It's NNUE era (sharing my thoughts)

[Dann Corbit]

1 ShashChess 15.0 64-bit 4CPU 3534 +27 −26 76.5% −175.1 46.9% 490
198‑199 OliThink 5.6.9 64-bit 2528 +27 −27 43.6% +46.2 30.3% 489
3534-2528=1006 Elo.

The eval of Olithink is mobility and a few simple tables.
Running on one thread, it is within one thousand Elo of the strongest program on the list:
https://ccrl.chessdom.com/ccrl/4040/

[Dann Corbit]

One should not imagine that I am an opponent or I dislike someone or their work because I do not agree with them.
I may very well admire both them and their work.

"The way I see it, an intelligent person who disagrees with me is
probably the most important person I'll interact with on any given
day."
--Billy Chambless

[tomitank]

You have to labour for the NNUE eval too. Unless you are a coward and reuse the code and weights that Stockfish already gave you. NNUE is not a magic bullet -- Stockfish NNUEs are a magic bullet. Everyone who has tried to replicate NNUE on their own will know this.

@Dann Corbit:
This is the most important part of what we have to say.

[Dann Corbit]

You have to labour for the NNUE eval too. Unless you are a coward and reuse the code and weights that Stockfish already gave you. NNUE is not a magic bullet -- Stockfish NNUEs are a magic bullet. Everyone who has tried to replicate NNUE on their own will know this.

@Dann Corbit:
This is the most important part of what we have to say.

Are you afraid to use the other data that SF produces? If it writes an EPD string will you use it?
Now, according to my understanding, the original SF NNUE creation code did not use its own eval at all. So it is the search of Stockfish producing the great nodes. Later, they used the SF eval because of the partially relying on SF eval for lopsided positions and it broke something. I do not know if they got that sorted out or not.

At any rate, I do not see data written by a chess program as sacred either. If you do, then I suggest you do not use it.

Of course if you use the data (and maybe if you do not) you will be banned from tournaments. This is ignorance on the part of the tournament organizers but so be it.

So if it is important for you to be allowed to play in tournaments then I suggest you use your own engine to create the data nodes. When your search is as good as Stockfish's search, then your nodes will be as good also.

[tomitank]

You have to labour for the NNUE eval too. Unless you are a coward and reuse the code and weights that Stockfish already gave you. NNUE is not a magic bullet -- Stockfish NNUEs are a magic bullet. Everyone who has tried to replicate NNUE on their own will know this.

@Dann Corbit:
This is the most important part of what we have to say.

Are you afraid to use the other data that SF produces? If it writes an EPD string will you use it?
Now, according to my understanding, the original SF NNUE creation code did not use its own eval at all. So it is the search of Stockfish producing the great nodes. Later, they used the SF eval because of the partially relying on SF eval for lopsided positions and it broke something. I do not know if they got that sorted out or not.

At any rate, I do not see data written by a chess program as sacred either. If you do, then I suggest you do not use it.

Of course if you use the data (and maybe if you do not) you will be banned from tournaments. This is ignorance on the part of the tournament organizers but so be it.

So if it is important for you to be allowed to play in tournaments then I suggest you use your own engine to create the data nodes. When your search is as good as Stockfish's search, then your nodes will be as good also.

Here is my solution:
http://talkchess.com/forum3/viewtopic.php?f=2&t=76359

I am as far from the SF solution as possible. Obviously, now all NN networks are suspicious after SF NNUE.

[maksimKorzh]

I agree with Andrew. Evaluation is the soul of the engine. LMR, Null move, etc is nothing without eval.

The real power of the modern chess engine is the branching factor, not the eval.

If that is the case, why are people taking SF's NNUE? The power is not in the eval you say, yet people gain hundreds to a thousand elo by copying it. Why don't we all just copy the Stockfish search tables? I'm so sick of "engine enthusiasts" that know nothing of what they speak.

I debate with myself what I should have done with Ethereal. Whether I should have sold copies. Or whether I should sell nets in the future. And my reoccurring thought is that Ethereal is at the top end, but it will never break the Leela / SF market hold. The only people using Ethereal are people that are quite invested in computer chess, and thus the only buyers would be the kinds of people on talkchess. And those people are the people who contribute ideas and insights that make things like Ethereal possible. But that is not true anymore.

But then I think of the sheer stupidity, ignorance, and otherwise degenerate views that so many on these forums hold. I don't think I view anyone on this forum as my friend anymore. I don't contribute my ideas to the programming section anymore. I've found the dozen or so people worth something on this forum, and collected them into a nice little discord. Maybe its time the Shut It Down movement goes all the way and burns this cesspool to the ground like it deserves.

NNUE added 100 Elo, not thousands.
The eval of Olithink would fit on a single page of paper.
Yet with a single thread it is within a thousand Elo of Stockfish

Wrong! Facts! Simple!

+180 to Ethereal and Komodo
+250 to Igel and Ruby
+300 to Andscas
+1000 to BBC

Only around 600 to BBC
And it's time to forget that experiment.
I'm no longer maintaining that.
That was just an experiment.
I'm now learning all from scratch.
Now playing around with Texel's tuning.
It would be much better to provide beginner's materials helping us noobs to learn.
No one would copypaste sf nnue if knows how to make your own.
Well at least it's true for me.

[JohnWoe]

Double standards.
You shall earn every +Elo with blood, sweat and tears!

Somehow it's ok to copy-paste syzygy egtb / probing code in folder.
But not NNUE.

Wait you still need to fish out the syzygy probing code from Stockfish. Also download 17TiB of 7 men somehow.

The problem is:
Except syzygy gives you +5 Elo.
NNUE gives +900 Elo.

The real improvement came from Shogi guys. Without them we would be squeezing +0.2 Elo out from razoring.

I use LMR, Null move, Magic bitboards, UCI, Qsearch, IID...
I have invented absolutely nothing new.

[dkappe]

Double standards.
You shall earn every +Elo with blood, sweat and tears!

It’s changed the engine improvement workflow and skill sets. To develop a 3000+ engine, you had to implement a good search and a good eval, then tune them so they worked well together.

With NNUE, you just can focus on search and learn a few tricks to train up a decent net. Gone is the HCE glory.

Now 3300 is the new 3000. In a few months, everyone will be looking for ways to goose the nets — through size, architecture, etc. Those that use existing tools like tensorflow and pytorch instead of developing them from scratch will win the race(s).

[jdart]

It is pretty well established I think that both search and eval are important. Stockfish is strong because it is especially strong in both areas. But a good eval is very important.

[Angle]

I only want to say that "NNUE Era" and "Copy-Paste SF NNUE Era" are two basicaly different rhings. The first is beautiful while the second is unworthy. In many discussions here, these two concepts are not separated, which is a great mistake.