Poll test

[hgm]

I think Chrome caches "deep code" code that's not on the immediate page for speed.

Indeed, FireFox does that too. And the problem was that the page links to some JavaScript on my website for substituting the [ d] and [ pgn] BBcodes in postings from before the upgrade by the corresponding diagrams. But this interfered with the vote submission, and I now changed it so that it no longer does. But because it is an off-page script it would not be automatically reloaded when it is already cached. So the fix only becomes active when you delete the old, faulty script from the cache.

[Ovyron]

People don't realize how fragile NNUE technology is. It was based on a genius idea, but tomorrow someone could have a different idea and when people switch to it NNUE could be obsolete and dead! It's just the best thing we have right now, but it has reached its limits and I don't see it getting any faster.

A classical eval will always be faster than NNUE, and remember, classical eval would have kept on improving if NNUE never existed and caught the second release of NNUE after two years, now that NNUE has stopped being a moving target people in the shadows could be improving classical eval and eventually surpass NNUE, by using classical eval's speed advantage, and because we can always see where classical eval and NNUE's eval vary in a position, all we need is a human that understands why, and if that person isn't on the Stockfish's team, get ready to see them dethroning Stockfish once and for all.

If this happens, ironically, the technology that saved 2 years of work for Stockfish will have been its downfall, as they halted seeking for the improvements that didn't depend on NNUE, but we don't know if other teams kept searching, because NNUE's slowness is its Achilles's heel.

[smatovic]

...

My three cents on this:

"I see..."
https://eta-chess.app26.de/post/i-see-/

"The next step for LC0?"
https://eta-chess.app26.de/post/eta-v06 ... p-for-lc0/

"Transhuman Chess with NN and RL..."
https://eta-chess.app26.de/post/transhu ... n-and-rl-/

--
Srdja

[Ovyron]

My three cents on this:

So what about a future without neural networks? At least not used on the final product. Currently the improvement we're seeing is done by humans that have ideas for new code, that code is tested and does better than the old code, and the engine improves.

What we need is, instead of a nn that improves game results and chess moves, one that improves the engine's code. That way we could have back a regular eval with full speed that uses it to beat the strongest net we have, because we'd use a neural network to program that classical eval.

[smatovic]

My three cents on this:

So what about a future without neural networks? At least not used on the final product. Currently the improvement we're seeing is done by humans that have ideas for new code, that code is tested and does better than the old code, and the engine improves.

What we need is, instead of a nn that improves game results and chess moves, one that improves the engine's code. That way we could have back a regular eval with full speed that uses it to beat the strongest net we have, because we'd use a neural network to program that classical eval.

I guess that is the next level, when a NN/AI starts to program a chess engine, a Technological Singularity Feedback Loop.

--
Srdja