Deathmatch: Leela vs. Stockfish

Quite funny how Lc0 is using a standard decent setup and Stockfish is over the top with huge resource availability and yet this is grounds for an equal match; Lc0 is already the winner here.

I don't think that Leela will win, because she played a little stronger than Andrew Tang (@penguingim1). I think it will be 6-0 to SF10. Well, Leela can take one draw, of course.

I want Lc0 to win so I can lol at the fish fanboy conspiracy theories

@sausage4mash Yes, but how is this possible with a ~2900-3000 level? :)
SF on ~3400 now

The chance of a draw is even very small.

@zevra_chess you seem to have no real information regarding Leela's strength. Leela is at least 3300+ in your scale, also Andrew played a match against a weaker version (your 2900 might be more or less correct). Several tests and tournaments have shown that the currently best Leela is at least the 4th strongest engine. It managed to defeat Komodo in a 30 game match and without a significant number of games it can beat any other engine (newest SF would be pretty unlikely though). On CCCC Leela has won multiple games against SF, in bonus tests on TCEC she also won at least one game and in the tests of people with weaker hardware she also manages to get wins against SF almost all the time.

Tl;dr: Leela will for sure get more than one draw and as 6 games is not much it might even win. In the end its just a too small sample where anything can happen. Still a fun idea though.

Really looking forward to Jerry's stream :)

Is it modified Stockfish?

@chess_tactic2017 it was already answered: "It will be the official Stockfish 10, not the multi-variant fork."

@Mlocik6 Sorry, my study was on Russian language, but modified Stockfish made a draw very good lichess.org/study/CSnzE4WS

So much misinformation here, lol. Lc0 scales badly; AlphaZero does not. Lc0 was built from the preliminary AlphaZero paper, which was missing some important details. The final paper fills in the gaps. In particular, now we know that, in AlphaZero, the balance between exploration and exploitation isn't constant but changes with search depth instead: more exploitation at lower plies and more exploration at higher plies. This will soon be included in Lc0, and it will make it scale much better.