Based on the playing style and progress in strength of the leela queen odds network here are my predictions for chess engines of the not-so-distant future. These are pure predictions based partly on evidence and partly just my intuition and are not necessarily meant as recommendations or a guide for chess engine development, or even something I will definitely attempt to achieve myself. I would be more than happy to be proven wrong and change my mind. My aim is an attempt to start a predictions thread for chess engines and is meant as a fun engaging activity, and to crowdsource predictions :P.
-
Judging by the self-play games of leela with search-contempt at 20000 nodes, which significantly favor black, it appears that the strength of chess engines is limited by the space of explored chess positions rather than the first move odds that white has. Subsequently, the proportion of draws in chess will reduce with either white or black winning a larger and larger fraction of the games, and the stronger of the two engines taking the win in increasing fraction of the games, and move odds will play a lesser role.
-
A lot of people in the chess community, myself included have previously held the opinion that current top chess engines play close to perfect chess and would easily draw even against an oracle of chess which plays every move perfectly. In light of the developments in the Leela Chess Zero community over the past few months, my opinion has been completely overturned. There are still plenty of positions that are totally unexplored by the current engines. To emphasize this point, I make a very specific prediction. It will be possible to beat stockfish from the starting chess position with black in tcec classical time controls. To be even more specific, the playing conditions would be the same binary and network architecture for both leela and stockfish, playing with the same time control, and using the same hardware that was used in the recent tcec Season 27 Superfinals. The only change would be that the leela network would be trained with games generated by search-contempt or other forms of contempt which expands the space of explored positions. Very likely, it will be by using training iterations similar to that used to train the leela queen odds network, without even using stockfish as an adversarial opponent, but just from asymmetric self play. It is important to note that this statement has no bearing whatsoever on whether future versions of Leela Chess Zero playing black will beat future versions of stockfish playing white in classical time controls from the starting chess position.
-
This point naturally follows from the previous point. Currently the popular opinion is that the starting chess position is dead and the forced openings used in top engine competitions is what is keeping engine chess alive. I believe it is the complete opposite. The starting chess position is alive and well and it is the forced openings which are dead. In some future editions of engine competitions, all games will begin from the standard starting chess position and the engines will create winning chances for themselves.
-
Leela Chess Zero of the future will beat the chess world champion in a match with classical time controls giving queen odds to the world champion. This will likely involve 5-10 selfplay iterations of training, with probably around 100k training games for each iteration and will probably be achieved before the end of 2025.
-
Mcts/uct based chess engines like Leela Chess Zero will move towards smaller network and deeper search for training or even for competitive games. They will be trained on fewer but more high-quality games trained using deeper search or other kinds of contempt. Current training games use 800 nodes per move. This might likely change more towards 10000-50000 nodes per move.
-
Chess still remains a highly tactical game. The distribution of chess openings played by top engines will drift towards more tactical and dynamic style than the more positional style currently played. We might be seeing a completely different set of openings than we are currently used to. Openings like kings gambit or stafford gambit or other newly discovered openings could see a resurgence. More and more games, even in classical time control will end in checkmate or draw by repetition or perpetual checks in the middlegame, before even hitting a single tablebase position. The focus of engine chess will transition in a direction of less endgame to more opening play. Also, Magnus might be right after all and we might see vanishingly fewer and fewer fortresses.
-
Leela Chess Zero or other self-play based engine's top preference on the first move will drift from d4 to e4 openings. This is deduced from the playing style of the current leela queen odds network which prefers e4 in the standard starting chess position (without the queen odds) playing as white.
A final note: Although a lot of the above points are heavily influenced by the discussions and results shared in the Leela Chess Zero discord channel, the above are purely and completely my own opinions and beliefs and none of them necessarily reflect those of its members or have even been expressed in the server. So they may be heavily biased and are subject to change frequently.
Link to the Leela Chess Zero website: