A few days ago, ChessBase released Fat Fritz 2, described on their website as the “new number 1” chess engine “with a massive new neural network, trained by Albert Silver with the original Fat Fritz.” They advertise Fat Fritz 2 as using novel strong ideas compared to existing chess engines, but in reality Fat Fritz 2 is just Stockfish with a different neural network and minimal changes that are neither innovative nor appear to make the engine stronger.

As this is not the first time something like this has happened involving both ChessBase and A. Silver, we would like to share our impression of these releases.

[EDITORS NOTE: Since this article's publication Chessbase has removed the claim that Fat Fritz is "the new #1" from their website. You can see an archive of the old version here.]

DeusX

In July 2018, Silver secretly sent the Leela Chess Zero engine with a custom neural network he had trained to participate in TCEC under the name DeusX. The network was trained using scripts from the Leela project, and the network architecture was the same as that used by the Leela engine (see this older Leela blog post for details). Training such a network is not unusual; members of the Leela project do so regularly to test ideas. But they have never pretended to have created a new engine by training a new network.

While an overwhelming part of DeusX’s strength is inherited from Leela, Silver downplayed the Leela work enormously in an interview, suggesting that he did in a few months what had taken other engine authors decades. Silver described himself as the engine author even though the engine itself was Leela without significant modification.

Fat Fritz

The following year, Silver released an updated version of the DeusX network under the name Fat Fritz, sold as a part of ChessBase’s Fritz package for €79.90. Once again, it used the Leela engine without functional changes (the changes made included modifying the name and author strings, and some default parameter values).

Fat Fritz was marketed as if it were an innovative engine, instead of being just a renamed Leela. As an example, the product description begins, “It’s a semi-secret development, an AlphaZero clone, engineered over the past nine months,” and doesn’t mention Leela. Probably the closest to what can be called an “attribution” is a brief mention in the middle of one of the articles, saying that Fat Fritz uses Leela “as a foundation.” In reality, Fat Fritz is Leela, but with a different net. Even this article begins by describing an “inspiring” talk given by a DeepMind employee to ChessBase programmers, supporting the false impression that ChessBase played a significant role in the development of the Fat Fritz code.

In ChessBase articles, the Fat Fritz “engine” was described in a way that implied it was stronger than Stockfish and Leela, but the evidence was questionable. Silver’s Stockfish comparison, for example, used an outdated version of Stockfish even though the development version was known to be considerably stronger. Similarly, when compared to Leela, the strongest configuration of Leela was not used.

Gian-Carlo Pascutto is an author of several strong Chess and Go engines and contributor to the Stockfish and Leela Chess Zero projects.

Fat Fritz 2

In 2020, Stockfish, Leela’s main competitor, started to support NNUE, fast neural networks that can run on a CPU. This feature improved Stockfish significantly, restoring its status as the strongest existing chess engine.

The Stockfish team had the same painful experience as the Leela team when Silver decided to jump on the hype train again, and released Fat Fritz 2, sold by ChessBase for €99.90. It is now Stockfish that has been copied instead of Leela, but the overall style is unchanged:

• As with Leela and FF1, only minimal changes have been made to the Stockfish engine (again, the name of the software and the authors, and some default parameters). Even though the Stockfish engine is critical for playing strength, it is mentioned only briefly and the impact of the Fat Fritz 2 neural network over the one used by Stockfish is greatly overstated. The product description says FF2 is “learning from the surgical precision of Stockfish’s legendary search”, but it isn’t learning from Stockfish, it is Stockfish.
• As before, Fat Fritz 2 is advertised as the strongest engine, but the only results presented are against an older version of Stockfish, and not the version used by FF2. Independent results show that current Stockfish versions on which FF2 is based are in fact stronger than FF2, suggesting that Silver’s net does not add playing strength.
• ChessBase has published an interview with Silver describing the work. In the text accompanying the interview, they describe Silver as the “inventor” of Fat Fritz 2, and say that he started the project “almost completely from scratch.” In reality only minimal changes were made and Silver likely did not author them.
• Silver describes FF2 as a “completely new” neural network, but it uses the Stockfish topology and differs from Stockfish’s network only in layer sizes. The interview article also says that Silver “came across a new neural network technology from Japan,” presumably because NNUE was originally implemented in Shogi engines. While it may sound as if Silver was responsible for bringing this innovation to chess, he did not implement NNUE and used mostly Stockfish tools to train the network.

Conclusion

It is sad to see claims of innovation where there has been none, and claims of improvement in an engine that is weaker than its open-source origins. It is also sad to see people appropriating the open-source work and effort of others and claiming it as their own. 

Everyone is permitted and encouraged to modify and improve code from Stockfish/Leela while giving credit; that is the intent of open-source software. Everyone is allowed to copy Stockfish/Leela and sell them, provided the terms of the Stockfish/Leela license are met. But don’t pretend that the product being sold is something it isn’t.