Branko Radovanovic wrote: ↑Tue Jun 29, 2021 1:17 pm I see no evidence NNUE is inherently stronger than a (very) highly developed HCE. For Stockfish, the advent of NNUE did deliver an equivalent of 5 or so years of HCE development almost instantly, but for all we know, in a parallel universe where NNUE was never invented there might yet appear a 2025 HCE version of Stockfish which is stronger than the 2025 NNUE version of Stockfish.
You might be right, but I remember Ed telling me that because search had increased so much he was removing knowledge from the eval (this would have been more than 20 years ago). If the search uncovers the knowledge, then removing that knowledge from the eval increases elo by increasing the speed of the eval, and hence increasing the amount of searching that can be done.
Going forward, what is likely to happen is that computers simply become unbeatable (computer v computer would then almost always end in a draw), and the incentive to try to improve the eval (or, indeed, any other aspect of a chess engine) will then be gone. Unfortunately, this is likely to happen before computer evaluation of chess positions is actually solved.
I can see good reasons for interest in chess engines with a search limit (1 ply would be a good limit for eval test), a time limit (anything over a millisecond would be pointless), and a program size limit (so that it doesn't just become a game of who can train the biggest NN). Unfortunately, the interest level in such a game would undoubtedly be much lower than computer chess has attracted in the past.
One other possibility: how about proving that the opening position of chess is a draw? I am in a tiny minority, but, per previous discussions, I think that it could actually be done with today's technologies, and once you can do that, the techniques could then be used to determine whether any position is drawn.