@gloriavictis101, not sure about those numbers but there is definitely a limit on Maia's capabilities.
Maia by design is a 'weak' chess AI as it completely disregards any depth in it's prediction of the moves. It makes all of its moves only based on the current position, move history and metadata like rating and time left, but it makes no effort to calculate ahead at any depth, unlike any other chess engine, which all employ some search algorithms too calculate ahead. So there would definitely a hard limit on Maia's performance because of this.
Maia is simply designed to match moves made by humans and not necessarily make the best moves possible. And it does it amazingly well. I'm certain it could be extended above the 1900 (actually the 1900 Maia predicts human moves very well into the 2500's, better than any other engine). Making good predictions does not necessarily make it play stronger as a matter of fact they tried to include some models where they included searching of moves ahead but it made their prediction accuracy worse.
I think the main reason they did not do that would probably be best explained by this: lichess.org/stat/rating/distribution/blitz
There is just so much more data available between 1100 and 1900 to have more confidence in their actual research. The higher rated player pool becomes so small that you could introduce certain biases in your models and having a much harder time with error margins.
Maia by design is a 'weak' chess AI as it completely disregards any depth in it's prediction of the moves. It makes all of its moves only based on the current position, move history and metadata like rating and time left, but it makes no effort to calculate ahead at any depth, unlike any other chess engine, which all employ some search algorithms too calculate ahead. So there would definitely a hard limit on Maia's performance because of this.
Maia is simply designed to match moves made by humans and not necessarily make the best moves possible. And it does it amazingly well. I'm certain it could be extended above the 1900 (actually the 1900 Maia predicts human moves very well into the 2500's, better than any other engine). Making good predictions does not necessarily make it play stronger as a matter of fact they tried to include some models where they included searching of moves ahead but it made their prediction accuracy worse.
I think the main reason they did not do that would probably be best explained by this: lichess.org/stat/rating/distribution/blitz
There is just so much more data available between 1100 and 1900 to have more confidence in their actual research. The higher rated player pool becomes so small that you could introduce certain biases in your models and having a much harder time with error margins.