Does Leela Odds bots know to play in a different way in different time controls?

[lkaufman]

The question is if Leela Knows to play in a different way at different time controls even if the expected result is the same because of different opponents.

My guess is that if we take humans who get 50% against LeelaQueenOdds at 1+1
and humans who get 50% against LeelaQueenOdds at 10+10 then you will find that they do different types of mistakes and I guess that you are going to see more one ply blunders at 1+1 and more positinal mistakes at 10+10(did not do a research to test if I am correct).

In theory we could have multiple nets based on time control, with each net trained against a bot that was itself trained to simulate humans of a specified level at a specified time control, but the benefit would not justify the massive amount of both human and machine time needed. We are continuing to make progress in multiple ways. The White Queen odds was updated a few days ago to a new net that performs slightly better against bots at least, with a new net for Black likely in a week or so. The rook odds and knight odds were transferred from a 4090 based machine to a 5090 based machine which allowed for a 50% increase in max nodes with a similar average time per move; it seems to be performing even better with this change. The two knights handicap was switched to the new queen odds net, which somehow tested as more of an improvement for two knights than for queen (!), so I particularly recommend two knights odds (using "LeelaPieceOdds") for very strong players who find queen odds a bit too easy (at their preferred time control) but who hardly ever win at rook odds. Perhaps soon two knights can be a fair match in 3'2" blitz for even rather strong grandmasters. That is a sentence I would never have imagined myself ever writing a year ago! Interestingly, the low-rated GM who set off the boom in interest in LeelaQueenOdds when he lost his first five blitz (3'2") at queen odds, "Mikulas", eventually pulled ahead after hundreds of games, and made a nearly even score after that at 2'2", and is now number 9 on the Leaderboard!

Thank you, Mr. Larry Kaufman, for your teachings. I've been analyzing the new humans who are steadily rising on the LeelaQueenOdds leaderboard, and I'm placing special emphasis on one man I find "super gifted." This is GM "FeegLood." Every day, belonging to or appearing on the leaderboard of the top 100 humans becomes more difficult and meritorious. I confess that I would be very sad if my ideal horse, "Catecan," and my own self were to be forgotten. I ask God for the strength to endure on the leaderboard. I would like to be on the same level as the giants Nakamura and Magnus; then I would be ready to travel to heaven on my horse, Catecan. But truth be told, this leaderboard describes another form of greatness, a giant in itself.

GM FeegLood has been identified as Jeffery Xiong, a young American star who has been over 2700 FIDE. He has nearly a 30% score in one minute no inc bullet at queen odds, Leela white, among the best 3 or 4 records, and near-even results in blitz at rook odds with alternating colors, perhaps the best record at rook odds.

Good evening, Mr. Larry Kaufman.
The GM and leader of the top 100 humans facing the robot LeelaQueenOdds has crossed the 3000-rating line inherent in the leaderboard. I thought 19th place belonged to GM Maghsoodloo, Parham, currently 38th in the world. The battles to advance upwards in the leaderboard are becoming very interesting and dynamic. Today I wonder if another giant will come to challenge Arka50!! Catecan and I are resting in an old inn we found on the road. We will continue walking with a sure and steady pace like an elephant.

LeelaQueenOdds Leaderboard

Next update in: 5m 06s
# Player Rating Games Last Game TC
1 Arka50 3035 150 2025-06-06 1+0
2 SuperEasy 2940 407 2025-06-06 8+3
3 verystablegenius69 2862 384 2025-05-07 1+0
4 wateenellende 2814 154 2025-03-26 1+0
5 VladimirMarkovic 2808 1403 2025-06-06 1+0
6 InvisibleGuest2023 2795 347 2025-06-06 1+0
7 Former_Player 2778 308 2025-06-01 3+2
8 JoanFluvia 2774 631 2025-06-04 2+1
9 Mikulas 2760 596 2025-06-06 2+1
10 edgelorde 2745 1644 2025-06-06 1+1
11 Hissha 2723 135 2025-05-27 1+10
12 Catecan 2706 6502 2025-06-06 1+0
13 jflup 2705 175 2025-06-04 3+1
14 Orichess88 2694 88 2025-04-03 2+3
15 Alex_Muezersky 2673 179 2025-04-05 1+1
16 dauANHbac 2669 74 2025-04-27 1+1
17 smeets 2662 29 2025-04-02 2+1
18 Phara0h1 2662 995 2025-05-28 4+0
19 FeegLood 2626 108 2025-06-04 1+0
20 Paulewski 2574 340 2025-06-06

Why do you think feeglood is Parham? I think there was convincing evidence posted that it is Xiong. Yesterday he made a perfect 11/11 score in freestyle friday, the first ever. Feeglood has been playing both LQO and LRO a lot lately, so perhaps that is why he achieved this result, assuming Feeglood = Xiong.

[lkaufman]

I don’t think there is much difference between classical games by 2600s and rapid games by 2800s, for example

Do you have data on this? I would have suspected that simple one-move blunders are much more common in the latter games.
(I'm not saying that Leela should take this into account--that might be too complex to implement to be worth it.)

I only have data that shows that "Accuracy" per chess.com does not differ much in that scenario, it doesn't give insight into the type of errors made. You are surely correct if we include games like 10 min 0 sec as Rapid, but I had in mind the standard FIDE rapid time control of 15' + 10", at which 2800 players would rarely make obvious gross one-move blunders, though it does happen.