TalkChess.com

I definitely see the value of "Let's Check" for very serious tournament players and sabermetric chess fans.

For me it does not hold much interest as it takes away some of what I find fun about chess and chess engines. Analyzing myself (rather weakly) and then checking it with some engines. I like that act. Checking with "Let's Check" would take some of that fun away for me.

But I do see it's value for various others.

Although technology won't yield to my wishes, I do hope chess doesn't come close to being solved in my lifetime. The various choices and possibilities and the as yet unkown are what make it fun for me.

Nelson Hernandez wrote:You laugh, but that only means you don't get it. We all know there is a finite but gigantic number of possible chess positions. This project does not mean to analyze every possible position: it cannot. It intends to analyze positions that have actually happened, and beyond that, positions that people contribute that are extensions of what has already happened. I can tell you that in terms of what positions have actually happened N>1 times that a very large database, i.e. something bigger than Chessbase's, would at this time consist of a few hundred million positions.

Now, it is very, very possible for any one person with enough dedicated hardware to analyze a few hundred million positions to some shallow depth, maybe one or two seconds per position. That's a provable fact. And if that is possible, then imagine what is possible if a lot of people work in a semi-coordinated, let alone systematic, manner. And then imagine what happens as the number of cores on a cheap machine multiplies again and again.

I don't quite get it. Few hundred million positions finally - yes, the present large human databases have several dozen million positions, and for a human player in the opening are probably more useful. If engine analysis is 1-2s, who needs it? Can't I just use infinite or deep analysis whenever I want for several seconds for every possible position? You claim that "any one person" can use ~1 billion seconds CPU time on some shallow analysis. First, 1 billion seconds CPU time even on 4 CPU machine means 10 years of publishable, unique, continuous analysis 24/24h, which is unrealistic. Second, who needs this shallow blunder check of opening positions? If you mean playchess engine room, then I might agree, but certainly not the majority of human players are in a serious need of this opening database.

Kai

Laskos wrote:

Nelson Hernandez wrote:You laugh, but that only means you don't get it. We all know there is a finite but gigantic number of possible chess positions. This project does not mean to analyze every possible position: it cannot. It intends to analyze positions that have actually happened, and beyond that, positions that people contribute that are extensions of what has already happened. I can tell you that in terms of what positions have actually happened N>1 times that a very large database, i.e. something bigger than Chessbase's, would at this time consist of a few hundred million positions.

Now, it is very, very possible for any one person with enough dedicated hardware to analyze a few hundred million positions to some shallow depth, maybe one or two seconds per position. That's a provable fact. And if that is possible, then imagine what is possible if a lot of people work in a semi-coordinated, let alone systematic, manner. And then imagine what happens as the number of cores on a cheap machine multiplies again and again.

I don't quite get it. Few hundred million positions finally - yes, the present large human databases have several dozen million positions, and for a human player in the opening are probably more useful. If engine analysis is 1-2s, who needs it? Can't I just use infinite or deep analysis whenever I want for several seconds for every possible position? You claim that "any one person" can use ~1 billion seconds CPU time on some shallow analysis. First, 1 billion seconds CPU time even on 4 CPU machine means 10 years of publishable, unique, continuous analysis 24/24h, which is unrealistic. Second, who needs this shallow blunder check of opening positions? If you mean playchess engine room, then I might agree, but certainly not the majority of human players are in a serious need of this opening database.

Kai

There is a Slav anti-Moscow line which is played in 5 games out of 15 in some tournament. Anand drops a novelty there, and 100,000 people around the world from 1800 to 2800 start analyzing that novelty for a quite some time. The idea here is that there is awful lot of repetition in the analysis, and if some of the people would like to share theirs, others would benefit.

You can question why would someone share their analysis, how can you be sure in the quality, how do you actually map results across hardware and different engines, and so on and so on.

But obviously the part that there are a lot of people analyzing the exact same position in a period of time is true.

If engine analysis is 1-2s, who needs it? Can't I just use infinite or deep analysis whenever I want for several seconds for every possible position? You claim that "any one person" can use ~1 billion seconds CPU time on some shallow analysis. First, 1 billion seconds CPU time even on 4 CPU machine means 10 years of publishable, unique, continuous analysis 24/24h, which is unrealistic. Second, who needs this shallow blunder check of opening positions? If you mean playchess engine room, then I might agree, but certainly not the majority of human players are in a serious need of this opening database.

...

From a tiny acorn does a mighty oak grow. What you call "unrealistic" is nonetheless possible; all that is needed is some proper scaling in terms of the scope of the job (# positions, # depth analysis). Maybe you wouldn't buy an expensive server and use it that way but other people might.

In terms of its value, look at Aquarium. A broadband analysis that looks at every historical variation, which is then minimaxed, can be a hell of a powerful tool. Not remotely close to infallible, of course, full of weaknesses--but perfection isn't the standard. Don't let perfection become the enemy of very good!

As for the value to human players, that will vary from player to player. For 2200 level players this will be the best training tool ever. For 2800 players you need analysis in the realm of what has never been seen before instead of the finite boundaries of what has already been seen.

The biggest problem with this approach is that while right now both information and misinformation are decentralized, we would now be going in the direction of centralizing both. There is a huge aggregate benefit from the information dissemination, but there will be some who derive benefits from exploiting misinformation (gross evaluative inaccuracies) in the system that has become conventional wisdom.

Nelson Hernandez wrote:If engine analysis is 1-2s, who needs it? Can't I just use infinite or deep analysis whenever I want for several seconds for every possible position? You claim that "any one person" can use ~1 billion seconds CPU time on some shallow analysis. First, 1 billion seconds CPU time even on 4 CPU machine means 10 years of publishable, unique, continuous analysis 24/24h, which is unrealistic. Second, who needs this shallow blunder check of opening positions? If you mean playchess engine room, then I might agree, but certainly not the majority of human players are in a serious need of this opening database.

...

From a tiny acorn does a mighty oak grow. What you call "unrealistic" is nonetheless possible; all that is needed is some proper scaling in terms of the scope of the job (# positions, # depth analysis). Maybe you wouldn't buy an expensive server and use it that way but other people might.

Am I crazy to buy an expensive server for this thing? I guess the return from the computing power generally in chess is logarithmic, and for finding novelties, traps or refutations in openings is a square root of a logarithm. So the 100 core system means only ~2 times more novelties and such than 1 core system. At the same time, my pocket deflator is linear with the number of cores, if not exponential

In terms of its value, look at Aquarium. A broadband analysis that looks at every historical variation, which is then minimaxed, can be a hell of a powerful tool.

In building playchess opening books.

Not remotely close to infallible, of course, full of weaknesses--but perfection isn't the standard. Don't let perfection become the enemy of very good!

I did not want perfection, just some usefulness of this opening database to humans. My guess is that after an initial enthusiasm, people will go back to old databases.

As for the value to human players, that will vary from player to player. For 2200 level players this will be the best training tool ever. For 2800 players you need analysis in the realm of what has never been seen before instead of the finite boundaries of what has already been seen.

My observation was that the novelties and refutations played by humans usually come from strong players or strong teams of players. These novelties and refutations are few and far between, occurring at a much slower pace than even playchess book novelties. I don't think that computers are determining a novelty or a refutation in human openings, it's probably mostly human effort with some help from comps. Why the scene isn't changed much since the advent of very strong comps?

The biggest problem with this approach is that while right now both information and misinformation are decentralized, we would now be going in the direction of centralizing both. There is a huge aggregate benefit from the information dissemination, but there will be some who derive benefits from exploiting misinformation (gross evaluative inaccuracies) in the system that has become conventional wisdom.

That's nice, I didn't even imagined that there could be some saboteurs posting analysis. Let's assume that nobody will sabotage the database, its life will be hard even without any sabotage, and it comes almost for free for some 3 years from Fritz.

Kai

Laskos wrote:

Nelson Hernandez wrote:If engine analysis is 1-2s, who needs it? Can't I just use infinite or deep analysis whenever I want for several seconds for every possible position? You claim that "any one person" can use ~1 billion seconds CPU time on some shallow analysis. First, 1 billion seconds CPU time even on 4 CPU machine means 10 years of publishable, unique, continuous analysis 24/24h, which is unrealistic. Second, who needs this shallow blunder check of opening positions? If you mean playchess engine room, then I might agree, but certainly not the majority of human players are in a serious need of this opening database.

...

From a tiny acorn does a mighty oak grow. What you call "unrealistic" is nonetheless possible; all that is needed is some proper scaling in terms of the scope of the job (# positions, # depth analysis). Maybe you wouldn't buy an expensive server and use it that way but other people might.

Am I crazy to buy an expensive server for this thing?

Assuredly, people such as the elite, who own powerful systems with 32 overclocked cores, will have little use for it, but they represent an absolute microcosm of the players.