Is Leela still zero ?

[bnst]

I find the Leela project to be very interesting for two reasons:
Perhaps the most important is that it shows what can be done with modern AI methods and perhaps can serve as a model to be used in many other and perhaps more important areas than chess.

Second and the most important here is that if Leela is still properly "zero" it gives us a sort of objective chess.
Most people who play chess and those who write books about chess have very personal biases on what is good or not and many of these biases are well integrated into common "knowledge". I hope that Leela still is properly "zero" even though the rules for the moves are hard-coded and I think the search when it plays has been taken from Stockfish.

Leela gives us a fresh and hopefully objective view on what is good or not. Take something we earlier thought as simple - the piece
values. What does Leela say about these ?
One thing I think has been learned is that it is possible to play more agressively against an opponents king even if it means opening your own king a bit. Perhaps also Leela has shown us that old-fashioned chess engines were too conserned about material and has shown us a way back to a more romantic chess with sacrifices and agressive play.

Someone should write a book about what has been learned from Leela (perhaps somebody already did ?).

So I end up here with two questions:
Is Leela still properly "zero" ?
What do people here feel that they have learned from Leela's "objective" view ?

Best regards
anst

P.S. I know that even Leela will have biases because just as us humans it will perhaps go in certain directions when it learns, but if I understand right the learning net has been restarted several times so at least we have got several views unaffected by our biases.

[Lazy_Frank]

1. Can't answer to question "Is Leela still "zero"?"

2.1. More often engines start to use and understand importance of so-called "fawn" -
pawn in front of opponent king, covered by pawns and restricting king mobility.
e.g. setups like: white pawn h6, black pawns h7-g6 and black king Kg8.

2.2. Pawn structures in early opening phase, especially in openings which is aggressive to pawn pushes on early phase.

[dkappe]

Leela uses a very different type of search than Stockfish. (See Monte Carlo Tree Search).

[kaissa]

Someone should write a book about what has been learned from Leela (perhaps somebody already did ?).

Not directly answering your question but I do not believe that there is a need for a book about Lc0 when there is already an amazing one titled Game Changer about AlphaZero.

[bnst]

1. Can't answer to question "Is Leela still "zero"?"

2.1. More often engines start to use and understand importance of so-called "fawn" -
pawn in front of opponent king, covered by pawns and restricting king mobility.
e.g. setups like: white pawn h6, black pawns h7-g6 and black king Kg8.

2.2. Pawn structures in early opening phase, especially in openings which is aggressive to pawn pushes on early phase.

Thank you for interesting insights. This supports my impressions too. Sadly I'm a very bad player and I love it when people with good skills share their knowledge. I think I read somewhere that Carlsen was using ideas from AI play.

Best regards
anst