Stockfish and Leela Chess Zero developers have joined
together with Lichess to condemn what they see as false advertising by ChessBase
of the recently released Fat Fritz 2 chess software. The promotional material
suggests the chess engine included with the software is “the new number 1”, and
that the key contribution is Albert Silver’s in training its neural network.
The Stockfish developers, meanwhile, point out that Fat Fritz 2, which is being
sold for around 100 euros, is the free Stockfish with a net that adds no
noticeable increase in playing strength.
Stockfish, which powers evaluations here on chess24, is the
current king of the computer chess world, playing at a level that’s literally
inhuman. It’s also a free, open-source engine. How can a commercial chess
engine compete? Well, that’s where Fat Fritz 2 comes in. The software, which
includes Stockfish, now complies with the Stockfish licensing requirements
(after the developers say they pointed
out a violation) so is perfectly legal to sell.
What has enraged the Stockfish developers, however, is the
perceived false advertising. Albert Silver’s contribution was to train a different
version of the neural network that works within the latest version of Stockfish,
but the promotional material around the release puts all the emphasis on that
work rather than Stockfish itself. To add insult to injury, the claims
of a leap in strength of Fat Fritz 2 relied on comparing it to Stockfish 12, but most if not all of the
difference could be explained away by the fact that Fat Fritz 2 was built from
a later developer version of Stockfish with a higher rating.
The good news for chess fans is that that advertising trick
led to the release of Stockfish 13 ahead of schedule, restoring a situation
where pure Stockfish is ahead of or very close to Fat Fritz 2 in rating, depending
on which list you check. You can download the latest version 13 here.
A computer chess revolution
An article at
ChessBase announcing the new software begins, “Fat Fritz 2.0 is the
successor to the revolutionary Fat Fritz, which was based on the famous
AlphaZero algorithms.” There have been revolutionary changes in chess computing
in the last few years, but Fat Fritz certainly wasn’t the driving force behind
them. DeepMind’s AlphaZero shook
the chess world in late 2017 by showing that its reinforcement learning algorithm needed just four hours of
playing against itself to go from zero knowledge to a level superior to Stockfish.
That four hours relied on having the phenomenal computing
power of Google to draw on, but the open-source Leela Chess Zero project showed
that the same principles could work if you drew on the computing power of an
army of volunteers over a much longer period of time. Fat Fritz 1.0 was not
based on a new adaptation of AlphaZero but on taking the Leela Chess Zero
code and tools and training a network differently.
The drawback of Leela Chess Zero is that to run it well
requires a high-powered graphics card, but a new development, NNUE (Efficiently
Updatable Neural-Network-based evaluation functions), allowed CPUs to
combine the speed of the traditional chess engine calculations with neural networks.
Stockfish adopted that new technology to regain the lead in the computer engine
Then, as with Leela Chess Zero, Albert Silver pivoted to using the new strongest
chess engine to “power” the latest version of Fat Fritz.
An open letter from Stockfish, LCZero and Lichess
The Stockfish, Leela Chess Zero and Lichess teams combined
to pen the article Fat
Fritz 2 is a rip off, which catalogues what they claim is a track record of
deceptive claims by Albert Silver. It concludes:
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.
One possible defence, or sales approach, for Fat Fritz 2 would
be to point out that the strongest current chess engine (the free Stockfish) is
bundled with a database, user interface and subscription to make a package that
might be attractive, especially to newer, less technically aware chess aficionados.
That wouldn’t allow you to hype it as a new invention that everyone needs to
ChessBase have now decided to double down on the original
claims in a new article that obliquely addresses the criticism, beginning, “To
say that Fat Fritz 2 has been making waves is an understatement”. “When you
purchase Fat Fritz you are certainly not paying for the free Lc0 and Stockfish
they come bundled with”, claims Albert Silver, while making a comparison
between Stockfish as the “car” and his altered neural network as the “driver”
that’s unlikely to make him any new friends in the open-source community.
The article focuses on the time and expense put in to
training the network while playing down the importance of rating now that it’s
become clear Fat Fritz 2 is behind or roughly equal to the latest Stockfish with its
default neural network, instead claiming “a powerful and
different perspective” based on training the network with Fat Fritz 1 rather
than Stockfish evaluations.
The computer chess wars
Such qualitative claims are much harder to disprove, but the history of
chess engines suggests they should always be treated with caution. When Rybka
reigned supreme in the chess world its author Vasik Rajlich used to talk about
how its intelligent search allowed it to operate at a lower depth than its
rival engines. Later it was claimed Rybka had copied another engine’s code without
attribution and manipulated the depth shown to hide the fact, with Rybka stripped of
the four World Computer Chess Championship titles it won from 2007-2010.
Adapting existing engines has, until the
appearance of AlphaZero, been an incomparably quicker means of creating a new
and competitive engine, and with open-source software that approach is
legitimate though sometimes ethically questionable if the resulting engine is
then sold for profit. Houdini 6, another high-flying engine, is currently
unable to compete in the TCEC computer championships after being accused of
being a Stockfish clone.
In the case of Fat Fritz 2 the use of Stockfish
isn’t hidden, though it is heavily downplayed in the ChessBase messaging. The
arguments are sure to rumble on, with chess engine experts assessing the real
added value or not of the new product. Caveat emptor, as they say – just how
much stronger do you really need your 3500+ rated chess assistant monster to