ChessBase's prophecy of 2004 fulfilled in 2019

[Laskos]

ChessBase posted in 2004 on its site the following announcement about the new engine called "Kimo" --- "The Knowledge Based Chess Program":

https://en.chessbase.com/post/kimo-an-i ... ach-to-che

Kimo.jpg

"Kimo" was kept secret for 15 years, until it surfaced to the public as Leela Chess Zero in 2019.
All the features described in 2004 are exactly as described, on 2004 hardware. This will be shown.
Kimo T40 on 3.0GHz Pentium core of 2004 achieves about 10 nodes/second. On a good GPU of 2004 (like NVIDIA GeForce 6800 GT) it achieves about 50 nodes/second.

The top engines of 2004 were Shredder 8, Junior 9 and Fritz 8, about 100 Elo points above Fruit 2.1 (on one core, most PCs were single core).

I simulated blitz games at 5 min + 3 seconds, node for node, for 2004 hardware.


CPU Kimo scored against Fruit 2.1 as:

Score of Kimo_42580 vs Fruit_21: 10 - 7 - 3 [0.575] 20
Elo difference: 52.51 +/- 150.97
Finished match

Which shows Kimo as being very competitive against top engines of 2004 on a 2004 CPU.


GPU Kimo scored against Fruit 2.1 as:

Score of Kimo_42580 vs Fruit_21: 13 - 3 - 4 [0.750] 20
Elo difference: 190.85 +/- 172.61
Finished match

Which shows Kimo as being above all top engines of 2004 on a 2004 GPU.

To longer time controls, the results will be even better against top engines of 2004 on 2004 hardware.

Other descriptions in the ChessBase prophecy are accurate too, proving that "Kimo" existed back then in 2004, was a top engine of its time, but was released only in 2019.

[Laskos]

The idea to check the ChessBase description of Kimo came to me after I performed a test with different versions of Kimo at 1 node per move (no search), and Texel 1.07 at 4000 nodes per move, which I estimate to be about 2100 +/- 100 FIDE Elo level at 45 min + 15 seconds time control. These show the latest Kimo as very strong even without the search, about a strong FM human level.

Code: Select all

Rank Name                           Elo     +/-   Games   Score   Draws
   1 Kimo 42580                      49      30     400   57.0%   24.0%
   2 Kimo 11248                      18      30     400   52.6%   23.8%
   3 Texel Elo 2100                   3      32     400   50.5%   14.5%
   4 Kimo 40b 119                    -8      29     400   48.9%   26.3%
   5 Kimo 32930                     -63      29     400   41.0%   30.5%
Finished match

40b is the 40 blocks Kimo (surprisingly, it didn't perform that well).

[jp]

Code: Select all

Rank Name                           Elo     +/-   Games   Score   Draws
   1 Kimo 42580                      49      30     400   57.0%   24.0%
   2 Kimo 11248                      18      30     400   52.6%   23.8%
   3 Texel Elo 2100                   3      32     400   50.5%   14.5%

So with no search, NN 42580 is only about 30 elo stronger than 11248. From what I've seen, I'd be surprised if old NNs are much above 2000 elo on a human longer TC scale. (This has been discussed in another thread, but I wasn't sure if newer nets were much better than older ones. It seems they aren't.)

[Laskos]

Code: Select all

Rank Name                           Elo     +/-   Games   Score   Draws
   1 Kimo 42580                      49      30     400   57.0%   24.0%
   2 Kimo 11248                      18      30     400   52.6%   23.8%
   3 Texel Elo 2100                   3      32     400   50.5%   14.5%

So with no search, NN 42580 is only about 30 elo stronger than 11248. From what I've seen, I'd be surprised if old NNs are much above 2000 elo on a human longer TC scale. (This has been discussed in another thread, but I wasn't sure if newer nets were much better than older ones. It seems they aren't.)

Yes, 2000 FIDE Elo for NN11248 might be reasonable, my estimation for Texel is 2100 +/- 100. I vaguely knew that 1xxxx nets were stronger than 3xxxx nets and early 4xxxx nets with a single node (no search), but I didn't know that late 4xxxx nets became the strongest at 1 node too. Anyway, it seemed strong enough that I figured out that using just several dozen or a hundred of nodes per move, 43580 might be stronger than Fruit 2.1 at blitz, a thing which turned out to be true. The thing is that we could have had this NN engine back in 2004, searching 300 nodes per move on an 2004 GPU at blitz TC (having about 50 nodes per second speed), and beating the crap out of Shredders, Fritzes and Juniors of those times. I remember the times and I was full of respect for those top engines.

[jorose]

Seems like a fun experiment. I do sometimes wonder how people would react if we teleported modern engines back in time. Potentially even further back, though before the 2000s you might have to send hardware back as well. Especially since people of the passed would have no idea how to actually play against engines. Imagine Tal trying one of his dubious sacrifices against Stockfish...

Several parts of Leela had not been invented yet, so even assuming we could have trained Leela there were still plenty of reasons we couldn't have her back then.

[Milos]

Yes, 2000 FIDE Elo for NN11248 might be reasonable, my estimation for Texel is 2100 +/- 100. I vaguely knew that 1xxxx nets were stronger than 3xxxx nets and early 4xxxx nets with a single node (no search), but I didn't know that late 4xxxx nets became the strongest at 1 node too. Anyway, it seemed strong enough that I figured out that using just several dozen or a hundred of nodes per move, 43580 might be stronger than Fruit 2.1 at blitz, a thing which turned out to be true. The thing is that we could have had this NN engine back in 2004, searching 300 nodes per move on an 2004 GPU at blitz TC (having about 50 nodes per second speed), and beating the crap out of Shredders, Fritzes and Juniors of those times. I remember the times and I was full of respect for those top engines.

That simply cannot be true. I remember playing against NN11xxx with no search (single node) and being able to comfortably draw and even beat it after couple of games onwards. And OTB I'm probably around 1600 Elo.
When running NN without search it simply has way too many tactical holes and it's too weak in endgame.

[Laskos]

Yes, 2000 FIDE Elo for NN11248 might be reasonable, my estimation for Texel is 2100 +/- 100. I vaguely knew that 1xxxx nets were stronger than 3xxxx nets and early 4xxxx nets with a single node (no search), but I didn't know that late 4xxxx nets became the strongest at 1 node too. Anyway, it seemed strong enough that I figured out that using just several dozen or a hundred of nodes per move, 43580 might be stronger than Fruit 2.1 at blitz, a thing which turned out to be true. The thing is that we could have had this NN engine back in 2004, searching 300 nodes per move on an 2004 GPU at blitz TC (having about 50 nodes per second speed), and beating the crap out of Shredders, Fritzes and Juniors of those times. I remember the times and I was full of respect for those top engines.

That simply cannot be true. I remember playing against NN11xxx with no search (single node) and being able to comfortably draw and even beat it after couple of games onwards. And OTB I'm probably around 1600 Elo.
When running NN without search it simply has way too many tactical holes and it's too weak in endgame.

I myself was surprised, but Texel 1.07 at 4000 nodes/ move I speculated pretty firmly to be in 2000-2100 FIDE Elo range at 45min + 15s. Can you beat or draw consistently Texel at 4000 nodes/move at this TC? If not, then the result might be an artifact due to Texel's inability to capitalize on tactics and endgame. Also, 42580 with a bit less than 100 nodes/move beat Fruit 2.1 at blitz. Fruit at blitz is about 2800 FIDE blitz level, a very tough nut. Again, Fruit level might not translate into human level against Lc0.

[Milos]

I myself was surprised, but Texel 1.07 at 4000 nodes/ move I speculated pretty firmly to be in 2000-2100 FIDE Elo range at 45min + 15s. Can you beat or draw consistently Texel at 4000 nodes/move at this TC? If not, then the result might be an artifact due to Texel's inability to capitalize on tactics and endgame. Also, 42580 with a bit less than 100 nodes/move beat Fruit 2.1 at blitz. Fruit at blitz is about 2800 FIDE blitz level, a very tough nut. Again, Fruit level might not translate into human level against Lc0.

I haven't tried Texel, but what I tried is SF with QS only i.e. depth=1 and I know I have struggled most of the time to even hold a draw (SF with QS only would punish each of my tactical blunders ruthlessly). And Lc0 with 11xxx net and no search was slightly better than SF with only QS.

[corres]

Please, see:
2004 APRIL FIRST!
It was a joke only...
(and nothing prophecy)

[supersharp77]

Please, see:
2004 APRIL FIRST!
It was a joke only...
(and nothing prophecy)

I was thinking.....'Kimo Chess Program..how did we all miss that one on Chessbase?' LMAO!!