Comments/Ratings for a Single Item

I had Fairy-Stockfish play against itself with this, and White won. So perhaps it's not as good of an example as imagined. But I think that may have been influenced by Black not being able to Promote to Queen.

I don't know if Fairy-Stockfish is any good at this? Does it include this as an officially supported, highly optimized variant, or is it something you configured yourself? I know that the ordinary Stockfish many years ago was terrible at this. I let it play against QueeNy (a derivative of the 2400 Elo engine Spartacus), and it got totally clobbered. Engines that do not know 2 Knights are far more valuable than a Queen will play like an idiot. You cannot expect engines that blunder away their Knights to have any success in playing with black. Can Fairy-Stockfish beat QueeNy?

I’ve recently seen a YouTube video with game between Levy Rozman (aka GothamChess) without his Knights VS Hikaru Nakamura with 7 Knights. Hikaru won, but it’s partially because he played by White and Levy blundered his Bishop on 4th move.

Maybe Bishops’ pair & 2 Rooks are better than 2 Queens.

Also note that the black King should play an active role; it should not be afraid to move to the centre, together with the pack of Knights. If the engine would use the standard king-safety evaluation, which strongly encourages it to cower away in a corner, and then keep a number of Knights there, and the best black can hope for is a draw. Instead black should go for the enemy King, and force the Queens to sacrifice themselves for a single Knight in order to pervent being checkmated.

I configured it myself. It is possible to train an NNUE for a custom variant for Fairy-Stockfish, so if someone were to do that, we could probably have much better version that should understand the value of the pieces and king safety properly. Unfortunately, I do not have the hardware to train it right now.

OK, that is what I thought. But for one, a single game is hardly statistically significant, and even in 10 games you can only draw a (weak) conclusion from a result like 10-0 or 9-1. But if the engine used is excessively weak, even that would be meaningless. Fairy-Stockfish is not automatically strong in every variant you configure it for, just because Stockfish is strong at orthodox Chess. It might have a world-class search, but strength depends on evaluation as well as search. With faulty evaluation good search only backfires, as the engine gets more clever in finding ways to force losing trades, or create trouble for itself in other ways. And the heavy pruning makes it blind for the lines that are actually winning, as it considers those poor play, and thus irrelevant.

QueeNy uses piece values Q=9.5 and N=5, which is not optimal with fewer Knights (where de Q/N ratio quickly goes up), but good enough to avoid getting there. But even then, to effectively clobber Stockfish required removal of the normal King Safety evolution. It could even beat Stockfish and many other top engines of those days with 6N vs 3Q (which is theoretically lost). The top engines never took the opportunity for making a 1-for-2 trade, (the first step for whie towards the win) until it was too late, and they had to sac 1-for-1.