Hacker News new | ask | show | jobs
by semi-extrinsic 3504 days ago
Regarding "38 cores": can anyone comment on whether that's sufficient to perfectly analyse the game in near-real-time, or is it just good enough to provide interesting analysis? I.e. would Stockfish running on 38 cores be competitive with / better than World Championship players?
3 comments
trontron 3504 days ago
modern chess programs are vastly superior to human players, regardless of one or 38 cores. stockfish has a elo rating of ~3350 (probably a little higher when running on 38 cores, not sure). carlsen has 2850. that corresponds with a 95% winning probability per game (or to be more precise: stockfish is expected to win 95% of all possible points over several chess matches)

it's hard to say if that's close to "perfect analysis", because we don't know what a perfect game will look like (chess is not a solved game). but compared to human level, one could say it's close to perfect.

link
kachnuv_ocasek 3503 days ago
I'm not saying Stockfish isn't stronger than any human player (even the world champion), but you can't compare ratings in two different systems like that. Magnus has never played a FIDE-rated game with Stockfish, nor has Stockfish played any rated games with a human opponent.
link
Scarblac 3504 days ago
Even mobile phones have been at World Championship level for years, but there's no idea to know how close to perfection we are. Top computers still lose games to other top computers.
link
throwaway_yy2Di 3503 days ago
It's easy to tell computer engines aren't analyzing perfectly, because when we tell to them to play themselves, they lose a large fraction of their games, as both colors. Perfect play would always force at least a draw from at least one side.

There's no known tractable way to solve chess. There's something like 10^120 move orders [0], and no known way to find perfect play without brute-forcing (almost) all of them. Chess engines can't solve to to the end of a game to see which moves are certain to win; they can only explore to a very shallow depth, and evaluate the horizon nodes by very human-like [1] approximate heuristics.

It looks perfect from a human PoV (the best human players have no chance of winning); but there's still an unimaginably large gulf between chess engines and mathematically perfect chess.

[0] https://en.wikipedia.org/wiki/Shannon_number

[1] https://github.com/official-stockfish/Stockfish/blob/master/...

link