lichess.org
Donate

Upgrade of artificial intelligence

Interestingly, the analysis does recognize the blunders. Like Toadofsky said, it looks like 500ms is not enough to avoid them at the moment. I hope they can make it work!

It was late when I made the announcement last night, so here is an addendum:

- I just bought 2 powerful servers, adding 32 CPUs to the A.I. cluster. We now have 82 instances of stockfish crunching moves and analysis. Is it the largest chess A.I. cluster ever? I don't know.

- Big thanks to Toadofsky for coordonating the developments on Stockfish lichess edition (enhanced with KotH and Threecheck), and to all coders involved.
Part of the problem could simply be that 3-check has a big horizon effect, which is amplified by the drastic pruning that Stockfish does normally. Whatever the problem is, it should not play 1. e4 e6 2. d4?? at level 8.

The most CPUs on a chess AI would probably be Lewis Stiller on the Connection Machine with 65536, though this was 1995 and they were deliberately very weak (a vector computer). Jonny currently runs on 800 or 2400 cores when it competes in the WCCC. Zappa ran on 512 cores one year, Deep Blue was 480, and Rybka cluster was 296 cores the last time public information was given. GPS Shogi ran on 3224 cores in 2012 when it won the Shogi championship. (Academics can borrow *huge* hardware.)

I have heard that some correspondence champions use thousands of cores on a regular basis. I don't know anyone who has tried to get chess to run sensibly on a GPU.
#15 I'm guessing that Stockfish is better at offense than at defense, regardless of level. Still, I expect that usually level 1 makes mistakes.
Yeah I found the new level 8 stockfish to sometimes play opening book moves that aren't good in three checks like the scandinavian
thanks for the update! I was hoping we could get computer analysis of three check games.

This topic has been archived and can no longer be replied to.