Hi all. I'm writing to update you on Leela Chess Zero, an open source re-creation of the AlphaZero chess engine.
We are currently at about IM strength or little better. You can track progress here:
If you would like to help train Leela, there are two easy ways you can do this for free. We are an open source, community project and rely on volunteers to provide computing power to train the neural network behind Leela. Luckily, you don't need to give up your computer to do this.
You will need to copy/paste the following URLs because it seems the forum interface doesn't like them.
You can track the number of games you've contributed per day using your username on the main page.
I don't think we are yet at IM level, i played today 2 games against stockfish and it's still does positional errors especially in the endgame allowing stockfish to easily develops passed pawns. It is probably at FM level or less.
I Like the project and i'm contributing with my GPU (about 1200 games per day)
It's way way way better than IM level, I can tell you that for sure.
It is not better than IM level at winning rook and pawn versus rook endgames, but in general, it certainly is.
Thanks for the comments. Indeed we seem to have weaknesses in the endgame, and sometimes even with simple tactics. This should be eliminated with more training.
Again, anyone who is interested should please refer to the links in the original post.
Man can still beat this computer. I played white, but missed in the winning position checkmate in 2. :P
This is great. Thanks for the links @BigGreenShrek
My desktop is running 24/7 with a decent GPU, but additional resources are great. I also get a kick out of Google-mooching given their lack of transparency with respect to A0.
@dudeski_robinson Someone on the Discord is working on a Colab script that will do engine tournaments to test various versions, too. I'll stick that here when it's tested and done.
@BigGreenShrek Would be great if we could get real-time updates on those results on some web page.
I'm just not sure how to interpret that weird ELO curve they have as a front page to the website.
@dudeski_robinson The script is now on the "Getting Started" page.
However, there's a bigger problem: everyone will need to update their clients soon to fix a bug where black could not promote pawns. (It is interesting we made it to ~2600 Elo despite this...) And the script will need to be changed to account for this.
Regarding the Elo curve on the front page, they started with a net that randomly chose moves, then trained for some time, and recorded the Elo difference between the two nets. Then a third net was trained, and the Elo difference between it and the second net was recorded on the graph. And so on. The rating differences are between successive IDs.
Currently 125 is the strongest in actual play, but it also contains the promotion bug.