Draws with Knight odds are possible against a top GM?

[lkaufman]

I found this blitz odds match between Leela and Naroditsky 2610 elo https://lichess.org/blog/XBsCBBMAACUA3C ... chess-zero

Leela - Naroditsky f1 bishop missing 3+1 1-0
Naroditsky - Leela f8 bishop missing 3+1 draw
Leela - Naroditsky b1 knight missing 3+1 1-0
Naroditsky - Leela g8 knight missing 3+1 0-1
Naroditsky - Leela g8 knight missing 5+2 draw

Overall 4-1 win for Leela

If you want to see all the games https://lichess.org/@/DrDragonitsky/all

Was there any information about either the GPU or the network Leela used? Anyway, I guess I can't expect to score in blitz at knight odds if Naroditsky loses badly even with knight and move or bishop odds!

[Laskos]

I found this blitz odds match between Leela and Naroditsky 2610 elo https://lichess.org/blog/XBsCBBMAACUA3C ... chess-zero

Leela - Naroditsky f1 bishop missing 3+1 1-0
Naroditsky - Leela f8 bishop missing 3+1 draw
Leela - Naroditsky b1 knight missing 3+1 1-0
Naroditsky - Leela g8 knight missing 3+1 0-1
Naroditsky - Leela g8 knight missing 5+2 draw

Overall 4-1 win for Leela

If you want to see all the games https://lichess.org/@/DrDragonitsky/all

Was there any information about either the GPU or the network Leela used? Anyway, I guess I can't expect to score in blitz at knight odds if Naroditsky loses badly even with knight and move or bishop odds!

Pretty amazing. Googling, it seems this Lichess Leela runs on 2080 + 1080Ti, so not much faster or better than your 2080 (don't even know how they combine fp16 with fp32, 1080Ti may run very slowly and actually harm).
Don't know the net ID number
Amazing results.

[EroSennin]

In the game chat Leela says the ID is the same Kaufman used 11248.

[lkaufman]

In the game chat Leela says the ID is the same Kaufman used 11248.

Thanks. I don't suppose Leela mentioned any other UCI settings? I used the ones recommended for 2080 in a talkchess post, plus a small temp setting for the first three moves. I think they must have used a too-high temp setting, since in some games Lc0 opened with 1...h6? Even this extra handicap didn't hurt. By the way Kai, it seems that a temp of 1.5 that you used is too high for knight odds games, it will play stupid moves too often in the first three moves. Even one is too high, 0.5 is fine though.
I ran a new test last night. Lc0 11248 (2080 gpu) vs Lc011248 (CPU only) knight odds! Time limit 4' + 2" (I tried 3' +1" first but CPU version was losing on time, I guess it needs 2 seconds per move min.). The result was that Black (CPU version) won by 8.5 to 1.5. Even the draws were miracle saves in lost or bad endings. The speed ratio was around 200 to 1, which should be enough to make this a reasonable match. I had the same experience with Komodo; even if it can give X handicap to an unrelated engine, it loses badly giving the same handicap to an older or crippled version of itself with a similar rating to the unrelated engine.
I also noticed that you can apparently get Lc0 to play from policy network only simply by setting depth to 1. Even on this level, with no search, it plays quite good chess, good enough to beat me in blitz in normal chess. Quite remarkable. It can find the right moves in fairly simple tactical exercises that I give to students, as long as they don't involve a large initial sacrifice.

[EroSennin]

In the game chat Leela says the ID is the same Kaufman used 11248.

Thanks. I don't suppose Leela mentioned any other UCI settings?

I don't know. You can ask him on Lichess.