Your best bet will be the LD2 net, if you're on a slow machine without decent GPU @Leon

It's a smaller net, so it's notably worse for deep analysis type business, and retains a lot of tactical mishaps, but it's general evaluation is similar to T40, and it will be stronger on a weak machine.

I usually make Komodo 10 make moves against Leela and try to see what happens. Many times the evaluation goes down when Komodo plays Leela in hypodermic (Modern/Pirc) or blocked (positional Winawer) lines. Quite often Leela can go down by half, sometimes even to only +0,20.

Or maybe Leela has a high White evaluation because it runs slower on my laptop before it can see Black's counterplay ¿ Who knows.

an ordinary chessplayer wrote on 09/12/19 at 19:40:39:


brabo wrote on 09/13/19 at 06:56:08:


I didn't mean that new opening preparations make me play worse in the opening. They make me play worse "overall". I do a rather sophisticated errors-by-phase-of-game meta analysis of my play, and stuffing my brain right before a game leads to later errors. During my school-days, cramming for a test was also a net negative strategy, despite obviously leading me to answer some specific questions correctly.

Winning a lot of opening games due to preparation is fine if you don't lose any additional middle-games or endgames due to the same preparation. But for some reason, I do. My best hypothesis is that intensive opening study messes with the evaluation function somehow. This is entirely aside from the missed opportunity for relaxation before a game, which is a big contributor to endgame play.

Unless I am woozy before the game, I *never* lose the opening. And I actually like it when strong players have obviously prepared against me. I try to verify that in the post-mortem. My performance rating in these encounters is better than when they just play chess.

fling wrote on 09/13/19 at 08:27:10:


You have to take into account that there is a scaling issue here. A faster GPU is the same as letting the engine think longer, i.e. a longer timecontrol. Small nets are good for "blitz" because the engine needs to be able to do some minimal searching to see enough tactics. But if the larger nets are better on faster hardware, that also means they are better for longer analysis.

Leon_Trotsky wrote on 09/12/19 at 22:21:03:

The question is: Who is correct?

Have you followed the given lines by Leela, figured out where she diverges from SF/Komodo/etc, and gotten the impression that she just misses all avenues of counterplay, or why she so grossly 'overestimates' the chances?

https://lichess.org/analysis/standard/2r1r3/1p3nkp/pP1pn1p1/P2N1p2/2P2P2/R3P2P/6... in this sample position out of a Dutch Leningrad, SF @depth32 spits out +0.5 with a variety of moves (Rc3, Rc1, Ra4, Bf1). White Advantage, but nothing dramatic.

https://www.chess.com/computer-chess-championship#game=44 Leelenstein did go for one of the SF suggestions (35.Rc3), but thinks White is completely winning (+2.6), and btw has been thinking so.. for pretty much the entire game!
The initial position after the opening book is evaluated by SF as +0.29, and by LS as a whopping +1.59

As late as move 54, SF believes to be in a tenable position (+0.76), while Leelenstein is long confused why there's no resignation coming (+4.26)..
After 55.Bf1, SFs eval suddenly explodes in White's favour, and never recovers.
Did Leelenstein overestimate White's chances here..?

MartinC wrote on 09/12/19 at 21:30:42:


I agree with Martin here. Lc0 does not run well on a laptop. Even on a desktop, like my old desktop with core i5-6600K and RTX 2060, it is not supposed to be great with the larger nets. EDIT: It should have been RTX 960.

I think you should try one of the nets that are claimed to be better for CPU in case you want to use Lc0 on your laptop.

The point is that you judge the engine based on a rig that is not good for it.

an ordinary chessplayer wrote on 09/12/19 at 19:40:39:

I can show you many of my won games in which I can prove that they were won largely due to new preparations made just before the games. Besides I also make analysis after a game, have files of likely opponents, study analysis from others and repeat before a game with chesspositiontrainer old files. My private library has in the past 15 years grown to several 1000 of lines which I can use. Still I notice my library lacks very often answers on lines my opponents play.

So I am convinced there is a link between success and preparing games. I play around 2300 level but I also believe this link applies to lower levels.

It makes no sense to talk about using Lc0 in preparing games just before the games if you are not convinced of that link.

I don't know your situation but personally I do quite some sport and puzzles during the year so during a tournament I can fully focus at the preparations and of course also the games.