Open Supply Neighborhood Important Of Chessbase, Fats Fritz 2

Matt Ginsberg

The event groups behind the 2 most profitable and influential open-source chess packages, Stockfish and Leela Chess Zero, have issued statements denouncing the business program Fats Fritz 2 and the corporate Chessbase that’s promoting this system for 99.90 euros.

The statements (Stockfish weblog, lichess announcement) assert that the engine in Fats Fritz 2 is Stockfish with minimal adjustments, that Fats Fritz 2 has violated the GNU Common Public License beneath which Stockfish is launched, and that Chessbase’s advertising has made false claims about Fats Fritz 2’s enjoying power.


Final week noticed the twenty fifth anniversary of GM Garry Kasparov‘s victory over IBM’s Deep Blue. Extra famously, the thirteenth world champion misplaced the second match a yr later. For a lot of, this marks the second in historical past when man’s superiority over the machine on the chessboard ended.

Kasparov’s defeat was synthetic intelligence’s victory. It was, in a method, the top of a journey that started in 1950 when Alan Turing completed the code for the first-ever written laptop program (though Ada Lovelace is seen as his predecessor there).

Turing died in 1954, two years earlier than the time period synthetic intelligence (A.I.) was coined for the primary time by John McCarthy. McCarthy appreciated to name chess the “drosophila of A.I.” as a result of fruit flies made genetics work in the identical method that chess in some ways has made A.I. work.

Enter Matt Ginsberg, an American laptop scientist, mathematician, and the creator of a pc program that performs bridge and one which solves crosswords. Since he’s additionally a contributor to Stockfish, Chess.com contacted him to debate the Leela Chess Zero/Stockfish vs. Fats Fritz controversy. Ginsberg, nevertheless, rightly identified that it is good to emphasise the significance of chess for A.I. and the other ways people and machines cope with problem-solving.

Matt Ginsberg
Matt Ginsberg.

The robust relationship between chess and A.I. resurfaced in December 2017 when Google’s DeepMind launched AlphaZero, the chess program that shocked the chess world as a result of it defeated Stockfish, at that time the world’s strongest chess-playing program, with a extremely dynamic enjoying type primarily based on neural community machine studying. For DeepMind, these early experiments with board video games have been the premise for groundbreaking analysis in different fields.

AlphaZero’s supply code was by no means launched, however the mission impressed the pc chess group to develop Leela Chess Zero (Lc0), utilizing the identical methods as AlphaZero. It was launched in 2018. In 2020, Stockfish started supporting NNUE, quick neural networks that may run on a single CPU.

The discharge of Fats Fritz 2, a business chess program utilizing a neural community, has provoked robust reactions from the open-source group. What follows is an try and untangle a fairly sophisticated story, following a chronological storyline. 

Albert Silver and Deus X

Shortly after DeepMind launched AlphaZero, Albert Silver began researching neural networks. Silver is a chess and laptop fanatic primarily based in Rio de Janeiro and a daily author and occasional photographer for Chessbase. The German firm has been the market chief in chess database and engine software program for the reason that late Eighties.

Albert Silver Chessbase Fat Fritz
Albert Silver. Picture: Chessbase.

A neural-network chess program consists of two parts: its search performance (often referred to as the engine) and the neural community. In his newest article for Chessbase, Silver used an analogy: “Within the chess engine world with neural networks, the search is [the] race automotive and engine, and the neural community is the motive force.”

Silver launched a neural-network-based chess program, referred to as Deus X, in 2018, the identical yr that Leela Chess Zero was launched. You will need to observe that Silver didn’t program the engine himself; as a substitute, he used Leela Chess Zero and skilled a neural community in a different way than the Lc0 staff had completed. 

As an alternative of getting it study by enjoying video games in opposition to itself, he skilled Deus X completely from human video games that he took from a database, with gamers rated from 2200 and above. “It was a private thought,” says Silver.

A schematic drawing of Deus X components, as created by Alexander Lyashuk.
A schematic drawing of parts in a system akin to Deus X, as created by Alexander Lyashuk. The modification to make use of knowledgeable video games seems in field C.

Deus X was invited to take part in season 13 of the Prime Chess Engine Championship (TCEC), and Silver was interviewed about his new chess program.

Within the days forward of the match, a number of the different members raised questions on Silver utilizing the Leela Chess Zero engine as a part of his program. Shortly earlier than the match began, all of the wrinkles have been ironed out with a joint assertion by the Leela Chess Zero staff, the match director, and Silver by which it was made clear that Deus X was “powered by” Leela Chess Zero.

Silver estimates that Deus X was about 100 Elo factors weaker than Lc0 on the time as a result of it was utilizing a small neural community: “I didn’t have any sources, and I used a really crappy laptop.”

Regardless of the Elo distinction, Deus X managed to beat Leela Chess Zero in a single recreation. In accordance with Silver, that was the beginning of his fallout with the open-source group, though the Leela staff attributes the loss to a {hardware} failure and says that no exhausting emotions resulted.

“This very day that it gained that recreation [August 7, 2018 – PD], they printed an assault on me, saying ‘we had a change of coronary heart, he’s scum, he’s rubbish, he’s been deceptive everyone,'” says Silver, referring to this submit on the Leela Chess Zero weblog. In actuality, the weblog submit doesn’t embody inflammatory language and is anxious largely with whether or not or not Silver credited the Lc0 work appropriately in describing Deus X.

The Lc0 staff wrote that Silver wanted “incomparably much less effort relative to the hassle that he reused from the LCZero mission, because the LCZero staff had all of the implementation work completed, together with constructing instruments for the supervised coaching and the engine itself, defining neural community structure, and so on.”

After season 13 ended, the TCEC match director took the facet of Leela Chess Zero, up to date its uniqueness standards, and didn’t invite Deus X again into subsequent seasons.

Deus X turns into Fats Fritz

Silver started growing his neural community dimension and including engine video games within the coaching course of. After getting monetary and sensible assist from Chessbase co-founder Frederic Friedel, he obtained entry to raised laptop {hardware}, and in 2019, he talked about the mission to a different co-founder, Matthias Wüllenweber.

Wüllenweber was concerned with turning Deus X right into a business chess program however insisted on first getting private approval from the Leela builders that the Leela engine might be used. Silver contacted Leela developer Alexander Lyashuk on August 15, 2018, apologizing for failing to credit score Lc0 correctly and asking for permission to host an unchanged Lc0 binary beneath the title of Lc0, with a weak (<2000 Elo) customized internet. The subsequent day Lyashuk emailed again, saying he had talked to different individuals concerned and that there have been “no objections.”

Silver’s program was commercially launched in November 2019 beneath the title Fats Fritz, which refers to Chessbase’s long-time engine Fritz that launched again in 1991. It was the market-leading engine all through the Nineteen Nineties. 

Fats Fritz was offered along with model 17 of Fritz for 79.90 euros. The bundle included an interface, a database, and a premium account.

Fat Fritz 2

On February 9, 2021, Fats Fritz 2 was launched, once more as a part of a bundle, now priced at 99.90 euros. Whereas the primary version (which was largely primarily based on Deus X) used Leela Chess Zero for its engine, Fats Fritz 2 makes use of Stockfish due to its assist of NNUE and has a neural community twice the dimensions: the place different nets use three layers of 256/32/32 neurons, Silver works with layers of 512/16/16 neurons.

“I spent six months just about renting out computer systems on the cloud to the tune of allowed $10K,” says Silver. “I rented these 24/7 growing information from the Fats Fritz neural community that was remodeled into information that I might use to coach a neural community in a NNUE atmosphere.”

Criticism from the open-source group

A day after Fats Fritz 2 was launched, an early criticism from the open-source group got here from Tord Romstad, one of many first builders of Stockfish. 

Joost VandeVondele, the present maintainer of the Stockfish mission, printed an announcement on February 15, 2021, the place he notes: “ChessBase’s communication on Fats Fritz 2, claiming originality the place there’s none, has shocked our group.”

In accordance with Stockfish builders, there is no such thing as a important distinction between Fats Fritz 2 and Stockfish, an opinion in addition they expressed in an article titled “Fats Fritz 2 is a rip-off” posted on lichess. It was co-written by authors of various open-source chess packages.

The article begins by stating that in July 2018, “Silver secretly despatched the Leela Chess Zero engine with a customized neural community he had skilled to take part in TCEC beneath the title Deus X.” It doesn’t point out the joint assertion printed earlier than the match the place Lc0 builders initially acknowledged that Silver had made it clear that Deus X was powered by Lc0.

The builders then transfer to the primary launch of Fats Fritz and observe that it used the Leela engine with some minor adjustments: “modifying the title and creator strings, and a few default parameter values.” They level out that within the product description Leela is not talked about anyplace, whereas Chessbase speaks of “an AlphaZero clone.”

Silver is sad with this description. “I complained about that, I had nothing to do with that,” he mentioned.

As for utilizing the official open-source binary of the Leela search, Silver says he wasn’t hiding it: “This was made public. Actually, in case you set up the business model of Fats Fritz and also you look within the UCI checklist of engines it says actually: ‘Fats Fritz (in Lc0). It may well’t get any clearer than that.”

The open-source builders offered related criticism towards Fats Fritz 2, declaring that solely minimal adjustments have been made to the Stockfish engine.

“Although the Stockfish engine is crucial for enjoying power, it’s talked about solely briefly and the affect of the Fats Fritz 2 neural community over the one utilized by Stockfish is tremendously overstated. The product description says FF2 is ‘studying from the surgical precision of Stockfish’s legendary search,’ nevertheless it isn’t studying from Stockfish, it’s Stockfish.”

Silver is not completely happy both with the suggestion that his program learns from Stockfish. He factors out that the neural community discovered from the video games of Fats Fritz 1 as a substitute. He disagrees with the argument that the affect of his neural community is overstated:

“Each engine and neural community have a really highly effective contribution to make,” Silver says. “In case you take for instance the NNUE and eliminated the neural community from it, it may possibly nonetheless play chess as a result of it has the basic engine concerned, however you’re going to drop 200 Elo so clearly the neural community is contributing greater than only a present of its face, to look fairly and make a pleasant banner.”

Silver says he has by no means denied that he makes use of Stockfish in Fats Fritz 2: “The search is completely one hundred pc Stockfish. Let’s not get any misunderstandings right here. I don’t declare in any other case, in any respect.”

Stockfish logo
The Stockfish emblem.

Authors checklist

Nonetheless, the individuals behind Stockfish are sad about one specific change. They declare that in Fats Fritz 2, the checklist of Stockfish authors has disappeared. (Till now, a complete of 180 authors have labored on Stockfish through the years as will be seen within the newest authors file.)

Silver says he would not perceive the issue there as a result of in latest Stockfish variations the authors should not talked about both: “It simply says Stockfish devs.”

It is somewhat bit extra nuanced than that. Whereas Fats Fritz 2 has the observe “Stockfish Devs and Albert Silver (neural community),” Stockfish offers “the Stockfish builders (see AUTHORS file).” In different phrases, Stockfish refers to an authors file, whereas Fats Fritz 2 doesn’t. For the typical consumer who merely installs Stockfish on his laptop, there is no such thing as a distinction as a result of that authors file doesn’t include the most well-liked obtain choices.

“This isn’t an enormous deal,” says Ginsberg. “It simply hurts our collective emotions.”

A element like that is typical of how refined the conflict between Silver and the open-source group is. One other instance is a phrase utilized by Silver in his newest article, the place he writes about Fats Fritz 2: “On an an identical binary to the one Stockfish 12 used, it was popping out forward of their head-to-head matches.”

As Ginsberg factors out, this assertion can’t be true for the easy motive that Silver used a in a different way formed neural community and subsequently needed to change the engine binary, telling it that the neural community was of a special dimension.

“I feel he’s actually nitpicking there,” says Silver. “Sure, I modified the Stockfish 12 binary however apart from the bigger neural community, every part else was unchanged. The search is actually the identical.”

Ginsberg disagrees it is nitpicking: “The phrase ‘an identical binary’ is extremely highly effective, it’s referring to a selected form of experiment that he merely didn’t run.”

Chessbase’s view on the matter

Chess.com additionally requested for a remark from Chessbase. Mr Wüllenweber offered the next assertion by e-mail:

• Albert has invested plenty of time, progressive concepts and {hardware} within the improvement of his neural community. Coaching a internet is well underestimated, possibly as a result of it does not imply to put in writing code. We consider that his totally different method to coaching provides demonstrable worth to the engine, particularly to its understanding of chess technique. Now we have some articles lined up which is able to illustrate that additional within the coming days.

• We remorse that the unimaginable achievements of the individuals behind the Stockfish mission weren’t acknowledged loud and clear within the first articles when Fats Fritz got here out. We all the time see this within the metaphor “On Shoulders of Giants.” Certain, it was an achievement to climb up there. Certain, Albert would possibly see the panorama farther from up there, or a minimum of in several colours. However with out the Stockfish big, one would not see something.

• We study from the crucial suggestions to not conduct comparable initiatives with out the shut coordination of all of the authors concerned. For Fats Fritz 1, there was an alternate with the authors of LCZero to explain our intentions. It was a mistake to not provoke the identical with Stockfish.

How robust is Fats Fritz 2?

In its advertising, Chessbase used the phrases “the brand new primary,” stating that in a comparative match of over 1552 video games Fats Fritz 2 “clearly beats Stockfish 12” and “outperforms the earlier prime engine by over 40 Elo factors (as of February 2021).”

Though Silver factors out that he prefers his personal phrase “one of the best of each worlds” over “the brand new primary,” he does declare that Fats Fritz 2 outperforms Stockfish in his personal non-public checks. 

Initially, Silver examined Fats Fritz 2 in opposition to Stockfish 12, which was launched September 2, 2020. Nevertheless, the engine of Fats Fritz 2 relies on a more moderen model of Stockfish 12—a developer model that’s nearer to Stockfish 13, which was launched on February 19, 10 days after Fats Fritz 2 got here out.

Evaluating video games between two packages that use totally different serps appears odd certainly. Why did he try this?

“Stockfish 12 is the one reference most individuals have, actually,” says Silver. “Sure, there’s a Stockfish dev however most individuals don’t learn about it and the issue is that Stockfish dev is a shifting goal, frankly. Even I don’t check in opposition to it in my private testing.”

One more reason for Silver to check in opposition to Stockfish 12 was that it was the benchmark on the pc ranking lists. Not like with human chess, for chess packages, there is no such thing as a single, universally accepted ranking checklist.

“Look, I used to be requested to offer a check mainly to point out the way it did in opposition to one of the best packages. On the CCRL [the Computer Chess Rating Lists – PD] and the CEGT [the Chess Engines Grand Tournament – PD], Stockfish 12 was on prime of the checklist. So I needed to present what they have been going to be getting and the way it in comparison with it. And that was the consequence I produced.”

As soon as Stockfish 13 got here out, Silver in contrast in opposition to it as a substitute and printed these leads to his February 21 article.

Fat Fritz 2 vs Stockfish
Silver’s check confirmed Fats Fritz 2 performing 16 Elo factors higher than Stockfish 13.

The Stockfish staff will get totally different outcomes. On a bigger pattern, the place each engines have been compiled from supply on the identical machine, their outcomes at 60+0.6s have been:

On the most recent CCRL blitz checklist (two minutes and a one-second increment), Fats Fritz 2 is the highest engine with two Elo factors greater than Stockfish 13. On the 40/2 checklist (video games with 40 strikes in two minutes), Fats Fritz 2 is forward of Stockfish 13 by six Elo factors, though the January 11, 2021 model of Stockfish on its flip is six factors forward of Fats Fritz 2.

The slowest time management CCRL seems to be at is video games at an equal of 40 strikes in quarter-hour. Testing for Stockfish 13 has solely simply begun and there are solely 24 video games up to now: a drawn match in opposition to Dragon by Komodo. On the prime, see once more see the January 11 model of Stockfish, though Fats Fritz 2 is as excessive because the second place.

CCRL 40+2 list Stockfish Fat Fritz 2
The February 20, 2021 CCRL 40+2 checklist.

The well-respected Stefan Pohl, seen as probably the most unbiased laptop program tester, at present has Stockfish 13 at 4 Elo factors stronger than Fats Fritz 2, with the January 11 Stockfish coming two factors increased.

Stefan Pohl computer chess rating list
Stefan Pohl’s newest checklist.

“I’ll be trustworthy with you, I’ve outcomes far and wide,” says Silver as he refers to a variety of lists. “I don’t know what to say. The whole lot goes to rely upon the openings, the circumstances, no matter.”

Concerning the query of why the outcomes are so totally different, Silver says: “Not all lists use the identical pattern dimension and there is likely to be totally different circumstances. I did a breakdown of one of many lists that produce the outcomes and I came upon that ECO A openings (largely English, Benoni and irregular openings – PD] represented 60 p.c of the video games. I believed: what if they’d a extra balanced distribution of the openings, wouldn’t it nonetheless appear like that?”

Ginsberg means that Silver’s divergent check outcomes are a mix of a comparatively previous processor and the time management. Silver used an AMD FX-8350 (4GHz), utilizing one thread (either side) with 512MB hash.

“This can be a processor that was launched in 2012,” says Ginsberg. “So a check with video games of 90 seconds and a one-second increment goes to be successfully hyperbullet. We’ve not completed any testing in any respect at these speeds, and all of us use fashionable CPUs.”

As of February 24, 2021, Chessbase has eliminated the “new primary” declare from its web site.

“We really feel that prospects shopping for Fats Fritz 2 get little or no added worth for cash,” wrote VandeVondele in his assertion. So what’s that added worth? Why would a buyer pay these 99.90 euros?

Silver: “Finally I don’t suppose it’s all about Elo. If it was, it will be actually boring, frankly. In case you’re speaking a couple of modest distinction it shouldn’t be absolutely the criterion. Certain, it’s a advertising side, there’s no query. But when I’m selecting an engine to research with, I do know tremendous GMs who nonetheless analyze with Leela though its Elo is worse than Stockfish NNUE. Why? As a result of it gives them with new views, totally different evaluation, totally different positions, totally different continuations. They relish that. They don’t care if on checklist X, Y, and Z it’s 50 Elo behind or forward. I’ve completed sufficient checks and sufficient video games to say that FF2 NNUE in comparison with Stockfish 13 NNUE in the identical search, utilizing the identical binary, makes use of totally different continuations, greater than sufficient to say that they’re clearly totally different.”

License points

Moreover Chessbase’s claims about Fats Fritz 2’s enjoying power, one other factor the open-source group is upset about is the alleged violation of the GPL license by Chessbase.

In his assertion on behalf of the Stockfish builders, VandeVondele wrote:

Promoting Stockfish derivatives is feasible with the GPLv3 license we grant, however not with out necessities. Particularly, the license states that if one redistributes a program derived from our work, the corresponding modifications of our sources and all data wanted to construct that program have to be made obtainable. Solely after explicitly informing Albert Silver (the creator of the web in Fats Fritz 2) of a license violation have matching C++ sources, however not the web weights, been made obtainable. Clearly, we condemn the method taken.

This a part of the conflict between Chessbase/Silver and the Stockfish builders appears to be the results of the best way Fats Fritz 2 was initially launched, together with a special interpretation of the GNU Common Public License (GPL) beneath which Stockfish has been launched.

In brief, the GPL permits for utilizing, sharing, and modifying software program, and that features the proper to take action commercially, so long as the by-product work is distributed beneath the identical or equal license phrases.

When Fats Fritz 2 got here out, the net obtain model consisted of a single set up file that may create a bundle on the consumer’s laptop that included the search engine and the neural community.

“I did it as a result of, one, the neural community is unusable in every other atmosphere, and the second motive is that this was what Stockfish was doing, this was customary process,” explains Silver.

Nevertheless, as a result of every part was supplied as one file, the entire bundle fell beneath the GPL. As VandeVondele defined, this meant that every part, together with the supply code and neural internet, needed to be made publicly obtainable.

Chessbase said that it had made the supply code of the search engine obtainable, however up to now, it hasn’t made clear the place it may be discovered. In accordance with Silver, the (hyperlink to the) code solely needs to be offered upon request. In actuality, the license says you must supply the supply code “in the identical method by means of the identical place at no additional cost,” whether or not a hyperlink is requested or not. Entry to the neural community wasn’t offered both.

When confronted with the GPL violation by Vandevondele, Chessbase up to date its product three days after its preliminary launch. Since then, the search engine and the neural community are offered individually.

“I used to be knowledgeable that the GPL permits for a breach to be corrected inside 30 days so that is only a technicality,” says Silver, who is not planning on freely giving the code and different data behind his neural community. He does supply a weaker, earlier neural community for customers who wish to compile their very own copy.

The explanation for separating the search engine and the neural community is that. in accordance with Silver, as soon as it has been separated from the binary, the neural community is not topic to the GPL.

Whereas acknowledging that such issues can solely be determined in court docket, Ginsberg says he disagrees: “You’ll be able to’t un-GPL one thing. As soon as it is on the market beneath the GPL, you may’t undo that.”

Ginsberg famous that he bought Fats Fritz 2 on February 22 and that his obtain nonetheless consisted of 1 set up file, saying: “The GPL applies to that file as a result of that file is object code that features the Stockfish binaries.”

It isn’t the primary time that the Stockfish builders have accused a business chess program of violating the GPL. A latest instance is Robert Houdart’s Houdini 6 that, they declare, is a Stockfish 8 by-product with out offering the sources on request. This has resulted within the TCEC revoking the championship outcomes of Houdini 6.

“I wish to suppose that they weren’t as cautious as they need to’ve been,” says Ginsberg about what he sees as a mistake made by Chessbase. “I don’t wish to say that anyone was evil right here however they didn’t act like scientists. I don’t wish to complain about them, I simply need your viewers to know the reality.”

Must you purchase Fats Fritz 2?

What stays is the query of whether or not it is smart for the typical chess fan to buy Fats Fritz 2. In accordance with Silver, it is the top consumer who must resolve: 

“Other than the interface and different perks you pay for, with Fats Fritz 2 you get a big, totally different neural community, working completely in Stockfish, no query about that. It is going to present a special analysis, it would present a special search. Do not forget that the search isn’t the identical. It’s guided by the neural community itself. Similar to you’d analyze strikes and also you is likely to be analyzing on the identical velocity as any person else, you’re going to be specializing in totally different strikes due to the way you take a look at chess. It’s the identical factor right here. Some traces will go deeper, some traces much less, and all of that’s the neural community that decides this. The velocity would be the identical however the traces and the strikes proven are particular to the neural community itself. Is it gonna play the identical? Generally sure, generally no. However it’s going to offer totally different concepts. Is that this sufficient worth for you? Is that sufficient to justify shopping for it? It’s a private name. I can’t make the decision. For me it’s in fact it’s. For some GMs it’s in fact it’s. For a mean participant who doesn’t care about such issues, the reply isn’t any. And if the GUI provides nothing to their expertise, then actually, no. It’s about totally different opinions, totally different views, totally different evaluation, totally different concepts. The worth of that he’s going to be as much as you to resolve.”

Ginsberg thinks the race automotive vs. driver analogy is truthful however finally would not work in Chessbase’s favor: “His neural community is worse as a result of it loses Elo. So that you’ve obtained the identical automotive, two drivers, and one man retains dropping. It’s the motive force. Claiming that the web is best may be very annoying to the Stockfish builders who’ve spent a lot time making one of the best internet, getting one of the best driver they may discover. So our view is: we have now one of the best automotive—he doesn’t argue with that—we even have one of the best driver, and it’s nonetheless our automotive.”

Apparently, Ginsberg would not thoughts in any respect that Chessbase is earning profits from Fats Fritz 2: “I am thrilled that they’re earning profits and I hope they make so much. As a result of if they’re earning profits it’s good for chess. I like chess. I would like chess to be standard. It doesn’t hassle me in any respect that they’re earning profits. Not within the slightest.”

The conclusion, subsequently, might be that ultimately, chess wins. Industrial and open-source chess packages will proceed to be developed and can get stronger as time passes by, main not simply to increased rankings however to extra discoveries, extra excellent video games performed by computer systems, and extra understanding of chess. Whereas computer systems want people to make new steps, people are helped by computer systems to understand the traditional, royal recreation much more.



Supply hyperlink

You may also like

Leave a Reply

Your email address will not be published. Required fields are marked *