Not only more interesting, it Imo might be to further improve (*) engine chess (thus possibly also at the highestEngines deliberately aimed to achieve good EAS scores (Chess System Tal EAS, Rebel EAS and Patricia etc) and a few others, are more interesting
levels, i.e. SF and two or three others). Note and the ratings of CST-2 and Rebel16.3 are already quite good.
While CST2, Rebel 16.3 (and certainly P3) have lower rating than SF, it wouldn't surprise me if SF could get
some further rating increase with similar methods (networks trained on lots of sharp positions eg. positional
sacrifices as from the Pohl tool etc). Idea: play lots of gambit games (eg. with Cutechess) eg. with my latest
gambit.bin (see below(***)) and subsequently extract the interesting positions with the Pohl tool. Now i'm not
an engine programmer nor a Nnue expert so in practice at highest levels it may be not so easy; key might be
to extract positions with sacrifices attacks etc with a favorable outcome (rather than just 'wild'/speculative
positions, whether the Pohl tool can be tuned for such a purpose, i don't know.
PS1 my latest gambiteer package (with the gambit.bin book included):
https://sourceforge.net/projects/chess-gambiteer/
Note these gambitlines, especially when going deeper eg. move 4/5/6 etc are not cast in iron, with some gambits
(eg. Smith Morra), using P3, i found some interesting new sacrifices (eg. with Nd5!? with still a play pawn on e6) and my
work on this gambitbook isn't finished yet, it's still an ongoing project. As for a separate version of P3 eg. P3-g playing
gambits by itself (without book), well that might be possible but often the compensation after a gambit goes
beyond the engine search, and thus I suspect that such a P3g (similar to P2) would have lower playing strenght (elo)
(*) PS2 some more thoughts:When I think of historical players in the past who played '(ultra) agressively', Nezhmetdinov and Tal come to mind;
but their sacrifices were not always correct. The players Morphy and Kasparov also played 'agressively' but usually more accurately,
less speculative. P3 is even more accurate, and -being able to quickly detect positional -or attacking- sacrifices,
imo has a good style, and already playing strong chess certainly compared with humans (although not at rating level of SF, ofcourse,
for other reasons maybe search related or otherwise). So the idea of using lots of positions with sacrifices etc for NN training (with the Pohl
tools) instead of zillions of drawish positions seems to pay off; while i first regarded P2 as just a curiosity, the new
version (P3) looks promising indeed. Wether this method (of going for 'aggression' will lead to a fundamentally stronger
engine remains to be seen; while it will be difficult to surpass SF in Elo, with similar methods I wouldn't be surprised
if the SF team might be able to add quite some Elo points again; my 2 cnts again. Maybe i should bring it up
in the Discord engine prog channel but like i said i'm not an engine -programming- expert.
(***) there may be other similar methods as well, I just gave an example (which might work, or not)
Ofcourse there's a limit for the SF rating, with chess being fundamentally a draw, so on these CCRL
lists there will be a limit for the top ratings, not exceeding eg. 4000 (or 3900), I predict.
