"Fat Fritz seemed to be from an entirely different plane of existence!"

by Tanmay Srinath
11/4/2019 – Fat Fritz for the club player: While the top players and grandmasters are leaping into the neural network boat, the question of value to amateurs is still left somewhat unanswered. Isn't one engine as good as another when it comes to sub-world-class players? We gave generous use to a former critic of Fat Fritz, TANMAY SRINATH, and not only was he extremely impressed, he was and was kind enough to share his experience with the readers.

ChessBase 15 - Mega package ChessBase 15 - Mega package

Find the right combination! ChessBase 15 program + new Mega Database 2020 with 8 million games and more than 80,000 master analyses. Plus ChessBase Magazine (DVD + magazine) and CB Premium membership for 1 year!


What sets Fat Fritz apart (part 1)

I wanted to race back into tournament play. However, because of a gap that exceeded a year and a half, I was incredibly rusty, and my opening preparation was non-existent. Then, I had to buy a new engine to analyse my openings in, and that is when my interest in analysis started to grow. It was also my first foray into the world of the aggressive MCTS [Monte Carlo Tree Search] engines, beginning with Komodo 12.

Believe it or not, most of the times the engines provide us wrong solutions to our problems, and are sometimes can be more useless than a rank beginner. The problem with machines is that they are non-interactive. For example, in a position of immense complexity, a human might struggle to continue in the best way, but an engine quickly finds a forced line leading to a piece vacuum and a draw. This problem, called the 0.00 conundrum, is so named because the engines fail to differentiate between lines that give static (aka boring) equality, and dynamic equilibrium, where there is a lot of life left in the position. I have on numerous occasions struggled with this, and it is only my intuition and understanding that helps me find a way through the dark forest and escape unscathed, more often than not thankfully. Another important thing they fail to take into account is the accuracy of play of their opponent, or the complexity of the position and the capabilities of the players, thus proving to be a good instrument to criticize the top players, something IM Sagar Shah looked at in detail recently.

It is said that the neural network simulates the human brain. Is this proof that we are yet to reach our potential? | Photo: Oracle Blogs

Believe it or not, but MCTS engines rarely suffer from this problem, because unlike the traditional engines which aim to find a move that wins the maximum number of pawns, the MCTS engines aim to find the move that has the greatest possible chance of winning. The former and the latter might not seem to make much of a difference, but believe me, it does — while Stockfish might assume that the opponent plays perfectly, Komodo MCTS will not, and thus more often than not it favours the side with the easier play. As an after-effect, it proves to be extremely good at handling gambits as well.

Leela Chess Zero is another example. I have found that in some positions where Stockfish sees nothing but a desert leading to a fight-less draw Leela poses immense practical problems, the sort which even the best engines on the planet sometimes fail to solve. While this might make you feel that MCTS engines are better than Stockfish, Houdini and their siblings, I would interject here and say you are right and wrong. Right in some cases and wrong in some. Instead of categorizing life into a slew of binary options of yes and no and right and wrong, why not find something in the middle?

Alpha Zero for the PC - Leela Chess Zero | Photo: Leela Chess Zero Blog

It was this thought process that ultimately resulted in me speaking to Albert Silver about his latest offering Fat Fritz. I had used the engine once, and was intrigued as to what it can offer, so Albert kindly lent me full use of Fat Fritz for a while. What Fat Fritz is, is something that the author himself has explained in great depth, so instead of going into that let me illustrate what exactly I felt about its capabilities and where it is most useful.

Let me first say that I have used engines extensively, and have found good and bad qualities in almost equal measure in each one. However, Fat Fritz seemed to be from an entirely different plane of existence! Why?

The first thing I noticed about Fat Fritz was the speed at which it judged positions. While normal engines took around five minutes to nearly fully understand any given position, Fat Fritz was astonishingly taking only a minute and a half at most to decide which move or plan it deemed best, and this I feel is not only because of the superior hardware it is run on, but also because of its innate intuition coming from millions of human games. 

This unique non-zero approach resulted in it often contradicting numerous top engines on various positions. Let us take one position for example:


I considered the position after 8...♞xe5 to be playable for White. Initially, that is what Stockfish seems to think too. The alternative 8...0-0! is not considered so seriously by it till beyond depth 30, by which point Fat Fritz is claiming an advantage of over 1 here for Black [i.e. -1].

What is special you might ask? Such quick accurate assessments make our analytical work easier as we need to spend less time in investigating a new position. Also, I noted the evaluation patterns for both engines. While Stockfish assessed both 0-0 and ♞xe5 almost equivalently as better for Black, Fat Fritz was giving a nearly winning advantage for Black after 0-0 and only a slightly better position after ♞xe5. See the difference? Fat Fritz believes fully in its main line, and assesses certain positions a lot better than its counterparts.

For those of you who feel that speed is not really an issue and feel that it is not worth spending money for it, I ask you to look at the following position closely.


For the Sicilian players out there, this is where you really start to sweat — White enters the Fried Liver version of the Wing Gambit after 1.e4 c5 2.b4!?. The reason the Wing Gambit lines after 2...cxb4 3.a3 e5(!) lines were not previously played was because of this particular setup for White, which gives him very dangerous attacking chances. However hard I looked at this with other engines, I was unable to find a suitable response for Black, with the very greedy ones wanting to take on d5.


It was Fat Fritz that put the matter to bed once and for all with the rare but best 7...b5!

This was a move I had never seriously considered before, but the more I looked at it the more I was convinced that White is in-fact closer to a loss than a draw or a win here — Black's aggressive counter-attack means that White doesn't get time to develop his famed assault on the f7 square. I feel White is a lot worse here — and that is not the end of his troubles — his winning chances are absolutely zero, and that's sad news for gambiteers across the globe.

Chess is perhaps as inscrutable for us as the vastness of the universe! | Photo: Delightful Knowledge

I was going through the game Shivananda vs Neverov once, and I stumbled upon this position.


Here the Ukrainian GM instead of playing 9...d6! (in the diagram) played 9...♞bc6 and won only after further errors from his Indian opponent. However, Fat Fritz immediately finds the refutation to White's risky opening play here with 9...d6!, and assesses it at a whopping -2.50! It is not only this move — it clearly shows a plan to blow open White's lacklustre position! This is what I had been waiting for a long time — an engine that can play a logical set of moves that we humans call "a plan"!

I am confident that this new development will take the game forward. In the second and concluding part I shall speak more about the unique features of Fat Fritz.

Continued in Part 2 (shortly)...

Tanmay is a young chess player from Bangalore, Karnataka, currently pursuing both chess and engineering at BMSCE Bangalore. Tanmay is also a Taekwondo Black Belt, who has represented the country in an International Tournament in Thailand. He is a big fan of Mikhail Tal and Vishy Anand, and sincerely believes in doing his bit to Power Chess in India!
Discussion and Feedback Join the public discussion or submit your feedback to the editors


Rules for reader comments


Not registered yet? Register

omavein omavein 11/5/2019 01:14
Stockfish on the first position : 8..0-0 , instant -1,3
LuckyLuke LuckyLuke 11/5/2019 12:00
In my opinion Shirov wasn't in danger, maybe he wasn't better at some point.
I agree with you it's not an easy line to face unprepared and b5 might be a better practical move.
It's astonishing how you can call it shallow analysis when you are just giving b5 and claim it is the best move by just saying look deep. When you have made the analysis to back up your claims please show. And this means you have to compare b5 with Nd5. Also it would be interesting how Fat Fritz evaluates b5 and Nxd5 at higher depths.
Karbuncle Karbuncle 11/5/2019 10:02
@scorpion29 How on Earth can you claim AB engines are "exhuasted" when Stockfish is literally improving almost daily with the new dev builds. How do you think Stockfish managed to win TWICE in back-to-back tournaments against the two best NN engines in the world? You have no platform for your argument other than WANTING to believe AB engines are no longer worth investing in. Seeing as how SF won the most recent tournaments, that would tend to suggest they still have superb strength and capabilities for refinement. Don't be a hater, appreciate that there's an unwarranted bias against AB engines. If NN engines are supposed to be RIGHT NOW the best on the planet, it makes no sense they would lose to SF. Both types will continue to compete and improve.
Scorpion29 Scorpion29 11/5/2019 05:20
It is not just about finding the move, it is also about evaluating it correctly! I am surprised that the level of superficiality here. I will address this in my upcoming article. Please read my comments in detail before making claims about the engine or the cpu. What I wanted to illustrate is that Fat Fritz quickly discards other moves and assesses 7...b5! as the best way to fight for an advantage. Let your smallfish run, it'll start liking Nxd5 more and more. It's about depth, contempt and all other such parameters. 7...b5 is considered equal by my stockfish atleast at a higher depth, say 40-45, so I would be delighted to be proven wrong. Of course, there are other moves, but what I wanted to show was that while something like asmfish or smallfish might take 3-4 mins to come to the correct decision, Fat Fritz takes atmost 15s!
I am always looking for ways to improve the content. I will address all your concerns in the upcoming article. Until then, please be patient, and ask all the questions you want!
Boisgilbert Boisgilbert 11/5/2019 03:51
7...b5 is found instantly by Smallfish on my iphone. I agree with ortsac2014 that something is off either with the author’s engines or cpu.
Scorpion29 Scorpion29 11/5/2019 02:26
After 7...Nxd5 the position is unclear after 8.Nxf7! 8.Qh5!? Is also a dangerous move to face unprepared. Study the shirov game properly
You'll notice Black was in serious danger at one point. Instead of shallow analysis try going in deep. Then you'll know.
LuckyLuke LuckyLuke 11/5/2019 12:28
How can white get a good position after 7....Nxd5 instead of 7...b5? I couldn't find a convincing line for white and Shirov won with 7...Nxd5 without much problems.
Scorpion29 Scorpion29 11/4/2019 04:39
it does find it.However, it quickly says -0.30 something after Bxb5 Qxd5 Ba4 with the contempt off. I suggest you go through it carefully before claiming that White is okay. He is not in that position atleast. It finds that endgame playable but I and Fat Fritz don't - Black is in plain sight atleast better if not more.
See, the thing people fail to understand sometimes both engines can concur. I have not attached the exact line after 7...b5! because I deemed it unnecessary. It was given only a cursory glance by Robert Ris and Marcus Reinke in their surveys. However, it should definitely be called the new main line - it gives Black serious winning chances. I suspect that in the other positions your engine is too slow to clearly understand what is best for both sides.
The next article will make it clearer to you why Fat Fritz is from another dimension!
ortsac2014 ortsac2014 11/4/2019 04:03
My Stockfish 10 found 7...b5 very fast. Perhaps the author's computer is very slow.
Frederic Frederic 11/4/2019 03:37
If you are suffering from chiropteral blindness, so am I. I have asked Albert and Maculey to correct the two apparently flawed diagrams.
Arminio12 Arminio12 11/4/2019 03:12
I may be as blind as a bat, but following the first diagram you write:

"I considered the position after 8...♞xe5 to be playable for White (…). The alternative 8...0-0! is not considered …"

In the diagram there's no black knight that can take on e5 (there's a black pawn on that square) and castling is just as impossible (Bf8 hasn't moved yet). Later on, after the second diagram you say 7(!)... b5 is the best move, i.e. we are now talking about move 7 instead of 8 and apart from that the aforementioned alternatives 7 (or 8) ...Nxe5 or ...0-0 are still impossible. What's missing here?
Scorpion29 Scorpion29 11/4/2019 02:13
boris25 There has been a diagram mistake from the backend. I will speak to them and try to rectify the situation. Thanks for pointing it out.
Karnbuncle there is still a lot of scope for improvement in NN engines, unlike those AB engines that have nearly exhausted their evolution capabilities. So instead of trying to debunk Leela you must understand why it lost...maybe a bad neural net...maybe hardware troubles. Of course AB engines are useful in brute force search, but the reason top GMs are moving away from them is due to the 0.00 problem that I have explained here. In the next article you will clearly understand why the present is probably not enough.
boris25 boris25 11/4/2019 01:11
In the first diagramm, how can Black play 8...0-0?
Karbuncle Karbuncle 11/4/2019 12:56
FYI Leela was defeated by Stockfish in the TCEC Cup knockout tournament. In fact Stockfish beat BOTH AllieStein AND Leela back-to-back to win that event. Stockfish also won the TCEC Season 16 Super Final by dominating AllieStein. Leela played so poorly against weaker engines that it didn't even qualify for the Super Final. So while GMs and other may be "jumping on the NN bandwagon", Stockfish is STILL the champion.
Marselos Marselos 11/4/2019 12:17
See it well , take off its opening book , see if it repeats the moves ...