Update 20210730 "Stockfish" Team suid Chessbase
I am no legal expert but this is how I interpret the situation. I will wait and see what happends and update this text when the case is settled. I have no clue or can predict what happends in court, both can win or semi win. It might also lead to that "Stockfish" team will get access to "Fat Fritz 2.0"s Lc0 generated data.
Team "Stockfish" suid Chessbase since they think they violeted copyright on both "Fat Fritz 2.0" and becouse Chessbase have the engine "Houdini" for sale at thier shop and they claim "Houdini" uses a lots of older "Stockfish" code. "Houdini" is also several years old and Chessbase sells it in a bundle with an older "Fritz" program. Many Chess shops that sales Chessbase products has both the "Houdini" bundle as well as "Fat Fritz 2.0".
"Stockfish" Team choosed to exclude Chessbase from thier copyright before suing Chessbase the but since the product was compiled when the licens was valid and dont think that they cant win on the "Fat Fritz 2.0" part but with "Houdini" part is an other matter since they did not fix some smaller GPL licens issues in time so I predict "Houdini" will no longer be sold. Chessbase did probly not even know anything about "Houdini"s sourche code or what its based on since they did never developed it themself or in any way modiefied it. Anonymous plagiaton rumors started only some year ago when the engine had already existed for several years. I will also be surprised if anyone actually buys or even uses "Houdini" today since many other engines outdeveloped it. However I belive that Chessbase can probebly not legally develop the "Fat Fritz 2.0" pure engine code any further but they can still train the NNUE network further.
"Stockfish" team claims no originalety by Chessbase but "Stockfish 14" more or less took Chessbase new ideas with a much larger NNUE network and use of "Lc0" generated data. Still not complety the same thing since "Fat Fritz 2.0" NNUE network is almost purly based on "Lc0" generated data and also a few data from top level human games. "Stockfish 14" NNUE network in turn uses both "Lc0" generated data and data from "Stockfish" vs "Stockfish" games but no data from human games. "Fat Fritz 2.0" evaluations are therefore more putting "Lco", with a completly different learing history, into "Stockfish" engines strong calculating power and no need of strong graphic card. Chessbase generated and tranined the NNUE network themself starting from "Fat Fritz 1" weaker network so it never used "Lc0" teams stronger neural network. So basicly "Fat Fritz 2.0" is a "Stockfish" NNUE network where "Stockfish" vs "Stockfish" games are not involved at all and was much larger than "Stockfish" original NNUE network where Team "Stockfish" original thougt that a larger NNUE network was a no gain approach and that a larger NNUE network even resultated in weaker play.
There are simular approaches of using other engines or human games as base for neural networks. For example there are many "Lc0" clones where some clones training are only based on other engines vs other engines matches and some clones training of are only based on human vs human games played at Lichess. But "Fat Fritz 2.0" is the first and ,last and only "Stockfish" clone that uses a NNUE network where "Stockfish" vs "Stockfish" games are not involved at all.
I predict that "Stockfish 15" will use an even larger NNUE network and that NNUE network size will grow even further after each new version.
I am not shure what will happends for custumers or Chess stores if "Stockfish" wins. This including those who bought or sold "Houdini" or "Fat Fritz 2.0" or both. Will for example Chessbase digital copys still be downloadable or not? Will it be possible to install discbased versions? Will the existing programs increase in value in the future becouse of the conflict? Will the programs be free? Will the programs still be usable?
Update 20210705 "Fat Fritz 2.0" now obselete design
With the realise of "Stockfish 14", "Fat Fritz 2.0" is now complety obselete design. "Stockfish" Team and "Lc0" Team worked togheter to improve "Stockfish". The size of the NNUE network has increased much and it learned from both same way as "Stockfish 13" but combined with "Lc0" thus "Fat Fritz 2.0" only original ideas with larger NNUE network and Lc0 based NNUE training are copyed except that the "Stockfish" NNUE network has not learned from top level human games but from "Stockfish" vs "Stockfish" games instead.
Becouse of this realise the rest of this articles text bellow are now obsolete but keept for history documentation reasons. If funny that "Fat Fritz 2.0" for a short period was engine nr 1 and that the engine realise made "Stockfish" Team and "Lc0" Team work togheter to create a even better "Stockfish" where "Fat Fritz 2.0" only original ideas over "Stockfish 13" with a larger NNUE network that learned from "Lc0" training was copyed and made even better. I prediced that this moment eventually would come that "Stockfish" eventuellt would outdevelop "Fat Fritz 2.0" but this realise came much earlier than I expected. I dont think that Chessbase can make much improvements. They could try to train the NNUE network even further or updated engine code towards latest "Stockfish" but there are little point. "Fat Fritz 2.0" did any rate prove that larger NNUE networks is not a dead end approch that "Stockfish" Team used to belive. In the future that NNUE network mihgt even be even larger.
Who knows if we will see a "Fat Fritz 3.0" in the future but it will be in unclear what it will based on and how to make engine uniqe and compitable and worth buying. Maybee it will be based on "Dragon" by "Komodo" Team in some way but different NNUE network.
Engines of the future
A completely new type of Chess engine was developed by Google called "Alpha Zero" and played against and outclassed a sort of handicapped and not latest version of "Stockfish" that existed back then. "Stockfish" was the best engine at that time and was completely free. "Stockfish" was not allowed to use openingbooks or use tablebases or use permanent brain and the event used a stupid time control that "Stockfish" was not designed for and "Stockfish" was setuped and forced to run way to many threads, more threads than "Stockfish" had been testet for. "Stockfish" got a way to low hashtable thus wasted most threads and was sort of killed by it self since each thread consumed calculation power with no gain since it used to many thread and way too low hashtable so I used multiple threads to calculate the same variations over and over again in different threads. "Stockfish" makers however still thinks that "Alpha Zero" would have won the match large but it would have been by less margin.
"Alpha Zero" used a grundbreaking self teaching alghoritm and a neural network teaching it self Chess. During the training it is used special hardware, with the theoritcal power of about 4000 i7 CPUs and during the match hardware with the theoritcal power of about 2000 i7 CPUS. The hardware it used was bases on a TPU computer instead of normal CPU computer. 1 TPU = 500 i7 CPUs, 4 TPUs was used during match and 8 TPUs was used during training. It was given the rules of Chess and trained itself by reapetly playing against it self forever, at a very fast speed per match since it run on powerfull hardware and slowly got stronger and stronger. It created its own openingbook and soon played openings that theory considered best today.
The project inspired many new engines developement to use a neural network but that could run on a PC and not only on special hardware. The first generations of the new engines was depending on that the computer had a powerfull graphic card and they was tricy to both install and setup but soon "Fat Fritz 2.0" appeared that was easy to install and could run on normal computers even thought the project has contreversals as its bases on a modiefied version of "Stockfish 12 NNUE" but with different and larger neural network trained in a different way and the modifcation was bascily to support a twice as large neural network.
The neural AI network "Alpha Zero" like engines are getting better every year and are slowly getting stronger than convential engines. As I understand even on normal computers the AI neural network engines are already better than convential engines now in 2021. I predict that by 2030 the top 10 strongest engines are all AI neural network based one way or an other.
"Fat Fritz 2.0" was based on the back then latest build of "Stockfish NNUE" but with a twice as large and differently trained AI neural network. The engine was tough Chess during supervision as well as self training unlike "Alpha Zero" who tough it self competly. Even through "Alpha Zero" can not run on normal hardware it has forever changed Chess engines development and is one of Googles most succefull projects and marks a new milestone in Chess engine development and is possible the best Chess engine ever. There has been specualtions that latest "Stockfish" has gained so much strength since the old version that "Alpha Zero" met so "Stockfish" would win this time but none knows for sure on how much "Alpha Zero" can ge improved in turn by self teaching.
"Fat Fritz 2.0" had sort of legal issues and many players are upset that it is basicly is a clone of "Stockfish NNUE" but with twice as large neural net and it was trained in a completely different way. So the products original part is basicly only the original trainined neural network and some slightly changes to the engine to support a larger neural network and some minor changes and bug fixes. The point was that "Fat Fritz 2.0" is bloody expensive and "Stockfish" is completely free, close to selling something for free for moneny. "Stockfish NNUE" was at that time a forked open sourche version of "Stockfish". Normal "Stockfish" is also open sourche. "Fat Fritz 2"s predesesor "Fat Fritz 1.0" was basicly a copy of a verson of the engine "Lc0" but with a different learning history. "Lc0" was probebly the first "Alpha Zero" clone and requered a strong graphic card. It is becoming better every monts. During "Fat Fritz 2.0" training it used "Fat Fritz 1" new neural network for start guidence but added by "Stockfish" calculaton power. This means that "Lc0" has part of the training. General engine experts says the AI engines strength is much depening much on how the neural net was trained. "Stockfish" NNUE network are are trained from last none NNUE "Stockfish" playing against itslef while most of "Fat Fritz 2" training comes from "Lc0" self play, using the orginal "Fat Fritz" network as base but trained much further and the network was also trained from high level human games so is uniqly trained.
Later "Stockfish 13" was realised. It is not that clear if "Stockfish 13" if is better than "Fat Fritz 2" or not. When this text was written 20210301 "Fat Fritz 2" was rated 1 elo point above "Stockfish 13". In terms of propaganda "Stockfish" won the war but in strengh is not clear since both engines are close related and both are improvments over "Stockfish 12 NNUE". "Stockfish" developers can perhaps learn something from "Fat Fritz 2.0" neural network training and also implement a twice as large neural network. "Stockfish" developers had priviusly ruled out a twice a large neural network as a dead end and no gain in their implementation of NNUE while "Fat Fritz 2.0" has showed that it works good with "Stockfish" engine to support and play good using a twice as large neural network, its the only engine development milestone that the "Fat Fritz 2.0" added appart from the idea of learning by combining self teaching games like "Lc0" and top level human games.
In one of my own Correspondence Chess games "Fat Fritz 2.0" suggested an other top move than "Stockfish 12" that is was based on, I used the back then latest build with NNUE. "Stockfish 12" did not have it as a top move even high extremly high depths prefering a different move. However after I entered that move "Stockfish 12" in the newest build saw is strength and evaluted the position as more favorable to me than the other move thus proving that "Fat Fritz 2" is stronger than "Stockfish 12" with NNUE but perhaps not stronger than new "Stockfish 13". I checked several positions and "Stockfish 13" and "Fat Fritz 12" often suggests different moves as the best from time to time, and even when the give the same move the evaluaton is different. Very often have the same top 4 candidate moves but in a different order so one can use both to get a second opinion. I dont know if Chessbase is going to realise updates to the "Fat Fritz 2" or not. If they dont "Stockfish" versions will soon dominated "Fat Fritz 2" since the communty regulary updates "Stockfish" and it becomes stronger every year. Its also possilbe that "Lc0" will become much better do to the new cooperation between the different crews.
I learned that in some Computer Chess Engine Championships clones are not allowed unless there are two major differences. Since "Fat Fritz 2.0" only difference towards "Stockfish" is twice as large and differently trained net and the only changed to the engine is to allow "Stockfish" to use a larger net I belive it that "Fat Fritz 2.0" can not play in every event. I also learned that some "Lc0" clones can not play in every event but some can by combining a differently trained net with an other major change of engine alghoritm that one clone has done and thus is a double clone and therefore are allowed to play.
Many "Stockfish" developers and fans become so upset about "Fat Fritz 2.0" that they created a fake promotion site on a new fake domain that says dony buy it and that it is garbage and they even used a DDoS attack to temporary disable a official Chessbase article about the product that described the training and some technikal things about how they made the neutal network and the engine. Many forums uses trolls that like fake negative facts about it etc etc.
"Stockfish" team also placed theire message at every Chess platform that existed, including at Lichess and claimed that "Fat Fritz 2.0" was a complete rip off and no originalety and weaker than "Stockfish" and dont buy it. After the post Lichess got "Stockfish 13" as build in engine in analyse mode.
"Stockfish" team also claims that they did not get any credit at all a fact that is false. On the box it even says that "Fat Fritz 2.0" is the neural network and not an engine and that "Stockfish" is the engine. They given full credit to "Stockfish" team and even give extra credit and names the persons who made the NNUE code.
Both "Stockfish" and also "Lc0" supporters also publiced clear fake videos on Youtube where "Fat Fritz 2.0" makes begginner mistakes in opening and idiotic middle game blunders. I have entered the moves in the games manuely and my "Fat Fritz 2.0" copy suggests completely different moves even at short time , the moves played often are not even on "Fat Fritz 2.0" top 10 candidate list. If the games realy are "Fat Fritz 2.0" it must be the must be a compilation from the free much weaker "Github version" that has a much weaker less trained neural network or else it is a human player that entered the moves pretending it is "Fat Fritz 2.0" playing witch I found more likely. A third possible option is that sone one put handicap function in the "Fritz" program on but I find it very unlikely. An other option I dont rule out is that is a completely different engine. An other possible option is that the opening book was altered to setup the game since in one of the games had so many opening blunderes where my "Fat Fritz 2.0" copy dont like any move in the opening or else they created a fixed start position since the videos dont show evalutions only moves played at a board that is not standard "Fritz" looking and time used per move or time left are not displayed either.
For a short time "Fat Fritz 2.0" free "Github version" was also on rating lists and it was rated higher than most engines but still lower than the top 5 engines while commercial version of "Fat Fritz 2.0" was rated as nr 1 and "Stockfish 12" as nr 2. When this text was written "Fat Fritz 2.0" is rated as nr 1, 1 Elo higher than "Stockfish 13". It could change anytime since newer versions of "Stockfish" are realised regulary getting stronger in every version while it is unclear if "Fat Fritz 2.0" will get any upgrades in the future like a even more trained neural network or if they will wait som years to create "Fat Fritz 3.0".
But one positive soon thing appered. Since "Fat Fritz 2.0" won over "Stockfish 12" according to Chessbase a new version on "Stockfish" , "Stockfish 13" was realised way much earlier than planned, it had been in delopment for a while, with a more than 30 elo point increase according to "Stockfish" developers. Also "Stockfish" and "Lc0" teams started to cooperate in a new way thus staring a war but in friendly and fair way. I also learned that the commercial engine "Komodo" has an experimental neural network version called "Dragon", so there are three sides that fight of becoming engine nr 1.
I have also learned the "Lc0" has got clones where the neural network learned by studing other engine matches games or similar other learning apporoches.
I have seen other Youtube videos where "Fat Fritz 2.0" completely outplayed latest "Lc0" and also videos where the last version of "Stockfish 12 NNUE" was totally crushed. I dont know if first version of "Stockfish 13" is that much better than last version of "Stockfish 12 NNUE" or if the videos are fake since no time or evalutions are displayed and a none standard look at the board.
Update added 20210504 Chessbase strikes back
Chessbase changed "Fat Fritz 2.0" into "Fat Fritz 2.0 SE". Everything is the same but installation program gives more clear credits to "Stockfish" crew and some additonal crew files are copyed to fix som legal stuff.
Than after waiting untill the storm was over Chessbase publiched 3 articles where top players promoted "Fat Fritz 2.0" over "Stockfish". In each article "Stockfish" gets credits and "Stockfish NNUE" is given as an improved evaluation function in for example in closed positions and many positions that was evaluted as 0.00 are not 0.00 evaluated anymore.
One article showed when "Stockfish 13" was defeated and one other article showed when "LC0" was crushed.
The third article was about analysing positions where the writer showed several examples where "Fat Fritz 2.0" has a more human view of the position since the net based on both "Fat Fritz 1" self trained parts and top level human games. The article also showed concrete moves where "Stockfish 13" move lead to equalety after horizon effect while "Fat Fritz 2.0"s move instead lead to a small advantage.
Engines developement and NNUE
A curiosety is that "Fat Fritz 2.0" project originally started and was developed as a much better trained version of the engine "Fat Fritz 1" but Shogi (a Japanise Chess like game) engine developers helped "Stockfish" developers to implement NNUE so half time in the projects development Chessbase started to use "Stockfish NNUE" engine as base but with the new developed and uppgraded "Fat Fritz 1" based trained neural network that was designed as twice as large as "Stockfish" neural network and they trained it even further for several monts using "Stockfish" powerfull search. The difference between the engines is that "Stockfish NNUE" neural network uses "Stockfish" based evalutions trained from playing "Stockfish" games against it self and "Fat Fritz 2.0" neural network uses "Fat Fritz 1" based evalutions and thus uses "Lc0" like evalution thus trained in a complelty different way even using top level human games as part of the training.
The reasons for choosing "Stockfish NNUE" as base was helped by the fact of no nead of strong graphic card and that it was open sourche code that was altered to support a larger neural network and some minor things. "Stockfish" is also extremtly good at reacing high ply fast much faster than "Lc0" that "Fat Fritz 1" was cloned from. It also makes it more easy to train the neural network.
Unless one is a engine expert one has no idea how the neural network effects the strength and style of the engine and that is the reason why "Stockfish" won the dirty propaganda war. Many players does not even understand the difference between engine and evalution function and how they work together.
The reason why "Lc0" are becoming stronger every year is related to more games played and analized and thus more and better training of is neural network and has not much connections to other developments of the engine that to my understanding only has got minor adjustments in every new version. I belive that the training and training methods of neural networks are the key to any engine in order to becoming rated as nr 1 and that the neural network engines will make every other engines obsolote in a near future.
Different engines are better than others in different openings and positions. I some positions are for exemple latest "Lc0" better than "Stockfish 13" dispete large rating gap. I think that one should compare what different engines says both when analysing or when playing Correspondence Chess.
Is "Fat Fritz 2.0" better than "Stockfish 13" and worth the money
In my humble opinion "Fat Fritz 2.0" is not better than "Stockfish 13". I rate the engines as equally strong. Both engines are improvments over last version of "Stockfish 12 NNUE".
To keep up to date with engines rating I recommend https://ccrl.chessdom.com/ccrl/404/.
"Fat Fritz 2.0" is expensive and unless one need the updated "Fritz" interface or an updated opening book or wants 6 free monts premium account at Chessbase its not worth the money. However my main favorite engine is "Fat Fritz 2.0" since I like its style of play best and it matches my own evaluations more often than "Stockfish 13".
"Fat Fritz 2.0" is however very useful for particullary Correspondence Chess players and also for professinal players since in many positions "Fat Fritz 2.0" suggests complelty different ideas than "Stockfish 13". The neural networks different training methods and what they are based on leads to wastly different playing styles and different ideas and evalutions. In my own personal experience the two engines often evalute the same positions differently even when they suggests the same move.
For ameuter players who dont know how to install and configure engines its also recommend to buy "Fat Fritz 2.0" to get it preinstalled, I had to help a friend to install "Stockfish 13" in hes older "Fritz" program at hes computer.
Update added 2021-05-18 "Komodo" realises "Dragon 2"
2021-05-04 the "Komodo" team realised a new version of their engine "Komodo" called "Dragon 2". It predessesor "Dragon 1:s" search was much based on last version of the traditional engine "Komodo" while "Dragon 2:s" uses a new search that unless I missintererpret the description is using an interesting design since sort of uses both "NNUE" as well as a traditional engine since it uses evaluations stored in the network depending on if the position already exist in the network or not.
"Dragon 2" NNUE network was trained using the last version of the traditional engine "Komodo", a former number 1 Chess engine, and its realy an interesting concept to sort of create a hybrid between "NNUE" and a traditional engine.
For ameuter players who dont know how to install and configure engines one can buy "Dragon" from Chessbase in a bundle with the latest "Fritz" interface program to get it preinstalled.
I dont think that "Dragon 2" is the best engine yet, but it is certainly a contender for the being the best Chess engine and is an engine that one can follow up how it develops in the future.
Old "Komodo" heavily relied on evaluation and with rather than depth rather than long variations ala "Stockfish" and I assume "Dragon 2" used the same approach.
Added 2021-05-23 Neural network training
The training of neural networks determents strength and style of the engines. "Lc0" neural network is based self teaching games alone. Some "Lc0" clones uses a different training methods like learning from other engines matches or other approaches. At least one clone are based on human games played at Lichess. This leeds to that "Lc0" and different clones has different elo and different evolutions.
"Dragon 2" NNUE network was trained using the last version of the traditional engine "Komodo", a former number 1 Chess engine, and its realy an interesting concept to sort of create a hybrid between "NNUE" and a traditional engine.
Engine with NNUE technolegy neural networks uses different training methods. "Stockfish" training is from latest none NNUE "Stockfish" based games against it self. "Fat Fritz 2.0" neural network training is mostly based on "Lc0" style self play learning and also learning from many top level human games. This leads to that "Fat Fritz 2.0" and "Stockfish 13" has different evolutions even throght they have the same base. "Fat Fritz 2.0" is almost a mix beetween "Lc0" and "Stockfish 12 NNUE".
"Dragon 2" neural network training are based on games from last version of "Komodo" playing games against it self. "Dragon 2" also stored evaluations in the neural network. "Dragon 2" is therefore almost a mix beetween "NNUE" and traditional engine.
It would be interesting if "Houdini" also came in an NNUE version trained from last version playing games against it self.
It has also appeared NNUE network making tools like https://github.com/dshawul/nnue-probe.
Other controversal Chess engines
Many older engines has also controversals. "Rybka" was at some point the best Chess engine for years. It became accoused and banned for plagiation of two completely different much weaker engines. After the banning the engine was slightly modified and developed by its maker and became a version of normal "Fritz". ¨ Normal "Fritz" versions are not based or evolved from each other but are different from scratch licensed to Chessbase.
An other controversal commercial former nr 1 engine was "Houdini". I read at some forum that it turned out to be based on some early "Stockfish" version with comments and variabels translated and also lots of code copy pasted from many other engines plus some major changes of evalution and some original alghoritms. The engine was never banned but other engines has become stronger than it.