dear, dear a tiny down tick

[Dann Corbit]

http://162.217.248.187/

If we see another downward spiral, then I must echo the words from Monty Python's Holy Grail:
I fart in your general direction.
Not to the project, but to whoever injected the bad change that started another tail-spin.

[mar]

I'm a bit skeptical about the "progress", if you only measure deltas from the previous version, error will accumulate, so nobody knows how strong she really is. They don't seem to do regression tests either.
Kai's recent tests show no progress since 290 or so while the "elo graph" shows +130 or so..., I see no progress going from 242=>303 so far (in fact I see a regression but not enough games yet...)
Training a buggy engine can produce either results, illusion seems to work well in this case, we see what we want to see and we won't know
until someone actually tests Leela properly.
Don't get me wrong, I would love to see Leela improve, but wishes often differ from reality.

[Guenther]

It doesn't matter anyway, because a rollback is announced since a while already.

[CMCanavessi]

It doesn't matter anyway, because a rollback is announced since a while already.

I doubt the rollback will ever happen. Leela is recovering pretty fast and already at a similar level of the "best network ever" in her short but action-packed history.

[Milos]

It doesn't matter anyway, because a rollback is announced since a while already.

I doubt the rollback will ever happen. Leela is recovering pretty fast and already at a similar level of the "best network ever" in her short but action-packed history.

You are delusional beyond belief.
ID303 is like 50+ Elo weaker than ID232. Do proper testing, stop selling BS.

[mar]

You are delusional beyond belief.
ID303 is like 50+ Elo weaker than ID232. Do proper testing, stop selling BS.

I measured -60 actually compared to 24x, so how on earth can their "elo graph" show +130 when it's actually regressing.
But people love fake graphs obviously...

[CMCanavessi]

It doesn't matter anyway, because a rollback is announced since a while already.

I doubt the rollback will ever happen. Leela is recovering pretty fast and already at a similar level of the "best network ever" in her short but action-packed history.

You are delusional beyond belief.
ID303 is like 50+ Elo weaker than ID232. Do proper testing, stop selling BS.

Who the hell is "selling BS"? Here are my ratings:

Code: Select all

   # PLAYER                                  :  RATING  PLAYED    W    D    L   (%)  D(%)  OppAvg  OppN  OppDiv
  90 Leela Chess Zero v0.8 ID 262 x64        :  2712.6     200  100   42   58    61    21  2634.9    25    25.0
  92 Leela Chess Zero v0.8 ID 258 x64        :  2705.0     200   94   50   56    60    25  2634.9    25    25.0
  96 Leela Chess Zero v0.7 ID 232 x64        :  2699.3     200   94   47   59    59    24  2634.9    25    25.0
  97 Leela Chess Zero v0.7 ID 236 x64        :  2697.4     200   85   64   51    59    32  2634.9    25    25.0
  98 Leela Chess Zero v0.8 ID 241 x64        :  2689.9     200   87   56   57    58    28  2634.9    25    25.0
  99 Leela Chess Zero v0.8 ID 245 x64        :  2680.7     200   89   47   64    56    24  2634.9    25    25.0
 101 Leela Chess Zero v0.8 ID 252 x64        :  2669.6     200   87   45   68    55    23  2634.9    25    25.0
 102 Leela Chess Zero v0.7 ID 227 x64        :  2667.4     328  141   95   92    57    29  2600.6    64    48.1
 103 Leela Chess Zero v0.7 ID 195 x64        :  2665.9     200   86   45   69    54    23  2634.9    25    25.0
 107 Leela Chess Zero v0.7 ID 219 x64        :  2653.1     200   74   62   64    53    31  2634.9    25    25.0
 108 Leela Chess Zero v0.10 ID 300 x64       :  2647.6     200   79   49   72    52    25  2634.9    25    25.0
 115 Leela Chess Zero v0.7 ID 210 x64        :  2634.8     200   71   58   71    50    29  2634.9    25    25.0
 118 Leela Chess Zero v0.9 ID 270 x64        :  2631.2     200   75   48   77    50    24  2634.9    25    25.0
 122 Leela Chess Zero v0.10 ID 296 x64       :  2622.1     200   72   49   79    48    25  2634.9    25    25.0
 126 Leela Chess Zero v0.7 ID 189 x64        :  2603.8     200   61   61   78    46    31  2634.9    25    25.0
 130 Leela Chess Zero v0.7 ID 185 x64        :  2572.2     200   58   50   92    42    25  2634.9    25    25.0
 131 Leela Chess Zero v0.10 ID 280 x64       :  2570.4     200   60   45   95    41    23  2634.9    25    25.0
 138 Leela Chess Zero v0.7 ID 176 x64        :  2521.4     400  188   80  132    57    20  2454.9    50    50.0
 140 Leela Chess Zero v0.10 ID 288 x64       :  2516.0     200   49   39  112    34    20  2634.9    25    25.0

Leela went down to around 2500 from 2700, and is (at ID 300) around 2650. Currently testing ID 305 which is a bit stronger. How is that not recovering? Yes, it's still 50-60 elo behind the best ever tested (262) but in 12 networks it gained 150 elo... don't know wtf you're taking about "delusions".

[Milos]

I doubt the rollback will ever happen. Leela is recovering pretty fast and already at a similar level of the "best network ever" in her short but action-packed history.

You are delusional beyond belief.
ID303 is like 50+ Elo weaker than ID232. Do proper testing, stop selling BS.

Who the hell is "selling BS"? Here are my ratings:

Code: Select all

   # PLAYER                                  :  RATING  PLAYED    W    D    L   (%)  D(%)  OppAvg  OppN  OppDiv
  90 Leela Chess Zero v0.8 ID 262 x64        :  2712.6     200  100   42   58    61    21  2634.9    25    25.0
  92 Leela Chess Zero v0.8 ID 258 x64        :  2705.0     200   94   50   56    60    25  2634.9    25    25.0
  96 Leela Chess Zero v0.7 ID 232 x64        :  2699.3     200   94   47   59    59    24  2634.9    25    25.0
  97 Leela Chess Zero v0.7 ID 236 x64        :  2697.4     200   85   64   51    59    32  2634.9    25    25.0
  98 Leela Chess Zero v0.8 ID 241 x64        :  2689.9     200   87   56   57    58    28  2634.9    25    25.0
  99 Leela Chess Zero v0.8 ID 245 x64        :  2680.7     200   89   47   64    56    24  2634.9    25    25.0
 101 Leela Chess Zero v0.8 ID 252 x64        :  2669.6     200   87   45   68    55    23  2634.9    25    25.0
 102 Leela Chess Zero v0.7 ID 227 x64        :  2667.4     328  141   95   92    57    29  2600.6    64    48.1
 103 Leela Chess Zero v0.7 ID 195 x64        :  2665.9     200   86   45   69    54    23  2634.9    25    25.0
 107 Leela Chess Zero v0.7 ID 219 x64        :  2653.1     200   74   62   64    53    31  2634.9    25    25.0
 108 Leela Chess Zero v0.10 ID 300 x64       :  2647.6     200   79   49   72    52    25  2634.9    25    25.0
 115 Leela Chess Zero v0.7 ID 210 x64        :  2634.8     200   71   58   71    50    29  2634.9    25    25.0
 118 Leela Chess Zero v0.9 ID 270 x64        :  2631.2     200   75   48   77    50    24  2634.9    25    25.0
 122 Leela Chess Zero v0.10 ID 296 x64       :  2622.1     200   72   49   79    48    25  2634.9    25    25.0
 126 Leela Chess Zero v0.7 ID 189 x64        :  2603.8     200   61   61   78    46    31  2634.9    25    25.0
 130 Leela Chess Zero v0.7 ID 185 x64        :  2572.2     200   58   50   92    42    25  2634.9    25    25.0
 131 Leela Chess Zero v0.10 ID 280 x64       :  2570.4     200   60   45   95    41    23  2634.9    25    25.0
 138 Leela Chess Zero v0.7 ID 176 x64        :  2521.4     400  188   80  132    57    20  2454.9    50    50.0
 140 Leela Chess Zero v0.10 ID 288 x64       :  2516.0     200   49   39  112    34    20  2634.9    25    25.0

Leela went down to around 2500 from 2700, and is (at ID 300) around 2650. Currently testing ID 305 which is a bit stronger. How is that not recovering? Yes, it's still 50-60 elo behind the best ever tested (262) but in 12 networks it gained 150 elo... don't know wtf you're taking about "delusions".

Your "at a similar level of the "best network ever"" is equal to 60Elo difference???
You play 200 games and you assume your "ranking" is accurate? Gee
You seems to be completely unaware that with gauntlet play error margin in your case is at least +/-60Elo...
What 150Elo are you talking about, in which world?
In your logic it also lost 200Elo in 16 networks which is even more ludicrous.

[CMCanavessi]

You are delusional beyond belief.
ID303 is like 50+ Elo weaker than ID232. Do proper testing, stop selling BS.

Who the hell is "selling BS"? Here are my ratings:

Code: Select all

   # PLAYER                                  :  RATING  PLAYED    W    D    L   (%)  D(%)  OppAvg  OppN  OppDiv
  90 Leela Chess Zero v0.8 ID 262 x64        :  2712.6     200  100   42   58    61    21  2634.9    25    25.0
  92 Leela Chess Zero v0.8 ID 258 x64        :  2705.0     200   94   50   56    60    25  2634.9    25    25.0
  96 Leela Chess Zero v0.7 ID 232 x64        :  2699.3     200   94   47   59    59    24  2634.9    25    25.0
  97 Leela Chess Zero v0.7 ID 236 x64        :  2697.4     200   85   64   51    59    32  2634.9    25    25.0
  98 Leela Chess Zero v0.8 ID 241 x64        :  2689.9     200   87   56   57    58    28  2634.9    25    25.0
  99 Leela Chess Zero v0.8 ID 245 x64        :  2680.7     200   89   47   64    56    24  2634.9    25    25.0
 101 Leela Chess Zero v0.8 ID 252 x64        :  2669.6     200   87   45   68    55    23  2634.9    25    25.0
 102 Leela Chess Zero v0.7 ID 227 x64        :  2667.4     328  141   95   92    57    29  2600.6    64    48.1
 103 Leela Chess Zero v0.7 ID 195 x64        :  2665.9     200   86   45   69    54    23  2634.9    25    25.0
 107 Leela Chess Zero v0.7 ID 219 x64        :  2653.1     200   74   62   64    53    31  2634.9    25    25.0
 108 Leela Chess Zero v0.10 ID 300 x64       :  2647.6     200   79   49   72    52    25  2634.9    25    25.0
 115 Leela Chess Zero v0.7 ID 210 x64        :  2634.8     200   71   58   71    50    29  2634.9    25    25.0
 118 Leela Chess Zero v0.9 ID 270 x64        :  2631.2     200   75   48   77    50    24  2634.9    25    25.0
 122 Leela Chess Zero v0.10 ID 296 x64       :  2622.1     200   72   49   79    48    25  2634.9    25    25.0
 126 Leela Chess Zero v0.7 ID 189 x64        :  2603.8     200   61   61   78    46    31  2634.9    25    25.0
 130 Leela Chess Zero v0.7 ID 185 x64        :  2572.2     200   58   50   92    42    25  2634.9    25    25.0
 131 Leela Chess Zero v0.10 ID 280 x64       :  2570.4     200   60   45   95    41    23  2634.9    25    25.0
 138 Leela Chess Zero v0.7 ID 176 x64        :  2521.4     400  188   80  132    57    20  2454.9    50    50.0
 140 Leela Chess Zero v0.10 ID 288 x64       :  2516.0     200   49   39  112    34    20  2634.9    25    25.0

Leela went down to around 2500 from 2700, and is (at ID 300) around 2650. Currently testing ID 305 which is a bit stronger. How is that not recovering? Yes, it's still 50-60 elo behind the best ever tested (262) but in 12 networks it gained 150 elo... don't know wtf you're taking about "delusions".

Your "at a similar level of the "best network ever"" is equal to 60Elo difference???
You play 200 games and you assume your "ranking" is accurate? Gee
You seems to be completely unaware that with gauntlet play error margin in your case is at least +/-60Elo...
What 150Elo are you talking about, in which world?
In your logic it also lost 200Elo in 16 networks which is even more ludicrous.

I know my error bars are quite big. Still they are better than yours, which are non-existant.

[shrapnel]

I know my error bars are quite big. Still they are better than yours, which are non-existent.

Yours is big, his is non-existent.