News about Leela Odds Bots
Moderator: Ras
-
lkaufman
- Posts: 6316
- Joined: Sun Jan 10, 2010 6:15 am
- Location: Maryland USA
- Full name: Larry Kaufman
News about Leela Odds Bots
Several major new developments regarding odds chess and the Leela Odds bots. First, an article by me about odds chess should appear soon on chess.com, which should spur interest in the topic (and in the bots). Second, two major bugs were found and fixed by ChatGPT Codex, which will avoid time-outs after individual engine crashes, and will avoid searching much deeper than intended at handicaps beyond queen odds, which should also speed up play a bit at all odds due to less demand on the single GPU. Third, although I have no training in AI or neural nets, I am now able to train nets by using ChatGPT Codex, and already in the first week have deployed much stronger bots for two specific handicaps, QueenForKnight, and TwoBishops (for both colors), which were using a rook odds net until now. Gain is in the ballpark of a hundred elo. Whether Codex will be able to improve on existing nets for the three main handicaps (N,R,Q) remains to be seen, but I am optimistic. Exciting times for the bots!
Komodo rules!
-
nmcrazyim5
- Posts: 42
- Joined: Sat Jan 28, 2023 12:03 am
- Full name: P
Re: News about Leela Odds Bots
This is fantastic! Thanks LK. Are the 100 elo gainers on all the other odds bots besides the 3 main odds bots on lichess?
Also, how does the training differ in imbalanced odds (on both sides, like Q for 2N)?
Also, how does the training differ in imbalanced odds (on both sides, like Q for 2N)?
-
lkaufman
- Posts: 6316
- Joined: Sun Jan 10, 2010 6:15 am
- Location: Maryland USA
- Full name: Larry Kaufman
Re: News about Leela Odds Bots
The chess.com article was just published; see "articles" on chess.com. The 100 elo gain is so far just on queen for knight odds and two bishops odds, those are the only ones trained so far. I expect similar gains for most of the LeelaPieceOdds bots, because except for two knights they didn't have their own nets until now. For the single piece handicaps, gains will be smaller as they already have good nets. We don't have a Q for 2N bot, that would be too small a handicap for any human (except maybe Magnus playing Classical TL). For Queen for knight odds, training is no different than for any other handicap.nmcrazyim5 wrote: ↑Thu Oct 29, 2026 5:42 am This is fantastic! Thanks LK. Are the 100 elo gainers on all the other odds bots besides the 3 main odds bots on lichess?
Also, how does the training differ in imbalanced odds (on both sides, like Q for 2N)?
Komodo rules!
-
lkaufman
- Posts: 6316
- Joined: Sun Jan 10, 2010 6:15 am
- Location: Maryland USA
- Full name: Larry Kaufman
Re: News about Leela Odds Bots
Now all of the multipiece handicaps (except two knights, which already had its own net) have nets trained by AI, with elo gains ranging from 100 to 200 over the nets previously used by the bots. Rook odds looks likely to get a significant upgrade from this AI training in a couple days. Still to go: knight odds, two knights odds, and queen odds. Some of the odds have even been promoted twice, first by AI training in the conventional way, and then again by training using "visit counts". Exciting times for this project.nmcrazyim5 wrote: ↑Thu Oct 29, 2026 5:42 am This is fantastic! Thanks LK. Are the 100 elo gainers on all the other odds bots besides the 3 main odds bots on lichess?
Also, how does the training differ in imbalanced odds (on both sides, like Q for 2N)?
Komodo rules!
-
nmcrazyim5
- Posts: 42
- Joined: Sat Jan 28, 2023 12:03 am
- Full name: P
Re: News about Leela Odds Bots
This is awesome! Looking forward to playing these bots.lkaufman wrote: ↑Wed Nov 11, 2026 4:47 amNow all of the multipiece handicaps (except two knights, which already had its own net) have nets trained by AI, with elo gains ranging from 100 to 200 over the nets previously used by the bots. Rook odds looks likely to get a significant upgrade from this AI training in a couple days. Still to go: knight odds, two knights odds, and queen odds. Some of the odds have even been promoted twice, first by AI training in the conventional way, and then again by training using "visit counts". Exciting times for this project.nmcrazyim5 wrote: ↑Thu Oct 29, 2026 5:42 am This is fantastic! Thanks LK. Are the 100 elo gainers on all the other odds bots besides the 3 main odds bots on lichess?
Also, how does the training differ in imbalanced odds (on both sides, like Q for 2N)?
-
lkaufman
- Posts: 6316
- Joined: Sun Jan 10, 2010 6:15 am
- Location: Maryland USA
- Full name: Larry Kaufman
Re: News about Leela Odds Bots
Rook odds has now been promoted with a new net, with an estimated gain of 100 elo! Knight odds training is underway, expected to be ready for the bot by Tuesday. Most of the multipiece odds have now been promoted twice. Second part of my article on odds chess went online yesterday at chess.com. Exciting times for the Leela Odds bots!
Komodo rules!
-
Father
- Posts: 2060
- Joined: Sun Mar 19, 2006 4:39 am
- Location: Colombia
- Full name: Pablo Ignacio Restrepo
Re: News about Leela Odds Bots
"Thank you, Dr. Larry Kaufman. Your analysis, studies, and performance tests on the odds bots are truly fascinating. Your question, 'How Strong Are The Leela Odds Bots?', opens up a whole universe of ideas. This work is chess and so much more!! It is a deeply meaningful contribution to humanity."lkaufman wrote: ↑Thu Oct 29, 2026 5:28 am Several major new developments regarding odds chess and the Leela Odds bots. First, an article by me about odds chess should appear soon on chess.com, which should spur interest in the topic (and in the bots). Second, two major bugs were found and fixed by ChatGPT Codex, which will avoid time-outs after individual engine crashes, and will avoid searching much deeper than intended at handicaps beyond queen odds, which should also speed up play a bit at all odds due to less demand on the single GPU. Third, although I have no training in AI or neural nets, I am now able to train nets by using ChatGPT Codex, and already in the first week have deployed much stronger bots for two specific handicaps, QueenForKnight, and TwoBishops (for both colors), which were using a rook odds net until now. Gain is in the ballpark of a hundred elo. Whether Codex will be able to improve on existing nets for the three main handicaps (N,R,Q) remains to be seen, but I am optimistic. Exciting times for the bots!
I am thinking chess is in a coin.Human beings for ever playing in one face.Now I am playing in the other face:"Antichess". Computers are as a fortres where owner forgot to close a little door behind. You must enter across this door.Forget the front.
-
nmcrazyim5
- Posts: 42
- Joined: Sat Jan 28, 2023 12:03 am
- Full name: P
-
lkaufman
- Posts: 6316
- Joined: Sun Jan 10, 2010 6:15 am
- Location: Maryland USA
- Full name: Larry Kaufman
Re: News about Leela Odds Bots
Yes, single rook. Specifically QR, the only rook odds used historically. Anyone wanting to play KR odds can do so on the LeelaPieceOdds bot, though it hasn't yet been updated with this new net. Testing on the new net for knight odds has just completed which also showed about a 100 elo gain. Furthermore we are now testing an improvement that should help both of these nets a bit more. I expect we'll also have a new net for queen odds soon, only issue is hardware availability for game gen and training and testing.
Komodo rules!
-
Nordlandia
- Posts: 2850
- Joined: Fri Sep 25, 2015 9:38 pm
- Location: Sortland, Norway
Re: News about Leela Odds Bots
Removing the rook as a handicap should not necessarily deprive the odds-receiving player of the right to castle. The king could therefore make the ordinary castling move, even though there was no physical rook to participate.
Would it be technically feasible to implement phantom castling with relative ease and minimal overhaul?.
Would it be technically feasible to implement phantom castling with relative ease and minimal overhaul?.