Your network blocks the Lichess assets!

lichess.org
Donate

Do Stronger Players Play More Accurately? I Analyzed 35,000 Games to Find Out

@Gleb_Momot said in #30:

@manoale50 disse em #23:

Yes, we should all take it as a joke, but what's the problem with ChatGPT? We love Stockfish, Maia, and some cheaters who don't know anything use the expression "it seems like the post/comments were generated by AI." I'm not saying that you or anyone else here is a cheater, so far no one has used that expression, but in other posts, yes, I insist that this is not the defense of a doctoral thesis... Should we stop posting or asking questions for fear of offending someone who feels diminished by us discussing anything about chess? Isn't that the purpose of the forum?

@Gleb_Momot said in #30: > @manoale50 disse em #23: Yes, we should all take it as a joke, but what's the problem with ChatGPT? We love Stockfish, Maia, and some cheaters who don't know anything use the expression "it seems like the post/comments were generated by AI." I'm not saying that you or anyone else here is a cheater, so far no one has used that expression, but in other posts, yes, I insist that this is not the defense of a doctoral thesis... Should we stop posting or asking questions for fear of offending someone who feels diminished by us discussing anything about chess? Isn't that the purpose of the forum?

@manoale50 said in #31:

Yes, we should all take it as a joke, but what's the problem with ChatGPT? We love Stockfish, Maia, and some cheaters who don't know anything use the expression "it seems like the post/comments were generated by AI." I'm not saying that you or anyone else here is a cheater, so far no one has used that expression, but in other posts, yes, I insist that this is not the defense of a doctoral thesis... Should we stop posting or asking questions for fear of offending someone who feels diminished by us discussing anything about chess? Isn't that the purpose of the forum?

I got hated on my blogs using Chat GPT, by people saying that using Chat GPT is not allowed on Lichess, and I dont know if that is true, and I want to make a person warning just like other do to me, that maybe you may even get banned.. Several people said to me that using Chat GPT in forum posts, and blogs may lead to ban.

@manoale50 said in #31: > Yes, we should all take it as a joke, but what's the problem with ChatGPT? We love Stockfish, Maia, and some cheaters who don't know anything use the expression "it seems like the post/comments were generated by AI." I'm not saying that you or anyone else here is a cheater, so far no one has used that expression, but in other posts, yes, I insist that this is not the defense of a doctoral thesis... Should we stop posting or asking questions for fear of offending someone who feels diminished by us discussing anything about chess? Isn't that the purpose of the forum? I got hated on my blogs using Chat GPT, by people saying that using Chat GPT is not allowed on Lichess, and I dont know if that is true, and I want to make a person warning just like other do to me, that maybe you may even get banned.. Several people said to me that using Chat GPT in forum posts, and blogs may lead to ban.

@Gleb_Momot said in #32:

I got hated on my blogs using Chat GPT, by people saying that using Chat GPT is not allowed on Lichess, and I dont know if that is true, and I want to make a person warning just like other do to me, that maybe you may even get banned.. Several people said to me that using Chat GPT in forum posts, and blogs may lead to ban.

I really understand your position now, it makes sense even to me, they are the purists, they think we should go back to using typewriters, because all these new things are diminishing them, anyway, for problems of insecurity and EGO there is no solution in sight, not even chatGPT solves it, Freud explains...

@Gleb_Momot said in #32: > I got hated on my blogs using Chat GPT, by people saying that using Chat GPT is not allowed on Lichess, and I dont know if that is true, and I want to make a person warning just like other do to me, that maybe you may even get banned.. Several people said to me that using Chat GPT in forum posts, and blogs may lead to ban. I really understand your position now, it makes sense even to me, they are the purists, they think we should go back to using typewriters, because all these new things are diminishing them, anyway, for problems of insecurity and EGO there is no solution in sight, not even chatGPT solves it, Freud explains...

Hello,
No, these are real statistical data. The reproduction method was explained very clearly, and it’s easy to reproduce if you don’t believe it. I have personally reproduced it several times on different datasets from different sources, and the conclusions were always the same.

GPT was used for two things:

To generate the image of the horse covering its eyes

To correct my spelling and grammar mistakes in English, since I’m not very good at it (it’s not my native language)

Regarding the curve: like most data scientists, I use the Python library NumPy. Since it’s the reference library for data science, if you ask GPT to create a graph, it won’t “generate an image” directly, but instead it will write Python code and execute it. That Python code will almost certainly use NumPy, because it’s the standard library for this purpose.

Once again, skepticism here is a good thing—it’s part of the job to challenge and question the results of your peers. But please, reproduce the experiment before drawing any conclusions, as is the proper practice.

Hello, No, these are real statistical data. The reproduction method was explained very clearly, and it’s easy to reproduce if you don’t believe it. I have personally reproduced it several times on different datasets from different sources, and the conclusions were always the same. GPT was used for two things: To generate the image of the horse covering its eyes To correct my spelling and grammar mistakes in English, since I’m not very good at it (it’s not my native language) Regarding the curve: like most data scientists, I use the Python library NumPy. Since it’s the reference library for data science, if you ask GPT to create a graph, it won’t “generate an image” directly, but instead it will write Python code and execute it. That Python code will almost certainly use NumPy, because it’s the standard library for this purpose. Once again, skepticism here is a good thing—it’s part of the job to challenge and question the results of your peers. But please, reproduce the experiment before drawing any conclusions, as is the proper practice.

@lucb3 if you are struggling to write a cool blog, because your English is not good enough, or you are not enough confident to make a cool image, of statistics it's ok, I just want to notify you that it might be a risk. I still believe that the statistics image are made directly by Chat GPT.

@lucb3 if you are struggling to write a cool blog, because your English is not good enough, or you are not enough confident to make a cool image, of statistics it's ok, I just want to notify you that it might be a risk. I still believe that the statistics image are made directly by Chat GPT.

I'd like to suggest another important topic deserving deep statistical analysis - do 100m Gold medalists run faster than the rest of the competitors?

I'd like to suggest another important topic deserving deep statistical analysis - do 100m Gold medalists run faster than the rest of the competitors?

Understanding Precision

In this context, "precision" refers to how closely a player's moves align with the optimal moves suggested by Stockfish. For each move, Stockfish evaluates the position before and after the move, quantifying any loss in advantage. The precision score reflects the average of these evaluations throughout the game, providing a percentage that indicates the overall accuracy of play. A score of 100% would mean every move matched the engine's top choice.

The above from the blog is the weakness in the results. The problem is that if there's only also 1 legal move then you match engine's top choice. In addition to things like opposite square bishop endgames where any move other than hanging bishop draws you'll also effortlessly match engine top move. This means that precision is a combination of accuracy and the positions reached in sample data so it's 2 variables rather than 1.

Precision would probably be more meaningful if you filtered all positions and only used ones where 3rd best move was evaluated as being at least 0.25 pawns worse than best move or found identical positions and compared how people of various ratings handled them. Exact figures could be tweaked but it would filter out moves in dead drawn positions along with only forced move positions where it's literally impossible to blunder. It would also be prudent to filter out repeated positions unless a different move is played. Since engine evaluates to depth 13 you may potentially get cases where if one move is a mate in 13 and other is a mate in 14 that it flags the mate in 14 as an error but that would apply to any depth unless you chose to limit it to 7 piece positions where tablebase can be used.

Having said that I think you would attain similar results with those changes as error rate is likely to be less important than rating impact and could be relatively uniform among rating ranges (I'm unsure if anyone has done stats to see if 100% accuracy forced moved are more common in any rating range). So while a 400 player spending 50 moves failing to convert a K+R vs K endgame might end up with really high accuracy in one game on average it wouldn't have a huge impact with enough games and higher rated games could also be boosted with people stubbornly defending opposite square bishop endgames etc. Main point is it's important to understand how precision works as otherwise it can be manipulated with edge cases.

Understanding Precision In this context, "precision" refers to how closely a player's moves align with the optimal moves suggested by Stockfish. For each move, Stockfish evaluates the position before and after the move, quantifying any loss in advantage. The precision score reflects the average of these evaluations throughout the game, providing a percentage that indicates the overall accuracy of play. A score of 100% would mean every move matched the engine's top choice. The above from the blog is the weakness in the results. The problem is that if there's only also 1 legal move then you match engine's top choice. In addition to things like opposite square bishop endgames where any move other than hanging bishop draws you'll also effortlessly match engine top move. This means that precision is a combination of accuracy and the positions reached in sample data so it's 2 variables rather than 1. Precision would probably be more meaningful if you filtered all positions and only used ones where 3rd best move was evaluated as being at least 0.25 pawns worse than best move or found identical positions and compared how people of various ratings handled them. Exact figures could be tweaked but it would filter out moves in dead drawn positions along with only forced move positions where it's literally impossible to blunder. It would also be prudent to filter out repeated positions unless a different move is played. Since engine evaluates to depth 13 you may potentially get cases where if one move is a mate in 13 and other is a mate in 14 that it flags the mate in 14 as an error but that would apply to any depth unless you chose to limit it to 7 piece positions where tablebase can be used. Having said that I think you would attain similar results with those changes as error rate is likely to be less important than rating impact and could be relatively uniform among rating ranges (I'm unsure if anyone has done stats to see if 100% accuracy forced moved are more common in any rating range). So while a 400 player spending 50 moves failing to convert a K+R vs K endgame might end up with really high accuracy in one game on average it wouldn't have a huge impact with enough games and higher rated games could also be boosted with people stubbornly defending opposite square bishop endgames etc. Main point is it's important to understand how precision works as otherwise it can be manipulated with edge cases.

@CarlosEstevez said in #37:

I'd like to suggest another important topic deserving deep statistical analysis - do 100m Gold medalists run faster than the rest of the competitors?

If you just look at the 100m Gold final, yes, but if you follow it from the qualifiers 1 year before the final, you will notice that just like chess, they also have 55% performance, this 55% is not a problem in chess, it happens in all sports, because it is basic mathematics, the famous rule of three, only engines can currently be above 55% and when they play against humans, among themselves, the same percentage remains... and no, this comment was not made by chatGPT, lol...

@CarlosEstevez said in #37: > I'd like to suggest another important topic deserving deep statistical analysis - do 100m Gold medalists run faster than the rest of the competitors? If you just look at the 100m Gold final, yes, but if you follow it from the qualifiers 1 year before the final, you will notice that just like chess, they also have 55% performance, this 55% is not a problem in chess, it happens in all sports, because it is basic mathematics, the famous rule of three, only engines can currently be above 55% and when they play against humans, among themselves, the same percentage remains... and no, this comment was not made by chatGPT, lol...