Hier die Einschätzung eines dev auf Discord:
<@!437624409923125268> Lc0 endgame play is very strong. The style is not beautiful but the results are very good. There is an experiment by <@!143240917602664448> that I confirmed at a different TC : against SF, lc0 is better at beginning of game and at endgame. She is relatively weaker when the number of piece is medium. With the same number of pieces, she is weaker if there is queens on board.
My interpretation of this is that the main issue remains tactics : when there is a medium number of pieces, there is a lot of open lines and so a lot of moves to investigate. That increase the probability to find a combination and SF is good at that.
Results in long TC tournaments or with strong hardware confirm this hypothesis (TCEC, CCC, <@!443676102775406592> stream) : Leela is better with balanced and sound start position (quasi unbeatable in this conditions) and weaker in unsound or unbalanced position especially with open lines (tactical blunder despite a good play).
I don't think that the solution for tactics is the net. I think it is likely the search.
There is also games at CCC right now where SF NNUE annouced mate in endgames before SF : that show that SF play in endgame is not so good, TB is not enough a compensation for the lack of knowledge in the eval function. SF, even in endgame, do not investigate the good lines. It lose a lot of time on lines that are not interesting.
Lothar