The players and conditions

In a deep match on top-of-the-line hardware, Fat Fritz 2 faced a special configuration of Leela, using its best and largest neural network inside a fast binary called Ceres, all while enjoying a dual GPU system that housed an RTX 3080. The RTX 3080 is practically a unicorn these days due to the bitcoin resurgence, and as a result sells for around $2000 on average. Fat Fritz 2 ran on a 14-core Intel CPU, and both were playing a rapid time control of roughly 35 mins. 

The Game

As is typical of such matches, the engines were provided the basic opening moves to set up the start, after which they were entirely on their own. In theory, Leela should enjoy an edge in this phase as the neural networks are notoriously strong in the opening, but Fat Fritz 2 is a different animal, even in the NNUE world. What makes it different is that instead of training on a traditional engine positions and evaluation set to a fixed depth as Stockfish's NN was, Fat Fritz 2 was actually trained on Fat Fritz 1.0 games, positions, and evaluations. For the Leela net aficionados, these were all at 1600 nodes/position, double what is typical for this level of data creation. So its opening knowledge is taken from Fat Fritz 1.0, though whether this will be enough to counter the much larger neural network it faces is up for debate.

This was the last book move provided for the engines. This positions makes its firs appearance in Mega 2021 in a game in 1957 with none other than Najdorf playing it with black, but it is still popular and recent players taking up the gauntlet with black include teen phenom GM Alireza Firouzja and GM Amin Bassem from Egypt. A number of continuations exist here, with the most popular being 10. Re1, but Fat Fritz 2 prefers the second choice...

This powerful shot first saw regular play by the Bulgarian master Kiril Ninov in the 1970s and 1980s, but it failed to get any traction, and it wasn't until 1990 when Yussupov played it against Kasparov that it caught the collective attention of grandmasters. Though Yussupov lost that day, Huebner came back in 1992 in Dortmund, and defeated the very same Kasparov with it. Karpov then became its next more serious proponent in the mid 1990s when he played it in the Amber tournaments. Even last year in the Cap d'Agde event, Bulgarian GM Cheparinov played it twice against Indian GM Bassem with each scoring a win. 

After 10...dxc5 11.dxe5 ♘e8 White played 12.e6

Though a good case can be made for 12.♘a4, this was the choice of Karpov, Huebner, and even Carlsen. Wait, Carlsen played this too? Sure, and of course we are referring to ... Torben Carlsen, the correspondence grandmaster! The point it fairly clear, the pawn is a goner anyhow, so why not use it to weaken Black's position?

Five moves later, this position presented itself.

The move is 17.e5! While Black does enjoy a beautifully centralized bishop, now it has been cut off from the kingside, which is starting to look like a piece of swiss cheese with those gaping holes in the dark squares.

For the record, this very position with 17.e5 was reached by GM Carlsen (Torben) in 2007 in an email tournament. It ended in a draw.

Fast forward the position until move 23 when White just played 23.♖f3-g3

The position is beginning to look almost catastrophic. Black's two bishops may be 'sweeping' the board, but they are biting dust and are cut off from where the action is really taking place. White is threatening to continue to build up the position with h4, ♗e4, ♘f3, etc. For lack of good options, Black played 23...h6 and was served up with the powerful sac that set in motion the end.

Here Fat Fritz 2 played 24.♘xe6!! and while other engines may indeed find it too, it brings to mind the quote by the late Rudolf Spielmann who once said that he could find every combination that Alekhine found. The problem wasn't finding the combinations but setting up those positions!

Speaking of which, by all means play the move on the diagram above to see how it looks.

After 24...♘xe6 (forced) 25.♕xg6+ ♘g7 26.♗e4 we can see why this looks so attractive. 

There is no immediate win or return of material, but White's initiative here is crystal clear. White continued to position his pieces and build up the attack while Black tried to organize a defense, until this position was reached after 32 moves. 

Can you guess what White played here? It is without question far from obvious. White played the astonishing move 33.♗h4!! 

In fact, this move deserves a diagram.

At first glance this move can seem so odd that one asks oneself whether this is some really odd computer move bringing to mind the strange plays in the pre-AlphaZero age, where the oddest moves would be justified by reams of calculations and billions of nodes, but otherwise inexplicable.

No, this not some tactical nuke dropped in Black's position, it is in fact a deep positional idea. The point here is that White plans to play ♕h7+, a  move that has been threatening for 10 moves now, followed by the g-pawn with g6, which opens the diagonal for the ♗h4, and prepares ♘f6. A beautiful concept, which explains the paradoxical looking 33rd move.

The play continued until move 41, when a cute last tactic presented itself:

This move invites Black to fall on his sword, since Bxd3 leads to mate after Qh8+. The game is clearly over though they played on until move 56.

Game with replayer