NNUE Poll

[smatovic]

[...]

[...]

I get that people are exited about NNUE, new tech toys for us, like Christmas (and we have an excuse to buy new, geeky, hardware ), but would people be that exited if they run Texel Tuning on an multidimensional PSQT?

--
Srdja

[Patrice Duhamel]

I may try NNUE in future, but there are things I want to try before, and I also lack of time, knowledge on how to implement it, and certainly lack of CPU power to train the network.

[JoAnnP38]

[...]

[...]

I get that people are exited about NNUE, new tech toys for us, like Christmas (and we have an excuse to buy new, geeky, hardware ), but would people be that exited if they run Texel Tuning on an multidimensional PSQT?

--
Srdja

I would be! In fact, the I am planning on adding that to Pedantic in release 0.3 God willing. Of course, I guess then my HCE becomes NNLite and maybe that's the point you are trying to make.

[smatovic]

I would be!

Beautiful, that's the spirit

I find it interesting to utilize the CPU SIMD unit for NN inference, or move generation, a new programming challenge.

--
Srdja

[dangi12012]

NNUE is a really bad term because people think it's one thing. If you read many engines source code you see it's really not.
Even Stockfish went through 5 architectures.

The shared idea is the incremental update for the input layer depending on: color, king SQ, piece, piece SQ. Everything else is different.

So the stigma around using NNUE is not warranted imo - it's another tool in the toolbox and it's really not 1 fixed thing that is set in stone.
Just call it incrementally updated eval NN and use it as drop in piece square table replacement.

"Someone" will soon share a single header file version with no external dependencies and just 100 lines of code or so. Bonus of code simplification is that the compiler emits mostly performant intrinsic like madd_epi16 from the scalar source on its own - so it's reasonably fast.

[Rebel]

NNUE is a really bad term because people think it's one thing. If you read many engines source code you see it's really not.
Even Stockfish went through 5 architectures.

We are already at 22 architectures.

[chrisw]

NNUE is a really bad term because people think it's one thing. If you read many engines source code you see it's really not.
Even Stockfish went through 5 architectures.

We are already at 22 architectures.

Very likely Stockfish went through many more, testing and throwing away. But unless you keep trying, start simple, testing and building, you cannot really hope to comprehend the characteristics of architectural changes, or even get a handle on the their time costs.

[Eelco de Groot]

Making an NNUE evaluation is not really programming. It requires some rather trivial standard code for running the NN. There is no creativity there, I might as well write a program to calculate my tax returns. What is the fun in that? Basically you let the computer write its own engine, and the only contribution you make yourself is paying the electricity bill. In the end you will get something that might play strong chess, but you would still have no clue why or how it does this. So as a method of getting some understanding is is also a bust.

If it is all so simple, why couldn't you get something trivial like this to work than? https://talkchess.com/forum3/viewtopic.php?f=7&t=76465
Maybe it is not so simple as you think it is. In the end all that matters is playing strength, nobody cares if it's been done with HCE, NNUE or whatever.

I am sure that many programmers would agree that it goes deeper than that. Computer chess is, in many senses, for me at least, still a Drosophila and a microcosm for what will happen, or could happen, in our relation with AI and how it could change our existence, fundamentally. As in a meeting with another intelligence. Like in Science Fiction, the Vulcans detecting Faster Than Light travel from the Phoenix engineered by Zephram Cochrane in 2063 and initiating First Contact.
Or the story of Prometheus (I think it was). The gods not wanting people to have fire and challenge them, like the Olympian gods overthrew themselves the Titans. I mean, these are existential questions.

Some related posts:
forum3/viewtopic.php?f=2&t=81876
https://prodeo.actieforum.com/t1150-alw ... ange#12906

"Full disclosure" as Srdja correctly asks for in the above thread, I think should no be some fleeting posts on Discord or even here but at the least some article, like in the IGCA journal. I mean, this could start a whole new science. And if you don't publish, somebody else will claim to be the first. That's how it works. Unfortunately, Ed and Chris are not on the best footing with the IGCA and Jaap van den Herik, if Jaap hasn't retired yet. There are other ways to publish of course. I'm sure Robert Hyatt could help bringing them into contact with the Association for Computing Machinery or some such. Does it still exist... But unfortunately, Ed and Chris are not on the best footing with Bob either... Maybe Peter Berger could help here...

[syzygy]

[...]

[...]

I get that people are exited about NNUE, new tech toys for us, like Christmas (and we have an excuse to buy new, geeky, hardware ), but would people be that exited if they run Texel Tuning on an multidimensional PSQT?

But people have been tuning their parameters for a very long time already (and obviously without any chance of "understanding" why the tuned values are better). This is how you get more Elo.

There are people who are not interested in Elo and never ever considered to tune their parameters, so yeah, then why switch to NNUE.

[smatovic]

I am excited, Uri Blass mentioned that he could read PSQTs and that they made sense to him but not the NNUE weights.

Not only cos of more Elo, more Elo verifies that NNUE is superior, but this is the next step, we programmers start to replace ourselves with a black box, philosophical implication, the Technological Singularity.

***edit***
Maybe it is about the narrative, the pov, for some people it just extended tuning of parameters, for others new tech toys, for others more Elo, for others the takeoff of the TS....

--
Srdja