to launch the BS, TCEC is designed to be the launch party, much extra BS will be produced about "on-going AI development, new ideas bla bla" and the Leela Authors will have been taken as the mugs who actually do all the creative work and get minimal credit. That's it.

Rights reserved to challenge BS and lies, whenever, wherever.

[Alexander Lyashuk]

Not touching the question whether Lc0 + alternative nets should be allowed at TCEC.

I know that that DeusX confusion is resolved now, but just to state my opinion for the record:

It is clear to me that failing to acknowledge usage of LCZero in the interview and announcement is not an oversight of some kind, but a deliberate attempt to hide the fact that it's derived from LCZero work, and to exaggerate an own share of work in the "DeusX" project. Simply speaking, it was more cheating than misunderstanding.

ASilver keeps telling that he described everything "perfectly clear" to Anton (TCEC admin), but given that I saw that interview, I strongly suspect that it was not so perfectly clear. It's easy to misrepresent the importance of different parts of the engine when talking to a non-specialist, and it seems to me that it was the case. I'm sure that if TCEC admins had any doubts, they would ask LCZero devs for an opinion.

There is also no doubt that ASilver should have notified LCZero devs somehow that he's going to submit Lc0+alternative net to TCEC as a separate entry. Again it seems to me that it was an explicit decision not to tell rather that something ASilver just forgot to think of.

That reason alone (trying to mislead TCEC administration and fans, not because of "no clones rule") would be enough to disqualify the engine, but as it would be bad PR for TCEC (because of that "mystery NN" hype), and because many people will be curious to see how they compare, it's actually better to keep it in the competition.

I do appreciate ASilver's effort to recover relationship between him and the LCZero community and I think we should continue to collaborate (i.e. ASilver did run lots of CLOPs which were very helpful and Lc0's default parameters are from ASilver's tunings), but I'd just like to say that what happened is not "okay".

You received this message because you are subscribed to the Google Groups "LCZero" group.
To unsubscribe from this group and stop receiving emails from it, send an email to lczero+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/4a279bba-4ef6-417d-ba0a-e4951ccb882b%40googlegroups.com.

[Albert Silver]

On Thu, Aug 2, 2018, 09:52 Alexander Lyashuk <moos...@gmail.com> wrote:

Not touching the question whether Lc0 + alternative nets should be allowed at TCEC.

I know that that DeusX confusion is resolved now, but just to state my opinion for the record:

It is clear to me that failing to acknowledge usage of LCZero in the interview and announcement is not an oversight of some kind, but a deliberate attempt to hide the fact that it's derived from LCZero work, and to exaggerate an own share of work in the "DeusX" project. Simply speaking, it was more cheating than misunderstanding.

Both the interviewer and Anton knew without a question of a doubt that the neural network was my work and that it would be running within lc0 ,which is not. This was explained explicitly and it was explained also that the neural network did not use any code or training material with in it that was from the lczero project. However to run, it could only run within the lc0 binary. I'm not sure how much clearer that can be. Anton then, I will assume, asked around or did his own homework to find out whether this was acceptable. And there was no misunderstanding either from the interviewer's part since the discussion was on the neural network itself. Again I'm not sure how that is cheating or even a misunderstanding. But to each his own.

ASilver keeps telling that he described everything "perfectly clear" to Anton (TCEC admin), but given that I saw that interview, I strongly suspect that it was not so perfectly clear.

Instead of 'strongly suspecting' why not simply ask him? Ask him what I told him and ask him whether any part of it, in hindsight, was not crystal clear. It would seem a far more honest approach than throwing around more unnecessary suspicion. If he declares that somehow he had not really understood and that my description, which was awfully simple, was misleading then so be it, I am the fink. But otherwise I will hope you withdraw this idea that somehow I hoodwinked him.

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/CACGjmsJbqgS1F-Keg-yTz-qSzjzey-aBNTFVdd-jkTwzPD9HFw%40mail.gmail.com.

[Curious]

Dkappe, a blog post on all the different parts of "leela" would probably help clarify some of the discussion. What is lc0, what do tensorflow do, all the other scripts, a database somewhere etc. A good old architectual overview.

(Getting even a midsized software company to make this documentation is hard enough, maybe asking a small opensource project to do it is a bit much)

[Alexander Lyashuk]

Okay, as you seem to think that you did everything right and it's "not even a misunderstanding", it would be interesting to understand at which exactly level our understanding of the situation differ.

I've written a post on chesstalk, asking for your vision for a few questions. I didn't ask anything that I "suspect" there (i.e. conversations between you and TCEC admin), only questions that I could observe myself and see myself as wrong.

Could you please respond to that message? Link: http://talkchess.com/forum3/viewtopic.php?f=2&t=68094&p=769966#p769966

Moving conversation there, because here it's already a bit too heated and personal.

Thank you in advance.

[Stephen Frost]

I don't have any vested interest, one way or another, but the way things were done is self-evidently not okay.

Disappointed in the deceptive behaviour.  But that's life.  Nice guys often finish last.

Not sure why the devs decided to continue with TCEC regardless; if it were me, I would have withdrawn Leela.

[Albert Silver]

Alexander,

I have pretty much said all I have to say on the matter, though I have no control over whether you accept my statements or not. I do not see any constructive outcome for any parties in pursuing a public argument on this, so I must decline this invitation. Should you have anything in private you want to talk about, you have my email. 

Peace,

Albert

[John Sidles]

Over on the TalkChess forum, Uri Blass expressed a view held by many: 

"The heart of the program is if then commands and not numbers that you change and I think that is absurd to call Deus a different engine."

To borrow a phrase from Tolkien, "the world has changed" ... the world of chess-programming, that is ... neural nets have changed it ... such that to retain their viability, institutions like TCEC can no longer embrace the newly-wrong principle that "the heart of a program is if-then commands".

As a concrete example, the typesetting community associated to TeX/LaTeX hosts a large set of public-domain fonts (the so-called "LaTeX Font Catalog") that differ solely in "numbers that the font-designer changes" ... and it would be absurd indeed, to claim that these fonts were all the same ... when the creative element of the fonts resides chiefly in the adjustment of "mere numbers" ... namely, the numbers that weight the splines and kernings of which the fonts are composed.

As with fonts, so with neural nets: vastly different nets can differ solely by "numbers that you change".  A good introduction to the vastly different programming techniques that are associated to neural nets is Michael Nielsen's free on-line book "Neural Networks and Deep Learning".

In summary:

• Many more chess programmers will creatively travel that path that Albert Silver has pioneered, and
  
• Institutions like TCEC do well to welcome and encourage these new NN pioneers, and
• old-guard programmers are well-advised to accept and adapt to this new creative reality.

[Chris Whittington]

Nobody is objecting to training another set of weights on different data. The Open Source nature of the project encourages it. It's fine. We all accept the nature of the new reality of Deep Learning. Nobody is being a dinosaur. Okay? Clear?

The objections come partly from the dishonesty and hype and naming and claiming authorship and developer status and and and that accompanied the announcement via cringe-inducing video interview that managed to avoid asking any pertinent questions.

Albert has gradually pulled back from hyped-up position at the beginning and it's now clear that "DeusX" consists of LC0.exe (C) Leela Authors and basically written by Crem (Alexander on this thread) plus a weight.txt file produced by Albert using Lc0 tools, basically Albert re-trained a net from scratch using human games. Over at LeelaChess they are retraining nets from scratch all the time, it's an established process for which the tools can be gotten from Leela Github and so on. Albert needed some help, so he asked devs at the Discord to help him. They helped him. Crem is obviously not happy, but the ZeusX entry seems to be going ahead anyway, despite.

Second objection is about "similarity of cloned engines", whether or not Albert has permission, whether or not he behaved badly and so on, regardless of all that, whether LC0+new weights is an okay entry given that LC0+Official weights is already entered. Because, TCEC rules, such as they are, are supposedly not to allow clone entries of Stockfish (for example).

Third objection is that nobody involved has had he time or space to work out a "rule" for this type of situation. It's been something of a coup de main. That objection centres around not allowing entry this time, but probably/possibly allowing it next season after reflective discussion.

Hope that's clear and fair summary.

Myself, I am just surprised that Albert doesn't see that the position of Crem (you can read it here and elsewhere) doesn't result in him (Albert) withdrawing the entry and re-working what he has done into something sufficiently "okay" for the next TCEC season, which I think is in about two months. Respecting the view of the author who did all of the work, or 95% or whatever it is, seems actually the most important point. But,as far as we know, this ZeusX is going ahead regardless.

[Chris Whittington]

You misrepresent Uri Blass's posting by omission. What he actually wrote was:

The heart of the program is if then commands and not numbers that you change and I think that is absurd to call Deus a different engine.

Unfortunately TCEC write misleading information:
http://www.chessdom.com/tcec-season-13- ... f-the-nns/
"For the first time ever the TCEC competition is going to see two NN engines competing. "

Instead of it it was better to say something like:
"For the first time ever the TCEC is going to see the same engine twice with the only difference that it get different weights that is equivalent to different value of many parameters like the value of pieces and value of positional things that engine has in their evaluation function for a normal engine."

Uri is saying two things. 

1. Just retraining the weights from scratch is insufficient to claim "New Engine". 

2. He proposes the text for a realistic and honest description of the "differences".

Crem has asked the following (rhetorical) questions, which Albert has significantly declined to answer. Crem is the main author of LC0. And not dismissable as an 'old-guard' programmer, obviously. I doubt there are any 'old-guard' programmers, actually, we are all very interested in NN technology as far I can tell. Some are using it, although it may be depressing for those with a lot of work invested in AB, nobody is trying to be Luddite, which I think you are suggesting.

It surprises me that Albert can't read the situation and pull his ZeusX from TCEC until next season, having addressed the issues in the meantime; also allowing everybody else to thrash out some sensible "rules". for this.

1. Do you agree that in "DeusX engine" entry submitted to TCEC, total amount of human effort inherited from LCZero and Leela Zero project, is greater than total amount of human effort that you personally contributed?

Reminder: LCZero effort includes, in particular, engine, including MCTS implementation, batching support, cudnn evaluation, movegen, uci support, optimizations and debugging; training script, both for supervised and unsupervised data; and time spent tuning and trying all that, e.g. network and training parameters, underspecified in papers from DeepMind.


2. Do you agree that given the amount of effort inherited from LCZero, and given that LCZero itself participates in TCEC too as a separate entity, it would be fair to announce which parts of your TCEC entry are a direct reuse of parts of LCZero project and how exactly they are different?


3. Why didn't you discuss possibility to use of Lc0 engine for your TCEC entry with LCZero developers? LCZero participates as a separate entity on it's own, and surely you cannot submit the engine that you didn't author without permission from the authors.


4. Why didn't you mention in the interview that you used and are using the code from LCZero project? You had lots of opportunities to do that.


5. You did not write a chess engine. Why didn't you ask TCEC admins to correct the title "Deus X – the NN chess engine by Albert Silver" of this announcement: http://www.chessdom.com/deus-x-the-nn-c ... rt-silver/ and a similar sentence in the article itself?

On Friday, 3 August 2018 01:31:40 UTC+2, John Sidles wrote:

[LittleWarrior]

I understand the anger about lc0 binary being used and in the spirit of TCEC IS the same engine but, it really isn't the same engine since the knowledge is totally different. Personally, I want to see them both and really don't care about the feuding. My only concern is whether the engine will be released and will there be a charge from the bulk of someone else's work.

[Jesse Jordache]

Claims that it's not a new engine because the heart of a Neural Net is the data informing it are understandable, but there are three problems:

1) The first is the old reductio ad absurdum, which is imagining 600 "Brand new NNs!" in the TCEC competition because someone has tweaked this or that input hyperparam.  Eventually it may come down to that, but on the maiden flight of Leela, it's Leela entering 600 different times.

2) More importantly, it's incredibly uncool to do: write a NN engine from, at the very least Leela boilerplate code, with the help of Leela devs, then going behind the back of said devs and entering it in a competition as your own work.  Remarks like Silver to the effect of "I can't control your (dev's) opinion so there's no further point in me arguing this publicly" are just willfully obtuse, like the whole thing is Alexander's issue.  He should of added "Thanks for writing my engine though".

3) The open source world relies on general decorum that separates "fair use" from just being sleazy.

[J Weston]

Perhaps Silver's previous deceptions make me over-cynical, but how is it possible for him to acquire the CPU resources to train a random-weighted LC0 engine (or other typical initial sets of weights) to the point of being competitive with the current state of LC0 and other conventional chess engines?   I had read he was asking for help in compiling LC0 just 2 months ago -- is that really enough time to train "his" NN?

Since he will not disclose his code or his NN weights, how do we know he didn't just run some post-optimization training with human games on a late-stage LC0 set of weights?  We have his "word" about all of the innovative ideas he has applied?   If he never discloses the internals, would we ever know?

I think this issue needs to be cleaned up before the next TCEC, to avoid a proliferation of such attention-grabbing efforts and the possible negative impact on the willingness of the community to contribute to open source efforts like Leela.

[David Larson]

I had the same thoughts. I think your hypothesis is very likely. It's what I would do. 

1. Grab the latest source code based on hundreds of hours of others peoples time.

2. Grab the latest training weights based on thousands of computing hours of other peoples setup time, hardware and electrical bills.

3. Get a database of games.

4. Provide additional training of Leela with #3, generating a new set of weights.

5. Change the name of the .exe to something else. Claim you wrote it...hope nobody notices the deceit.

[Chris Whittington]

From a programming and intellectual property point of view it's the same engine. From an enduser or chess player point of view, you can dream up that it's "different" because of the "human" games, but I could write a long essay, which I am not going to, that the chessic style and strength of a self-play learner as opposed to a human games learner will converge to pretty much the same thing, when run down the same LC0 training tube. You can however expect to read reams and reams of marketing BS from whoever publishes and/or supports this thing pretending newstyle bla di bla. Because that is for sure the aim. Probably ChessBase. It already started with the video.

From programming point of view, both "engines" can use exactly the same lc0.exe. All the decision algorithms to select next move to expand or prune are identical. Only difference is the number out of the PolicyHead, one of several otherwise identical numbers used by the UCT algorithm.The evaluation for a node from the Value will be different, but the decisions of what to do with that value will be identical. The entire Searhc algorithm is identical between each program.

The neural net itself will do exactly the same calculations and perform exactly the same algorithm through exactly the same structure network, only difference will be the weightings applied.

You can bet your bottom dollar this thing is going to get released. For money. In a wave of BS.

[Jesse Jordache]

"You can bet your bottom dollar this thing is going to get released. For money. In a wave of BS."

I think the only way that would be legal would be if it also included the source code and the standard GPL.  Otherwise not.  There's a reason MacOS is built on top of Unix rather than Linux.

[RachelHikes]

On Monday, August 6, 2018 at 5:52:01 AM UTC-7, Jesse Jordache wrote:

I think the only way that would be legal would be if it also included the source code and the standard GPL.  Otherwise not.  There's a reason MacOS is built on top of Unix rather than Linux.

MacOS is built on top of the Mach kernel, because that's what NextStep was built on, and NextStep was built on the Mach kernel because a Mach architect, Avie Trevanian was in charge of NextStep. Still, you are right in that the absence of a restrictive license does come into play in that Mach was written as a replacement kernel for BSD, which had no restrictive license.

Just a little computer science history from the olden dayes. No big deal.

[Moritz Buchty]

as he is using LC0 and mostly just changed weights, his network file must be compatible with the weights files in the LC0 archive...

let an image comparision run over them to see how identical they are.

similar to the programs that check if an image you uploaded to the net is being used elsewhere (with changed colors, blurred here, sharpened there, maybe even liquified to give you big eyes and funny noses).

tweak that 2D comparision to be used on each of the CNN planes and you'll see if there are similarities to one of them

[Jesse Jordache]

I stand corrected.  Thank you for that orthogonal, but still interesting fact.  And posted on my birthday too.

[Andrew T.]

MacOS is built on top of Darwin which is a BSD variant which, like Linux, is Unix-like not Unix.  All of these OS are posix compatible but still - if Linux isn’t Unix, neither is Darwin or OSX.

--

You received this message because you are subscribed to the Google Groups "LCZero" group.
To unsubscribe from this group and stop receiving emails from it, send an email to lczero+un...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/lczero/7391d8a0-f4b2-435b-9498-2ca239cdce87%40googlegroups.com.

[ad]

This reminds me a little bit on the startup of Facebook ...
https://www.theguardian.com/technology/2007/jul/25/media.newmedia