An actual interesting computer chess read about FF2

[connor_mcmonigle]

Albert has certainly lied or, at the very least, severely bent the truth surrounding the networks he's trained. Calling him a liar is on the edge of being justified imho.

Just listen to his original interview about "DeusX":
https://youtu.be/CpjvvcfbdR4

No one said it was impossible to train a neural network on human games. In fact, the tools developed by the Leela developers that Albert used were designed to do exactly this. The claim was that it was unlikely to be possible to produce a stronger network than that produced through self play. I guess Albert found this to be true for himself as FF is largely trained on self play games.

Also, he manages to go a straight 26 minutes without mentioning the Leela developers once. He was even directly asked how it was possible for one developer to produce a strong neural network chess engine. This is just one example, of many, of his downplaying or neglecting to mention the significance of the work of others.

In interviews, he consistently tries to characterize himself as some forefather of using neural networks in chess engines, but he even lacks the technical understanding to just provide remotely accurate explanations of neural networks in those very interviews.

Is calling him a liar justified? I'm not sure, but it's certainly closer to the truth than many of Albert's statements.

[carldaman]

In all the lighthearted frivolity about the law, it seems we’ve missed an article on how an ICCF GM evaluates an analysis partner and comes to prefer FF2. Warning: contains actual chess.

https://en.chessbase.com/post/discussin ... al-network

Good read - thanks for the link.

One nagging little thought: if it's better at suggesting good moves, then why doesn't it play better chess? Why isn't at the top of the independent rating lists?

The Fat Fritz 2 compile of Stockfish has the default contempt set to zero, resulting in higher quality and less biased evaluations, but on rating lists weaker engines have an easier time drawing Fat Fritz 2 due to contempt being equal to zero.

Actually, Fat Fritz default contempt is 24. Even though it makes little difference.

Ever since NNUE was implemented contempt makes little difference for SF. Contempt-less NNUE can be an impediment against winning more with Black vs weaker engines, so it fails to maximize Elo potential.

But even if SF's contempt was fully functional, as it used to be before NNUE, it would be correct to have it on, since the engine has to compete against mostly weaker opponents on the rating lists and competing is about winning the most games.

[Nay Lin Tun]

Obviously observer biased error with cherry pick positions.
Give him unknown copy of FF2 and SF13 and let him analyse about 10 to 20 positions.
Let us know what he think is better.
( Sponsored articles are already Redflag in critical apprasial btw)

[dkappe]

After all of this Sturm und Drang I think two things are clear:

1) everyone has heard of FF2
2) the SF/Leela/lichess communities’ words and actions speak for themselves.

[Milos]

After all of this Sturm und Drang I think two things are clear:

1) everyone has heard of FF2
2) the SF/Leela/lichess communities’ words and actions speak for themselves.

3) Two ppl have serious lack of shame. Alberto Plata "author" of FF2 and his "devil advocate" here.

[jefk]

initially the CB post looked interesting but after the upheaval here I
copied the game for some check/analysis on my (not so fast) desktop.
(wondering if anyone else here actually bothered to do that, ie.
look at the actual moves (wrt. the comments made by S. Ham)

First (and worst) anomaly: in a sideline with SF analysis, SH
claims that SFNNue wants to play an obvious error, 19...g5??
But i wasn't able to reproduce that (maybe, just *maybe*, after a much
deeper analysis SF could come up with 19...g5, although i doubt that...
(the response 20.f4 stays stable at approx 2.46 after play >33.

Then about the 'positional' difference with 16.Rab1 vs 16.Rfb1,
in the SF line with 16.Rfb1, SH claims that Sf plays 18.Bxd7 but i
got (at play 43 a 'better' (?) move 18.Bg3! (0.45).
In the line with 16.Rab1, SF first considers 20...f6! (+0.3)at ply 39
but then indeed switches to 20...Bc6 (also the FF2 move),
but then the eval of the resulting position after 21.e4 also is
going down, to approx 0.39 after 21...b5 at 31 ply.

So, although SH prefers the line with 16.Rab1, it's not very clear
if this is really better. When i look at the evals with Komodo Dragon,
it gives after the line 16.Rfb1 /18.Bxd7 (*) /21.e4 an eval of approx 0.49 (ply34)
and after 16.Rab1 b5 (and 19.h4) approx 0.50 at ply 36.
(*) Kom/Dragon prefers this move instead of Bg3...
Preliminary conclusion: hardly a significant difference between the two lines, and in both
situations it still would be very difficult to win for White (against a strong engine).

Another fundamental issue in the analysis, about correspondence chess
in the age with Nnue (without opening book); personally i wouldn't play
Nc3 in the SH line, the Nimzovitch which is strong for Black,
and instead of 4...d5 (a book line) would prefer 4...b6!! or 4...0-0!.
And the latter moves are exactly which SF resp. Kom/Dragon are
preferring at > 34 ply analysis or so (no book needed!).

As for the S.Ham CB story, as correspondence GM i don't think he's deliberately
twisting the truth, yet with some prejudice (and/or bias) and some rather
superficial (yet pseudo-deep) analysis seems to have exaggerated differences
between SfNnue and FF2, and apparently also unaware that you don't need
'old' theory, i.e. (sloppy) book-lines anymore with the Nnue and deep analysis.

PS as for the whole 'open source' debate, to my knowledge it's perfectly
legal to grab open source code, 'package it' and try to sell it with
some advertising etc. But it would be decent business ethics of-course
if the seller would notify the producer(s) of the source code of such intentions;
and then would be reminded of-course to include the source code with the GPL
or other license with the commercial 'package'. The whole anti CB atmosphere
(who still have the awesome yet expensive product of CB16) reminds me
a bit of the 'comrades/Decembrists' in the Ippoloito/Robbolito times
whereby i'm surprised that also many Western programmers would have
such severe (apparently) anti-capitalist sentiments.. Just my 2 cnts but
that's my impression so-far at least, unless i would be convinced otherwise
(eg. because of some valid legal issues). Nevertheless anyway i hope to
have contributed now in a positive way to this current controversion(s)

[Cornfed]

Ever since NNUE was implemented contempt makes little difference for SF. Contempt-less NNUE can be an impediment against winning more with Black vs weaker engines, so it fails to maximize Elo potential.

But even if SF's contempt was fully functional, as it used to be before NNUE, it would be correct to have it on, since the engine has to compete against mostly weaker opponents on the rating lists and competing is about winning the most games.
[/quote]

Of course, this line of thinking presumes the main purpose of an engine is to....'fight it out' against another engine during engine v engine play.
Truly, I think most people just switch on an engine to get evaluation of certain positions. That is something way too many people tend to forget...

[towforce]

PS as for the whole 'open source' debate, to my knowledge it's perfectly
legal to grab open source code, 'package it' and try to sell it with
some advertising etc.

I think everyone agrees with that, and the GPL FAQ explicitly says that you can. CB have published their modified version of SF, and they have published a set of NN weights - but the NN weights they have published are not the ones they're selling as their FF2 product.

The questions under dispute are:

1. Can an NN be copyrighted?

2. Does the SF GPL license oblige them to publish the NN weights file that they're selling with FF2?

IMO, in this case, it's of no consequence (people who want the strongest engine will download SF 13, people who buy FF2 will be getting a lot more than just SF + new weights), but we cannot seem to come to agreement about the above two questions.

[MonteCarlo]

...

First (and worst) anomaly: in a sideline with SF analysis, SH
claims that SFNNue wants to play an obvious error, 19...g5??
But i wasn't able to reproduce that (maybe, just *maybe*, after a much
deeper analysis SF could come up with 19...g5, although i doubt that...
(the response 20.f4 stays stable at approx 2.46 after play >33.

...

He didn't word it very clearly, but the way I read that section he is not claiming that SF will play g5 in the Rab1 line.

He's asking and answering the question "After Rab1, why can't we play the same defense SF gives for Rfb1? What's the difference?" and the idea is that in the Rfb1 line SF gives g5 as a good defense for black, but it's unplayable in the same line after Rab1.

In the Rab1 line, SF of course avoids g5?? and sees f4 as crushing immediately, as you note

Otherwise, I came to similar conclusions.

I spent a few hours analyzing the position, and It's not at all clear to me that Rab1 is any better than Rfb1; to be fair, he mostly sticks to saying that he likes the move better, not that it's objectively stronger, which is fair enough, I suppose (although playing devil's advocate, I'm not sure that picking an analysis engine that tends to agree with your own biases is necessarily the best idea).

In terms of getting a win against a strong ICCF opponent though, even if they're an unassisted engine, I don't think that choice is going to make a difference (and he does admit that whatever advantage exists is pretty small).

Cheers!

[willmorton]

Albert has certainly lied or, at the very least, severely bent the truth surrounding the networks he's trained. Calling him a liar is on the edge of being justified imho.

Just listen to his original interview about "DeusX":
https://youtu.be/CpjvvcfbdR4

No one said it was impossible to train a neural network on human games. In fact, the tools developed by the Leela developers that Albert used were designed to do exactly this. The claim was that it was unlikely to be possible to produce a stronger network than that produced through self play. I guess Albert found this to be true for himself as FF is largely trained on self play games.

Also, he manages to go a straight 26 minutes without mentioning the Leela developers once. He was even directly asked how it was possible for one developer to produce a strong neural network chess engine. This is just one example, of many, of his downplaying or neglecting to mention the significance of the work of others.

In interviews, he consistently tries to characterize himself as some forefather of using neural networks in chess engines, but he even lacks the technical understanding to just provide remotely accurate explanations of neural networks in those very interviews.

Is calling him a liar justified? I'm not sure, but it's certainly closer to the truth than many of Albert's statements.

I totally agree and I think that's the reason we have had such a strongreaction from the open source community this time.