Fat Fritz

     Recently I noticed that Fritz 17 is out (Price: $89.95) and in addition to the usual addition and or improvement of various bells and whistles, it comes with the Fat Fritz engine. 
     One improvement is the 3D boards based on real-time ray tracing, but they require a powerful graphics card with a NVIDIA chip or otherwise they run on a CPU only for demonstration purposes. 
     The Fat Fritz engine is an extremely strong neural net engine inspired by Alpha Zero, which produces “human-like strategic analyses of world class quality.” Fat Fritz is based on the technology that created AlphaZero and is based on the open-source project Leela Chess Zero. BTW, Leela can be difficult to install unless you are computer literate. 
     One of the key tenets of the Leela Chess Zero is that it uses nothing except what it learns of its own accord. With Fat Fritz an attempt has been to make it the strongest and most versatile neural network by including material from sources such as millions of the best games in history played by humans, games by the best engines including Stockfish, Rybka, Houdini, and more, endgame tablebases, openings and millions of self-play games, etc. 
     One reviewer found that the strongest Leela engines and development versions of Stockfish are marginally stronger than Fat Fritz. That being the case, why get excited about paying $90 for a commercial neural net engine when Leela and Stockfish are both free to download and are slightly stronger? 
     The reviewer stated that if you are only looking for blunder-checking, then Stockfish is OK, but if you want understanding and want to generate your own ideas in conjunction with the engines, “...you need to use all the tools at your disposal, and you need to understand their strengths and weaknesses.” 
     That’s because engines like Stockfish, while they excel at calculation, defense, and tablebase endgames, they lack positional understanding of things like the initiative and compensation in situations where there is a material imbalance. For this reason the reviewer observed that Fat Fritz feels more comfortable in messy positions. 
     Neural network engines like Leela and Fat Fritz have the positional understanding, but their weakness is that they can sometimes miss concrete continuations and their endgame play is not always very accurate.  Consequently, different engines must be used together if one wants to generate insights that they might not otherwise have. As the reviewer pointed out, neural net engines are valuable because they bring new ideas into the game and new notions of what’s playable. 
     He also pointed out that when you see a 0.00 evaluation by Stockfish it’s not considering factors like the initiative and who has more ways to win which is something the neural engines, like strong human players, also take into consideration. 
     All this is interesting, but my feeling is that unless you are a very strong player or are playing correspondence chess at the highest level, there’s not much reason to invest in Fritz 17 just for the engine. However, if you are interested in any of the other features, that is another matter.